Seminars

Event : Thesis Defence
Title : Techniques for estimating the direction of pointing gestures using depth images in the presence of orientation and distance variations from the depth sensor
Speaker : Shome Subhra Das
Degree Registered :PhD<
Date : 17/11/2021
Venue : Online
Abstract : Currently, we interact with computers, robots, drones, and virtual reality interfaces using pointing devices such as mouse, touchpad, joystick, virtual reality wand, drone controller, etc. These devices have one or more of the following limitations: being cumbersome, non-immersive, immobile, and having a steep learning curve. The target of this work is to explore ways to replace existing pointing devices with pointing gesture-based interfaces.
This thesis addresses two problems, namely estimating the direction being indicated by a pointing gesture (PDE) and detection of pointing gestures. The proposed techniques use a single depth sensor and use only the hand region. To our knowledge, this is the maiden attempt at creating depth and orientation tolerant, accurate methods for estimating the pointing direction using only depth images of the hand region. The proposed methods achieve accuracies comparable to or better than those of existing methods while avoiding their limitations.
Significant contributions of the thesis:

Event : Seminar
Title :
Speaker : Dr. Kannan Tirugnanam
Date : 20/20/2021
Venue : Online
Abstract :
Speaker Biodata :

Event : Thesis Colloquium
Title : Modelling, Stabilization Methods and Power Amplification for Power Hardware-in-Loop Simulation with Improved Accuracy
Speaker : Kapil Upamanyu
Degree Registered :PhD
Advisor : Prof. G Narauyanan
Date : 13/10/2021
Venue : Online
Abstract : Simulations of physical systems are conducted extensively for research and design purpose. Potential of a simulation can be extended significantly by conducting it in real-time. Real-time simulation allows for hardware-in-loop (HIL) simulation of a system, i.e., a part of the system represented by a mathematical model in conventional simulation is replaced by a physical hardware. When the power level of the physical hardware is considerably higher than that of RTS signals, the output of RTS is applied to the physical hardware through a power amplifier (PA), and such an HIL is known as power HIL (PHIL). PHIL simulation allows for the testing of a physical hardware (device under test), in a safe and controlled environment, without the rest of the system being available. Among other factors, the dynamics of the PA, which are not present in the actual system, cause the response of the PHIL simulation to differ from that of the actual system. Conventional switched-mode PAs, which utilize passive filters to attenuate the switching harmonics, have limited dynamic response. These PAs, when employed in a PHIL simulation, are unable to accurately replicate the fast transients of the system. A high-bandwidth switched-mode PA without output filter is proposed for improved accuracy of PHIL simulations. Even with fast PAs, inaccuracies are present in a PHIL simulation due to finite computation time of RTS, the transport lag of signals and the sampling effects of RTS. The inaccuracy can be so significant that the PHIL simulation of a system can be unstable even though the actual system is stable, and vice versa. A discrete-time domain modelling is proposed, which estimates the instability of a PHIL simulation much better than the conventional continuous-time domain approach. An unstable PHIL simulation can be stabilized by employing compensation algorithms. Novel compensation algorithms are proposed for stabilizing those PHIL simulation which cann ot be stabilized using conventional compensation algorithms. An output filter-less voltage source inverter is proposed as a PA suitable to be interfaced with inductive loads (e.g., most of the power system loads). Such a PA has a reference tracking bandwidth comparable to the switching frequency. Unlike the output of the conventional PAs, the output of the proposed PA is completely unaffected by the sudden changes in the current drawn by the loads. The proposed PA, realized through an IGBT-based converter stack, is utilized to emulate the transients of synchronous generator, including the fast transient corresponding to the field excitation controller, while feeding a passive linear load. A modified synchronous generator emulation method is proposed, which is capable of replicating the stator transients accurately for balanced, unbalanced and non-linear loads. The applicability of the proposed PA is extended for it to be interfaced to PWM converters by proposing an in-phase synchronization of PWM carriers of the PA and the converter under test. The PA is utilized for emulating unbalanced and harmonic (up to 23rd order) grid voltages while testing the control of a three-phase 415 V, 3kW PWM rectifier. Accurate current responses are also obtained when the step changes in the grid voltage and the rectifier dc bus reference are considered. For various applications, a PA is required to have power sinking capability, which can be achieved by supplying it from a grid-connected bidirectional PWM rectifier. A simple input voltage sensor-less vector control of PWM rectifier is proposed. While the performance of the proposed method, in terms of THD and power factor, is comparable to the sensor-based method and existing sensor-less methods, its computation time requirement is much lower than those for these methods. The proposed control is validated through simulations and experiments on a three-phase 415 V, 3 kW grid-connected PWM rectifier, generating 800 V dc supply. A discrete-time domain modelling of the PI-controlled current loop of PWM converters, which is accurate even when the bandwidth of the loop is close to the switching frequency, is presented. Conventionally, the continuous-time domain models approximate the time delay in the system and hence are inaccurate for high-bandwidth current loops. The discrete-time domain model is used to derive closed-form time-domain expressions of the current for step changes in the current reference and the disturbance voltage, for a given set of controller and hardware parameters. Based on the derived expressions, a pre-filter to the PI-controlled current loop is proposed to achieve the dead-beat response (two-switching cycle settling time) for the reference tracking, while having a settling time much lesser than those with the existing dead-beat control methods for a step change in the disturbance. Conventionally, continuous-time domain methods are used to model PHIL simulations. For PHIL simulation having fast PAs, the effects of sampling at the input and zero-order hold at the output of RTS are significant and are not adequately modelled through continuous-time domain approach. Since, a PHIL simulation consists of discrete-time RTS, a more accurate discrete-time domain modelling approach is proposed. The proposed model is used to conduct stability analysis of a PHIL simulation. The stability limits in terms of the parameters of the simulated and physical quantities are more accurately estimated through the proposed method as compared to the conventional methods. The stability limits of PHIL simulation are verified through simulations and experiments. Compensation algorithms, such as feedback current filtering method, are employed for stabilizing an unstable PHIL simulation. The proposed modelling method is further utilized to conduct the stability analysis of PHIL simulation employing existing compensation algorithms. The analysis provides the stability limits in terms of the cut-off frequency of the filters used for the compensation. The proposed discrete-time model can also be used to design the parameters of the compensation. The effectiveness of the proposed model in predicting the stability of PHIL simulation and the design of compensation are validated through simulations and experiments. Further, novel compensation algorithms, based on lag compensator and cross coupled compensator, are proposed. These are used to stabilize PHIL simulations which are originally unstable without and with existing compensation algorithms. A novel fault-tolerant synchronous inverter, i.e., a renewable energy fed grid-connected voltage source inverter emulated as a synchronous generator, is realized based on the proposed cross-coupled compensator.
Speaker Biodata :

Event : Seminar
Title : Sustainability through High Voltage Engineering and Research
Speaker : Dr. Prem Ranjan
Date : 14/10/2021
Venue : Online
Abstract : Through this talk, we will take a glance at applications of high voltage engineering in three different sustainable technologies. (a) Direct application of high voltage and pulsed power will be shown for economical generation of nanoparticles (NPs) through wire explosion process (WEP). Control of NPs size, phase and formation mechanism will be discussed through modelling studies and different material characterisation techniques. Application of WEP-synthesized semiconductor NPs will be discussed for wastewater treatment. (b) Then, we will go through the need of high voltage electric system in more electric aircraft (MEA) to reduce the carbon footprint. Different tools available to evaluate the arc faults and damage caused to the neighbouring systems will be detailed through mathematical and experimental tools. (c) Drive towards sustainable environment is leading to search of SF6 alternatives in power equipment, which are responsible for more than 80% of total SF6 emission. Research towards SF6 alternatives in gas insulated systems to reduce the global warming potential will be discussed in brief. Finally, the prospective of high voltage engineering and research in some other areas will be discussed.
Speaker Biodata : Prem Ranjan is working as a postdoc researcher at High Voltage Lab, The University of Manchester, UK, since Nov. 2019. He obtained the B. Tech. degree in Electrical and Electronics Engineering from NIT Calicut, India in 2015 and the integrated MS, PhD degrees in Electrical Engineering from IIT Madras, Chennai, India in 2019. He worked as an exchange researcher at Nagaoka University of Technology, Japan for 4 months during 2016 and 2017. His research interests are focused on sustainable applications of high voltage engineering including exploding wire, gas insulation (SF6 alternatives), arc-tracking in more electric aircraft and condition monitoring of power apparatus.

Event : Thesis Defence
Title : Analysis and Enhancement of Stability of Power Systems with Utility-scale Photovoltaic Power Plants
Speaker : Indla Rajitha Sai Priyamvada
Degree Registered : PhD
Advisor : Dr. Sarasij Das
Date : 14/10/20221
Venue : Online
Abstract : Owing to the negative impact of carbon emissions on the environment, power systems are experiencing a paradigm shift in power generation. The fossil fuel-based generators that utilize synchronous machines are increasingly being replaced by the renewables such as Photovoltaic (PV) generators. Utility-scale PV power plants are coming up in the various parts of the world. Power electronic interface, control strategies and lack of inherent rotational element are the main factors that distinguish PV generation from Synchronous Generators (SGs). In addition, the time constants of the PV control loops and Phase Locked Loop (PLL) are of the same order unlike the SGs. The power electronic interface offers a better control over the electrical energy generated by the PV generators. However, the power electronic interface brings new challenges to power system stability. This research work focuses on addressing transient and small-signal stability issues of grid connected utility scale PV power plants.
In conventional power systems, swing equation of SGs and (extended) equal area criterion are used to assess the transient stability of power system. However, the same analysis techniques may not be applicable for PV generators. In this research work, transient stability assessment criteria are developed for grid connected PV generator with two different control strategies viz., Vdc-Q control and PQ control (with/without support functionalities). The proposed criteria are developed considering the outer and inner control loop, PLL and filter dynamics of PV generator. PSCAD simulations are carried out on a two-bus system and a modified IEEE-39 bus system to validate the proposed criterion. The stability criteria are found to effectively assess the stability of grid connected utility scale PV generators.
The power transfer capability of transmission network is limited by thermal limits, voltage limits and stability limits. Power transfer capability of transmission lines emanating from PV generators considering thermal and voltage limits is explored well in the literature. However, there is a lack of literature on stability constrained power transfer capability limit. In this research work, adaptive control-based tuning laws are proposed for grid connected PV generators to improve the stability constrained power transfer capability. The adaptive tuning laws are derived based on the Lyapunov energy function analysis. The Lyapunov functions are formulated using the summation of squares of the PI block errors and difference between the PI parameter values from their optimal values. Time domain simulations are carried out on a two-bus system and a modified IEEE-39 bus system to validate the proposed tuning laws. From time domain simulations, it is observed that the proposed tuning laws are found to effectively improve the stability limit on power transfer to the voltage limit.
The increased penetration of PV generations into power systems has also brought qualitative changes in small signal stability of power systems. Two new categories of oscillation modes are introduced into power systems which have participation from PV state variables. As the mode shape of the two new categories of oscillation modes is different from that of SG modes, the power system stabilizer design should be revisited. In this research work, control-based power system stabilizer is developed considering the controllability and observability of the new categories of oscillation modes. The effectiveness of the developed stabilizer in providing sufficient damping to the new categories of oscillation modes is validated through PSCAD simulations on a modified IEEE-39 bus system.
As power systems are large interconnected systems, the increased penetration of PV generation has resulted in notable interaction among PV generators and SGs. Investigation of the interaction among generators is important to understand the dynamic behaviour of overall power system when subjected to disturbances. This research work is carried out to understand the interaction among PV and SGs. The interaction is analysed through investigation of interaction among oscillation modes of PV generation and SG. A mathematical formulation to quantify the interaction among the oscillation modes of PV generations and SGs is proposed. A modified IEEE-39 bus system is considered to carry out the interaction study and validate the results obtained from mathematical formulations.

Event : Thesis Defence
Title : Acoustic-Articulatory Mapping: Analysis and Improvements with Neural Network Learning Paradigms
Speaker : Aravind Illa
Degree Registered :PhD
Advisor : Prof. Prasanta Kumar Ghosh
Date : 29/09/2021
Venue : Online
Abstract : Human speech is one of many acoustic signals we perceive, which carries linguistic and paralinguistic (e.g: speaker identity, emotional state) information. Speech acoustics are produced as a result of different temporally overlapping gestures of speech articulators (such as lips, tongue tip, tongue body, tongue dorsum, velum, and larynx) each of which regulates constriction in different parts of the vocal tract. Estimating speech acoustic representations from articulatory movements is known as articulatory-to-acoustic forward (AAF) mapping i.e., articulatory speech synthesis. While estimating articulatory movements back from the speech acoustics is known as acoustic-to-articulatory inverse (AAI) mapping. These acoustic-articulatory mapping functions are known to be complex and nonlinear.
Complexity of this mapping depends on a number of factors. These include the kind of representations used in the acoustic and articulatory spaces. Typically these representations capture both linguistic and paralinguistic aspects in speech. How each of these aspects contributes to the complexity of the mapping is unknown. These representations and, in turn, the acoustic-articulatory mapping are affected by the speaking rate as well. The nature and quality of the mapping varies across speakers. Thus, complexity of mapping also depends on the amount of the data from a speaker as well as number of speakers used in learning the mapping function. Further, how the language variations impact the mapping requires detailed investigation. This thesis analyzes few of such factors in detail and develops neural network based models to learn mapping functions robust to many of these factors.
Electromagnetic articulography (EMA) sensor data has been used directly in the past as articulatory representations (ARs) for learning the acoustic-articulatory mapping function. In this thesis, we address the problem of optimal EMA sensor placement such that the air-tissue boundaries as seen in the mid-sagittal plane of the real-time magnetic resonance imaging (rtMRI) is reconstructed with minimum error. Following optimal sensor placement work, acoustic-articulatory data was collected using EMA from 41 subjects with speech stimuli in English and Indian native languages (Hindi, Kannada, Tamil and Telugu) which resulted in a total of ~23 hours of data, used in this thesis. Representations from raw waveform are also learnt for AAI task using convolutional and bidirectional long short term memory neural networks (CNN-BLSTM), where the learned filters of CNN are found to be similar to those used for computing Mel-frequency cepstral coefficients (MFCCs), typically used for AAI task. In order to examine the extent to which a representation having only the linguistic information can recover ARs, we replace MFCC vectors with one-hot encoded vectors representing phonemes, which were further modified to remove the time duration of each phoneme and keep only phoneme sequence. Experiments with phoneme sequence using attention network achieve an AAI performance that is identical to that using phoneme with timing information, while there is a drop in performance compared to that using MFCC.
Experiments to examine variation in speaking rate reveal that, the errors in estimating the vertical motion of tongue articulators from acoustics with fast speaking rate, is significantly higher than those with slow speaking rate. In order to reduce the demand for data from a speaker, low resource AAI is proposed using a transfer learning approach. Further, we show that AAI can be modeled to learn acoustic-articulatory mappings of multiple speakers through a single AAI model rather than building separate speaker-specific models. This is achieved by conditioning an AAI model with speaker embeddings, which benefits AAI in seen and unseen speaker evaluations. Finally, we show the benefit of estimated ARs in voice conversion application. Experiments revealed that ARs estimated from speaker independent AAI preserves linguistic information and suppress speaker-dependent factors. These ARs (from unseen speaker and language) are used to drive target speaker specific AAF to synthesis speech, which preserves linguistic information and target speaker’s voice characteristics.

Event : Thesis Colloquium
Title : Robust Nonconvex Penalties for Solving Sparse Linear Inverse Problems and Applications to Computational Imaging
Speaker : Praveen Kumar Pokala
Degree Registered : PhD
Advisor : Prof. Chandra Sekhar Seelamantula
Date : 06/08/2021
Venue : Online
Abstract : Sparse linear inverse problems require the solution to the l-0-regularized least-squares cost, which is not computationally tractable. Approximate and computationally tractable solutions are obtained by employing convex/nonconvex relaxations of the l-0-pseudonorm. One such approximation is obtained by considering the l-1-norm, which is a convex relaxation of the l-0-pseudonorm. However, l-1 regularization is known to result in biased estimates due to over-relaxation of the l-0-pseudonorm but it comes with the advantage of convexity of the regularized least-squares cost. Several nonconvex approximations of the l-0 pseudonorm have been proposed to overcome the bias introduced by the l-1-norm and to ensure better sparsity. However, certain aspects of nonconvex sparse regularization have not been explored. Some of these are as follows:
Nonconvex sparse priors have been explored in the synthesis-sparse framework, but not in the analysis-sparse framework due to the unavailability of proximal operators in closed-form in the analysis setting. Existing nonconvex approaches attach the same regularization weights across all the components of a sparse vector and treat them as fixed hyperparameters. Considering different weights for the entries and adapting them iteratively is likely to result in a superior performance.
Prior learning networks based on deep-unfolded architectures for solving nonconvex penalties have not been explored. This thesis addresses the above aspects in three parts and considers applications to various computational imaging problems.
Part-1: Nonconvex Analysis-sparse Recovery
In this part, we solve the analysis-sparse recovery problem based on three regularization approaches:
Convexity-preserving nonconvex regularization: We propose the analysis variants of the generalized Moreau envelope and generalized minimax concave penalty (GMCP) over a complex domain. Since the cost is a real-valued function defined over a complex domain, it is nonholomorphic, i.e., it does not satisfy Cauchy-Riemann (CR) conditions. To circumvent this problem, we rely upon on Wirtinger calculus to derive the proximal operator for the analysis l-1 prior and develop an efficient optimization strategy employing projected proximal algorithms. The projection transform maps the analysis-sparse recovery problem into an equivalent constrained synthesis-sparse formulation.
Nonconvex sparse regularization: We consider the problem of nonconvex analysis sparse recovery in which the signal is assumed to be sparse in a redundant analysis operator. Standard nonconvex sparsity promoting priors do not have a proximal operator in closed-form under a redundant analysis operator and therefore, proximal approaches cannot be applied directly. This led us to develop two alternatives -- Moreau envelope regularization and projected transformation.
Generalized weighted l-1 regularization: We develop a generalized weighted l-1 regularization strategy, which allows for efficient weight-update strategies for iteratively reweighted l-1-minimization under tight frames. Further, we impose sufficient conditions on the weight function that leads to a reweighting strategy, which follows the interpretation originally given by Candès et al., but is more efficient than theirs. Since the objective function is nonholomorphic, we resort to Wirtinger calculus for deriving the update equations. We develop an algorithm called generalized iteratively reweighted soft-thresholding algorithm (GIRSTA) and its fast variant, namely, generalized fast iteratively reweighted soft-thresholding algorithm (GFIRSTA). We provide convergence guarantees for GIRSTA and empirical convergence results for GFIRSTA.
We demonstrate the efficacy of the proposed regularization strategies in comparison with the benchmark techniques considering compressive-sensing magnetic resonance image (CS-MRI) reconstruction under a redundant analysis operator, more specifically, shift-invariant discrete wavelet transform (SIDWT).
Part-2: Weighted Minimax Concave p-pseudonorm Minimization
In this part, we develop techniques for accurate low-rank plus sparse matrix decomposition (LSD) and low-rank matrix recovery. We proposed weighted minimax-concave penalty (WMCP) as the nonconvex regularizer and show that it admits a certain equivalent representation that is more amenable to weight adaptation. Similarly, an equivalent representation to the weighted matrix gamma norm (WMGN) enables weight adaptation for the low-rank part. The optimization algorithms are based on the alternating direction method of multipliers. The optimization frameworks relying on the two penalties, WMCP and WMGN, coupled with a novel iterative weight-update strategy, result in accurate low-rank plus sparse matrix decomposition and low-rank matrix recovery techniques. Further, we derive an algorithm, namely, iteratively reweighted MGN (iReMaGaN) algorithm, which has a superior low-rank matrix recovery performance. The proposed algorithms are shown to satisfy descent properties and convergence guarantees. On the applications front, we consider the problems of foreground-background separation and image denoising. Simulations and validations on standard datasets show that the proposed techniques outperform the benchmark techniques. Next, we extended the idea to obtain a generalized l-p-penalty, namely, minimax concave p-pseudonorm (MCpN) based on a novel p-Huber function as the sparsity promoting function, and its weighted counterpart, weighted MCpN (WMCpN) as a regularizer for solving the sparse linear inverse problem. WMCpN is a generalization of which several penalties, namely, l-1-norm, minimax concave penalty (MCP), l-p penalty, weighted l-1-norm, and weighted l-p penalty become special cases. However, MCpN and WMCpN regularizers do not have closed-form proximal operators, which makes the optimization problem challenging. To overcome this hurdle, we develop an equivalent representation that is more amenable to optimization and allows for an analytical weight-update strategy. MCpN is a special case of WMCpN where all the weights are fixed and equal. The optimization algorithms are based on the alternating direction method of multipliers. Considering the application of interferometric phase estimation, we demonstrate that MCpN and WMCpN result in accurate interferometric phase estimation. Simulations and experimental validations on standard datasets show that the proposed techniques outperform the benchmark techniques.
Part-3: Nonconvex Sparse Regularization and Deep-Unfolding
In the final part, we transition from fixed analytical priors to data-driven priors. To begin with, we develop a deep-unfolded architecture, namely, FirmNet, for sparse recovery. FirmNet has two parameters -- one that controls the noise variance, and the other that allows for explicit sparsity control. We show that FirmNet is better than Learned-ISTA (LISTA) by at least three-fold in terms of the probability of error in support (PES), and about 2 to 4 dB higher reconstruction SNR. Further, we solve the problem of reflectivity inversion, which deals with estimating the subsurface structure from seismic data through FirmNet. As an application, we consider the problem of seismic reflectivity inversion. We demonstrate the efficacy of FirmNet over the benchmark techniques for the reflectivity inversion problem by testing on synthetic 1-D seismic traces and 2-D wedge models. We also report validations on simulated 2-D Marmousi2 model and real data from the Penobscot 3D survey off the coast of Nova Scotia, Canada. Next, we propose convolutional FirmNet (ConFirmNet), which is an extension of the FirmNet approach to solve the problem of convolutional sparse coding. As an application, we build a ConFirmNet based sparse autoencoder (ConFirmNet-SAE) and demonstrate suitability for image denoising and inpainting. Further, we also show that training ConFirmNet-SAE with the Huber loss imparts robustness to outliers. ConFirmNet-SAE also proves to be robust to mismatch between training and test noise conditions than convolutional learned iterative soft-thresholding algorithm (CLISTA). Finally, we propose a sparse recovery formulation that employs a nonuniform, nonconvex synthesis sparse model comprising a combination of convex and nonconvex regularizers, which results in accurate approximations of the l-0 pseudo-norm. The resulting iterative optimization employs proximal averaging. When unfolded, the iterations give rise to a nonuniform sparse proximal average network (NuSPAN) that can be optimized in a data-driven fashion. We demonstrate the efficacy of NuSPAN also for solving the problem of seismic reflectivity inversion.
Speaker Biodata : Praveen Kumar Pokala received his B.Tech. degree in Electronics and Telecommunication Engineering from Jawaharlal Nehru Technological University, Hyderabad, India, in 2006 and M. Tech degree in Signal Processing from Indian Institute of Technology (IIT), Guwahati, India, in 2009. Subsequently, he worked as an Assistant Professor in LPU university, Jalandhar, India and GITAM university, Hyderabad, India. He is currently pursuing Ph.D. in the Department of Electrical Engineering, Indian Institute of Science, Bangalore. His current research interests are machine learning, deep learning, and nonconvex optimization algorithms, with applications to inverse problems in computational imaging.

Event : Thesis Defence
Title : Robust learning under label noise with Neural networks
Speaker : Deep B Patel
Degree Registered :M.Tech. (Research)
Advisor : Prof. P S Sastry
Date : 23/07/2021
Venue : Online
Abstract : Label noise is inevitable when employing supervised learning based algorithms in practice. In many applications involving neural networks one needs a large training set and the process of obtaining such labelled data (e.g., crowd sourcing, employing automatic web searches etc.) often lead to training set labels being noisy. In the context of neural networks, it is demonstrated that standard algorithms (such as minimizing empirical risk with cross entropy loss function) are susceptible to overfitting in the presence of noise. This thesis explores the problem of robust learning under label noise. There are many approaches proposed for designing learning algorithms that are robust to label noise. We look at the sample reweighting methods where in one tries to assign weights to different examples so that examples with noisy labels are assigned small or zero weights. This can be viewed as a kind of curriculum learning where in the clean (easy) samples are to be given more weightage than the corrupted (hard) samples. Based on such heuristics, we propose a simple, adaptive curriculum based learning strategy called BAtch REweighting (BARE). The statistics of loss values of all samples in a mini-batch are used to decide which examples in each mini-batch would be allowed to update the weights. This yields an adaptive curriculum where the sample selection is naturally tied to current state of learning. Our algorithm does not need any clean validation data, needs no knowledge at all of the noise rates and also does not have any hyperparameters. We empirically demonstrate the effectiveness of our algorithm on benchmark data sets such as MNIST, CIFAR-10 and Clothing-1M, and show that it is much more efficient in terms of time and has as good or better robustness compared to other current algorithms based on sample reweighting. We next consider another aspect of the susceptibility of deep networks to label noise. It is shown recently that deep networks trained on data with random labels can memorize the data in the sense of being able to drive the training error to zero. This phenomena of memorization is confirmed by multiple studies and it is seen that none of the standard regularization techniques can mitigate it. This depends on the kind of local minima that SGD can take the network to. Hence it could depend on the topography of the empirical risk that is minimized. Thus, the choice of loss function can be critical in determining this. However, none of the studies on memorization investigate the role of loss function. We present extensive empirical results to show that while standard loss functions like CCE and MSE result in memorization, symmetric loss functions such as RLL can resist such memorization to a good degree. We formally define what resisting memorization means and then provide some theoretical justification for the empirical results.

Event : Thesis Defence
Title : Developmental Studies of a Solid-State Pulsed Power System for Liquid Food Sterilisation
Speaker : Tanmaya Pradhan
Degree Registered :M.Tech. (Research)
Advisor : Dr. Joy Thomas
Date : 22/07/2021
Venue : Online

Event : Thesis Defence
Title : Speaker verification using whispered speech
Speaker : Abinay Reddy Naini
Degree Registered :MTech (Research
Advisor : Prof. Prasanta Kumar Ghosh
Date : 13/07/2021
Venue : Online
Abstract : Like neutral speech, whispered speech is one of the natural modes of speech production, and it is often used by speakers in their day-to-day life. For some people, such as laryngectomees, whispered speech is the only mode of communication. Despite the absence of voicing in whispered speech and difference in characteristics compared to the neutral speech, previous works in the literature demonstrated that whispered speech contains adequate information about the content and the speaker.
In recent times, virtual assistants have become more natural and widespread. This led to an increase in the scenarios, where the device has to detect the speech and verify the speaker even if the speaker whispers. Due to the noise-like characteristics, detecting whispered speech is a challenge. On the other hand, a typical speaker verification system, where neutral speech is used for enrolling the speakers but whispered speech for testing, often performs poorly due to the difference in acoustic characteristics between the whispered and the neutral speech. Hence, the aim of this thesis is two-fold: 1) develop a robust whisper activity detector specifically for speaker verification task, 2) improve whispered speech based speaker verification performance.
The contributions in this thesis lie in whisper activity detection as well as whispered speech based speaker verification. It is shown how an Attention-based average pooling in a speaker verification model can be used to detect the whispered speech regions in noisy audio more accurately than the best of the baseline schemes available. For improving speaker verification using whispered speech, we proposed features based on formant gaps, and we showed that these features are more invariant to the modes of the speech compared to the best of the existing features. We also proposed two feature mapping methods to convert the whispered features to neutral features for speaker verification. In the first method, we introduced a novel objective function, based on cosine similarity, for training a DNN, used for feature mapping. In the second method, we iteratively optimized the feature mapping model using cosine similarity based objective function and the total variability space likelihood in the i-vector based background model. The proposed optimization provided a more reliable mapping from whispered features to neutral features resulting in an improvement of speaker verification equal error rate by 44.8% (relative) over an existing DNN based feature mapping scheme.

Event : Thesis Colloquium
Title : Spectrotemporal Processing of Speech Signals Using the Riesz Transform
Speaker : Jitendra Kumar Dhiman
Degree Registered :Ph.D.
Advisor : Prof. Chandra Sekhar Seelamantula
Date : 12/07/2021
Venue : Online
Abstract : Speech signals have time-varying spectra. Spectrograms have served as a useful tool for the visualization and analysis of speech signals in the joint time-frequency plane. In this thesis, we consider 2-D analysis of speech spectrograms. We consider a spectrotemporal patch and model it as a 2-D amplitude-modulated and frequency-modulated (AM-FM) sinusoid. Demodulation of the spectrogram yields the 2-D AM and FM components, which correspond to the slowly varying vocal-tract envelope and the excitation, respectively. For solving the demodulation problem, we rely on the complex Riesz transform, which is a 2-D extension of the 1-D Hilbert transform. The demodulation viewpoint brings forth many interesting properties of the speech signal. The spectrotemporal carrier helps us identify the regions that are coherent and those that are not. Based on this idea, we introduce the coherencegram corresponding to a given spectrogram. The temporal evolution of the pitch harmonics can also be characterized by the orientation at each time-frequency coordinate, resulting in the orientationgram. We show that these features collectively enable solutions for the important problems of voiced/unvoiced segmentation, aperiodicity estimation, periodic/aperiodic signal separation, and pitch tracking. We compare the performance of the proposed methods with benchmark methods. The spectrotemporal amplitude characterizes the time-varying magnitude response of the vocal-tract filter. We show how the formants and their bandwidths manifest in the spectrotemporal amplitude. It turns out that the formant bandwidths are mildly overestimated, which are perceptible when one performs speech synthesis using the estimated parameters. We propose a method for correcting the formant bandwidths, which also restores the speech quality. Finally, we use the curated spectrotemporal amplitude, pitch, aperiodicity, and voiced/unvoiced decisions for the task of speech reconstruction in a spectral synthesis model and a neural vocoder, namely, WaveNet. We show that conditioning WaveNet on the spectrotemporal features results in high-quality speech synthesis. The quality of the synthesized speech is assessed using both objective and subjective measures.
We rely on the Perceptual Evaluation of Speech Quality (PESQ) measure and standard Mean Opinion Score (MOS) test for objective and subjective evaluation, respectively. The performance of the proposed parameters is evaluated in a vocoder framework that uses the spectral synthesis model for speech reconstruction. The objective evaluation shows that the performance of the Riesz transform-based speech parameters is on par with the baseline systems. Using the spectral synthesis model, we report an average PESQ score in the range from 2.30 to 3.45 over a total of 200 speech waveforms taken from the CMU-ARCTIC database comprising both male and female speakers. In comparison, WaveNet-based speech reconstruction gave an average PESQ score of 3.65.
Subjective evaluation was carried out through listening tests conducted in an acoustic test chamber on volunteers in the age group of 21 to 30. The average MOS score was 4.30 when the Riesz transform-based features were used in WaveNet for speech reconstruction, which was also comparable with the baseline systems: STRAIGHT and WORLD. Both objective and subjective evaluations also showed that the quality of reconstructed speech waveforms was superior with the proposed features in a WaveNet vocoder than in the spectral synthesis model.
An audio demonstration is available at the GitHub link: http://jitendradhiman.github.io/Demo
Speaker Biodata : Jitendra Kumar Dhiman received his B.Tech. degree in Electronics and Telecommunication Engineering from the Institution of Electronics and Telecommunication Engineering, Delhi, India, in 2010, and M.Tech. degree in Signal Processing from Indian Institute of Technology Hyderabad in 2013. Subsequently, he joined as a project assistant in Spectrum Lab (EE Department, IISc) and worked on prosody modification of speech signals, and then as a PhD student working on spectrotemporal models for speech processing. His research interests include speech and audio signal processing and machine learning.

Event : Thesis Defence
Title : Dynamics of a Stratified Population of Optimum Seeking Agents on a Network
Speaker : Nirabhra Mandal
Degree Registered :M. Tech (Research)
Advisor : Dr. Pavankumar Tallapragada
Date : 06/07/2021
Venue : Online
Abstract : Very large scale multi-agent systems occur both naturally and in engineering applications. In many of these systems, the agents are either selfish or act with varying levels of coordination. Understanding the evolution of large populations of such agents has been of interest in diverse domains such as biology, ecology, sociology, economics, transportation engineering, robotics and control engineering. One specific area that has not been studied enough is that of evolution of large populations on networks of choices. This specific setting has potential applications in the contexts of fleet redistribution of ride sharing services, evolution of transportation mode choices of a population, opinion dynamics, human and insect swarm migrations and robotics swarms.
In this thesis, we consider a population composed of a continuum of agents that seek to maximize a payoff function by moving on a network. The nodes in the network may represent physical locations or abstract choices. The population is stratified and hence agents opting for the same choice may not get the same payoff. In particular, we assume payoff functions that model diminishing returns, that is, agents in ``newer'' strata of a node receive a smaller payoff compared to ``older'' strata. Moreover, at each time instant, the network imposes constraints on the set of choices that an agent can revise to.
We first model the population dynamics under three choice revision policies, each having varying levels of coordination - (i). no coordination and the agents are selfish, (ii). coordination among agents in each node and (iii). coordination across the entire population. To model the case with selfish agents, we generalize the Smith dynamics to our setting, where we have a stratified population and network constraints. We refer to this dynamics as stratified Smith dynamics or SSD. To model nodal coordination, we allow the fraction of population in a node, as a whole, to take the `best response' to the state of the population in the node's neighborhood. We call this as nodal best response dynamics or NBRD. For the case of population-wide coordination, we explore a dynamics where the population evolves according to centralized gradient ascent of the social utility, though constrained by the network. We call this dynamics as network restricted payoff maximization or NRPM. In each case, we show that the dynamics has existence and uniqueness of solutions. We also show that the solutions from any initial condition asymptotically converge to the set of Nash equilibria and the social utility converges to a constant.
We then study the steady state of the population and the steady state social utility for the three dynamics SSD, NBRD and NRPM. We provide sufficient conditions on the network parameters under which there exists a unique Nash equilibrium. We then utilize positive correlation properties of the dynamics to in order to provide an upper bound on the steady state social utility as a function of the initial population configuration. Finally, we extend the idea behind the sufficient condition for the existence of a unique Nash equilibrium to partition the graph appropriately in order to provide a lower bound on the steady state social utility. We then illustrate interesting cases as well as our results using simulations.
Lastly, we discuss some preliminary ideas to utilize this framework to control the population to a desired configuration. We try to achieve this by changing the payoff functions at a much slower rate than the rate at which the population converges. We also look at ways to compute an optimal control action that achieves this.

Event : Thesis Defence
Title : Probabilistic source-filter model of speech
Speaker : Achuth Rao M V
Degree Registered :PhD
Advisor : Prof. Prasanta Kumar Ghosh
Date : 28/06/2021
Venue : Online
Abstract : The human respiratory system plays a crucial role in breathing and swallowing. However, it also plays an essential role in speech production, which is unique to humans. Speech production involves expelling air from the lungs. As the air flows from the lungs to the lips, some kinetic energy gets converted to sound. Different structures modulate the generated sound, which is finally radiated out of the lips. The speech consists of various information such as linguistic content, speaker identity, emotional state, accent, etc. Apart from speech, there are various scenarios where the sound is generated in the human respiratory system. These could be due to abnormalities in the muscles, motor control unit, or the lungs, which can directly affect generated speech as well. A variety of sounds are also generated by these structures while breathing including snoring, Stridor, Dysphagia, and Cough.
The source filter (SF) model of speech is one of the earlier models of speech production. It assumes that speech is a result of filtering an excitation or source signal by a linear filter. The source and filter are assumed to be independent. Even though the SF model represents the speech production mechanism, there needs to be a tractable way of estimating the excitation and the filter. The estimation of both of them given speech falls under the general category of signal deconvolution problem, and, hence, there is no unique solution. There are several variations of the source-filter model in the literature by assuming different structures on the source/filter. There are various ways to estimate the parameters of the source and the filter. The estimated parameters are used in various speech applications such as automatic speech recognition, text to speech, speech enhancement etc. Even though the SF model is a model of speech production, it is used in applications including Parkinson's Disease classification, asthma classification.
The existing source filter models show much success in various applications, however, we believe that the models mainly lack two respects. The first limitation is that these models lack the connection to the physics of sound generation or propagation. The second limitation of the current models is that they are not fully probabilistic. The inherent nature of the airflow is stochastic because of the presence of turbulence. Hence, probabilistic modeling is necessary to model the stochastic process. The probabilistic models come with several other advantages: 1) systematically inducing the prior knowledge into the models through probabilistic priors, 2) the estimation of the uncertainty of the model parameters, 3) allows sampling of new data points 4) evaluation of the likelihood of the observed speech.
We start with the governing equation of sound generation and use a simplified geometry of the vocal folds. We show that the sound generated by the vocal folds consists of two parts. The first part is because of the difference between the subglottal and supra glottal pressure difference. The second part is because of the sound generated by turbulence. The first kind is dominant in the voiced sounds, and the second part is dominant in the unvoiced sounds. We further assume the plane wave propagation in the vocal tract, and there is no feedback from the vocal tract on the vocal folds. The resulting model is the excitation passing through an all-pole filter, and the excitation is the sum of two signals. The first signal is quasi-periodic, and the shape of each cycle depends on the time-varying area of the glottis. The second part is stochastic because the turbulence is modeled as a white noise passed through a filter. We further convert the model into a probabilistic one by assuming the following distribution on the excitations and filters. We model the excitation using a Bernoulli Gaussian distribution. Filter coefficients are modeled using the Gaussian distribution. The noise distribution is also Gaussian. Given these distributions, the likelihood of the speech can be derived as a closed-form expression. Similarly, we impose an appropriate prior to the model’s parameters and make a maximum a posteriori estimation of the parameters. But the model assumption can be changed/approximated with respect to the application and resulting in different estimation procedures. To validate the model, we apply this model to seven applications as follows:
1. Analysis and Synthesis: This application is to understand the representation power of the model.
2. Robust GCI detection: This shows the usefulness of estimated excitation, and the probabilistic modeling helps to incorporate the second-order statistics for robust the excitation estimation.
3. Probabilistic glottal inverse filtering: This application shows the usefulness of the prior distribution on filters.
4. Neural speech synthesis: We show that the model’s reformulation with the neural network results in a computationally efficient neural speech synthesis.
5. Prosthetic esophageal (PE) to normal speech conversion: We use the probabilistic model for detecting the impulses in the noisy signal to convert the PE speech to normal speech.
6. Robust essential vocal tremor classification: The usefulness of robust excitation estimation in pathological speech such as essential vocal tremor.
7. Snorer group classification: Based on the analogy between voiced speech production and snore production, the derived model is applicable for snore signals. We also use the parameter of the model to classify the snorer groups.

Event : Thesis Colloquium
Title : Spacially Adaptive Regularization for Image Restoration
Speaker : Sanjay Viswanath
Degree Registered : PhD
Advisor : Dr. Muthuvel Arigovindan
Date : 07/06/2021
Venue : Online
Abstract : Image restoration/reconstruction refers to the estimation of the underlying image from measurements generated by imaging devices. This problem is generally ill-posed due to the fact that measurements are corrupted because of the physical limitations of the imaging device, and the inherent noise involved in the measurement process. There are three main classes of methods in the current literature. The first class of methods is based on the regularization framework that enforces an ad-hoc prior on the restored image. The second class of methods uses regression-based learning paradigms, where a training set of clean images and the corresponding distorted measurements are used to generate a trained prior. The third class of methods adopts trained priors similar to the ones utilized in the second class of methods but within the regularization framework. This third class of methods, the trained regularization methods, are getting increasing attention because of their versatility as the regularization methods, while also encompassing natural priors obtained from training. However, the need for training data can limit their applicability. In this thesis, we propose spatially adaptive regularization methods where the adaptation information is retrieved from the measured data that undergoes reconstruction. Due to the adaption, the enforced prior is more natural than the existing regularization methods. At the same time, our methods do not require training data.
Summary of Contributions:
In the first part, we propose a novel regularization method that adaptively combines the well-known second-order regularization, called Hessian-Schatten (HSN) norm regularization, and first-order TV (TV-1) functionals with spatially varying weights. The relative weight involved in combining the first- and second-order terms becomes an image and this weight is determined through the minimization of a composite cost function, without user intervention. Our contributions in this part can be summarized as follows:
• We construct a composite regularization functional containing two parts: (i) the first part is constructed as the sum of TV-1 and HSN with spatially varying relative weights; (ii) the second part is an additional regularization term for preventing rapid spurious variations in the relative weights. The total composite cost functional is convex with respect to either the required image or the relative weight, but it is non-convex jointly.
• We construct a block coordinate descent method involving minimizations w.r.t. the required image and the relative weight alternatively with the following structure: the minimization w.r.t. the required image is carried out using the Alternating Direction Method of Multipliers (ADMM), and the minimization w.r.t. the relative weight is carried out as a single-step exact minimization using a formula that we derive.
• Since the total cost is non-convex, the reconstruction results are highly dependent on the initialization for block-coordinate descent method. We handle this problem using a multi-resolution approach, where a series of coarse-to-fine reconstructions are performed by minimization of cost functionals defined through upsampling operators. Here, minimization w.r.t. the relative weight and the required image is carried out alternatively, as we progress from coarse to final resolution levels. At the final resolution level, the above-mentioned block coordinate descent method is applied.
• The sub-problem of minimization w.r.t. to the required image involves spatially varying relative weights. Further, this sub-minimization problem in the above-mentioned multi-resolution loop involves upsampling operators. Hence, the original ADMM method proposed by Papafitsoros et al. turns out to be unsuitable. We propose an improved variable splitting method and computational formulas to handle this issue.
• We prove that the overall block coordinate descent method converges to a local minimum of the total cost function using Zangwill’s convergence theorem.
We name our method Combined Order Regularization with Optimal Spatial Adaptation (COROSA). We provide restoration examples involving deconvolution of TIRF images and reconstruction of Magnetic Resonance Imaging (MRI) images from under-sampled Fourier data. We demonstrate that COROSA outperforms existing regularization methods and selected deep learning methods.
In the second part, we make COROSA more adaptive by replacing the HSN with a spatially varying weighted combination of Eigenvalues of the Hessian. This means that the resulting regularization will be in the form of a spatially varying weighted sum of three terms involve the gradient and two Eigenvalues of Hessian. This allows the function to restore fine image structures through directional weighting, in terms of the local Eigenvalues. We again adopt a BCD scheme that alternates between the spatially varying weight estimation and image computation, as done in the first part. However, both steps are more complex with the new form. The first task of weight estimation is more complex as it involves three terms. The second task of image computation is more complex because there is no known proximal operator for regularization involving unequally weighted Hessian Eigenvalues. We solve the first problem by constructing a novel iterative method, and the second problem by deriving a novel proximal formula. Here too, we adopt a multi-resolution approach to initialize the BCD method. We call our method the Hessian Combined Order Regularization with Optimal Spatial Adaptation (H-COROSA). We experimentally compare H-COROSA with well-known regularization methods and selected learning-based methods for MRI reconstruction from under-sampled Fourier data.
Compressive Sensing based methods have shown the advantage of l0 based sparsity enforcing functionals in restoration. For practical applications, lp, 0 < p ≤1 functionals have been found to perform better than l1 functionals. In the last part, we propose an lp-based generalization of the previous COROSA and H-COROSA formulations. We replace the corresponding l1 based functionals with lp norm enforced on the combined multi-order functionals. Additionally for H-COROSA, we also consider two forms of penalty for the spatial weights. We construct an iteration scheme that is a merging of the majorization-minimization method for lp norm and BCD method used in the first two parts of the thesis. Again, we use a similar multi-resolution method for initialization. We demonstrate the advantage of using lp norm using MRI reconstruction examples involving severe undersampling in the Fourier domain.

Event :
Title :
Speaker :
Degree Registered :
Advisor :
Date :
Venue :
Abstract :
Speaker Biodata :

Event : Thesis Colloquium
Title : Novel Regularized Image Reconstruction Methods for Sparse Photoacoustic Tomography
Speaker : Rejesh N A
Degree Registered : PhD
Advisor : Dr. Muthuval Arigovindan
Date : 13/05/2021
Venue : Online
Abstract : Among all tissue imaging modalities, photoacoustic tomography (PAT), has been getting increasing attention in the recent past due to the fact that it has high contrast, high penetrability, and has the capability of retrieving high resolution. By using the combination of optical absorption and acoustic wave propagation, PAT has been able to image tissues at relatively large depths with high resolution compared to purely optical modalities. Upon shining with a laser pulse, the substance under investigation absorbs optical energy and undergoes thermoelastic expansion; as a result, the spatial distribution of the concentration of the substance gets translated into the distribution of pressure-rise. This initial pressure rise travels outwards as ultrasound waves which are collected by ultrasound transducers placed at the boundary. From the ultrasound signal measured by the transducers as a function of time, a PAT reconstruction method recovers an estimate of the initial pressure-rise by solving the associated inverse problem. The inverse problem is however challenging. It is challenging because the image has to be recovered for the entire cross-sectional plane, whereas the samples of the acoustics pressure are available only from the points lying in the periphery of the imaging specimen where the transducers are located. In this thesis, we make contributions in two widely used types of reconstructions methods known as the time-reversal method, and the model-based method.
Summary of Contributions:
In the first part, we develop an improved model-based method. Model-based reconstruction methods in PAT express the measured pressure samples as a linear transformation on the initial pressure-rise and perform a regularized reconstruction. Model-based methods yield superior image quality even in the situation where measured data size is small. We propose a model-based image reconstruction method for PAT involving a novel form of regularization and demonstrate its ability to recover good quality images from datasets of significantly reduced size. The regularization is constructed to suit the physical structure of typical PAT images. We construct it by combining second-order derivatives and intensity into a non-convex form to exploit a structural property of PAT images that we observe: in PAT images, high intensities and high second-order derivatives are jointly sparse. This regularization is combined with a data fidelity cost, and the required image is obtained as the minimizer of this cost. As this regularization is non-convex, the efficiency of the minimization method is crucial in obtaining artefact-free reconstructions. We develop a custom minimization method for efficiently handling this non-convex minimization problem. Further, as non-convex minimization requires a large number of iterations and the PAT forward model in the data-fidelity term has to be applied in the iterations, we propose a computational structure for efficient implementation of the forward model with reduced memory requirements. We evaluate the proposed method on both simulated and real measured data sets and compare them with a recent reconstruction method that is based on well-known total variation regularization.
Appropriate tuning of the regularization weight, λ, plays a crucial role in determining the quality of reconstructed images in PAT. To make any regularization method practicable, we need to have a way to determine the λ from the measured data. Unfortunately, an appropriately tuned value of the regularization weight varies significantly with the variation in the noise level, as well as, with the variation in the high-resolution contents of the image, in a way that has not been well understood. In the part of the work described above, we did not address this problem as the focus has been to demonstrate the suitability of the intensity-augmented regularization for PAT image recovery; in the experimental demonstration, we determined the required regularization weight by using the models that generated data. In the second part of the thesis, we develop a semi-automatic method for determining the regularization weight from measured data. As a first step, we introduce a relative smoothness constraint with a parameter; this parameter computationally maps into the actual regularization parameter, but its tuning does not vary significantly with variation in the noise level, as well as with the variation in the high-resolution contents of the image. Next, we construct an algorithm that integrates the task of determining this mapping along with obtaining the reconstruction. Finally, we demonstrate experimentally that we can run this algorithm with a nominal value of the relative smoothness parameter---a value independent of the noise level and the structure of the underlying image---to obtain good quality reconstructions. We compare the structural similarity (SSIM) scores of reconstructions obtained this way to that of reconstructions in which the regularization weight was determined using the models themselves; we show that the SSIM scores are comparable. This means that, from a practical point of view, our work solves the problem of determining the required regularization weight from measured images.
In the first two parts, we assumed that the forward model that measures the signal from the target object be ideal. In particular, we assumed that the excitation pulse and transducers' impulse response are Dirac deltas. We focused only on the non-ideality of the transducer configuration, i.e., we handled the case where the transducer locations do not densely sample the detection surface as required by the well-known back-projection method to work. Both excitation pulse and transducer impulse response have a finite width, and this leads to some distortions in the reconstructed image. In the last part of the thesis, we propose a pre-processing method for correcting the distortions in the context of using time-reversal methods which are similar to the back-projection method. To this end, we formulate the broadening of the PA signals as a convolution between the impulse response of the system and the input excitation pulse. A deconvolution method using Tikhonov regularization is proposed to correct the PA signals before applying the time-reversal method. This resulted in improved resolution in the reconstructed images. A two-level deconvolution with the Tikhonov regularization method is also proposed to remove the blurring caused by the finite bandwidth of transducers and by the broad excitation pulses. We evaluate the usefulness of our method using numerical simulations and demonstrate that the reconstructed images from the deconvolved PA signals remain unaffected by changes in pulse widths or pulse shapes, as well as by the limited bandwidth of the ultrasound detectors.

Event : Thesis Colloqium
Title : Techniques for estimating the direction of pointing gestures using depth images in the presence of orientation and distance variations from the depth sensor
Speaker : Shome Subhra Das
Degree Registered :PhD
Advisor : --
Date : 26/04/2021
Venue : Online
Abstract : Currently the interaction with computers, robots, drones, and virtual reality interfaces is through the use of mouse, touch-pad, joystick, virtual reality wand, drone controller etc. These devices have one or more of the following limitations: being cumbersome, non-immersive, immobile, and having a steep learning curve. The target of this work is to explore ways to replace existing pointing devices with pointing gesture-based interfaces.
This thesis addresses two problems, namely estimating the direction being indicated by a pointing gesture (PDE) and detection of pointing gestures. The proposed techniques use a single depth sensor and use only the hand region. To our knowledge, this is the maiden attempt at creating a depth and orientation tolerant, accurate method for estimating the pointing direction using only depth images. The proposed methods achieve accuracies comparable to or better than those of existing methods while avoiding their limitations.
Significant contributions of the thesis:
(i) Proposing a real-time technique for estimating the pointing direction using a nine-axis inertial motion unit (IMU) and an RGB-D sensor. It is the first method to compute the pointing direction (PD) by finding the axis vector of the index finger. It is also the first method to fuse information from the IMU and depth sensor to obtain the PD. Further, this is the first method to obtain the accurate ground-truth pointing direction of index finger-based pointing gestures.
(ii) Creation of a large (100k+ samples) data-set with accurate ground truth for PDE from depth images. Each sample consists of the segmented depth image of a hand, the fingertip location (2D + 3D), the pointing vector (as a unit vector and in terms of the yaw and pitch values), and the mean depth of the hand. This is the first public data-set for depth image-based PDE that has accurate ground truth and a large number of samples.
(iii) Proposing a new 3D convolutional neural network-based method to estimate pointing direction. This is the first deep learning-based method for PDE that uses only the depth images for PDE, without the use of RGB data. It is tolerant to variation in orientation and depth of the hand with respect to the camera and is suitable for real-time applications.
(iv) Proposing another technique for estimating the pointing direction using global registration of the test data point cloud with a pointing hand model captured using Kinect fusion based method. It is tolerant to the variation in the orientation and depth of the hand w.r.t. the RGB-D sensor. It does not have the limitation of the previously proposed methods since it does not require the attachment of any device such as IMU nor does it require any data-set for training. It achieves less net angular error than most techniques in the literature using only the hand region.
(v) Creation of a large data-set of positive and negative samples for detection of pointing gestures from depth images of the hand region. A technique is also proposed using deep learning to distinguish pointing gestures from other hand gestures. This achieves higher accuracy than the only other technique in the literature by Cordo et al. for all the depths of the hand covering the entire range of the depth sensor.

Event : Thesis Colloquium
Title : Study of Robust Learning under Label Noise with Neural Networks
Speaker : Deep Patel
Degree Registered :MTech (Research)
Advisor : Prof. P S Sastry
Date : 23/04/2021
Venue : Online
Abstract : Label noise is inevitable when employing supervised learning based algorithms in practice. In many applications involving neural networks one needs a large training set and the process of obtaining such labelled data (e.g., crowd sourcing, employing automatic web searches etc.) often lead to training set labels being noisy. In the context of neural networks, it is demonstrated that standard algorithms (such as minimizing empirical risk with cross entropy loss function) are susceptible to overfitting in the presence of noise. This thesis explores the problem of robust learning under label noise. There are many approaches proposed for designing learning algorithms that are robust to label noise. We look at the sample reweighting methods where in one tries to assign weights to different examples so that examples with noisy labels are assigned small or zero weights. This can be viewed as a kind of curriculum learning where in the clean (easy) samples are to be given more weightage than the corrupted (hard) samples. Based on such heuristics, we propose a simple, adaptive curriculum based learning strategy called BAtch REweighting (BARE). The statistics of loss values of all samples in a mini-batch are used to decide which examples in each mini-batch would be allowed to update the weights. This yields an adaptive curriculum where the sample selection is naturally tied to current state of learning. Our algorithm does not need any clean validation data, needs no knowledge at all of the noise rates and also does not have any hyperparameters. We empirically demonstrate the effectiveness of our algorithm on benchmark data sets such as MNIST, CIFAR-10 and Clothing-1M, and show that it is much more efficient in terms of time and has as good or better robustness compared to other current algorithms based on sample reweighting. We next consider another aspect of the susceptibility of deep networks to label noise. It is shown recently that deep networks trained on data with random labels can memorize the data in the sense of being able to drive the training error to zero. This phenomena of memorization is confirmed by multiple studies and it is seen that none of the standard regularization techniques can mitigate it. This depends on the kind of local minima that SGD can take the network to. Hence it could depend on the topography of the empirical risk that is minimized. Thus, the choice of loss function can be critical in determining this. However, none of the studies on memorization investigate the role of loss function. We present extensive empirical results to show that while standard loss functions like CCE and MSE result in memorization, symmetric loss functions such as RLL can resist such memorization to a good degree. We formally define what resisting memorization means and then provide some theoretical justification for the empirical results.

Event : Thesis Colloquium
Title : Lipschitz Regularization of Convolutional Neural Networks
Speaker : Anirudh Singh
Degree Registered :MTech (Res)
Advisor : Prof. Kunal Narayan Chaudhury
Date : 16/04/2021
Venue : Online
Abstract : We consider the problem of estimating the Lipschitz constant of (the end-to-end mapping in) convolutional neural networks (CNNs) and regulating it during the training process. This has applications ranging from stable training of generative models and robustness to adversarial attacks to stability analysis of CNN-based closed-loop systems and the derivation of generalization bounds.
The challenge in this regard is that computing the exact Lipschitz constant of even simple neural networks is known to be NP-hard. In practice, the product of the largest singular values of the convolutional layers is used to bound the Lipschitz constant. However, even computing the singular values for a single convolutional layer is computationally prohibitive; moreover, this has to be done for each layer in every backpropagation pass. Methods not using an explicit SVD still rely on the power method, which requires inner iterations within the already iterative training procedure. In this context, our contributions are the following:
(i) We propose a simple bound for the singular value of a convolutional layer, which can be computed efficiently using FFTs and does away with the need for SVD or power iterations.
(ii) We study the effect of Lipschitz regularization on constrained convolutional architectures which reveals the role played by overparameterization.
We test the performance of Lipschitz regularization using the proposed bound on the CIFAR-10 image classification dataset. Though our bound is loose, quite remarkably, we can improve the generalization accuracy by more than 2% without practically introducing any computational overhead. In comparison to algorithms that use the exact singular values, we perform marginally better in terms of accuracy while taking significantly less time to complete the training. In addition, we study convolutional architectures with significantly fewer trainable parameters than regular convolutions. This study reveals that CNNs may be able to model certain distributions using multiple parameterizations, and it is this redundancy in parameterization that allows Lipschitz regularization to be successful.

Event : Thesis Collowquium
Title : Deep Learning Methods for Audio EEG Analysis
Speaker : Jaswanth Reddy Katthi
Degree Registered :MTech (Research)
Advisor : Dr. Sriram Ganapathy
Date : 15/04/2021, 11.00am
Venue : Online
Abstract : The perception of speech and audio is one of the defining features of humans. Much of the brain's underlying processes as we listen to acoustic signals are unknown, and significant research efforts are needed to unravel them. The non-invasive recordings capturing the brain activations like electroencephalogram (EEG) and magnetoencephalogram (MEG) are commonly deployed to capture the brain responses to auditory stimuli. But these non-invasive techniques capture artifacts and signals not related to the stimuli, which distort the stimulus-response analysis. The effect of the artifacts becomes more evident for naturalistic stimuli. To reduce the inter-subject redundancies and amplify the components related to the stimuli, the EEG responses from multiple subjects listening to a common naturalistic stimulus need to be normalized. The currently used normalization and pre-processing methods are the canonical correlation analysis (CCA) models and the temporal response function based forward/backward models. However, these methods assume a simplistic linear relationship between the audio features and the EEG responses and therefore, may not alleviate the recording artifacts and interfering signals in EEG. This talk proposes novel methods using machine learning advances to improve the audio-EEG analysis.
We propose a deep learning framework for audio-EEG analysis in intra-subject and inter-subject settings. The deep learning based intra-subject analysis methods are trained with a Pearson correlation-based cost function between the stimuli and EEG responses. This model allows the transformation of the audio and EEG features that are maximally correlated. The correlation-based cost function can be optimized with the learnable parameters of the model trained using standard gradient descent-based methods. This model is referred to as the deep CCA (DCCA) model. Several experiments are performed on the EEG data recorded when the subjects are listening to naturalistic speech and music stimuli. We show that the deep methods obtain better representations than the linear methods and results in statistically significant improvements in correlation values.
Further, we propose a neural network model with shared encoders that align the EEG responses from multiple subjects listening to the same audio stimuli. This inter-subject model boosts the signals common across the subjects and suppresses the subject-specific artifacts. The impact of improving stimulus-response correlations are highlighted based on multi-subject EEG data from speech and music tasks. This model is referred to as the deep multi-way canonical correlation analysis (DMCCA). The combination of inter-subject analysis using DMCCA and intra-subject analysis using DCCA is shown to provide the best stimulus-response in audio-EEG experiments.
The talk will conclude with a discussion on future challenges in audio-EEG analysis. We will highlight how much of the audio signal can be recovered purely from the non-invasive EEG recordings with modern machine learning methods.
Speaker Biodata : Jaswanth Reddy Katthi is a third year M. Tech (Research) student in EE department under the program System Science and Signal Processing. He is a part of LEAP lab, and pursuing his research under the guidance of Dr. Sriram Ganapathy. He did his bachelors from Jawaharlal Nehru Technological University, Anantapur in Electronics and Communication Engineering, during 2013-2017. His research interests involve machine learning, speech processing and cognitive neuroscience. He is an IEEE student member. He is from Kurnool, Andhra Pradesh.

Event : Seminar
Title : Novel condition monitoring and diagnostic techniques for insulating materials
Speaker : Ashiwn Desai
Date : 11/03/2021
Venue : Online
Abstract : With the growing demand for uninterrupted power supply, the reliable operation of the power system is very important. The reliability of power system depends on its insulation design and maintenance. Partial discharges (PD) such as corona, developed from a weak link, at the edge of the winding or due to the protrusion from the winding conductor, surface discharges due to high tangential electric field, and discharges due to particle movement tend to cause characteristic changes in the insulation properties of the material and can result in catastrophic failure of the insulation system. The trend is changing from time-based maintenance to condition-based maintenance so that system can be immediately repaired as and when required, which will avoid any catastrophic damage. Hence, it is essential to deploy continuous, non-intrusive, onsite condition monitoring technique to ensure trouble-free and reliable operation of the power system.
Amongst different PD detection techniques, Ultra High Frequency (UHF) method is superior in identifying the incipient discharge activity. Condition monitoring involves classification and localization of these incipient discharges which has become a challenge to researchers worldwide. The first part of the talk describes the investigation of the partial discharge activity using the UHF technique. Phase resolved partial discharge (PRPD) measurement by IEC 60270 method in conjunction with statistical parameters like kurtosis and skewness of the pulse height or phase-distribution, number of discharges per period, etc. are generally used for classification. The main limitation of such an analysis is that the parameters are purely descriptive and are not physical parameters like duration of the signal or rise time of the signal. Time-Frequency transformation for discharge classification has been proposed which uses true physical parameters of the UHF signals: duration of the signal, bandwidth of the signal and rise time of the signal.
In addition, the use of Cross Recurrence Plot (CRP) for estimation of Time Difference Of Arrival (TDOA) of UHF signals has been proposed. This method of TDOA estimation is shown to be superior to traditional methods under very noisy conditions thereby increasing localization accuracy remarkably.
Further, application of novel, non-intrusive and non-destructive techniques like Laser Induced Breakdown Spectroscopy (LIBS ) and Optical Emission Spectroscopy (OES) for condition monitoring and diagnosis of outdoor polymeric insulators will be discussed. These methods are shown to be a quick and reliable way of ranking the performance of the insulators.
Grounding is an important aspect of the power system from the viewpoint of power system reliability, equipment protection, and human safety. Designing a suitable grounding system requires reasonably accurate quantification of soil parameter values; particularly conductivity, σ, and for high frequency (transients) the relative permittivity, εr. The second part of the talk will be on characterization of non-linearities in soil conductivity and permittivity. The effect of frequency and current magnitude on ground impedance should be accounted for in grounding system design. Practical verification of grounding installations is commonly achieved by measuring ground resistance or ground impedance using low voltage AC or switched DC test equipment. Such equipment may inject only a few amperes of current into the grounding system, sometimes of the order of mA, which represents a small fraction of the actual current that may flow under fault conditions. IEEE standards 80, 81 recognize the effect of current magnitude on soil resistivity and grounding impedance in terms of thermal effects at high current causing drying out and soil ionization; however, no reference is made to the non-linear characteristics of resistivity (or permittivity) and ground electrode (rod) impedance over a low magnitude current range. The investigation revealed the dependency of soil parameters on current density and values obtained at low current density are overestimated compared to that observed under fault conditions. This dependency is due to the interface formed between the metallic electrodes and soil, further, this was found to significantly affect the measured values at low frequencies. A theoretical model has been developed to account for the observed dispersion in the soil parameters.
Speaker Biodata : Dr. Ashwin Desai is a Postdoctoral fellow, Khalifa University, Abu Dhabi, UAE. He got his PhD from IIT Madras, Chennai, India in 2019. His area of research is Condition monitoring of high voltage insulation systems and electrical grounding systems.

Event : Thesis Defence
Title : Emulation of wind turbine and sensorless control of doubly-fed induction generator for wind energy application
Speaker : Ramu Nair
Degree Registered :PhD
Advisor : Prof. G Narayanan
Date : 08/03/2021
Venue : Online
Abstract : Wind energy utilization has been growing at a rapid rate, fuelling research and development in wind turbine – generator systems. Doubly fed induction generator (DFIG) driven by a wind turbine is a commonly used wind energy conversion system. To advance research and education in wind energy conversion systems, a controlled test bed is necessary that does not depend on wind availability. Hence this thesis deals with emulation of wind turbine using a power electronic controlled squirrel cage induction motor (SCIM). The thesis also concerns improved control techniques for the DFIG to enhance the performance of the wind energy system.
Wind turbine emulation involves controlling the SCIM drive such that the motor exhibits the characteristics of a wind turbine. The inertia of the motor being much lower than the inertia of the emulated turbine is an important challenge in such wind turbine emulation schemes. This thesis proposes one degree of freedom (1-DOF) and two degree of freedom (2-DOF) control structures for wind turbine emulation, overcoming this challenge.
Regarding control of DFIG in wind energy applications, position sensorless operation of DFIG is desirable from considerations of cost, maintenance, cabling and reliability. This thesis proposes two stator flux based model reference adaptive observers (SF-MRAOs) for estimation of rotor speed and position in stand-alone DFIG. One of the proposed SF-MRAO is shown to work with good dynamic performance in vector control of stand-alone DFIG. The thesis also proposes a PLL based MRAO (PLL-MRAO) which does not require integration of sensed quantities, unlike other existing MRAOs. The linearized SF-MRAO is further utilized to propose a modified direct voltage control (DVC) of stand-alone DFIG with simplified design of controllers compared to existing DVC.
Grid integration of DFIG system requires synchronization of stator induced voltages with grid voltages before grid connection, and active and reactive power (PQ) control after grid connection. This thesis also proposes a unified control algorithm for synchronization and power control, enabling a seamless transfer from synchronization mode to PQ control mode. Synchronization requires initial rotor position information, obtained through either a novel rotor parking scheme or PLL-MRAO. It is shown that the unified control has negligible transients during grid integration of DFIG.
All the proposed observers and control algorithms are validated through simulations and experiments, performed on a 7.5 kW doubly-fed induction generator coupled to a 5.5 kW squirrel cage induction motor, available in the laboratory.
All are welcome

Event : Thesis Defense
Title : Parallel Computing Techniques for High Speed Power System Solutions
Speaker : R. Gnanavignesh
Degree Registered :PhD
Advisor : Dr. U J Shenoy
Date : 12/03/2021
Venue : Online
Abstract : Modern power systems are enormously large and complex entities. Increase in system size, introduction of complex controls, uncertainties in forecasting, etc. necessitate faster software tools to handle power system planning, operation and operator training. This thesis aims to improve the performance of power system software tools by proposing parallel algorithms with the objective of reducing their execution time.
Solution of a sparse set of linear algebraic equations is one of the most essential modules used in almost all power system software tools. The thesis addresses the issue of reducing the execution time of sparse linear algebraic solver by parallelizing sparse matrix factorization. A LU factorization algorithm which is more amenable for parallelization is identified and chosen. In this work, the structural symmetry property of power system sparse matrices is exploited to maximize the column or node level parallelism. Results obtained from the implementation of the proposed algorithm on Graphical Processing Units (GPUs) corroborate its efficacy by achieving significant reduction in the solution time when compared with state of the art CPU based sequential sparse linear solvers.
Power flow algorithm is one of the most frequently executed algorithms with respect to the steady state realm of the power system. Reduction in the solution time for the power flow algorithm would further boost other applications like contingency analysis, optimal power flow, dynamic studies, etc. This thesis proposes a parallel power flow algorithm based on Newton-Raphson method. Inclusion of reactive power limit constraints at generator buses in the problem formulation stage itself eradicates the need to use heuristic techniques. In this work, the given power system network for which the power flow solution is desired, is decomposed into smaller sub-networks and processed in an independent as well as in a concurrent manner. Partial results from the sub-networks are consolidated to arrive at the solution of original network. Results obtained indicate preservation of the superior convergence property of Newton-Raphson method and a significant reduction in the solution time required for the parallel version of the power flow when compared with the sequential version.
Transient stability assessment is an important module within the Dynamic Security Assessment application. Time domain simulation for the stability assessment by solving thousands of Differential Algebraic Equations (DAEs), even though is the preferred method, is computationally intensive and becomes a major computing challenge as system size increases. The thesis proposes a parallel algorithm based on spatial domain decomposition employing relaxation conditions to speedup the transient stability simulation to handle the aforementioned challenge. A convergence enhancing mechanism through selection of appropriate admittance parameters for the network emulating fictitious buses which mimic the remainder of the system for each sub-network is derived. Results obtained corroborate the scalability and improved speedup features of the methodology which achieves a significant reduction in the simulation execution time.

Event : Thesis Defence
Title : Price of Privacy of Smart Meter Data
Speaker : Soumyajit Gangopadhyay
Degree Registered : M. Tech (Research)
Advisor : Dr. Sarasij Das
Date : 17/02/2021
Venue : online
Abstract : Smart meter data form the backbone of various data-driven distribution system applications ranging from demand response, outage management, power theft detection, capacitor bank switching to distribution network planning and asset management. However, privacy concerns often inhibit consumers from sharing their smart meter data. Privacy preserving algorithms have been proposed in the literature to address the privacy concern of consumers. However, privacy preservation often tends to reduce the usability of data by limiting the information that can be extracted from the data. Consumers incur privacy loss if they give up their privacy right over smart meter energy data to the utility. Compensating the consumers for their privacy loss can be an alternate approach to encourage smart meter data sharing. The advantage of this approach is that it does not reduce the data usability. There is a lack of literature on the price of privacy of smart meter data.
In this work, a novel privacy pricing framework is proposed which can be used by the power utility to determine the price of privacy of consumers' energy data. The proposed framework is intended to compensate the consumers who agree to give up their privacy right over their smart meter energy data. At first, the notion of smart meter data privacy is presented. Estimation of the price of privacy of energy consumption data of smart meters is a challenging task due to the lack of a mathematical privacy model. So, a privacy model is proposed to quantify the private information content of consumers' energy consumption data based on the privacy notion. Then, a nonlinear convex optimization problem is formulated using the proposed privacy model to determine the price of privacy of the consumers for a given budget of the utility. A constraint is incorporated in the optimization problem to ensure that the privacy price does not exceed the electricity bill of a consumer. The impact of the utility's budget (allocated for privacy pricing) on a consumer's price of privacy is shown using practical smart meter data. The results show that a consumer's price of privacy also depends on the total number of consumers (participating in privacy pricing) and the privacy content of their data. The results also indicate that the proposed pricing framework can be effective in incentivizing consumers to share their smart meter data.

Event : Thesis Colloquium
Title : Analysis and Enhancement of Stability of Power Systems with Grid-scale Photovoltaic Power Plants
Speaker : Indla Rajitha Sai Priyamvada
Degree Registered :PhD
Advisor : Dr. Sarasij Das
Date : 09/02/2021
Venue : Online
Abstract : Owing to the negative impact of carbon emissions on environment, power systems are experiencing a paradigm shift in power generation. The fossil fuel-based generators that utilize synchronous machines are increasingly being replaced by the grid scale PhotoVoltaic (PV) generators. The energy conversion technology and the power electronic interface are the main factors that differentiate grid scale PVs from Synchronous Generators (SGs). The PV generators are equipped with power electronic converters which act as an efficient interface between generators and grid. The power electronic interface offers a better control over the electrical energy generated by the PV generators. However, the power electronic interface brings new challenges to power system stability. The small-signal stability and transient stability must be analyzed in detail to ensure reliable operation of power grid. This research work focuses on addressing transient and small-signal stability issues of grid connected grid scale PV generators.
In conventional power systems, swing equation of SGs and (extended) equal area criterion are used to assess the transient stability of power system. However, the same analysis techniques may not be applicable for PV generators. In this research work, transient stability assessment criteria are developed for grid connected PV generator (with two different control strategies viz., P-Q control and Vdc-Q control with/without support functionalities). The proposed criteria are developed considering the outer and inner control loop, Phase Locked Loop, and filter dynamics of PV generator. PSCAD simulations are carried out on a two-bus system and a modified IEEE-39 bus system to validate the proposed criterion. The stability criteria are found to effectively assess the stability of grid connected PV generators.
The power transfer capability of transmission network is limited by thermal limits, voltage limits and stability limits. Power transfer capability of transmission lines emanating from PV generators considering thermal and voltage limits is explored well in the literature. However, there is a lack of literature on stability constrained power transfer capability limit. In this research work, adaptive control-based tuning laws are proposed for grid connected PV generators to improve the stability constrained power transfer capability. The adaptive tuning laws are derived based on the Lyapunov energy function analysis. The Lyapunov functions are formulated using the summation of squares of the PI block errors and difference between the PI parameter values from their optimal values. PSCAD simulations are carried out to validate the proposed criterion. From the studies, it is observed that the proposed tuning laws are found to effectively improve the stability limit on power transfer to be equal to the voltage limit.
The increased penetration of PV generations into power systems has also brought qualitative changes in the small signal stability of power systems. Two new categories of oscillation modes are introduced into power systems that have participation from PV state variables. As the mode shape of the two new categories of oscillation modes is different from that of SG modes, the power system stabilizer design should be revisited. In this research work, H-infinity control-based power system stabilizer is developed considering the controllability and observability of the new categories of oscillation modes. The effectiveness of the developed stabilizer in providing sufficient damping to the new categories of oscillation modes is validated through PSCAD simulations on a modified IEEE-39 bus system.
As power systems are large interconnected systems, the increased penetration of PV generation has resulted in notable interaction among PV generators and SGs. Investigation of the interaction among generators is important to understand the dynamic behavior of overall power system when subjected to disturbances. This research work is carried out to understand the interaction among PV and SGs. The interaction is analyzed through investigation of interaction among oscillation modes of PV generation and SG. A mathematical formulation to quantify the interaction among the oscillation modes of PV generations and SGs is proposed. A modified IEEE-39 bus system is considered to carry out the interaction study and validate the results obtained from mathematical formulations.
Speaker Biodata :

Event : Thesis Defence
Title : Neural Representation Learning of Speech and Audio Signals
Speaker : Purvi Agrawal
Degree Registered : PhD
Advisor : Dr. Sriram Ganapathy
Date : 28/01/2021
Venue : Online
Abstract : Representation learning is the branch of machine learning consisting of techniques that are capable of automatically discovering meaningful representations from raw data for efficient information extraction. In the recent years, following the trends in other streams of machine learning, representation learning using neural networks has attracted significant interest. In the speech processing field, representation learning has been a challenging task. This talk is focused on developing neural methods for representation learning of speech and audio signals, with the goal of improving downstream applications that rely on these representations. For representation learning, we pursue two broad directions - supervised and unsupervised. In the case of speech/audio signals, we identify two stages of representation learning. The first stage is the learning of a time-frequency representation (equivalent of spectrogram) from the raw audio waveform. The second stage is the learning of modulation representations (filtering the time-frequency representations along the temporal domain, called rate filtering and spectral domain, called scale filtering). In the first part of the talk, we propose an interpretable supervised representation learning framework for speech and audio data. Here, a two-stage representation learning approach from raw waveform is proposed, consisting of acoustic filter-bank learning (time-frequency representation learning) from raw waveform followed by a modulation representation learning. This two-stage learning is directly optimized for the task at hand. The key novelty in the proposed framework consists of a relevance weighting mechanism that acts as a feature selection module. This is inspired by gating networks and provides a mechanism to weight the relevance of the acoustic and modulation representations for the task involved. The relevance weighting network can also utilize feedback from the previous predictions of the model for tasks like automatic speech recognition (ASR). The proposed relevance weighting scheme is shown to provide significant performance improvements for ASR task and UrbanSound audio classification task. A detailed analysis yields insight into the interesting properties of the relevance weights that are captured by the model at the acoustic and modulation stages for speech and audio signals. The second part of the talk deals with representation learning methods for speech data in an unsupervised manner. Using the modulation representation learning as the goal, we explore various neural architecture for unsupervised learning. These include restricted Boltzmann machines (RBM), variational auto-encoders (VAE) and generative adversarial networks (GAN). For learning modulation representations that are distinct and irredundant, we propose different learning frameworks like external residual approach, skip-connection based approach, and a modified cost function-based approach. The methods developed for rate and scale representation learning are benchmarked using an automatic speech recognition (ASR) task on noisy and reverberant conditions. We also illustrate that the unsupervised representation learning can be extended to the first stage of learning time-frequency representations from raw waveforms. The talk will conclude with extensions of the proposed techniques for speech and audio processing.
Speaker Biodata : Purvi Agrawal is a PhD scholar in Learning and Extraction of Acoustic Patterns (LEAP) lab with Dr. Sriram Ganapathy, Dept. of Electrical Engineering, Indian Institute of Science (IISc), Bengaluru. Prior to joining IISc, she obtained her Master of Technology in Communications from DA-IICT, Gandhinagar in 2015. She has also interned in Sony R&D Audio Labs, Tokyo in 2017. Her research interests include interpretable deep learning, raw waveform modeling, low-resource data modeling, multi-modal and biologically inspired deep learning.

Details of Seminars/Colloquium/Defence

Event : Thesis Defence
Title : Tuning of Multi-Band Power System Stabilizers in Multi-Machine Power Systems
Speaker : Aditya Shankar kar
Degree Registered :MTech(Res)
Advisor : Dr. Gurunath Gurrala
Date : 18/12/2020
Venue : Online
Abstract : Intermittent nature of renewables acts as a frequent trigger for small signal oscillations in power grid. These oscillatory modes correspond to either the rotor modes associated with the synchronous machines of conventional generators or control interactions due to the power electronic interfaces in renewable sources. Frequent tuning of power system stabilizers (PSS) of synchronous machines becomes inevitable to maintain small signal stability of the grid over a wide range of operating and system conditions. Multi-band power system stabilizers (MB-PSS or IEEE-PSS4B) play a very important role in such scenarios as they provide separate compensators for different frequency bands covering a wide frequency range. However, tuning the compensators of MB-PSS becomes very challenging due its complex structure.
The MB-PSS uses three separate compensator blocks for low frequencies (0.01Hz to 0.1 Hz), intermediate frequencies (0.1Hz to 0.8 Hz) and high frequencies (0.8Hz to 4 Hz). Electrical power and speed are used as inputs to the MB-PSS. In this thesis each band of an MB-PSS is viewed as a conventional speed input PSS. A systematic approach for tuning MB-PSS is proposed in this thesis based on conventional tuning approach using phase compensation of GEP(s) transfer function. The gain of each band is selected to compensate the GEP(s) in the respective frequency band so that pure damping torque is achieved. A new gain, common for all bands, is introduced for achieving desired damping of the rotor modes. The effectiveness of the tuning PSS under weak and strong system conditions is evaluated using a single machine infinite bus test system (SMIB). It is found that phase compensation under strong system and gain selection to produce 10% to 15% damping under weak system conditions is found to provide better damping performance over a wide range of operating and system conditions. The tuning methodology is extended to multi machine power systems by representing each machine as an equivalent single machine infinite bus. Widely used test systems 4 Generator 9 bus system, 5 generator 10 bus system and 10 generator 39 bus systems are used to evaluate the proposed tuning approach.
This thesis also focuses on the development of a generalized FPGA platform for implementing different IEEE PSS types defined in IEEE STD 421.5. A low-cost open source credit card sized supercomputer called Parallella has been used in the development of PSS. It contains a dual-core ARM-A9 + FPGA Zynq SoC and a 16-core Epiphany co-processor. PSS types which use speed and electric power as input are considered for realization in FPGA as they cover majority of the PSS types. Implementation of PSS is broken down to smaller independent structural blocks which are common across different PSS types. Development of the structural blocks and a process controller, which assembles a desired PSS type functionality, is described. A separate module is built to interface the PSS-FPGA controller and any external micro-controller of an excitation system. All the developments are optimized to ensure minimum resource utilization. The developed framework in FPGA is utilized to implement a speed PSS for a laboratory synchronous machine with a buck converter excitation system.

Event : Thesis Colloquium
Title : Acoustic-Articulatory Mapping: Analysis and Improvements with Neural Network Learning Paradigms
Speaker : Aravind Illa
Degree Registered : PhD
Advisor : Prof. Prasanta Kumar Ghosh
Date : 18/12/2020
Venue : Online
Abstract : Human speech is one of many acoustic signals we perceive, which carries linguistic and paralinguistic (e.g: speaker identity, emotional state) information. Speech acoustics are produced as a result of different temporally overlapping gestures of speech articulators (such as lips, tongue tip, tongue body, tongue dorsum, velum, and larynx) each of which regulates constriction in different parts of the vocal tract. Estimating speech acoustic representations from articulatory movements is known as articulatory-to-acoustic forward (AAF) mapping i.e., articulatory speech synthesis. While estimating articulatory movements back from the speech acoustics is known as acoustic-to-articulatory inverse (AAI) mapping. These acoustic-articulatory mapping functions are known to be complex and nonlinear.
Complexity of this mapping depends on a number of factors. These include the kind of representations used in the acoustic and articulatory spaces. Typically these representations capture both linguistic and paralinguistic aspects in speech. How each of these aspects contributes to the complexity of the mapping is unknown. These representations and, in turn, the acoustic-articulatory mapping are affected by the speaking rate as well. The nature and quality of the mapping varies across speakers. Thus, complexity of mapping also depends on the amount of the data from a speaker as well as number of speakers used in learning the mapping function. Further, how the language variations impact the mapping requires detailed investigation. This thesis analyzes few of such factors in detail and develops neural network based models to learn mapping functions robust to many of these factors.
Electromagnetic articulography (EMA) sensor data has been used directly in the past as articulatory representations (ARs) for learning the acoustic-articulatory mapping function. In this thesis, we address the problem of optimal EMA sensor placement such that the air-tissue boundaries as seen in the mid-sagittal plane of the real-time magnetic resonance imaging (rtMRI) is reconstructed with minimum error. Following optimal sensor placement work, acoustic-articulatory data was collected using EMA from 41 subjects with speech stimuli in English and Indian native languages (Hindi, Kannada, Tamil and Telugu) which resulted in a total of ~23 hours of data, used in this thesis. Representations from raw waveform are also learnt for AAI task using convolutional and bidirectional long short term memory neural networks (CNN-BLSTM), where the learned filters of CNN are found to be similar to those used for computing Mel-frequency cepstral coefficients (MFCCs), typically used for AAI task. In order to examine the extent to which a representation having only the linguistic information can recover ARs, we replace MFCC vectors with one-hot encoded vectors representing phonemes, which were further modified to remove the time duration of each phoneme and keep only phoneme sequence. Experiments with phoneme sequence using attention network achieve an AAI performance that is identical to that using phoneme with timing information, while there is a drop in performance compared to that using MFCC.
Experiments to examine variation in speaking rate reveal that, the errors in estimating the vertical motion of tongue articulators from acoustics with fast speaking rate, is significantly higher than those with slow speaking rate. In order to reduce the demand for data from a speaker, low resource AAI is proposed using a transfer learning approach. Further, we show that AAI can be modeled to learn acoustic-articulatory mappings of multiple speakers through a single AAI model rather than building separate speaker-specific models. This is achieved by conditioning an AAI model with speaker embeddings, which benefits AAI in seen and unseen speaker evaluations. Finally, we show the benefit of estimated ARs in voice conversion application. Experiments revealed that ARs estimated from speaker independent AAI preserves linguistic information and suppress speaker-dependent factors. These ARs (from unseen speaker and language) are used to drive target speaker specific AAF to synthesis speech, which preserves linguistic information and target speaker’s voice characteristics.

Event : Thesis Defence
Title : AlGaN/GaN Heterojunctions Based Hall Sensors for Magnetic Field Sensing over Wide Temperature Range
Speaker : Sagnik Kumar
Degree Registered :MTech(Res)
Advisor : Prof. G. Narayanan
Date : 16/12/2020
Venue : Online
Abstract : Most commercially available Hall sensors are based on silicon and have a limited operating temperature range. For field sensing at cryogenic temperatures or at extremely high temperatures, wide band-gap based materials offer a viable alternative. This thesis evaluates the viability of Hall-effect sensing using AlGaN-GaN based heterojunctions grown on Si substrates for extreme temperatures. In particular, this thesis deals with the fabrication methodologies of Hall sensing elements of different geometries and areas, fabrication of square and Greek-cross shaped Hall sensors, experimental investigations on the variations of sensitivity, offset voltage and geometrical correction factors of the Hall sensing elements with temperature. Additionally, the design of a micro-controller based electronic subsystem for offset compensation, signal amplification, filtering, data acquisition, and display are also part of this study.
Hall-effect sensors are fabricated using AlGaN/GaN heterojunctions grown on Si substrates. Several geometries are possible for the Hall sensing elements. The square-shape is the simplest to fabricate while the Greek-cross shape has a very high geometrical correction factor. Sensors with large square-shaped active areas of side 1.25 cm are realised using a simple fabrication methodology. On the other hand, an array of small sensors is batch fabricated on a single, large wafer employing a 14-step process. The batch fabricated samples are of the Greek-cross shaped geometry with active areas of the order of 0.1 mm x 0.1 mm.
The fabricated Hall sensors are characterised extensively over temperatures ranging from 75 K to 500 K and at applied magnetic field strengths up to 2 Tesla. Sheet resistance and Hall voltages are measured at various operating conditions for multiple Hall sensing samples. Despite sample-to-sample variations, the sensitivity of each sample is observed to remain fairly stable over the wide temperature range of 75 K to 500 K. Temperature compensation can be provided for the minor variations in sensitivity. However, a temperature sensor is not required for this purpose; one of the transresistances or the sheet resistance itself (obtainable through terminal measurements) can be used as a measure of temperature. The geometrical correction factor, another parameter of interest in this study, is shown to remain fairly constant over wide operating ranges. The offset in the sensed Hall signal is also shown to depend on the operating conditions and bias current. However, the offset is shown to get nullified by averaging the output voltages obtained by interchanging the current injection and voltage sense terminals. The method of current spinning is used for offset compensation and is shown to be effective at all operating conditions.Thus, the 2DEG in an AlGaN/GaN heterojunction proves to be a viable alternative for magnetic field sensing both in cryogenic environments as well as in extremely high temperature conditions. The operating range spans those of the military grade, industrial grade and commercial grade Hall-effect devices.
A micro-controller based electronic subsystem has been designed to cancel out the offsets in the Hall signal using current spinning techniques. In addition, the electronic subsystem performs the necessary signal amplification and provides a visual read-out of the sensed magnetic field. The square-shaped sensing element along with the signal conditioning circuit has been suitably packaged for use as a field sensing probe. The probe is tested and calibrated using known magnetic fields in the air-gap of an electromagnet.

Event : Thesis Defence
Title : Protection Schemes for maintaining the Coordination Time Interval between the relays in Micro-grids
Speaker : Vinod V
Degree Registered : PhD
Advisor : Dr U Jayachandra Shenoy
Date : 14/12/2020
Venue : Online
Abstract : Conventional power system alone cannot meet the ever growing needs of electrical power in the world in a reliable manner. Under these circumstances, large penetration of Distributed generators in the distribution level of the power system network is a great boon to the consumers. Due to the penetration of distributed generators (DGs) of large capacities, the difference of current seen by protective relays, during faults in their primary and backup regions is considerable. As a result, the settings of the relay curve suitable for its primary region, is not capable of tripping within the desired time during fault in its backup region. This could lead to several issues with the protection system such as relay miscoordination, delay in relay operation and also failure in the relay operation(blinding). This thesis discusses a methodology to solve such issues by identifying the current status of micro-grid configuration. The proposed communication based method works well by maintaining the Coordination Time Interval (CTI) between the relays, irrespective of the source strength (weak or strong), or during different mode of operation of the micro-grid (islanded or grid connected) as well as for different micro-grid configuration (radial or looped). Even when fault current is less than the load current, the suggested scheme is able to trip within the desired time and provides proper relay coordination. The performance of the communication based relay coordination is validated by comparing the updated relay tripping time with conventional operating time of the relay. The proposed scheme has been emulated in hardware and the results have been validated with the simulation results.
The thesis also explains a relaying scheme, which uses a classification approach using multi-class Support Vector Machine (SVM) for improving the accuracy in relay coordination. Without using any extensive communication facility, the proposed scheme utilizes the local information to identify the corresponding SVM class to coordinate the relays so as to maintain the CTI within the desired limits. The methodology includes a simplified multi-class SVM to classify different classes within the sampling interval. Suitable relay logic based on One Against One (OAO) multi-class SVM classifier has been designed to find the SVM class with appropriate relay trip time. Simulation results show that irrespective of micro-grid mode of operation and the configuration, the proposed scheme achieves better relay coordination as compared to the conventional overcurrent relays. Prototype of the proposed relay is tested on a 400V 4-Bus laboratory type micro-grid system by creating different types of faults and the results obtained demonstrate the efficacy of the proposed method.

Event : Thesis Defence
Title : Strategies for Handling Large Vocabulary and Data Sparsity Problems for Tamil Speech Recognition
Speaker : Madhavaraj A
Degree Registered : PhD
Advisor : Prof. A. G. Ramakrishnan
Date : 08/12/2020
Venue : Online
Abstract : In this work, we focus on the design and development of an automatic speech recognition (ASR) system for Tamil and various experiments conducted in order to enhance its performance. We develop various techniques ranging from subword modeling, multilingual training, speaker adaptation, and scattering transform-based feature extraction/acoustic modeling to improve the performance of the Tamil ASR system.
Subword modeling:
(i) We handle the unlimited vocabulary problem arising due to the deep morphology, agglutination, and inflection properties of Tamil language. (ii) We develop various techniques based on maximum likelihood (using the expectation-maximization algorithm) and Viterbi estimation techniques to segment each word into its constituent subwords.
(iii) We also develop a novel word segmentation technique using weighted finite-state transducer framework by manually encoding the word-formation rules of Tamil. Such a technique can also be easily extended for other NLP applications. Multilingual training for handling data sparsity problem: (i) We address the data sparsity problem using multilingual training of deep neural network (DNN) based acoustic model by leveraging acoustic information from transcribed speech corpus of other source languages.
(ii) In the first method named "Data pooling with phone mapping", we train the DNN acoustic model with features and senones from target language as well as from other source languages by pooling them together.
(iii) In the second method named "Multitask DNN", we train a DNN with features from source as well as target languages to predict the senones of each language in separate output layers.
(iv) We propose techniques to modify phones/senones to appropriately train the DNN with pooled data, and also propose modifications to the cross-entropy loss function to train the multitask-DNN.
Speaker adaptation by modeling coactivation statistics:
(i) We develop techniques based on asymptotic Bayesian approximation and derive a loss function and use it to adapt the speaker-independent DNN (SI-DNN) using very little speaker-specific adaptation data.
(ii) We derive a novel loss function based on the prior coactivation statistics of hidden nodes of the SI-DNN so that during adaptation, the adapted DNN does not deviate far away from the SI-DNN.
(iii) The proposed loss function uses the advantage of cross-entropy loss and it can as well be viewed as a generalization of the well-known center loss function.
Scattering transform-based feature extraction and acoustic modeling:
(i) We study the effect of using scattering transform in the feature extraction stage of ASR.
(ii) We also propose different DNN architectures using 1-D and 2-D convolution layers to predict the senones directly from the raw speech waveform.
(iii) The convolution layers are initialized with 1-D and 2-D Gabor filterbank coefficients such that the intermediate layers of the DNN learn a representation similar to that of scattering transform-based features.

Event : Thesis Colloquium
Title : Shape-constrained Biomedical Image Segmentation and Applications
Speaker : Harish Kumar J R
Degree Registered :PhD
Advisor : Prof. Chandra Sekhar Seelamantula
Date : 20/11/2020
Venue : Online
Abstract : The detection, segmentation, and delineation of the targeted regions of interest in biomedical images are fundamental steps for computer-aided assessment and prescreening. In this thesis, we focus on shape-constrained active contours explicitly due to their shape-specificity, efficiency and low computational cost. In particular, they need fewer degrees-of-freedom than other active contour approaches. They are simpler to formulate and optimize. Variants of concentric shape-constrained active contour representations include circular active disc, elliptical active disc, and active oblong. The degrees of freedom include scaling, translation and rotation. The energy is specified using a contrast function and the optimization is carried out using gradient-descent/ascent technique or its faster variants; use Green’s theorem to make the optimization computationally efficient.
The thesis addresses segmentation problems for two image classes: retinal fundus images in ophthalmology and ultrasound images of the carotid artery. In retinal fundus image analysis, we consider optic disc, optic cup segmentation, glaucoma diagnosis, creation of a glaucoma specific database, fovea segmentation, oxygen saturation measurement from dual-wavelength oximetry images. In the carotid artery segmentation problem, we consider both transverse-mode and longitudinal model images and develop methods for segmenting the lumen intima and the media adventitia boundary.
To start with, we present a reliable and fully automated method for the segmentation and outlining of the optic disc and cup using retinal fundus images with relevant parameters for glaucoma prescreening. We calculate the cup-to-disc ratio (CDR) and rim-to-disc ratio (RDR) from the segmented optic disc and cup. We perform two-stage glaucoma classification using CDR following the International Classification of Diseases (ICD) rules and RDR value following the new disc-damage-likelihood-scale (DDLS) rule. We categorize the glaucomatous condition as normal, moderate, or severe. In addition, we incorporate a check for the pattern of decreasing rim-widths of inferior, superior, nasal, and temporal (ISNT) regions to differentiate normal fundus from a glaucomatous one. We have validated our glaucoma prescreening technique on publicly available as well as locally obtained fundus image databases. The number of retinal fundus images used for the validation of optic disc and cup segmentation is 1597 and 436, respectively, and is 436 for glaucoma severity grading. The algorithm performance is validated against expert clinician outlining and quantitative comparison is provided using Jaccard and Dice similarity measures. The tool is Java/Android/iOS-based, repeatable, easy to use, provides quantitative analysis, and takes only a few seconds per image for the diagnosis. Our solution is available as a smartphone app, “NAYANA,” which assesses severity of glaucoma by computing CDR, RDR, and ISNT parameters. The app is available in both Android and iOS versions. We have tested these versions in a hospital setup with the aforementioned set of algorithm parameters. The results showed that the system could diagnose the severity of glaucoma reliably.
As part of this dissertation, we have created a new, comprehensive, and the largest glaucoma-specific retinal fundus images database (with 1500 images) containing images of both glaucomatous and normal eyes. We provide optic disc and cup manual segmentation ground-truth and a decision on glaucoma by five expert ophthalmologists. The database has been created with the help of research funding from Govt. of India’s IMPacting Research, InnovatioN and Technology (IMPRINT) - India initiative. It is a result of collaboration with five expert ophthalmologists to support comparative studies on automatic OD and OC segmentation algorithms using retinal fundus images and subsequent decision on glaucoma diagnosis.
We present an automated technique for the measurement of oxygen saturation in retinal arterioles and venules using dual-wavelength retinal oximetry images. The technique is based on segmenting an optic disc centered ring-shaped region of interest and subsequent analysis of the oxygen saturation levels. The two dominant peaks in the histogram of the oxygen saturation levels correspond to arteriolar and venular oxygen saturations from which the arterio-venous saturation difference (AVSD) can be calculated. For evaluation, we have created a normative database of Asian Indian eyes containing forty four dual- wavelength retinal oximetry images. Validations against expert manual annotations exhibit high consistency across the dataset indicating that the automated technique is an accurate alternative to the manual procedure.
Proceeding further, we also address the problem of fovea segmentation and develop a technique for delineation of macular regions based on the active disc formalism. We report validation results on three publicly available fundus image databases, amounting to a total of 1370 fundus images for automatic fovea localization and 370 fundus images for fovea segmentation and macular region delineation.
With regard to the carotid artery segmentation problem, we develop an automated outlining technique based on circular active disc for outlining the media adventitia boundary and elliptical active disc for outlining the lumen intima boundary of the common carotid artery. We report results of experimental validation on Brno University’s Signal Processing (SP) lab database, which contains 971 transverse mode ultrasound images of the carotid artery. The database provides manual annotations, which are also circular and serve as the ground-truth/reference. The segmentation of the lumen intima layer has been accomplished with the use of elliptical active disc and the performance has been validated on the SP lab database. Comparisons with other state-of-the-art techniques are also reported.
We then consider the longitudinal-mode ultrasound images of the common carotid artery and address the segmentation of the lumen intima layer. The method is hybrid in the sense that a coarse segmentation is first achieved by optimizing a locally defined contrast function of an active oblong with five degrees-of-freedom. Subsequently, fine segmentation and delineation of the carotid artery are achieved by post-processing the portion of the ultrasound image spanned by the annulus region of the optimally fitted active oblong followed by a cubic curve-fitting technique to delineate the lumen intima boundary. The algorithm has been validated on 84 common carotid artery longitudinal-mode ultrasound images provided by the SP lab, Brno university. The segmentation results of the proposed technique exhibit a good correlation with the ground-truth annotations provided by two expert radiologists.
Finally, we develop a generalized formulation to derive circle/ellipse/rectangle and other shapes using active lp-balls. The concentric-shape strategy, energy function and optimization techniques follow the active disc model. By using one more parameter than required for a rectangle or an ellipse, we are able to significantly increase the degree of flexibility and cover a parametric family of shapes including rectangles and ellipses. We have demonstrated the suitability of active lp-balls for shape-specific segmentation using synthetic images. Applicability of active lp-balls for suitable biomedical applications remains to be explored.
In summary, this thesis focuses on various aspects of shape-constrained active contours and demonstrates concrete applications to retinal fundus image segmentation, retinal oximetry assessment, and common carotid artery segmentation in transverse and longitudinal mode ultrasound images.
Speaker Biodata :

Event : Thesis Colloquium
Title : Investigation on the Influence of soil's electrical parameters on the lightning stroke-current evolution and the fields in close range
Speaker : Rupam Pal
Degree Registered : PhD
Advisor : Prof. Udaya Kumar
Date : 20/11/2020
Venue : Online
Abstract : The lightning return stroke forms one of the severest natural sources of electromagnetic interference for systems, both in the air and soil. Several physical fields govern this complex physical phenomenon, and most of the engineering applications resort to much simplifications. Several pertinent aspects are somewhat unclear, and it is not practical to conduct the field measurements to answer them. One such important aspect, which is of practical relevance, is the influence of soil's electrical properties on the stroke current evolution and the fields in the close range. It is investigated in the present work.

Among different models for the lightning return stroke, only the 'Self-consistent return stroke' model is found to be suitable for the intended work. This model employs a macroscopic electrical representation of the underlying physical phenomenon to emulate the stroke current evolution. However, this model has considered only a perfectly conducting earth and relied on the time-domain thin-wire formulation to solve for the associated dynamic electromagnetic fields. On the other hand, a more realistic representation of the soil, including its dispersive nature and non-linearity, is required for the present work. This necessitated suitable adoption of the 'Finite difference time domain' (FDTD) method for modeling the channel and its corona sheath, soil-ionization, and soil-dispersion.

The developed FDTD formulation is used to investigate and ascertain the role of soil's electrical properties on the stroke current evolution and the field in the soil. For the first time, it is shown that the soil's electrical conductivity has some noticeable influence on the stroke current magnitude, and the ionization phenomenon in soil tends to reduce this influence. The dispersive nature of the soil's conductivity, and permittivity to a lesser extent, significantly reduces the field in the soil. The current concentration near the surface, which is expected for the skin-effect phenomenon, is altered at later periods by the field produced by the channel current. Also, the normal field in the soil changes its polarity. The vertical stratification of the soil, which is expected near the water body-soil interface, influences the field in the soil quite significantly. A strike to a model mountain leads to an entirely different field structure beyond its base. Similarly, a strike to a tall tower produces a field in the soil, which is bipolar near the tower base. These are quite novel findings, and many of them are somewhat unexpected.

In summary, significant contributions have been made towards the FDTD formulations for modeling lightning phenomena and finding the role of soil's electrical parameters on lightning stroke current evolution and the resulting field.

Event : Thesis Colloquium
Title : Parallel Computing Techniques for High Speed Power System Solutions
Speaker : R Gnanavignesh
Degree Registered : PhD
Advisor : Dr. U J Shenoy
Date : 19/10/2020
Venue : Online
Abstract : Modern power systems are enormously large and complex entities. Increase in system size, introduction of complex controls, uncertainties in forecasting, etc. necessitate faster software tools to handle power system planning, operation and operator training. This thesis aims to improve the performance of power system software tools by proposing parallel algorithms with the objective of reducing their execution time.
Solution of a sparse set of linear algebraic equations is one of the most essential modules used in almost all power system software tools. The thesis addresses the issue of reducing the execution time of sparse linear algebraic solver by parallelizing sparse matrix factorization. A LU factorization algorithm which is more amenable for parallelization is identified and chosen. In this work, the structural symmetry property of power system sparse matrices is exploited to maximize the column or node level parallelism. Results obtained from the implementation of the proposed algorithm on Graphical Processing Units (GPUs) corroborate its efficacy by achieving significant reduction in the solution time when compared with state of the art CPU based sequential sparse linear solvers.
Power flow algorithm is one of the most frequently executed algorithms with respect to the steady state realm of the power system. Reduction in the solution time for the power flow algorithm would further boost other applications like contingency analysis, optimal power flow, dynamic studies, etc. This thesis proposes a parallel power flow algorithm based on Newton-Raphson method. Inclusion of reactive power limit constraints at generator buses in the problem formulation stage itself eradicates the need to use heuristic techniques. In this work, the given power system network for which the power flow solution is desired, is decomposed into smaller sub-networks and processed in an independent as well as in a concurrent manner. Partial results from the sub-networks are consolidated to arrive at the solution of original network. Results obtained indicate preservation of the superior convergence property of Newton-Raphson method and a significant reduction in the solution time required for the parallel version of the power flow when compared with the sequential version.
Transient stability assessment is an important module within the Dynamic Security Assessment application. Time domain simulation for the stability assessment by solving thousands of Differential Algebraic Equations (DAEs), even though is the preferred method, is computationally intensive and becomes a major computing challenge as system size increases. The thesis proposes a parallel algorithm based on spatial domain decomposition employing relaxation conditions to speedup the transient stability simulation to handle the aforementioned challenge. A convergence enhancing mechanism through selection of appropriate admittance parameters for the network emulating fictitious buses which mimic the remainder of the system for each sub-network is derived. Results obtained corroborate the scalability and improved speedup features of the methodology which achieves a significant reduction in the simulation execution time.

Event : Thesis Colloquium
Title : Probabilistic source-filter model of speech
Speaker : Achuth Rao M V
Degree Registered : PhD
Advisor : Prof. Prasanta Kumar Ghosh
Date : 16/10/2020, 2.30 pm
Venue : online
Abstract : The human respiratory system plays a crucial role in breathing and swallowing. However, it also plays an essential role in speech production, which is unique to humans. Speech production involves expelling air from the lungs. As the air flows from the lungs to the lips, some kinetic energy gets converted to sound. Different structures modulate the generated sound, which is finally radiated out of the lips. The speech consists of various information such as linguistic content, speaker identity, emotional state, accent, etc. Apart from speech, there are various scenarios where the sound is generated in the human respiratory system. These could be due to abnormalities in the muscles, motor control unit, or the lungs, which can directly affect generated speech as well. A variety of sounds are also generated by these structures while breathing including snoring, Stridor, Dysphagia, and Cough.
The source filter (SF) model of speech is one of the earlier models of speech production. It assumes that speech is a result of filtering an excitation or source signal by a linear filter. The source and filter are assumed to be independent. Even though the SF model represents the speech production mechanism, there needs to be a tractable way of estimating the excitation and the filter. The estimation of both of them given speech falls under the general category of signal deconvolution problem, and, hence, there is no unique solution. There are several variations of the source-filter model in the literature by assuming different structures on the source/filter. There are various ways to estimate the parameters of the source and the filter. The estimated parameters are used in various speech applications such as automatic speech recognition, text to speech, speech enhancement etc. Even though the SF model is a model of speech production, it is used in applications including Parkinson's Disease classification, asthma classification.
The existing source filter models show much success in various applications, however, we believe that the models mainly lack two respects. The first limitation is that these models lack the connection to the physics of sound generation or propagation. The second limitation of the current models is that they are not fully probabilistic. The inherent nature of the airflow is stochastic because of the presence of turbulence. Hence, probabilistic modeling is necessary to model the stochastic process. The probabilistic models come with several other advantages: 1) systematically inducing the prior knowledge into the models through probabilistic priors, 2) the estimation of the uncertainty of the model parameters, 3) allows sampling of new data points 4) evaluation of the likelihood of the observed speech.
We start with the governing equation of sound generation and use a simplified geometry of the vocal folds. We show that the sound generated by the vocal folds consists of two parts. The first part is because of the difference between the subglottal and supra glottal pressure difference. The second part is because of the sound generated by turbulence. The first kind is dominant in the voiced sounds, and the second part is dominant in the unvoiced sounds. We further assume the plane wave propagation in the vocal tract, and there is no feedback from the vocal tract on the vocal folds. The resulting model is the excitation passing through an all-pole filter, and the excitation is the sum of two signals. The first signal is quasi-periodic, and the shape of each cycle depends on the time-varying area of the glottis. The second part is stochastic because the turbulence is modeled as a white noise passed through a filter. We further convert the model into a probabilistic one by assuming the following distribution on the excitations and filters. We model the excitation using a Bernoulli Gaussian distribution. Filter coefficients are modeled using the Gaussian distribution. The noise distribution is also Gaussian. Given these distributions, the likelihood of the speech can be derived as a closed-form expression. Similarly, we impose an appropriate prior to the model’s parameters and make a maximum a posteriori estimation of the parameters. But the model assumption can be changed/approximated with respect to the application and resulting in different estimation procedures. To validate the model, we apply this model to seven applications as follows:

Event : Thesis Colloquium
Title : Tuning of Multi-Band Power System Stabilizers in Multi-Machine Power Systems
Speaker : Aditya Shankar Kar
Degree Registered : MTech (Research)
Advisor : Dr. Gurunath Gurrala
Date : 31/08/2020
Venue : Online
Abstract : Intermittent nature of renewables acts as a frequent trigger for small signal oscillations in power grid. These oscillatory modes correspond to either the rotor modes associated with the synchronous machines of conventional generators or control interactions due to the power electronic interfaces in renewable sources. Frequent tuning of power system stabilizers (PSS) of synchronous machines becomes inevitable to maintain small signal stability of the grid over a wide range of operating and system conditions. Multi-band power system stabilizers (MB-PSS or IEEE-PSS4B) play a very important role in such scenarios as they provide separate compensators for different frequency bands covering a wide frequency range. However, tuning the compensators of MB-PSS becomes very challenging due its complex structure. The MB-PSS uses three separate compensator blocks for low frequencies (0.01Hz to 0.1 Hz), intermediate frequencies (0.1Hz to 0.8 Hz) and high frequencies (0.8Hz to 4 Hz). Electrical power and speed are used as inputs to the MB-PSS. In this thesis each band of an MB-PSS is viewed as a conventional speed input PSS. A systematic approach for tuning MB-PSS is proposed in this thesis based on conventional tuning approach using phase compensation of GEP(s) transfer function. The gain of each band is selected to compensate the GEP(s) in the respective frequency band so that pure damping torque is achieved. A new gain, common for all bands, is introduced for achieving desired damping of the rotor modes. The effectiveness of the tuning PSS under weak and strong system conditions is evaluated using a single machine infinite bus test system (SMIB). It is found that phase compensation under strong system and gain selection to produce 10% to 15% damping under weak system conditions is found to provide better damping performance over a wide range of operating and system conditions. The tuning methodology is extended to multi machine power systems by representing each machine as an equivalent single machine infinite bus. Widely used test systems 4 Generator 9 bus system, 5 generator 10 bus system and 10 generator 39 bus systems are used to evaluate the proposed tuning approach. This thesis also focuses on the development of a generalized FPGA platform for implementing different IEEE PSS types defined in IEEE STD 421.5. A low-cost open source credit card sized supercomputer called Parallella has been used in the development of PSS. It contains a dual-core ARM-A9 + FPGA Zynq SoC and a 16-core Epiphany co-processor. PSS types which use speed and electric power as input are considered for realization in FPGA as they cover majority of the PSS types. Implementation of PSS is broken down to smaller independent structural blocks which are common across different PSS types. Development of the structural blocks and a process controller, which assembles a desired PSS type functionality, is described. A separate module is built to interface the PSS-FPGA controller and any external micro-controller of an excitation system. All the developments are optimized to ensure minimum resource utilization. The developed framework in FPGA is utilized to implement a speed PSS for a laboratory synchronous machine with a buck converter excitation system.

Event : Thesis Second Colloquium
Title : AlGaN/GaN Heterojunctions Based Hall Sensors for Magnetic Field Sensing over Wide Temperature Range
Speaker : Sagnik Kumar
Degree Registered : MTech (Research)
Advisor : Prof. G Narayanan
Date : 17/08/2020
Venue : Online.
Abstract : Most commercially available Hall sensors are based on silicon and have a limited operating temperature range. For field sensing at cryogenic temperatures or at extremely high temperatures, wide band-gap based materials offer a viable alternative. This thesis evaluates the viability of Hall-effect sensing using AlGaN-GaN based heterojunctions grown on Si substrates for extreme temperatures. In particular, this thesis deals with the fabrication methodologies of Hall sensing elements of different geometries and areas, fabrication of square and Greek-cross shaped Hall sensors, experimental investigations on the variations of sensitivity, offset voltage, and geometrical correction factors of the Hall sensing elements with temperature. Additionally, the design of a micro-controller based electronic subsystem for offset compensation, signal amplification, filtering, data acquisition, and display are also part of this study.
Hall-effect sensors are fabricated using AlGaN/GaN heterojunctions grown on Si substrates. Several geometries are possible for the Hall sensing elements. The square-shape is the simplest to fabricate while the Greek-cross shape has a very high geometrical correction factor. Sensors with large square-shaped active areas of side 1.25 cm are realised using a simple fabrication methodology. On the other hand, an array of small sensors is batch fabricated on a single, large wafer employing a 14-step process. The batch fabricated samples are of the Greek-cross shaped geometry with active areas of the order of 0.1 mm x 0.1 mm.
The fabricated Hall sensors are characterised extensively over temperatures ranging from 75 K to 500 K and at applied magnetic field strengths up to 2 Tesla. Sheet resistance and Hall voltages are measured at various operating conditions for multiple Hall sensing samples. Despite sample-to-sample variations, the sensitivity of each sample is observed to remain fairly stable over the wide temperature range of 75 K to 500 K. Temperature compensation can be provided for the minor variations in sensitivity. However, a temperature sensor is not required for this purpose; one of the transresistances or the sheet resistance itself (obtainable through terminal measurements) can be used as a measure of temperature. The geometrical correction factor, another parameter of interest in this study, is shown to remain fairly constant over wide operating ranges. The offset in the sensed Hall signal is also shown to depend on the operating conditions and bias current. However, the offset is shown to get nullified by averaging the output voltages obtained by interchanging the current injection and voltage sense terminals. The method of current spinning is used for offset compensation and is shown to be effective at all operating conditions. Thus, the 2DEG in an AlGaN/GaN heterojunction proves to be a viable alternative for magnetic field sensing both in cryogenic environments as well as in extremely high-temperature conditions. The operating range spans those of the military-grade, industrial-grade and commercial -grade Hall-effect devices.
A micro-controller based electronic subsystem has been designed to cancel out the offsets in the Hall signal using current spinning techniques. In addition, the electronic subsystem performs the necessary signal amplification and provides a visual read-out of the sensed magnetic field. The square-shaped sensing element along with the signal conditioning circuit has been suitably packaged for use as a field sensing probe. The probe is tested and calibrated using known magnetic fields in the air-gap of an electromagnet.

Event : Thesis Colloquium
Title : Efficient Algorithms for Learning Restricted Boltzmann Machines
Speaker : Vidyadhar Upadhya
Degree Registered : PhD
Advisor : Prof. P S Sastry
Date : 07/08/2020
Venue : Microsoft Teams
Abstract : The probabilistic generative models learn useful features from unlabeled data. The Restricted Boltzmann Machine (RBM) is an important probabilistic generative model which also forms the building block for several deep generative models. It is difficult to train and evaluate RBMs mainly because the normalizing constant (known as the partition function) for the distribution that they represent is computationally hard to evaluate. Therefore, various approximate methods (based on noisy gradient of the log likelihood estimated through sampling) are used to train RBMs. In this thesis we discuss some new algorithms for efficient learning of both binary-binary and Gaussian-binary RBMs.
We propose a new algorithm for learning binary-binary RBMs by exploiting the property that the log-likelihood function is a difference of convex functions with respect to its parameters (coordinate-wise). We devise a stochastic variant of the difference of convex functions programming, termed S-DCP, where the convex optimization problem at each iteration is approximately solved through a few iterations of stochastic gradient descent. The resulting algorithm is simple and the contrastive divergence (CD) algorithm, the current standard algorithm for learning RBMs, can be derived as a special case of the proposed algorithm. Moreover, the hyperparameters of the S-DCP algorithm can be chosen such that the computational complexity of S-DCP can be made comparable to that of other CD based algorithms. We further modify this algorithm to accommodate centered gradients. We then propose a second order learning algorithm on the convex S-DCP objective function using diagonal approximation of the Hessian which, we show, can be easily computed with the gradient estimates. To compensate for the noise in the Hessian estimate and to make the algorithm stable, we use an exponential averaging over these estimates. Through extensive empirical studies on a number of benchmark datasets, we demonstrate the superior performance of the proposed algorithms. Our empirical results show that the centered S-DCP as well as the diagonally scaled S-DCP are highly effective and efficient methods for learning RBMs.
We extend the S-DCP algorithm to learn Gaussian-binary RBMs by proving that the Gaussian-binary RBM log-likelihood function is also a difference of convex functions with respect to the weights and the biases of the hidden units under the assumption that the conditional distribution of the visible units have a fixed variance. We further modify the algorithm to allow for learning this variance also. Through extensive empirical studies on a number of benchmark datasets, we demonstrate that S-DCP learns good models more efficiently compared to CD and PCD, the current standard algorithms for learning Gaussian-binary RBMs.
Calculation of the log likelihood of an RBM model on a set of validation samples is intractable due to the partition function. Currently one uses sampling based methods for approximating the log likelihood. We provide an empirical analysis of these sampling based algorithms to estimate the log-likelihood and suggest some simple techniques to improve these estimates.
Speaker Bio :Mr. Vidyadhar Upadhya is a Ph.D. student in the department of Electrical Engineering working with Prof. P. S. Sastry

Event : Thesis Defence
Title : Soft Switched Multilevel Unidirectional High Frequency Link DC to AC Converter for Medium Voltage Grid Integration of Solar Photovoltaics
Speaker : Manmohan Mahapatra
Degree Registered :MTech (Research)
Advisor : Dr. Kaushik Basu
Date : 05/08/2020
Venue : Microsoft Teams
Abstract : Grounding the frame of photovoltaic (PV) panel is a necessity for the safety of humans. This leads to the formation of large capacitance between PV cells and ground. Hence, to reduce leakage current due to parasitic capacitance between PV cells and ground, the DC output voltage of a photovoltaic panel is normally kept below 1 kV. Conventionally, for medium voltage (3.3 kV-66 kV) AC grid integration of PV panel, the DC output of PV is first converted to 400V AC and is connected to a 400V collection grid through a line frequency transformer (LFT). This LFT provides isolation and limits circulating current among the PV modules. Another step-up LFT is used to connect 400V AC grid to the medium voltage (MV) transmission grid. These line frequency transformers are bulky and expensive. The line side filters are placed on the low voltage side of the first LFT and hence, experience high currents leading to higher copper losses. To avoid the limitations of LFTs, power converters with high frequency transformer (HFT) are becoming popular. The HFT is fed from a DC side inverter (DSI) and the output of HFT (which is high frequency AC) is converted to line frequency AC using power electronic converters. This type of converter is known as high frequency link (HFL) DC to AC converter. State-of-the-art HFL DC to AC converters mostly employ a multi-stage power conversion technique where an isolated DC to DC converter is cascaded with an inverter. The stages are controlled independently. The inter-stage voltage stiff DC-link is maintained with large electrolytic capacitor. But such an approach requires higher amount of filtering and use of electrolytic capacitor affects long-term reliability. Moreover, the capacitor voltage needs to be tightly regulated to protect the devices. The grid interfaced inverter is high frequency hard-switched resulting in reduced efficiency. These drawbacks are overcome in a single-stage power conversion approach where the inter-stage filter capacitor is removed and all the power devices are either soft or line frequency switched resulting in reduction in switching loss and improvement in efficiency.
In literature, to replace the step-up LFT and to directly integrate the converter to the medium voltage grid, a popular solution is the usage of cascaded multilevel power conversion. Generally, the above discussed multi-stage converter is employed as modules in a cascaded multilevel configuration to produce medium voltage. Moreover, some existing topologies use single-stage converters in a cascaded multilevel configuration to produce medium voltage, but the grid side converters are high frequency switched, leading to higher loss. In this thesis, a new topology is proposed to overcome the drawbacks of existing cascaded multilevel power conversion topologies.
In the thesis, a new single-stage high frequency link cascaded multilevel converter topology is proposed for MV grid integration of solar power. A single-stage high frequency link DC to AC converter is used as a module. The DC side of each module is connected to a PV source. The AC sides of multiple such modules are connected in series in a cascaded fashion to interface with the MV AC grid. Proposed modulation of the DC to AC module results in zero voltage switching (ZVS) of the DC side converter and line frequency switching of the AC side converter. ZVS happens for most part of the line cycle. Over a switching cycle, the operation of this module is similar to a phase-shifted full bridge (PSFB) DC to DC converter. In the PSFB converter, during switching transition, the parasitic capacitance of AC side diode bridge along with leakage inductance of HFT forms a resonating circuit. This resonating circuit leads to high voltage stress on the secondary side devices. An active snubber is designed to restrict the voltage overshoot. The operation of PSFB converter, considering all parasitic, is not explored in literature. In this thesis, a detailed analysis of the operation of the PSFB and step-by-step design methodology is given. The hardware is designed and tested with DC input voltage of 400 V, DC output voltage of 1240 V, output power of 1.5 kW and switching frequency of 20 kHz. Experimental results validate the analysis.
A method is proposed to observe medium voltage waveforms with the standard low-voltage probe. A method to remotely control the medium voltage converter is developed to ensure safety.

Event : Thesis Colloquium
Title : Protection Schemes for maintaining the Coordination Time Interval between the relays in Micro-grid
Speaker : V Vinod
Degree Registered : PhD
Advisor : Dr. U J Shenoy
Date : 30/07/2020
Venue : Microsoft Teams
Abstract : Conventional power system alone cannot meet the ever growing needs of electrical power in the world in a reliable manner. Under these circumstances, large penetration of Distributed generators in the distribution level of the power system network is a great boon to the consumers. Due to the penetration of distributed generators (DGs) of large capacities, the difference of current seen by protective relays, during faults in their primary and backup regions is considerable. As a result, the settings of the relay curve suitable for its primary region, is not capable of tripping within the desired time during fault in its backup region. This could lead to several issues with the protection system such as relay miscoordination, delay in relay operation and also failure in the relay operation(blinding). This thesis discusses a methodology to solve such issues by identifying the current status of micro-grid configuration. The proposed communication based method works well by maintaining the Coordination Time Interval (CTI) between the relays, irrespective of the source strength (weak or strong), or during different mode of operation of the micro-grid (islanded or grid connected) as well as for different micro-grid configuration (radial or looped). Even when fault current is less than the load current, the suggested scheme is able to trip within the desired time and provides proper relay coordination. The performance of the communication based relay coordination is validated by comparing the updated relay tripping time with conventional operating time of the relay. The proposed scheme has been emulated in hardware and the results have been validated with the simulation results.
The thesis also explains a relaying scheme, which uses a classification approach using multi-class Support Vector Machine (SVM) for improving the accuracy in relay coordination. Without using any extensive communication facility, the proposed scheme utilizes the local information to identify the corresponding SVM class to coordinate the relays so as to maintain the CTI within the desired limits. The methodology includes a simplified multi-class SVM to classify different classes within the sampling interval. Suitable relay logic based on One Against One (OAO) multi-class SVM classifier has been designed to find the SVM class with appropriate relay trip time. Simulation results show that irrespective of micro-grid mode of operation and the configuration, the proposed scheme achieves better relay coordination as compared to the conventional overcurrent relays. Prototype of the proposed relay is tested on a 400V 4-Bus laboratory type micro-grid system by creating different types of faults and the results obtained demonstrate the efficacy of the proposed method.

Event : Thesis Defence
Title : Power Electronic Converters for Condition Monitoring and Voltage Equalization of Batteries
Speaker : Shimul Kumar Dam
Degree Registered :PhD
Advisor : Prof. Vinod John
Date : 28/07/2020
Venue : Microsoft Teams
Abstract : Power converters are used in battery-based storage systems in many applications. Apart from the task of regulating the charging and discharging, the power electronic converters can also help to monitor battery condition and to avoid over-charge or over-discharge of any battery cell. One approach to monitoring the cell condition is by measuring its impedance. The power converter for charging and discharging of the cell stack can be used for online measurement of cell impedances. The challenges involved in control, measurement, and the hardware requirements for impedance measurement are analyzed in this work, and solutions are proposed. A Proportional Integral Resonant (PIR) controller-based control scheme and a DAC based measurement method are proposed for impedance measurement over the required frequency range. Two different approaches are proposed to achieve sufficient output voltage resolution for generating small amplitude voltage perturbation. One approach achieves high voltage resolution by replacing the single-leg buck converter with a multi-leg interleaved converter. The other approach uses a low-power rated auxiliary converter in series with the main converter to achieve high voltage resolution. Both of the methods are experimentally verified and compared with commercial equipment and the advantages of each approach are evaluated.
A voltage equalizer is a power electronic circuit that equalizes the cell voltages in a series-connected cell stack to avoid over-charge and over-discharge of any individual cell. A low-cost voltage equalizer using selection switches for a cell to cell equalization is proposed. This equalizer uses capacitive voltage level shifting to avoid bulky and lossy isolation transformer and to reduce cost. A new approach with a lower number of low-frequency selection switches further reduces the equalizer cost. A high-performance voltage equalizer is also proposed to achieve fast equalization by direct multi-cell to multi-cell charge transfer. This topology is shown to provide soft-switching with high efficiency. The equalizer is controlled in an open loop. The equalization currents do not reduce with progress in voltage equalization, making this topology faster than the existing open-loop multi-cell to multi-cell topologies. A modularization method is proposed for this topology to provide a direct path for charge transfer from any cell in one module to any cell in another module. The operation of both the equalizers and the modularization technique are experimentally verified which confirms the theoretical analysis.

Event : Thesis Defence
Title : Unidirectional High-Frequency-Link DC to Three Phase AC Conversion: Topology, Modulation and Converter Design.
Speaker : Anirban Pal
Degree Registered :PhD
Advisor : Dr. Kaushik Basu
Date : 27/07/2020
Venue : Microsoft Teams
Abstract :

Event : Thesis Defence
Title : Speech task-specific representation learning using acoustic-articulatory data
Speaker : Mannem Renuka
Degree Registered : MTech(Research)
Advisor : Prof. Prasanta Kumar Ghosh
Date : 17/07/2020
Venue : Online
Abstract : Human speech production involves modulation of the air stream by the vocal tract shape determined by the articulatory configuration. Articulatory gestures are often used to represent the speech units. It has been shown that the articulatory representations contain information complementary to the acoustics. Thus, a speech task could benefit from the representations derived from both acoustic and articulatory data.
A typical acoustic representation consists of spectral and temporal characteristics e.g., Mel Frequency Cepstral Coefficients (MFCCs) Line Spectral Frequencies (LSF), and Discrete Wavelet Transform (DWT). On the other hand, articulatory representations vary depending on how the articulatory movements are captured. For example, when Electro-Magnetic Articulography (EMA) is used, the recorded raw movements of the EMA sensors placed on the tongue, jaw, upper lip, and lower lip and tract variables derived from them have often been used as articulatory representations. Similarly, when real-time Magnetic Resonance Imaging (rtMRI) is used, articulatory representations are derived primarily based on the Air-Tissue Boundaries (ATB) in the rtMRI video. The low resolution and SNR of the rtMRI video makes the ATB segmentation challenging. In this thesis, we propose various supervised ATB segmentation algorithms which include semantic segmentation, object contour detection using deep convolutional networks. The proposed approaches predict ATBs better than the existing baselines, namely, Maeda Grid and Fisher Discriminant Measure based schemes. We also propose a deep fully-connected neural network based ATB correction scheme as a post processing step to improve upon the predicted ATBs. However, articulatory data is not directly available in practice, unlike the speech recording. Thus, we also consider the articulatory representations derived from acoustics using an Acoustic-to-Articulatory Inversion (AAI) method.
Generic acoustic and articulatory representations may not be optimal for a speech task. In this thesis, we consider the speech rate (SR) estimation task, useful for several speech applications and propose techniques for deriving acoustic and articulatory representations for the same. SR is defined as the number of syllables per second in a given speech recording. We propose a Convolutional Dense Neural Network (CDNN) to estimate the SR from directly given as well as learnt acoustic and articulatory representations. In the case of acoustics, the SR is estimated directly using MFCCs. When raw speech waveform is given as input, one-dimensional convolutional layers are utilized to estimate the SR specific acoustic representations. The center frequencies of the learned convolutional filters range from 200 to 1000 Hz unlike MFCC filter bank frequencies which range from 0 to 4000 Hz. The task-specific features are found to perform better in SR estimation compared to the MFCCs.
The articulatory features also help in accurate SR estimation since the characteristics of articulatory motion significantly vary with the changes in the SR. To estimate the SR specific articulatory representations, both the AAI and CDNN models are jointly trained using a weighted loss function which includes loss for the SR estimation and loss for estimating articulatory representations from acoustics. Similar to the acoustics case, the task-specific articulatory representations derived from acoustics perform better in SR estimation compared to the generic articulatory representations. Even though the task-specific articulatory representations derived from acoustics are not identical to the generic articulatory representations, both are found to be low pass in nature. The CDNN based approach using both generic and learnt representations perform better than the temporal and selected subband correlation (TCSSBC) based baseline scheme for SR estimation task.

Event : Thesis Colloquium
Title : Neural Representation Learning for Speech and Audio Signals
Speaker : Purvi Agrawal
Degree Registered :PhD
Advisor : Dr. Sriram Ganapathy
Date : 03/07/2020
Venue : C 241 MMCR, EE
Abstract : Representation learning is the branch of machine learning consisting of techniques that are capable of automatically discovering meaningful representations from raw data for efficient information extraction. In the recent years, following the trends in other streams of machine learning, representation learning using neural networks has attracted significant interest. For example, deep representation learning in the text domain using word embeddings has shown interesting semantic properties that make them widely useful for many natural language processing applications. In the speech processing field, representation learning has been a challenging task. This talk is focused on developing neural methods for representation learning of speech and audio signals, with the goal of improving downstream applications that rely on these representations.
For representation learning, we pursue two broad directions - supervised and unsupervised. In the case of speech/audio signals, we also identify two stages of representation learning. The first stage is the learning of a time-frequency representation (equivalent of spectrogram) from the raw audio waveform. The second stage is the learning of modulation representations (filtering the time-frequency representations along the temporal domain, called rate filtering and spectral domain, called scale filtering).
In the first part of the talk, we propose representation learning methods for speech data in an unsupervised manner. Using the modulation representation learning as the goal, we explore various neural architecture for unsupervised learning like restricted Boltzmann machines (RBM), variational auto-encoders (VAE) and generative adversarial networks (GAN). For learning modulation representations that are distinct and irredundant, we propose different learning frameworks like external residual approach, skip connection based approach, and cost function based approach. The methods developed for rate and scale representation learning are benchmarked using an automatic speech recognition (ASR) task on noisy and reverberant conditions. We also illustrate that the unsupervised representation learning can be extended to the first stage of learning time-frequency representations from raw waveforms.
The second part of the talk deals with supervised representation learning. Here, we propose a two-stage representation learning approach from raw waveform consisting of acoustic filterbank learning (time-frequency representation learning) from raw waveform followed by a modulation representation learning. The key novelty in the proposed framework consists of a relevance weighting mechanism that acts as a feature selection module. This is inspired by gating networks and provides a mechanism to weight the relevance of the acoustic and modulation representations in predicting the target class. The relevance weighting network can also utilize feedback from the previous predictions of the model for tasks like ASR. The proposed relevance weighting scheme is shown to provide significant performance improvements for ASR task and UrbanSound audio classification task. A detailed analysis yields insights into the interesting properties of the relevance weights that are captured by the model at the acoustic and modulation stages for speech and audio signals. The talk will conclude with a comparison between unsupervised and supervised learning schemes for neural representation learning of speech signals.

Event : Thesis Colloquium
Title : Strategies for Handling Large Vocabulary and Data Sparsity Problems for Tamil Speech Recognition
Speaker : Madhavaraj A
Degree Registered :PhD
Advisor : Prof. A. G. Ramakrishnan
Date : 26/06/2020
Venue : Online
Abstract : Significant contributions of the thesis: In this work, we focus on the design and development of an automatic speech recognition (ASR) system for Tamil and various experiments conducted in order to enhance its performance. We develop various techniques ranging from subword modeling, multilingual training, speaker adaptation, and scattering transform-based feature extraction/acoustic modeling to improve the performance of the Tamil ASR system.

Event : Thesis defense
Title : One-Shot Coordination of First and Last-Mode Service in Multi-Modal Transportation
Speaker : Subhajit Goswami
Degree Registered : MTech (Res)
Advisor : Dr. Pavankumar Tallapragada
Date : 12/06/2020
Venue : C 241 MMCR, EE
Abstract : In this thesis, we propose a coordination of the first and last mode connectivity in a multi- modal transportation system. In particular, we consider a one-shot problem wherein the passengers must be transported to or from a common location before a fixed deadline. We consider a macroscopic model, wherein we model a region of interest by a graph and consider flows of vehicles and volumes of passenger demand and vehicle supply. We then consider the problem of operator's profit maximization through optimal pricing and allocations of feeder vehicles given the knowledge of demand and supply distributions. Due to the specific nature of the problem, we can first determine the optimal prices, subsequent to which the overall problem becomes a linear program. We first study the problem of "one-shot feed-in". Given the large scale or scope of the problem and the need for a near real-time implementation, we first seek to reduce the problem size. We propose an off-line route elimination algorithm that given a route- set returns another "reduced" route-set by eliminating the routes that would never be used in an optimal solution irrespective of the demand and supply distributions. In simulations on a 24 node graph, our proposed route reduction algorithm reduced the number of optimization variables to nearly one third of the original number. Such a reduced route-set could then be utilized for faster computation of the optimal solutions of the feed-in problem when the demand and supply distributions are revealed. We then analyse the supply optimization problem for "one-shot feed-in" for a given demand distribution. In this problem, given a total supply, we are interested in optimally distributing the supply so as to maximize the operator's profits. With this formulation, we give the closed form expression of the "absolute maximum profits" of the operator over all supply distributions given the demand distribution. Next, we go on to show that the "one-shot feed-out" problem is equivalent to the supply optimization problem for "one-shot feed-in" and that similar results can be drawn using the equivalence analysis. Finally, we propose a simple framework for determining the prices. With this framework, we are able to analyse the cost of a first or last mode feeder service relative to the best alternative transportation at which the feeder service becomes profitable. We demonstrate our analytical results with a suite of simulations, including a comparative study of our model with respect to a macroscopic single-depot single-window routing problem.

Event : Thesis Colloquium
Title : Price of Privacy of Smart Meter Data
Speaker : Soumyajit Gangopadhyay
Degree Registered : MTech (Research)
Advisor : Dr. Sarasij Das
Date : 03/06/2020
Venue : Online
Abstract : Smart meter data form the backbone of various data driven distribution system applications ranging from demand response, outage management, power theft detection, capacitor bank switching to distribution network planning and asset management. However, privacy concerns often inhibit consumers from sharing their smart meter data. Privacy preserving algorithms have been proposed in the literature to address the privacy concern of consumers. However, privacy preservation often tends to reduce the usability of data by limiting the information that can be extracted from the data. Consumers incur privacy loss if they give up their privacy right over smart meter energy data to the utility. Compensating the consumers for their privacy loss can be an alternate approach to encourage smart meter data sharing. The advantage of this approach is that it does not reduce the data usability. There is a lack of literature on the price of privacy of smart meter data.
In this work, a novel privacy pricing framework is proposed which can be used by the power utility to determine the price of privacy of consumers' energy data. The proposed framework is intended to compensate the consumers who agree to give up their privacy right over their smart meter energy data. At first, the notion of smart meter data privacy is presented. Estimation of the price of privacy of energy consumption data of smart meters is a challenging task due to the lack of a mathematical privacy model. So, a privacy model is proposed to quantify the private information content of consumers' energy consumption data based on the privacy notion. Then, a nonlinear convex optimization problem is formulated using the proposed privacy model to determine the price of privacy of the consumers for a given budget of the utility. A constraint is incorporated in the optimization problem to ensure that the privacy price does not exceed the electricity bill of a consumer. The impact of the utility's budget (allocated for privacy pricing) on a consumer's price of privacy is shown using practical smart meter data. The results show that a consumer's price of privacy also depends on the total number of consumers (participating in privacy pricing) and the privacy content of their data. The results also indicate that the proposed pricing framework can be effective in incentivizing consumers to share their smart meter data.

Event : Thesis Defence
Title : Novelty Detection in Computer Vision.
Speaker : Supritam Bhattacharjee
Degree Registered :MTech (Research)
Advisor : Prof. Soma Biswas
Date : 12/05/2020
Venue : Online
Abstract : In image classification, conventional supervised classifiers are trained using training samples and their corresponding class labels. During testing, the goal is to classify a given query image into one of the classes encountered during training. This is based on the assumption that the query will belong to one of the classes seen during training, which is a "closed set" setting. In an ever changing and dynamic world, the sets of training and testing classes might not be same. Thus, during testing, the query data may come from a class that was not present during the training stage. Novelty detection tries to answer the pertinent question: given a test sample, should we even try to classify it? Multi-class novelty detection is increasingly becoming an important area of research due to the continuous increase in the number of object categories. This is an extremely challenging task as it is an ill-posed problem, where there is no information about the distribution of the novel classes, and there is no limit to the number of novel classes that can be encountered during testing. In this thesis, we have proposed two approaches for novelty detection in the context of image classification. Our third contribution is another novelty detection approach for the application of generalized zero-shot learning.
In the first work, we propose a model termed as Segregation Network, which addresses the problem of novelty detection using the idea of mix-up technique. Here, during training, a pair of data samples is utilized along with a generated data-point, which is formed by their convex combination. The proposed model is trained with a novel loss to reveal the constituent classes of the interpolated points, which is then used to determine whether the query sample is novel or not.
In the second work, we propose a novel two-stage approach for novelty detection. In the first stage, for a given query, analyzing the top retrieved samples of the base classification network gives us a set of probable categories for that query. For the second stage, a comparator network is designed which can compare features from two samples and classify them into similar or dissimilar pairs. Finally, the scores from the above two stages are fused to form a final score, which can be used to determine if the query is novel.
Finally, we also look into an interesting application of novelty detection for Generalized Zero Shot Learning (GZSL), where the goal is to classify the input data, which can belong to seen or unseen classes, by utilizing some high-level categorical information or attributes. Most existing GZSL algorithms are biased towards the seen classes, as the algorithms are trained using these classes. In order to remove this bias, we introduce a conditional autoencoder based novelty detector, where the attributes of the seen classes are used as the condition. The proposed algorithm does not require any prior information about the unknown classes, and thus can be seamlessly used with any existing GZSL approach to improve its performance, as shown in extensive experiments across many benchmark datasets.

Event : Thesis Defence
Title : Fault Location in Double Wye Shunt Capacitor Banks
Speaker : Polisetty Sai Pavan
Degree Registered :Mtech (Research)
Advisor : Dr. Sarasij Das
Date : 24/04/2020
Venue : C 241 MMCR, EE
Abstract : Shunt capacitor banks (SCB) are commonly used to provide reactive power support in both transmission and distribution systems. Outage of any SCB is crucial and hence, should be well protected against various types of internal faults. Double wye connected SCBs are widely used at high voltage levels. Neutral current compensation method is commonly used to detect internal faults in double wye SCBs. Existing protection algorithms mostly fail to detect simultaneous faults happening within a protection pass. Existing algorithms either fail to detect a faulted condition or detect one type of fault as some other type of fault. In this work, a novel method for detecting different types of internal faults in grounded and ungrounded double wye SCBs has been developed. One of the main advantages of the proposed method is that it can detect simultaneous faults happening within a protection pass. The proposed method can detect different types of internal faults under supply voltage unbalance and impedance unbalance (during healthy condition) situations. A laboratory scale ungrounded double wye SCB test setup is developed to test the working of proposed method.

Event : Thesis Defence
Title : Pronunciation assessment and semi-supervised feedback prediction for spoken English tutoring
Speaker : Chiranjeevi Yarra
Degree Registered :PhD
Advisor : Prof. Prasanta Kumar Ghosh
Date : 20/04/2020
Venue : C 241 MMCR, EE
Abstract : Spoken English pronunciation quality is often influenced by the nativity of a learner, for whom English is the second language. Typically, the pronunciation quality of a learner depends on the degree of following four sub-qualities: 1) phoneme quality 2) syllable stress quality 3) intonation quality, and 4) fluency. In order to achieve a good pronunciation quality, learners need to minimize their nativity influences in each of the four sub-qualities, which can be achieved with effective spoken English tutoring methods. However, these methods are expensive as they require highly proficient English experts. In cases, where a cost-effective solution is required, it is useful to have a tutoring system which assesses a learner's pronunciation and provides feedback in each of the four sub-qualities to minimize nativity influences in a manner similar to that of a human expert. Such kind of systems are also useful for learners who can not access high quality tutoring due to their demographic and physical constraints. In this thesis, several methods are developed to assess pronunciation quality and provide feedback for such a spoken English tutoring system for Indian learners.
Most of the existing works on automatic pronunciation assessment predict an overall pronunciation quality. However, feedback prediction has typically been done separately in each of the four sub-qualities. Both pronunciation assessment and feedback prediction require annotations on a large set of recordings from learners. While the former requires ratings for overall pronunciation quality, the later needs feedback specific labeling. Unlike ratings, obtaining labels for feedback prediction requires highly skilled annotators. Such annotators are not available in large numbers and labeling with their expertise is also costly. Due to this paucity of labels, it is challenging to design a tutoring system in a cost effective manner particularly for Indian nativity, which is known for its large accent variabilities. With regard to these challenges, the key contributions in this thesis are: 1) building models for estimating parameters for providing meaningful feedback without using any labelled data, 2) building models for estimating overall pronunciation quality using annotated data, and 3) developing voisTUTOR, a system for learners to train themselves with neutral accent of English with the help of a spoken English expert. The feedback prediction is semi-supervised in nature as no feedback-specific labels are used for building feedback prediction models.
Feedback in each of the four sub-qualities is predicted by analyzing mismatches in the respective parameters between learners' and an expert's speech. In the phoneme category, phoneme errors made by a learner are provided as feedback, where the phonemes are estimated using rule based pronunciation dictionary. These rules are deduced from the errors made by the Indian learners while speaking English. For demonstrating the correct pronunciation, an articulatory video is synthesized using an expert's speech. Further, the effect of accents on the uttered phonemes is assessed using goodness of pronunciation measure, which is computed in a deep neural network-hidden Markov model (DNN-HMM) based automatic speech recognition (ASR) framework. In the stress category, mismatches in the estimated stressed syllable locations are provided as feedback. For this, stress-specific features are computed by exploring linguistic parameters, such as sonority, from every syllable when the ground truth syllable information is available. Its performance is analysed when the syllable information is estimated as in a real scenario. The stress locations are also estimated in an ASR framework without computing any stress-specific features. In the intonation category, feedback is provided based on the local and global mismatches in pitch patterns. For this, models are proposed to estimate the pitch values and their associated confidence scores. It is observed that the global mismatches depend on temporal variations in the pitch and its patterns. These mismatches are identified better when the confidence scores along with the pitch values are used in the models, based on HMMs and long-short term memory (LSTM) networks. Both the global and local mismatches are identified using knowledge driven template matching approach, that performs confidence score based median filtering and pitch stylization. In the fluency category, mismatches in the pause locations are provided as feedback. The pause locations are estimated using features based on speech acoustics only without considering any canonical stress markings because the learners' pronunciation do not often match the canonical pronunciation. Further, analysis is performed to estimate speaking rate directly from the speech acoustics, where speaking rate has been shown to be correlated with the fluency of a learner's pronunciation.
Overall pronunciation rating is estimated using a joint model considering DNNs and LSTM networks. For this, studies are conducted to find out differences between the speech rhythm of Indian languages and that of English. Features based on speech rhythm are used for estimating the rating along with the features based on the parameters used for the feedback in all four sub-qualities. Further, in order to create an interactive learning environment in voisTUTOR, these feedback and the ratings are displayed using audio-visual aids including line and bar graphs and text messages. All of these are made available in an android app using a web-server with LAMP (Linus, Apache, MySQL, PHP) stack on Ubuntu 14.04 LTS system.

Event : Thesis Colloquium
Title : Improved Generative Models for Zero-shot object recognition
Speaker : Pambala Ayyappa Kumar
Degree Registered :Mtech(Research)
Advisor : Dr. Soma Biswas
Date : 04/03/2020
Venue : C 241, MMCR

Event : Seminar
Title : Analysis and Control Innovations for Future Grids
Speaker : Prof. Sairaj V. Dhople
Date : 25/02/2020
Venue : C 241 MMCR, EE
Abstract : Driven largely by the rapid integration of renewable resources and energy storage technologies, future grids will be dominated by power-electronics interfaces. Long-standing practices in analysis and control that were conceptualized and tuned for fossil-fuel-driven synchronous generation would therefore need to be revisited and revised. In this context, this talk focuses on advances to two problems that are fundamental to and underlie the analysis and control of power grids: power flow and economic dispatch. For the power-flow problem, we examine the distributed slack bus formulation whereby the active-power output of each generator is modeled with three elements: a nominal injection modulated by a fraction of the net-load imbalance allocated via a participation factor.
This setup acknowledges system dynamics and controllers more accurately than the conventional single slack bus, but it has long been plagued by ambiguous and inconsistent interpretations of its constituent elemental quantities. In the talk, we demonstrate that with the: i) nominal active-power injections set to be the economic dispatch setpoints, ii) participation factors fixed to be the ones used in automatic generation control, and iii) net-load imbalance considered to be the total load and loss unaccounted in economic dispatch, the power flow solution best matches results from a simulation of the system differential algebraic equation model. Shifting gears, the primary goal of real-time power system operation is to economically dispatch generation to meet net system load while regulating frequency. Presently, this is accomplished by modulating generator setpoints via a systematically engineered combination of: i) online proportional-integral control based on frequency and tie-line-flow errors (automatic generation control); and ii) offline optimization that minimizes cost of generation based on the load forecast (economic dispatch). We outline a continuous-time economic dispatch problem and a companion solution algorithm to reconcile the temporal gap that separates look-ahead dispatch and real-time control, improve market efficiency, and enhance power quality. To ensure and demonstrate compatibility, contributions in these two problems are presented first in the context of conventional grids with synchronous generation and then translated to future low-inertia grids with inverter-based resources.
Speaker Bio :Sairaj V. Dhople received the B.S., M.S., and Ph.D. degrees in electrical engineering, in 2007, 2009, and 2012, respectively, from the University of Illinois, Urbana-Champaign. He received the National Science Foundation CAREER Award in 2015 and the Outstanding Young Engineer Award from the IEEE Power and Energy Society in 2019. He currently serves as an Associate Editor for the IEEE Transactions on Energy Conversion and the IEEE Transactions on Power Systems.

Event : Seminar
Title : Towards Dynamic Capsule Networks: How Neuroscience Can Help AI
Speaker : Dr. K.P. Unnikrishnan
Date : 19/02/2020
Venue : C 241 MMCR, EE
Abstract : The recently introduced Capsule Network (CapsNet) formalism has many unique features. We describe these and point out some shortfalls. Recent work from our group, addressing some of these, has improved the performance of CapsNets. We conclude by describing a new formalism that incorporates some of the unique features of mammalian sensory systems. These networks could provide real-time sensory processing needed for autonomous vehicles.
Speaker Bio :KP Unnikrishnan is the founder and scientific director of eNeuroLearn, an Ann Arbor based AI/DL startup. eNeuroLearn brings architectures and algorithms from Neuroscience to enhance Deep Learning. He has worked in Neural Networks, Computational Neuroscience, Data Mining, and Deep learning for the past 35 years. He has a PhD in Physics from Syracuse University and has worked at Bell Labs, Caltech, University of Michigan, General Motors Reasearch, NorthSore University Health System and Ford Motor Company. He started at IISc as a PhD student in Physics and has been visiting the institution since about 1985.

Event : Seminar
Title : Smart Sampling: Super-Resolution and Learning Strategies
Speaker : Dr. Satish Mulleti
Date : 17/02/2020
Venue : C 241 MMCR, EE
Abstract : Analog to digital conversion enables efficient processing of natural signals on digital signal processors. The Shannon-Nyquist sampling theorem is a widely used method for discrete representation of analog bandlimited signals. In many practical applications, either the signals are nonbandlimited or have large bandwidth, and accurate samplers are not feasible in such cases.
In applications such as RADAR, ultrasound imaging, optical coherence tomography (OCT), wideband communication, etc., the signals to be sampled have a certain structure, which could be of in a form such as: (i) sparsity in certain basis function; (ii) signals belonging to a shift-invariant space; (iii) multiband structure of the signal spectrum; (iv) a finite rate of innovation (FRI), etc. By using the structure to advantage, one can sample and reconstruct such signals at much lower rates than the Nyquist rate.
In this talk, we shall consider in detail the FRI signal models. Specifically, we consider the signals that are modeled as a linear combination of amplitude-scaled and time-shifted copies of a known pulse. We shall present how these signals could be exactly reconstructed by their filtered measurements taken at sub-Nyquist rate. On the application front, we will show the applicability of the FRI signals to OCT, curve fitting, and nuclear magnetic resonance (NMR) imaging. By establishing the FRI property of the measurements, we show that one can attain super-resolved tomograms and NMR spectra by using fewer measurements than the standard recovery techniques.
We extend sub-Nyquist sampling framework to a stream-of-pulses model where the pulse shape in unknown due to distortion during transmission and propagation. In this case, we consider a multichannel setup and show that the reconstruction problem can be posed as a sparse multichannel blind-deconvolution (MBD) problem. We will discuss the identifiability conditions to uniquely recover the signals from compressible measurements. We will show that the proposed compressive MBD results require fewer measurements and fewer channels for identifiability compared to previous results, which aids in building cost-effective receivers.
In the above-mentioned applications, the sampling strategy, weather deterministic or random, is fixed and does not adapt according to the parameters of the input signal. We shall discuss our results where we learn the sampling pattern that cognitively changes according to the input signals. We apply the learning-based sampling method to the problem of sparse array selection problems and show improvement in the resolution compared with a random sampling mechanism
Speaker Bio :Satish Mulleti received a Bachelor of Engineering degree from the Electronics and Communication Engineering Department, Jalpaiguri Government Engineering College, India. He obtained a Master of Engineering degree in Electrical Engineering from the Department of Electrical Engineering, Indian Institute of Technology Kanpur, India. Subsequently, he worked as a researcher in the Indian Space Research Organization (ISRO), India and Tata Consultancy Services (TCS) Innovation Labs, Mumbai, India. In August 2011, he joined the Spectrum Lab, Department of Electrical Engineering, Indian Institute of Science, Bangalore for his PhD. He received Prof. DJ Badkas medal from IISc for his PhD work. From April 2017 to April 2019 he was a postdoc at the Department of Electrical Engineering, Technion - Israel Institute of Technology, from April 2017 to April 2019. He is presently a postdoctoral fellow in the SAMPL lab, Faculty of Mathematics and Computer Science, Weizmann Institute of Science, Israel. His research interests include sampling theory, in particular, finite-rate-of-innovation signal sampling, compressive sensing, machine learning, blind deconvolution, sparse array signal processing, and spectral estimation.

Event : Seminar
Title : Machine Learning in Health Monitoring of Rotating Machines, IT Systems, and Humans
Speaker : Dr. Rahul Kumar Sevakula
Date : 14/02/2020
Venue : C 241 MMCR, EE
Abstract : “If you can’t measure it, you can’t improve it” is a popular quote in the management circles. A testimony to this quote lies in the fact that now millions are using smartwatches to measure their sleep, heart rate, etc. These days, condition monitoring systems form a part of many of the expensive machines/systems, namely airplanes, cars, air compressors, power systems, satellites etc. In this talk, Dr. Sevakula would provide a brief introduction on the challenges faced while building such health monitoring systems and how machine learning can be appropriately used case by case. Over the last few years, he had the opportunity of working on health monitoring systems across a wide range of use-cases, and he made novel contributions in each of these fields. During his PhD he worked on the condition monitoring of Air compressors, during his stay at IBM he worked on the predictive maintenance of IT systems, and currently at Harvard Medical School (HMS) he is working on the detection and prediction of cardiac arrhythmias in the ICU setup.
Speaker Bio :Dr. Rahul K Sevakula received the BTech degree from the National Institute of Technology Warangal, India, in 2009, and the PhD degree in Electrical Engineering from the Indian Institute of Technology Kanpur, India, in 2017. He is currently a Postdoctoral Research Fellow within the Cardiovascular Research Center at the Massachusetts General Hospital and Harvard Medical School (HMS), and is being funded by the American Heart Association. Prior to joining HMS, he worked at IBM as a Data Scientist between 2016 - 2018. Over the last 8 years, he has developed novel classifiers and has successfully used signal processing, machine learning and deep learning algorithms for predictive maintenance of machines, predictive maintenance of IT systems, cardiac health monitoring, and cancer classification. His current research interests lie in machine learning and predictive maintenance/health monitoring systems.

Event : Thesis Defence
Title : Knowledge-driven training of deep models for better reconstruction and recognition.
Speaker : Ram Krishna Pandey
Degree Registered :PhD
Advisor : Prof. A G Ramakrishnan
Date : 13/02/2020
Venue : C 241 MMCR, EE
Abstract : The thesis shows that the performance of any deep architecture can be improved at three different levels: (i) input, (ii) architecture and (iii) at the objective or loss function levels. The following are the key contributions of the thesis:

Event : Seminar
Title : Handwritten Image Recognition in Low-resource Indic Scripts
Speaker : Prof.Partha Pratim Roy
Date : 30/01/2020
Venue : C 241 MMCR, EE
Abstract : Handwritten image analysis has long been an active research area because of its complexity and challenges due to a variety of handwritten styles. Dataset is a necessary and important resource to develop any recognition/word-spotting system for bench-marking. It has been observed that the availability of training data is not uniformly distributed for all scripts. Hence, analysis of low-resource handwritten scripts is difficult as sufficient training data is not available and it is often expensive for collecting data of such scripts. Much of the existing literature comprises preprocessing strategies which are seldom sufficient to cover all possible variations. This research presentation will highlight few solutions towards effective analysis of low resource scripts. One of the framework exploits a large resource of dataset for training and uses it for cross-language learning of low-resource scripts. Also, adversarial learning based framework is designed to augment training datasets in order to introduce unseen variations; boosting the capability to better learn informative features from handwritten images.
Speaker Bio : Dr. Partha Pratim Roy is presently working as Associate Professor in the Department of Computer Science and Engineering, Indian Institute of Technology (IIT), Roorkee. He received his Masters in 2006 and PhD in 2010 from Universitat Autonoma de Barcelona, Spain. He did postdoctoral stay in France and Canada from 2010 to 2013. Dr. Roy gathered industrial experience while working in TCS and Samsung. In Samsung, he was a part-leader of Computer Vision research team. He is the recipient of "Best Student Paper" awarded by International Conference on Document Analysis and Recognition (ICDAR), 2009, Spain. His research interest includes Pattern Recognition, Document Image Processing, Biometrics and Human Computer Interaction. He is presently serving as Associate Editor of Springer Nature Computer Science.

Event : Thesis Colloquium
Title : Speaker verification using whispered speech
Speaker : Abinay Reddy Naini
Degree Registered :MSc(Engg)
Advisor : Dr. Prasanta Kumar Ghosh
Date : 28/01/2020
Venue : C 241 MMCR, EE
Abstract : Like neutral speech, whispered speech is one of the natural modes of speech production, and it is often used by speakers in their day-to-day life. For some people, such as laryngectomees, whispered speech is the only mode of communication. Despite the absence of voicing in whispered speech and difference in characteristics compared to the neutral speech, previous works in the literature demonstrated that whispered speech contains adequate information about the content and the speaker. In recent times, virtual assistants have become more natural and widespread. This led to an increase in the scenarios, where the device has to detect the speech and verify the speaker even if the speaker whispers. Due to the noise-like characteristics, detecting whispered speech is a challenge. On the other hand, a typical speaker verification system, where neutral speech is used for enrolling the speakers but whispered speech for testing, often performs poorly due to the difference in acoustic characteristics between the whispered and the neutral speech. Hence, the aim of this thesis is two-fold: 1) develop a robust whisper activity detector specifically for speaker verification task, 2) improve whispered speech based speaker verification performance.
The contributions in this thesis lie in whisper activity detection as well as whispered speech based speaker verification. It is shown how an Attention-based average pooling in a speaker verification model can be used to detect the whispered speech regions in noisy audio more accurately than the best of the baseline schemes available. For improving speaker verification using whispered speech, we proposed features based on formant gaps, and we showed that these features are more invariant to the modes of the speech compared to the best of the existing features. We also proposed two feature mapping methods to convert the whispered features to neutral features for speaker verification. In the first method, we introduced a novel objective function, based on cosine similarity, for training a DNN, used for feature mapping. In the second method, we iteratively optimized the feature mapping model using cosine similarity based objective function and the total variability space likelihood in the i-vector based background model. The proposed optimization provided a more reliable mapping from whispered features to neutral features resulting in an improvement of speaker verification equal error rate by 44.8% (relative) over an existing DNN based feature mapping scheme.

Event : Thesis Colloquium
Title : Novelty Detection in Computer Vision
Speaker : Supritam Bhattacharjee
Degree Registered :M.Tech (Res)
Advisor : Prof. Soma Biswas
Date : 27/01/2020
Venue : C 241 MMCR, EE
Abstract : In image classification, conventional supervised classifiers are trained using training samples and their corresponding class labels. During testing, the goal is to classify a given query image into one of the classes encountered during training. This is based on the assumption that the query will belong to one of the classes seen during training, which is a "closed set" setting. In an ever changing and dynamic world, the sets of training and testing classes might not be same. Thus, during testing, the query data may come from a class that was not present during the training stage. Novelty detection tries to answer the pertinent question: given a test sample, should we even try to classify it? Multi-class novelty detection is increasingly becoming an important area of research due to the continuous increase in the number of object categories. This is an extremely challenging task as it is an ill-posed problem, where there is no information about the distribution of the novel classes, and there is no limit to the number of novel classes that can be encountered during testing. In this thesis, we have proposed two approaches for novelty detection in the context of image classification. Our third contribution is another novelty detection approach for the application of generalized zero-shot learning.
In the first work, we propose a model termed as Segregation Network, which addresses the problem of novelty detection using the idea of mix-up technique. Here, during training, a pair of data samples is utilized along with a generated data-point, which is formed by their convex combination. The proposed model is trained with a novel loss to reveal the constituent classes of the interpolated points, which is then used to determine whether the query sample is novel or not.
In the second work, we propose a novel two-stage approach for novelty detection. In the first stage, for a given query, analyzing the top retrieved samples of the base classification network gives us a set of probable categories for that query. For the second stage, a comparator network is designed which can compare features from two samples and classify them into similar or dissimilar pairs. Finally, the scores from the above two stages are fused to form a final score, which can be used to determine if the query is novel.
Finally, we also look into an interesting application of novelty detection for Generalized Zero Shot Learning (GZSL), where the goal is to classify the input data, which can belong to seen or unseen classes, by utilizing some high-level categorical information or attributes. Most existing GZSL algorithms are biased towards the seen classes, as the algorithms are trained using these classes. In order to remove this bias, we introduce a conditional autoencoder based novelty detector, where the attributes of the seen classes are used as the condition. The proposed algorithm does not require any prior information about the unknown classes, and thus can be seamlessly used with any existing GZSL approach to improve its performance, as shown in extensive experiments across many benchmark datasets.

Event : Thesis Colloquium
Title : Pitch-synchronous Discrete Cosine Transform Features For Speaker Identification and Verification
Speaker : Amit Meghanani
Degree Registered : M. Tech (Research)
Advisor : Prof. A G Ramakrishnan
Date : 27/01/2020
Venue : C 241 MMCR, EE
Abstract : Extracting speaker-specific information from speech is of great interest as speaker recognition technology finds application in a wide range of areas such as forensics and biometric security systems. In this work, we propose a new feature named Pitch-synchronous Discrete Cosine Transform (PS-DCT) for speaker identification (SID) and verification (SV). This feature exploits the time-domain, quasi-periodic structure of the voiced phones. Our experiments on standard databases demonstrate that the proposed features supplement the handcrafted classical features, such as MFCC. Improvement in performance is more prominent in the case of limited data speaker verification tasks.
We have also tested the efficacy of the proposed features for other speech-based tasks. For the task of emotion recognition, pitch-synchronous based training of LSTM network with PS-DCT and MFCC gives improved performance over regular MFCC based trained network. The proposed features also work for the vowel recognition task.

Event : Thesis Colloquium
Title : AlGaN/GaN Heterojunctions based Hall Sensors for Magnetic Field Sensing over Wide Temperature Range
Speaker : Sagnik Kumar
Degree Registered : MTech (res)
Advisor : Prof. G. Narayanan
Date : 22/01/2020
Venue : C 241 MMCR, EE
Abstract : Hall-effect sensors have been used extensively for magnetic field sensing in applications such as electric machines, power transmission and consumer electronics. Most commercially available Hall sensors are based on silicon and have a limited operating temperature range. For field sensing at cryogenic temperatures or at extremely high temperatures, wide band gap-based materials offer a viable alternative. This thesis demonstrates Hall-effect sensing using AlGaN-GaN based heterojunctions for extreme temperatures.
Hall-effect sensors have been fabricated using AlGaN/GaN heterojunctions grown on Si and SiC substrates. A process flow for batch fabrication of a number of such sensing elements on a large GaN-on-Si wafer has been proposed. The fabricated samples have been extensively characterized at temperatures ranging from 75 K to 500 K and at magnetic field strengths upto 2 Tesla.
The sensed Hall signal is reported to be quite weak with a large offset. An electronic subsystem has been designed to cancel out the offsets in the Hall signal using current spinning techniques. In addition, the electronic subsystem performs the necessary signal amplification and provides a visual read-out of the sensed magnetic field. The sensing element along with the signal conditioning circuit has been tested and calibrated using a Helmholtz coil setup.

Event : Seminar
Title : Addressing R&D Challenges in Emerging Power Systems: Innovative Research Facilities and Some Research Results
Speaker : Prof. Tarlochan Sidhu
Date : 17/01/2020
Venue : C 241 MMCR, EE
Abstract : Power systems have been evolving and continue to evolve. Starting from deregulation to emergence of renewable sources and their integration into existing power systems, adoption of electrified transportation and, a holistic approach to production and use of energy are some of the factors that have created interesting R&D challenges for power system researchers. This has created a need for development of innovative research infrastructure to solve issues for the emerging power systems. This talk will describe the research infrastructure and laboratories developed at Ontario Tech University. At the same time, innovative solutions including interdisciplinary approaches are needed to address challenges for the emerging power systems. Results from two such industrial research projects (one related to arcing fault detection in electrical equipment and the other related to use of wireless LAN in substations) will be discussed.
Speaker Bio :Prof Tarlochan Sidhu, Ontario Tech University, Canada

Event : Thesis Colloquium
Title : Speech task-specific representation learning using acoustic-articulatory data
Speaker : Mannem Renuka
Degree Registered : MSc (Engg)
Advisor : Prof. Prasanta Kumar Ghosh
Date : 28/01/2020
Venue : C 241 MMCR, EE
Abstract : Human speech production involves modulation of the air stream by the vocal tract shape determined by the articulatory configuration. Articulatory gestures are often used to represent the speech units. It has been shown that the articulatory representations contain information complementary to the acoustics. Thus, a speech task could benefit from the representations derived from both acoustic and articulatory data.
A typical acoustic representation consists of spectral and temporal characteristics e.g., Mel Frequency Cepstral Coefficients (MFCCs) Line Spectral Frequencies (LSF), and Discrete Wavelet Transform (DWT). On the other hand, articulatory representations vary depending on how the articulatory movements are captured. For example, when Electro-Magnetic Articulography (EMA) is used, the recorded raw movements of the EMA sensors placed on the tongue, jaw, upper lip, and lower lip and tract variables derived from them have often been used as articulatory representations. Similarly, when real-time Magnetic Resonance Imaging (rtMRI) is used, articulatory representations are derived primarily based on the Air-Tissue Boundaries (ATB) in the rtMRI video. The low resolution and SNR of the rtMRI video makes the ATB segmentation challenging. In this thesis, we propose various supervised ATB segmentation algorithms which include semantic segmentation, object contour detection using deep convolutional networks. The proposed approaches predict ATBs better than the existing baselines, namely, Maeda Grid and Fisher Discriminant Measure based schemes. We also propose a deep fully-connected neural network based ATB correction scheme as a post processing step to improve upon the predicted ATBs. However, articulatory data is not directly available in practice, unlike the speech recording. Thus, we also consider the articulatory representations derived from acoustics using an Acoustic-to-Articulatory Inversion (AAI) method. Generic acoustic and articulatory representations may not be optimal for a speech task. In this thesis, we consider the speech rate (SR) estimation task, useful for several speech applications and propose techniques for deriving acoustic and articulatory representations for the same. SR is defined as the number of syllables per second in a given speech recording. We propose a Convolutional Dense Neural Network (CDNN) to estimate the SR from directly given as well as learnt acoustic and articulatory representations. In the case of acoustics, the SR is estimated directly using MFCCs. When raw speech waveform is given as input, one-dimensional convolutional layers are utilized to estimate the SR specific acoustic representations. The center frequencies of the learned convolutional filters range from 200 to 1000 Hz unlike MFCC filter bank frequencies which range from 0 to 4000 Hz. The task-specific features are found to perform better in SR estimation compared to the MFCCs.
The articulatory features also help in accurate SR estimation since the characteristics of articulatory motion significantly vary with the changes in the SR. To estimate the SR specific articulatory representations, both the AAI and CDNN models are jointly trained using a weighted loss function which includes loss for the SR estimation and loss for estimating articulatory representations from acoustics. Similar to the acoustics case, the task-specific articulatory representations derived from acoustics perform better in SR estimation compared to the generic articulatory representations. Even though the task-specific articulatory representations derived from acoustics are not identical to the generic articulatory representations, both are found to be low pass in nature. The CDNN based approach using both generic and learnt representations perform better than the temporal and selected subband correlation (TCSSBC) based baseline scheme for SR estimation task.

Event : Seminar
Title : Using deep symmetry_sensitive network for detecting diseases in symmetric organs
Speaker : Dr. Arko Barman
Date : 10/01/2020
Venue : C 241 MMCR, EE
Abstract : Spatial symmetry is commonly used by clinicians in the diagnosis and prognosis of diseases involving multiple organs such as brain, prostate, breasts, and lungs. Anomalies in symmetry can be indicative of patient-specific disease-related features that are less sensitive to inter-patient variability. However, quantifying these symmetric anomalies is challenging as the symmetries in the human body are not exact mirrored copies.
This study involves the design and development of a novel deep learning architecture, named Deep Symmetry-sensitive Network (DeepSymNet), capable of learning anomalies in symmetry from minimally processed 2D or 3D images. Besides a synthetic dataset, DeepSymNet was evaluated for detection of Large Vessel Occlusion (LVO) in the brain for ischemic stroke patients (using 3D CTAngiography images) and detection of breast cancer (using 2D mammogram images). DeepSymNet is less sensitive to noise, rotation and translation when compared against a "symmetry-naive" deep convolutional neural network (CNN) with a similar number of parameters. Additionally, the DeepSymNet architecture is efficient during training and achieved similar or better performance compared to a symmetry-naive CNN with orders of magnitude fewer training examples.
Finally, an interpretation of the decision-making process of the DeepSymNet model using activation maps is presented. In our experiments, DeepSymNet automatically learned to focus on the regions of the brain/breast that clinicians observe for detecting LVO/breast cancer without any prior knowledge apart from symmetry and outcome variable.
Speaker Bio : Arko Barman is a postdoctoral research fellow at The University of Texas Health Science Center at Houston, where his current research focuses on developing imaging-based computer-aided diagnosis systems for stroke, Alzheimer's and other diseases. He received his B.E. degree in Electrical Engineering from Jadavpur University in 2009 and his M.E. degree in Signal Processing from Indian Institute of Science in 2011. He received his Ph.D. in Computer Science at the University of Houston in 2018. Dr. Barman has worked at Broadcom Corporation and PARC (A Xerox Company), and has also served as an Assistant Professor at a 4-year engineering college in India. His research interests include Medical Image Computing, Deep Learning, Computer Vision, Machine Learning, Data Mining, and Heuristic Optimization Algorithms. He has also been involved in curriculum design and has taught a diverse range of courses at different levels.

Event : Seminar
Title : Utilizing Real-time MRI to Investigate Speech Articulation Disorders
Speaker : Dr. Christina Hagedorn
Date : 10/01/2020
Venue : C 241 MMCR, EE
Abstract : Over the past two decades, real-time Magnetic Resonance Imaging (rtMRI), elaborating traditional medical MRI, has played a critical role in studying a variety of biological movement patterns. Through collaboration between engineers and speech scientists, rtMRI technology has been applied to the study of speech production. Through semi-automatic detection of air-tissue boundaries and estimation of articulatory kinematics using pixel intensity time functions, rtMRI can be used to quantitatively analyze speech production patterns in both typical and disordered populations. In this work, rtMRI is demonstrated to shed light on aspects of speech produced by individuals with tongue cancer and individuals with Apraxia of Speech that would not be possible using tools providing more limited spatiotemporal information about vocal tract shaping.
Speaker Bio :Christina Hagedorn is an assistant professor of Linguistics and director of the Motor Speech Laboratory at the City University of New York (CUNY) – College of Staten Island. Her research focuses primarily on disordered speech production. Her work aims to shed light on the precise nature of articulatory breakdowns in disordered speech and how this can inform theories of unimpaired speech production, as well as lead to refinement of the therapeutic techniques used to address these speech deficits. She received her Ph.D. in Linguistics from the University of Southern California, where she was a member of the Speech Production and Articulation kNowledge (SPAN) Group, the USC Phonetics and Phonology Group, and was a Hearing and Communication Neuroscience pre-doctoral fellow. She received her clinical training in Communicative Sciences and Disorders at New York University, and holds a certificate of clinical competency in Speech and Language Pathology (CCC-SLP).

Event : Thesis Defence
Title : Class-specific and noise-specific speech enhancement approaches
Speaker : Nazreen P M
Degree Registered :PhD
Advisor : Prof. A G Ramakrishnan
Date : 27/12/2019
Venue : C 241 MMCR, EE
Abstract : This thesis proposes, implements and analyzes various speech sound class-specific and noise-specific enhancement approaches and frame-wise selection methods for class-specific and noise-specific models.
The thesis work has three major contributions:
We have analyzed the performance of our enhancement scheme, where we use various speech-sound class-specific dictionaries to enhance noisy speech. By removing the contribution from the bases of those classes that correlate well with noise, one could improve the enhancement performance. We achieve this by learning different dictionaries for different classes and select a particular dictionary for a frame. We explore a class-specific enhancement approach, where we use a sparse coding, dictionary-based approach to learn dictionaries of various speech classes and noises. When we analyze the performance of our various class-specific approaches in terms of phoneme recognition, we obtain performances superior to class-independent case, even when we use estimated (approximate) labels for enhancement.
Any error in the estimated class label in the class-specific approach results in the selection of an erroneous dictionary for enhancement. The joint enhancement-decoding (JED) algorithm that we propose tries to overcome this issue by jointly optimizing for labels of all the frames and the decoding path to improve the phoneme recognition accuracy. The current noisy speech frame is enhanced by multiple (N) phoneme-specific dictionaries close to the approximate label of that frame. These N enhanced frames are then fed into the JED algorithm. The algorithm accepts these N observations and chooses the best for each frame such that the overall likelihood is maximized to obtain the final recognized labels. The Viterbi decoding algorithm used in speech recognition is integrated with the class label selection to develop the JED algorithm.
We also propose a method of picking the best DNN model in the scenario, where multiple noise-specific DNN models are available for enhancement, using the Monte Carlo (MC) dropout proposed by Gal and Ghahramani. The variance measure of the output signal vectors, resulting from different MC dropout trials, is used as a measure of the model precision to select one out of the multiple models for each frame. We find this method to be particularly useful for unseen noisy scenario, where the noise corrupting the test speech is different from those with which the available DNN models are trained. For the unseen noisy scenario, this method performs better than selecting the model using a DNN classifier. We observe some promising results with the enhancement performance of the algorithm on speech corrupted with a mixture of multiple noises and for the case, where different segments of speech are corrupted by different noises. This approach also has worked on speech corrupted with real noise recorded from a road (C V Raman Road).
Speaker Bio :

Event : Seminar
Title : From compressed sensing to deep learning: tasks, structures, and models
Speaker : Prof. Yonina Eldar
Date : 18/12/2019
Venue : C 241 MMCR, EE
Abstract : The famous Shannon-Nyquist theorem has become a landmark in the development of digital signal and image processing. However, in many modern applications, the signal bandwidths have increased tremendously, while the acquisition capabilities have not scaled sufficiently fast. Consequently, conversion to digital has become a serious bottleneck. Furthermore, the resulting digital data requires storage, communication and processing at very high rates which is computationally expensive and requires large amounts of power. In the context of medical imaging sampling at high rates often translates to high radiation dosages, increased scanning times, bulky medical devices, and limited resolution.
In this talk, we present a framework for sampling and processing a large class of wideband analog signals at rates far below Nyquist in space, time and frequency, which allows to dramatically reduce the number of antennas, sampling rates and band occupancy.
Our framework relies on exploiting signal structure and the processing task. We consider applications of these concepts to a variety of problems in communications, radar and ultrasound imaging and show several demos of real-time sub-Nyquist prototypes including a wireless ultrasound probe, sub-Nyquist MIMO radar, super-resolution in microscopy and ultrasound, cognitive radio, and joint radar and communication systems. We then discuss how the ideas of exploiting the task, structure and model can be used to develop interpretable model-based deep learning methods that can adapt to existing structure and are trained from small amounts of data. These networks achieve a more favorable trade-off between increase in parameters and data and improvement in performance, while remaining interpretable.
Speaker Bio : Yonina C. Eldar received the B.Sc. degree in Physics in 1995 and the B.Sc. degree in Electrical Engineering in 1996 both from Tel-Aviv University (TAU), Tel-Aviv, Israel, and the Ph.D. degree in Electrical Engineering and Computer Science in 2002 from the Massachusetts Institute of Technology (MIT), Cambridge. From January 2002 to July 2002 she was a Postdoctoral Fellow at the Digital Signal Processing Group at MIT.
She is currently a Professor in the Department of Mathematics and Computer Science, Weizmann Institute of Science, Rehovot, Israel. She was previously a Professor in the Department of Electrical Engineering at the Technion, where she held the Edwards Chair in Engineering. She is also a Visiting Professor at MIT, a Visiting Scientist at the Broad Institute, and an Adjunct Professor at Duke University and was a Visiting Professor at Stanford. She is a member of the Israel Academy of Sciences and Humanities (elected 2017), an IEEE Fellow and a EURASIP Fellow.
Dr. Eldar has received numerous awards for excellence in research and teaching, including the IEEE Signal Processing Society Technical Achievement Award (2013), the IEEE/AESS Fred Nathanson Memorial Radar Award (2014), and the IEEE Kiyo Tomiyasu Award (2016). She was a Horev Fellow of the Leaders in Science and Technology program at the Technion and an Alon Fellow. She received the Michael Bruno Memorial Award from the Rothschild Foundation, the Weizmann Prize for Exact Sciences, the Wolf Foundation Krill Prize for Excellence in Scientific Research, the Henry Taub Prize for Excellence in Research (twice), the Hershel Rich Innovation Award (three times), the Award for Women with Distinguished Contributions, the Andre and Bella Meyer Lectureship, the Career Development Chair at the Technion, the Muriel & David Jacknow Award for Excellence in Teaching, and the Technion’s Award for Excellence in Teaching (twice). She received several best paper awards and best demo awards together with her research students and colleagues including the SIAM outstanding Paper Prize and the IET Circuits, Devices and Systems Premium Award, and was selected as one of the 50 most influential women in Israel.
She was a member of the Young Israel Academy of Science and Humanities and the Israel Committee for Higher Education. She is the Editor in Chief of Foundations and Trends in Signal Processing, a member of the IEEE Sensor Array and Multichannel Technical Committee and serves on several other IEEE committees. In the past, she was a Signal Processing Society Distinguished Lecturer, member of the IEEE Signal Processing Theory and Methods and Bio Imaging Signal Processing technical committees, and served as an associate editor for the IEEE Transactions On Signal Processing, the EURASIP Journal of Signal Processing, the SIAM Journal on Matrix Analysis and Applications, and the SIAM Journal on Imaging Sciences. She was Co-Chair and Technical Co-Chair of several international conferences and workshops.
She is author of the book “Sampling Theory: Beyond Bandlimited Systems” and co-author of the books “Compressed Sensing” and “Convex Optimization Methods in Signal Processing and Communications,” all published by Cambridge University Press.

Event : Thesis Defence
Title : Modeling, Analysis, and Control of Reconfigurable Battery/Grid-Tied Solar Photo-Voltaic Inverter
Speaker : Venkatramanan D
Degree Registered : PhD
Advisor : Prof. Vinod John
Date : 12/12/2019
Venue : B 303, EE
Abstract : Grid reliability and power outages are key concerns today, due to the ever-increasing energy demand. Traditionally, Uninterruptible Power Supplies (UPS) with battery storage have been employed to contend with grid outages. For renewable power production, Solar-Photovoltaics (SPV) based Distributed Energy Resources (DERs) have been integrated with the grid using a power electronic Grid-Tied Inverter (GTI). A conventional GTI by design engages in power conversion only when the grid is present, and ceases operation during a power outage to avoid a local unintentional island formation. Thus, solar energy is left unutilized by the GTI during a power outage, while the UPS steps in, to power critical loads. Recently, hybrid-PV or dual-mode inverter systems, that combine the complementary functions of UPS and GTI, have been in research focus due to their ability for standalone system operation during an outage while accessing solar power. Such a hybrid-PV system adopts an integrated approach for the design, sizing, and control of the overall system. Although this approach meets the desired operational objectives, it does not offer a retrofitting ability that allows up-gradation of an existing UPS system with PV capability. Hence, the hybrid-PV approach requires replacement of the existing UPS backup system along with its battery banks.
This work enhances the existing methods of solar energy access during a power outage, where the GTI system is kept independent of the UPS. A reconfigurable battery/grid-tied inverter (RGTI) architecture is proposed, that ties to the grid when it is available and functions as a DC-AC inverter to inject maximum solar power. However, during a power outage, it reconnects to the battery-bank of an existing UPS present in a facility, where it functions as a DC-DC converter to provide PV based energy support. Such a scheme facilitates independent design and control of the RGTI without requiring high-bandwidth communication, and can function with any generic UPS. This allows for independent retrofitting and scaling of the PV installation and also leads to extension of cycle-life of the UPS batteries. However, such an operation of the RGTI with UPS requires several questions to be resolved in terms of system architecture, islanding behavior, maximum-power point tracking (MPPT), battery management, and overall system control, which are addressed in this work.
For the islanding behavior in grid-tied mode, a novel dynamic-phasor based GTI system model is proposed that accurately captures the system dynamics under unintentional islanding conditions. The equilibrium solutions of the model are shown to yield the steady-state operating frequency of the islanded system for arbitrary GTI power-factor and load quality-factors. The stability of the multiple possible frequency solutions of the islanded system is analytically established using state-space formulation and eigenvalue analysis.
Under battery-tied conditions, a small-signal model of the PV fed RGTI system is proposed by considering the dynamic model of the solar PV source. The conditions for model-order reduction is analytically established that facilitates simplified control analysis. Based on the transfer-functions derived, the minimum perturbation period, which is a critical parameter for any perturbative MPPT algorithm for convergence and stable operation, is analytically obtained as a function of control and system parameters.
A battery emulation control scheme is proposed for the RGTI that facilitates the seamless functioning of the RGTI in parallel with the physical UPS battery by emulating a high ampacity battery bank using PV. This sub-mode is shown to minimize the discharge current on the battery during a grid outage whenever solar irradiation is adequate. A system-level supervisory control scheme for the overall operation and power management is proposed to handle the dynamic changes in solar irradiation and UPS load variations during the day, such that the battery discharge burden is always minimized.
A discrete IGBT converter hardware platform is developed along with rooftop solar PV installation, where the proposed analytical models, circuit-level and supervisory controls, and the RGTI operation are experimentally verified on a 4.5 kW experimental setup.

Event : Thesis Colloquium
Title : Unidirectional High-Frequency-Link DC-Three Phase AC Conversion: Topology, Modulation and Converter Design
Speaker : Anirban Pal
Degree Registered :PhD
Advisor : Kaushik Basu
Date : 10/12/2019
Venue : C 241 MMCR, EE
Abstract : Conventionally, for grid integration of a large scale photovoltaic (PV) system, a three-phase voltage source inverter followed by a line frequency transformer (LFT) is used. The LFT provides galvanic isolation and thus reduces the circulation of leakage current and ensures safety. Few limitations with the conventional system are (a) huge volume as the LFT is bulky, (b) quite expensive due to large amount of iron and copper used in LFT and (c) the inverter is hard switched impacting the converter efficiency. The converter topologies with high-frequency galvanic isolation have attractive features like high power density and are less expensive. Hence these converters are promising alternatives to the conventional solution.
In this work, we have investigated five new unidirectional three-phase inverter topologies with high-frequency galvanic isolation. The main objective is to achieve converter power loss independent of switching frequency. Hence converter can be switched at a higher frequency which results in smaller magnetics and improves power density.
To achieve the converter power loss independent of switching frequency, the active switches of the introduced topologies are either line frequency switched or high-frequency soft-switched. With soft switching, the converter has negligible switching loss. The soft switching is achieved without additional snubber circuit. The pulse width modulation is implemented on the input DC side converters which are soft switched. The active switches of the grid interfaced converter are low frequency switched and thus enabling the use of high voltage blocking inherently slow semiconductor devices for direct medium voltage grid integration. The topologies are gradually improved to achieve soft switching of the DC side converters throughout the line cycle. The soft switching conditions are derived through detailed circuit analysis. The operations of these topologies are experimentally verified on hardware prototypes with power range 2-6kW. Out of five introduced topologies, three topologies can support only unity power factor operation. An additional shunt compensator is needed for any reactive power support. The remaining two topologies can support up to 0.866 power factor operation. The performances of the introduced topologies are compared with conventional and existing high frequency isolated solutions. Though the new topologies have relatively higher switch count, the converter power losses, filter requirements are comparable or smaller than the conventional solutions and have higher power density.

Event : Thesis Colloquium
Title : Power Electronic Converters for Condition Monitoring and Voltage Equalization of Batteries
Speaker : Shimul Kumar Dam
Degree Registered :PhD
Advisor : Prof. VInod John
Date : 10/12/2019
Venue : C 241 MMCR, EE
Abstract : Power converters are used in battery-based storage systems in many applications. Apart from the task of regulating the charging and discharging, the power electronic converters can also help to monitor battery condition and to avoid over-charge or over-discharge of any battery. One approach to monitoring the battery condition is by measuring its impedance. The power converter for charging and discharging of the battery bank can be used for online measurement of battery impedance. The challenges involved in control, measurement, and the hardware requirements for impedance measurement are analyzed in this work, and suitable solutions are proposed. A Proportional Integral Resonant (PIR) controller-based control scheme and a DAC based measurement method are proposed for impedance measurement over the required frequency range. Two different approaches are proposed to achieve sufficient output voltage resolution for generating small amplitude voltage perturbation. One approach achieves high voltage resolution by replacing the single-leg buck converter with a multi-leg interleaved converter. The other approach uses a low-power rated auxiliary converter in series with the main converter to achieve high voltage resolution. Both of the methods are experimentally verified and compared with commercial equipment and the advantages of each approach are evaluated.
A voltage equalizer is a power electronic circuit that equalizes the cell voltages in a series-connected battery stack to avoid over-charge and over-discharge of any individual cell. A low-cost voltage equalizer using selection switches for a cell to cell equalization is proposed. This equalizer uses capacitive voltage level shifting to avoid bulky and lossy isolation transformer and to reduce cost. A new approach with a lower number of low-frequency selection switches further reduces the equalizer cost. A high-performance voltage equalizer is also proposed to achieve fast equalization by direct multi-cell to multi-cell charge transfer. This topology is shown to provide soft-switching with high efficiency. The equalizer is controlled in an open loop. The equalization currents do not reduce with progress in voltage equalization, making this topology faster than the existing open-loop multi-cell to multi-cell topologies. A modularization method is proposed for this topology to provide a direct path for charge transfer from any cell in one module to any cell in another module. The operation of both the equalizers and the modularization technique are experimentally verified which confirms the theoretical analysis.

Event : Seminar
Title : An Approach Towards an Extreme Fast Charging Station Architecture for Electric Vehicles
Speaker : Vishnu Mahadeva Iyer
Date : 12/12/2019
Venue : C 241 MMCR, EE
Abstract : This talk introduces an extreme fast charging (XFC) station architecture concept for simultaneous charging of multiple electric vehicles (EVs). The XFC station architecture comprises of several power electronic converters interfaced to form a DC microgrid. The first part of the talk proposes the use of partial rated power electronic converters for enabling XFC. Partial power processing enables independent charging control over each EV, while processing only a fraction of the total battery charging power. System level benefits of the proposed approach include lower capital investments, lower operational costs, lower footprint and improved power and energy efficiency. Experimental results from a down-scaled laboratory test-bench are provided to validate the control aspects, functionality and effectiveness of the proposed approach. The second part of the talk addresses the system level stability challenges that exist in such an XFC station architecture-based DC microgrid. An active voltage stabilizer is proposed to stabilize the voltage oscillations in the DC microgrid. The functionality of the active stabilizer, based on a bidirectional DC-DC power converter, is elucidated and a suitable control strategy is discussed. The proposed approach is validated through analytical models and hardware results from the laboratory test-bench.
Speaker Bio :Vishnu Mahadeva Iyer received the B.Tech degree in Electrical & Electronics Engineering from College of Engineering, Trivandrum in 2011, M.E degree in electrical engineering from the Indian Institute of Science, Bengaluru in 2013. He completed his Ph.D. from the NSF FREEDM Systems Center, NC State University, Raleigh in November, 2019. He worked with GE Research, Bengaluru from 2013 to 2015 as a power electronics engineer and he will be joining as a lead power electronics engineer at GE Research, Niskayuna by January 2020. His current research interests include power electronics for automotive applications, grid-connected power converters, resonant and soft-switched power converters and control and stability of power electronic systems.

Event : Seminar
Title : Mining Frequent Serial Episodes with Simultaneous events
Speaker : Santhosh Gandreti
Degree Registered :PhD
Advisor : Prof. P. S. Sastry
Date : 29/11/2019
Venue : C 241 MMCR, EE
Abstract : Frequent Episode Mining is one of the popular topics in Data mining. The main idea of this branch of data mining is to deal with large amounts of temporal heterogeneous data, Episodes are sequential pattern that capture the temporal structures present in the data. Most of the traditional episode mining algorithms deal with serial, parallel or general episodes, which deals with restrictions on order of occurrence of events. Apart from these categories of episodes, we also find patterns containing simultaneous occurrence of events. Such patterns are mostly observed in data obtained from multiple sensors, spatiotemporal data etc. There are a few papers that address this issue of simultaneous events, but they do not cover all frequency notions. Moreover, they concentrate mainly on head and total frequency definitions and use depth-first search, thereby consuming a lot of space. In today’s talk, we propose a BFS algorithm for mining frequent serial episodes with simultaneous events. We present a formal definition for episodes with simultaneous events and a hierarchical lattice structure for an a-priori based candidate generation. We also extended Automata based frequency counting algorithm for the scenario of simultaneous events under minimal and non-overlapping occurrence frequency notions and compare its performance with already existing DFS algorithms.
Speaker Bio :Santhosh is currently pursuing his PhD with Prof. P. S. Sastry, in LSM Lab, EE, IISc. His research interests are Multivariate episode mining and Machine learning. He received his B.Tech degree in Electrical Engineering from IIT, Bhubaneswar in 2016.

Event : Seminar (Cancelled)
Title : Next Generation Power Electronic Systems: A Paradigm Shift in Design Philosophy
Speaker : Ayan Mallik
Date : 09/12/2019
Venue : C 241 MMCR, EE
Abstract : Power electronics has emerged as an enabling technology in deployment of next generation of systems including transportation systems, motor drives, robotics, renewable energies, smart grids and data centers, among many others. The demands for higher efficiency, enhanced reliability higher power density, specific power and better thermal management pose stringent challenges for the power electronic converters to accommodate. This presentation will put forward some of the emerging research areas, which have been pursued with the emphasis on more-electric-aircrafts (MEA), next generation onboard chargers for electric vehicles and data center power supplies. A novel approach to replace transformer rectifier units (TRUs) by actively controlled and lightweight Regulated Transformer Rectifier Units (RTRUs) employing the advantages of emerging wide band gap (WBG) semiconductor technology will be presented. In addition, this presentation will focus on an innovative topology to combine an onboard charger and an auxiliary power module with the objective of enhancing the performance, while adding extra functionalities and enhancing the operation efficiency in light and medium loads. Also, a new direction in multi-objective optimization and reliability assessment for power electronic systems will be discussed in this presentation.
Speaker Bio : Ayan received his Integrated Bachelor’s and Master’s degree (dual degree) in Electrical Engineering from Indian Institute of Technology (IIT), Kharagpur, India in 2014. He received his M aster of Science (MS) and Doctor of Philosophy (PhD) degrees in Electrical Engineering from the University of Maryland, College Park in 2018. His major research interests include the design, modelling, control and optimization of power electronic converters, characterizations and applications of wide bandgap (WBG) semiconductors, highly efficient & high-power density solutions for power conversions in the applications of more-electric-aircrafts, electric vehicles, wireless charging and data centers. He is an author/co-author of over thirty journal and conference papers. He has worked on research, development and testing of regulated transformer rectifier units for more-electric- aircrafts, integrated bidirectional onboard charger design for electric vehicles, high density DC- DC conversion for data centers, among many other projects. Dr. Mallik is the recipient of various awards and recognitions including first place in Dean’s doctoral dissertation competition at the University of Maryland, ECE distinguished dissertation award, University of Maryland's (UMD) Invention of the Year award (2018), Jimmy H.C. Lin invention award (2018), UMD’s outstanding graduate student award (2016), among many others.

Event : Thesis Colloquium
Title : Sampling of Structured Signals -- Algorithms and Imaging Applications
Speaker : Sunil Rudresh
Degree Registered : PhD
Advisor : Prof. Chandra Sekahr Seelamantula
Date : 05/12/2019
Venue : C 241 MMCR, EE
Abstract : Sampling establishes a link between the analog world and its digital counterpart that operates on the acquired discrete measurements of the underlying continuous-time signals. The classical Shannon sampling framework provides the key tool to faithfully go back and forth between the two representations for bandlimited signals, starting from measurements taken at the Nyquist rate. During the past few decades, the sampling paradigm has been extended to accommodate a larger class of signals such as signals belonging to shift-invariant spaces, multiband signals, finite-rate-of-innovation (FRI) signals, signals that admit sparse representation, etc.
FRI sampling considers signals that are not necessarily bandlimited, but are fully specified by a finite number of parameters per unit interval of time. The goal of this thesis is to extend the sampling for a larger class of FRI signals exploiting the structure present in the signal to devise efficient sampling and reconstruction strategies. Specifically, we consider signals that are a sum-of-weighted and time-shifted pulses that have undergone some convolutive modifications.
In the frst part, we consider the sampling and perfect reconstruction of three signal classes: asymmetric FRI pulse trains, modulated FRI signals, and 2D FRI signals.

Event : Seminar
Title : Semi-supervised Learning for Amazon Alexa
Speaker : Dr. Sivaram Garimella and Kishore Nandury
Date : 21/11/2019
Venue : C 241 MMCR, EE
Abstract : State-of-the-art Acoustic Models (AM) are large, complex deep neural networks that typically comprise millions of model parameters. Deep neural networks can express highly complex input-output relationships and transformations, but the key to getting the best performance out of them is the availability of large amounts of matched acoustic data – matched to the desired dialect, language, environmental/channel condition, microphone characteristic, speaking style, and so on. Since it is both time consuming and expensive to transcribe large amounts of matched acoustic data for every desired condition, we leverage Teacher/Student based Semi-Supervised Learning technology for improving the AM. Our training leverages vast amount of un-transcribed data in addition to multi-dialect transcribed data yielding up to 7% relative word error rate reduction over the baseline model, which has not seen any unlabelled data.
Speaker Bio :Sri Garimella is a Senior Manager heading the Alexa Machine Learning/Speech Recognition group in Amazon, India. He has been associated with Amazon for more than 7 years. He obtained PhD from the Department of Electrical and Computer Engineering, Center for Language and Speech Processing at the Johns Hopkins University, Baltimore, USA in 2012. And Master of Engineering in Signal Processing from the Indian Institute of Science, Bangalore, India in 2006. Kishore Nandury is an Applied scientist in Alexa ASR team in Amazon Bangalore. Prior to Amazon, he has worked in Intel, Sling media & NVIDIA graphics. He has obtained Masters degree in Signal Processing from Indian Institute of Science in 2005.

Event : Seminar
Title : Low-Coherent Imaging in Scattering Media.
Speaker : Dr. Bettina Heise
Date : 13/11/2019
Venue : C 241 MMCR, EE

Event : Seminar
Title : Neuroplasticity and the Musical Experience & Human Bipedal Locomotion:Temporally Coordinated Mechanisms of Balance Control
Speaker : Prof. Shihab Shamma & Prof. John Jeka
Date : 07/11/2019
Venue : C 241 MMCR, EE
Abstract : I. Neuroplasticity and the Musical Experience Why are humans so enamored by musical sounds in all their forms? What makes music special in the human experience in ways that are absent in all other animals? In this talk I will address the experimental and computational algorithms we developed or used to study the encoding of music in the brain. I will also explain how with these algorithms we can tap into the emotional engagement with music and reveal our dynamic interactions with music on a moment by moment basis. All these studies shed new light on the emotive bases of music and the ways in which it can be enhanced and harnessed for enjoyment and health. II. Human Bipedal Locomotion:Temporally Coordinated Mechanisms of Balance Control Human locomotion is a complex motor behavior in which multiple constraints are satisfied that include moving the body through space, balancing to remain upright, moving a foot through space to a stepping location, avoiding obstacles in the foot’s path, and modulating the movement pattern to achieve desired movement direction and speed of the body with a desired leg cadence and step length. Current theoretical modeling suggests how the biomechanics of the body, muscle physiology, and spinal reflex loops may generate stable walking patterns but does not yet address important features of human locomotion. One of these features is the neural control of balance, which has been studied extensively in standing, using a variety of techniques with quiet unperturbed stance as well as sensory and mechanical perturbations. Despite the vast knowledge gained regarding balance control during standing, such findings do not necessarily translate to balance control during walking. The main reason is the gait cycle. While responses to disturbances during standing follow a short-medium-long latency response pattern over 50-200 ms involving a proximal-to-distal pattern (or vice versa) of muscular activation, responses to disturbances during walking can occur anytime over the much longer (600 ms) gait cycle of steady state walking. Critically, body configuration changes dramatically over the gait cycle (e.g., double vs. single stance), necessitating different mechanisms to maintain upright balance at different points of the cycle. One common principle to maintain upright balance during standing and walking is that the base of support must (on average) be kept under the body’s center of mass (CoM). The locomotion literature has focused extensively on one particular mechanism to achieve this: foot placement control. When the central nervous system (CNS) senses a movement of the CoM to the right, e.g., it changes the foot placement of the next step to the right. However, recent studies have suggested that to achieve flexible control of upright stance during walking, foot placement control is only one in a series of several temporally coordinated control actions. Consider the demands on upright stability while crossing a busy intersection with many other pedestrians. Continuous small changes in direction are required to avoid pedestrians walking towards you while progressing under the time constraints of the crosswalk signal. These small changes, essentially responses to disturbances during steady-state walking, can occur at any time during the gait cycle. To adapt flexibly, using only one balance mechanism (i.e., foot placement change) available only during a short time window of the gait cycle, is not sufficient. Instead, multiple balance mechanisms which are temporally coordinated lead to the most flexible response to disturbances while walking. I will discuss how task, environmental constraints and neurological deficits may change the availability of these mechanisms and how loss of any one of these mechanisms requires compensation from a remaining mechanism that is not optimal for flexible, stable locomotion.
Speaker Bio :Prof. Shihab Shamma received his B.S. degree in 1976 from Imperial College, in London, U.K. He received his M.S. and Ph.D. degrees in Electrical Engineering from Stanford University in 1977 and 1980, respectively. Dr. Shamma received his M.A. in Slavic Languages and Literature in 1980 from the same institution. Dr Shamma has been a member of the University of Maryland faculty since 1984, when he started as an Assistant Professor in the Electrical Engineering Department. He has been associated with the Institute for Systems Research since its inception in 1985, and received a joint appointment in 1990. He is a fellow of the Acoustical Society of America and the Institute of Electrical and Electronics Engineers. Dr. Shamma's research deals with issues in computational neuroscience, euromorphic engineering, and the development of microsensor systems for experimental research and neural prostheses. Primary focus has been on studying the computational principles underlying the processing and recognition of complex sounds (speech and music) in the auditory system, and the relationship between auditory and visual processing. Prof. Shamma also holds the Pratiksha Trust Distinguished Chair position at IISc. Prof. Jeka was an NIMH predoctoral fellow and received his PhD in Neuroscience from Florida Atlantic University. He then received an NIH NRSA postdoctoral fellowship with the Ashton Graybiel Spatial Orientation laboratory at Brandeis University. After 18 years at the University of Maryland – College Park, he moved to Temple University as Professor and Chair of the Department of Kinesiology from 2013-2017 and is currently Chair of Kinesiology & Applied Physiology at the University of Delaware. His research interests include multisensory fusion for the control of human postural/locomotion and its application to the rehabilitation of individuals with balance disorders including cerebral palsy, vestibular loss and traumatic brain injury. He has received over $10 million in funding from the NIH, NSF, NASA and private sources such as the Erickson Foundation and the Shriners Foundation. He currently has two patents pending and is the Chief Scientific Officer of Treadsense, Inc., which develops technology for enhanced mobility.

Event : Seminar
Title : Latent Dirichlet Allocation
Speaker : Dr. Hemant Misra
Date : 31/10/2019
Venue : C 241 MMCR, EE
Abstract : Topic models such as Latent Dirichlet Allocation (LDA: https://en.wikipedia.org/wiki/Latent_Dirichlet_allocation) have been used extensively in the last decade for tasks such as information retrieval, topic discovery, dimensionality reduction etc. In the current presentation, the application of LDA for the task of text-segmentation (https://en.wikipedia.org/wiki/Text_segmentation#Topic_segmentation) has been explained. Results on multiple datasets are shown to demonstrate the performance of the proposed LDA based system vis-a-vis other standard methods. The talk also uncovers challenge faced by the dynamic programming (DP) algorithm used in proposed LDA based segmentation and how it was overcome. Talk will also cover some of the exciting things we are doing at Swiggy in the Applied Research team in the areas of speech, computer vision (CV) and natural language processing (NLP).
Speaker Bio :Dr. Hemant Misra is an active researcher in the areas of text and signal processing, speech/speaker recognition, machine learning, healthcare applications and education. He did his MS (1999) from IIT, Madras, and PhD from EPFL (2006). Then he held post-doc positions at Telecom ParisTech, University of Glasgow, and Xerox Research Centre Europe. After having successful stints at Philips (Healthcare) Research, IBM's India Research Lab and Citicorp Services India, currently Hemant is 'VP - Head of Applied Research' at Swiggy.
Host Faculty : Dr. Sriram Ganapathy (EE)

Event : Seminar
Title : Design of system integrity protection schemes using synchrophasor measurements
Speaker : Mr. Aditya Nadkarni
Date : 23/10/2019
Venue : C 241 MMCR, EE
Abstract : The talk will focus on controlled islanding and early detection of voltage instability in power systems. In controlled islanding, out-of-step prediction, early cutset determination and a transfer tripping scheme will be discussed. In another application, a unified early warning scheme (EWS) for detecting voltage instability will be presented. The approach shows that, by extrapolating real-time trends in PMU time series, limit violations in bus voltages and generator reactive power output can be estimated well in advance.
Speaker Bio :Aditya Nadkarni obtained his B.Tech in Electrical Engineering from VJTI (Mumbai University) in 2010, and M.S. in Power Systems, from Arizona State University, USA in 2013. He is currently pursuing his Ph.D. in Electrical Engineering from Indian Institute of Technology, Bombay. His areas of interest are design of real-time data-analytics and power system optimization.

Event : Thesis Defence
Title : Analysis of whispered speech and its conversion to neutral speech
Speaker : G Nisha Meenakshi
Degree Registered : PhD
Advisor : Dr. Prasanta Kumar Ghosh
Date : 22/10/2019
Venue : C 241 MMCR, EE
Abstract : Whispering is an indispensable form of communication that emerges in private conversations as well as in pathological situations. In conditions such as partial or total laryngectomy, spasmodic dysphonia etc, alaryngeal speech such as esophageal, tracheo-esophageal speech and hoarse whispered speech are common. Whispered speech is primarily characterized by the lack of vocal fold vibrations, and, hence, pitch. In recent times, applications such as voice activity detection, speaker identification and verification and speech recognition have been extended to whispered speech as well. Several efforts have also been undertaken to convert the less intelligible whispered speech into a more natural sounding neutral speech. Although supported by literature, research towards gaining a better understanding of whispered speech largely remains unexplored. Hence, the aim of the thesis is two-fold, 1) to analyze different characteristics of whispered speech using both speech and articulatory data, 2) to perform whispered speech to neutral speech conversion using the state-of-the-art modelling techniques.
In the first part of this thesis, we analyze whispered speech using both audio data (recorded via microphone) and articulatory data (recordings of movements of articulators, such as lips, tongue, jaw etc, using Electromagnetic Articulography synchronous with audio data). Specifically, we experimentally analyze how the pitch-less whispered speech encodes information such as speaker's gender and voicing, that are typically pitch-dependent in neutral speech. We find that whispered speech does retain speaker's gender and voicing related information. This could be attributed to the exaggerated movements of the articulators that typically occur while trying to maintain intelligibility in the absence of pitch. Therefore, we next investigate for the optimal transformation function that relates whispered articulatory movements with those of neutral speech. Experiments reveal that an affine transformation could relate the two articulatory movements better than other candidate functions considered. In addition, we also find how much the acoustics of whispered speech carries information about the corresponding articulatory movements compared to that of neutral speech.
In the second part, we design a feature that is necessary for segmenting whispered speech from a long recording of noisy whispered speech interleaved with silence/noise segments, as a per-processing step in the conversion/reconstruction framework. In order to reconstruct neutral speech from whispers, we follow a voice conversion-based approach which requires an appropriate parametrization of the whispered speech spectrum. For this, we experimentally find an optimal choice of parameters that is robust, both, for representation and to handle modelling errors. This representation is employed in the proposed bi-directional long short-term memory based whispered to neutral speech conversion system that yields a perceptually more natural sounding speech compared to the state-of-the-art conversion systems.

Event : Thesis Defence
Title : Kernel-Based Image Filtering: Fast Algorithms and Applications
Speaker : Sanjay Ghosh
Degree Registered :PhD
Advisor : Dr. Kunal Narayan Chaudhury
Date : 21/10/2019
Venue : C 241 MMCR, EE
Abstract : Image filtering is a fundamental task in computer vision and image processing. Various linear and nonlinear filters are routinely used for enhancement, superresolution, sharpening, restoration, etc. The focus of this thesis is on kernel-based filtering that has received significant attention in recent years. The basic idea of kernel filtering is quite straightforward, namely, each pixel in the image is replaced by a weighted average of its neighboring pixels. The weighting is performed using an affinity kernel, which is generally non-negative, symmetric positive definite. Depending on the choice of the kernel function, there could be different filters: Gaussian, bilateral filter, nonlocal means, guided filtering, etc. While the dominant applications of kernel filtering are enhancement and denoising, it can also be used as a powerful regularizer for image reconstruction. In general, the brute-force implementation of kernel filtering is prohibitively expensive. Unlike convolution filters, they cannot be implemented efficiently using recursion or the fast Fourier transform. In fact, their brute-force implementation is often too slow for real-time applications. The key motivation of this work was to develop fast approximation algorithms for kernel filtering and explore their applications.
We have focused on two popularly used kernel filters, bilateral filter and nonlocal means, in this thesis. In the context of bilateral filtering, we demonstrated that by using Fourier approximation of the underlying kernel, we can obtain state-of-the-art fast algorithm for filtering of gray images. The main idea is to express the filtering as a series of fast convolutions, which are applied to simple nonlinear transforms of the input data. We achieved around 50x speedup using our proposed method. In relation to existing works, a unique aspect of our method is that we are able to analyze and provide theoretical guarantees on the filtering error incurred by the approximation. We extended this to color images, texture smoothing, low-light image enhancement, etc. In a different direction, we have developed a fast algorithm for symmetrized nonlocal means, which can be used as a regularizer (denoiser) in plug-and-play image restoration. Plug-and-play is a recent paradigm where a powerful denoiser is used to regularize the inversion of the measurement model within an iterative framework. The attractive aspect of symmetrized nonlocal means is that the associated plug-and-play iterations are fast and provably convergent. In practice, the proposal algorithm can significantly speedup various restoration tasks such as deblurring, inpainting, superresolution, and single-photo imaging--what would typically take minutes can now be done in seconds.

Event : Thesis Defence
Title : Modelling, Optimisation and Control of Photovoltaic Energy Conversion Systems
Speaker : Pallavi Bharadwaj
Degree Registered : PhD
Advisor : Prof. Vinod John
Date : 10/10/2019
Venue : B 303, EE
Abstract : Uncertainty in global fossil fuel supplies and rising climate change concerns call for an urgent need to switch to renewable energy resources. There are several challenges associated with the harnessing of solar power, which have so far limited its contribution to only 8% of India's energy mix. This work is an attempt to understand and overcome some of the fundamental challenges associated with the utilisation of the solar power.
Measurement of input and output of PV systems is the first challenge which involves expensive pyranometers. The development of in-house, low-cost, high-performance irradiation meters, with performance standardised on the basis of ISO 9060 standard, facilitates input irradiation and temperature input measurement. For output measurement, a switched mode power conversion based closed-loop controlled PV characterisation setup is developed in this work which reduces the ripple in PV current measurement as compared to open loop response.
Being a stochastic source of energy, it is difficult to predict its behavior but with the novel sequential parameter extraction method, this work presents a way of predicting the energy output of PV systems under varying ambient conditions with an improvement of 10% in the PV output prediction compared to the datasheet-based existing methods. The sequential optimisation solves the second challenge of modelling PV modules under steady state which is further enhanced for transient modelling by experimental evaluation of PV capacitance.
The third major challenge in PV energy conversion is imposed by shading, which is addressed by the first time introduction of subcell model to partial shading analysis. Subcell model is experimentally verified to be 93% accurate for opaque shading and 95% accurate for translucent shading. This model further facilitates the understanding of hotspot formation and dust induced reliability issues.
The fourth major problem is the global maximum power point tracking, to which the PV fraternity is still looking for a solution. This is addressed in this work using a fundamental shading fraction based GMPPT algorithm, wherein the shading versus global peak correlations are derived using module voltage information. This method is scalable from small to large PV strings and provides high maximum tracking speeds and improved energy capture as compared to popular existing MPPT algorithms.

Event : Thesis Colloquium
Title : Knowledge-driven training of deep models for better reconstruction and recognition
Speaker : Ram Krishna Pandey
Degree Registered :PhD
Advisor : Prof. A G Ramakrishnan
Date : 16/09/2019
Venue : C 241 MMCR, EE
Abstract : The thesis shows that the performance of any deep architecture can be improved at three different levels: (i) input, (ii) architecture and (iii) at the objective or loss function levels. The following are the key contributions of the thesis:
(i) Techniques are proposed to enhance the quality of low-resolution, binary document images for better human readability and OCR performance. The mean opinion score of native readers increases from 4.5 to 9.6 on a scale of 0 to 10 and the OCR accuracy increases from 26% to 64% for a 100-dpi binary Tamil input document. (Input and architecture exploration)
(ii) New architectures are proposed for superresolution of natural images. Multiple interpolations are fused in a deep network to obtain better reconstruction. Results are comparable to the state of the art. (Input and architecture exploration)
(iii) Mean square Canny error (MSCE) is proposed as a ``new loss function" that improves the performance of state of the art deep architectures (super-resolution or denoising) that originally use mean square error (MSE) as loss function. (Objective exploration)
(iv) It is shown that feeding the gradient and/or the Laplacian of the input image improves the performance of facial emotion classifiers by a good margin, with no additional computational overhead during inference. This is used to find a lightweight model without compromising the classification accuracy. (Input, architecture and objective explorations)

Event : Seminar
Title : Critical Clearing Time Calculation of Single Photovoltaic Generator Infinite Bus (SPIB) System using Energy Functions
Speaker : Indla Rajitha Sai Priyamvada
Degree Registered : PhD
Advisor : Dr. Sarasij Das
Date : 13/09/2019
Venue : C 241 MMCR, EE
Abstract : With the concern of environmental impact, the power systems are experiencing a shift from fossil fuel-based generation to Renewable Energy (RE) based generation such as wind and Photovoltaic (PV). The transient stability of the power systems must be maintained to ensure reliability of the power systems. The Critical Clearing Time (CCT) is an important index in determining the transient stability of power systems. The swing equation of Synchronous Generators (SGs) and equal area criterion are used to determine the CCT of SGs. However, the method used for SGs is not applicable for PV generators. A novel method to estimate CCT of a Single PV generator - Infinite Bus (SPIB) system will be presented in this talk.
Speaker Bio : Rajitha is currently pursuing her PhD in Electrical Engineering under the guidance of Dr. Sarasij Das. Currently, she is working towards developing methods to analyse the transient stability of grid connected PV generators.

Event : Thesis Defence
Title : Hardware Emulation of a Long Transmission Line by High Frequency Power Electronic Converter for the study of Switching Transients
Speaker : Sushmit Mazumder
Degree Registered :MTech(Research)
Advisor : Dr. Kaushik Basu
Date : 04/09/2019
Venue : C 241 MMCR, EE
Abstract : Hardware emulators, simulating a test environment in real-time, are an essential tool in testing power system equipment. This work presents a hardware emulator of a programmable transmission line capable of simulating high frequency transients. Emulation of, energization of long transmission line, requires real-time solution of transmission line equations. The work identifies a continuous time model that captures the wave nature and suitable for real-time implementation. The discrete time model to be solved in the observer of the hardware emulator is derived. A step by step procedure is developed to determine two key emulator parameters: observer sampling frequency and switching frequency of the power amplifier considering different hardware, event and software related constraints. For high bandwidth requirement, a high switching frequency, 100kHz, 12 kVA SiC based voltage source converter is designed to operate under a resonant based current controller. A SoC that combines an ARM processor along with a FPGA, is used to implement the observer. Experimental results verify accuracy of the designed emulator in the event of energization of a long transmission line.

Event : Thesis Defence
Title : Fast and Robust Biomedical Image Reconstruction from Nonuniform Samples
Speaker : Bibin Francis
Degree Registered :PhD
Advisor : Dr. Muthuvel Arigovindan
Date : 29/08/2019
Venue : C 241 MMCR, EE
Abstract : We consider the problem of reconstructing images from non-uniformly under-sampled spatial point measurements with emphasis on robustness to noise. The computational methods that deals with this problem are known as scattered data approximation (SDA) methods. Among these, well-performing methods achieve the reconstruction by minimizing a cost that is a weighted sum of data fidelity term measuring the accuracy of the fit at the measurement locations, and a regularization term. The latter term incorporates certain smoothness, and is constructed by summing the squared derivative values of a chosen order. The relative weight between these two terms is known as the smoothing parameter. Prominent methods in this category are known as thin-plate spline (TPS) and radial basis function (RBF) methods, and they require solving large numerically ill-conditioned and/or dense linear system of equations. Subspace variational method alleviates the numerical instability and the computational complexity associated with the TPS and RBF methods. However, this approach involves solving large and sparse linear system of equation requiring specialized numerical methods.
In the first part of the thesis, we propose a novel method for SDA that eliminates the need for solving dense linear system of equations, and even the need for storing matrix representing linear system. This is achieved by handling the reconstruction problem in two stages. In the first stage, the given non-uniform data are transformed into a pair of regular grid images, where, one image represents the measured samples and the other represents the sample density map. In the second stage, the required image is computed as the minimizer of a cost that is completely expressed in terms of regular grid discrete operations. It is expressed as a sum of weighted quadratic data fitting term involving the transformed image pair, and and discrete quadratic roughness functional. Computing the minimizer of this cost involves solving a well-conditioned sparse linear system of equations, where system matrix is represented in terms of filtering and array multiplications without the need for storing it explicitly. We demonstrate that the proposed method, which is named as regular grid weighted smoothing (RGWS), has much lower computational complexity than TPS and RBF methods, with only a little compromise in the reconstruction quality.
RGWS uses quadratic regularization, which is known to yield over-smoothed images under the presence of noise. We extend the RGWS method by incorporating non-quadratic regularization which is constructed by applying a square root on the sum of squares of derivative values (known as l1 regularization). We propose a reconstruction method using this l1 regularization, which we name as the l1-RGWS. We perform extensive set of reconstruction experiments with various levels of under-sampling and noise and compare the performances of l1-RGWS and the original RGWS, which we also call l2-RGWS.
When the sampling density becomes low, the performance of l1-RGWS degrade abruptly. This behavior is referred as phase transition in the literature. For the present problem, we observed that the quality of reconstruction with l1 regularization falls below l2-RGWS for very low sampling densities . We analyze this in a probabilistic viewpoint and infer that the prior probability model corresponding to l1-regularization is based on the assumption that probability of a pixel location taking certain derivative value is independent of the derivative values of its neighboring pixel locations, which is clearly not true. We developed a probability model where error incurred by this independence assumption is compensated by means of a multi-resolution based re-weighting scheme. In this scheme, the desired reconstruction is formulated as a series of coarse-to-fine multi-resolution reconstructions, and re-weighting of the prior probability for each resolution level is derived from the reconstruction of previous resolution level. We demonstrate that the new method, which we name the multiresolution based scattered data approximation (MSDA), performs better than l1-RGWS and l2-RGWS under wide range of sampling densities, with slightly increased computational complexity.
We then developed an extended method, where, instead of re-weighting the form of prior probability model corresponding to l1 regularization, the probability model itself is determined using maximum entropy principle. Specifically, at each resolution level in the multi-resolution reconstruction, the required probability model is determined as the maximizer of entropy subject to the information extracted from the lower resolution reconstruction as constraints. To further enhance the performance, we use directional second derivative operators to define the probability model. Moreover, to control the variance of this probability model, we also propose to use a modified multiresolution scheme, where the image sizes increase by a fractional factor, instead of doubling. We demonstrate that the new method, which we call the maximum entropy regularized reconstruction (MERR), outperforms both MSDA and l1-RGWS for a wide range of sampling densities and noise levels.

Event : Thesis Colloquium
Title : Detection of Faults in Ungrounded Double Wye Shunt Capacitor Banks
Speaker : Polisetty Sai Pavan
Degree Registered :MTech (research)
Advisor : Dr. Sarasij Das
Date : 27/08/2019
Venue : C 241 MMCR, EE
Abstract : Shunt capacitor banks (SCB) are commonly used to provide reactive power support in both transmission and distribution systems. Outage of any SCB is crucial and hence, should be well protected against various types of faults. Double wye connected SCBs are widely used at high voltage levels. Neutral current compensation method is commonly used to detect internal faults in ungrounded double wye SCB. Existing protection algorithms mostly fail to detect simultaneous faults (with same fault starting time). Existing algorithms either fail to detect a faulted condition or detect one type of fault as some other type of fault. In this work, a novel method for detecting different types of internal faults in ungrounded double wye SCBs has been developed. One of the main advantages of the proposed method is that it can detect simultaneous faults with same fault starting time. The proposed method can detect different types of internal faults under supply voltage unbalance and impedance unbalance (during healthy condition) situations. A laboratory scale ungrounded double wye SCB test setup is developed to test the working of proposed method.

Event : Seminar
Title : Pitch-synchronous discrete cosine transform features for speaker identification and verification
Speaker : Amit Meghanani
Degree Registered :M.Tech (Res)
Advisor : Prof. A. G. Ramakrishnan.
Date : 02/08/2019
Venue : C 241 MMCR, EE
Abstract : A new feature named Pitch-synchronous discrete cosine transform (PS-DCT) for speaker identification (SID) and verification (SV) is proposed here. This feature exploits the time-domain, quasi-periodic structure of the voiced phones. To capture the waveform shape of the voiced phones, we employ discrete cosine transform (DCT). Our experiments on various datasets demonstrate that the proposed features can be supplemented to other classical features. Improvement is more prominent in limited data speaker verification tasks.
Speaker Bio :Amit Meghanani is M.Tech (Res) student in the Department of Electrical Engineering, IISc, working with Prof. A. G. Ramakrishnan. He did his B.Tech in ECE from NIT Silchar. His research interest is in Speaker identification and verification, Natural language processing.

Event : Seminar
Title : Unified Generator Classifier for efficient Zero-Shot Learning
Speaker : Ayyappa Kumar
Degree Registered :M.Tech (Res)
Advisor : Dr. Soma Biswas
Date : 19/07/2019
Venue : C 241 MMCR, EE
Abstract : Generative models have achieved state-of-the-art performance for the zero-shot learning problem, but they require re-training the classifier every time a new object category is encountered. The traditional semantic embedding approaches, though very elegant, usually do not perform at par with their generative counterparts. In this talk, we propose an unified framework termed GenClass, which integrates the generator with the classifier for efficient zero-shot learning, thus combining the representative power of the generative approaches and the elegance of the embedding approaches. End-to-end training of the unified framework not only eliminates the requirement of additional classifier for new object categories as in the generative approaches, but also facilitates the generation of more discriminative and useful features.
Speaker Bio :Ayyappa is an M.Tech (Research) student in the Department of Electrical Engineering at IISc working with Dr. Soma Biswas.

Event : Seminar
Title : Location of High Impedance Faults Using Smart Meters
Speaker : Asha Radhakrishnan
Degree Registered :PhD
Advisor : Dr. Sarasij Das
Date : 26/07/2019
Venue : C 241 MMCR, EE
Abstract : A High Impedance Fault (HIF) occurs either when an overhead conductor breaks and falls to the ground or when an energized primary conductor makes contact with an object of high impedance such as a tree or building. HIF is a safety hazard when left undetected as the energized conductor remains exposed. HIFs involve fault currents with low magnitudes which make HIFs difficult to be detected. HIF detection has thus been extensively explored by research community. Numerous efforts have been made in the literature to address HIF detection. However, location of HIFs has remained as a challenge to the protection engineers. A novel method has been proposed here to locate HIFs using Smart Meters (SM) in distribution systems. The performance of the proposed method has been evaluated considering EAFs, DGs and power electronics interfaced loads. The proposed algorithm has also been implemented on a commercial smart energymeter to demonstrate the feasibility.
Speaker Bio :Asha Radhakrishnan is a doctoral student in the Department of Electrical Engineering, Indian Institute of Science, Bangalore working with Dr. Sarasij Das

Event : Seminar
Title : Class-specific and noise-specific speech enhancement approaches
Speaker : Nazreen
Degree Registered :PhD
Advisor : Prof. A G Ramakrishnan
Date : 09/07/2019
Venue : C 241 MMCR, EE
Abstract : This thesis proposes, implements and analyzes various speech sound class-specific and noise-specific enhancement approaches and frame-wise selection methods for class-specific and noise-specific models. We have analyzed the performance of our enhancement scheme, where we use various speech-sound class-specific dictionaries to enhance noisy speech. By removing the contribution from the bases of those classes that correlate well with noise, one could improve the enhancement performance. We achieve this by learning different dictionaries for different classes and select a particular dictionary for a frame. We explore a class-specific enhancement approach, where we use a sparse coding, dictionary-based approach to learn dictionaries of various speech classes and noises. When we analyze the performance of our various class-specific approaches in terms of phoneme recognition, we obtain performances superior to class-independent case, even when we use estimated (approximate) labels for enhancement.
An error in the estimated class labels in the class-specific approach results in the selection of erroneous dictionaries for enhancement. The joint enhancement-decoding (JED) algorithm that we propose tries to overcome this issue by jointly optimizing for labels of all the frames and the decoding path to improve the phoneme recognition accuracy. The current noisy speech frame is enhanced by multiple (N) phoneme-specific dictionaries close to the approximate label of that frame. These N enhanced frames are then fed into the JED algorithm. The algorithm accepts these N observations and chooses the best for each frame such that the overall likelihood is maximized to obtain the final recognized labels. The Viterbi decoding algorithm used in speech recognition is integrated with the class label selection to develop the JED algorithm.
We also propose a method of picking the best DNN model in the scenario, where multiple noise-specific DNN models are available for enhancement, using the Monte Carlo (MC) dropout proposed by Gal and Ghahramani. The variance measure of the output signal vectors, resulting from different MC dropout trials, is used as a measure of the model precision to select one out of the multiple models for each frame. We find this method to be particularly useful for unseen noisy scenario, where the noise corrupting the test speech is different from those with which the available DNN models are trained. For the unseen noisy scenario, this method performs better than selecting the model using a DNN classifier. We observe some promising results with the enhancement performance of the algorithm on speech corrupted with a mixture of multiple noises and for the case, where different segments of speech are corrupted by different noises.

Event : Thesis Colloquium
Title : Pronunciation assessment and semi-supervised feedback prediction for spoken English tutoring
Speaker : Chiranjeevi Yarra
Degree Registered :PhD
Advisor : Prasanta Kumar Ghosh
Date : 08/07/2019
Venue : C 241 MMCR, EE
Abstract : Spoken English pronunciation quality is often influenced by the nativity of a learner, for whom English is the second language. Typically, the pronunciation quality of a learner depends on the degree of following four sub-qualities: 1) phoneme quality 2) syllable stress quality 3) intonation quality, and 4) fluency. In order to achieve a good pronunciation quality, learners need to minimize their nativity influences in each of the four sub-qualities, which can be achieved with effective spoken English tutoring methods. However, these methods are expensive as they require highly proficient English experts. In cases, where a cost-effective solution is required, it is useful to have a tutoring system which assesses a learner's pronunciation and provides feedback in each of the four sub-qualities to minimize nativity influences in a manner similar to that of a human expert. Such kind of systems are also useful for learners who can not access high quality tutoring due to their demographic and physical constraints. In this thesis, several methods are developed to assess pronunciation quality and provide feedback for such a spoken English tutoring system for Indian learners.
Most of the existing works on automatic pronunciation assessment predict an overall pronunciation quality. However, feedback prediction has typically been done separately in each of the four sub-qualities. Both pronunciation assessment and feedback prediction require annotations on a large set of recordings from learners. While the former requires ratings for overall pronunciation quality, the later needs feedback specific labeling. Unlike ratings, obtaining labels for feedback prediction requires highly skilled annotators. Such annotators are not available in large numbers and labeling with their expertise is also costly. Due to this paucity of labels, it is challenging to design a tutoring system in a cost effective manner particularly for Indian nativity, which is known for its large accent variabilities. With regard to these challenges, the key contributions in this thesis are: 1) building models for estimating parameters for providing meaningful feedback without using any labelled data, 2) building models for estimating overall pronunciation quality using annotated data, and 3) developing voisTUTOR, a system for learners to train themselves with neutral accent of English with the help of a spoken English expert. The feedback prediction is semi-supervised in nature as no feedback-specific labels are used for building feedback prediction models.
Feedback in each of the four sub-qualities is predicted by analyzing mismatches in the respective parameters between learners' and an expert's speech. In the phoneme category, phoneme errors made by a learner are provided as feedback, where the phonemes are estimated using rule based pronunciation dictionary. These rules are deduced from the errors made by the Indian learners while speaking English. For demonstrating the correct pronunciation, an articulatory video is synthesized using an expert's speech. Further, the effect of accents on the uttered phonemes is assessed using goodness of pronunciation measure, which is computed in a deep neural network-hidden Markov model (DNN-HMM) based automatic speech recognition (ASR) framework. In the stress category, mismatches in the estimated stressed syllable locations are provided as feedback. For this, stress-specific features are computed by exploring linguistic parameters, such as sonority, from every syllable when the ground truth syllable information is available. Its performance is analysed when the syllable information is estimated as in a real scenario. The stress locations are also estimated in an ASR framework without computing any stress-specific features. In the intonation category, feedback is provided based on the local and global mismatches in pitch patterns. For this, models are proposed to estimate the pitch values and their associated confidence scores. It is observed that the global mismatches depend on temporal variations in the pitch and its patterns. These mismatches are identified better when the confidence scores along with the pitch values are used in the models, based on HMMs and long-short term memory (LSTM) networks. Both the global and local mismatches are identified using knowledge driven template matching approach, that performs confidence score based median filtering and pitch stylization. In the fluency category, mismatches in the pause locations are provided as feedback. The pause locations are estimated using features based on speech acoustics only without considering any canonical stress markings because the learners' pronunciation do not often match the canonical pronunciation. Further, analysis is performed to estimate speaking rate directly from the speech acoustics, where speaking rate has been shown to be correlated with the fluency of a learner's pronunciation.
Overall pronunciation rating is estimated using a joint model considering DNNs and LSTM networks. For this, studies are conducted to find out differences between the speech rhythm of Indian languages and that of English. Features based on speech rhythm are used for estimating the rating along with the features based on the parameters used for the feedback in all four sub-qualities. Further, in order to create an interactive learning environment in voisTUTOR, these feedback and the ratings are displayed using audio-visual aids including line and bar graphs and text messages. All of these are made available in an android app using a web-server with LAMP (Linus, Apache, MySQL, PHP) stack on Ubuntu 14.04 LTS system.

Event : Seminar
Title : Novelty Detection in Generalized Zero-Shot Learning
Speaker : Devraj Mandal
Degree Registered :PhD
Advisor : Dr. Soma Biswas
Date : 05/07/2019
Venue : C 241 MMCR, EE
Abstract : Generalized Zero-Shot Recognition is a challenging problem, where the task is to recognize new categories that are unavailable during the training stage, in addition to the seen categories. Existing approaches suffer from the inherent bias of the learned classifier towards the seen categories. As a consequence, unseen category samples are incorrectly classified as belonging to one of the seen categories. In this talk, I will discuss some strategies by which we can mitigate this problem. In one approach we will discuss how a simple center loss trained along with the standard classification loss can help to detect the out-of-distribution samples. In the next part of my talk, I will discuss how an out-of-distribution classifier can be potentially trained using a generative based approach to mitigate this bias in the classifier performance.
Speaker Bio :Devraj Mandal received the B. Tech degree in Electronics & Communication Engineering from West Bengal University of Technology, Kolkata, in 2011 and the MTech degree from Jadavpur University, Kolkata, in 2014. He is currently a doctoral student in the Department of Electrical Engineering, Indian Institute of Science, Bangalore, India. His research interests are in image processing, computer vision, and pattern recognition.

Event : Thesis Colloquium
Title : Estimating the variation of NOx in diesel exhaust treated with discharge-plasma/ozone injection: a case study on modelling with ANN & Dimensional-Analysis
Speaker : Dipanwita Sinha M
Degree Registered : PhD
Advisor : Prof. B S Rajanikanth
Date : 05/07/2019
Venue : HVE Seminar Hall
Abstract : Controlling the gaseous pollutants in the diesel engine exhaust is getting more challenging due to the stringent limits imposed by the Government authorities. Amongst these pollutants the one which contributes towards greenhouse effect and acid rain are the oxides of nitrogen (NOx). Two different approaches to reduce NOx have been practiced namely Pre-combustion and Post-combustion. While the Pre-combustion technique is more or less saturated, the Post-combustion technology is mainly dependent on catalysis/adsorption. The use of high expensive materials and shorter life span of catalyst makes the catalysis process expensive. On the other hand, continuous regeneration processes make the adsorbents another expensive option.
In this scenario, research dimension is being shifted to newer and economical alternatives and one such option is electrical discharge based non-thermal plasma (NTP), which is gaining significance owing to the success that was achieved at the laboratory level in the reduction of NOx/VOCs. Several non-thermal plasma techniques exist for NOx reduction in the diesel exhaust namely pulsed plasma processing, dielectric barrier discharges (DBD), surface discharge and electronic beam based gas cleaning. Among all of them, DBD is the popular one where discharges are self-contained allowing operation of the plasma reactor at higher voltages without causing a complete breakdown. In all the NTP treated diesel exhaust there will be copious generation of oxidizing and reducing radicals by high energetic electrons which in turn are dependent on the applied high voltage parameters. The radicals thus generated induce mixed chemical reactions in the environment where the gases are getting treated and cause overall reduction of NOx. A review of literature survey observed that for the past one decade, injection of Ozone into the exhaust gas stream could also result in desirable chemical reactions prevailed by oxidation of the pollutants. However, the works on ozone injection technique to cleanse the diesel exhaust is rather scarce. In this thesis an effort has been made to remove NOx from diesel engine exhaust by causing the chemical reactions with ozone which was generated by a separate DBD reactor followed by a quasi - experimental model based on dimensional analysis to predict the removal of NOx upon mixing with ozone. This method involves a mathematical derivation incorporating variation of flow rates and amounts of ozone injected to estimate the nitric oxide (NO) and nitrogen dioxide (NO2) concentrations in the presence of ozone. The predicted results were then validated with experimental observations.
It has been observed that the injection of ozone into the diesel exhaust critically depends on the ratio of nitric oxide present in the exhaust to the amount of ozone injected. As the concentration of NO varies with the loading, accordingly ozone levels also should be varied and under such circumstances the need is felt to predict or estimate the ozone concentration for a given set of parameters. In the next part of the thesis, an attempt has been made to predict the DBD plasma generated ozone using artificial neural network (ANN) approach. A multilayer feed-forward ANN model was developed while considering physical and electrical parameters as inputs to ozone production. The predicted values were then validated.
The last part of the thesis is about treatment of engine exhaust with direct exposure to DBD plasma and subsequent prediction of converted gases. DBD reactor has been extensively used in several gas treatment applications. During the application of DBD plasma for gas cleaning, several parameters play a major role in the conversion/oxidation of gaseous pollutants namely physical parameters like reactor configuration, gas flow rate, gas temperature, permittivity of dielectric material and electrical parameters. ANN plays a major role in this kind of situations where the underlying phenomena is complex and time consuming to model it theoretically. In this work, an attempt has been made to model the DBD plasma assisted exhaust treatment with ANN to estimate the reduction of NO as well as production of NO2 when subjected to several parametric variations such as, applied power, frequency of the applied pulses, gas flow rate and engine load, without involving plasma physics and chemical kinetics. The non-linearity between the dependent (NO and NO2) and the independent parameters (physical and electrical) have been analyzed through the model. The predicted results agree well with the experimental ones. The thesis thus presents couple of predictive techniques that may help in the long run to easily set the input parameters to suit the requirements of non-thermal plasma or electric discharge plasma-based exhaust cleaning or ozone generation.

Event : Thesis Colloquium
Title : Soft Switched Multilevel Unidirectional High Frequency Link DC-AC converter for Medium Voltage Grid Integration
Speaker : Manmohan Mahapatra
Degree Registered :MTech(Research)
Advisor : Dr. Kaushik Basu
Date : 03/07/2019
Venue : C 241 MMCR, EE
Abstract : With decreasing fossil fuel reserve, solar power is gaining huge popularity. It is harvested in the form of DC up to 1 kV because of safety regulations and to limit solar panels' leakage current. For grid integration of solar power, a 3ph voltage source inverter (VSI) converts the DC to 400 V line frequency AC. Then it is connected to 400 Volt grid through a line frequency transformer (LFT). The LFT ensures safety regulations and limits circulating current. The 400 V grid is connected to medium voltage (MV) transmission grid through a step up LFT. These LFTs are bulky, heavy and costly. Further, in conventional PV-grid integration, line filters are placed in low voltage side of the transformer and the current flowing through these filter is high leading to higher copper losses. An alternate solution is the use of high frequency transformer (HFT). The HFT is fed from a DC side converter (DSC) and the output of HFT is fed to a power electronics converters to convert LF-AC from HF-AC. These type of topologies are known as high frequency link (HFL) DC/AC converters. State of the art HFL DC-AC converters are commonly multi-stage solutions where an isolated DC/DC converter is cascaded with a 3ph VSI through inter-stage filter capacitor. This filter capacitor reduces the reliability of the converter and the 3ph VSI is hard switched. Alternate solution for multi-stage conversion is single-stage conversion where inter-stage filter capacitor is not required, hence improves reliability. In these topologies, the step LFT is still there.
For MV grid integration without step-up LFT, multilevel converters are used. Multilevel converters like diode clamped and flying capacitor are used for power conversion from a single DC source whereas cascaded H-bridge (CHB) is used for multiple isolated DC sources. In a CHB topology, isolated DC supplies are generally produced by using phase shifted full bridge (PSFB) based DC/DC converters with a filter capacitor at the output of the PSFB converters. This is similar to multiple multi-stage HFL connected in series to achieve higher voltage.
In this thesis, a new CHB based unidirectional single stage multilevel converter topology is proposed for MV grid integration. The topology does not employ any inter-stage filter capacitor and uses HFT for isolation and voltage step-up. The DSC is zero voltage switched (ZVS) for most part of the line cycle without any additional snubber elements. The CHB’s switches interfacing the MV grid are switched at line frequency incurring negligible switching loss. The converter has modular structure which helps in easy repair-replacement. Modulation strategy of the converter and key simulation results are given assuming the converter is ideal.
As the MV side switches are LF switched, over a switching cycle the proposed topology behaves as an MV step-up PSFB converter. In MV step-up PSFB converter, the diode bridge parasitic capacitance affects the circuit operation and hardware design. There is no known literature that describes MV step-up PSFB converter considering all parasitics. In this thesis, a detailed analysis of the proposed converter’s operation is given when it is run as an MV step-up PSFB converter. The hardware is designed and tested with 400 V DC input, 1240 V DC output voltage, 1.5 kW output power and switching frequency 20 kHz. Analysis and experimental results are matched which validates the analysis.

Event : Seminar
Title : Coordinated Conveying
Speaker : Prof. Shivakumar Sastry
Date : 28/06/2019
Venue : C 241 MMCR, EE
Abstract : We present a well-structured, mixed-criticality, system that exhibits rich spatio-temporal behaviors that arise from the interactions of mobile conveying units. The system-level objective is for the collection of decentralized, autonomous, mobile units to transport entities from some input port to an output port when each entity has its own destination, deadline and Quality of Service constraints. Entities move by riding on the mobile units and transfer from one unit to another when two units rendezvous. We propose a systematic approach to transform the spatio-temporal patterns of interactions to a graph from which we can obtain the shortest paths that can help to address the system objective. In the future, such systems can be used to investigate a variety of intelligent, embedded, mechatronic systems issues. This system can also be integrated with other models for processes and systems to investigate some of the challenges envisaged in Industry 4.0.
Speaker Bio : Dr. Shivakumar Sastry (Alumni, EE, IISc, MSc Engineering, 1987) is a Professor with the Department of Electrical and Computer Engineering, The University of Akron. He received his Ph.D. degree in Computer Engineering and Science from Case Western Reserve University and holds Masters Degrees in Computer Science from University of Central Florida and in Electrical Engineering from the Indian Institute of Science. His research interests are in Networked Embedded Systems, Real-time systems, Graph algorithms and Network Analysis. Prior to joining Akron, he was a Senior Research Scientist with Rockwell Automation. He is currently also serving as the Director of Strategic Initiatives.
The talk is jointly hosted by IEEE Signal Processing Society Bangalore Chapter and Department of Electrical Engineering, IISc.

Event : Thesis Defence
Title : Face Recognition in Unconstrained Environment
Speaker : Sivaram Prasad Mudunuri
Degree Registered :PhD
Advisor : Dr. Soma Biswas
Date : 27/06/2019
Venue : C 241 MMCR, EE
Abstract : The increasing use of surveillance cameras for addressing security concerns has led to increased demand for robust face recognition systems. The images captured by the surveillance cameras usually have poor resolution, uncontrolled pose and illumination conditions which makes the task of recognizing these faces extremely challenging. Significant attention has been devoted to addressing one or more of the different challenges like poor illumination, non-frontal pose, expression, etc. But addressing all these challenges together is essential in many applications like recognizing faces from surveillance cameras. The research in this thesis is motivated by this need.
The main focus of this thesis is to develop algorithms to match low-resolution (LR) facial images captured under wide range of pose variations and poor illuminations with high-resolution (HR) facial images captured in frontal pose and good illumination conditions which are often available during enrolment. The contributions of our work are as follows: 1) We addressed the problem of low-resolution face recognition across pose and illumination variations 2) Addressed the problem of matching LR NIR faces with HR VIS faces to address the recognition task in low-light conditions.
In the first part of this thesis, we address LR face recognition task using metric learning which does not require localizing facial landmarks in non-frontal face images at low resolution during testing. We then explored the LR face verification task using the deep learning framework. We generalized this deep learning framework for LR object recognition in which the testing objects have not been seen during the training.
The work in the second part of this thesis is motivated by the fact that recognizing the faces of uncontrolled subjects captured in night-time/low-light conditions using near-infrared (NIR) cameras is quite challenging since the images have very different visual characteristics from the gallery images that are captured using RGB camera. Several approaches have been proposed to address this task, but most of them deal with good resolution images. The NIR images captured by the surveillance cameras usually have low resolution in addition to considerable variations in pose. This makes the problem even more challenging, since now the low-resolution, uncontrolled NIR probe images needs to be matched against the high-resolution controlled VIS gallery images. But this scenario is relatively less explored in literature. We proposed a dictionary alignment approach for addressing this problem. We also proposed a re-ranking approach to further improve the recognition performance for each probe by combining the rank list given by the proposed algorithm with that given by another complementary feature/algorithm. Finally, we have also collected our own database HPR (Heterogeneous face recognition across Pose and Resolution) which has facial images captured from two surveillance quality NIR cameras and one HR visible camera, with significant variations in head pose and resolution.

Event : Thesis Defence
Title : Attention Feedback and Representations in OCR
Speaker : Shiva Kumar H R
Degree Registered :PhD
Advisor : Prof. A G Ramakrishnan
Date : 24/06/2019
Venue : C 241 MMCR, EE
Abstract : The thesis work has three major contributions: 1. Design and development of an industry-grade OCR system for Kannada that performs better than Google’s Tesseract OCR on a challenging dataset of 250 images created by the candidate, and made publicly available with the ground truth. OCR’s are mainly required to digitize very old and legacy printed documents, which have many challenges such as old script, broken and merged characters and interspersed English words. The dataset has a good proportion of pages with all the above issues, and the OCR is still able to perform better than the already high baseline accuracy of > 95% of Tesseract.
2. Proposing the problem of segmentation of overlapping text lines from printed and handwritten documents as a computer science problem of representing each text line as a node in a red-black tree and assigning the connected components in the page image to different nodes. It also uses bipartite graph representation to assign corresponding line segments from different vertical partitions of the document page and eliminate line merges and splits. Rather than cutting the image into different line images, the algorithm builds (assembles) each line image by adding the connected components assigned to each particular node.
3. Inspired by the rich feedback in the mammalian visual neural pathway, he has proposed improvements to image enhancement, binarization, segmentation and hence the recognition, using feedback from the latter modules such as SVM, Viterbi decoder and Unicode formation. This significantly improves the recognition performance of the OCR, in the presence of split and merged characters and also interspersed English words in the document page. On document pages of Kannada, Tulu, Konkani and Sanskrit text printed in Kannada script, this attention-feedback strategy improves the word recognition accuracy by 4.5%, 2.4%, 2.8% and 6.3%.

Event : Seminar
Title : Challenges in speaker verification with whispered speech
Speaker : Abinay Reddy Naini
Degree Registered :M.Tech(Res)
Advisor : Dr. Prasanta Kumar Ghosh
Date : 21/06/2019
Venue : C 241 MMCR, EE
Abstract : Whispering is an indispensable form of communication that emerges in private conversations as well as in pathological situations. In conditions like laryngectomy and neurogenic disorders such as Vocal chord Paresis and Spasmodic Dysphonia one or both of the vocal cords are affected, leading to the patient’s voice becoming breathy and rough whisper. A typical speaker verification system, where neutral speech is used for enrolling the speakers when tested with whispered speech often degrades the performance of speaker verification systems due to the difference in acoustic characteristics of whispered and neutral speech. So I will be presenting about different features and feature mapping techniques on whispered speech, to handle the performance degradation.
Speaker Bio : Abinay is currently pursuing his M.Tech(Res) with Dr. Prasanta Kumar Ghosh, in SPIRE Lab, EE, IISc. He is working on challenges to speaker verification with varying vocal effort. Previously, he worked as a Developer in Polaris,Chennai. He received his B.Tech in EE from National Institute of Technology, Warangal in 2015.

Event : Thesis Defence
Title : Hindi Online Handwritten Character Recognition
Speaker : Mr. Anand Sharma
Degree Registered : PhD
Advisor : Prof. A G Ramakrishnan
Date : 20/06/2019
Venue : C 241 MMCR, EE
Abstract : The thesis deals with the recognition of isolated Devanagari characters written online with a stylus. The work has three major contributions: 1. Proposing and effective extraction of sub-units of characters. A sub-unit is a sub-stroke of a character such that all the points in it satisfy a common geometric property. It is shown that Hindi ideal online character can be uniquely represented in terms of sub-units. A method of extraction of sub-units from actual Hindi handwritten characters is developed such that the extracted sub-units are similar to the sub-units in the corresponding ideal characters. 2. New features are developed that are independent of variations in the direction and order of strokes in the characters. These features, called HPOD features, spatially map co-ordinates, orientation of a stroke, and dynamics of orientation of the stroke at each point in a character. These features are used to represent a character at local sub-unit level and global character level in the sub-unit based classifier developed in this thesis. Accuracies of the traditional second order statistics (SOS), sub-space (SS), Fisher discriminant (FD), feedforward neural network (FNN), and support vector machines (SVM) classifiers increase when trained with HPOD features. 3. A sub-unit based (SUB) spatio-structural statistical classifier is developed that models handwritten characters in terms of the joint distribution of local HPOD features, global HPOD features and the number of sub-units. The classifier uses latent variables to model the structure of sub-units. The parameters of the model are estimated using the maximum likelihood method. The use of HPOD features and the assumption of independent generation of sub-units given the number of sub-units, make the classifier independent of variations in the direction and order of strokes in characters. The SUB classifier has the highest classification accuracy among the classifiers considered in this study.

Event : Seminar
Title : Physics-Based Vision and Learning
Speaker : Dr. Achuta Kadambi
Date : 14/06/2019
Venue : C 241 MMCR, EE
Abstract : Today, deep learning is the de facto approach to solving many computer vision problems. However, in adopting deep learning, one may overlook a subtlety: the physics of how light interacts with matter. By exploiting these previously overlooked subtleties, we will describe how we can rethink the longstanding problem of 3D reconstruction. Using the lessons learned from this prior work, we will then discuss the future symbiosis between physics and machine learning, and how this fusion can transform many application areas in imaging.
Speaker Bio :Achuta Kadambi is an Assistant Professor of Electrical and Computer Engineering at UCLA, where he directs the Visual Machines Group. The group blends the physics of light with artificial intelligence to give the gift of sight to robots. Achuta received his BS from UC Berkeley and his PhD from MIT, completing an interdepartmental doctorate between the MIT Media Lab and MIT EECS. Please see his group web page for research specifics: http://visual.ee.ucla.edu

Event : Thesis Defence
Title : Theoretical and Algorithmic Aspects of Rigid Registration
Speaker : Aditya Vikram Singh
Degree Registered :MTech (Research)
Advisor : Dr. Kunal Narayan Chaudhury
Date : 12/06/2019
Venue : C 241 MMCR, EE
Abstract : In this thesis we consider the rigid registration problem, which arises in applications such as sensor network localization, multiview registration, and protein structure determination. The abstract setup for this problem is as follows. We are given a collection of labelled points in d-dimensional Euclidean space. There are observers, each of whom assigns coordinates to a subset of points in their local reference frame. For each observer, we know which points they observe, and the (possibly noisy) local coordinates assigned to these points. Based on this information, we wish to infer the global coordinates of the points.
We investigate the following questions in this context:

Event : Seminar
Title : Comparison Between Different Notions of Stability for Laurent Systems
Speaker : Dr. Chirayu Athalye
Date : 07/06/2019
Venue : C 241 MMCR, EE
Abstract : Most of the systems in modern-day engineering applications are governed by either partial differential/difference equations (n-D systems) or delay-differential equations. Such systems can be modeled as infinite dimensional dynamical systems. A crucial question regarding the dynamical systems defined over an infinite dimensional state-space is that of stability. However, as the underlying state-space is infinite dimensional, generalization of results about the stability of finite dimensional systems is not straightforward and can be counter-intuitive.
In this talk, we examine a particular family of infinite dimensional discrete autonomous systems, which are governed by a Laurent polynomial matrix in the shift operator. We call this family of systems as Laurent systems. A Laurent system emerges in many interesting scenarios - namely, time-relevant discrete 2-D systems, the formation problem of infinite chains of agents, repetitive processes encountered in coal-cutting and metal-rolling industries, discrete quantum mechanics, etc. In this talk, we compare four different notions of stability for Laurent systems, and explain how some of the stability results are counter-intuitive when compared with the case of finite dimensional systems.
Speaker Bio :Chirayu obtained his M.Tech. and Ph.D. degrees from IIT Bombay. Currently he is an INSPIRE faculty at IISc. His research interests include infinite dimensional systems, n-D systems, stability analysis of dynamical systems, optimal control, and convex optimization.

Event : Seminar
Title : AllGoVision – Scalable DL based video analytics for Surveillance applications
Speaker : Mr. Ashwin
Date : 31/05/2019
Venue : C 241 MMCR, EE
Abstract : This talk covers AllGoVision which is an advanced video analytics solution with automation to improve operations, safety and security. AllGoVision has been a pioneer in the industry since 2009. AllGoVision comprises of 40+ basic and advanced video analytics features with varied application in Smart Cities, Smart Buildings, Smart Retail and Smart Traffic. AllGoVision software integrates with Visual sensors which are IP Surveillance cameras, gets the video , analyses it, detects events using AI algorithms and sends real time event. Effectively it is IOT software with the capability to process real time video data. With the use of Deep Learning and Artificial Intelligence based algorithms, AllGoVision has best in class accuracy and much lower false detections. With 200+ installations across 30+ countries, AllGoVision has very effective solutions for Security features for Perimeter Protection, Intelligent Traffic Solution including License plate recognition for Entry/Exit, Parking & Pathway, Business Intelligence features like Counting, Demographic analysis, People presence to improve operations and Facial Recognition for Entry/Exit and Indoor surveillance. This a scalable architecture which connects 10-5000 cameras. AllGoVision offers plug and play solution with your existing surveillance infrastructure. The software uses advanced deep learning concepts for development of technology which is briefly covered.
Speaker Bio : Ashwin started AllGoVision video analytics product group under AllGo Embedded Systems which grew into an independent company with business in 30+ countries. He is instrumental in making AllGoVision a well-known Video analytics brand with his technical abilities and business acumen. He brings to AllGoVision the rich experience of 20+ years and outstanding knowledge in Multimedia, Computer vision and machine learning. He has many publications and patents to his credit in the various field of Multimedia. Prior to AllGoVision, he served in NXP semiconductors in the role of an Architect. He was also associated with Motorola and was instrumental in developing Indian Language Speech Recognition solutions for mobiles. Ashwin holds a Master Degree from IISc Bangalore with a specialization in Signal Processing and System Science.

Event : Seminar
Title : On- Demand Dynamic First-mile Connectivity for Shared Multi-Modal Transportation
Speaker : Subhajit Goswami
Degree Registered :MTech (Research)
Advisor : Dr. Pavankumar Tallapragada
Date : 10/05/2019
Venue : C 241 MMCR, EE
Abstract : With the advent of networking, the area of transportation has attracted fresh interest. An interesting area is the coordinated Multi-modal transportation, which has the potential to effectively improve travel times and operational efficiency. In this talk, we will explore the problem of first-mile connectivity and in particular we consider a “one-shot” version with coordination given a demand. We will delve into details of a macroscopic model, which has certain advantages over more microscopic models. We describe an algorithm that reduces the complexity of the problem further and explore the limits on the profits that can be earned with such a system.
Speaker Bio :Subhajit Goswami received his B.E. degree from Jadavpur University, India, in 2017 and is currently pursuing his Masters in Technology by Research at the Dept. of Electrical Engineering at Indian Institute of Science Bangalore, India, from 2017 under Dr. Pavankumar Tallapragada. His interests lie in the Domain of Control Systems and Optimization with applications to Intelligent Transportation systems.

Event : Seminar
Title : Tightness of Semidefinite Relaxation
Speaker : Aditya Vikram Singh
Degree Registered :M.Tech (Research)
Advisor : Dr. Kunal Narayan Chaudhury
Date : 03/05/2019
Venue : C 241 MMCR, EE
Abstract : We consider a rank-constrained semidefinite program (REG-SDP), which arises in applications such as sensor network localization, multiview registration, and protein structure determination. The rank constraint makes REG-SDP nonconvex and computationally hard to solve. One way to make the problem computationally tractable is to drop the rank constraint, which yields a convex semidefinite program (called a convex ‘relaxation’ of REG-SDP). Empirically, it has been observed that, under certain conditions, global optimum of this convex relaxation is also a global optimum of REG-SDP (i.e., the convex relaxation is ‘tight’). In this talk, we will present an explanation of this tightness phenomenon using Lagrange duality.
Speaker Bio :Aditya is an M.Tech (Research) student in the Department of Electrical Engineering at IISc working under Prof. Kunal N. Chaudhury.

Event : Thesis Defence
Title : Algorithms for Processing RGBD Images and Videos for Depth-Based 3D Video Systems
Speaker : Suraj K
Degree Registered :PhD
Advisor : Dr. SOma Biswas
Date : 29/04/2019
Venue : C 241 MMCR, EE
Abstract : With increased availability of depth sensing cameras, the demand for depth-based 3D video systems is on the rise, which have been a natural choice for immersive media. This thesis addresses problems that are relevant at various stages of the depth-based 3D video system such as acquisition, representation, coding and display. We mainly address four distinct problems: Image-guided depth map upsampling, segmentation of RGBD images, salient object detection in RGBD images and virtual view synthesis for multiview-plus-depth videos.
The first two contributing chapters (Chapters 2 and 3) address the problem of depth map upsampling using a guidance image. Upsampling is performed to increase the resolution of the depth map and obtain per-pixel depth information. While the approach described in Chapter-2 doesn’t use any learning techniques, a deep learning based method is proposed in Chapter-3.
The next contributing chapter (Chapter-4) proposes an unsupervised algorithm to perform segmentation of a given RGBD image. The algorithm performs segmentation in a multi-stage manner and is based on hierarchical agglomerative clustering. Information extracted from the color image, albedo, depth map, surface normals, plane information and the edge maps obtained from both the color and depth images are utilized to perform clustering. This is followed by addressing the problem of detecting salient object in a given RGBD image (Chapter-5). The RGBD image is segmented first and scores are calculated for each segment. The superpixels belonging to the segment having the highest score are used as query to perform graph-based manifold ranking to obtain the final saliency map.
Finally, in Chapter-6, a fast yet effective algorithm to synthesize the virtual video from multiple synchronized RGBD videos is proposed. The input video frames are first 3D warped and then blended. A modified non-local means filtering based technique that uses both spatial and temporal information is proposed fill the disocclusion holes.

Event : Thesis Colloquium
Title : Analysis of whispered speech and its conversion to neutral speech
Speaker : Nisha Meenakshi
Advisor : Dr. Prasanta Kumar Ghosh
Degree Registered : PhD
Date : 25/04/2019
Venue : C 241 MMCR, EE
Abstract : Whispering is an indispensable form of communication that emerges in private conversations as well as in pathological situations. In conditions such as partial or total laryngectomy, spasmodic dysphonia etc, alaryngeal speech such as esophageal, tracheo-esophageal speech and hoarse whispered speech are common. Whispered speech is primarily characterized by the lack of vocal fold vibrations, and, hence, pitch. In recent times, applications such as voice activity detection, speaker identification and verification and speech recognition have been extended to whispered speech as well. Several efforts have also been undertaken to convert the less intelligible whispered speech into a more natural sounding neutral speech. Although supported by literature, research towards gaining a better understanding of whispered speech largely remains unexplored. Hence, the aim of the thesis is two-fold, 1) to analyze different characteristics of whispered speech using both speech and articulatory data, 2) to perform whispered speech to neutral speech conversion using the state-of-the-art modeling techniques.
In the first part of this thesis, we analyze whispered speech using both audio data (recorded via microphone) and articulatory data (recordings of movements of articulators, such as lips, tongue, jaw etc, using Electromagnetic Articulography synchronous with audio data). Specifically, we experimentally analyze how the pitch-less whispered speech encodes information such as speaker's gender and voicing, that are typically pitch-dependent in neutral speech. We find that whispered speech does retain speaker's gender and voicing related information. This could be attributed to the exaggerated movements of the articulators that typically occur while trying to maintain intelligibility in the absence of pitch. Therefore, we next investigate for the optimal transformation function that relates whispered articulatory movements with those of neutral speech. Experiments reveal that an affine transformation could relate the two articulatory movements better than other candidate functions considered. In addition, we also find how much the acoustics of whispered speech carries information about the corresponding articulatory movements compared to that of neutral speech.
In the second part, we design a feature that is necessary for segmenting whispered speech from a long recording of noisy whispered speech interleaved with silence/noise segments, as a pre-processing step in the conversion/reconstruction framework. In order to reconstruct neutral speech from whispers, we follow a voice conversion-based approach which requires an appropriate parametrization of the whispered speech spectrum. Such a parameterized representation of the spectrum is employed to build several whispered to neutral speech conversion systems. Among them, we find that the proposed bi-directional long short-term memory based whispered to neutral speech conversion system yields a perceptually more natural sounding speech compared to the state-of-the-art conversion systems.

Event : Seminar
Title : High-frequency Magnetics: Enabling Frequency Innovation in Passive Devices
Speaker : Dr. Ranajit Sai
Date : 16/04/2019
Venue : C 241 MMCR, EE
Abstract : From connected thermostats to self-driving vehicles, billions of connected devices will be transforming our lives in the era of IoT and 5G. This burst in device number demands more efficient power electronics. A (r)evolution in power electronics is on the horizon as the demand for smart power management systems, both in profile and performance, for all devices is skyrocketing. With the wide bandgap semiconductor technology maturing, power semiconductors are now driving a paradigm shift in power electronics – power-supply-on-board to power-supply-on-chip (PwrSoC). However, one of the biggest obstacles is the inductor – both for its size and performance. A proper design of the magnetics can solve some of the problems and push the technology further. On the other hand, hard-switching power converters are a known source of high-frequency harmonic noise that is strong enough even in the GHz frequency regime to affect the sensitive analog blocks of the RF front-end. As inter-component spacing is diminishing due to the increase in device density, electromagnetic interference among devices is on the rise. Therefore, a proper EMI/EMC design along with a proper understanding of radiated and conducted noise is essential – both from the perspective of noise source and noise victim. Here again, magnetics can be the savior – by absorbing radiated noise owing to the loss component of permeability, thereby addressing on-chip EMC/EMI issues. This talk will sketch the recent trends and pain-points of both inductor design for PwrSoC and electromagnetic noise handling in the age of IoT. In the first part of the talk, I’ll discuss the design and development of a ferrite-core on-chip inductor. I will demonstrate that it was necessary not only to delve into the crystal structure of ferrites to achieve desired core characteristics but also to develop a new deposition technique to get that ferrite deposited on-chip in a CMOS-compatible manner. In the second part of the talk, I will focus on the assessment of the impact of EM noise radiated from an EV-grade power converter to a nearby mobile communication system. I’ll also demonstrate the development of an EM noise suppressor and integration of the same as a critical component of a communication chip. Finally, in the third part of the talk, I’ll illustrate the immediate and long-term direction of my research, drawing upon my training in the field of magnetics and microfabrication, before concluding my talk with a teaching plan. Brief Biography: Dr. Ranajit Sai received his PhD in an interdisciplinary program – Nanoengineering for Integrated Systems from IISc in 2014, in which he designed and developed an on-chip ferrite-core inductor for which an US patent has been granted. After his graduation, Dr. Sai joined Tohoku University in Sendai, Japan as an assistant professor – elevated from his brief six-month stint as a postdoc there – and served in the Dept. of Electrical Engg. for three years and in an interdisciplinary centre, NICHe, for eight months. Since February 2018 he has been affiliated to CeNSE at IISc as a visiting professor. His area of research includes high frequency passive devices such as power and RF inductors, electromagnetic noise suppressors, skin-effect-free meta-conductors. Dr. Sai has more than 10 years’ experience in the design and development of ferrite-core on-chip inductors and demonstrated the first such inductor to be operated at up to 10 GHz with the highest inductance density ever reported. He is active also in assessing the impact of electromagnetic noise radiated from the EV-grade power converters on 5G bands and in providing solutions for the suppression of that radiated noise. He has more than 40 journal and conference publications, and one patent to his name.
Speaker Bio :Dr. Ranajit Sai is Visiting professor at CeNSE, IISc, Bengaluru.

Event : PhD Colloquium
Title : Modeling, Characterization, Control and Designof Switched Reluctance Machines
Speaker : Syed Shahjahan Ahmad
Advisor : Prof. G. Narayanan
Date : 16/04/2019
Venue : B 303, EE
Abstract : Switched reluctance machines (SRM) are permanent magnet free, and have a simple rotor construction with no current carrying parts. These are particularly suitable for high-temperature and high speed applications. However, modeling and control of SRM are challenging on account of phase inductance and back-emf being dependent on phase current and rotor position. This thesis addresses modeling for motoring and generation, characterization of SRM, current control for low speed operation, single pulse control for high speed motoring and generation, power converter for feeding SRM, and characterization of prospective magnetic materials for high speed SRM. The thesis also discusses design, fabrication and light load testing of two high-speed SRM prototypes:
Delta modulation and variable gain PI based current control are well known techniques for current control in an SRM. This thesis proposes and validates a fixed gain PI control with back-emf compensation for current control of SRM. A novel model predictive based current control is also proposed, which has better current tracking ability. Then a novel constant current injection based characterization method is proposed, which can yield the flux-linkage characteristics of the SRM without the requirement of blocking the rotor at known positions.
The thesis derives a mathematical model of SR generation (SRG) system, and utilizes this model to study voltage build-up during stand-alone operation of SRG system. A new high-speed optimal single pulse controller for SRG is also reported. Unlike the existing methods, the proposed real-time technique does not require any prior knowledge of the SRM characteristics or any off-line optimization procedure, and would be useful for self commissioning of SRM drives.
High-speed SRM requires high switching frequency power converter for effective control. Hence SiC devices based 50 kHz, 800 Vdc, 50 Arms power converter (asymmetric H-bridge) is developed, which is suitable for 20 kW 3-phase SRM. A fast fault detection and protection technique is part of the gate drive circuit of the above power converter.
Design and performance prediction of high-speed machines require knowledge of magnetic properties of materials over a wide range of frequency and excitation, which are often not available. A novel linear precision power amplifier (PPA) is developed for characterization of magnetic materials, which does not need a coupling transformer. This is a multi-stage, direct-coupled amplifier with low output offset, rated for 70 V peak, 10 A peak, DC-5 kHz. Using this PPA, the magnetic properties of numerous ferromagnetic alloys have been studied experimentally.
Finally, design and fabrication of two high-speed SRM prototypes, namely, (a) 10000 rpm, 5 kW, air cooled and (b) 40000 rpm, 10 kW, liquid cooled, are presented. No-load test results of the two prototypes are presented at different speeds. The results including phase current, rotor position and no load losses.

Event : Seminar
Title : Battery Charging Systems: Challenges and Topologies 9
Speaker : Mr. Manish Parmar
Date : 09/04/2019
Venue : C 241 MMCR, EE
Abstract : Battery plays a vital role in all the electronic devices and is becoming a necessary component in all modern handheld or wireless devices. As more and more complex devices start getting powered with batteries, the charging of these batteries gets increasingly challenging. When keeping the overall battery management solution safe and convenient to the user, different kinds of challenges start to drive the final solution. This talk will focus on key challenges and techniques for designing Battery Charging System and ICs. The presentation will focus on mobile devices based on lithium-based chemistry.
Speaker Bio :Manish Parmar is a Design & Development Manager for Battery Charging Products at Texas Instruments, India. He is responsible for designing and developing products for Battery Management Systems. Manish has joined Texas Instruments in 2004 and has worked through multiple Power management products ranging from DC-DC, Solar-modules, LED drivers, etc.

Event : Seminar
Title : Matrix commutator conditions for stability of switched linear systems
Speaker : Dr. Atreyee Kundu
Date : 05/04/2019
Venue : C 241 MMCR, EE
Abstract : Switched systems find wide applications in modelling and analysis of a large class of complex systems. Stability of switched systems has attracted considerable research attention over the past few decades. In this talk I will describe various research questions that arise in stability theory of switched systems, and solve some of them by employing commutation relations between the subsystem matrices.These techniques are relatively new in the literature and offer robustness with respect to smallperturbations in the elements of the subsystem matrices. I will end the talk with open questions involving matrix commutator based characterization of stability of switched systems.
Speaker Bio : Dr. Atreyee is an INSPIRE Faculty fellow at the EE dept.

Event : Seminar
Title : Research on Breath Sounds - from sensing to signal processing
Speaker : Dr. Prasanta Kumar Ghosh
Date : 29/03/2019
Venue : C 241 MMCR, EE
Abstract : Breath sound is often used as a biomarker for several pulmonary diseases. In this talk, I shall present an overview of the research on breath sounds. A summary of different normal and abnormal breath sounds will be presented. Advancements in sensing and processing the breath sounds will also be summarized. Finally, the talk will highlight the challenges that need to be addressed in the future research.
Speaker Bio :Prasanta Kumar Ghosh is an assistant professor in the department of EE, IISc.

Event : Seminar
Title : Finite Difference Time-domain method for lightning return-stroke simulation
Speaker : Rupam Pal
Advisor : Prof. Udaya Kumar
Date : 22/03/2019
Venue : C 241 MMCR, EE
Abstract : LIghtning is an important natural source of electromagnetic interference (EMI) for electrical and electronics systems. Among different phases of the lightning-strike, return-stroke phase assumes most importance in this regard. As the measurement of fields due to lightning return-stroke is an expensive procedure and is quite difficult, numerical modeling has been extensively used over the years. Finite difference time-domain method is a numerical tool which solves Maxwell's equations in time-domain and has become quite popular for solving different electro-dynamic problems due to its simplicity and flexibility. In this presentation, a brief introduction to the return-stroke phenomenon is to be discussed along with the basic principals of FDTD method. Also, application of FDTD method in lightning-related studies will be discussed.
Speaker Bio :The speaker is currently working for his PhD thesis in the Department of Electrical Engineering, Indian Institute of Science. He is working under the guidance of Prof. Uday Kumar. He has completed his ME from Department of Electrical Engineering, IISc, and BEE from the Department of Electrical Engineering, Jadavpur University. The research interest includes computational electrodynamics, , FDTD method, sub-station grounding, lightning return-stroke modeling.

Event : Seminar
Title : Understanding the loss surface of DNNs
Speaker : Ruturaj G. Gavaskar
Advisor : Dr. Kunal Narayan Chaudhury
Date : 15/03/2019
Venue : C 241 MMCR, EE
Abstract : The unprecedented success of deep neural networks (DNNs) has led to a revived interest in understanding the loss function used to train such networks. This is usually highly nonconvex due to the presence of nonlinear activations and the compositional nature of the network. In theory, this makes it difficult to compute the optimal network parameters. Nevertheless, the common wisdom is that the loss function is generally well-behaved and that optimization algorithms are often able to find the optimal parameters. In fact, Baldi and Hornik had conjectured back in 1989 that under realistic assumptions on the data, every local minimum is a global minimum for the squared loss. This was recently settled by Kawaguchi in a breakthrough work (NIPS '16). However, the proof technique is quite complicated and several unrealistic assumptions are used. At the core of the proof is a reduction to a so-called deep linear network (DLN) that has no activations and is simpler to analyze. Later, it was shown by Laurent and Brecht (ICML '18) that all local minima of a DLN can be shown to be global without any assumptions whatsoever on the data, provided the loss is convex and differentiable. We will discuss some of these findings in this talk. We will also give a simple proof of the result of Laurent and Brecht for scalar DLNs.
Speaker Bio :Ruturaj G. Gavaskar is a PhD student in the Department of Electrical Engineering, IISc. He completed his Masters degree in 2018 (IISc) and Bachelors in 2014 (BITS Pilani)

Event : Seminar
Title : Improving the signal to noise ratio of our brain and body
Speaker : Prof. A G Ramakrishnan
Date : 08/03/2019
Venue : C 241 MMCR, EE
Abstract : The talk will convince you that conscious, deep breathing is an amazing cardiac exercise. Your relationship with breathing will transform after this talk. In any field or situation, we call something a “signal”, if it is desirable or useful and “noise”, if it is undesirable or not useful. In the context of the brain and/or body, therefore, anything that improves the health, efficiency and/orp erformance can be considered “signal” and something that causes disease, decay or reduction in efficiency, “noise”. A very recent study has shown that breathing modulates the circulation of cerebrospinal fluid (CSF) in the brain. This is a finding with far-reaching implications. Improved circulation of CSF in the brain has the potential to better eliminate the metabolic waste products such as beta-amyloid from the brain, thus reducing the risk of Alzheimer’s disease. Boosting CSF has been linked to improvement in cognitive function and reduction of cognitive concerns in the elderly. Deep breathing also results in increased oxygen and reduced carbon-dioxide levels in the blood and consequently, reduced heart rate and pumping force of the heart due to the feedback is given by the chemoreceptors at the aorta and carotid artery. More on the brain's signal to noise ratio in the talk...
Speaker Bio :Prof. A G Ramakrishnan is a professor in the department of Electrical Engineering, IISc.

Event : Seminar
Title : Simplifying the Brain
Speaker : Prof. V. Srinivasa Chakravarthy
Date : 06/03/2019
Venue : C 241 MMCR, EE
Abstract : This talk will be based on the book published by Prof. Srinivasa Chakravarthy through Springer in 2019.
Speaker Bio :Prof. SRINIVASA CHAKRAVARTHY heads the Computational Neuroscience Laboratory in the Indian Institute of Technology, Madras. He obtained his B Tech in Electrical Engineering from IIT Madras, MS and Ph D from University of Texas at Austin and did his postdoc in Neuroscience, at the Baylor College of Medicine, Houston. His major research interests are in modeling basal ganglia, modeling motoneuron recruitment in skeletal muscle using oscillatory networks, developing a neuromuscular model of handwriting generation involving oscillatory neural networks, developing the consequences of our hypothesis that chaotic vasomotion is conducive to efficient oxygenation and the study of cardiac memory.
In addition, he has worked on developing handwritten character recognition systems in Hindi, Telugu, Kannada and Tamil.
His book, "Demystifying the brain" has been published by Springer this year and "Computational Neuroscience Models of the Basal Ganglia" was published last year. He has also translated any number of popular science books into Telugu.

Event : Seminar
Title : Bharati Script – A universal script for Indian languages
Speaker : Prof. V. Srinivasa Chakravarthy
Date : 06/03/2019
Venue : C 241 MMCR, EE
Abstract : We present Bharati, a simple, novel script that can represent the characters of a majority of contemporary Indic scripts. Indian language writing systems typically have a alphasyllabic structure, in which a character, known as an Akshara, is constituted by multiple consonants and a vowel. The vowels and consonants of different Indic scripts have a significant overlap, offering a natural possibility for unification. Bharati is a realization of this possibility. The shapes/motifs of Bharati characters are drawn from some of the simplest characters of existing Indic scripts. Bharati characters are designed such that they strictly reflect the underling phonetic organization, thereby attributing to the script qualities of simplicity, familiarity, ease of acquisition and use. The simplicity of Bharati script renders it ideal also for character recognition technology applications like Optical Character Recognition and Handwritten Character Recognition. For more information please visit: For more information please visit: www.bharatiscript.com

Event : Seminar
Title : Event-triggered control: When and what to communicate in networked control systems
Speaker : Dr. Pavan Tallapragada
Date : 01/03/2019
Venue : C 241 MMCR, EE
Abstract : Networked control systems (NCS) are systems in which the signals for feedback control are communicated over a communication channel or network. Such systems appear in a wide variety of important applications. However, NCS pose a unique set of challenges compared to classical control systems, such as that of limited communication resources. Thus, the questions of when and what to communicate in NCS are of a major concern. These questions have a natural relation to sampling and quantization. Time-triggered (periodic) sampling and uniform static quantization are still the de facto standards in practice. However, such sampling and quantization methods are often very conservative and indeed it is difficult to provide analytical guarantees for them except in very simple scenarios. Faced with this, in the last decade, a major theme in NCS has emerged - that of event-triggered control, wherein sampling and quantization are treated as online decision variables rather than offline parameters. This allows for efficient use of communication resources with the added benefit of provable guarantees for a wide array of problems. In this talk, I will first introduce NCS through applications and highlight some of the challenges. Then, I will present some of our important contributions in event-triggered control over the past several years.
Speaker Bio :Pavan obtained a PhD at the University of Maryland, College Park in 2013 and then did a Postdoc at the University of California, San Diego before he joined IISc as Assistant Professor in 2017.

Event : Seminar
Title : Localized Matrix Sketching with Applications to Active Array Imaging
Speaker : Rakshith Sharma
Date : 27/02/2019
Venue : C 241 MMCR, EE
Abstract: Active array imaging of far-field targets is a well studied problem with many applications. Imposing a range-limit on such targets offers an opportunity to carry-out the imaging process with far-fewer spatial array measurements compared to conventional imaging techniques, when broadband excitation is used. We present a novel trade-off between the number of array measurements and the bandwidth of excitation, leading to the possibility of very sparse arrays. We model the proposed trade-off as a matrix-sketching problem, a popular area of research in the numerical linear algebra community. Using such a model, we provide theoretical guarantees on the quality of image reconstruction from far fewer measurements compared to conventional imaging. We will also elaborate on matrix sketching and related problems in that area.
Speaker Bio : Rakshith Sharma is a fourth year PhD student in the ECE department at Georgia Institute of Technology, advised by Prof. Justin Romberg. His research interests lie in the areas of randomized linear algebra, high dimensional statistics, array signal processing and machine learning.

Event : Seminar
Title : High Speed Permanent Magnet Machines
Speaker : Brij B. Bhargava
Date : 21/02/2019
Venue : C 241 MMCR, EE
Abstract: In this talk Mr. Bhargava will talk about the history of high speed machines. Special materials that are used in the desing of high speed machines will be covered. Reliability and characterisics of the special materials and its impact on the machine design is highlighted. The implications on the availability of these materials in India will also be discussed. The system desing aspects of variable voltage, variable frequency, veriable speed, high power desity machines will be covered in the talk.
Speaker Bio : Mr. Brij B. Bhargava has worked in a number of companies reated to the design and development of electric machines. He was the Chief Engineer, Motors and Generators for Pacific Scientific, Danaher Corp. Brij has designed more than twenty rare earth PM engine-dedicated alternators for aircraft turbine engines for customers such as GE, Pratt and Whitney and Rolls Royce. He was part of the team for the development of Low Loss Electrical Steel for high efficiency machines, and initiated Direct Drive Starter/Generator Programs using the PM Design approach and designed and developed 85K to 145K rpm, 3 kW to 45 kW micro-turbine PM starter/generator for GM, Ford and Elliott. He started the Ashman Consulting and Ashman Technologies to bring this critical and rare high speed, high efficiency electrical machine technology for commercial and aerospace applications. Brij has over 35 years of experience designing PM motors, alternators, generators for aerospace and commercial applications.

Event : Thesis Colloquium
Title : Total Electric field due to a singel electron avalanche and its coupling to transmission line conductor
Speaker : Debasish Nath
Degree Regstered : PhD
Advisor : Prof. Udaya Kumar
Date : 27/02/2019
Venue : C 241 MMCR, EE
Abstract: Transmission of bulk electric power from the generating stations to the load centres can be carried out only through high voltages transmission lines. One of the main issues in the design and maintenance of extra and ultra-high voltage transmission system is the phenomenon named corona. It is the local electrical breakdown of air in the vicinity of the line conductors and hardware. Even though the design and dimensions of these elements are made considering the corona onset, surface abrasions arising either during installation or operation can lead to intolerable corona. Apart from producing some insignificant chemical reactions and noticeable acoustic noise, they can be a significant source of electromagnetic interference. In the early days, this interference was of concern only to radio and television receptions, however, with extensive use of wide frequency bands for modern applications, it has assumed prime importance.
The EMI due to the transmission line corona has been extensively studied and reliable empirical formulas have been proposed. The basis for all the earlier studies was the experimentally measured corona currents. This approach fails for new line designs especially with higher and higher voltages being employed due to non-availability of experimental data. A second approach assumed corona current to be injected into the line and subsequent analysis was carried out based on transmission line model. However, there were assumptions made on the mode of corona current injection into the conductor and the frequency range involved were also not adequate for the modern-day applications. Applicability of transmission line model for analysis is also questionable.
From a theoretical perspective, the coupling of the field produced by corona to the conductor was hardly investigated and the total field produced by the corona itself was not quantified. In order to address these serious lacunae, the present work was taken up and it can be considered as the first leap towards the correct picturization, as well as, quantification of the problem.
The field produced by the electron avalanche involves noticeable retardation effects. In the literature, only the field produced by arbitrarily moving point charge of fixed strength is given by the Heaviside-Feynman equation. On the contrary, the avalanche involves an arbitrarily moving charge of time varying strength at its head with trailing positive charge, which is almost stationary. Starting from the basics, an analytical expression for the electric field due to an arbitrarily moving point charge of time varying strength is derived which forms a fundamental contribution to Electrodynamics. This is extended to deduce an expression for the total electric field due to an avalanche for the very first time. Suitable validation of the expression is provided through numerical simulation of electric field integral equation.
Corona discharge is a complex phenomenon having many distinctly different modes which differ in their visual, as well as, electrical characteristics. Innumerable electron avalanches contribute to the measured corona current with their space-charge acting as a moderator. Therefore, in order to model the corona on conductors, an indirect approach based on linear system theory is proposed. An equivalent spatio-temporal dipole distribution was obtained to produce the measured current on the conductor. The general expression derived for the isolated avalanche is extended for this purpose.
Using the above, the means of induction, spatial decay rate of corona current in the close range, its propagation mode and field produced by both avalanche/equivalent dipole and that due to induced current in the conductor, have all been investigated and quantified.
In summary, the contributions made in this work are more of fundamental in nature and would be of significant interest to the high voltage power transmission line, as well as, to the communication engineers.

Event : Thesis Colloquium
Title : Multisource Subnetwork Level Transfer in Deep CNNs Using Bank of Weight Filters
Speaker : Suresh Kirthi K
Degree Regstered : PhD
Advisor : P S Sastry and KR Ramakrishnan
Date : 21/02/2019
Venue : C 241 MMCR, EE
Abstract: The convolutional neural networks (CNNs) have become important part of addressing several open problems in the areas of computer vision, speech, text and others. One concern about CNNs has always been their need for large amount of training data, large computational resources and long training time. In this regard the transfer learning promises to address this concern of CNN training through reuse of pretrained networks (CNNs). In this thesis we discuss transfer learning in CNNs where the transfer is from multiple source CNNs and done at subnetwork levels. The subnetwork multisource transfer is attempted for the first time hence we begin by showing the existence of such a transfer. We consider subnetworks at various granularities for the transfer. These granularities begin at a whole network-level then proceed to layer-level and further filter-level. In order to realize this kind of transfer we create a set called bank of weight filters (BWF) which is a repository of the pretrained subnetworks that are used as candidates for transfer. We also experiment to show the effectiveness of the subnetwork level transfer learning against training from the scratch while using fraction (eg., 10%, 30%, 50%, 70% and 80%) of the training data. In the second part of this work we show the usefulness of the filter-level multisource transfer for the cases of transfer from natural to non-natural (hand drawn sketches) image datasets, transfer across different CNN architectures having different number of layers, filter dimensions and others. In this part we also discuss the transfer from the CNNs trained on high-resolution images to the CNNs needed for the low-resolution images and vice-versa. In the multisource transfer the prelearnt weights are chosen from different CNNs and hence had to be finetuned. In this regard we feel it would be beneficial and efficient if the finetuning of transferred weights can be completely avoided. This is the theme in the third part of our work where we conceptualize filter-trees as the complete feature generation entity that can be used for transfer without finetuning. Similar to BWF we create a repository of pre-learnt filter-trees called bank of filter-trees to realize the transfer using filter-trees. In this part through experimental results we show that the transfer using BFT has the performance on par with the training from scratch, which is the best achievable performance. The selection methods for choosing the subnetworks from BWF or BFT so far was done uniformly randomly. For the sake of completion we introduce a method to do informed choice in the last part of our work. We propose a learnable auxilliary layer called choice layer whose learnt weights give an idea of importance of the subnetwork (filter-trees here) for the target task.

Event : Thesis Colloquium
Title : Fast and Robust Biomedical Image Reconstruction from Nonuniform Samples
Speaker : Bibin Francis
Degree Regstered : PhD
Advisor : Dr. Muthuvel Arigovindan
Date : 19/02/2019
Venue : B 303, EE
Abstract: we consider the problem of reconstructing images from non-uniformly under-sampled spatial point measurements with emphasize on the robustness to noise. The computational methods that deals with this problem are known as scattered data approximation (SDA) methods. Among these, well-performing methods achieve the reconstruction by minimizing a cost that is a weighted sum of data fidelity term measuring the accuracy of the fit at the measurement locations, and a regularization term. The latter term incorporates certain smoothness, and is constructed by summing the squared derivative values of a chosen order. The relative weight between these two terms is known as the smoothing parameter. Prominent methods in this category are known as radial basis function (RBF) methods, and they require solving large numerically ill-conditioned and/or dense linear system of equations. Subspace variational method alleviates the numerical instability and the computational complexity associated with the RBF methods. However, this approach involves solving large and sparse linear system of equation requiring specialized numerical methods. In the first part of the thesis, we propose a novel method for SDA that eliminates the need for solving dense linear system of equations, and even the need for storing matrix representing linear system. This is achieved by handling the reconstruction problem in two stages. In the first stage, the given non-uniform data is transformed into a pair of regular grid images, where, the one represents the measured samples and the other represents the sample density map. In the second stage, the required image is computed as the minimizer of a cost that is completely expressed in terms of regular grid discrete operations. It is expressed as a sum of weighted quadratic data fitting term involving the transformed image pair, and and discrete quadratic roughness functional. Computing the minimizer of this cost involves solving a well-conditioned sparse linear system of equations, where system matrix is represented in terms of filtering and array multiplications without the need for storing it explicitly. We demonstrate that the proposed method, which is named as regular grid weighted smoothing (RGWS), has much lower computational complexity than the RBF methods, with only a little compromise in the reconstruction quality. RGWS uses quadratic regularization, which is known to yield over-smoothed images under the presence of noise. We extend the RGWS method by incorporating non-quadratic regularization which is constructed by applying a square root on the sum of squares of derivative values (known as l1 regularization). We propose a reconstruction method using this l1 regularization, which we name as the l1-RGWS. We perform extensive set of reconstruction experiments with various levels of under-sampling and noise and compare the performances of l1 -RGWS and the original RGWS, which we also call l2 -RGWS. When the sampling density becomes low, the performance l1 -RGWS degrade abruptly and becomes worse than the l2 -RGWS. This behavior is known as the phase transition in the literature. We analyze this in a probabilistic viewpoint and infer that the prior probability model corresponding to l1 -regularization is based on the assumption that probability of a pixel location taking certain derivative value is independent of the derivative values of its neighboring pixel locations, which is clearly not true. We developed a probability model where error incurred by this independence assumption is compensated by means of a multi-resolution based re-weighting scheme. In this scheme, the desired reconstruction is formulated as a series of coarse-to-fine multi-resolution reconstructions, and re-weighting of the prior probability for each resolution level is derived from the reconstruction of previous resolution level. We demonstrate that the new method, which we name the multiresolution based scattered data approximation (MSDA), performs better than l1 -RGWS and l2 -RGWS under wide range of sampling densities, with slightly increased computational complexity. We then developed an extended method, where, instead of re-weighting the form of prior probability model corresponding to l1 regularization, the probability model itself is determined using maximum entropy principle. Specifically, at each resolution level in the multi-resolution reconstruction, the required probability model is determined as the maximizer of entropy subject to the information extracted from the lower resolution reconstruction as constraints. To further enhance the performance, we use directional second derivative operators to define the probability model. Moreover, to control the variance of this probability model, we also propose to use a modified multiresolution scheme, where the image sizes increase by a fractional factor, instead of doubling. We demonstrate that the new method, which we call the maximum entropy regularized reconstruction (MERR), outperforms both MSDA and l1 -RGWS for a wide range of sampling densities and noise levels.

Event : Seminar
Title : Power Electronics and Drives for Electrified Vehicles
Speaker : Srikanthan Sridharan
Date : 19/02/2019
Venue : C 241 MMCR, EE
Abstract: Demanding fuel-economy requirements and emission regulations have resulted in an increasing market for electrified vehicles (HEVs / EVs) across the world. To achieve faster market penetration of electrified vehicles, automotive industries are driving towards increased efficiency of the e-drive system with volume and cost reduction. This talk will outline and focus on system level loss minimization in the electric machine drive system to enhance drive efficiency for widely varying duty cycles. The total drive losses obtained from the proposed minimization will be compared with the existing strategies. Keeping in line with the loss minimization objective, a method to design an improved lossless and dynamic active damping controller will also be presented, to suppress undesired resonant oscillations in the motor voltages and currents, due to the use of LC filters. Following this, as an effort towards volume reduction, appropriate component sizing considerations in the HEV/EV electric drive system will also be discussed.
Speaker bio: Srikanthan Sridharan received his Bachelor’s degree in electrical and electronics engineering from the College of Engineering, Guindy, Anna University, Chennai, India, and the Master’s degree in electrical Engineering from the Indian Institute of Technology Madras, India and the doctoral degree in electrical engineering from the University of Illinois at Urbana–Champaign, USA. He is currently working as a Research Engineer in the Electric Machines and Drive Systems group of Ford Motor Company, MI, USA. His research interests include power electronics, machines and drives control and electric transportation He received the Best Paper award at the IEEE Transportation Electrification Conference and Expo in 2015. He also served as the Publications Chair for the IEEE Power and Energy Conference at Illinois (PECI) in 2015.

Event : Thesis Defence
Title : Stationary diesel exhaust treatment by blending discharge plasma/ ozone with industry wastes: a study on abatement of NOx and THC
Speaker : Apeksha Madhukar
Degree Registered: PhD
Advisor : B S Rajanikanth
Date : 27/02/2019
Venue : HV Seminar Hall, EE
Abstract: Increased usage of fossil fuels, especially diesel, has made a large impact on the environment in the form of rise in global temperature, increased acidity in the rain water, decreased yield in vegetation and numerous health related issues. Diesel has become a foremost and inevitable source of energy in day to day life, be it in stationary engines or in automobile engines. In the past three decades the usage of diesel has doubled up, particularly in third world countries like India, resulting in increased soot and gaseous particle emission. Of importance is the emission of oxides of nitrogen (NOx) and total hydrocarbons (THC), together accounting for 50% of NOx/THC emission. Though there exists an efficient system for controlling the solid soot particulate of diesel exhaust, the same for handling the gaseous pollutants is still at large. Therefore, any effort towards treating these gaseous pollutants efficiently and cost effectively is a welcome step before letting the same into the atmosphere.
The treatment involved in controlling these gaseous pollutants can be met at engine level (pre-combustion) or at the exhaust stream (aftertreatment/post-combustion). While the former technique has limited scope and been saturated with engine design modifications, the latter is currently being handled by catalyst/adsorbents. The latter scenario is more or less similar, be it engine exhaust or industrial exhausts, from the point of treating gaseous pollutants. However, the usage of catalysts/adsorbents has several drawbacks such as short life, high cost, storage, leakages and limited efficiency thereby motivating the researchers to look for alternate means of mitigating these gaseous pollutants. It is at this juncture, seeing the success of high voltage driven electric discharge-based precipitators, the thought of exploring the chemical potential of this electrical discharge plasma (also known as non-thermal plasma, NTP) came up almost three decades ago for controlling the gaseous pollutants at the downstream of the exhaust. Across the globe there was a spur in this NTP treatment of gaseous pollutants in a controlled condition and many successful reports came out at laboratory level. It was realized that NTP mainly results in oxidation of the pollutants due to the oxygen rich environment of the exhaust thus, necessitating usage of additional treatment involving catalysts/adsorbents. Since the exercise of introducing NTP is to provide an alternative to the commercially available expensive catalysts/adsorbents, the attention was shifted to utilize materials which are available abundantly and at a lesser price. The solid industrial waste is one such material satisfying this requirement and is being explored in the current thesis work.
In the current thesis work, gaseous pollutants from a stationary diesel engine exhaust were exposed to electrical discharge plasma shower in a carefully controlled laboratory condition. Oxides of nitrogen and total hydrocarbons are the two components that were studied amongst the gaseous pollutants. Since NTP is known for oxidation of the pollutants, in the current work, the exhaust was treated with discharge plasma/ozone injection and the oxidized pollutants were then adsorbed in pellets made out of solid industrial wastes such as fly ash, red mud, oyster shell, coffee husk, foundry sand and rice husk, the latter two being explored for the first time for their adsorption properties. The barrier plasma was either volume discharge type or surface discharge type during the study. The thesis then progresses through utilizing a novel way of treating the exhaust by cascading the barrier discharge plasma with ozone injection and vice-versa to enhance the overall oxidation of the gaseous pollutants be it NOx or THC.
It was observed that among the solid industry wastes studied, the red mud and foundry sand showed better NOx removal efficiencies compared to oyster shell, coffee husk and fly-ash, when cascaded with plasma treated exhaust. Further, foundry sand and red mud (as catalyst) performed equally well in controlling increased concentrations of NOx (associated with higher loading of the engine) in the post-plasma treated exhaust. Combined plasma+industrial waste-based adsorbents provide an efficient and economic option for NOx mitigation in diesel exhaust with appropriate scaling. Combined plasma+ozone-based technique provides a possible option for reduction/conversion of THC in diesel exhaust. This approach is first of its kind in the NTP fraternity. The results have been discussed at length in this thesis from the point of possible reaction pathways associated with conversion/reduction of NOx/THCs under plasma/O3 injection. The trapped NO2 in the adsorbents can be used as potential raw material for nitric acid/fertilizer industries. Through this research work another pathway for managing the ever-accumulating solid waste has been shown.
Speaker bio:

Event : Thesis Colloquium
Title : Hindi Online Handwritten Character Recognition
Speaker : Anand Sharma
Guide : Prof. A G Ramakrishnan
Date : 18/02/2019
Venue : C 241 MMCR, EE
Abstract: The thesis deals with the recognition of isolated Devanagari characters written online with a stylus. The work has three major contributions:
1. Proposing and effective extraction of sub-units of characters. A sub-unit is a sub-stroke of a character such that all the points in it satisfy a common geometric property. It is shown that Hindi ideal online character can be uniquely represented in terms of sub-units. A method of extraction of sub-units from actual Hindi handwritten characters is developed such that the extracted sub-units are similar to the sub-units in the corresponding ideal characters.
2. New features are developed that are independent of variations in the direction and order of strokes in the characters. These features, called HPOD features, spatially map co-ordinates, orientation of a stroke, and dynamics of orientation of the stroke at each point in a character. These features are used to represent a character at local sub-unit level and global character level in the sub-unit based classifier developed in this thesis. Accuracies of the traditional second order statistics (SOS), sub-space (SS), Fisher discriminant (FD), feedforward neural network (FNN), and support vector machines (SVM) classifiers increase when trained with HPOD features.
3. A sub-unit based (SUB) spatio-structural statistical classifier is developed that models handwritten characters in terms of the joint distribution of local HPOD features, global HPOD features and the number of sub-units. The classifier uses latent variables to model the structure of sub-units. The parameters of the model are estimated using the maximum likelihood method. The use of HPOD features and the assumption of independent generation of sub-units given the number of sub-units, make the classifier independent of variations in the direction and order of strokes in characters. The SUB classifier has the highest classification accuracy among the classifiers considered in this study.

Event : Seminar
Title : Overvie of Qualcomm Innovation Fellowship (QIF) India 2019 program
Speaker : Researchers from Qualcomm
Date : 14/02/2019
Venue : C 241 MMCR, EE
Abstract: The Qualcomm Innovation Fellowship (QIF) program invests in University PhD/masters students and their forward thinking ideas. The QIF program is focused on recognizing, rewarding, and mentoring innovative research students across a broad range of technical research areas, based on Qualcomm’s core values of innovation, execution and teamwork. QIF enables graduate students to be mentored by our engineers and supports them in their quest towards achieving their research goals. Details on the QIF India 2019 program can be found on the attached flyer and the following link, including details of the application process, proposal areas and team composition/eligibility: https://www.qualcomm.com/invention/research/university-relations/innovation-fellowship/2019-india . The call for proposals opens from Jan-15 and the submission deadline is Mar-1. This talk will provide an overview of Qualcomm Research and details on the QIF India 2019 program.

Event : Seminar
Title : Analysing auditory representation formation during language learning: an exploration with EEG
Speaker : Akshara Soman
Degree Registered: PhD
Advisor : Dr. Sriram Ganapathy
Date : 15/02/2015
Venue : C 241 MMCR, EE
Abstract: Humans are inherently capable of distinguishing between sounds from familiar and unfamiliar languages when they listen to human speech. But very little is known regarding the changes in the brain during a new language learning in terms of when these changes occur, and how they reflect in the learning. In my research, I am attempting to address some of these questions at the representation level using electroencephalogram (EEG) recordings. In the talk, we will discuss the use of a classification approach to discriminate the two languages from the EEG signal recorded during listening state, to quantify the representation level differences for the auditory stimuli of two different languages. This talk will further discuss the analysis of the EEG signals and the spoken audio signals to understand the underlying neural-behavioural correlates using a pronunciation rating technique and inter-trial distance based measures.
Speaker bio: Akshara Soman is a Ph.D. student at the Department of Electrical Engineering, Indian Institute of Science, working under the guidance of Dr. Sriram Ganapathy.

Event : Seminar
Title : Plasma-Assisted Combustion for Aerospace Applications
Speaker : Mr. Ravi B. Patel
Date : 01/02/2019
Venue : C 241 MMCR, EE
Abstract: Plasma-assisted combustion (PAC) technology has attracted much attention ie recent years as an innovative technique to improve the combustion efficiency. PAC has been studied to improve combustion performance of various engines used for aerospace propulsion like scramjet engines, gas turbine engines and new combustor technologies like MILD or lean burn combustion. Plasma can enhance the combustion process mainly through two modes - thermal and non-thermal (kinetic). In the last few years several efforts have been made to understand the kinetic enhancement mode by using non-thermal plasma produced using high voltage, short duration and high repetition rate pulses. Nanosecond pulse discharge has been demonstrated to maximize kinetic combustion enhancement due its effective energy loading into dissociative and ionization reactions. Radicals can be produced at such rate that fundamental properties of a fuel like flame speed, ignition threshold temperature, ignition delay and flammability limits can be improved which is not possible through thermal pathways.
In this talk, Fundamentals of plasma assisted combustion and its potential applications will be discussed. He will also narrate his M.Sc. thesis work studying the effect of non-thermal plasma on methane/air premixed combustion. Combustion parameters like flame speed and low-temperature ignition have been studied with respect to various parameters like applied voltage, pulse repetition rate (PRR), the the reduced electric field (E/N) and total energy input for plasma generation. Gas chromatography and optical emission spectroscopy techniques have been used for major species and reduced electric field measurements accordingly. Overall analysis suggests toward an s an optimum range of reduced electric field to maximize kinetic combustion enhancement which can be used in designing efficient plasma assisted burner for aerospace applications.
Speaker bio: Ravi B. Patel was born in Gujarat, India in 1994. He received his B.E. degree in Mechanical engineering from the L. D. College of Engineering in 2015. He completed his M.Sc (Engg) at the Department of Aerospace Engineering, Indian Institute of Science, Bangalore, India. His areas of interest are plasma assisted combustion, aerospace propulsion, non-thermal plasma and combustion spectroscopy and diagnostics.

Event : Thesis Colloquium
Title : Modelling, Optimisation and Control of Photovoltaic Energy Conversion Systems
Speaker : Pallavi Bharadwaj
Degree Registered: PhD
Advisor : Prof. Vinod John
Date : 30/01/2019
Venue : EE-B303
Abstract: Owing to its tropical position, India's annual solar energy input exceeds the combined capacity of all known fossil fuel reserves in the country. Still a large number of people live without electricity. This gap between the demand and the supply is due to the several challenges associated with the harnessing of the solar power. This work addresses four major challenges associated with the optimal utilisation of the solar energy resource. Measurement of input and output of PV systems is the first challenge which involves expensive solutions. The development of in-house, low-cost, high-performance irradiation meters, with performance standardised on the basis of ISO 9060 standard, facilitates input irradiation and temperature input measurement. For output measurement, a switched mode power conversion based closed-loop controlled PV characterisation setup is developed in this work which reduces the ripple in PV current measurement by at least 100% as compared to open loop response. Being a stochastic source of energy, it is difficult to predict its behavior but with the novel sequential parameter extraction method, this work presents a way of predicting the energy output of PV systems under varying ambient conditions with an improvement of 10% in the PV output prediction compared to the datasheet-based existing methods. The sequential optimisation solves the second challenge of modelling PV modules under steady state which is further enhanced for transient modelling by experimental evaluation of PV capacitance. The third major challenge in PV energy conversion is imposed by shading, which is addressed by the first time introduction of subcell model to partial shading analysis. Subcell model is experimentally verified to be 93% accurate for opaque shading and 95% accurate for translucent shading. This model further facilitates the understanding of hotspot formation and dust induced reliability issues. The fourth major problem is the global maximum power point tracking, to which the PV fraternity is still looking for a solution. This is addressed in this work using a fundamental shading fraction based GMPPT algorithm, wherein the shading versus global peak correlations are offline computed and stored in lookup tables for faster online tracking of global maximum. This algorithm is simple to implement, high in efficiency and can retrofit with existing PV converters.

Event : Seminar
Title : Protection and Safety on Power Systems with Highly Renewable Distributed Energy Resources
Speaker : Dr. Nirmal Kumar Nair
Date : 28/01/2019
Venue : HV Seminar Hall
Abstract: The United Nations Climate Change Paris Agreement (COP21) amongst 196 countries has triggered the rise in penetration of renewable distributed energy resources (DER) into AC interconnected power systems. Integration of different technologies and varying scales of renewable generation across interconnected transmission and distribution grids will accelerate resulting in pressure on ensuring safety and operational integrity of existing reliable power system operation. In this context one of the critical aspects impacted is the practice of power system protection and safety which this plenary will address. Existing knowledge of traditional safety and protection philosophies, schemes, practices and related system impacts will continue to be revisited to factor unique DER fault/abnormal characteristics and how they are electrically seen by the existing AC system and Intelligent Electronic Devices (IEDs) protection devices. This talk will highlight forward-looking and global transformative power system protection/safety concepts that helps understand challenges that highly distributed DER penetration causes and identify possible solutions using existing and newer technologies. Specifically, I will address the following items:

Event : Thesis Colloquium
Title : Attention-Feedback and Representations in OCR
Speaker : Shiva Kumar H R
Degree Registered: PhD
Advisor : Prof. A G Ramakrishnan
Date : 28/01/2019
Venue : C 241 MMCR, EE
Abstract: The thesis work has three major contributions: Design and development of an industry-grade OCR system for Kannada that performs better than Google’s Tesseract OCR on a challenging dataset of 241 images created by the candidate, and made publicly available with the ground truth. OCR’s are mainly required to digitize very old and legacy printed documents, which have many challenges such as old script, broken and merged characters and interspersed English words. The dataset has a good proportion of pages with all the above issues, and the OCR is still able to perform better than the already high baseline accuracy of >95% of Tesseract. Proposing the problem of segmentation of overlapping text lines from printed and handwritten documents as a computer science problem of representing each text line as a node in a red-black tree and assigning the connected components in the page image to different nodes. It also uses bipartite graph representation to assign corresponding line segments from different vertical partitions of the document page and eliminate line merges and splits. Rather than cutting the image into different line images, the algorithm builds (assembles) each line image by adding the connected components assigned to each particular node. Inspired by the rich feedback in the mammalian visual neural pathway, he has proposed improvements to image enhancement, binarization, segmentation and hence the recognition, using feedback from the latter modules such as SVM, Viterbi decoder and Unicode formation. This significantly improves the recognition performance of the OCR, in the presence of split and merged characters and also interspersed English words in the document page. On document pages of Tulu, Konkani and Sanskrit text printed in Kannada script, this attention-feedback strategy improves the word recognition accuracy by 2%, 2% and 6% over the already high values of 90 to 95%.

Event : APSIPA Distinguished Lecture
Title : Spoofing Attacks in Automatic Speaker Verification (ASV): Analysis and Countermeasures
Speaker : Prof. Hemant A. Patil
Date : 22/01/2019
Venue : C 241 MMCR, EE
Abstract: Speech is most natural way of communication between humans and it carries various levels of information, such as linguistic content, emotion, acoustic environment, language, speaker’s identity and health conditions, etc. Speaker recognition verifies or identifies a speaker via his/her voice. Automatic Speaker Verification (ASV) involves verifying the claimed speaker’s identity. In practice, we would like a speaker verification system to be robust against variations, such as microphone and transmission channel, intersession, acoustic noise, speaker ageing, etc. This robustness makes ASV system to be vulnerable to various spoofing attacks as it tries to nullify these effects and make spoofed speech more close to the natural speech. Hence, we would like the system to be secure against spoofing attacks. In this talk, difference issues concerning the robustness and security of a speaker verification system were discussed. We also discuss the latest progress and the research activities in anti-spoofing countermeasures against voice conversion (VC), speech synthesis (SS), replay, twins and professional mimics. In particular, brief details of risk and technological challenges associated with each of these attacks were discussed. The talk will also gave brief overview of three international challenge campaigns, namely, ASV Spoof 2015, ASV Spoof 2017 and ASV Spoof 2019 organized during INTERSPEECH 2015, INTERSPEECH 2017 and INTERSPEECH 2019, respectively. Finally, the talk concludes with overall summary of current state-of-the-art in this field and discusses future research directions.
Speaker bio: Hemant A. Patil received Ph.D. degree from the Indian Institute of Technology (IIT), Kharagpur, India, in July 2006. Since 2007, he has been a faculty member at DA-IICT Gandhinagar, India since 2007 and developed Speech Research Lab at DA-IICT recognized as ISCA speech labs. Dr. Patil is member of IEEE, IEEE Signal Processing Society, IEEE Circuits and Systems Society, International Speech Communication Association (ISCA), EURASIP and an affiliate member of IEEE SLTC. He is regular reviewer for ICASSP and INTERSPEECH, Speech Communication, Elsevier, Computer Speech and Language, Elsevier and Int. J. Speech Tech, Springer, Circuits, Systems and Signal Processing, Springer. He has published around 225 research publications in national and international conferences/journals/book chapters. He visited department of ECE, University of Minnesota, Minneapolis, USA (May-July, 2009) as short-term scholar. He has been associated (as PI) with three MeitY sponsored projects in ASR, TTS and QbE-STD. He was co-PI for DST sponsored project on India-Digital Heritage (IDH)-Hampi. His research interests include speech and speaker recognition, TTS, infant cry analysis. He has received DST Fast Track Award for Young Scientists for infant cry analysis. He has coedited a book on Forensic Speaker Recognition with Dr. Amy Neustein (EIC, IJST Springer). Presently, he is coediting two books in speech technology for medical-domain. Dr. Patil has taken a lead role in organizing several ISCA supported events, such as summer/winter schools/CEP workshops (such as speaker and language recognition, speech source modeling, text-to-speech synthesis, speech production-perception link, advances in speech processing) and progress review meetings for two MeitY consortia project all at DA-IICT Gandhingagar. Dr. Patil has supervised 04 doctoral (including doctoral thesis supervision in spoofing attacks) and 42 M.Tech. theses. Presently, he is supervising 03 doctoral students. Recently, he offered a joint tutorial with Prof. Haizhou Li during Asia Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) 2017 and INTERSPEECH 2018. In addition, he will offered a joint tutorial with Prof. H. Kawahara on the topic, “Voice Conversion: Challenges and Opportunities,” during APSIPA ASC 2018, Honolulu, USA. He has been selected as APSIPA Distinguished Lecturer (DL) for 2018-2019 and he delivered 15 APSIPA DLs in three countries, namely, India, China and Canada. Email: hemant_patil@daiict.ac.in, Google Scholar: https://scholar.google.co.in/citations?user=9kTiuPIAAAAJ

Event : Seminar
Title : Cross-Modal Matching
Speaker : Dr. Soma Biswas
Date : 25/01/2019
Venue : C 241 MMCR, EE
Abstract: Due to increase in the number of sources of data, research in cross-modal matching is becoming an increasingly important area of research. It has several applications like matching text with image, matching near infra-red images with visible images for night-time or low-light surveillance, matching sketch images with pictures for forensic applications, etc. This is an extremely challenging task due to significant differences between data from different modalities. In this talk, I will discuss about the different challenges of this problem and also some of the approaches we are working on for addressing this. In addition, I will also touch upon some related problems, like zero-shot learning, etc.
Speaker bio: Dr. Soma Biswas is an Assistant Professor in the Electrical Engineering department in IISc. She received her PhD degree in Electrical and Computer Engineering from University of Maryland, College Park, in 2009. Then she worked as a Research Assistant Professor at University of Notre Dame and as a Research Scientist at GE Research before joining IISc. Her research interests include image processing, computer vision, and pattern recognition.

Event : Seminar
Title : Interdependent Electric and Water Infrastructure Modeling, Optimization and Control
Speaker : Prof. Vijay Vital
Date : 17/01/2019
Venue : C 241 MMCR, EE
Abstract: The phrase water-energy nexus is commonly used to describe the inherent and critical interdependencies between the electric power system and the water delivery system (WDS). In this work, an analytical framework capturing the interactions between these two critical infrastructures is examined and a mathematical model to describe the associated dynamics is developed. Based on the time-scale of the associated dynamics, the electric network behavior is represented using time-series power flows following unit commitment and optimal power flow solutions. The WDS is represented using continuity equations at the delivery junction nodes and energy equations around the delivery loops and from tanks and reservoirs to the network. An integrated simulation engine of the interdependent infrastructure systems is formulated to conduct long-term simulations. Test cases have been conducted to show the impact of mega droughts and electric supply disruptions on the two interdependent systems.

Event : Seminar
Title : Power Switch ICs – What (are they), why (they are needed) & the challenges
Speaker : Subrato Roy
Date : 17/01/2019
Venue : C 241 MMCR, EE
Abstract: Power switches are quintessential in any system requiring power domain control to increase efficiency or load sequencing to ensure proper operation. They are also used for input source and output load protection against various system faults like output short, input surge, environmental conditions, etc. This will be illustrated with a deep dive inside a real world equipment to understand the value power switches add to the system. We will discuss different classes of power switches - from simple load switch to eFUSE providing a host of protection as well as telemetry.
We will then look at some of the innovations achieved in development of these devices which keep on pushing the boundaries on performance and cost. (1) In design meeting class A surge performance, thermal based power limiting. (2) In packaging achieving low footprint and low cost MCM (Multi Chip Module) device. (3) In test supporting 16 site parallel factory testing with low cost tester platform while guaranteeing <1% voltage measurement accuracy error for critical specs like current monitoring and limiting.
Speaker bio: Subrato Roy works as Design Manager in TI’s Analog Power Products Division. He has worked on power management circuits for embedded systems working on deep sub-micron CMOS processes. He currently leads development of eFUSE and Hotswap Controller ICs to deliver best in class devices in this segment. He has authored / co-authored 9 publications in TI’s internal technical conferences. He has 3 granted patents and another 5 filed at USPTO. He is a Member of Group Technical Staff (MGTS) of TI.

Event : Seminar
Title : Modulation of Inverter Fed Split-phase Induction Motor Drive
Speaker : Mr. Sayan Paul
Date : 18/01/2019
Venue : C 241 MMCR, EE
Abstract: Multi-phase induction machine is attractive for high power applications due to reduced power rating of individual phase-drive unit. Six-phase induction machine, one of the most common multi-phase machines, is of two types: Symmetrical and Asymmetrical six-phase. The later one is also known as split-phase induction machine (SPIM). SPIM has two sets of three-phase windings spatially shifted by 30 degree electrical. This winding arrangement makes SPIM advantageous over Symmetrical six-phase machine due to its less susceptibility towards stator excitation harmonics. In this presentation, modulation of inverter fed SPIM drive will be discussed.
The first part of this presentation describes the working principle, advantages and applications of multi-phase machines. Modeling and steady-state equivalent circuit of SPIM will be discussed in order to understand the appropriate voltages which should be applied for proper operation of SPIM. Two existing modulation techniques of inverter fed SPIM drive will be discussed. Subsequently, two novel modulation techniques will be discussed which are the original research contribution of this work. These proposed techniques enable us to achieve higher DC-bus utilization of inverter fed SPIM drive without creating pulsating torque and better phase current harmonic distortions compared to the existing modulation technique.
All the modulation techniques have been verified through simulations in Matlab Simulink and experiments performed on 5 kW hardware prototype.
Speaker bio: Sayan Paul received his B.tech. degree in Electrical Engineering from NIT Durgapur in 2014. He had worked in Haldia Petrochemicals Limited (HPL) for one year (2014-2015). From 2015 to 2018, he did M.Sc.(Engg.) from Electrical Engineering Dept. of Indian Institute of Science and is currently pursuing PhD in the same department. His current research interests include modulation of multi-phase machine drives.

Event : Seminar
Title : Development of Control Schemes to Enhance Stability and Dynamic Performance of Islanded Inverter-based Microgrids
Speaker : Dr. P E S N Raju
Date : 11/01/2019
Venue : C 241 MMCR, EE
Abstract: Renewable energy sources (RESs) such as solar, wind/micro wind and hydro/micro hydro-based power generations have gained considerable attention worldwide due to global warming, fast depletion of fossil fuels along with growing energy demand. Generally, power generation from these RESs is in the range of tens of kilowatts to the fraction of megawatts and due to this, these energy sources are usually connected at distribution level in order to reduce power losses in long transmission. Therefore these sources are called distributed generations (DGs). RESs based DG units produce fluctuating active power due to their intermittent nature. Further, the output of these DG units is either a DC or a variable frequency AC. Therefore, these DG units are interfaced to the distribution network or the local loads through a front-end inverter, named as inverter-interfaced distributed generation (IIDG). A recently evolved concept is to group a few of these IIDG units and a cluster of loads together to form a small local power system, called an inverter-based or AC microgrids (ACMGs).
AC microgrids (ACMGs) can be operated either in an island mode or in a grid connected mode of operation. Stability of ACMGs is not a critical issue in the grid-connected mode of operation as the stiff grid would be responsible for their stable operation. However, in the island mode of operation, it is an important concern due to the intermittent and low-inertial nature of IIDG units. The reason for this lies in the fact that in the islanded mode, IIDG units are responsible for maintaining the frequency and the voltage within their specified limits while sharing the load among the IIDG units in a stable manner. Therefore, the stability of islanded inverter-based microgrids (IIMGs) largely depends on the power-sharing control algorithm. A widely accepted droop-based power-sharing control approach has been used to share real and reactive powers among IIDG units. The low inertial nature of IIMGs makes them more vulnerable to instability even under a small change in operating conditions. Apart from this, there are several other factors which may further degrade the stability of these IIMGs. These factors are the interaction between generation and load dynamics, poor damping of low-frequency modes associated with the droop controllers. Thus, stability and dynamic performance enhancement of the IIMG is an important concern for its satisfactory and reliable operation.
The main focus of this talk is to discuss on the development of control schemes to enhance the stability and dynamic performance of IIMGs feeding multiple types of passive loads, rectifier interfaced active load (RIAL) and dynamic induction motor (IM) load, simultaneously. Various types of passive loads include resistive (R) load, inductive (RL) load and constant power load (CPL). In this thesis, a generalized model of IIMGs feeding R load, RL load, CPL, RIAL and dynamic IM load has been developed. The developed model has been used to investigate the impact of load dynamics as well as load sharing among IIDG units on the stability and dynamic performance of IIMGs. Based on these investigations, decentralized, centralized and two-level hierarchical controllers have been developed. The proposed controllers are designed based on a robust extended linear quadratic Gaussian (LQG) control, which combines the Kalman estimator with the linear quadratic regulator with prescribed degree-of-stability (LQRPDS). In order to obtain the optimal values of the diagonal weighting matrices of Kalman estimator and LQRPDS, a bi-objective optimization problem has been formulated and solved using a fast and elitist multiobjective non-dominated sorting genetic algorithm-II (NSGA-II). The proposed controllers produce supplementary control signals to the local power-sharing controller of each IIDG unit and DC voltage as well as AC current controllers of the RIAL. Eigenvalue analysis and time-domain simulations have been performed to validate the effectiveness of the proposed controllers. The proposed controllers provide robust control performance under various load configurations as well as small load disturbances. Further, to enhance the dynamic performance of IIMGs under large load disturbances, a control algorithm for the battery energy storage system (BESS) has been developed and evaluated for the load-leveling application.
Speaker bio: Dr. P E S N Raju is a Postdoctoral Research Associate in the School of Electrical and Electronic Engineering, The University of Manchester.

Event : Seminar
Title : Aggregate Reduced-order Models for Grid-tied Inverters
Speaker : Prof. Sairaj Dhople
Date : 07/01/2019
Venue : B 303, EE
Abstract: Rapid adoption of renewable sources of generation has increased the number of power electronics inverters installed on the ac power grid. Scalable models that present limited computational burden will be critical to model and analyze the collective dynamics of large numbers of inverters in next-generation power networks. This talk presents reduced-order aggregate dynamical models for grid-connected inverters. We place no restrictions on the converter topology and merely assume that the ac-side switch-averaged voltage can be controlled via pulse width modulation. The ac output of each inverter is connected to the ac grid through an LCL filter. The closed-loop system contains a phase locked loop for grid synchronization, and real- and reactive-power controllers realized with inner and outer PI current- and power-control loops. We derive a reduced-order state-space model for an arbitrary number of such paralleled inverters, and prove that it has the same model order and structure as any single inverter. This implies that the parallel collection can be modeled equivalently as one aggregate inverter unit that has the same physical and control structure-albeit with different parameters-as the individual inverters. We then extend this result to the setting where a distribution network interconnects the inverters. Applications are presented with regard to modeling the collective dynamics of photovoltaic- and wind-energy-conversion systems as well as stability analysis of mixed-machine-inverter systems.
Speaker bio: Sairaj Dhople received the B.S., M.S., and Ph.D. degrees in Electrical Engineering from the University of Illinois at Urbana-Champaign in 2007, 2009, and 2012, respectively. He is currently an Associate Professor at the Department of Electrical and Computer Engineering, University of Minnesota, where he is affiliated with the Power and Energy Systems research group. His research interests include modeling, analysis, and control of power electronics and power systems with a focus on renewable integration. He was the recipient of the National Science Foundation CAREER Award in 2015. He currently serves as an Associate Editor for the IEEE Transactions on Energy Conversion and the IEEE Transactions on Power Systems.

Event : Seminar
Title : Modeling and Control of Modular Multilevel Converter for Photovoltaic Application
Speaker : Mr. Anirudh Acharya
Date : 04/01/2019
Venue : C 241 MMCR, EE
Abstract: The trend of utility PV plant is clearly to move towards more distributed architecture. The Modular Multilevel Converter (MMC) is a suitable topology to upgrade of the PV plant with the energy storage or to scale the PV plant to higher power levels. The MMC is reliable, scalable and provides inherent modularity. The MMC is built using basic power electronic block referred as the Sub-Module (SM). The conversion of power from dc to ac is achieved without any intermediate power conversion stage similar to the classical central PV inverter. The MMC topology increases the number of MPPT by '6N' compared to two/three level inverter, where 'N' is the number of SMs connected in series. Different control strategies are investigated and a novel control approach in stationary frame is proposed which enables asymmetric power generation with balanced power injection to the ac grid. The proposed converter topology is investigated for uneven irradiance and a shade-tolerant variant of the topology using differential power processing converters is proposed.
The aim of the research project are:

Speaker bio: Anirudh Acharya received his MSc (Engg.) degree from Department of Electrical Engineering, Indian Institute of Science, Bangalore in 2011. He is currently pursuing his PhD degree at Department of Electric Power Engineering, NTNU, Norway. From 2011 to 2015, he was with ABB as Scientist working on specialized power converters for HVDC and FACTS. His main research interest is in design and control of converters in motor drives, renewable energy and storage systems. He is currently exploring the feasibility of using Modular Multilevel Converters for grid connected photovoltaic plant and control strategy for effective harvest of solar power.

Event : Thesis Colloquium
Title : Kernel-Based Image Filtering: Fast Algorithms and Applications
Speaker : Sanjay Ghosh
Degree Registered: PhD
Advisor : Dr Kunal Narayan Chaudhury
Date : 09/01/2019
Venue : C 241 MMCR, EE

Event : Seminar
Title : Deep Learning for recommender systems
Speaker : Dr. Anoop Deoras
Date : 17/12/2018
Venue : C 241 MMCR, EE
Abstract: In this talk we will survey latent models, starting with shallow and progressing towards deep, as applied to personalization and recommendations. After providing an overview of the Netflix recommender system, we will discuss research at the intersection of deep learning, natural language processing and recommender systems and how they relate to traditional collaborative filtering techniques. We will discuss techniques for embedding discrete user action events into continuous latent space for building a context aware collaborative filtering model for personalization and recommendations. Finally, we will highlight promising new directions in this space.
Speaker bio: Dr. Anoop Deoras is a Lead Researcher at Netflix where he leads the algorithmic innovation and productization of deep learning based recommender system models. He is interested in building the next generation of Machine Learning algorithms to drive the Netflix experience. Before that, he was a Lead Researcher at Microsoft, working on Cortana, an AI based virtual personal assistant for Windows OS. He holds a PhD from Johns Hopkins University where he proposed innovative algorithms for the first ever successful integration of Recurrent Neural Network based language models in Large Vocabulary Continuous Speech Recognition and Statistical Machine Translation.

Event : Seminar
Title : Fast Plug-and-Play Restoration
Speaker : Unni V S
Degree Registered: PhD
Advisor : Dr. Kunal Narayan Chaudhury
Date : 14/12/2018
Venue : C 241 MMCR, EE
Abstract: In plug-and-play image restoration, the regularization is performed using powerful denoisers such as nonlocal means (NLM) or BM3D. This is done within the framework of alternating direction method of multipliers (ADMM), where the regularization step is formally replaced by an off-the-shelf denoiser. Each plug-and-play iteration involves the inversion of the forward model followed by a denoising step. In this talk, we present a couple of ideas for improving the efficiency of the inversion and denoising steps. First, we propose to use linearized ADMM, which generally allows us to perform the inversion at a lower cost than standard ADMM. Moreover, we can easily incorporate hard constraints into the optimization framework as a result. Second, we develop a fast algorithm for doubly stochastic NLM, originally proposed by Sreehari et al. (IEEE TCI, 2016), which is about 80 times faster than brute-force computation. This particular denoiser can be expressed as the proximal map of a convex regularizer and, as a consequence, we can guarantee convergence for linearized plug-and-play ADMM. We demonstrate the effectiveness of our proposals for super-resolution and single-photon imaging.
Speaker bio: Unni V. S. is a PhD student at the Department of Electrical Engineering, Indian Institute of Science, working under the guidance of Dr. Kunal Narayan Chaudhury, EE Dept.

Event : Seminar
Title : Localised Solar Energy Solutions for Sustainability
Speaker : Prof. Chetan Singh Solanki
Date : 12/12/2018
Venue : C 241 MMCR, EE
Abstract: Today’s world is at the crossroads of a contradictory energy scenario wherein, on the one hand, energy access has to be provided to billions while, on the other hand, increasing demand and usage of energy is causing catastrophic climate change. As per IPCC report 2018, world is already hotter by nearly 1°C and that “limiting global warming to 1.5°C would require “rapid and far-reaching” transition in energy. In the context, Solar Urja through Localization for Sustainability (SoULS) initiative of IIT Bombay provides appropriate solutions for energy sustainability. The advancement in solar technology together with cost reduction, makes it possible to localise solar energy generation and consumption where local communities take care of assembly, supply, repairs and even manufacturing of solar products. An idea in line with Gandhian philosophy of "self-reliance" and "swaraj". The SoULS initiative has involved more than 5000 local women and provided first order solar energy solutions to more than 3 million students and households, spread across 22,000 villages in India.
Speaker bio: (click here for pdf brief bio) Dr. Chetan Singh Solanki, is a Professor at the Department of Energy Science and Engineering, IIT Bombay. He received his Ph.D. from IMEC (Ketholik University) Leuven, Belgium. He is currently leading two projects of national importance on the dissemination of affordable solar technology. The National Center for Photovoltaic Research and Education (NCPRE), is funded by the MNRE, Govt. of India. Prof. Solanki is also the Principal Investigator in the Solar Urja through Localization for Sustainability (SoULS) project at IIT Bombay. He also started kWatt Solutions Pvt. Ltd. Dr. Solanki has taken several initiatives at social front as well. He is the founder of Education Park, an initiative in school education. He founded ROSE, an organization for supporting education in rural India during his doctoral study in Belgium. Dr. Solanki’s SoULS project implemented in Rajasthan received the Prime Minister’s. He has won the European Material Research Society’s young scientist award in 2003 and IIT Bombay’s Young Investigator Award in 2009. He has published over 100 research papers in reputed international journals. He has 4 US patents to his credit. Prof. Solanki has authored 4 books on solar and renewable energy.

Event : Thesis Colloquium
Title : Algorithms for Processing RGBD Images and Videos for Depth-Based 3D Video Systems
Speaker : Suraj K
Degree Registered: PhD
Advisor : Dr. Soma Biswas
Date : 12/12/2018
Venue : C 241 MMCR, EE
Abstract: With increased availability of depth sensing cameras, the demand for depth-based 3D video systems is on the rise, which have been a natural choice for immersive media. This thesis addresses problems that are relevant at various stages of the depth-based 3D video system such as acquisition, representation, coding and display. We mainly address four distinct problems: Image-guided depth map upsampling, segmentation of RGBD images, salient object detection in RGBD images and virtual view synthesis for multiview-plus-depth videos.
The first two contributing chapters (Chapters 2 and 3) address the problem of depth map upsampling using a guidance image. Upsampling is performed to increase the resolution of the depth map and obtain per-pixel depth information. While the approach described in Chapter-2 doesn’t use any learning techniques, a deep learning based method is proposed in Chapter-3.
The next contributing chapter (Chapter-4) proposes an unsupervised algorithm to perform segmentation of a given RGBD image. The algorithm performs segmentation in a multi-stage manner and is based on hierarchical agglomerative clustering. Information extracted from the color image, albedo, depth map, surface normals, plane information and the edge maps obtained from both the color and depth images are utilized to perform clustering. This is followed by addressing the problem of detecting salient object in a given RGBD image (Chapter-5). The RGBD image is segmented first and scores are calculated for each segment. The superpixels belonging to the segment having the highest score are used as query to perform graph-based manifold ranking to obtain the final saliency map.
Finally, in Chapter-6, a fast yet effective algorithm to synthesize the virtual video from multiple synchronized RGBD videos is proposed. The input video frames are first 3D warped and then blended. A modified non-local means filtering based technique that uses both spatial and temporal information is proposed fill the disocclusion holes.

Event : Thesis Defence
Title : Power Electronic Technologies for Medium and High Power High Voltage Power Supplies
Speaker : Subhash Joshi T G
Degree Registered: PhD
Advisor : Prof. VInod John
Date : 07/12/2018
Venue : B 303 EE
Abstract: High Voltage Power Supplies (HVPS) are integral parts in many applications, whose performance depends on the high dc output voltage and a small ripple. The magnitude of ripple demanded by most of the HVPS applications is challengingly small. Vacuum tubes form common loads in high voltage applications. Often the fault event reported in these vacuum tubes is due to internal arcing. During this fault event, the vacuum tube will fail if it accumulated fault energy is above a specified limit. This is due to the high stored energy in HVPS. Hence, HVPS of low stored energy is attractive. The output voltage ripple and stored energy are inversely interdependent for a given operating frequency. This demands either protection for the vacuum tube during the fault events or small output voltage ripple without increasing the stored energy in the HVPS.
The topologies used in HVPS can be classified based on the power rating. For megawatt (MW) power level the preferred topology is mains frequency rectifier. Resonant topologies operating at higher frequency are preferred when the power level is in tens of kilowatt (kW). This thesis investigates a protection device for a vacuum tube that operates with MW power level HVPS, and design method to reduce the output voltage ripple for medium power HVPS.
A crowbar is an energy diverting device connected in parallel with the vacuum tube. It protects the tube during fault by providing an alternative path for the flow of energy. Conventional crowbars are built using either mercury or nitrogen gas based switches. Due to the environmental concern and higher operational cost, the state-of-art is to replace these devices with semiconductor devices such as thyristors. Crowbar built with semiconductor devices are referred to as solid state crowbar (SSC). This research, models and designs the subcomponents of a SSC, including: (i) Modelling of fault current, and a fuse wire that is used to emulate a vacuum tube during internal arc (ii) Design of inductor to limit di/dt in crowbar current (iii) Design of static and dynamic voltage balancing network for the thyristors (iv) Mechanical assembly design that ensures meeting the required crowbar electrical characteristics (v) Selection of cost-effective semiconductor device for crowbar application (vi) Thermal modelling of crowbar for pulse power applications (vii) Selection of cable for the pulse power application.
In a switched converter topology, the causes of output voltage ripple are: the switch action, the presence of input dc ripple, and variations in the load. In this thesis, the influence of input voltage ripple on the output dc voltage, called Audio Susceptibility (AS), is discussed. AS of load resonant converters has not been widely studied in literature. This research uses exact discretization method to obtain: (i) The analytical large signal and cyclic steady state model of the Series Resonant Converter (SRC) considering the resonant tank and output filter states (ii) The analytical small signal AS model of the SRC, and resonant gain condition for input ripple (iii) The design of an SRC for superior AS performance considering the selection of (a) the resonant components (b) the switching frequency (iv) A comparison of SRC design for (a) superior AS performance and (b) maximum power transfer capability (v) A selection of SRC components including the high voltage high frequency magnetics and selection of the MosFET semiconductor switching devices.
All the modeling and design method considered in this work has been verified by experimental studies on two 10MW, 10kV peak power SSC and a 10kW, 10kV SRC that has been fabricated as a part of the research.

Event : Seminar
Title : Computational Imaging with Few Photons, Electrons, or Ions
Speaker : Prof. Vivek Goyal
Date : 03/12/2018
Venue : C 241 MMCR, EE
Abstract: LIDAR systems use single-photon detectors to enable long-range reflectivity and depth imaging. By exploiting an inhomogeneous Poisson process observation model and the typical structure of natural scenes, first-photon imaging demonstrates the possibility of accurate LIDAR with only 1 detected photon per pixel, where half of the detections are due to (uninformative) ambient light. I will explain the simple ideas behind first-photon imaging. Then I will touch upon related subsequent works that mitigate the limitations of detector arrays, withstand 25-times more ambient light, allow for unknown ambient light levels, and capture multiple depths per pixel. The philosophy of modeling at the level of individual particles is also at the root of current work in focused ion beam microscopy.
Related paper DOIs:
10.1126/science.1246775
10.1109/TSP.2015.2453093
10.1109/LSP.2015.2475274
10.1364/OE.24.001873
10.1038/ncomms12046
10.1109/TSP.2017.2706028
10.1126/science.aat2298
Speaker bio: Vivek Goyal received the M.S. and Ph.D. degrees in electrical engineering from the University of California, Berkeley, where he received the Eliahu Jury Award for outstanding achievement in systems, communications, control, or signal processing. He was a Member of Technical Staff at Bell Laboratories, a Senior Research Engineer for Digital Fountain, and the Esther and Harold E. Edgerton Associate Professor of Electrical Engineering at MIT. He was an adviser to 3dim Tech, winner of the 2013 MIT $100K Entrepreneurship Competition Launch Contest Grand Prize, and consequently with Nest Labs 2014-2016. He is now an Associate Professor of Electrical and Computer Engineering at Boston University.
Dr. Goyal is a Fellow of the IEEE. He was awarded the 2002 IEEE Signal Processing Society (SPS) Magazine Award, the 2017 IEEE SPS Best Paper Award, an NSF CAREER Award, and the Best Paper Award at the 2014 IEEE International Conference on Image Processing. Work he supervised won student best paper awards at the IEEE Data Compression Conference in 2006 and 2011, the IEEE Sensor Array and Multichannel Signal Processing Workshop in 2012, and the IEEE International Conference on Imaging Processing in 2018 as well as five MIT thesis awards. He currently serves on the Editorial Board of Foundations and Trends and Signal Processing, the IEEE SPS Computational Imaging SIG, and the IEEE SPS Industry DSP TC. He previously served on the Scientific Advisory Board of the Banff International Research Station for Mathematical Innovation and Discovery, as Technical Program Committee Co-chair of Sampling Theory and Applications 2015, and as Conference Co-chair of the SPIE Wavelets and Sparsity conference series 2006-2016. He is a co-author of Foundations of Signal Processing (Cambridge University Press, 2014).

Event : Thesis Defence
Title : Visual Speech Recognition
Speaker : Abhilash Jain
Degree Registered: MSc (Engg)
Advisor : Dr. G N Rathna
Date : 05/12/2018
Venue : C 241 MMCR, EE
Abstract: Visual speech recognition (VSR), or automatic lip-reading, is the task of extracting speech information from visual input. The addition of visual speech has been shown to improve the performance of traditional audio speech recognition (ASR) systems, and hence has been active area of research since it’s inception. This thesis proposes a new VSR system for isolated word recognition tasks, with focus on the feature extraction methodology. A novel two-stage feature extraction technique is proposed in this thesis. Image transform based features -- discrete cosine transform (DCT) and local binary patterns (LBP) -- are used. The use of difference images for temporal feature extraction is also proposed. A new region of interest (ROI), which consists of the throat and lower jaw along with the mouth, is also introduced. For ROI extraction, the Viola-Jones algorithm is used. Classification is done using a multi-class Support Vector Machine (SVM) model. The system provides a simple, yet effective way to extract features from the video input, and performs comparably to some recent VSR systems, which employ more complicated techniques, like lip modelling or deep learning, to extract visual features.

Event : Seminar
Title : High-Performance Power System Dynamics (Transient Stability) Simulation
Speaker : Dr. Shrirang Abhyankar
Date : 07/12/2018
Venue : C 241 MMCR, EE
Abstract: This talk will present our research group¹s efforts on accelerating dynamics (transient stability) simulation of very large-scale power grids. Using scalable numerical solvers and adaptive time-stepping, our preliminary work shows that real-time simulation of large-scale power grids (tens of thousands of buses) is achievable. This talk will also cover basics on high-performance computing for power system analysis, history of HPC applications for power system analysis, and an introduction to the open-source numerical library 'Portable Extensible Toolkit for Scientific Computing¹ (PETSc) will be presented.
Speaker bio: Shrirang Abhyankar is a Power Systems Scientist in the Energy Systems Division at Argonne National Laboratory. In this role, Dr. Abhyankar several keys projects for the U.S. Department of Energy ranging from transmission-distribution co-simulation for assessing the impact of high PV penetration, cascading failure analysis, and stability-constrained optimization. Prior to this position, he was a post-doctoral appointee in the Mathematics and Computer Science Division at Argonne National Laboratory. He received M.S. and Ph.D. degrees from Illinois Institute of Technology in 2007 and 2011, respectively. His research interests include high-performance computing techniques for power system analysis, combined transmission-distribution simulation, and power system visualization. Since 2009, he is a developer of the high performance numerical library Portable Extensible Toolkit for Scientific Computing. (PETSc).

Event : Thesis Colloquium
Title : Theoretical and Algorithmic Aspects of Rigid Registration
Speaker : Aditya Vikram Singh
Degree Registered: MTech (Research)
Advisor : Dr. Kunal Narayan Chaudhury
Date : 06/12/2018
Venue : C 241 MMCR, EE
Abstract: In this thesis we consider the rigid registration problem, which arises in applications such as sensor network localization, multiview registration, and protein structure determination. The abstract setup for this problem is as follows. We are given a collection of N labelled points in d-dimensional Euclidean space. There are M observers, each of whom observes a subset of points and assigns coordinates to them in their local frame of reference. For each observer, we know which points they observe, and the (possibly noisy) local coordinates assigned to these points. Based on this information, we wish to infer the global coordinates of the N points. We investigate the following questions in this context.

Event : Thesis Defence
Title : Degradation Studies on Polymeric Insulators used for EHV and UHV Transmission
Speaker : Alok Ranjan Verma
Degree Registered: PhD
Advisor : Dr. Subba Reddy B
Date : 05/12/2018
Venue : High Voltage Seminar Hall, EE
Abstract: High voltage insulators used in overhead power transmission systems are of key importance for safe, reliable, and efficient operation in transferring huge amount of electrical power. Conventionally, ceramic/glass insulators were used in electrical power transmission, recently, in the country and elsewhere composite/ polymeric insulators are being used due to their promising advantages. These insulators are of recent origin and organic in nature, their material properties like surface electrical resistance and long-term performance are still under consideration by the international technical committees CIGRE, IEC, IEEE etc.
The present research work aimed at the development of newer experimental facilities and conducting investigations on the performance and analysis of polymeric insulators in service life, we have carried out the

Event : Thesis Colloquium
Title : Modeling, Analysis and Control of Reconfigurable Battery/Grid Tied Solar Photovoltaic Inverter
Speaker : Venkatramanan D
Degree Registered: PhD
Advisor : Prof. Vinod John
Date : 30/11/2018
Venue : C 241 MMCR, EE
Abstract: Grid reliability and power outages are key concerns today, due to the ever-increasing energy demand. Traditionally, Uninterruptible Power Supplies (UPS) with battery storage have been employed to contend with grid outages. For renewable power production, Solar-Photovoltaics (SPV) based Distributed Energy Resources (DERs) have been integrated with the grid using a power electronic Grid-Tied Inverter (GTI). A typical GTI by design engages in power conversion only when the grid is present, and ceases operation during a power outage to avoid a local unintentional island formation. Thus, solar energy is left unutilized by the GTI during a power outage, where the UPS steps in, to power critical loads. Recently, hybrid-PV or dual-mode inverter systems, that combine the complementary functional properties of UPS and GTI, have been in the focus of research due to their ability of standalone system operation during an outage while accessing solar power. Such a hybrid approach, although meets the desired operational objectives, requires the design, sizing, and control of the entire system, that comprises of multiple power converters, battery-banks, and SPV, to be carried out in a unified manner.
This work enhances the existing methods of solar energy access during a power outage, where the GTI is kept as an independent system from the UPS. A reconfigurable battery/grid tied inverter (RBGTI) scheme is proposed, that ties to the grid and functions as a regular DC-AC GTI when the grid is present. However, during a power outage, it reconnects to the battery-bank of an existing UPS present in a facility, where it functions as a DC-DC converter to provide PV based energy support. However, such an operation of RBGTI requires several questions to be resolved in terms of hardware configuration, islanding behavior, battery management, and overall system control, which are addressed in this work.
For the islanding behavior in grid-tied mode, a dynamic-phasor based GTI system model is proposed that captures the system dynamics accurately after unintentional islanding and allows systematic stability study based on eigenvalue analysis. In the battery-tied mode, a dynamic model of the PV fed battery charge-controller system is proposed which facilitates the systematic design of a maximum-power-point tracking (MPPT) controller and a load current tracking controller for the RBGTI, that achieves the effect of a virtual PV based battery-bank in parallel with the physical UPS battery. A supervisory RBGTI control scheme is proposed that ensures stable system operation during dynamic conditions of load power and solar insolation changes while reducing discharge burden on the UPS battery. A discrete IGBT converter hardware platform is developed, where the proposed analytical models, controls and the RBGTI performance are verified on a 4.5kW experimental setup.

Event : Seminar
Title : A New Perspective to Cross-modal Retrieval
Speaker : Mr. Titir Dutta
Degree Registered: PhD
Advisor : Dr. Soma Biswas
Date : 30/11/2018
Venue : C 241 MMCR, EE
Abstract: Cross-modal retrieval essentially deals with the problem of retrieving relevant (same class) information from a set of data (say, images, video etc.), given a query of some other media format (say, text documents, tags etc.). It has become an important research area in recent times due to the abundance of multi-media data. Quite a number of algorithms have been proposed in literature to address this problem, formulated as unsupervised, supervised or semi-supervised manner. However, all the algorithms perform well under the assumption that the query data encountered by the system is always from a class for which sufficient training samples are available. Recently, with the increase in popularity of research in the direction of unknown- class identification problem, the question arises as what happens when the query class is unknown to the system in case of cross-modal retrieval as well. We feel that it is the right time to take a step back and analyse the current status of the research in this area.
Our work focuses on the cross-modal retrieval problem under such a generalized scenario. Towards this goal, we have defined a generalized experimental protocol and established benchmark on multiple datasets. We have proposed a ranking- based algorithm to be used as an add-on with any other baseline cross-modal algorithm to improve the performance under such generalized cases as well as a semantic-aware latent-space learning method which outperforms the recent state- of-the-arts.
In this talk, we’ll present a description of this problem statement as well as our attempt to address it.
Speaker bio: Titir Dutta is a PhD student at the Department of Electrical Engineering, Indian Institute of Science, working under the guidance of Dr. Soma Biswas, EE Dept.

Event : Seminar
Title : Can Deep Learning Solve Inverse Problems?
Speaker : Dr. Angshul Majumdar
Date : 07/12/2018
Venue : C 241 MMCR, EE
Abstract: The vast majority of applications in deep learning arise in predictive analytics; this includes tasks like classification, regression, clustering etc. There are many talks, tutorials and courses on these topics. Therefore our talk does not focus on such practiced topics. We would concentrate on the niche area of inverse problems. A large number of problems ranging from denoising, deblurring to super-resolution, inverse half toning and reconstruction can be formulated as inverse problems. In the recent past, Compressive Sensing was the go to approach for addressing such problems. However, in recent times, more and more researchers are utilizing deep learning for solving them. This talk discusses some recent approaches based on CNNs, autoencoders, deep dictionary learning and domain adaptation for solution of generic inverse problems.
Speaker bio: Angshul Majumdar (S’07-M’12-SM’16) received his BE degree in electronics and communication engineering from the Bengal ultant in Technology Advisory. His current research interests include signal processing and Compressed Sensing Based Magnetic Resonance Imaging Reconstruction, published by Cambridge University Press in 2015 and Compressed Sensing for Engineers, published by CRC Press (in print expected December 2018). He has co-edited two book titled MRI: Physics, Reconstruction and Analysis (2015) and Deep Learning in Biometrics (2018), both published by CRC Press. His work has been funded by several national and international agencies including Department of Science and Technology (Govt. of India), CEFIPRA (Indo-French), IC-IMPACTS (Indo-Canada) and Air Force Research Labs (US). He is the founding chair of IEEE SPS Delhi chapter (2015-18). He currently serves as the chair of the IEEE SPS Chapter’s Committee (2016-18) and will be chairing the IEEE SPS E ducation committee from 2019. He has organized three IEEE SPS Sponsored Workshops and has been the Finance Chair for IEEE ISBA’17 (Flagship conference of Biometrics Council).

Event : Thesis Colloquium
Title : Emulation of Transients in a Long Transmission Line by Power Electronic Converter
Speaker : Sushmit Mazumdar
Degree Registered: MTech (Research)
Guide : Dr. Kaushik Basu
Date : 22/11/2018
Venue : C 241 MMCR, EE
Abstract: With increasing number of distributed power sources through power electronic converters in recent times, power quality and grid stability are becoming important day by day. But direct on-field tests of high power grid connected power converters are generally not possible. With the advent of Power-Hardware-in-the-Loop (PHiL) testing, we can get around this problem. PHiL requires real-time hardware emulator (HE), usually realized with a power amplifier (a power electronic converter) and a real-time digital simulator (RTDS). The available off-the shelf RTDS are expensive and capable of addressing a wide range application. One of the objective of this work is to develop an application specific stand-alone hardware emulator that will simulate a transmission line in real time and will be useful for the testing of high power grid tied power converters. A Silicon Carbide (SiC) based power electronic converter has been developed in the laboratory to implement the power amplifier of the transmission line emulator (TLE), which can emulate medium and long lines both during the transient as well as in steady state. To match the actual characteristics of a physical transmission line with the developed TLE, digital implementation of the distributed parameter line model has been done on a System-on-Chip (SoC) platform. A comprehensive analysis has been made to choose the switching frequency of the power electronic converter and the sampling frequency of the line model solver, while adhering to the power and digital hardware constraints (maximum switching frequency limit, clock speed, etc.). The stability aspects of the TLE with reference to the grid impedance has also been addressed. Finally, the relevant simulation and experimental results are matched which validates the developed TLE test bench setup.

Event : Seminar
Title : GaN and SiC Power Electronics Converters design and test using digital twin
Speaker : Prof. Stig Munk Nielsen
Date : 21/11/2018
Venue : C 241 MMCR, EE
Abstract: The semiconductor package laboratory at Aalborg University was founded in 2014 and inaugurated in 2017. Today we are a group of 13 power electronics engineers and material physicists researching on: (i) packaging issues regarding up to 10kV/15kV SiC semiconductor power modules and 650V GaN components (ii) design of SiC based power converters (iii) monitoring of IGBTs condition. In the presentation is focused on a few challenges using faster switching transistors as MOSFETS and the impact of medium voltage SiC devices. The new power module packaging facilities at Aalborg University and the converters based on medium voltage SiC MOSFETs are presented. The on-state voltage monitoring of IGBTs are implemented on commercial power stacks. The IGBT monitoring equipment are installed at one megawatt wind turbine test-site for failure mode investigations. Running for more than a year generating Thera bytes of data the analysis results of the direct power module condition monitoring are presented.
Speaker bio: Stig Munk-Nielsen (S’92–M’97) received the M.Sc. and Ph.D. degrees from Aalborg University, Aalborg, Denmark, in 1991 and 1997, respectively. He is currently Professor WSR at the Department of Energy Technology, Aalborg University. His research interests include LV and MV Si, SiC and GaN converters, packaging of power electronic devices, electrical monitoring apparatus for IGBTs, failure modes and device test systems. In the last ten years, he has been involved or has managed 10 research projects, including national and European Commission projects. Published 200 international power electronic papers being co-author or author.

Event : Thesis Colloquium
Title : Stationary diesel exhaust treatment by blending discharge plasma/ozone with industry wastes: a study on abatement of NOx and THC
Speaker : ApekshaMadhukar
Degree Registered: PhD
Advisor : Prof.B.S. Rajanikanth
Date : 28/11/2018
Venue : HVE Seminar Hall, EE
Abstract: Increased usage of fossil fuels, especially diesel, has made a large impact on the environment in the form of rise in global temperature, increased acidity in the rain water, decreased yield in vegetation and numerous health related issues. Diesel has become a foremost and inevitable source of energy in day to day life, be it in stationary engines or in automobile engines. In the past three decades the usage of diesel has doubled up, particularly in third world countries like India, resulting in increased soot and gaseous particle emission. Of importance is the emission of oxides of nitrogen (NOx) and total hydrocarbons (THC), together accounting for 50% of NOx/THC emission. Though there exists an efficient system for controlling the solid soot particulate of diesel exhaust, the same for handling the gaseous pollutants is still at large. Therefore, any effort towards treating these gaseous pollutants efficiently and cost effectively is a welcome step before letting the same into the atmosphere.
The treatment involved in controlling these gaseous pollutants can be met at engine level (pre-combustion) or at the exhaust stream (aftertreatment/post-combustion). While the former technique has limited scope and been saturated with engine design modifications, the latter is currently being handled by catalyst/adsorbents. The latter scenario is more or less similar, be it engine exhaust or industrial exhausts, from the point of treating gaseous pollutants. However, the usage of catalysts/adsorbents has several drawbacks such as short life, high cost, storage, leakages and limited efficiency thereby motivating the researchers to look for alternate means of mitigating these gaseous pollutants. It is at this juncture, seeing the success of high voltage driven electric discharge-based precipitators, the thought of exploring the chemical potential of this electrical discharge plasma (also known as non-thermal plasma, NTP) came up almost three decades ago for controlling the gaseous pollutants at the downstream of the exhaust. Across the globe there was a spur in this NTP treatment of gaseous pollutants in a controlled condition and many successful reports came out at laboratory level. It was realized that NTP mainly results in oxidation of the pollutants due to the oxygen rich environment of the exhaust thus, necessitating usage of additional treatment involving catalysts/adsorbents. Since the exercise of introducing NTP is to provide an alternative to the commercially available expensive catalysts/adsorbents, the attention was shifted to utilize materials which are available abundantly and at a lesser price. The solid industrial waste is one such material satisfying this requirement and is being explored in the current thesis work.
In the current thesis work, gaseous pollutants from a stationary diesel engine exhaust were exposed to electrical discharge plasma shower in a carefully controlled laboratory condition. Oxides of nitrogen and total hydrocarbons are the two components that were studied amongst the gaseous pollutants. Since NTP is known for oxidation of the pollutants, in the current work, the exhaust was treated with discharge plasma/ozone injection and the oxidized pollutants were then adsorbed in pellets made out of solid industrial wastes such as fly ash, red mud, oyster shell, coffee husk, foundry sand and rice husk, the latter two being explored for the first time for their adsorption properties. The barrier plasma was either volume discharge type or surface discharge type during the study. The thesis then progresses through utilizing a novel way of treating the exhaust by cascading the barrier discharge plasma with ozone injection and vice-versa to enhance the overall oxidation of the gaseous pollutants be it NOx or THC.
It was observed that among the solid industry wastes studied, the red mud and foundry sand showed better NOx removal efficiencies compared to oyster shell, coffee husk and fly-ash, when cascaded with plasma treated exhaust. Further, foundry sand and red mud (as catalyst) performed equally well in controlling increased concentrations of NOx (associated with higher loading of the engine) in the post-plasma treated exhaust. Combined plasma+industrial waste-based adsorbents provide an efficient and economic option for NOx mitigation in diesel exhaust with appropriate scaling. Combined plasma+ozone-based technique provides a possible option for reduction/conversion of THC in diesel exhaust. This approach is first of its kind in the NTP fraternity. The results have been discussed at length in this thesis from the point of possible reaction pathways associated with conversion/reduction of NOx/THCs under plasma/O3 injection. The trapped NO2 in the adsorbents can be used as potential raw material for nitric acid/fertilizer industries. Through this research work another pathway for managing the ever-accumulating solid waste has been shown.

Event : Thesis Defense
Title : Target Detection and Tracking under Non-ideal Conditions in Airborne Radars:
Speaker : Narasimhan R S
Degree Registered: PhD
Advisor : Prof. P S Sastry
Date : 26/11/2018
Venue : C 241 MMCR, EE
Abstract: Signal processing for target detection, parameter estimation and tracking in airborne radars is a challenge given the complexity of the operational environment. In this thesis we investigate the problems posed by non-ideal operational conditions for radar signal detection and tracking and propose novel, efficient and computationally light solutions for realization of robust signal processing techniques. This work aims to identify the lacunae of the current techniques for Pulse Doppler radar processing and proposes to improvise them to achieve practical and readily deployable solutions. Towards this we develop novel algorithms for adaptive threshold detectors for non-homogeneous background, suppression of ground clutter returns emanating from antenna sidelobes, target parameter estimation for Medium Pulse Repetition Frequency (MPRF) waveforms and suppression of spurious plots and tracking under dense clutter background.
Firstly, we bring out the non-idealities of the interference background, such as multiple interfering targets and clutter edge and discuss the design of adaptive threshold detector for such situations. In this context we propose censored cell averaging CFAR, switching censored cell averaging and greater of CFAR and robust variability index CFAR detectors. In the next problem, we focus on clustering the detected range-Doppler cells of a CFAR image applying connected component analysis technique and a valley detection logic to improve radar resolution. Suppression of clutter leaks emanating from antenna sidelobes is the next important aspect analyzed in the thesis. Here, we bring out the limitations of conventional sidelobe blanking and propose a novel sidelobe blanking technique based on quadrant subarrays of the main antenna. In our next study, we propose an efficient algorithm for range-Doppler unfolding as airborne radars employ medium pulse repetition frequency waveforms and measured range and Doppler is simultaneously ambiguous. The approach uses novel clustering technique. We conclude our work with the design of range rate tracking filter to simultaneously achieve conflicting requirements of low lag and low variance and use range rate information for efficient maneuver tracking under dense clutter background.

Event : Thesis Defense
Title : Binaural source localization using subband reliability and interaural time difference patterns
Speaker : Girija Ramesan Karthik
Degree Registered: MSc (Engg)
Advisor : Dr. Prasanta Kumar Ghosh
Date : 28 11/2018
Venue : C 241 MMCR, EE
Abstract: Binaural source localization using subband reliability and interaural time difference patterns Abstract: Machine localization of sound sources is necessary for a wide range of applications, including human-robot interaction, surveillance and hearing aids. Robot sound localization algorithms have been proposed using microphone arrays with varied number of microphones. Adding more microphones helps increase the localization performance as more spatial cues can be obtained based on the number and arrangement of the microphones. However, humans have an incredible ability to accurately localize and attend to target sound sources even in adverse noise conditions. The perceptual organization of sounds in complex auditory scenes is done using various cues that help us group/segregate sounds. Among these, two major spatial cues are the Interaural time difference (ITD) and Interaural level/intensity difference(ILD/IID). An algorithm inspired by binaural localization of humans would extract these features from the input signals. Popular algorithms, for binaural source localization, model the distributions of ITD & ILD in each frequency subband (typically in the range of 80Hz-5kHz for speech source) using Gaussian Mixture Models (GMMs) and perform likelihood integration across the time-frequency plane to estimate the direction of arrival (DoA) of the sources. In this thesis, we show that the localization performance of a GMM based scheme varies across subbands. We propose a weighted subband likelihood scheme in order to exploit the subband reliability for localization. The weights are computed by applying a non-linear warping function on subband reliabilities. Source localization results demonstrate that the proposed weighted scheme performs better than uniformly weighing all subbands. In particular, the best set of weights closely correspond to the case of selecting only the most reliable subband. We also propose a new binaural localization technique in which templates, that capture the direction-specific interaural time difference patterns, are used to localize sources. These templates are obtained using histograms of ITDs in each subband. DoA is estimated using a template matching scheme, which is experimentally found to perform better than the GMM based scheme. The concept of matching interaural time difference patterns is also extended to binaural localization of multiple speech sources.

Event : Thesis Defense
Title : Speech enhancement using deep mixture of experts
Speaker : Pavan Subhaschandra Karjol
Degree Registered: MSc (Engg)
Advisor : Dr. Prasanta Kumar Ghosh
Date : 28 11/2018
Venue : C 241 MMCR, EE
Abstract: Speech enhancement is at the heart of many applications such as speech communication, automatic speech recognition, hearing aids etc. In this work, we consider the speech enhancement under the framework of multiple deep neural network (DNN) system. DNNs have been extensively used in speech enhancement due to its ability to capture complex variations in the input data. As a natural extension, researchers have used variants of a network with multiple DNNs for speech enhancement. Input data could be clustered to train each DNN or train all the DNNs jointly without any clustering. In this work, we propose clustering methods for training multiple DNN systems and its variants for speech enhancement. One of the proposed works involves grouping phonemes into broad classes and training separate DNN for each class. Such an approach is found to perform better than single DNN based speech enhancement. However, it relies on phoneme information which may not be available for all corpora. Hence, we propose a hard expectation- maximization (EM) based task specific clustering method, which, automatically determines clusters without relying on the knowledge of speech units. The idea is to redistribute the data points among multiple DNNs such that it enables better speech enhancement. The experimental results show that the hard EM based clustering performs better than the single DNN based speech enhancement and provides similar results as that of the broad phoneme class based approach.

Event : Seminar
Title : Low-cost Phasor Measurement Unit Development
Speaker : Puneet Kumar
Degree Registered: M.Tech(Research)
Advisor : Dr. Gurunath Gurrala
Date : 09/11/2018
Venue : C 241 MMCR, EE
Abstract: Phasor Measurements Units (PMUs) are the key sensor technologies which got significant penetration in bulk power systems in the last decade. They have demonstrated significant value in improving state estimation, real-time wide area monitoring, control, and protection. However, most of the existing PMU devices have huge deployment cost. This talk primarily focuses on the development of a low-cost PMU on a credit card sized supercomputer (PARALLELA) and challenges faced in the implementation. At first, an overview of Phasor Measurement Unit will be given. Implementation details on FPGA of the PARALLELA will be presented. Two kinds of phasor measurement algorithms, Energy-based and PLL based, will be discussed. Implementation of a Teager Energy Operator (TEO) based PMU algorithm on FPGA and a comparison of some of the PLL methods in meeting the IEEE C37.118 standard will be presented.
Speaker bio: Puneet has obtained his B.Tech from Indian Institute of Technology, Roorkee in 2015. After that, he worked as Graduate Engineer Trainee in Jindal Steel and Power Limited from June 2015 to November 2015. After that, he worked as a project assistant under Dr. Gurunath Gurrala from April 2016 to June 2016. He then joined the Department of Electrical Engineering at Indian Institute of Science, Bangalore in August 2016 as an M.Tech(Research) student in Power Systems.

Event : Thesis Defence
Title : Modeling, Analysis and Control of Ultracapacitor based Bidirectional DC-DC Converter Systems
Speaker : Saichand K
Degree Registered: PhD
Advisor : Prof. Vinod John
Date : 08.11.2018
Venue : C 241 MMCR, EE
Abstract: With high penetration of Distributed Energy Resources (DERs), there is an increase in demand for Energy Storage Systems (ESS) to handle peak power demands. Energy storage systems also address the intermittency of the renewable sources. In this regard, ultracapacitor based energy storage systems are being increasingly preferred due to their high cycle life, high power density, and lower losses. However, ultracapacitor based storage systems have a relatively lower energy density, greater self-discharge and undergo wider terminal voltage variation in comparison to a battery based backup system. Therefore, ultracapacitor based backup systems are useful for addressing surge and peak power demands, and for improving the dynamic performance of the primary energy source. These systems are also used in stand-alone systems scenarios such as in load leveling and peak shaving of battery banks in electric/Hybrid electric vehicles. However, ultracapacitor based backup systems typically need an interfacing bidirectional dc-dc converter to handle the power flow as well as wide voltage variation of ultracapacitor stack. The dc-dc converter has two operating modes - charging and discharging modes.
This work focuses on the design and control of an ultracapacitor based bidirectional dc-dc converter for backup applications. In this regard, since the two control modes share the same power electronic converter, accurate mode identification and seam-less mode transition for stand-alone scenarios is an important challenge. This work proposes a switch control method based mode identification algorithm which assists in an accurate identification of control modes. A smooth and seamless transition between control modes is achieved using PWM blocking and an alternate virtual resistance based control which in turn is based on droop control used for dc-dc converter paralleling. This work also explores the possibility of utilizing the simplified static voltage source model typically used for battery banks for ultracapacitors as well. The work also considers an adaptive control based approach for handling wide input voltage variation and for alleviating Right Hand Plane (RHP) zero issue associated with it. All the control methods proposed in this work are validated on an experimental laboratory power converter prototype.

Event : Seminar
Title : Motion Averaging in 3D Vision
Speaker : Prof. Venu Madhav Govindu
Date : 16/11/2018
Venue : C 241 MMCR, EE
Abstract: Reconstructing the 3D world from multiple images is a long standing problem in computer vision. A key aspect of this problem is estimating the camera motions. We address this problem through a general framework of motion averaging that is applicable to both regular cameras as well as depth scanners. Our approach exploits the rich geometric structure of the problem (Lie group representations) and yields state-of-the-art algorithms for large-scale problems. In this talk I will present an elementary and accessible introduction to motion averaging and illustrate its application to problems of large-scale 3D reconstruction.
Speaker bio: Venu Madhav Govindu is with the Department of Electrical Engineering, Indian Institute of Science, Bengaluru. His primary area of research interest is geometry in computer vision.
Webpage: http://www.ee.iisc.ac.in/people/faculty/venu/index.html

Event : Seminar
Title : Consensus Equilibrium and Plug & Play Priors for Inverse Imaging Problems
Speaker : Dr. Suhas Sreehari
Date : 12/11/2018
Venue : C 241 MMCR, EE
Abstract: In this talk, we introduce the notion of Consensus Equilibrium (CE), which generalizes regularized inversion to include a much wider variety of both data fidelity components and regularization components without having to express these as cost functions. In this framework, the problem of MAP estimation in regularized inversion is replaced by the problem of solving these equilibrium equations. The key contribution of CE is to provide a new framework for fusing multiple sensor and image models. We introduce the formulation of the CE equations and prove that the solution of the CE equations generalizes the standard MAP estimate under appropriate circumstances. We also outline some algorithms for solving the CE equations, including a version of the Douglas-Rachford algorithm and Newton’s method (including a Jacobian-free form using Krylov subspaces). We then introduce Alternating Direction Method of Multipliers (ADMM) as a special case, segueing into Plug & Play (P&P) Priors. Finally, we demonstrate the strength and utility value of these methods in computational imaging.
*Work done while at Purdue University, West Lafayette, IN.
Speaker bio: Dr. Sreehari is a Quantitative Associate with Decision Science & Artificial Intelligence at Wells Fargo. He earned his PhD in electrical engineering from Purdue University – West Lafayette. His research interests lie in signal processing, distributed optimization, machine learning, quantitative modeling, estimation, and prediction for a variety of applications ranging from finance to computational imaging.

Event : Seminar
Title : Fast and Scalable Estimation of Uncertainty using Bayesian Deep Learning
Speaker : Dr. Emtiyaz Khan
Date : 02/11/2018
Venue : C 241 MMCR, EE
Abstract: Uncertainty estimation is essential to design robust and reliable systems, but this usually requires more effort to implement and execute compared to maximum-likelihood methods. In this talk, I will summarize some of our recent work that enables fast and scalable estimation of uncertainty using deep models, such as Bayesian neural network. The main feature of our method is that they are extremely easy to implement within existing deep-learning softwares. I will also summarize some of the current challenges faced by the Bayesian deep-learning community and how real-world applications can be useful for our research.
Joint work with Wu Lin (UBC), Didrik Nielsen (RIKEN), Voot Tangkaratt (RIKEN), Yarin Gal (UOxford), Akash Srivastva (UEdinburgh), Zuozhu Liu (SUTD).
Speaker bio: Dr. Emtiyaz Khan is a team leader (equivalent to Full Professor) at the RIKEN center for Advanced Intelligence Project (AIP) in Tokyo where he leads the Approximate Bayesian Inference (ABI) Team. Since April 2018, he is a visiting professor at the EE department in Tokyo University of Agriculture and Technology (TUAT) and also a part-time lecturer at Waseda University.

Event : Seminar
Title : Emulation of Switching Transients of a Transmission Line by High Frequency Power Electronic Converter
Speaker : Sushmit Mazumdar
Degree Registered: M.Tech(Research)
Advisor : Dr. Kaushik Basu
Date : 02/11/2018
Venue : C 241 MMCR, EE
Abstract: The flexible ac transmission line emulator (TLE) is the key element to bridge the gap between the emulator elements like generator and load for grid emulation purpose. A Silicon Carbide (SiC) based power electronic converter has been developed to implement the TLE, which can emulate both medium and long lines during the switching transients as well as in steady state. To match the actual characteristics of a physical transmission line with the developed TLE, digital implementation of the distributed parameter line model has been done on an SoC platform. A comprehensive analysis has been made to choose the switching frequency of the power electronic converter and the sampling frequency of the line model solver, while adhering to the power and digital hardware constraints (maximum switching frequency limit, clock speed, etc.). The stability aspects of the TLE with reference to the grid impedance has also been addressed. Finally the relevant simulation and experimental results are verified which validates the developed TLE.
Speaker bio: Sushmit has obtained B.E(Hons) from Indian Institute of Engineering Science and Technology, Shibpur in 2015. After that he worked as a Graduate Engineer Trainee in M.N Dastur & Company Ltd. from August 2015 to July 2016. He then joined the Electrical Engineering Department at Indian Institute of Science, Bangalore in August 2016 as a M.Tech(Research in Power Electronics) student.

Event : Thesis Defence
Title : Upper frequency bound on circuit based models for transformer windings
Speaker : Santosh J
Degree Registered: PhD
Advisor : Prof. Udaya Kumar
Date : 26/10/2018
Venue : C 241 MMCR, EE
Abstract: Power transformers are critical elements in any electric power system. Apart from their huge cost and long replacement time, the loss of revenue due to the outage of the transmission line can prove to be quite expensive. Therefore, all efforts are made from the design stage to their service in field to ensure their safety. According to a CIGRE report, close to 27 % of transformer failures are due to the winding insulation failures. Classically, power transformers, apart from the stress due to the operating voltage, are subjected to temporary overvoltages, switching and lightning surges. Due to their larger frequency content, which can range up to hundreds of kHz, winding behaves differently thereby augmenting the electrical stress at certain locations. Classically, the associated slow-wave phenomenon has been analyzed by either ladder network model or by distributed circuit models. With the introduction of the gas insulated substations, a new kind of internally generated surge called Very Fast Transient Overvoltages (VFTO) appeared, which stresses the winding differently. The associated frequency contents range from tens to a few hundreds of MHz. The chopped lightning surge, which is also frequent on the long lines, have frequency content in the range of tens of MHz. Further, a study on propagation of partial discharge induced currents in the winding involves frequencies up to hundreds of MHz. The classical electrical modes, which inherently assume quasi-static or transverse electromagnetic mode of propagation, are not really suitable for frequencies beyond couple of MHz. In view of the above, the present work is taken up, which aims to deduce the upper frequency limit for the applicability of the circuit-based approach. For the intended work, a fully field-based approach is essential. Firstly, a detailed investigation on the commonly employed numerical schemes for solution of the time-domain thin-wire formulation for the full-wave equation is made. It is shown that in spite of using suggested corrective measures, the marching-on-time schemes are not stable over long time spans. To overcome this, marching-on-degree scheme, suggested in a recent literature is chosen and efforts are made to minimize the computational burden. As the main concern in this investigation is the power transformer windings, detailed investigation on the upper frequency limit for the circuit-based models are carried out for pertinent ranges of geometrical dimension of single layer helical winding, disc winding and their combinations. Through detailed numerical simulations, empirical equations have been provided linking the upper frequency limit for the circuit-based models to the windings. Also, relevant observations are made on the series stress in the winding, as well as, the current distribution. The contribution made, while useful to the area of power transformers, is believed to have a deeper implication.

Event : Seminar
Title : What we did with the Riesz transform
Speaker : Prof. Chandra Sekhar Seelamantula
Date : 26/10/2018
Venue : C 241 MMCR, EE
Abstract: The Riesz transform is a two-dimensional extension of the Hilbert transform. In this talk, we shall examine some interesting properties of the transform and present applications to

Event : Thesis Defence
Title : Fast total variation minimizing image restoration under mixed
Poisson-Gaussian noise
Speaker : Manu Ghulyani
Degree Registered: MSc (Engg)
Advisor : Dr. Muthuvel Arigovindan
Date : 25/10/2018
Venue : C 241 MMCR, EE
Abstract: Fast total variation minimizing image restoration under mixed Poisson-Gaussian noise Image acquisition in many biomedical imaging modalities is corrupted by Poisson noise followed by additive Gaussian noise. Maximum Likelihood Estimation (MLE) based restoration methods that use the exact Likelihood function for this mixed model with non-quadratic regularization are very few. In particular, while it has been demonstrated that total variation (TV) based regularization methods give better results, such methods that use exact Poisson-Gaussian Likelihood are slow. Here, we propose an ADMM based fast algorithm for image restoration using exact Poisson-Gaussian Likelihood function and TV regularization. Specifically, we propose a novel variable splitting approach that enables isolating the complexity in the exact log-likelihood functional from the image blurring operation, allowing a fast Newton-like iteration on the log-likelihood functional. This leads to a significantly improved convergence rate of the overall ADMM iteration. We give sufficient conditions for convergence of this algorithm. We also propose Expectation-Minimization based iterations to further exploit the proposed splitting approach. The effectiveness of the proposed methods is demonstrated using restoration examples. Next, we extend this method for super-resolved image reconstruction for structured illumination microscopy (SIM). In SIM, extension of resolution beyond diffraction limit is achieved by illuminating the sample with a sinusoidal pattern. While known practical methods achieve reconstruction for SIM by modifying the measured data with sinusoidal modulation followed by a regularized multi-PSF deconvolution, our approach achieves reconstruction by means of TV penalized MLE with exact likelihood composed of raw measured data.

Event : Thesis Defence
Title : Robust Risk Minimization under label Noise
Speaker : Himanshu Kumar
Degree Registered: MSc(Engg)
Advisor : Prof. P.S. Sastry
Date : 23/10/2018
Venue : C 241 MMCR, EE

Event : Seminar
Title : Importance of Inscription Stones and the application of technology in their preservation
Speaker : Vinay Kumar and Udaya Kumar P L
Date : 12/10/2018
Venue : C 241 MMCR, EE
Abstract: Inscriptions stones (shila shaasanas) in the Bengaluru region are original documentation of the region’s people, culture, religion and language dating back to as early as 750CE. These stones give us a picture of the social fabric of the past including linguistic plurality amidst people, construction of lakes, tax practices, donations, grants, governance and suchlike. Rampant urbanization in Bengaluru has led to destruction of a majority of the 150 stones in the old ‘Bangalore’ region documented by B.L. Rice and others from 1894 to 1905 in the remarkable twelve-volume series Epigraphia Carnatica.
#InscriptionStonesOfBangalore is a civic activism project to raise awareness and protect ancient inscription stones found in the Bengaluru region. The project has been using technology (social media, mapping, 3D scanning, 3D printing, OCR) to protect preserve & restore the dignity of the last few remaining ‘Inscription Stones Of Bangalore’.
Speaker bio: Vinay’s interests range from Mars to Mohenjodaro. He has a master’s degree in Aerospace Engineering from University of Texas at Arlington. He is a patent engineer who was previously with the medical device research team at Novo Nordisk. He is also a recipient of the Govt. of India – Department of Biotechnology Foldscope research grant, to explore possibilities of using Foldscope as a research tool. He currently runs Sqvare Peg Labs, a non-profit with a mission to advance public understanding of science & technology. Udaya is a passionate Bangalorean and an accidental historian conservationist. He has a master’s degree in Engineering Mechanics from IIT Madras and has earlier worked in various capacities for the Tatas and General Electric. He currently heads the Software Delivery Centre, India at Schneider Electric, delivering industrial automation solutions to clients worldwide.

Event : Seminar
Title : Supervised I-vector Modeling: Theory and Applications
Speaker : Shreyas R
Degree Registered: Ph.D.
Advisor : Dr. Sriram Ganapathy
Date : 05/10/2018
Venue : C 241 MMCR, EE
Abstract: Over the last decade, the factor analysis based modeling of a variable length speech utterance into a fixed dimensional vector (termed as i-vector) has been prominently used for many tasks like speaker recognition, language recognition and even in speech recognition. The i-vector model is an unsupervised learning paradigm where the data is initially clustered using a Gaussian Mixture Universal Background Model (GMM-UBM). The adapted means of the Gaussian mixture components are dimensionality reduced using the Total Variability Matrix (TVM) where the latent variables are modeled with a single Gaussian distribution. In this talk, I will be giving an overview of i-vector modeling, and describe a supervised i-vector modeling framework where the speech utterances are associated with a label. I will also discuss a simple strategy of weighting the prior distribution of the latent variables to make the model more discriminative. Finally, I will discuss the applications of this model on language and accent recognition problems.
Speaker bio: Shreyas R is a Ph.D. student at the Department of Electrical Engineering, Indian Institute of Science, working under the guidance of Dr. Sriram Ganapathy

Event : Thesis Defence
Title : Sparsity Driven Solutions to Linear and Quadratic Inverse Problems
Speaker : Subhadip Mukherjee
Degree Registered: PhD
Advisor : Prof. Chandra Sekhar Seelamantula
Date : 26/09/2018
Venue : C 241 MMCR, EE
Abstract: The problem of signal reconstruction from inaccurate and possibly incomplete set of linear/non-linear measurements occurs in a variety of signal processing applications. In this thesis, we develop reconstruction algorithms that exploit signal sparsity in such settings. The assumption of sparsity is practically relevant, as most signals encountered in practice admit a sparse representation in an appropriately chosen bases. In particular, we consider the linear and quadratic measurement models, which give rise to the so-called sparse coding and phase retrieval problems, respectively.
The thesis consists of the following three parts:
1. Dictionary Learning: We first address the problem of dictionary learning when the training examples are corrupted by impulsive noise. The data error is measured in terms of the $\ell_p$ norm, $0 2. Sparse Coding: In the context of sparse coding, we first consider a setting where the linear measurements are revealed sequentially and develop an online reconstruction scheme. We demonstrate experimentally that our algorithm achieves a reconstruction performance that is on par with the batch mode reconstruction and requires progressively less computation as more measurements are acquired. Subsequently, we develop a feed-forward deep neural network (DNN) architecture motivated by the iterative shrinkage-thresholding algorithm (ISTA) and propose to model the nonlinear activation function of the network using a linear expansion of thresholds (LET), which has been shown to be successful in several image denoising and deconvolution problems. We show that such a parametrization is economical and induces a rich variety of sparsity promoting regularizers. The network architecture corresponding to the fast ISTA (FISTA) algorithm is also considered and is shown to be a residual network, which is easier to train using the gradient-descent algorithm. The proposed DNN architecture is extended to encompass the case where the dictionary is unknown, leading to the dictionary learning problem considered before. The underlying dictionary can be estimated by optimizing the weights and biases of the network such that the training examples are accurately reproduced at the output.
3. Phase Retrieval: We first consider the problem of phase retrieval (PR) from Fourier magnitude measurements subject to the sparsity constraint. We develop two algorithms that iteratively apply a combination of projections and reflections on to the magnitude and sparsity constraint sets. The proposed algorithms essentially generalize the classical Fienup's algorithm for PR. An error reduction property is established for the proposed algorithms, meaning that the error in the measurement domain remains non-increasing as the iterations progress. The reconstruction problem in frequency-domain optical coherence tomography is considered as an application, and it is demonstrated that the proposed algorithms lead to superior artifact-free reconstruction as compared with the state-of-the-art PR techniques. Subsequently, we generalize the framework to accommodate generic quadratic measurements, not necessarily Fourier intensity measurements. Further, the measurements are considered to be encoded with a finite precision to make the setting practically relevant. As an extreme case, we also consider quadratic measurements encoded using a binary alphabet. We develop a reconstruction algorithm that leverages the principles of lifting and consistent reconstruction and applies an accelerated projected-gradient descent algorithm to minimize the associated cost function. It is possible to accommodate the sparsity prior, if available, in the proposed algorithm. Experimental results demonstrate the superiority of the proposed algorithm to the state-of-the-art PR techniques that are not tailored to handle finite-precision measurements. In the presence of additive noise prior to encoding, the Cramer-Rao bound (CRB) for phase retrieval is derived and used for benchmarking. It is shown that the proposed algorithm yields reconstruction mean-squared errors (MSEs) that are within 2 to 3 dB of the corresponding CRB at all values of the input signal-to-noise ratio.
Speaker bio:

Event : Seminar
Title : Euler: His Life and Work
Speaker : Dr. Kaushik Basu
Date : 28/09/2018
Venue : C 241 MMCR, EE
Abstract: Euler is considered as one of the greatest mathematicians of all time. He made seminal discoveries in the long-established areas of mathematics like number theory, geometry, algebra, analysis. He ventured into uncharted territories of analytical number theory, graph theory etc. He was also a first rate applied mathematician and made significant contribution in mechanics, optics etc. Euler is widely considered as most prolific mathematician. He wrote mathematics faster than one can understand it. As an expositor Euler has no parallel. Euler was virtually blind in last decade of his life, but it had zero impact both on the quality and the quantity of his mathematics. His life is a history of the triumph of human spirit. In the first part of this talk we will take a close look at his life both as a man and his work. Next, we will survey few of his magnificent discoveries. In the last part of this talk we will see Euler in action- one of the famous classical number theory problem that he solved. The talk is aimed at general audience with some familiarity with college level mathematics.
Speaker bio: Kaushik Basu, is an assistant professor in the department of EE, IISc.

Event : Thesis Defence
Title : Denoising and Refinement Methods for 3D Reconstruction
Speaker : Sk. Mohammadul Haque
Degree Registered: PhD
Advisor : Prof. Venu Madhav Govindu
Date : 26/09/2018
Venue : C 241 MMCR, EE
Abstract: Capturing raw 3D data from the real world is the initial step for many 3D reconstruction pipelines in different computer vision applications. However, due to inaccuracies in measurement and oversimplification in mathematical assumptions during capture, 3D data remain contaminated with significant amounts of errors.
In this thesis, we investigate two different types of errors that are invariably encountered in 3D reconstruction. The first type of errors comprises the random measurement error or noise that is inevitably present in raw 3D data obtained from the real world. The second type of errors comprises those that are geometrically-structured. Specifically, we consider non-rigid alignment errors that arise in multi-view scenarios where complete 3D reconstructions of real-world objects are obtained from observations taken from multiple viewpoints.
Firstly, we consider random measurement errors, modelled as an additive 3D Gaussian noise. We consider the task of denoising 3D data obtained in two different modalities, i.e. 3D meshes and 3D point clouds and establish the important factors that dictate the quality of denoising. We develop denoising schemes that account for these factors and provide evidence of superior denoising performance on synthetic and real datasets over existing approaches.
Secondly, we consider non-rigid errors that are encountered in a multi-view 3D reconstruction pipeline. In particular, we address the problem of multi-view surface refinement for high quality 3D reconstruction, where low quality reconstructions obtained from consumer-grade depth cameras are enhanced using additional photometric information. We show that non-rigid estimation discrepancies that emerge in such tasks is a major issue limiting the quality of reconstruction. We systematically delineate the underlying factors and show that existing refinement methods in the literature do not consider these factors, hence, failing to carry out a proper refinement. Considering these factors, we develop a complete multi-view pipeline for high quality 3D reconstruction. We show the efficacy of our pipeline on synthetic and real datasets, as compared to other existing methods.
Speaker bio:

Event : Seminar
Title : Practical challenges in the design of power electronics devices for plasma cutting applications
Speaker : Dr. Girish Kamath
Date : 24/09/2018
Venue : C 241 MMCR, EE
Abstract: The speaker has a background in power electronics design and has been designing power supplies for plasma cutting applications for the past 17 years. Plasma cutting in a nutshell, is a popular technology that converts gas such as air or oxygen into the plasma state to cut metal. In this two-part talk, the speaker will begin with an overview of a typical plasma cutting system. The ignition circuit, which is a usually a Tesla coil based High-Voltage High- Frequency circuit is used to initiate the conversion of gas to plasma at the beginning of the cut cycle. This circuit arguably makes Plasma cutting unique among the list of industrial electronics applications. The speaker will present details of a typical gas ignition circuit and the role it plays in the metal cutting process. The application of this circuit brings with it certain practical challenges, mainly related to radiated emissions. This will be the subject matter for the second part of the talk. Here, the speaker will describe briefly the mechanics of radiated emissions due to its operation. It is seen that cables play an important role in EM noise radiation. To address this problem, a low-cost diagnostic tool to measure the antenna effects of cables is being developed. The proposed method is being formulated with a simple test cable case and verified with its 3D RF physics model. Details of the investigation along with experimental and modeling results will be presented. The talk will end with future scope of work in this area.
Speaker bio: Dr. Girish Kamath graduated with an MS degree in 1996 followed by a Ph.D. in Electrical Engineering from the University of Minnesota in 1998. He has been working in the Motor Drives and Plasma Cutting industries since then. He currently designs power electronics systems and controls for plasma cutting power supplies. His main areas of interest are multi-physics modelling approach to power supply component design, High Voltage circuits, Electromagnetic Compatibility and Digital Control.

Event : Seminar
Title : Standing out in speech: Prosodic prominence in speech production and perception
Speaker : Prof. Jennifer Cole
Date : 12/09/2018
Venue : C 241 MMCR, EE
Abstract: Across languages, variation in the prosodic form of utterances relates to information structure (IS)- words that convey new information or focus are distinguished from words that are discourse-given through the phonological specification of prosodic features and/or their phonetic implementation. Yet decades of research on the prosodic encoding of IS across languages leaves fundamental questions unanswered. This talk focuses on speaker variation in the production of prosody, and the reflection of that variation in listeners’ perception of IS meaning. Evidence from prominence perception studies of English, French, Hindi, Spanish, and Russian shows that listeners perceive prominence in unscripted speech not only in relation to acoustic prosodic cues, but also as a function of the syntactic, semantic, and pragmatic context, and calibrated by speech style. Additional evidence shows differences among individual listeners in the weighting of acoustic cues to prosody, and differences in the strength of association between prosodic expressions and particular types of IS meaning—with many different prosodic expressions accepted as instances of broad (new-information) focus, and narrower criteria for prosodic expressions accepted as marking contrastive focus. Implications for the acoustic modeling of prosody and its relationship to IS meaning will be discussed as motivation for the development of a novel, information-theoretic model of the link between sound and meaning.
Speaker bio: Prof. Cole received her Ph.D. in Linguistics from the Massachusetts Institute of Technology in 1987 and was on the faculty at the UIUC (Linguistics and Cognitive Science, 1990-2016) and Yale University (Linguistics, 1987-1989) prior to joining Northwestern in 2016. She has served as elected chair of the AAAS Section Z [Linguistics and Language Science], on the National Research Council Board on Behavioral, Cognitive & Sensory Sciences, on the Board of the American Institute of Pakistan Studies, and on the Board of the Linguistic Society of America. Dr. Cole was the founding editor for the journal Laboratory Phonology (2000-2005), and was on the editorial board for Language, Phonology, and the Oxford Research Reviews in Linguistics. Dr. Cole has received research funding from the NSF, NIH, Department of Education, US National Security Education Program, and the Volkswagen Stiftung.

Event : Seminar
Title : Flux linkages: the concept, its usage, and limitations
Speaker : Prof. Udaya Kumar
Date : 14/09/2018
Venue : C 241 MMCR, EE
Abstract: The principle of electromagnetic induction has been employed in most of the electrical engineering applications. The original statement of the law (stated in 1831) by its discoverer “The Great Michael Faraday” was modified later with an intention of generalization. The two distinct forms of the phenomena namely the flux-cutting and the flux-linkages have been sometimes the source of confusion. There are several attempts to address the problem, which has not entered the mainstream.
On the other hand, the basic concept of flux-linkages by itself can lead to a very limited picturization of the actual phenomena and perhaps the role of the conductor in both the forms of the law seems to have been undermined. Taking examples from electrical engineering, this presentation aims to intuitively build a better description of the phenomena and further its evaluation. Couple of examples, which were once classified under “paradoxes of electromagnetic induction ”, will also be dealt at the end.

Event : Thesis Defence
Title : Modulation of Power Electronic Converter Fed Split-phase Induction Machine Drive.
Speaker : Sayan Paul
Degree Registered: MSc (engg)
Advisor : Dr. Kaushik Basu
Date : 07/09/2018
Venue : B-303, 2nd floor, EE
Abstract: Multi-phase induction machine (IM) is attractive for high power applications due to reduced power rating of individual phase-drive unit. Six-phase induction machine, one of the most common multi-phase machines, is of two types: Symmetrical and Asymmetrical six-phase. The later one is also known as split-phase induction machine (SPIM). SPIM has two sets of three-phase windings spatially shifted by 30 degree electrical. This winding arrangement makes SPIM advantageous over Symmetrical six-phase machine due to its less susceptibility towards stator excitation harmonics. In this work, modulation of power-electronic converter fed SPIM drive has been investigated.
The power-electronic converters considered in this work are of two types: DC-AC (Inverter) and AC-AC (Matrix Converter). Inverter requires an active front-end rectifier to interface with the conventional three-phase grid whereas matrix converter (MC) can be directly integrated to the three-phase grid. Though relatively complex to control MC provides a high power-density solution. The modulation strategies of MC fed SPIM hasn't been still explored in the literature. So this thesis is aimed to devise the modulation techniques of inverter and MC fed SPIM drive.
This work proposes a new way of modeling the six-phase inverter driving a SPIM. This new model enables us to unify seemingly different existing modulation techniques. The theoretical maximum modulation index achievable without injecting any harmonics both in phase currents and torque have been derived with the help of this new model. Two novel modulation techniques have been devised beyond this limit. These techniques don’t produce pulsating torque and achieves significantly better phase current harmonic distortions compared to the existing modulation technique. This implies about 8% increase in the converter gain or voltage conversion ratio.
The thesis also discusses a modulation strategy of MC fed SPIM drive without injecting any harmonics in phase current. The maximum output voltage obtainable with this strategy for a given input voltage has been derived in the thesis. Modulation of 3 phase-3 phase MC has been discussed in this work.
All the modulation techniques have been verified through simulations in Matlab Simulink and experiments performed on a 5 kW hardware prototype.

Event : Seminar
Title : Will the Spring of AI Last Forever? Past and Future of Speech Research
Speaker : Prof. Dirk Van Compernolle
Date : 07/09/2018
Venue : C 241 MMCR, EE
Abstract: In this talk we will review the evolution in the field of speech recognition over the past 50 years. We start with the early years where some of our basic algorithms were invented and small systems were handcrafted by experts. 1974 was about the first time that someone quoted "speech recognition to be a done deal". Quickly the field went 100% the machine learning path and for almost 20 years the HMM technology seemed to ride on Moore's law. With limited added sophistication the HMM framework made handsome use of the ever-growing computing resources. However, the ten years after that there was relatively slow progress. Performance seemed to settle on a "plateau". Remarkably or not it is during this winter that modern top technology players such as Google, Apple, Facebook, Microsoft, … started introducing speech recognition products with "old technology". The last decade saw the surfacing of "deep learning" which take advantage of large amount of data and the gigantic amounts of computing power delivered by modern GPUs. As a consequence speech recognition products have transitioned from a niche to a mainstream technology with acceptable performance in many situations. To my knowledge, no one in the field predicted that the impact of deep learning was going to be so dramatic. Neither did anyone foresee that the relevance of human expertise in a particular field would become so irrelevant so quickly. So where is this "spring of AI" going? Machine learning has surpassed human learning because it can handle more data but it does not (yet) generalize better from small amounts of data nor does it surpass millions of years of evolution in biological processes.
Speaker bio: Dirk Van Compernolle received the electrical engineering degree from the KU Leuven (Belgium) in 1979 and a Ph D in 1985 from Stanford University with a thesis on multichannel speech processing strategies for cochlear implants. After a postdoc at IBM, working on robust speech recognition, he joined the Electrical Engineering Department (ESAT) of the KU Leuven in 1987, where he became a full professor in 1994. In 1994 he joined the Belgian speech technology startup Lernout Hauspie Speech Products as Vice President in charge of the speech recognition division. He stayed with L&H till 2000, a period in which the company grew from 100 to over 2000 employees. After L&H he returned to KU Leuven where he has held various positions. Apart from his research activities in speech & language technology, he has been active as a business angel and investor in a wider range of technologies. His research interests include robust speech recognition, speech enhancement and non-parametric machine learning applications. His recent research mostly focused on speech recognition for low resource languages.

Event : Seminar
Title : Distributed optimization for multiagent systems
Speaker : Dr. Ashish Cherukuri
Date : 30/08/2018
Venue : C 241 MMCR, EE
Abstract: Multiagent systems are capable of providing innovative services to citizens and improving the efficiency of existing infrastructures. However, realizing the full potential of such systems is quite challenging. In this talk we will outline some of these challenges pertaining to the area of optimization.
We will see how design of efficient coordination algorithms help address some of these issues. In addition, we will investigate distributed algorithms for stochastic optimization problems. The talk will conclude with an overview of the future research directions.
Speaker bio: Ashish Cherukuri is currently a postdoctoral researcher in the Automatic Control Laboratory, ETH Zurich. He obtained his PhD from University of California, San Diego. Prior to that, he received Bachelors of Technology in 2008 and Masters of Science in 2010, both in Mechanical Engineering, from IIT Delhi and ETH Zurich, respectively. He will be joining University of Groningen as a tenure-track Assistant Professor in February 2019. His webpage is http://people.ee.ethz.ch/~cashish/

Event : Seminar
Title : Neural networks in speech data mining
Speaker : Prof. Jan "Honza" Cernocky, Dr.Oldrich Plchot, and Mr. Karel Benes
Date : 31/07/2018
Venue : C 241 MMCR, EE
Abstract: In the last decade, neural architectures have consumed a significant portion of speech data mining, and the BUT Speech@FIT group has always been at the forefront of the developments. In this talk, we will give an overview of our current activities in automatic speech recognition (ASR), speaker recognition (SR) and spoken language recognition (SLR) based on neural networks. In ASR, we will summarize our work on rapid prototyping of systems for new languages and domains, including multi-lingual and semi-supervised training of acoustic models and subspace adaptation of neural language models. In SR and SLR, we will cover the current developments of NNs replacing individual building blocks of classical i-vector systems, NN-derived embeddings (so called x-vectors) and end-to-end neural systems. We will also mention current open basic research questions related to NNs in speech data mining and NLP.
Speaker bio: Dr. Honza Cernocky, Assoc. Prof. is the founder and managing director of BUT Speech@FIT, he is responsible for projects, funding and international relations of the group. He is also the Head of Department of Computer graphics and Multimedia at FIT BUT. http://www.fit.vutbr.cz/~cernocky/
Oldrich Plchot, Ph.D. is senior researcher in BUT Speech@FIT and is responsible for R&D of speaker and language recognition, including numerous successes in NIST and DARPA evaluations. http://www.fit.vutbr.cz/~iplchot/
Karel Benes is MSc. graduate of BUT and is currently PhD student in BUT Speech@FIT. His research concentrates on NN architectures in language modelling for ASR and historical text analysis. With Karthick Baskar, he has won ISCA Best student paper prize at Interspeech 2017 in Stockholm. http://www.fit.vutbr.cz/~ibenes/

Event : Thesis Colloquium
Title : Degradation Studies on Polymeric Insulators used for EHV and UHV Transmission
Speaker : Alok Ranjan Verma
Degree Registered: PhD
Advisor : Dr. Subba Reddy B
Date : 23/08/2018
Venue : High Voltage Seminar Hall, EE
Abstract: High voltage insulators used in overhead power transmission systems are of key importance for safe, reliable, and efficient operation in transferring huge amount of electrical power. Conventionally, ceramic/glass insulators were used in electrical power transmission, recently, in the country and elsewhere composite/ polymeric insulators are being used due to their promising advantages. These insulators are of recent origin and organic in nature, their material properties like surface electrical resistance and long-term performance are still under consideration by the international technical committees CIGRE, IEC, IEEE etc. The present research work focuses majorly on the performance of polymeric insulators in service life, we have carried out the (1) Investigation on material properties specifically to find the surface resistance against Electrical tracking and Erosion of polymeric material. (2) Long term performance of composite insulators subjected to Multistress and rotating wheel and dip test and its analysis. The investigation pertaining to the surface electrical resistance of polymeric insulating material is performed using Inclined Plane Tracking and Erosion (IPT) method on flat samples. However, for the long-term performance full scale samples with limited creepage length are evaluated. Rotating wheel & dip test (RWT) and Multistress experimentation that involves cyclic application of UV (1 W/m2), thermal (50°C), humidity (up to 85% RH) and with electrical stress. These studies are carried out under proposed environmental condition simulating acid rain as per observed acid rain information, with AC and DC voltages. The long-term performance is evaluated for the experimental duration of 1000 hours under each condition with minimum of three test specimen. During the experimentation, leakage current is measured and recorded at regular intervals of time using data acquisition system. After the experiments the degraded samples are evaluated using different physico-chemical techniques, which involve Fourier Transform Infra-Red (FTIR) spectroscopy, Scanning Electron Microscope (SEM) and Energy Dispersive X-ray Analysis (EDAX) analysis, and Thermo-gravimetric Analysis (TGA) etc. Further, the recurrence plot analysis is performed on measured leakage current and different quantification parameters are computed. Some interesting inferences which are drawn based on the observation for specific patterns in the leakage current waveforms which were not reported earlier will also be presented.

Event : Thesis Colloquium
Title : Detection and Imaging of Buried Landmines Using Ultra Wide Band
Speaker : Vijayakumar Solaiselvam
Degree Registered: PhD
Advisor : Dr. Joy Thomas M.
Date : 20/08/2018
Venue : High Voltage Seminar Hall, EE
Abstract: Conflicts between nations are on the rise and remnants of the past conflicts have left behind more than 100 million unexploded ordnance (UXO) across the world and these are invariably lodged in the ground. Most of the UXOs belong to three categories – (i). Anti personal / anti-tank landmines, (ii). Improvised explosive devices (IED) and (iii) Remnants of explosive devices used in wars like cluster bombs, gun shells etc. On an average, every year more than 4,000 people lose their life/limbs around the world due to these UXOs. Renewed conflict across the globe led to the highest causalities in the year 2017, which recorded more than 10,000 causalities. India is also one of the UXO infested country. India has millions of anti-personal/anti-tank mines buried across its borders and improvised explosive devices in the internal conflict zones. In the last decade, India lost more than 400 army and para military personal and it is the most important cause for peacetime causality of the Indian forces. India’s neighbours hold the highest landmine stockpile in the world. It is reported that for every 2000 mines cleared, one deminer loses his life. So there is an urgent need for developing better equipment/technologies for efficient and reliable detection of buried unexploded landmines. This thesis is an attempt to contribute in this area of great importance to the nation’s security. Anti-personal landmines and IEDs are generally buried at a shallow depth of up to 40 cm. The reduced and sometimes non-existent metal content in these devices make it difficult to detect those using metal detectors. Varied nature of soil medium like soil moisture, humus and soil texture make the detection process complex. Many electromagnetic (EM) wave based subsurface landmine detection systems have been developed by different countries. But most of them failed to satisfy the requirements of the United Nation Mine Action Standard (UNMAS). Many EM based techniques like complex natural frequency classification and traditional sub-GHz detection system were proposed in various literatures. In the recent times high frequency (>1GHz) subsurface imaging is proposed as a simplified solution for detection of both anti-personal landmines and improvised explosive devices. Moisture in the soil medium attenuates most of the higher frequencies, making it difficult to detect the UXOs. So for a successful use of high frequency based UXO detection system, a detailed understanding of the soil medium and its effect on the incident EM signal bandwidth and mean frequency is required. Also most of the earlier research lacks a detailed study related to the effect of soil medium on the detection range and the detection resolution. Hence in this thesis an exhaustive EM model for varied types of soil for different moisture content and clay fractions has been developed. Systems using EM signals of sub-GHz frequencies offer poor detection resolution whereas use of frequencies above 1GHz leads to reduced detection range. Collaborative attempts have been made with the EE signal processing group to circumvent this problem using low frequency super resolution signal processing technique. The results have been verified experimentally using ground penetrating radars (GPR) with the collaborative effort of Geotechnical group at IISc . The results will be presented and discussed during the seminar.

Event : Thesis Defence
Title : Multiview 3D reconstruction using rank-constrained ADMM
Speaker : Sk Miraj Ahmed
Degree Registered: MScEngg)
Advisor : Dr. Kunal Narayan Chaudhury
Date : 14/08/2018
Venue : C 241 MMCR, EE
Abstract: We consider the problem of reconstructing a 3D surface from its multiview scans. Typically, the computational pipeline for this problem has two phases: (I) finding point-to-point correspondences between overlapping scans, and (II) registration of the scans based on the correspondences. The focus of this thesis is on phase II. In particular, we work with a global registration model, where the scans are registered in one-shot using rotations and translations. We consider a least-squares formulation of global registration, where the variables are the transforms (rotations and translations) associated with the scans. The present novelty is that we reduce this intrinsically nonconvex problem to an optimization over the positive semidefinite cone, where the objective is linear but the constraints are nonconvex (a rank constraint is involved). We propose to solve this using variable splitting and the alternating direction methods of multipliers (ADMM). Due to the linear objective and the structure of constraints, the ADMM sub-problems turn out to be projections with closed-form solutions. In particular, for m scans, the per-iteration cost is the partial eigendecomposition of a 3m x 3m matrix, and (m-1) singular value decompositions of 3 x 3 matrices. We empirically show that for appropriate parameter settings, the proposed solver has a large convergence basin and is stable under perturbations. This is in keeping with recent empirical results on the effectiveness of ADMM for nonconvex programming (the convergence theory is still in its infancy though). We use the proposed ADMM algorithm to align 3D scans, where we determine the pairwise correspondences (in phase I) using the standard ICP algorithm. We present results on simulated and real datasets to demonstrate the effectiveness of our method. A remarkable feature of our method is that it can tolerate heavy amount of outliers in the correspondences. In particular, our method has better noise robustness than existing methods, where by noise we mean both perturbations in measurements and correspondences. The proposed method therefore has a wider scope of application beyond 3D reconstruction. An interesting open problem in this regard is establishing convergence (optimality) for the ADMM iterations; this is not covered by exisiting results.

Event : Thesis Colloquium
Title : Analysis of whispered speech and its conversion to neutral speech
Speaker : Nisha Meenakshi
Degree Registered: PhD
Advisor : Prasant Kumar Ghosh
Date : 13/08/2018
Venue : C 241 MMCR, EE
Abstract: Whispering is an indispensable form of communication that emerges in private conversations as well as in pathological situations. In conditions such as partial or total laryngectomy, spasmodic dysphonia etc, alaryngeal speech such as esophageal, tracheo-esophageal speech and hoarse whispered speech are common. Whispered speech is primarily characterized by the lack of vocal fold vibrations, and, hence, pitch. In recent times, applications such as voice activity detection, speaker identification and verification and speech recognition have been extended to whispered speech as well. Several efforts have also been undertaken to convert the less intelligible whispered speech into a more natural sounding neutral speech. Although supported by literature, research towards gaining a better understanding of whispered speech largely remains unexplored. Hence, the aim of the thesis is two-fold, 1) to analyze different characteristics of whispered speech using both speech and articulatory data, 2) to perform whispered speech to neutral speech conversion using the state-of-the-art modelling techniques. In the first part of this thesis, we analyze whispered speech using both audio data (recorded via microphone) and articulatory data (recordings of movements of articulators, such as lips, tongue, jaw etc, using Electromagnetic Articulography synchronous with audio data). Specifically, we experimentally analyze how the pitch-less whispered speech encodes information such as speaker's gender and voicing, that are typically pitch-dependent in neutral speech. We find that whispered speech does retain speaker's gender and voicing related information. This could be attributed to the exaggerated movements of the articulators that typically occur while trying to maintain intelligibility in the absence of pitch. Therefore, we next investigate for the optimal transformation function that relates whispered articulatory movements with those of neutral speech. Experiments reveal that an affine transformation could relate the two articulatory movements better than other candidate functions considered. In addition, we also find how much the acoustics of whispered speech carries information about the corresponding articulatory movements compared to that of neutral speech. In the second part, we design a feature that is necessary for segmenting whispered speech from a long recording of noisy whispered speech interleaved with silence/noise segments, as a per-processing step in the conversion/reconstruction framework. In order to reconstruct neutral speech from whispers, we follow a voice conversion-based approach which requires an appropriate parametrization of the whispered speech spectrum. For this, we experimentally find an optimal choice of parameters that is robust, both, for representation and to handle modelling errors. This representation is employed in the proposed bi-directional long short-term memory based whispered to neutral speech conversion system that yields a perceptually more natural sounding speech compared to the state-of-the-art conversion systems.

Event : Seminar
Title : Brain Computer Interfaces as Experimental Tools
Speaker : Dhruv Jawali
Degree Registered: PhD
Advisors : Prof. Chandra Sekhar Seelamantula and Dr. Supratim Ray
Date : 10/08/2018
Venue : C 241 MMCR, EE
Abstract: Our brains are a vast network of specialized, electrically active cells called neurons. To study how these neuronal networks perform computations, neuroscientists place a dense array of electrodes in the area of interest. For example, hand movement in monkeys is studied by placing micro-electrode arrays in the motor cortex. Over the years, it has become clear that the activity of the network as a whole rather than individual neurons is responsible for such movements. Making sense of the individual responses of a sparse subset of neurons from which recordings can be practically obtained is both challenging and frustrating. A major problem is that a direct, causal connection between the recorded signal and behaviour is difficult to establish.
In this talk, we will look at how Brain Computer Interfaces can be employed to understand how the brain learns. I will present the results reported in the paper titled "Neural constraints on learning", published in Nature in 2014, and discuss the algorithms used to learn and perturb the internal representations of information within the brain.
Speaker bio: Dhruv received his bachelor’s degree in Computer Science from NIT Goa in 2014, after which he worked in Samsung Research Labs, Bengaluru for one year. He joined the PhD programme at the National Mathematics Initiative (NMI) Department, IISc in 2015 under the joint guidance of Prof. Chandra Sekhar Seelamantula and Dr. Supratim Ray. He works on sparse coding of neural signals and dictionary learning algorithms.

Event : Thesis Colloquium
Title : Total electric field due to an electron avalanche and its coupling to the transmission line conductors
Speaker : Debasish Nath
Degree Registered: PhD
Advisor : Prof. Udaya Kumar
Date : 10/08/2018
Venue : High Voltage Seminar Hall, EE
Abstract: Transmission of bulk electric power from the generating stations to the load centers can be carried out only through high voltages transmission lines. One of the main issues in the design and perhaps maintenance of extra and ultra-high voltage transmission system is the corona, a local electrical breakdown of air on the line conductors and hardware. Even though the dimensioning of these elements is made considering the corona onset, surface abrasions arising either during the installations or during the operation can lead to the intolerable corona. Apart from producing some insignificant chemical reactions and noticeable acoustic noise, they can be a significant source of electromagnetic interference. In the early days, this interference was of concern only to radio and television receptions, however, with extensive use of wide frequency bands for modern applications, it has assumed prime importance.
The EMI due to the transmission line corona has been extensively studied and reliable empirical formulas have been proposed. The basis for all the earlier studies was the experimentally measured corona currents. On one hand, there were assumptions on the mode of corona current injection on to the conductor and on the other, the frequency range involved were not adequate for the modern-day applications. From the theoretical perspective, the coupling of the field produced by corona to the conductor was hardly investigated and the total field produced by the corona itself was not quantified. In order to address these serious lacunae, the present work was taken up and it can be considered as the first leap towards the correct picturization, as well as, quantification of the problem.
The field produced by the electron avalanche involves noticeable retardation effects. In the literature, only the field produced by arbitrarily moving point charge of fixed strength is available. On the contrary, the avalanche involves growing spherical electron cloud with trailing positive charge, which is almost stationary. Starting from the basics, an analytical expression for the total field due to an avalanche has been derived for the first time. Suitable validation has been provided through numerical simulation of the electric field integral equation.
Indeed, corona discharge is a complex phenomenon having many distinctly different modes which differ in their visual, as well as, electrical characteristics. Innumerable electron avalanches contribute to the measured corona current with their space-charge acting as a moderator. Therefore, in order to model for the corona on conductors, an indirect approach based on linearity is proposed. An equivalent spatio-temporal dipole distribution was obtained to produce the measured current on the conductor. The general expression derived for the isolated avalanche is extended for this purpose.
Using the above, the means of induction, the spatial decay rate of corona current in the close range, its propagation mode, and field produced by both avalanche/equivalent dipole and that due to induced current in the conductor, have all been investigated and quantified.
In summary, the contribution made in this work is more basic in nature and would be of significant interest to the high voltage power the transmission line, as well as, the communication engineers.

Event : Thesis Colloquium
Title : Face Recognition in Unconstrained Environment
Speaker : Sivaram Prasad Mudunuri
Degree Registered: PhD
Advisor : Dr. Soma Biswas
Date : 17/07/2018
Venue : C 241 MMCR, EE
Abstract: The increasing use of surveillance cameras for addressing security concerns has led to increased demand for robust face recognition systems. The images captured by the surveillance cameras usually have poor resolution, uncontrolled pose and illumination conditions which makes the task of recognizing these faces extremely challenging. Significant attention has been devoted to addressing one or more of the different challenges like poor illumination, non-frontal pose, expression, etc. But addressing all these challenges together is essential in many applications like recognizing faces from surveillance cameras. The research in this thesis is motivated by this need. The main focus of this thesis is to develop algorithms to match low-resolution (LR) facial images captured under wide range of pose variations and poor illuminations with high-resolution (HR) facial images captured in frontal pose and good illumination conditions which are often available during enrolment. The contributions of our work are as follows: 1) We addressed the problem of low-resolution face recognition across pose and illumination variations 2) Addressed the problem of matching LR NIR faces with HR VIS faces to address the recognition task in low-light conditions. In the first part of this thesis, we address LR face recognition task using metric learning which does not require localizing facial landmarks in non-frontal face images at low resolution during testing. We then explored the LR face verification task using the deep learning framework. We generalized this deep learning framework for LR object recognition in which the testing objects have not been seen during the training. The work in the second part of this thesis is motivated by the fact that recognizing the faces of uncontrolled subjects captured in night-time/low-light conditions using near-infrared (NIR) cameras is quite challenging since the images have very different visual characteristics from the gallery images that are captured using RGB camera. Several approaches have been proposed to address this task, but most of them deal with good resolution images. The NIR images captured by the surveillance cameras usually have low resolution in addition to considerable variations in pose. This makes the problem even more challenging, since now the low-resolution, uncontrolled NIR probe images needs to be matched against the high-resolution controlled VIS gallery images. But this scenario is relatively less explored in literature. We proposed a dictionary alignment approach for addressing this problem. We also proposed a re-ranking approach to further improve the recognition performance for each probe by combining the rank list given by the proposed algorithm with that given by another complementary feature/algorithm. Finally, we have also collected our own database HPR (Heterogeneous face recognition across Pose and Resolution) which has facial images captured from two surveillance quality NIR cameras and one HR visible camera, with significant variations in head pose and resolution.

Event : Seminar
Title : Voice based Asthmatic patient and healthy subject classification
Speaker : ShivaniYadav
Degree Registered: Direct PhD
Advisors : Dr. Prasanta Kumar Ghosh and Dr. Dipanjan Gope
Date : 17/08/2018
Venue : C 241 MMCR, EE
Abstract: In this work, we consider the task of automatic classification of asthmatic patients and healthy subjects using voice stimuli. Cough and wheeze have been used as voice stimuli for this classification task in the past. In this work, we focus on sustained phonations, namely /A:/, /i:/, /u:/, /eI/, /oU/ and compare their classification performances with the cough and wheeze. Classification experiments using 35 asthmatic patients and 36 healthy subjects show that sustained vowel /i:/ achieves the highest classification accuracy of 80.79% among five vowels considered. However, it is found to be higher and lower than the classification accuracies of 78.72% and 90.25% obtained using cough and wheeze respectively. This suggests that for speech-based asthma classification, /i:/ would be a better choice compared to other vowels considered in this work. However, when non-speech sounds are included for classification, wheeze is a better choice than sustained /i:/.
Speaker bio: Shivani Yadav received the bachelor’s degree from NIT Jalandhar, Punjab from Instrumentation and Control department in 2015. She joined the direct PhD programme in BSSE , 2015 under the joint guidance of Dr. Prasanta Kumar Ghosh and Dr. Dipanjan Gope.

Event : Seminar
Title : Distributed Control of Large-Scale Infrastructure Networks: Looking Beyond Stability
Speaker : Sivaranjani Seetharaman
Date : 02/08/2018
Venue : C 241 MMCR, EE
Abstract: With the proposed development of smart cities around the world, research into novel scalable control techniques to ensure performance and safety in large-scale infrastructure networks is becoming increasingly important. The first half of the talk will focus on distributed control policies for disturbance management in networked systems by exploiting integrated communication infrastructure. In the second half of this talk, dissipativity-based scalable compositional control techniques to ensure robustness and performance of large-scale interconnected systems will be discussed. The talk will draw upon applications in power grids and transportation networks to demonstrate the implementation of the proposed distributed control designs.
Speaker bio: Sivaranjani Seetharaman is a graduate student in the Department of Electrical Engineering, University of Notre Dame, where her research focuses on distributed control for large-scale infrastructure networks, with emphasis on power grids and transportation networks.

Event : Thesis Colloquium
Title : Performance Improvement of a High Voltage Power Converter for Microwave Power Module
Speaker : P. Sidharthan
Degree Registered: PhD
Advisors : Prof. G. Narayanan, Dr. S. K. Datta, MTRDC, DRDO
Date : 31/07/2018
Venue : C 241 MMCR, EE
Abstract: Microwave Power Module (MPM) is a medium-power microwave amplifier finding applications in space & terrestrial communication, radar and microwave imaging systems. The MPM has a traveling wave tube (TWT) as the main power amplifier, fed by a solid-state power amplifier (SSPA), sharing the overall gain. An Electronic Power Conditioner (EPC) within the MPM, powers the SSPA and the TWT, apart from housekeeping subsystems of the MPM. This PhD work attempts to improve the performance characteristics of an MPM by improving the performance of the EPC, and more particularly, by improving the performance of the high voltage power converter, which powers the accelerating anode of the TWT.
Experimental investigations carried out on a state-of-the art EPC, employing a series resonant converter (SRC) for the accelerating anode of the TWT, revealed that the RF output of the TWT carried spurious components beyond permissible limits for certain communication applications. Further studies proved that the spurious components generated are due to the presence of high voltage ripple present on the accelerating anode of the TWT. The requirement of the MPM for an airborne application also demanded weight and volume reduction.
A passive RC-filter,deployedfor the accelerating anode voltage,has reduced the spurious components by 12dB. Investigations on mounting of power MOSFETs for the full-bridge high voltage power converter led to studies on common-mode EMI generated while a power MOSFET is switching. Studies conducted on the common-mode current drawn by a MOSFET while switching a resistive load have proved that an improved mounting is possible which could reduce the conducted EMI by 20dB and the common-mode current by 10dB.
Studies conducted with non-invasive infrared (IR) imaging techniques have demonstrated that proper selection of the dielectric and shielding materialsfor isolating the MOSFET-tab from the heatsink could reduce the thermal resistance between the taband sink by an order of magnitude. Novel techniques with IR imaging are used for comparison of high voltage,fast reverse recovery diodes used in the high voltage power converter. It is shown that SiC diodes perform significantly better thanultra-fast and super-barrier diodes in terms of reverse recovery and diode losses.
The conventional method of detecting the RF output level using an RF directional coupler and an RF detector diode is improved by a solution, which has yielded 20:1 volume reduction and 40:1 weight reduction. A high voltage planar transformer has beendeveloped for delivering 400W at 4.3kV to the accelerating anode voltage, weighing 75 gram. The proposed solutions are useful to reduce the size andweight and to improve the performance of MPM.

Event : Seminar
Title : Fast total variation minimizing image restoration under mixed Poisson-Gausssian noise (Click here for the poster)
Speaker : Mr. Manu Ghulyani
Date : 27/07/2018
Venue : C 241 MMCR, EE
Abstract: Image acquisition in many biomedical imaging modalities is corrupted by Poisson noise followed by additive Gaussian noise. Maximum Likelihood Estimation (MLE) based restoration methods that use the exact Likelihood function for this mixed model with non-quadratic regularization are v ery few. In particular, while it has been demonstrated that total variation (TV) based regularization methods give better results, such methods that use exact Poisson-Gau ssian Likelihood are slow. Here, we propose an ADMM based fast algorithm for image restoration using exact Poisson-Gaussian Likelihood function and TV regularization. Specifically, we propose a novel variable splitting approach that enables isolating the complexity in the exact log-likelihood functional from the image blurring operation, allowing a fast Newton-like iteration on the log-likelihood functional. This leads to a significantly improved convergence rate of the overall ADMM iteration.
We give sufficient conditions for convergence of this algorithm. We also propose Expectation-Minimization based iterations to further exploit the proposed splitting approach.
The effectiveness of the proposed methods is demonstrated using restoration examples.
Speaker bio: Manu Ghulyani received the B.E. (Hons.) degree from BITS, Pilani, India, in 2012. After completing the B.E., he worked as an Operations engineer at NTPC Ltd from 2012 till 2015. Since 2015 he has been an M.Sc(Engg.) student under the supervision of Dr. Muthuvel Arigovindan at the Department Of Electrical Engg., IISc. His research interests include Image Restoration, Optimization, Machine Learning and Statistical Modeling

Event : Thesis Defence
Title : A Fast Constant-Time Approximation for Locally Adaptive Bilateral Filtering
Speaker : Ruturaj G. Gavaskar
Degree Registered: MTech(Res)
Advisor : Kunal Narayan Chaudhury
Date : 27/07/2018
Venue : C 241 MMCR, EE
Abstract: Smoothing is a fundamental task in low-level image processing that is used to suppress irrelevant details while preserving salient image structures. The simplest smoothing mechanism is to average neighboring pixels using a spatial kernel. While this works well when the kernel is narrow, it inevitably results in blurring of edges when the kernel is wide. This problem can be alleviated using a range kernel along with the spatial kernel. The range kernel automatically damps out the smoothing action near an edge and is turned off in homogeneous regions where greater smoothing is required. A canonical prototype in this regard is the bilateral filter in which both kernels are Gaussian. A flip side of the range kernel is that it makes the bilateral filter non-linear and computationally expensive. However, several fast algorithms have been proposed in the literature that allow the filter to be implemented in real-time.
The focus of this thesis is on a generalization of the classical bilateral filter in which the center and width of the range kernel are allowed to change from pixel to pixel. The so-called adaptive bilateral filter was originally proposed for image sharpening and noise removal, but it can also be used for other applications. Similar to the classical bilateral filter, its brute-force implementation requires intense computations. However, most fast algorithms for classical bilateral filtering require the range kernel to be fixed, and hence cannot be extended for the adaptive counterpart.
For the first time, we propose a fast algorithm for adaptive bilateral filtering. The algorithm is constant-time in that the computational complexity does not scale with the the width of the spatial kernel. At the core of the algorithm is the observation that the filtering can be performed purely in range space using an appropriately defined local histogram. By replacing the histogram with a polynomial and the finite sum in range-space with an integral, we can approximate the filter using a series of definite integrals. We derive an efficient algorithm from this analytic approximation using the following innovations: the polynomial is fitted by matching its moments to those of the target histogram (this is done using fast convolutions), and the integrals are recursively computed using integration-by-parts. The proposed algorithm can achieve at least 20X acceleration over the brute-force computation, without perceptible distortions in visual quality. We demonstrate the effectiveness of our algorithm for sharpening, removal of compression artifacts, texture filtering, and saliency-driven detail enhancement.

Event : Seminar
Title : Technologies for Indian Languages – solving real problems of real people (Click here for the poster)
Speaker : Prof. A G Ramakrishnan
Date : 20/07/2018
Venue : C 241 MMCR, EE
Abstract: The talk will give an overview of the work carried out in the Medical Intelligence and Language Engineering Laboratory over the past two decades. In 2001, the lab created the vision of creating automated book readers for the blind in Indian languages. The idea was that people with visual disability must be able to access any text in Indian languages that people with normal vision are able to read. This led the lab to Optical character recognition, script recognition at the level of the word to deal with bilingual and trilingual printed text, text-to-speech conversion, text detection and recognition from scene and born-digital images, transcription between any pair of Indian languages and then to recognition of handwriting online. Currently, we are working on automated speech recognition and document image superresolution using deep learning. Our technique of augmenting deep learning with conventional image processing techniques has led to a patent filing in superresolution. Our recent interests are in EEG studies on comatose patients and multimodal recognition of emotion from video images, speech, EEG, ECG, EMG and GSR.
Speaker bio: A G Ramakrishnan obtained Sir Andrew Watt Kay Young Investigator Award from the Royal College of Physicians and Surgeons, Glasgow for his Ph D work on the nerve conduction & evoked potential studies on leprosy patients; Manthan Award in 2014 for the impact of his Tamil and Kannada OCR on blind students; Manthan Award in 2015 for the use of his Tamil and Kannada TTS; Prof. Anandakrishnan Award from INFITT for developing Tamil handwriting recognition. www.kannadapustaka.org provides Braille and audio books of Kannada school texts using his Kannada OCR and TTS. He is a member of the Karnataka Knowledge Commission and FICCI Indian Languages Internet Alliance.

Event : Seminar
Title : Optimal Energy Extraction from a Flywheel
Speaker : Dr. S R Gurumurthy
Date : 18/07/2018
Venue : C 241 MMCR, EE
Abstract: Concern of increasing energy demand, exhausting fossil fuel reserves and consequences of climate change, urges us to design energy efficient systems to conserve the existing energy resources. Energy storage is one of the area where there is a scope to improve the efficiency. One of the devices which can store energy is flywheel and this has the merits of less maintenance, higher energy density, environmental friendly, longer life and unlimited charge and discharge cycles. Flywheel energy storage is now considered a viable technology for stationary applications like UPS, solar and wind power systems particularly when short time back up is required. This talk covers the discussion on the design methodology of an efficient Flywheel Energy Storage (FES) system which includes Bidirectional Power Converter (BDC), Brushless DC (BLDC) and a flywheel. The energy extracted from an FESS is limited by various factors like gain of the boost converter, source resistance of the generator and losses in the system. A brief discussion on these factors including a method for source-wise apportioning of the losses, techniques of reducing these losses, factors which influence the extractions of energy harvested will be covered in the talk. A novel scheme of using a multi-armature winding PM-BLDC generator and a buck converter for the enhancement of the harvested energy will also be discussed. The presentation also covers the applications of FESS in the Battery less UPS.
Speaker bio: Dr. S R Gurumurthy has obtained his B.E. in Electronics & Communication Engineering in 1984 from National Institute of Engineering, Mysuru and M Sc (Engg) in Power Electronics from Indian Institute of Science, Bengaluru and PhD on Flywheel Energy Storage systems from HBNI, Mumbai. As a Scientific Officer in BARC, he has been engaged in Design & Development, Testing & Commissioning of Variable high-speed drives, Power supplies, Special purpose instruments and various control systems for nuclear facilities for past three decades. Some of the important R&D contributions include High speed drives for special purpose machines and BLDC machines, Power supplies, Special instruments required for high speed rotor balancing like, Vibration Analyzers, True Power monitors, Controller for BLDC generators, Flywheel Energy Storage Systems. Also carried out the installation, commissioning and maintenance of VFDs, UPS, Control panels for the plant. Currently in Bhabha Atomic Research Centre Mysuru as Scientific Officer – H+ and engaged in the development of integrated FES based UPS and other instruments required for the plant.

Event : Seminar
Title : Soft Magnetic Thin Film Applications for integrated electronics
Speaker : Masahiro Yamaguchi
Date : 16/07/2018
Venue : C 241 MMCR, EE
Abstract: Ferromagnetic thin film materials, having high permeability at radio frequencies and above are candidate materials for use in inductive passive components in power electronics. This lecture begins with a review of new soft magnetic thin film materials, followed by discussions that include: (1) Development of international cross measurements of RF permeameters to evaluate RF permeability and related FMR profiles of magnetic films; (2) small signal lossy application to CMOS integrated electromagnetic noise suppressor; (3) large current permeable application to Point-of-Load type one-chip DC-DC converters. The lecture will conclude with an outlook that provides a perspective on the future of on-chip RF magnetics.
Speaker bio: Masahiro Yamaguchi received his BSc., MSc. and Ph D. in electrical engineering from Tohoku University, Japan in 1979,1981 and 1984, respectively. In 1984, Dr. Yamaguchi joined the Department of Electrical Engineering, Tohoku University as Assistant Professor. In 1990, Dr. Yamaguchi became Associate Professor at the Research Institute of Electrical Communication, Tohoku University. During October-December 1995, he joined Department of Electrical and Computer Engineering, University of Wisconsin-Madison, USA, as a Visiting Associate Professor. Since 2003, Dr. Yamaguchi works for the Department of Electrical Engineering, Faculty of Engineering, Tohoku University as a full Professor. His interest covers high frequency magnetic materials and applications, and EMC. He is awarded Best symposium paper award at APEMC2015. He was the 2016-2017 Chair of IEEE EMC Society Sendai Chapter, and awarded 2017 Chapter of the Year award. He is 2018-2019 Chair of IEEE Magnetics Society Sendai/Sapporo Joint Section Chapter, a TPC member of 2018 IEEE PELS Power-Supply-on-Chip Workshop, EMC Compo 2019, Joint MMM/Intermag 2019, and the Secretary/Treasurer of IEEE Magnetics Society. He may be reached at yamaguti@ecei.tohoku.ac.jp

Event : Seminar
Title : Speech enhancement using deep mixture of experts based on hard expectation maximization.
Speaker : Pavan Subhaschandra Karjol
Degree Registered: MSc (Engg.)
Advisor : Dr. Prasanta Kumar Ghosh
Date : 13/07/2018
Venue : C 241 MMCR, EE
Abstract: We consider the problem of deep mixture of experts based speech enhancement. The deep mixture of experts, where experts are considered as deep neural network (DNN), is difficult to train due to the network structure. In this work, we propose a pre-training method for individual DNN in deep mixture of experts. We use hard expectation maximization (EM) to pre-train the individual DNNs. After pre-training, we take a weighted combination of outputs of individual DNN experts and jointly train the whole system. We compare the proposed method with a single DNN based speech enhancement scheme. Speech enhancement experiments, in four SNR conditions, show the superiority of the proposed method over the baseline scheme. The average improvements obtained for four seen noise cases over single DNN scheme are 0.08, 0.59 dB and 0.015 in terms of objective measures viz perceptual evaluation of speech quality (PESQ), segmental signal to noise ratio (seg SNR) and short time objective intelligibility (STOI) respectively.
Speaker bio: Pavan is currently pursuing his MSc (Engg.) in SPIRE Lab, EE, IISc. Previously, he worked as an Associate Software Engineer at Robert Bosch, Bengaluru. He received his B.E in electronics and communication from R V College of Engineering, Bengaluru, in 2014. His area of interest is speech enhancement using deep learning techniques.

Event : Seminar
Title : Delivering efficiencies in health care and manufacturing.
Speaker : Prof. Svetha Venkatesh
Date : 13/07/2018
Venue : C 241 MMCR, EE
Abstract: This talk considers what to do when confronted with failures with current data or analysis. What can we do?
1. When current predictions for rare events are poor?
Instead of focusing on rare event classification, for example, suicide prediction, we focus on identifying the riskiest events with minimal error. Such events are likely precursors to outliers of interest. We demonstrate our results through outlier detection in surveillance (leading to our start-up company iCetana, Australia) and in suicide risk prediction (implemented in in Barwon Health, Geelong, Australia). We discuss the challenges in data modeling, pitfalls and our outcomes.
2. When data has special characteristics?
We predict cancer toxicity risk, and show how we leverage the special characteristics of the data to build better predictive models. We share our insights we have learnt in our path from such data to models.
3. When data is limited?
We use Bayesian optimization based methods to demonstrate how to accelerate the experimental process, the foundation of both product ad process design. We show how we have been able to impact the discovery of novel materials and alloys.
Speaker bio: Prof. Svetha Venkatesh (Alfred Deakin Professor and ARC Australian Laureate Fellow, and Director of Centre for Pattern Recognition and Data Analytics, Deakin University, Australia).

Event : Defence
Title : Application of Semi-Analytical Methods for Large Power System Simulations
Speaker : Disha L Dinesha
Degree Registered: MSc (Engg)
Advisor : Dr. Gurunath Gurrala
Date : 10/07/2018
Venue : C 241 MMCR, EE
Abstract: Power system dynamics can be accurately modeled in time domain by using nonlinear differential and algebraic equations (DAEs). The challenge lies in solving the large number of nonlinear DAEs faster than real-time in order to provide the operator with sufficient information on the unfolding critical contingencies to take preventive control measures. Research in these regards is focused on the use of new computing architectures, faster numerical solvers and efficient parallelization techniques. Semi analytical methods are frequently used for numerical simulations of real-world systems in the applied sciences and engineering including nonlinear ODE, PDE and DAE problems. Applicability of two widely used semi-analytical methods called Adomian Decomposition Method (ADM) and Homotopy Analysis Method (MHAM) have been explored for the time domain simulations of large power systems in this thesis. These methods have very narrow convergence region, i.e., the solution is close to the exact solution only when the time t is small. Applying them over successive time intervals as a sequence of initial value problems overcomes this limitation. These are called multi-stage methods. The multi-stage ADM and HAM are tested on 7 widely used test systems ranging from 10 generators, 39 buses to 4092 generators, 13659 buses. Impact of the step size and the number of terms is investigated on the stability and accuracy of the method. HAM gives a family of solutions and it is shown that ADM becomes a special case of HAM, which gives the best accuracy. This special case is also called as Homotopy Perturbation Method (HPM). The numerical stability and accuracy of the MADM and MHAM are found to be better than the modified Euler and weaker than the trapezoidal method. An average speed up of 42% and 26% is observed in the solution time of ODEs alone using the MADM when compared to the midpoint-trapezoidal method and the modified-Euler method respectively. MADM is also used for real-time simulation of buck, boost power electronic converters and 3-ph single machine infinite bus system. For quantifying the error generated at every time step of real-time simulation using MADM, a z-transform based method is used and the errors are compared with different numerical methods. The errors are found to be better than most of the widely used numerical methods, forward Euler, Backward Euler etc. for real-time simulation.

Event : Thesis Colloquium
Title : Target Detection and Tracking under Non-ideal Conditions in Airborne Radars
Speaker : Narasimhan R S
Degree Registered: PhD
Advisor : Prof. P S Sastry
Date : 09/07/2018
Venue : C 241 MMCR, EE
Abstract: Signal processing for target detection, parameter estimation and tracking in airborne radars is a challenge given the complexity of the operational environment. In this thesis we investigate the problems posed by non-ideal operational conditions for radar signal detection and tracking and propose novel, efficient and computationally light solutions for realization of robust signal processing techniques. This work aims to identify the lacunae of the current techniques for Pulse Doppler radar processing and proposes to improvise them to achieve practical and readily deployable solutions. Towards this we develop novel algorithms for adaptive threshold detectors for non-homogeneous background, suppression of ground clutter returns emanating from antenna sidelobes, target parameter estimation for Medium Pulse Repetition Frequency (MPRF) waveforms and suppression of spurious plots and tracking under dense clutter background.
Firstly, we bring out the non-idealities of the interference background, such as multiple interfering targets and clutter edge and discuss the design of adaptive threshold detector for such situations. In this context we propose censored cell averaging CFAR, switching censored cell averaging and greater of CFAR and robust variability index CFAR detectors. In the next problem, we focus on clustering the detected range-Doppler cells of a CFAR image applying connected component analysis technique and a valley detection logic to improve radar resolution. Suppression of clutter leaks emanating from antenna sidelobes is the next important aspect analyzed in the thesis. Here, we bring out the limitations of conventional sidelobe blanking and propose a novel sidelobe blanking technique based on quadrant subarrays of the main antenna. In our next study, we propose an efficient algorithm for range-Doppler unfolding as airborne radars employ medium pulse repetition frequency waveforms and measured range and Doppler is simultaneously ambiguous. The approach uses novel clustering technique. We conclude our work with the design of range rate tracking filter to simultaneously achieve conflicting requirements of low lag and low variance and use range rate information for efficient maneuver tracking under dense clutter background.
Speaker bio:

Event : Thesis Colloquium
Title : Fast total variation minimizing image restoration under mixed Poisson-Gaussian noise
Speaker : Manu Ghulyani
Degree Registered: MSc (Engineering)
Advisor : Dr. Muthuvel Arigovindan
Date : 27/06/2018
Venue : B 303, EE
Abstract: Fast total variation minimizing image restoration under mixed Poisson-Gaussian noise Image acquisition in many biomedical imaging modalities is corrupted by Poisson noise followed by additive Gaussian noise. Maximum Likelihood Estimation (MLE) based restoration methods that use the exact Likelihood function for this mixed model with non-quadratic regularization are very few. In particular, while it has been demonstrated that total variation (TV) based regularization methods give better results, such methods that use exact Poisson-Gaussian Likelihood are slow. Here, we propose an ADMM based fast algorithm for image restoration using exact Poisson-Gaussian Likelihood function and TV regularization.
Specifically, we propose a novel variable splitting approach that enables isolating the complexity in the exact log-likelihood functional from the image blurring operation, allowing a fast Newton-like iteration on the log-likelihood functional. This leads to a significantly improved convergence rate of the overall ADMM iteration. We give sufficient conditions for convergence of this algorithm. We also propose Expectation-Minimization based iterations to further exploit the proposed splitting approach. The effectiveness of the proposed methods is demonstrated using restoration examples.
Next, we extend this method for super-resolved image reconstruction for structured illumination microscopy (SIM). In SIM, extension of resolution beyond diffraction limit is achieved by illuminating the sample with a sinusoidal pattern. While known practical methods achieve reconstruction for SIM by modifying the measured data with sinusoidal modulation followed by a regularized multi-PSF deconvolution, our approach achieves reconstruction by means of TV penalized MLE with exact likelihood composed of raw measured data.

Event : Seminar
Title : Adaptive Threshold Detectors for Airborne Radars
Speaker : Narasimhan R S
Degree Registered: PhD
Advisor : Prof. P S Sastry
Date : 15/06/2018
Venue : C 241 MMCR, EE
Abstract: Radar signal detection using adaptive threshold detectors (also called CFAR detector) in airborne radars encounter several non-ideal conditions that make it difficult to characterize the background interference. The situations which often occur in practice are (I) presence of multiple targets (or outliers) leading to degraded detection of primary target, (II) presence of abrupt transition in interference power leading to excessive false alarms and (III) the combination of (I) & (II). Designing an efficient adaptive detector that caters to all possible combinations of non-homogeneity is a non-trivial problem. Many algorithms based on ordered statistics, censoring of outliers using sorting, removal of extraneous samples using sample by sample hypothesis testing on sorted samples, ordered data variability index and few others are proposed in this context. These approaches require sorting or prior information on depth of censoring. In this talk, we discuss three censored CFAR techniques which do not need sorting and are computationally light. Censored Cell Averaging CFAR, Switching Censored Cell Averaging Greater Of CFAR and Robust Variability Index CFAR will be discussed.
Speaker bio: Narasimhan R S received BE degree in Electronics from Visveswaraya Technological University in the year 2003 and M.E degree in System sciences & Automation from Indian Institute of Sciences (IISc), Bangalore, India in 2009. He is pursuing doctorate in philosophy from Electrical Engineering Department, Indian Institute of Sciences under the supervision of Prof. K. R. Ramakrishnan in the field of radar signal processing and radar tracking algorithms. He is a scientist in Electronics and Radar Development Establishment (LRDE), Bangalore since 2003. He has worked in the field of radar system modelling, simulation, radar data and signal processing algorithms for airborne and ground based radar systems.

Event :Seminar
Title : Breaking Performance Limits in Digitally Controlled Hi-Frequency DC-DC Converters: Beyond Small-Signal Perspectives
Speaker : Dr. Santanu Kapat
Date : 14/06/2018
Venue : C 241 MMCR, EE
Abstract: Digitally controlled high-frequency DC-DC converters have been gaining widespread acceptance because of offering various technical benefits, such as real-time efficiency optimization, online controller tuning, insensitivity to noise and component variations, and many more. However, the use of linearized small-signal models (SSM) in existing design approaches can neither fully explore the potential performance benefits in DC-DC converters due to ignoring fast switching dynamics nor capture underlying nonlinear behavior due to finite sampling and quantization.
This presentation shows that the use of fast dynamics indeed helps to achieve near time-optimal recovery by real-time voltage-controller tuning in current-mode digital-pulse-width-modulators (DPWMs). The finite sampling and quantization effects under uniformly sampled DPWM often lead to border-collision bifurcations which are unlikely using analog PWM. These significantly increase the RMS inductor current, thereby increasing conduction losses and violate the output-voltage ripple constraints. Event-based variable-frequency ripple-based digital modulators offer superior performance and stability benefits over a DPWM, and experimental case studies are demonstrated along with some applications, like LED array driving, envelope tracking, dynamic voltage scaling, and 48V DC grid. Finally, future digital control challenges are highlighted.
Speaker bio: Santanu Kapat received the M.Tech. and Ph.D. degrees in Electrical Engineering from the IIT Kharagpur, India, in 2006 and 2010, respectively.
From 2009 to 2010, he was a Visiting Scholar in the Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign. From 2010 to 2011, he was a Research Engineer at GE Global Research, Bangalore, India. Since 2011, he has been with the Department of Electrical Engineering, IIT Kharagpur, where he is presently an Associate Professor.
His research interests include modeling, analysis and design of digital and nonlinear control in high-frequency DC-DC converters, and applications to dynamic voltage scaling, LED driving, DC nano-grid, bi-directional DC/AC converters for renewable energy applications.
Dr. Kapat is the recipient of the INSA Young Scientist Medal and INAE Young Engineering Award in 2016. He has been serving as Associate Editors for the IEEE Transactions on Power Electronics since 2015 and IEEE Transactions on Circuits and Systems II: Express Briefs since 2018. He is a Senior Member of IEEE.

Event : Seminar
Title : 3D embedding from sparse distance-constraints and its application to structural biology
Speaker : Niladri Ranjan Das
Degree Registered: PhD
Advisor : Dr. Kunal Narayan Chaudhury
Date : 08/06/2018
Venue : C 241 MMCR, EE
Abstract: Embedding points in a fixed dimension from a given set of distance constraints (equality/upper/lower) is a known NP-complete problem. This follows from reduction of graph realization in a fixed dimension (NP-hard) into the aforementioned problem. One of the many applications of this problem is determining 3D structure of protein from Nuclear magnetic spectroscope (NMR) experiments. The computational challenge is primarily due to the paucity of distance-geometry information extracted from the NMR spectra. We propose an optimization paradigm for protein structure determination that primarily relies on the experimental distance constrains and the classical covalent-bond geometry of the molecule. Our method can confidently model regions of the protein molecule with larger density of experimental bounds. Such regions are broken down into smaller fragments. Their 3D structures are efficiently resolved by solving smaller optimization problems (semidefinite programming). The solutions are finally combined in one-shot using convex optimization techniques. Post processing steps completes the algorithm. We have tested the method for NMR experimental data available online. The qualitative analysis of protein structures obtained shows improvement over contemporary distance geometric methods. The method is also found to be scalable and robust to data deletion.
Speaker bio: Niladri Ranjan Das received the B.E degree Computer Science and Engineering in 2011. Following this he has worked for two years as software developer in mainframe at IBM. He joined IISC in the year 2014 as direct PhD in mathematical science (NMI) jointly advised by Kunal Narayan Chaudhury (EE) and Debnath Pal (CDS). His research interest includes convex optimization, and linear nonlinear optimization.

Event : Defence
Title : Dual Comparison One Cycle Control for Grid Connected Converters
Speaker : Nimesh V
Degree Registered: Ph D
Advisor : Prof. Vinod John
Date : 11/06/2018
Venue : C 241 MMCR, EE
Abstract: Grid connected converters are widely used as front end rectifiers, interface between renewable energy and grid, and power quality applications. Popular control techniques, to generate gating signals for active devices, reported in literature are voltage oriented control, direct power control and conventional one cycle control(C-OCC). Light load instability and steady state dc offset phenomenon are the major concerns with of the conventional one cycle control reported in literature. These issues has been addressed in the literature by treating them independently.
C-OCC uses peak detection comparison method, so the peak of the current always confines with the grid voltage. In this thesis a method to control both the peak and valley of the current is proposed, such that the converter changes its state when the expected value of current is reached. Valley of the current in each carrier cycle is decided such that the current has no steady state dc offset in it. To control the peak and valley of current, one more comparison is required. So it is named as dual comparison one cycle control(DC-OCC). A generalized approach for controlling average current, in a carrier cycle, for grid connected converter is also proposed. Stability of the inherent current loop in DC-OCC, using propagation dynamics of small disturbance, shows that the proposed control strategy does not suffer from localized sub-harmonic instability of C-OCC. A novel method to compensate for the inductive drop is proposed this thesis.
The sensed input current is added with a fictitious current, generated from gating signal of active device, to enable bi-directional power flow in converters controlled by DC-OCC. A second order band pass filter(BPF) is used to generate the fictitious current from the gating signal. Effects of BPF corner frequency in quality of the current drawn or injected into grid is used in the design of the filter. Control modifications also enables the converter to operate as a STATCOM. Average modeling technique is used to derive the model of the converter controlled by DC-OCC. Further, the non-linear average model is linearized using small signal analysis. The small signal model shows the presence of an inherent current loop and a proportional controller. Gain of the proportional controller is the effective resistance seen by the current loop. As the gain of the phase shifted current loop approaches 0.5, the closed poles of the inherent current loop crosses over to the right half of s-plane, causing instability in current loop. Design of voltage loop controller parameters is also presented in this thesis.
All of the above modifications are validated in simulations and experiments. Simulation and experimental results are presented in this thesis for converters in the range from 600 W to 2 kW.

Event :Seminar
Title : Modular Multilevel Converter Based Medium-Voltage Induction Motor Drive for Wide Speed Range
Speaker : Prof. Gautam Poddar
Date : 04/06/2017
Venue : C 241 MMCR, EE
Abstract: The talk deals with the issues of modular multilevel converter (MMC) based medium voltage (MV) induction motor drive for variable speed applications. Sub-module (SM) capacitors of conventional MMC suffer from large voltage fluctuations as the motor speed decreases with rated torque. Finally, it becomes impractical to operate the motor at zero frequency. A modified configuration of MMC based induction motor drive and its new control technique have been proposed to limit these voltage fluctuations at all frequencies of the drive with rated load torque. An experimental set-up has been built to test the proposed control technique. Experimental results show that MMC based drive can be industrially viable from 0Hz to 50Hz.
Speaker bio: Prof. Gautam Poddar graduated with an M.Tech Power Electronics and Drives Indian Institute of Technology, Kharagpur in 1994, and PhD Electrical Engineering Indian Institute of Science, Bangalore 2002. Prior to joining IIT Kharagpur, he has had significant experience in industrial power electronic systems research and development at CDAC, Thiruvananthapuram. His main areas of interest are high-power drives, multilevel converter topologies, and active power filters.

Event :Defence
Title : Analysis, Control and Applications of Active Phase Converters for Single-Phase Power Grids
Speaker : Anil K Adapa
Degree Registered: PhD
Advisor : Prof. Vinod John
Date : 04/06/2018
Venue : C 241 MMCR, EE
Abstract: Single-phase distribution is a preferred approach for supplying power to rural or remote locations and microgrids due to its lower infrastructure costs when compared to a three-phase grid. However, the unavailability of a three-phase power supply limits the application of three-phase induction motors in small industries in urban and rural areas, and in the agricultural sector. Increasingly, due to the modernization of the technologies used in agriculture and mechanization of production processes, the demand for electrical energy in single-phase distribution grids has grown considerably. Three-phase induction machines exhibit superior characteristics and have a higher power density compared to their single-phase counterparts. In this scenario, there is a need for phase converters to facilitate the use of three-phase induction machines in single-phase grids. This work is on reduced switch count active phase converter topologies that process only a fraction of the load power while maintaining power quality at the grid and load ends. An auxiliary capacitor assisted active phase converter (AC-APC) topology is shown to have reduced current stress in the switching devices and the dc-link capacitors. The proposed control structure allows asymmetric control of the three legs of the APC. It also facilitates independent selection of optimized components for higher order filters, to meet the distinct design requirements at the grid and load terminals of the APC. The control is implemented independently as two single-phase converter controllers that generate modulation signals for the APC with a shared leg, in a decoupled manner. A method for soft-starting an induction motor is incorporated in the APC. Moreover, the bidirectional power flow capability of the APC also facilitates injection of power to a single-phase grid. A common-mode filter and its design method are presented for the asymmetric APC, which significantly reduces the effect of common-mode voltage in the system. All the methods proposed in this work are validated on an experimental 5 kVA laboratory APC prototype.

Title : Unidirectional Single-stage High Frequency Link Inverter Topologies for Grid Integration of Renewables
Speaker : Anirban Pal
Advisor : Dr. Kaushik Basu
Date : 01/06/2018
Venue : C 241 MMCR, EE
Abstract: With the depletion of fossil fuel reserve and the present global warming scenario, power utilities all over the world are focusing on renewable and alternative energy source-based power generation. Power electronic converters are essential for efficient conversion, control and conditioning of the power from the renewable sources before connecting to the AC transmission grid. Commercially available state-of-the-art power electronic converters use three phase voltage source inverter (VSI) to obtain controllable magnitude and grid frequency AC output. The output of the VSI is connected to the grid through a three-phase line frequency transformer (LFT). Beside the voltage matching the LFT ensures safety by providing galvanic isolation, avoiding DC current injection into the grid and helps to reduce leakage current to conform with the standards. The LFT is bulky and expensive and increases footprint of the overall system. Also, the VSI is high frequency hard-switched resulting in reduced efficiency. High frequency link (HFL) based isolated converter topologies with attractive features like high power density, small and compact footprint, low system cost are becoming popular as an alternative solution to the LFT based state of the art converters. This work presents two new HFL based three phase inverter topologies. In both the topologies, the active switches are either line frequency switched or high frequency soft-switched for most part of the line cycle result in improved converter efficiency. The operation of the proposed topologies are verified in laboratory scale hardware prototypes.
Speaker bio: Anirban Pal received the M.E. degree in electrical engineering from the Indian Institute of Science, Bangalore, India in 2015. He is currently pursuing the Ph.D. degree at the Electrical Engineering Department, Indian Institute of Science, Bangalore, India. His research interests include general area of power electronics.

Title : Industry Trends in Variable Speed Drives
Speaker : Mr. Navaneeth Kumar N
Date : 31/05/2018
Venue : C 241 MMCR, EE
Abstract: The talk is intended to provide a better perspective to the students on how OEM's design and what factors influence the architecture. Variable speed drives have been in the market for long and has been undergoing constant changes to meet up with the market requirements, IEC standards and technological advancements in power, computational devices, diagnostics and communication interfaces. This session will focus on Market trends, different power stage architecture (Current sensing, Gate driver and Isolation) used in drive, functional safety in drives, EMC standards and application of WBG devices in drives.

Speaker bio: N. Navaneeth Kumar is a systems manager at Texas Instruments, where he is primarily responsible for definition and development of subsystem solutions for industrial motor control equipment’s like AC Inverters, Servo drives, Soft starters and other related equipment’s. He has extensive experience in high voltage - high power electronics, EMC, analog and mixed signal designs. He has system-level product design experience in drives, solar inverters, UPS, and protection relays. He was elected to TI’s technical ladder in 2017 for his contributions in the area of Motor drives. Prior to joining TI, he has held various roles in HCL technologies and Wipro Technologies. He has got 5+ years of European work exposure by developing products at various OEM’s. Navaneeth earned his bachelor of electronics and communication engineering from Bharathiar University, India and his master of science in electronic product development from Bolton University, UK.

Event : MSc(Engg)Colloquium
Title : Binaural source localization using subband reliability and interaural time difference patterns
Speaker : Girija Ramesan Karthik
Advisor : Dr. prasanta Kumar Ghosh
Date : 29/05/2018
Venue : C 241 MMCR, EE
Abstract: Machine localization of sound sources is necessary for a wide range of applications, including human-robot interaction, surveillance and hearing aids. Robot sound localization algorithms have been proposed using microphone arrays with varied number of microphones. Adding more microphones helps increase the localization performance as more spatial cues can be obtained based on the number and arrangement of the microphones.
However, humans have an incredible ability to accurately localize and attend to target sound sources even in adverse noise conditions. The perceptual organization of sounds in complex auditory scenes is done using various cues that help us group/segregate sounds. Among these, two major spatial cues are the Interaural time difference (ITD) and Interaural level/intensity difference(ILD/IID). An algorithm inspired by binaural localization of humans would extract these features from the input signals. Popular algorithms, for binaural source localization, model the distributions of ITD & ILD in each frequency subband (typically in the range of 80Hz-5kHz for speech source) using Gaussian Mixture Models (GMMs) and perform likelihood integration across the time-frequency plane to estimate the direction of arrival (DoA) of the sources. In this thesis, we show that the localization performance of a GMM based scheme varies across subbands. We propose a weighted subband likelihood scheme in order to exploit the subband reliability for localization. The weights are computed by applying a non-linear warping function on subband reliabilities. Source localization results demonstrate that the proposed weighted scheme performs better than uniformly weighing all subbands. In particular, the best set of weights closely correspond to the case of selecting only the most reliable subband.
We also propose a new binaural localization technique in which templates, that capture the direction-specific interaural time difference patterns, are used to localize sources. These templates are obtained using histograms of ITDs in each subband. DoA is estimated using a template matching scheme, which is experimentally found to perform better than the GMM based scheme. The concept of matching interaural time difference patterns is also extended to binaural localization of multiple speech sources.

Title : Deep Speech Revolution - Opportunities and Challenges
Speaker : Dr. Sriram Ganapathy
Date : 18/05/2018
Venue : C 241 MMCR, EE
Abstract: The automatic processing of speech signals has seen a recent revolution with various commercial engines approaching human performance on some of the speech tasks. The market research suggests that in a couple of years the number of voice-assisted devices would outnumber the number of people on earth. However, this area also has a rich history of five decades with many algorithms originally inspired by human speech processing. In this talk, I will attempt to look under the hood of some of these engines and how they achieve the performance with deep learning. In particular, I will provide a biased sampling of the techniques used in the field of automatic speech recognition and speaker verification technology. I will also highlight the exciting opportunities and research challenges for signal processing and machine learning engineers that will take us to the next wave of speech systems.
Speaker bio: Sriram Ganapathy is an Assistant Professor at the EE Dept., IISc. His research interests are in signal processing, machine learning, deep learning and neuroscience with applications to robust speech recognition, speech enhancement, speech coding and audio analytics including biometrics. Before joining as a faculty member at the Indian Institute of Science, he spent 4 years as a Research Staff Member at the IBM T.J. Watson Research Center in Yorktown Heights, NY, USA. He completed his Ph.D. with Prof. Hynek Hermansky, at Center for Language and Speech Processing, Dept. of ECE, Johns Hopkins University, USA. Dr. Ganapathy is a Senior Member of the IEEE Signal Processing Society.

Title : An Optical-flow Based Framework for Video Quality Assessment
Speaker : Dr. Sumohana S. Channappayya
Date : 11/05/2018
Venue : C 241 MMCR, EE
Abstract: This talk presents a framework for video quality assessment that is based on features derived from first and second order moments of optical-flow magnitude, and off-the-shelf image quality assessment algorithms. The efficacy of the optical-flow based features in discerning temporal distortions is demonstrated first. Spatial distortions are estimated using off-the-shelf image quality assessment algorithms. Overall video quality is estimated using a combination of these spatial and temporal quality estimates. This framework delivers state-of-the-art performance on popular video quality assessment databases in the both full-reference and no-reference video quality assessment settings. The talk will discuss several interesting features of this video quality assessment framework.
Speaker bio: Sumohana S. Channappayya received the B.E. degree in ECE from the University of Mysore, India, in 1998, the M.S. degree in Electrical Engineering from the Arizona State University, Tempe, in 2000, and the Ph.D. degree in Electrical and Computer Engineering from The University of Texas at Austin, in 2007. He is currently an Associate Professor of Electrical Engineering at IIT Hyderabad where he directs the Laboratory for Video and Image Analysis (LFOVIA). His research interests include image and video quality assessment, multimedia communication, and biomedical imaging. He was a recipient of the Excellence in Teaching Award at IIT Hyderabad in 2013 and 2017.

Title : Unsupervised Representation Learning For Noise Robust Speech Recognition
Speaker : Purvi Agrawal
Advisor : Dr. Sriram Ganapathy
Date : 11/05/2018
Venue : C 241 MMCR, EE
Abstract: Speech recognition software has dealt with its fair share of ridicule, what with the countless memes on its funny, and occasionally inappropriate, misinterpretations of certain voice commands. Intelligent personal assistants like Siri, Echo or Cortana find it all the more difficult to understand what we say in a noisy background, like in a car, at a restaurant, at the airport, etc. Yet, we pursue our efforts to build the perfect speech-recognition program that one day will listen to everything we say anywhere. In this quest, where else can we look for motivation other than our brain -- the perfect ‘program’ that listens and understands a voice in different environments, such as train, street, etc. ? And that is what we are pursuing to benefit the machine understanding of speech by designing a data-driven filtering approach. The modulation filtering approach to robust automatic speech recognition (ASR) is based on enhancing perceptually relevant regions of the modulation spectrum while suppressing the regions susceptible to noise. In our work, we explore the derivations of key modulations of speech signal purely from a data-driven perspective using deep generative models. The proposed approach motivated from neuroscience-based studies indicates significant improvement. We attempt to address the question whether big data can provide neuro-scientific cues for speech processing.
Speaker bio: Purvi Agrawal is a Ph.D. scholar in Learning and Extraction of Acoustic Patterns (LEAP) lab, working with Dr. Sriram Ganapathy, at the Electrical Engineering Dept., Indian Institute of Science (IISc), Bengaluru. Prior to joining IISc, she did her M.Tech. from Dhirubhai Ambani Institute of Information and Communication Technology (DA-IICT), Gandhinagar, in 2015. Her research interests are primarily into machine learning in signal processing, on incorporating machine learning aspects into speech signal processing and applications.

Event : Thesis Defence
Title : Fast High-Dimensional Filtering.
Speaker : Pravin Nair
Degree Registered: MSc (Engineering
Advisor : Dr. Kunal Narayan Chaudhury
Date : 08/05/2018
Venue : C 241 MMCR, EE

Title : Pushing the frontiers of Functional Neuroscience - touch with tech!
Speaker : Dr. Sharan Srinivasan
Date : 04/05/2018
Venue : Faculty Hall
Abstract: A brain injury of any nature or cause is like the hard disc of your computer crashing. You will have to reload all the softwares and then try and restore all the saved files and folders so that your computer is back to where it was before the crash! But the biggest challenges are... who has these softwares? How do we reload it, and in what sequence? How do we restore the saved files? How do we even know what all files were saved? What do we do if the original programs don’t take off or run well? Brain Injury Rehabilitation (BIR) has remained an enigma for most people, including the Neuroscience professionals and experts. The physical healing of brain injuries never really translates to an automatic and complete functional recovery of brain functions to the premorbid levels of functioning. The complexity and uniqueness of every individual's’ brain functions (it is like each of us having our own customised operating systems!) and the fact that most of it cannot be measured easily and in a standardised way make this task onerous and near impossible! This lack of clarity at all levels resulted in the creation of a large number of neurologically disabled individuals who are now a ‘liability’ to the family, community and country.
Speaker bio: Dr. Sharan Srinivasan is a Senior Neurosurgeon and the HOD of the Department of Neurosciences at the Bhagwan Mahaveer Jain Hospital, Bengaluru. He has performed over major 8000 brain and spine surgeries. His special interests are in Stereotactic & Functional neurosurgery (neurosurgeries for Movement Disorders, Spasticity, Pain, Psychiatry) and Stereotactic Radiosurgery. Along with Dr. Sanjiv (a neurologist), he runs an exclusive centre called Jain Institute of Movement Disorders & Stereotactic Neurosurgery. He was in the international news recently for performing the ‘FIRST’ Vo thalamotomy for a guitarist dystonia in India. The patient had 100% relief of his symptoms on the operating table. Media around the world (BBC, CNN, ToI, NDTV, Fox news, Dubai times, Singapore times, ANI, few FM radio stations in the US, etc.) covered this. He is also very passionate about Neuro rehabilitation. After seeing the severe and complex neurological disabilities that critical patients who survived severe stroke, head or spinal cord injuries had, he started NewRo, a dedicated state-of-the-art neuro rehab center in 2012 to help such patients achieve their maximum possible level of functioning potential. This is probably the first ‘stand-alone’ neuro rehabilitation center in India.

Title : Quantification and comparison of neural information at different scales of recording.
Speaker : Sidrat Tasawoor
Advisor : Dr. Supratim Ray
Date : 27/04/2018
Venue : C 241 MMCR, EE
Abstract: Neuronal activity from the brain can be recorded in several ways across multiple scales e.g. single unit spiking, local field potentials, electrocorticogram and electroencephalogram. These signals represent the information about the external environment at different neuronal population levels in the brain. Though they have been studied and characterised separately, a unified approach to analyse them together is rarer, but important for various reasons. To compare them with each other, a common measure has to be used. Shannon’s mutual information is a suitable measure in this case. I will talk about how we record simultaneously from the different scales, and how we can use information theory in a neuroscience context to draw inferences from these different neural responses.
Speaker bio: Sidrat is a 3rd year PhD student from IMI, IISc, and is working with Dr. Supratim Ray from Centre for Neuroscience and Dr. C S Seelamantula from Deptt. of Electrical Engineering. She received her B.Tech degree from NIT Srinagar, and M.E. from BITS-Pilani, Hyderabad. Her research interests include signal processing, vision and information theory.

Title : Renewable Energy Aggregation: A Coalitional Game Theory Approach
Speaker : Dr. Pratyush Chakraborty
Date : 20/04/2018
Venue : C 241 MMCR, EE
Abstract: Aggregation of geographically diverse renewable energy resources has significant potential to reduce the variability of renewable integration in a power grid. In this work, we focus on developing an aggregation strategy using cooperative game theory. It has been previously shown that using a joint power contract, the cooperative game of expected profit with renewable resources has a non-empty core. But we show that the game of realized profit has empty core using that contract. Here, we propose a new contract that generates non-empty core and thus stability for the game of realized profit. Next, we allocate the realized cost of variability. The allocation is based on cost causation principle and has a closed form expression. Finally we propose a strategy taken by the benevolent aggregator that will enable the resources to maintain their short-term term stability goals of aggregation as well as improve the long-term profit. Thus our strategy strongly promotes market aggregation of renewable energy.
Speaker bio: Pratyush Chakraborty received the B.E. degree in electrical engineering from Jadavpur University, India, in 2006. From 2006 to 2009, he worked in the Industrial Solution and Services Division in Siemens Limited, Kolkata, India. He received the M.Tech. degree in electrical engineering from Indian Institute of Technology, Bombay, India in 2011, the M.S. and PhD. Degrees in electrical and computer engineering from the University of Florida in 2013 and 2016 respectively. He is currently a Postdoctoral Research Scholar at the University of California, Berkeley. His research interests include game theory, mechanism design, and power system with deep renewable penetration.

Title : Algorithmic construction of stabilizing switching signals for switched systems
Speaker : Dr. Atreyee Kundu
Date : 03/04/2018
Venue : C 241 MMCR, EE
Abstract: Switched systems find wide applications in power systems and power electronics, automotive control, aircraft and air traffic control, network and congestion control, etc. It is well-known that a switched system does not necessarily inherit qualitative properties of its constituent subsystems. Consequently, characterization of stabilizing switching signals constitutes a key topic in the literature. In this talk I will describe an algorithm to construct stabilizing switching signals under pre-specified restrictions on admissible switches and dwell time on subsystems. The results employ Lyapunov functions and graph theoretic tools.
Speaker bio: Dr. Atreyee Kundu obtained her Ph D from IIT Bomaby in Systems and Control Engineering in 2015. She was then a postdoc, first at Control Systems Technology Group, Department of Mechanical Engineering, Eindhoven, University of Technology, The Netherlands for a year, and then at Centre for Research in Automatic Control of Nancy, University of Lorraine, France for 6 months. Since Jan 2017, she has been a Member of Technical Staff at the Robert Bosch Centre for CPS.

Title : Bernstein’s proof of the Weierstrass approximation theorem ( Click here for the Poster )
Speaker : Dr. Kunal Narayan Chaudhury
Date : 06/04/2018
Venue : C 241 MMCR, EE
Abstract: We will look at Bernstein’s constructive proof of the Weierstrass approximation theorem. In particular, we will see how the language of probability provides us with a clear insight into the proof mechanism. Time permitting, some generalizations of the theorem will also be discussed. The talk is based on the standard material.
Speaker bio: Dr. Kunal Narayan Chaudhury is currently an Assistant Professor in the Department of Electrical Engineering, Indian Institute of Science, Bangalore, India. His research interests include image processing, computer vision, and numerical optimization. He is a member of the Society for Industrial and Applied Mathematics and a Senior Editor of the SPIE Journal of Electronic Imaging.

Title : Towards Phonologically Motivated Sign Language Processing
Speaker : Dr. Mathew Magimai Doss from Idiap, Martigny and EPFL, Switzerland
Date : 03/04/2018
Venue : C 241 MMCR, EE
Abstract: Sign language is a mode of communication commonly employed by the Deaf community to communicate with each other as well as to communicate with the Hearing community. SMILE is a Swiss NSF funded Sinergia project involving sign language technologists and sign linguists that aims to develop a sign language learning system for Swiss German sign language (DSGS). In this talk, I will present recent developments in the SMILE project. More specifically, I will present (a) development of SMILE DSGS dataset and (b) development of a phonologically motivated sign language assessment approach using hidden Markov models and artificial neural networks, akin to articulatory feature-based speech processing.
Speaker bio: Dr. Mathew Magimai Doss received the Bachelor of Engineering (B.E.) in Instrumentation and Control Engineering from the University of Madras, India in 1996; the Master of Science (M.S.) by Research in Computer Science and Engineering from the Indian Institute of Technology, Madras, India in 1999; the PreDoctoral diploma and the Docteur dès Sciences (Ph.D.) from Ecole Polytechnique Fédérale de Lausanne (EPFL), Switzerland in 2000 and 2005, respectively. He was a postdoctoral fellow at International Computer Science Institute (ICSI), Berkeley, the USA from April 2006 until March 2007. Since April 2007, he has been working as a permanent researcher in the Speech and Audio Processing group at Idiap Research Institute, Martigny, Switzerland. He is also a lecturer at EPFL. He is a senior area editor of the IEEE Signal Processing Letters. His main research interest lies in signal processing, statistical pattern recognition, artificial neural networks and computational linguistics with applications to speech and audio processing and multimodal signal processing.

Title : Stability Analysis of Infinite Dimensional Systems Described by Block Laurent Operators
Speaker : Dr. Chirayu D. Athalye
Date : 28/03/2018
Venue : C 241 MMCR, EE
Abstract: In this talk, I will discuss the following two related problems. ℓ2-Stability and ℓ¥-Stability Analysis of Laurent Systems Infinite dimensional systems, i.e., dynamical systems defined over an infinite dimensional state-space – arise as a natural mathematical model for numerous engineering applications. In fact, any system that is modeled by partial differential/difference equations (a.k.a. n-D systems) or by delay-differential equations can be cast as an infinite dimensional dynamical system. One of the major questions concerning such systems is that of stability; however, owing to the infinite dimensionality of the state-space, extension of results on stability of finite dimensional systems is often not possible. Stability of the corresponding infinite dimensional autonomous system (a.k.a. zero input system) naturally provides a first step towards a rigorous study of stability of the general infinite dimensional state space equations with inputs. In this talk, we will discuss about the stability analysis of a particular type of infinite dimensional discrete autonomous systems which are closely related to discrete 2-D autonomous systems. We analyze ℓ2-stability and ℓ¥-stability of the following type of infinite dimensional discrete autonomous systems, which we refer as Laurent systems: xk+1(·) := A(s,s−1)xk(·), (1) where A(s,s−1) 2 Rn×n[s,s−1], s is the shift operator, and xk 2 R¥(Z,Rn), 8k 2 N. A system such as (1) appears in many practical scenarios, namely, time-relevant discrete 2-D systems, repetitive processes, discrete quantum mechanics where space-time is discrete, formation problemof infinite chains of kinematic points, etc. We provide necessary conditions, sufficient conditions, necessary and sufficient conditions for ℓ2-stability and ℓ¥-stability of systems given by Eq. (1). We also explain how the stability analysis of systems given by Eq. (1) is equivalent to the stability analysis of discrete 2-D autonomous systems. ℓ¥-Stability Analysis of n Infinite Chains of Kinematic Points As a result of continuously growing vehicular traffic and in order to reduce frequent occurrence of accidents, research interest has been increasing in the problemof formation control of a group of autonomous vehicles and unmanned vehicle technologies. Also, formation control of a group of drones/quadcopters has become a very active area of research because of applications in areas like remote monitoring, military surveillance, exploration activities, target tracking, etc. For ease of analysis, vehicles/ drones/quadcopters are usually modeled as point masses; and hence, they can be referred as kinematic points. Now, the above mentioned problems can be grouped under one heading “the formation problem of kinematic points”. One novel approach to study the formation problem of a large chain of kinematic points is to model it as an infinite chain of kinematic points. In this talk, we analyze the behavior of n infinite interacting chains of kinematic points (subjected to bounded perturbations) with the immediate-neighbors interaction dynamics, where kinematic points can move in a two dimensional plane. The immediate-neighbors interaction dynamics for n infinite interacting chains of kinematic points is a special type of continuous-time analogue of systems given by Eq. (1). We show that if an initial perturbation is bounded, then such an autonomous system will converge to an equilibrium point; moreover, under some additional conditions the autonomous system will in fact converge to the same equilibrium point in which it was before the perturbation. These results provide good insights in the problem of formation control of extremely large chains of vehicles/drones/quadcopterswith movements in a two dimensional plane.
Speaker bio: Chirayu Athalye received his B.E. degree in Electrical Engineering from Mumbai University; M.Tech and Ph.D. degrees from the EE-Department of IIT Bombay. Currently he is a postdoctoral fellow in the EE-Department of IISc. His research interests are infinite dimensional systems, multidimensional systems, stability analysis, optimal control, and convex optimization.

Title : Visual Speech Recognition
Speaker : Abhilash Jain
Advisor : Dr. G. N. Rathna
Date : 23/02/2018
Venue : C 241 MMCR, EE
Abstract: Visual Speech Recognition (VSR) deals with the task of extracting speech information from visual cues from a person’s face while speaking. Accurate lip segmentation and modelling are essential in any VSR algorithm for good feature extraction. However, lip modelling is a complicated task and is not very robust in natural conditions. A technique for limited vocabulary visual-only speech recognition that does not use lip modelling and can account for variation in length of spoken words is introduced. For visual feature extraction, Discrete Cosine Transform (DCT) and Local Binary Pattern (LBP) have been tested. An Error-Correcting Output Codes (ECOC) multi-class model using Support Vector Machine (SVM) binary learners is used for recognition and classification of words.
Speaker bio: Abhilash Jain is currently pursuing his MSc (Engg.) degree in Department of Electrical Engineering, under the guidance of Dr. Rathna G N. He is a member of the DSP Lab. He completed his B.E. (Hons.) from Birla Institute of Technology and Science, Pilani in 2014. His research interests include signal processing and machine learning.

Title : Deep Representations, Adversarial Learning and Domain Adaptation for Some Computer Vision Problems
Speaker : Prof. Rama Chellappa
Date : 22/02/2018
Venue : C 241 MMCR, EE
Abstract: Recent developments in deep representation-based methods for many computer vision problems have knocked down many research themes pursued over the last four decades. In this talk, I will discuss methods based on deep representations, adversarial learning and domain adaptation for designing robust computer vision systems with applications in unconstrained face and action verification and recognition, expression recognition, subject clustering and attribute extraction. The face and action recognition system being built at UMD is based on fusing multiple deep convolutional neural networks (DCNNs) trained using publicly available still and video face data sets. I will then discuss some new results on generative adversarial learning and domain adaptation for improving the robustness of the recognition system.
Speaker bio: Prof. Rama Chellappa is a Distinguished University Professor, a Minta Martin Professor of Engineering and Chair of the ECE department at the University of Maryland. His current research interests span many areas in image processing, computer vision, machine learning and pattern recognition. Prof. Chellappa is a recipient of an NSF Presidential Young Investigator Award and four IBM Faculty Development Awards. He received the K.S. Fu Prize from the International Association of Pattern Recognition (IAPR). He is a recipient of the Society, Technical Achievement and Meritorious Service Awards from the IEEE Signal Processing Society. He also received the Technical Achievement and Meritorious Service Awards from the IEEE Computer Society. Recently, he received the inaugural Leadership Award from the IEEE Biometrics Council. At UMD, he received college and university level recognitions for research, teaching, innovation and mentoring of undergraduate students. In 2010, he was recognized as an Outstanding ECE by Purdue University. He received the Distinguished Alumni Award from the Indian Institute of Science in 2016. Prof. Chellappa served as the Editor-in-Chief of PAMI. He is a Golden Core Member of the IEEE Computer Society, served as a Distinguished Lecturer of the IEEE Signal Processing Society and as the President of IEEE Biometrics Council. He is a Fellow of IEEE, IAPR, OSA, AAAS, ACM and AAAI and holds six patents.

Title : Introduction to graph convolutional neural networks
Speaker : Dr. Tijmen Tieleman
Date : 21/02/2018
Venue : C 241 MMCR, EE
Abstract: This talk introduces the audience to graph convolutional neural networks, which allow a neural network to take as input a graph, and to analyze a graph for e.g. classification or regression. The focus will be on the drug development and general chemistry application, where the graph describes a molecule, with a node for each atom and an edge for each molecular bond. We'll start with the basic concept, we'll go over success stories from recent literature, and we'll seek to engage the audience in a lively discussion around the many open questions about this type of neural networks, such as: should a bond also be represented as a node in the graph? Can the concepts of dilated convolution or frequency domain implementation be translated to the domain of graph convolutions? Can we formulate a generic convolution method that works well on both graphs and images? How can operations on such a highly variably shaped and sized input be efficiently batched on a GPU? What applications can be made using this technology? This talk and the intended discussion will be most beneficial for audience members who have a basic understanding of Deep Learning, preferably including an understanding of convolutional neural networks for image classification, and who are curious about adapting the core technologies of Deep Learning to new applications. However, an effort will be made to make it interesting also for those who come less than ideally prepared.
Speaker bio: Dr. Tijmen Tieleman graduated from Geoff Hinton's lab at the University of Toronto with a PhD in machine learning, in 2014. After obtaining his PhD, Dr. Tieleman joined minds.ai, and now serves as its CTO. His main interests, besides Deep Learning, are in algorithms, probability theory, and theory of mathematics & computer programming. Mrinal (Ishant) Haloi holds a BTech in ECE from IIT Guwahati with international publications in ECCV, IV, ICIP. He's also a contributor to Tensorflow (master) and OpenCV (master). This is a presentation by Mrinal Haloi (Ishant) and Tijmen Tieleman (co-developer of RMSprop algorithm), of minds.ai.

Title : UWB Type High Power Electromagnetic Radiating Systems as an Electromagnetic Weapon
Speaker : Bhosale Vijay Hiralal
Advisor : Dr Joy Thomas M
Date : 16/02/2018
Venue : C 241 MMCR, EE
Abstract: Use of VLSI circuits based electronic systems for compactness and faster operation is ever increasing. Such sensitive electronics can get easily affected functionally or physically by Intentional Electromagnetic Interference (IEMI) sources. Short duration Ultra Wide Band (UWB) type pulse is one of such intentionally generated EMI source. To test the electronic system’s susceptibility a UWB source of appropriate rating is required. The same high power UWB source can as well function as an electromagnetic weapon. The high power UWB pulse generation requires a high voltage pulsed power source called pulser along with a high bandwidth antenna. The pulser has an energy storage device followed by a fast discharge switch whose role is very important in the UWB system operation, as the switch performance parameters like rise time and dielectric recovery decide the intensity of radiated electric field and the system energy output. Most UWB systems developed worldwide have used pressurised dielectric gas as the switching medium in the pulser. In this work, gas at sub-atmospheric pressure has also been tried as the switching medium. For enhanced overall system energy output, energy per switching shot is enhanced by optimising switch breakdown voltage instead of improving the Pulse Repetition Rate (PRR) as attempted by previous researchers.

Overall, the work consists of design, analytical and numerical simulation studies along with experimental evaluation of various performance parameters of the pulser, the switch, impulse radiating antenna (IRA) and the UWB system as a whole. The system developed under this work is on par with similar systems developed worldwide and sometimes even exceed in some of the performance parameters like Figure of Merit (FOM) and PRR.

Speaker bio: BHOSALE VIJAY HIRALAL received his Bachelors and Masters degrees in Electrical Engineering (B. E. and M. E.) from the University of Pune, India in 1999 and 2001 respectively with the first university rank for both the degrees. He joined Electronics and Radar Development Establishment (LRDE), one of the premier research laboratories of the Defence Research and Development Organization (DRDO), India as Scientist in 2004. He has worked in the Electro-Magnetic Interference / Electro-Magnetic Compatibility (EMI/EMC) group of LRDE where he has actively participated in the development of a Nuclear Electro Magnetic Pulse (NEMP) simulator. He has been instrumental in the development of high power pulsed sources in the voltage regime of up to a MV. At LRDE the pulsed power EMP simulation facilities were set up by him for the radiated and conducted susceptibility evaluation of critical electronic systems to be deployed in various defence platforms. He has worked in the EMC analysis, evaluation and ruggedization of many subsystems and systems of military radars. For his contributions in the area of enabling technologies for high voltage pulsed power he was awarded with the National Science Day award 2010. He is a life member of Society of EMC engineers (India) and a member of the IEEE EMC Society. His current research interest areas include pulsed power systems for high power UWB systems, analysis and optimization of gaseous switching for pulsed power sources and analysis and evaluation of electronics susceptibility for various HPEM signals.

Title : GaN Power Devices: Characteristics, Design Considerations and Applications
Speaker : Salil Chellappan
Date : 18/01/2018
Venue : C 241 MMCR, EE
Abstract: In this presentation, the characteristics, design considerations and a typical application of GaN power devices will be discussed. The first half of the presentation will focus on the application related characteristics of GaN power devices. A comparative study of available device architectures (like enhancement and depletion mode), specifications and packaging vis-à-vis power conversion applications will be presented. The second half will focus more on a practical implementation using GaN devices – a “Zero Voltage Switched Interleaved Critical Conduction Mode Totem pole Bridgeless PFC”, that extends the switching frequency range to the MHz region. The background, implementation aspects and results will be presented.
Speaker bio: Salil Chellappan (Systems Manager – Power Delivery, Industrial Systems, Texas Instruments) in his present role drives the growth of TI’s business in Power Delivery end equipment sector by executing strategic reference designs and other collateral for release on TI website. Prior to this role, Salil was Lead Engineer in Power Design Services where he was responsible for developing customer driven power designs for the growth of TI’s business in India. He was elected to TI’s technical ladder in 2012 for his contributions in the area of power conversion. Salil has more than 27 years’ experience in power conversion and analog design that includes ten years in TI and the rest in various high-profile organizations like GE, Power Integrations, Lucent Technologies and Bharat Electronics. He has a Bachelor’s Degree in Electronics & Communication Engineering from Kerala University. He dedicates his spare time to urban farming and aquascaping.

Title : Least-Squares Registration of Point Sets over SE(d) using Closed-Form Projections
Speaker : Sk Miraj Ahmed
Advisor : Dr. Kunal Narayan Chaudhury
Date : 19/01/2018
Venue : C 241 MMCR, EE
Abstract: We consider the problem of aligning multiple point sets (dimension d ) by computing their relative motions at once. Typically, the computational pipeline for this problem has two phases: (I) finding point-to-point correspondences between overlapping scans, and (II) registration of the scans based on the correspondences. The focus of this thesis is on phase II. In particular, we work with a global registration model, where the scans are registered in one-shot using rotations and translations. We have shown the application of it in 3D reconstruction problem where d=3 for each point. We consider a least-squares formulation of global registration, where the variables are the transforms associated with the scans. The present novelty is that we reduce this intrinsically non-convex problem to an optimization over the positive semidefinite cone, where the objective is linear but the constraints are nevertheless non-convex. We propose to solve this using variable splitting and the alternating methods direction of multipliers (admm). Due to the linear objective and the structure of constraints, the admm sub-problems turn out to be projections with closed-form solutions. We empirically show that for appropriate parameter settings, the proposed solver has a large convergence basin and is stable under perturbations. This is in keeping with recent empirical results on the effectiveness of admm for non-convex problems (the convergence theory is still in its infancy though).
Speaker bio: Sk Miraj Ahmed received the BE degree in Electrical Engineering from Jadavpur University, Kolkata, India in 2015. He is currently doing his masters (MSc(Engg)) in Electrical Engineering from Indian Institute of Science, Bangalore, India. He is a IEEE student member. His main research interests include Computer vision and convex optimization.

Title : Introduction to MCMC method and applications
Speaker : Prof. Krishna B Athreya
Date : 12/01/2018
Venue : C 241 MMCR, EE
Abstract: A very useful result in probability theory as applied to the real world is the law of large numbers . It says that the sample mean of iid observations converges to the population mean in some sense. The CLT is a refinement of this. About half a century ago this method was extended to Markov chains and a new tool known as MCMC was born. In this talk we shall outline this method with an application. This is a Chalk and board talk.
Speaker bio: K. B. Athreya is a visiting professor in the Math department here at IISc.. He is an emeritus faculty in mathematics and statistics at Iowa State University, Ames, Iowa, USA. His areas of research are probability theory, stochastic processes and mathematical analysis. He enjoys teaching mathematics at all levels. Besides numerous research papers he has also written many popular articles on Mathematical topics.

Title : Computational Auditory Scene Analysis: Binaural Source Localization
Speaker : Karthik Girija Ramesan
Date : 05/01/2018
Venue : C 241 MMCR, EE
Abstract: Ever wondered how we hear what we hear? In this talk, we will discuss the ability of the human auditory system to perform auditory scene analysis (ASA). Humans have an incredible ability to accurately localize and attend to target sound sources even in adverse noise conditions. The perceptual organization of sounds in complex auditory scenes is done using various cues that help us group/segregate sounds. In particular, we will see how spatial cues can be used to perform this task. We will then see how the distributions of these cues can be modelled and used in various Computational Auditory Scene Analysis (CASA) systems to perform binaural source localization.
Speaker bio: Karthik is currently pursuing his MSc(Engg.) in SPIRE Lab, EE, IISc. He is working on Source Localization and Separation Algorithms which are inspired by human audition. Previously, he worked as an Associate System Engineer at IBM, Pune. He received his B.Tech in ECE from Amrita School of Engineering, Coimbatore, in 2013

Title : Performance of Transformers Connected to Wind Energy Plants under High-Frequency Fast Transient Voltages
Speaker : Shesha Jayaram
Date : 22/12/2017
Venue : C 241 MMCR, EE
Abstract: The increasing spread of renewable energy resources, both in their types and numbers, along with the ever-growing demand in electricity have brought about many technical and operational challenges to power grids. Additionally, electrification of vehicles is expected to bring a significant pressure on the grid due to the sizable amount of electricity required to charge electric vehicles.Different aspects of above transitions; such as, system stability, mitigation of low frequency harmonics, optimization of use of renewable sources to allow greater demand participation, better usage of assets and active network management and increased reliability have been studied. However, the impact of new technologies and changing loads on the electrical insulation system of power grid components have not been studied to the same extent. Studies related to the performance of the electrical insulation systems and their integrity in the future grid that is enriched with penetration of renewable energy sources and plug-in electric vehiclesfrom the perspective of developing suitable diagnostics is therefore essential.
One of the factors related to the reported problems is the presence of high-frequency high-dV/dt voltages that are created by switching operations in wind energy plants. This talk presents research undertaken to investigate the performance of wind turbine step-up transformers under distorted voltages, with consideration of internal resonance phenomena and high-frequency effects.
Speaker bio: Dr. Shesha Jayaram is a Professor and University Research Chair in the Department of Electrical and Computer Engineering, and Director of the High Voltage Engineering Laboratory at the University of Waterloo, Canada. Dr. Jayaram’s research emphasizes solution-based outputs and focuses on four main areas: high voltage engineering and insulation diagnostics, high voltage engineering applied to environment, nanocomposite materials, and pulse power applications. She has published extensively, and holds many patents in HV applications to biotechnology and nanotechnology. She has been an active member of the IEEE Dielectrics and Electrical Insulation and Industry Applications Societies, and the Electrostatic Society of America. She is a Fellow of the IEEE, and a registered Professional Engineer in the Province of Ontario, Canada.

Title : Positive trigonometric polynomials: Application to spectral super-resolution
Speaker : Kumar Vijay Mishra
Date : 20/12/2017
Venue : C 241 MMCR, EE
Abstract: We address the problem of super-resolution frequency recovery using prior knowledge of the structure of a spectrally sparse, undersampled signal with frequencies lying anywhere in the continuous domain [0, 1]. We devise a general semidefinite program (SDP) to recover these frequencies using theories of positive trigonometric polynomials (PTP). Our theoretical analysis shows that given sufficient prior information, perfect signal reconstruction is possible using signal samples no more than thrice the number of signal frequencies. We extend our PTP formulations to solve an open problem on the formulation of an equivalent positive semidefinite program for atomic norm minimization in recovering signals with d-dimensional (d greater than or equal to 2) off-thegrid frequencies. Finally, we combine SDP and l1-minimization to develop fast versions of our algorithms.
Speaker bio: Dr. Mishra obtained Ph.D. in electrical engineering and M.S. in mathematics from The University of Iowa in 2015 and M.S. in electrical engineering from Colorado State University in 2012 while working on NASA GPM-GV mission weather radars. He received B. Tech. summa cum laude (Hons., Gold Medal) in electronics and communication engineering from the National Institute of Technology, Hamirpur in 2003. During 2003-2007, he worked as a research scientist at eLectronics and Radar Development Establishment (LRDE), Defence Research and Development Organization, Bengaluru on military surveillance radars. He was a research intern at Mitsubishi Electric Research Labs (Cambridge) and Qualcomm (San Jose) in 2015. During 2015-2017, he was Andrew and Erna Finci Viterbi and Lady Davis postdoctoral fellow at the Faculty of Electrical Engineering, Technion - Israel Institute of Technology. He is the recipient of Royal Meteorological Society Quarterly Journal Editor's Prize (2017), Lady Davis Fellowship (2016-17), Andrew and Erna Finci Viterbi Fellowship (twice awarded, 2015, 2016), Technion EE Excellent Undergraduate Adviser Award (2017), DRDO LRDE Scientist of the Year Award (2006), Cornell Base-of-Pyramid Narrative Competition Prize (2009), Altera Forum Guru Challenge Winner (2008), and NITH Best Student Award (2003). His research interests include radar systems-theory-and-hardware, signal processing, radar polarimetry, and electromagnetics. His recent research is focused on radar applications in sub-Nyquist processing, spectrum sharing, autonomous vehicles, MIMO, cognitive radars and deep learning.

Title : Control of Power Electronics Systems using Predictive Switching Sequences and Switching Transitions (poster)
Speaker : Prof. Sudip Majumder
Date : 22/12/2017
Venue : C 241 MMCR, EE
Abstract: This presentation provides a fundamentally different perspective to the control of solid-state semiconductor-device-based switching power-electronic systems (PESs). It is based on controlling the time evolution of the feasible switching sequences and controlling the switching transitions of PESs. The former - that is, the switching-sequence-based control (SBC) - yields rapid response under transient condition, optimal equilibrium response, and yields seamless transition between the two dynamical modes. Further, by enabling integration of modulation and control, SBC precludes the need for ad-hoc offline modulation synthesis. In other words, an optimal switching sequence for a PES is generated dynamically without the need for prior determination of a modulation scheme (which generates a pre-determined switching sequence) as evident in most conventional approaches. This presentation will provide the mechanism to carry out SBC synthesis and how it leads to multi-scale optimality leading to enhanced PES performance. Subsequent to the outline of SBC, the presentation will focus on switching-transition control (STC). The primary objective of STC is to demonstrate how key PES parameters including dv/dt and di/dt stress, switching loss, electromagnetic noise emission can be controlled dynamically by modulating the dynamics of the power semiconductor devices. Both electrical and newly developed optical-control mechanisms to achieve STC will be briefly outlined. Finally, envisioned mechanism for monolithic integration of SBC and STC will be illustrated. This presentation will demonstrate, along with results, multiple practical applications (currently of high priority in the power/energy space) where the radically new control concepts make a tangible and substantive difference.
Speaker bio: Sudip K. Mazumder received his Ph.D. degree from Virginia Tech in 2001. He is a Professor and the Director of Laboratory for Energy and Switching-Electronics Systems in the Department of Electrical and Computer Engineering at the University of Illinois, Chicago (UIC). He also serves as the President of the small business NextWatt LLC. He has over 25 years of professional experience and has held R&D and design positions in leading industrial organizations and has served as a Technical Consultant for several industries. His current areas of interests are switching-sequence and switching-transition based control of power-electronics systems and interactive-power networks; power electronics for renewable energy, micro/smart grids, energy storage; wide-bandgap (GaN/SiC) power electronics; and optically-triggered wide-bandgap power semiconductor devices. His research has attracted about 45 sponsored-research projects from leading federal agencies and industries, and yielded over 200 peer-reviewed publications in prestigious tier-one international journals and conferences, 10 patents, 10 book chapters and 1 (pending) book, and 85 invited/plenary/keynote lectures and presentations. He has guided/guiding 11 post-doctoral researchers and 15 Ph.D. and 10 M.S. students. He is the recipient of UIC’s Inventor of the Year Award (2014), University of Illinois’ University Scholar Award – university’s highest award (2013), IEEE International Future Energy Challenge Award (2005), ONR Young Investigator Award (2005), NSF CAREER Award (2003), and IEEE PELS Transaction Paper Award (2002). In 2016, he was elevated to the rank of an IEEE Fellow and he was invited to serve as a Distinguished Lecturer for IEEE PELS beginning in 2016. He served/serving as the Guest Editor-in-Chief/Editor for IEEE PELS/IES Transactions between 2013-2014 and 2016-2017, as the first Editor-in-Chief for Advances in Power Electronics (2006-2009), and as an Associate Editor for IEEE IES/PELS/TAES/TII/JESTPE Transactions (2003-/2009-/2008-/2016-/2016-). Currently, he serves as the Chair for IEEE PELS TC on Sustainable Energy Systems.

Title : Robust Learning of Classifier in presence of Label Noise
Speaker : Himanshu Kumar
Advisor : Prof. P S Sastry
Date : 15/12/2017
Venue : C 241 MMCR, EE
Abstract: Pattern Recognition(PR) is concerned with discovery of regularities in the data and using these for taking actions such as classifying the data into different categories. We focus on a type of PR problems, supervised classification learning, where a classifier is learnt given a labeled training data. In practice, labels in the training data are corrupted due to inadvertent errors, crowd sourcing, human biases etc. Performance of the many existing classifier learning algorithms is shown to be adversely affected under label noise. This problem is more relevant now-a-days as huge training data is required for training deep networks. We propose sufficient conditions on loss function, under which risk minimization is inherently tolerant to label noise for multiclass classfication. We also propose new robust losses which are easier to optimize compared to existing robust loss MAE(Mean Absolute Error).
Speaker bio: Himanshu Kumar is currently pursuing MSc(Engg) in Electrical Engineering. He is a member of the Learning Systems and Multi-Media Lab. He has completed his B.Tech. from Indian School of Mines, Dhanbad in Electronics and Communication Engineering. His research interests include Machine Learning and Pattern Recognition.

Title : Reinforcement Learning Techniques for Controlling Power Networks
Speaker : Dr. Anupama Kowli
Date : 07/12/2017
Venue : C 241 MMCR, EE
Abstract: As power grids transition towards increased reliance on renewable generation, energy storage and demand response resources, an effective control architecture is required to harness the full functionalities of these resources. A stumbling block to the development of such an architecture is the limited understanding of the uncertainty and dynamics that come into play when renewable generation and demand response are involved. This talk presents algorithms which allow control synthesis in settings wherein the precise distribution of the uncertainty and its temporal statistics are not known. These algorithms are based on recent developments in Markov decision theory, approximate dynamic programming and reinforcement learning. They impose minimal assumptions on the system model and allow the control to be "learned" based on the actual dynamics of the system. Furthermore, they can accommodate complex constraints on generation, storage and demand resources. Representative studies demonstrating applications of these algorithms to practical control problems in power systems are discussed.
Speaker bio: Anupama Kowli is an Assistant Professor at Indian Institute of Technology Bombay. She is a researcher in the area of electricity markets, energy economics, resource planning and power system operation and control. She received her Masters and Ph.D. from University of Illinois at Urbana Champaign in 2009 and 2013, respectively. Anupama was a visiting scholar at University of Florida. She interned as an energy consultant at KEMA Inc and as a control engineer at the Pacific Northwest National Laboratory.

Title : Consensus over Digraphs: Robustness and Reachable Sets
Speaker : Dr. Dwaipayan Mukherjee
Date : 08/12/2017
Venue : C 241 MMCR, EE
Abstract: In this talk, I will be discussing two aspects of consensus in multi-agent systems. Consensus is a widely researched topic in the domain of multi-agent systems. The central idea in consensus is to achieve agreement in the states of agents while these agents communicate over a directed or undirected topology. In graph theoretic terms, the agents are the nodes of the network while the edges connecting the nodes depict the flow of information among the agents. A lot of work has been done on consensus of networks where the communication links are bidirectional (represented by undirected graphs). However, when the information flows over a directed graph, the analysis is rendered difficult owing to a lack of symmetry in the interconnections. The matrices associated with the directed graph, such as the Adjacency and Laplacian are no longer symmetric in such cases either. In the first part of my talk, I will be considering such directed networks and looking at an edge version of the consensus protocol. It will be shown that this particular interpretation helps in analyzing the robustness of the network. Agents will be modeled as single integrators. The perturbation in our system appears in the form of uncertainty in edge weights. The system will be cast in the M-Δ form and a Nyquist criteria based bound on the stability of the same will be presented. A general result will be derived that is applicable to all digraphs having a globally reachable node. Subsequently, two special digraphs, the directed cycle and the directed acyclic graph will be considered, where the tolerable limit on the perturbation will be given a graph theoretic interpretation. The highlight of this study is that we shall use tools from control theory to obtain the stability bounds, instead of carrying out a spectral analysis of the Laplacian or other related matrices. Thereafter, a double integrator model for agents will be considered and a general consensus protocol presented, along with control theoretic tools for obtaining the bounds on perturbation. Finally, as a dual to the above problem, we shall look at the design problem, where the objective is to choose suitable edge weights that will ensure consensus of double integrators over a digraph. The second part of the talk will focus on the cycle digraph, which is at the heart of cyclic pursuit, and we shall obtain necessary and sufficient conditions for convergence of discrete time heterogeneous cyclic pursuit in both synchronous and asynchronous modes. We shall see how the set of points where the agents rendezvous may expand due to the heterogeneity in the gain of cyclic pursuit and explore the possibility of using negative gains. Finally, the talk will conclude by looking at another variant of discrete time cyclic pursuit- heterogeneous deviated cyclic pursuit. In this case, we shall discuss sufficient conditions of stability and show how the reachable set may also expand when the deviations of the agents are heterogeneous.
Speaker bio: Dwaipayan Mukherjee is currently a Post-doctoral fellow at the Faculty of Aerospace Engineering, Technion- Israel Institute of Technology. His research is funded by a fellowship of the Israel Council for Higher Education. He received his B. E. (2007) from Jadavpur University, Kolkata, in Electrical Engineering, and M.Tech. (2009) in Control Systems Engineering (Department of Electrical Engineering) from the Indian Institute of Technology, Kharagpur. In 2014, he defended his doctoral thesis titled ‘Cyclic Pursuit- Variants and Applications’ at the Indian Institute of Science, Bangalore, Dept. of Aerospace Engineering. His research interests include networked dynamic systems, co-operative control, cyber-physical systems, and control theory.

Title : Exploiting Covariance Structure in Sparse Recovery from Noisy Linear Measurements
Speaker : Praveenkumar Pokala
Advisor : Prof. Chandra Sekhar Seelamanthula
Date : 17/11/2017
Venue : C 241 MMCR, EE
Abstract: In this talk, we shall focus on the problem of sparse recovery from noisy linear measurements, based on the dictionary learned from the clean speech signals using k-SVD algorithm. In doing so, we explore various greedy approaches like orthogonal matching pursuit (OMP), covariance assisted matching pursuit (CAMP), and proposed formulation, affine subspace matching pursuit (ASMP). Next, we discuss the importance of selecting appropriate regularization function that takes into account the covariance structure of the observations. We demonstrate optimal regularization based on cross-validation within the framework of elastic-nets.
Speaker bio: Praveenkumar received M.Tech in signal processing from IIT Guwahati. Currently, he is a research scholar in Electrical Engineering Department, Indian Institute of Science, Bangalore, India. His research interests include sparse signal processing, speech processing, and machine learning.

Title : The impact of speaking rate on Acoustic to Articulatory Inversion (AAI)
Speaker : Illa Aravind
Advisor : Dr. Prasanta Kumar Ghosh
Date : 10/11/2017
Venue : C 241 MMCR, EE
Abstract: Acoustic-to-articulatory inversion (AAI) is the task of recovering the position of articulators used for speech production, namely tongue, jaw and lips from the acoustic representations of speech. AAI is often used for deriving production motivated features from speech which has a number of applications in automatic speech recognition, speaker verification, language learning and pronunciation evaluation. Increase or decrease in speech rate is achieved by controlling the articulatory movements, which in turn, changes the speech characteristics and hence, the AAI mapping. In this talk, the experimental results indicating the alteration of AAI performance by differences in speech rate will be presented.
Speaker bio: Illa Aravind is currently pursuing PhD in SPIRE lab, Electrical Engineering Department. He completed his Master's with specialization in signal processing from NIT, Calicut. His area of interest is machine learning with application to speech processing.

Title : Lab To Market (A link between academic research and commercial market place)
Speaker : Prof. S K Sinha
Date : 03/11/2017
Venue : C 241 MMCR, EE
Abstract: Academic institutions and Govt. funded R&D labs in India work on variety of research problems and keep generating new ideas year after year. While the main objective of academic institutions is to provide high quality manpower to the society, the research programs often result in ideas and solutions which may have commercial value. Laboratory prototypes are also build to prove the concept proposed. However, a vast majority of these ideas and prototypes remain confined to the laboratories and do not reach market. This is a well known issue – over the years I have myself listened to many intellectuals, industry experts and my colleagues IISc, talk about them at length. However, in this presentation, I intend to share with you my understanding of why these issues become the barriers between academia and industry and offer a possible solution.
Speaker bio: Prof. SK Sinha graduated in Electrical Engineering from Bihar College of Engineering (now NIT-Patna) in 1967. After a two year stint in Industry, he joined the same institution as a lecturer and worked there till 1977. He then came to the Indian Institute of Science (IISc), Bangalore, to pursue higher education and obtained ME degree in 1979 and PhD in 1984, both from the Department of Electrical Engineering. On completion of his doctoral program, he was offered a faculty position at EE, IISc, where he worked till 1995. His main area of research was in the domain of Parallel and Distributed Computing, targeted towards fail-safe systems. In 1995, he moved to the Centre for Electronics Design and Technology (CEDT, now renamed DESE) at IISc, where he set up the first Embedded System Laboratory in the country, specially organized for teaching the subject. Most of his R&D effort have been directed towards solving real life problems from industry and contributions have been made to the transportation sector, mining sector, steel rolling mills, lighting electronics and embedded computing systems. He retired from IISc in 2011, and after three years of preparation, he has incubated a technology company at IISc with a vison to ‘Generate Wealth from Academic Research’. His venture, named LAB TO MARKET INOVATIONS PVT LTD, is supported by IISc and has four serving faculty members of IISc as advisors.

Title : Consistency for Re-identification and Saliency for Explainability
Speaker : Dr. Abir Das
Date : 02/11/2017
Venue : C 241 MMCR, EE
Abstract: One of the fundamental goals of computer vision is to understand a scene. Towards this goal, we want the system to answer several questions - who, what, when, why, how much, etc. pertaining to the visual scene. Re-identifying persons over a network of cameras addresses questions involving the identity of the persons i.e., it deals with questions involving the word ‘who’. Similarly asking the question why a complex model gives rise to a particular decision enables one to make the models explainable and thus more trustworthy by making them more compatible with human reasoning.

Person re-identification is the task of identifying and monitoring people moving across a number of non-overlapping cameras. Several factors like significant changes in viewing angle, lighting, background clutter, and occlusion cause features to vary a lot from camera to camera. The first part of the talk will be about the following research questions about person re-identification. The first question is - Can we model the way features get transformed between cameras? Can we also learn the way feature ‘does not’ get transformed and tell if a image pair (from separate cameras) is coming from the same person or not? The similarity between the feature histograms and time series data motivated us to apply the principle of Dynamic Time Warping to study the transformation of features by warping the feature space. The warped space not only allowed us to model feasible transformation between pairs of instances of the same target, but also to separate them from the infeasible transformations between instances of different targets. Existing person re-identification methods are camera pairwise where the focus is on finding similarities of persons between pairs of cameras. While this works well for a 2 camera network, it introduces inconsistency of re-identification when a network consisting of 3 or more cameras are considered. The next part of the talk will address two important research questions. Can the results be made consistent? and Will re-identification performance be improved by enforcing consistency? We addressed the problem by posing re-identification as an optimization that minimizes the global cost of associating pairs of targets on the entire camera network constrained by a set of consistency criterion.

Supervised deep learning methods have already enjoyed enormous success in computer vision and language research and have the potential to revolutionize robotics. Yet, it remains largely unclear about how the system comes to a decision, how certain the model is about its decision, if and when it can be trusted or when it has to be corrected. Due to this opacity, it becomes worse when such models fail. The next part of the talk will address the ‘why’ question for a video description system. We explored a top-down approach to capture spatio-temporally salient regions corresponding to generated video descriptions for an off-the-shelf video-to-text generation system. The work is motivated by the need to explain the word generation mechanism. For example, for a generated video description ‘A woman is cutting a piece of meat’, the quest is to see if the word “woman” is generated because the model recognized a woman or merely because “A woman” is a likely way to start a sentence? The saliency is estimated by measuring the drop in word probabilities when only one small part of the input video is fed into the network.

The talk will be concluded with some insight into possible future directions leveraging on the strengths of explainable spatio-temporal saliency towards getting rational feedback from human agents and to use the feedback iteratively to get better and trustworthy models.

Speaker bio: Dr. Abir Das received his B.E. degree in Electrical Engineering from Jadavpur University, India in 2007. He received his M.S. and Ph.D. degrees in the same subject from University of California, Riverside, USA in 2013 and 2015 respectively. He is currently a post doctoral researcher at the Computer Science Department at Boston University, USA. His main research interests include multi-camera person re-identification and video summarization, end-to-end video description and activity detection as well as explainable AI using machine learning based methods.

Title : Industrial Alarm Systems: Overview and Challenges
Speaker : Dr. Sarasij Das
Date : 27/10/2017
Venue : C 241 MMCR, EE
Abstract: Alarm systems play critically important roles in the safe and efficient operation of various industrial sectors, such as power, manufacturing, and process industries. However, most of the existing industrial alarm systems perform poorly. This talk primarily focuses on the challenges faced by the industrial alarm systems. At first, an overview of industrial alarm systems will be presented.Industrial alarm systems often suffer from alarm overloading. Main causes responsible for alarm overloading will be discussed. The current research status of industrial alarm systems will be summarized.
Speaker bio: Dr. Sarasij Das is Assistant Professor in the department of Electrical Engineering, IISc.

Title : Operational Risk Metric for Dynamic Security Assessment of Renewable Generation
Speaker : Prof. VIjay Vittal
Date : 24/10/2017
Venue : C 241 MMCR, EE
Abstract: The objective of this talk is to explore the efficacy of applying risk-based security assessment (RBSA) to define reliability standards for electricity grids with high penetration of converter-interfaced generation. A novel approach to estimate the impact of transient instability is presented in this paper by modeling several important protection systems in the transient stability analysis. In addition, a probabilistic model is developed to capture the uncertainty of increased converter-interfaced renewable penetration. A synthetic test case is derived from a realistic power system to verify the proposed method. The simulation results show that RBSA not only provides significantly relaxed security limits, but also helps in identifying critical aspects of system reliability that are not possible using conventional deterministic methods.

Title : Interval Tree based Text Line Segmentation of Offline Handwritten Documents
Speaker : Shiva Kumar H R
Advisor : Prof. A. G. Ramakrishnan
Date : 13/10/2017
Venue : C 241 MMCR, EE
Abstract: Text line segmentation is an important pre-processing step in document image analysis such as word/character recognition, word spotting and text/image alignment. Line segmentation methods used for printed documents often fail on offline handwritten documents due to unconstrained writing. In handwritten documents, different lines could have different skew angles, and even along the same line the skew may vary. Ascenders and descenders of adjacent lines might overlap and sometimes even touch each other. All these make text line segmentation in offline handwritten documents a challenging task. This talk presents an elegant and unique algorithm for the segmentation of text-lines from handwritten documents exploiting interval tree (IT) data structure. We construct an IT by inserting the row-interval of each connected component (CC) into the tree. While inserting an interval, we recursively merge all intervals that have significant overlap into a single enclosing interval. Tall CCs, which may arise due to the touching of components from adjacent lines, are inserted into the tree after cutting if needed. Non-overlapping short components, which may include diacritical marks, are inserted into the closest intervals. Once all the CCs are inserted, the IT has one node for each segmented text-line and we do in-order tree traversal to get the lines in sorted order. The algorithm is efficient since each CC is processed only once in creating the IT and the time complexity of IT search/edit operations is of the order of the logarithm of the number of lines. Results on ICDAR-2013 Handwriting-Segmentation-Contest dataset (English, Greek, Bangla) show that our approach outperforms the state-of-the-art text-line segmentation methods tested on this dataset. Results on ICDAR-2009 and PBOK datasets (French, German, Kannada, Oriya) show that it also scales to these Indic and European languages.
Speaker bio: Shiva Kumar, H. R. received the B.E. degree in Computer Science from R. V. College of Engineering, Bangalore in 2003. Since then he has been working with IBM India Software Labs on various technologies like compiler development, JavaEE server development, MobileFirst server development, Bluemix and Watson. His work on open source JavaEE server - Apache Geronimo earned him the coveted Committer status in Apache Geronimo in 2007. The impact of his work with IBM clients fetched him the IBM Client Value Outstanding Technical Achievement Award in 2016. In parallel, he is pursuing his Ph.D. from Dept. of Electrical Engineering, Indian Institute of Science (IISc), Bangalore. As part of his research work, he has developed high performance optical character recognition (OCR) and Text to Speech (TTS) systems for Kannada and Tamil languages. The impact of these systems in making printed Kannada/Tamil books accessible to blind students won him awards at national level (Gandhian Young Technological Innovation Award 2015) and South Asia level (Manthan Award 2014 and 2015).

Title : Electromagnetic transient and phasor domain hybrid simulation
Speaker : Prof. Vijay Vittal
Date : 17/10/2017
Venue : C 241 MMCR, EE
Abstract: This presentation proposes an approach for hybrid time domain simulation that combines electromechanical transient stability analysis and electromagnetic transient analysis (time domain analysis). This capability will provide the ability to represent desired portions of the system in greater detail and allow for the analysis of phenomena that require attention to unbalance in phases, unsymmetrical faults, and devices that are represented on a single phase basis. While the presentation will demonstrate the hybrid method for the study of Fault Induced Delayed Voltage Recovery phenomena, the proposed method is general and applicable to a number of problems that require variable detail at different parts of the system such as Geomagnetically Induced Currents, HV, ACDC systems, inverter interfaced generation, and others. It will enable electromagnetic transient analysis utilizing the entire model of systems at the required locations.

Title : Design of a compulsator to drive a railgun used for Electro Magnetic Launch (EML) applications
Speaker : Apurva Kulkarni
Advisor : Dr. Joy Thomas M
Date : 06/10/2017
Venue : C 241 MMCR, EE
Abstract: Electromagnetic launch (EML) means accelerating the projectiles using electromagnetic energy instead of chemical explosives. Electromagnetic railgun is the EML device which is used to accelerate the projectiles up to the hyper-velocity range which is having applications in the field of defense. Also, in the absence of chemical explosives (especially on the naval ships), safety of the operator is ensured. A 20 kJ railgun is built and tested in Pulsed power and EMC lab of IISc Bangalore. Compulsator is one of the pulsed powered sources being used for the railgun. It is basically an ac generator specifically designed in order to maximize the current for a short duration. The scope of the current work is to design and develop a compulsator for the existing railgun. The seminar mainly covers the characteristic study and the design of a compulsator.
Speaker bio: Apurva Kulkarni is currently pursuing Ph.D in the department of electrical engineering, IISc, Bangalore. He has completed his Bachelor’s from NIT, Nagpur and Masters from VJTI Mumbai. His research interests include pulsed power engineering, design of rotating machines, electromagnetic launch.

Title : Transient Stability Analysis of an all Converter Interfaced Generation (CIG) WECC system
Speaker : Prof. Vijay Vittal
Date : 03/10/2017
Venue : C 241 MMCR, EE
Abstract: In this talk, transient stability analysis of an 18205 bus Western Electricity Coordinating Council (WECC) system has been carried out when all conventional sources have been replaced with converter interfaced generation (CIG). In this 100% CIG system, the only rotating machines directly connected to the network are wound rotor induction generator wind turbines and induction motor loads. The inertia contribution of these rotating devices is small. With close to zero inertia in the system, the dynamic performance of the system under different contingencies is examined and discussed. The analysis conducted demonstrates that while such a futuristic system can survive certain contingencies, well designed coordinated wide-area converter control action may have to be incorporated to enhance the reliability of the system.
Speaker bio: Vijay Vittal received the Ph.D. degree in electrical engineering from Iowa State University, Ames, IA, USA, in 1982. Dr. Vittal is a member of the U.S. National Academy of Engineering (NAE), the director of Power System Engineering Research Center (PSerc), and the Ira A. Fulton Chair Professor, Foundation Professor of Power System Engineering at Arizona State University. Dr. Vittal was the recipient of the Presidential Young Investigator (PYI) Award in 1985, the IEEE PES Outstanding Power Engineering Educator Award in 2000 and the IEEE Herman Halperin Electric Transmission and Distribution Award in 2013. From 2005 to 2011, he served as the Editor-in-Chief for the IEEE Transactions on Power Systems.

Title : Distribution Systems with High Photovoltaic (PV) Penetration- Modeling, Analysis and Main Findings
Speaker : Prof. Vijay Vittal
Date : 26/09/2017
Venue : C 241 MMCR, EE
Abstract: This talk will address topics related to the integration of a significant percentage of solar photo voltaic (PV) generation penetration in a 12.47 kV distribution feeder in Flagstaff, Arizona. Several different aspects of the project will be discussed and described. This will include steps to develop a model of the feeder for analysis and comparison with actual measurements, the types of analyses conducted with the developed model and the comparisons with measured data, new methods of anti-islanding methods for solar inverters, selection and choice of energy storage systems for the feeder and the impact of such storage in conjunction with the use of smart inverters on system performance, voltage regulation using smart inverters, impact of PV penetration on feeder protection and the need for improved protection coordination.
Speaker bio: Vijay Vittal received the Ph.D. degree in electrical engineering from Iowa State University, Ames, IA, USA, in 1982. Dr. Vittal is a member of the U.S. National Academy of Engineering (NAE), the director of Power System Engineering Research Center (PSerc), and the Ira A. Fulton Chair Professor, Foundation Professor of Power System Engineering at Arizona State University. Dr. Vittal was the recipient of the Presidential Young Investigator (PYI) Award in 1985, the IEEE PES Outstanding Power Engineering Educator Award in 2000 and the IEEE Herman Halperin Electric Transmission and Distribution Award in 2013. From 2005 to 2011, he served as the Editor-in-Chief for the IEEE Transactions on Power Systems.

Title : Total Electric Field due to an Electron Avalanche and its Coupling to Transmission Line Conductors
Speaker : Debasish Nath
Advisor : Prof. Udaya Kumar
Date : 15/09/2017
Venue : C 241 MMCR, EE
Abstract: Transmission of bulk electric power from the generating stations to the load centers can be carried out only through high voltages transmission lines. One of the main issues in the design and perhaps maintenance of extra and ultra-high voltage transmission system is the corona, a local electrical breakdown of air on the line conductors and hardware. In the early days, this interference was of concern only to radio and television receptions, however, with extensive use of wide frequency bands for modern applications, it has assumed prime importance.

The EMI due to the transmission line corona has been extensively studied and reliable empirical formulas have been proposed. The basis for all the earlier studies was the experimentally measured corona currents. Corona current was assumed to be injected into the conductor and on the other, the frequency range involved were not adequate for the modern-day applications. From the theoretical perspective, the coupling of the field produced by corona to the conductor was hardly investigated and the total field produced by the corona itself was not quantified. In order to address these serious lacunae, the present work was taken up and it can be considered as the first leap towards the correct picturization, as well as, quantification of the problem.

The field produced by the electron avalanche involves noticeable retardation effects. In the literature, only the field produced by arbitrarily moving point charge of fixed strength is available. On the contrary, the avalanche involves growing spherical electron cloud with trailing positive charge, which is almost stationary. Starting from the basics, an analytical expression for the total field due to an avalanche has been derived for the first time. Suitable validation has been provided through numerical simulation of electric field integral equation.

Indeed, corona discharge is a complex phenomenon having many distinctly different modes which differ in their visual, as well as, electrical characteristics. Innumerable electron avalanches contribute to the measured corona current with their space-charge acting as a moderator. Therefore, in order to model for the corona on conductors, an indirect approach based on linearity is proposed. An equivalent spatio-temporal dipole distribution was obtained to produce the measured current on the conductor. The general expression derived for the isolated avalanche is extended for this purpose.

Speaker bio: Debasish Nath is currently doing his PhD in High Voltage Lab under the guidance of Prof. Udaya Kumar. He has completed his Master's and Bachelor's from IISc, Bangalore and BESU, Shibpur respectively. His research interest is in Electrodynamics.

Title : Fast high-dimensional filtering using clustering
Speaker : Pravin Nair
Advisor : Dr. Kunal N. Chaudhury
Date : 08/09/2017
Venue : C 241 MMCR, EE
Abstract: Several algorithms in image processing involve spatio-range filtering of images and high-dimensional data derived from them. The exact computation of these so-called high-dimensional filters is challenging, especially for real-time processing of high-resolution images. We recently showed that a simple yet accurate approximation of high-dimensional filters can be obtained using a mix of clustering, fast convolutions and interpolation. The resulting algorithm is competitive with state-of-the-art methods and importantly comes with guaranteed error bounds. I will discuss the algorithm in detail in this talk. To emphasize the practical speed-up obtained using our fast algorithm, I will present a demo at the end.
Speaker bio: Pravin Nair is currently pursuing MSc(Engg) in electrical engineering. He is a member of the Laboratory for Imaging Sciences and Algorithms (LISA). He has completed his B.Tech. from Amrita Institute of Technology, Coimbatore. His research interests include signal and image processing.

Title : Introduction to Monte Carlo Markov Chain Methods.
Speaker : Prof Krishna B Athreya
Date : 01/09/2017
Venue : B 308, EE
Abstract: A very useful result in probability theory as applied to the real world is the law of large numbers. It says that the sample mean of iid observations converges to the population mean in some sense. The CLT is a refinement of this. About half a century ago this method was extended to Markov chains and a new tool known as MCMC was born. In this talk, we shall outline this method with an application. This is a Chalk and board talk.
Speaker bio: K. B. Athreya is a visiting professor in the Math dept here at IISc. He is an emeritus faculty in mathematics and statistics at Iowa State University, Ames, Iowa, USA. His areas of research are probability theory, stochastic processes and mathematical analysis. He enjoys teaching mathematics at all levels. Besides numerous research papers, he has also written many popular articles on Mathematical topics.

Title : Optimal Design of Line-frequency 12-pulse Transformer - Modeling Tools and Techniques
Speaker : Dr. Girish Kamath
Date : 28/08/2017
Venue : C 241 MMCR, EE
Abstract: Multi-pulse rectification is a well-known line-harmonic current reduction method used with medium to high power 3-phase rectifier type loads. It is rugged and economical because of its simplicity ? the circuit mainly consists of a line-frequency transformer and a set of rectifiers. Since the transformer is a significant cost driver, a cost-effective multi-pulse solution naturally calls for its optimal design. This poses several interesting challenges. In this talk, the speaker provides a practitioner?s perspective on how the design process has evolved to address these challenges. The first step in the design process is to realize that the transformer?s performance, cost, and size are significantly influenced not only by its electromagnetic properties but also by cooling and thermal factors. The talk will discuss key components of the optimal design process using three practical examples. The speaker will begin with the presentation of a novel 12-pulse auto-transformer for a 20kVA motor drive application. The concept of multi-pulse rectification and the novelty of the proposed auto-transformer rectifier circuit will be introduced. This will be followed by an EM-fields based method to estimate the winding losses in a 12-pulse 33kW transformer and end with the thermal model of a 6-pulse 25kW transformer. The trends and factors influencing design practices in the industry will be discussed. It is seen that a holistic design approach that considers the key physics namely, electromagnetics, flow and heat transfer and their interactions would naturally lead to an optimally designed product.
Speaker bio: Dr. Girish Kamath graduated with an MS degree in 1996 followed by a Ph.D. in Electrical Engineering from the University of Minnesota in 1998. He has been working in the Motor Drives and Plasma Cutting industries since then. He currently designs power electronics systems and controls for plasma cutting power supplies. His main areas of interest are multi-physics modeling approach to power supply component design, High Voltage circuits, Electromagnetic Compatibility and Digital Control.

Title : Grid Integration of Renewables in the UK
Speaker : Dr. Balarko Chaudhury
Date : 18/08/2017
Venue : C 241 MMCR, EE
Abstract: Decarbonizing the electric power sector is crucial towards meeting UK’s ambitious greenhouse gas emission reduction targets. In doing so, the network operators face several challenges due to increasing penetration of renewables (solar photovoltaic, wind power etc.) and electric vehicles into their networks. This talk would focus on two specific technical issues to be addressed in order to facilitate grid integration of renewables without compromising the security of supply. These are: 1) grid frequency regulation with low system inertia and 2) poor utilization of transmission assets with increased volatility in power transfers. In this talk, I would discuss the role of 1) fast demand response through thermostatic loads and 2) coordinated control of flexible AC transmission system (FACTS) devices and high voltage direct current (HVDC) systems (including DC grids) in addressing the above challenges. The complexity and value of deploying the above solutions for additional flexibility would be covered through representative case studies.
Speaker bio: Dr Balarko Chaudhuri is a Reader in Power Systems in the Department of Electrical and Electronic Engineering at Imperial College London. Before joining Imperial College as a Lecturer in 2006, he worked in General Electric Global Research. His research interests includes power systems stability, grid integration of renewable energy, wide-area control through HVDC/FACTS and demand response. He has been as an external consultant for Electric Power Research Institute (EPRI), USA, and National Grid, UK, and served on consultation committees responsible for recommendations to the UK regulator (Ofgem).
Dr Chaudhuri has published over 85 research papers in IEEE Transactions, IET Proceedings and leading international conferences. He has co-authored two books on Robust Control in Power Systems (Springer) and Multi-terminal Direct Current Grids (Wiley IEEE) and a book chapter on 'Integrating HVDC into AC Grid' in EPRI's HVDC Reference Book (2016 edition).
Dr Chaudhuri is a Fellow of the IET, Senior Member of IEEE and a member of Cigre. He serves as an Editor of IEEE Transactions on Smart Grid and an Associate Editor of IEEE Systems Journal and Elsevier Control Engineering Practice.

Title : Online Domain Adaptation through Known Features and Pro-active Learning
Speaker : Dr. Raghu Krishnapuram
Date : 18/08/2017
Venue : C 241 MMCR, EE
Abstract: Domain adaptation algorithms have gained much interest in recent years due to their empirical success. However, most of the papers in the literature assume the availability of a pool of unlabeled (or sometimes partially labeled) target domain data. In many practical situations, this is an unrealistic assumption. In this talk, we consider the problem of online domain adaptation where the target (test) data arrives one instance at a time. We assume the availability of a fixed budget for obtaining their labels, as well as a set of fallible (unreliable) oracles in addition to a reliable oracle. We propose an online domain adaptation algorithm that uses the budget judiciously by balancing the cost and reliability of the oracles to adapt the model in an incremental fashion. One of the advantages of the proposed algorithm is that it can seamlessly integrate previously unseen features during the adaptation process by using a probabilistic model. Experiments on several benchmark real-world datasets in the text domain as well as image domain empirically establish the efficacy of the proposed algorithm.
Speaker bio: Dr. Raghu Krishnapuram is currently Head, R&D and IP Cell, as well as Professor, Computer Science and Engineering, Ramaiah Institute of Technology, Bangalore. He was until recently Program Manager, Financial Services, Xerox Research Centre ? India. Earlier, he worked at IBM T J Watson Center, Yorktown Heights, New York, where he was a technical leader for cognitive computing research. From 2000 to 2013, Raghu held various leadership positions at IBM Research India. During the last 4 years of his tenure at IBM Research India, he served as Associate Director, where he led projects in the area of ?Knowledge, Information, and Smarter Planet Solutions?, with a particular focus on emerging markets. He also served as a relationship manager for IBM?s services divisions such as IBM Global Process Services and IBM Business Services. Raghu is an alumnus of IIT-Bombay. After graduating with a Ph.D. from Carnegie Mellon University in 1987, he worked in the academia in the US for 13 years, initially at the University of Missouri and later at Colorado School of Mines, where he was a full professor. Raghu has published about 170 papers in journals and conferences, many with a very high citation count. He has filed over 40 patent disclosures out of which 11 have been granted by the US Patent Office. Raghu has been recognized as a Master Inventor by IBM and has served on the Technology Council of the IBM Academy of Technology. He is also a Fellow of IEEE and the Indian National Academy of Engineers (INAE).

Title : Person of Interest Tracking in Camera Networks
Speaker : Shiva Kumar K A
Advisor : Dr. G N Rathna
Date : 11/08/2017
Venue : C 241 MMCR, EE
Abstract: Due to development in smart cameras, nowadays wide area camera networks are ubiquitous with many applications like surveillance, security, and sports analysis. In these applications, tracking plays an important role. Tracking is the process of following the targets (persons) continuously and consistently with the same label throughout the camera network. In general, targets could be any objects like persons, vehicles, animals etc.
In many surveillance applications, tracking a specific person among many persons or objects throughout the camera network is an important problem. A typical example is a famous American television series: "Person of Interest" (POI) in which an ultra-intelligent "The machine" is developed which uses all possible media (text, audio, video) to detect any possible crime event (perpetrator or victim). The machine can also track a specific person throughout the city. In this work, given a person (among many persons), the goal is to follow the person throughout the camera network (only cameras are used as sensors).
Tracking a person in camera network is a difficult task because of: Intra and inter-camera (re-identification) challenges. Intra-camera challenges include scale and illumination variation, occlusion. Inter-camera (re-identification) challenges include view/pose variation, background, high illumination variation, multiple feasible paths between cameras, similar appearance of persons, large blind gaps, tracking termination of the person.
In this talk, person re-identification (re-id) in camera networks will be discussed and the specific challenges involved in re-identification for on-line person tracking are identified. An on-line update scheme, specific to the person, is proposed for person re-id using metric learning algorithm. Also, a distributed person of interest tracking algorithm will be explained. The important aspects of our algorithm are i) Spatio-temporal constraints which are used to reduce the number of potential tracks during re-identification stage. ii) Learning a generic distance metric between each pair of cameras during off-line, which is fine-tuned to the interested person while the person is being tracked by a camera. iii) During re-identification, cameras share information with neighbors and perform distributed minimum, to accurately re-identify the person. The proposed algorithm is easy, simple to implement.
Speaker bio: Shiva Kumar K A is currently doing his Ph.D. in CVAI lab under the supervision of Prof. K R Ramakrishnan (retd) and G N Rathna. He has completed his Master's from NIT, Warangal. His research interests include object tracking in camera networks.

Title : Total Electric Field due to an Electron Avalanche and its Coupling to Transmission Line Conductors.
Speaker : Debasish Nath
Advisor : Prof. Udaya Kumar
Date : 04/08/2017
Venue : C 241 MMCR, EE
Abstract: Transmission of bulk electric power from the generating stations to the load centers can be carried out only through high voltages transmission lines. One of the main issues in the design and perhaps maintenance of extra and ultra-high voltage transmission system is the corona, a local electrical breakdown of air on the line conductors and hardware. In the early days, this interference was of concern only to radio and television receptions, however, with extensive use of wide frequency bands for modern applications, it has assumed prime importance.

Using the above, the means of induction, spatial decay rate of corona current in the close range, its propagation mode and field produced by both avalanche/equivalent dipole and that due to induced current in the conductor, have all been investigated and quantified.
Speaker bio: Debasish Nath is currently doing his PhD in High Voltage Lab under the guidance of Prof. Udaya Kumar. He has completed his Master?s and Bachelor?s from IISc, Bangalore and BESU, Shibpur respectively. His research interest is in Electrodynamics.

Title : Design and development of large vocabulary continuous speech recognition for Tamil
Speaker : Madhavaraj A
Advisor : Prof. A G Ramakrishnan
Date : 28/07/2017
Venue : C 241 MMCR, EE
Abstract: In the last 40 years, we have seen steady progress in speech recognition. This progress can be attributed to two factors: (i) the use of hidden Markov model (HMM) in modeling the temporal variations in speech and (ii) the increasing computational power of modern computers. In the past 10 years alone, we have seen many low-cost commercial interactive speech recognition applications developed by Apple, Microsoft, Amazon, etc. Large vocabulary continuous speech recognition system (LVCSR) forms the heart of such applications. Researchers from these companies report a word recognition accuracy ranging from 90% to 95% for vocabulary size of about 1,50,000. It is also well known that automatic speech recognition (ASR) research is mainly focused on English and other European languages. It can be said that no substantial progress has been made for speech recognition for South Indian languages due to the unavailability of standard speech and text corpora. Our research focuses on overcoming these limitations to build a reasonably good Tamil LVCSR system. In this talk, I will describe the design and development of a DNN based large vocabulary, continuous speech recognition for Tamil. I will discuss in detail the steps involved in building acoustic model (AM), language model (LM) and pronunciation dictionary (or lexicon), which are then combined to build an end-to-end ASR system. I will also go through the practical implementation of the steps, where the components of the ASR system can be viewed as a graph in the form of a weighted finite state transducer, and how this graph can be used to recognize speech during the testing phase. Finally, I will talk about a specific problem in ASR namely, suppressing speaker variation, so as to improve the accuracy of a speaker-independent ASR system. We address this problem by building a deep neural network with gradient reversal layer to recognize both phones and speaker identity. The network is trained such that it minimizes phone recognition loss and maximizes speaker recognition loss. This way, we can assume that the intermediate layer’s output suppresses the speaker information and can be used as speaker-independent features for our final ASR system. This technique gives an improvement of 5% over the baseline model.
Speaker bio: Madhavaraj A is currently doing his PhD in MILE lab under the supervision of Prof. A G Ramakrishnan. He has completed his Master’s and Bachelor’s from IIT Guwahati and Anna University respectively. His research interests include Speech recognition and Machine learning. He has worked as an intern in Amazon on deep neural network based acoustic modeling for Echo speech recognition system.

Title : High-Performance, Energy-Efficient, EMI-Aware Mixed-Signal Dynamic Power Management Architectures
Speaker : Dr. Santanu kapat
Date : 18/07/2017
Venue : C 241 MMCR, EE
Abstract: Dynamic power management is a useful technique to optimize performance and efficiency in embedded systems, IoT devises, digital processors, display devices, wireless sensor networks, and many more, in which DC-DC converters are the key elements. However, the design using existing DC-DC converter architectures and pulse width modulation (PWM) techniques is confronted with the problem to simultaneously achieve high performance, high efficiency, and improved power spectrum.

This presentation introduces novel mixed-signal DPWM solutions which can improve transient response, efficiency, and power density over a wide operating range along with the provision of custom harmonic reduction without considerable performance and efficiency impacts. A discrete-time framework is introduced for analysis and design of stable digital controllers with fast response and high efficiency. Further, new DC-DC converter architectures are proposed to achieve ultra-fast transient response under dynamic voltage scaling.
Speaker bio: Santanu Kapat received the M.Tech. and Ph.D. degrees in Electrical Engineering from the IIT Kharagpur, India, in 2006 and 2010, respectively. From 2009 to 2010, he was a Visiting Scholar in the Department of Electrical and Computer Engineering, University of Illinois at Urbana-Champaign. From 2010 to 2011, he was a Research Engineer at GE Global Research, Bangalore, India. Since 2011, he has been with the Department of Electrical Engineering, IIT Kharagpur, where he is an Assistant Professor. His research interests include analysis and design of digital and nonlinear control in high-frequency DC-DC converters, and applications to dynamic voltage scaling, LED driving, DC nanogrid, bi-directional DC/AC converters for renewable energy applications. Dr. Kapat received the INSA Young Scientist Award and INAE Young Engineering Award in 2016. He has been serving as an Associate Editor for the IEEE TRANSACTIONS ON POWER ELECTRONICS since 2015. He is a Senior Member of IEEE.

Title : Flying Cars – Challenges and Propulsion Strategies
Speaker : Prof. Kaushik Rajashekara
Date : 17/07/2017
Venue : C 241 MMCR, EE
Abstract: The technology and interest in the flying cars is as old as airplanes and automobiles. However with the rapid advancement and commercialization of airplanes and automobiles, and with many technical challenges associated with flying cars, the interest in flying cars declined. In recent years, with the advances in technology of engines, electric motors, power converters, and communications, there is an increasing interest in flying vehicles and more electrification of these vehicles. Several companies are already developing these vehicles with the intent of commercialization. In this presentation, the history of flying cars including some of the on-going developments will be presented. The technical challenges, particularly related to lift and propulsion, and the problems related to making it a wide scale adoption will be discussed. The challenges, requirements of developing a hybrid or a pure electric flying car, and propulsion strategies for operating like an automobile, airplane with vertical take-off and landing will also be presented.
Speaker bio: Kaushik Rajashekara received his PhD (1984) degree in Electrical Engineering from Indian Institute of Science. From 1977-1984, he was a Senior Scientific Officer in CEDT, Indian Institute of Science. In 1989, he joined Delphi division of General Motors Corporation in Indianapolis, IN, USA as a staff project engineer. In Delphi and General Motors, he held various lead technical and managerial positions, and was a Technical Fellow and the chief scientist for developing electric machines, controllers, and power electronics systems for electric, hybrid, and fuel cell vehicle systems. In 2006, he joined Rolls-Royce Corporation as a Chief Technologist for More Electric architectures and power conversion/control technologies for aero, marine, defense, and energy applications. In August 2012, he joined as a Distinguished Professor of Engineering at the University of Texas at Dallas. Since September 2016, he is a Distinguished Professor of Engineering in University of Houston.

Prof. Rajashekara was elected as a Member of the National Academy of Engineering in 2012 for contributions to electric power conversion systems in transportation. He was also elected as 2015 Fellow of the National Academy of Inventors and 2013 Fellow of Indian National Academy of Engineering. He is the recipient of the IEEE Richard Harold Kaufmann award for outstanding contributions to the advancement of electrical systems in transportation; IEEE Industry Applications Society Outstanding Achievement Award, and EEE IAS Gerald Kliman award for contributions to the advancement of power conversion technologies through innovations and their applications to industry. He is a Fellow of IEEE and a Fellow of SAE International.

Prof. Rajashekara has published more than 150 papers in international journals and conferences, and has over 45 patents. He has given more than 150 invited presentations in international conferences and universities. He has co-authored one IEEE Press book on sensorless control of ac motor drives and contributed individual chapters to six published books. His research interests are in the area of power electronics, drives, transportation electrification, and energy management of microgram systems.

Go top

Title : Image Fusion using Optimization Framework
Speaker : Dr. Ketan Kotwal
Date : 14/07/2017
Venue : C 241 MMCR, EE
Abstract: The advancement in the image sensory technology has enabled us to /see/the objects beyond the visible range of human eyes. With the help of hyperspectral imaging systems, one can capture the scene response across nearly 200â€“250 spectral bands with a very fine bandwidth as low as 10 nm that reveal various features in the scene at different wavelengths. Research in hyperspectral imaging is growing due to its ability of providing robust, accurate, and multi-dimensional information. In this talk, I will first brief about hyperspectral imaging and its usefulness in various fields. As hyperspectral image contains far more bands than those can be displayed on a standard display device, one has to go through all 200+ bands to visualize or process the contents of the data. However, this process is time consuming, inconsistent, and unreliable. Image fusion provides effective solution to the problem of visualization of hyperspectral images by providing a single image representing most features of the image. I will describe how image fusion can be posed as an optimization problem. This fusion technique focuses on the desired characteristics of the output image, rather than those of input images. I will also talk about how the optimization-based framework can be useful in generalizing the fusion problem for image or non-image data.
Speaker bio: Ketan Kotwal obtained his MTech and PhD from Indian Institute of Technology Bombay (IIT Bombay) in 2012 with specialization in image processing. His PhD work develops new approaches for multi-band image visualization and their evaluation. His work on optimization-based fusion framework was selected in top 10 papers in the International Conference on Information Fusion in 2011. Dr. Kotwal is the recipient of Best Thesis Award by the Computer Society of India (CSI), as well as Excellence in Thesis Award by IIT Bombay. He is the co-author of research monograph â€œHyperspectral Image Fusionâ€ published by Springer, US. Dr. Kotwal was a part of multimedia group at Samsung Research India where he developed camera features for Samsung's flagship mobiles. For last 2 years, he has consulted to several companies regarding R&D problems in image processing, computer vision, and machine learning.

Go top

Title : A Story of Sub-Nyquist Sampling: Theory and Applications
Speaker : Mr. Sunil Rudresh
Advisor : Prof. Chandra Sekhar Seelamantula
Date : 30/06/2017
Venue : C 241 MMCR, EE
Abstract: Human beings perceive everything around us (speech, vision, touch, heat, etc.) in analog domain and let the machines process the sampled data in digital domain. Analog-to-digital and digital-to-analog converters (ADCs and DACs) act as bridges between the analog and digital worlds. The link between the two worlds is directed by the well known Shannon-Nyquist sampling theorem, which states that a bandlimited signal has to be sampled at least at the rate, which is twice its bandwidth for it to be reconstructed perfectly. In this talk, we ask specific questions such as the following: (a) Do we really need to sample signals at Nyquist rates? (b) Could we sample below the Nyquist rate (sub-Nyquist) and reconstruct signals perfectly? (c) For what class of signals sub-Nyquist sampling and perfect reconstruction is possible? In this connection, we consider a class of signals called finite-rate-of-innovation (FRI) signals, which can be sampled and reconstructed in the sub-Nyquist sampling regime and they need not be bandlimited. FRI signals are sampled within a kernel-based sampling framework, which is in line with the famous adage “Think analog, act digital” (M. Unser). We explore how sampling and reconstruction of FRI signals can be carried out. We demonstrate applications of the FRI sampling to ultrasound and RADAR imaging, where we achieve super-resolution by acquiring samples at sub-Nyquist rates.
Speaker bio: Sunil R. obtained his Bachelor of Engineering degree from the PES Institute of Technology (Department of Electronics and Communication Engineering), Bangalore, India. For two years, he worked as an Analog Design engineer in Cypress Semiconductor Corporation, India. Since August 2014, he is working in the Spectrum Lab, Department of Electrical Engineering, Indian Institute of Science towards his Ph.D. His research interests include sampling theory, in particular, finite-rate-of-innovation signal sampling, compressive sensing, and spectral estimation.

Go top

Title : Computational mechanisms underlying the control of simple and complex movements
Speaker : Prof. Aditya Murthy
Date : 23/06/2107
Venue : C 241 MMCR, EE
Abstract: A fundamental computation that our brains must perform is the conversion of a stimulus into a motor act. This operation implicitly requires decision-making and motor planning. Using fast eye movements called saccades that rapidly direct our gaze to points of interest in the visual scene we investigate the computational architecture underlying flexible motor planning and control. Using the insights from gained from these experiments we will describe results from recent experiments that provide insights into how the brain might coordinate and control simultaneous eye and hand movements.
Speaker bio: Prof A. Murthy obtained his bachelor's degree from St. Xavier's college, Mumbai and Master's degree from Bombay University. His doctoral training was with Dr. Allen Humphrey in the Department of Neurobiology at the University of Pittsburgh where I examined the neural mechanisms involved in the processing of motion in the visual system. During his postdoctoral training, he worked with Dr. Jeffrey Schall at Vanderbilt University studying the primate visuomotor system to more directly relate neural activity to psychological functions and behavior. Currently, he is the Chairperson at the Centre For Neuroscience, Indian Institute of Science.

Go top

Title : Future Distribution System Operation - Theory and Practice
Speaker : Dr. Yashodhan Agalgoankar
Date : 15/06/2017
Venue : C 241 MMCR, EE
Abstract: The distribution system infrastructure around the world is facing ever increasing challenges due to constant system changes. The rising penetration of intermittent renewable resources, ageing infrastructure, the proliferation of new loads such as electric vehicles, and demand response are some of the emerging issues, which system operators need to manage. Despite these changes, distribution companies are expected to maintain reliability, resiliency, and power quality. This necessitates research into various directions to achieve seamless operation of future low and medium voltage utility distribution networks. Typically, distribution system operators are facing challenges such as network voltage control and distribution system protection in the presence of distributed energy resources (DERs). In order to tackle these challenges, research into modelling of distribution systems including power electronic interfaced DERs and demand response is critical. These models need to consider inherent characteristics of power distribution networks such as unbalanced power flows. To alleviate impacts of DERs on the systems requires designing new operational techniques. This seminar will present a couple of representative Volt-Var operation techniques based on stochastic optimisation. Many distribution utilities are considering deployment of Advanced Distribution Management Systems (ADMS) to improve operational efficiency and resiliency of the networks. The advancement of research is necessary into various ADMS algorithms such as Outage Management Systems (OMS) and Fault Location Isolation and Service Restoration (FLISR). Evaluation of the practical and realistic long-term benefits of implementation of FLISR based on the reliability analysis is also a critical challenge for distribution system utilities. The research in distribution system protection and resiliency can be critical for utilities. Further, DERs in the form of microgrids can operate autonomously and assist in alleviating power grid disturbance and improving distribution system resiliency. Also, the comprehensive approach to Power system security is necessary considering the possibility of the cyber threats. The seminar will try to offer an insight into the modern distribution system operational challenges, DER integration challenges, distribution system cyber security challenges, and propose mathematical optimisation based strategies to achieve a seamless operation. This seminar will also discuss different future research directions, which I intend to undertake through collaborative work.
Speaker bio: Yashodhan P. Agalgaonkar received Bachelors in Electrical Engineering from Walchand College of Engineering, Shivaji University, India, in May 2003, an M.Sc. in Electrical Power Engineering from the Chalmers University of Technology, Gothenburg, Sweden, in February 2006, and a Ph.D. in Electrical Power Engineering from Imperial College London, London, U.K., in March 2014. He was a Postdoctoral Researcher at Imperial College, London, until January 2015. From April 2006 to October 2010, he was with Crompton Greaves, India, and with Converteam (now GE Power Conversion) Chennai, India, as a Senior Research Engineer. He worked at Converteam GmBh research Center in Berlin, Germany for 2.5 years. Since February 2015 he has been a Mid-career staff scientist in the Energy and Environment Division of Pacific Northwest National Laboratory, Richland, Washington, USA. He conducts research on diverse areas of power system operation for the US government Department of Energy programs.

Go top

Title : Data Driven Conversational Dialog
Speaker : Prof. Alan Black
Date : 02/06/2017
Venue : C 241 MMCR, EE
Abstract: Historically, successful spoken dialog systems were hand crafted sets of explicit rules that defined a set of paths through potential turns between a user and machine. Although often very successful, these are expensive to develop and require substantial work to expand to new domains. Recently there have been attempts to try to use databases of existing conversations to learn dialog structure thus making the build process easier. There are some successes here, but there are also significant problems. Finding the right data is hard, or may even be impossible, solutions to finding the "right" data has become a research goal in itself. This talk will present the current techniques in statistical and neural conversational models in dialog systems, their successes and their limitations as well as potential research directions to addressing these short comings.
Speaker bio: Alan W Black is a Professor in the Language Technologies Institute at Carnegie Mellon University. He was born in Edinburgh, Scotland, and did his bachelors in Coventry, England, and his masters and doctorate at the University of Edinburgh. Before joining the faculty at CMU in 1999, he worked in the Centre for Speech Technology Research at the University of Edinburgh, and before that at ATR in Japan. He is one of the principal authors of the free software Festival Speech Synthesis System, the FestVox voice building tools and CMU Flite, a small footprint speech synthesis engine, that is the basis for many research and commercial systems around the world. He also works in spoken dialog systems, the LetsGo Bus Information project and mobile speech-to-speech translation systems, and recently doing work in using speech processing techniques for unwritten languages. Prof Black was an elected member of ISCA board (2007-2015). He has over 200 refereed publications and is one of the highest cited authors in his field.

Go top

Title : Defining and Enabling Resiliency of the Electric Grid
Speaker : Anurag K Srivastava
Date : 01/06/2017
Venue : C 241 MMCR, EE
Abstract: Keeping the power on to critical facilities such as hospitals and fire department during extreme weather events, cyber events and other electric grid disruptions is essential. Microgrids improve the reliability of the critical loads in natural disasters and grid disturbances. With additional planning and design, microgrid can also help to restore critical loads outside microgrid and hence increase the system resiliency. There is a need for formal metrics to quantify resiliency of the different distribution system, or different configurations of the same network. This talk presents a tool to study the cyber-physical resiliency of the microgrid for planning phase and operational phase. The microgrid resiliency metric is formulated based on graph theoretic metrics and power system constraints. The information from these two phases is provided to the operator to make informed and proactive decisions to ensure the resilient operation of the electric power system.
Speaker bio:

Go top

Title : Modeling of Distributed Energy Resources and Their Limiting Conditions
Speaker : Prof. Mahesh S. Illindala
Date : 29/05/2017
Venue : C 241 MMCR, EE
Abstract:
Speaker bio: Prof. Mahesh S Illindala completed his B Tech (Electrical) in 1995 from REC Calicut (now NIT Calicut). He obtained his MSc(Engg) from EE, IISc, in 1999. He then graduated with PhD in 2005 from Univ Wisconsin, Madison, USA. He worked in Caterpillar Inc. for 6 years, researching on electric drive train, UPS, PV and fuel cells. Since 2011, he has been on the faculty of Department of Electrical and Computer Engg, Ohio State University. His present research interests are micro-grids and distributed energy resources. Dr Illindala won the Young Investigator Award from the Office of Naval Research in 2016, and the IAS Magazine Prize Article Award in 2016.

Go top

Title : Speeding Up of Dynamic Simulations of Large Power Systems
Speaker : Disha L Dinesha
Advisor : Gurunath Gurrala
Date : 26/05/2017
Venue : C 241 MMCR, EE
Abstract: Power grid is one of the key infrastructures, which significantly influences nation’s economic growth. Blackouts occur in power grid rarely but when they happen huge economic losses and social distress will occur. Preventing blackouts is very important in order to avoid huge losses. The power system undergoes several phases before a complete blackout occurs whose duration varies from few seconds to several hours. Identifying the unfolding cascading events in the initial phase beforehand for predicting the blackout behaviour is very important. This requires faster than real time simulation of large power systems. This talk discusses various stages of blackout, the timelines involved, the simulation requirements and few approaches for speeding up of the dynamic simulations of large power grids for predicting cascading events.
Speaker bio: Disha L D received her B.E. in Electrical and Electronics Engineering from R.V.College of Engineering and has a work experience of 2 years. She is currently working towards her MSc(Engg) degree in the Department of Electrical Engineering at Indian Institute of Science. Her research interests are in Power System Dynamics and Stability Analysis.

Go top

Title : Enhancement of low resolution document images for improved OCR recognition
Speaker : Ram Krishna Pandey
Advisor : Prof. A.G. Ramakrishnan
Date : 19/05/2017
Venue : C 241 MMCR, EE
Abstract: Recognition of document images has important applications in restoring old and classical texts. The problem involves quality improvement before passing it to a properly trained OCR to get accurate recognition of the text. The image enhancement and quality improvement constitute important steps as subsequent recognition depends upon the quality of the input image. There are scenarios when high-resolution images are not available and our experiments show that the OCR accuracy reduces significantly with the decrease in the spatial resolution of document images. Thus the only option is to improve the resolution of such document images. The goal is to construct a high-resolution image, given a single low-resolution binary image, which constitutes the problem of single image super-resolution. Most of the previous work in super-resolution deal with natural images which have more information content than the document images. To solve this problem of document image super-resolution, we have used convolution neural network (CNN) to learn a function which maps low resolution patches to high-resolution patches. We experiment with different number of layers in the CNN, settings of weight parameters, learning strategies and non-linear functions to build a fast end-to-end framework for document image super-resolution. We have investigated various architectures with different complexities and obtained a novel CNN based model which can improve the quality of document images in terms of PSNR, perceptual quality and OCR character and word level accuracy.
Speaker bio: Ram Krishna Pandey received his B.Tech. in Computer Science and Engineering from GKV Hardwar in 2012 and M.Tech. in Computer Science and Engineering from IIIT Bhubaneswar in 2014. He is currently working toward the PhD degree in the Department of Electrical Engineering at Indian Institute of Science, Bangalore, India. His research interests are in image processing, machine learning and document image analysis.

Title : Grid-tied Inverters for Renewable Energy Applications
Speaker : Dr. Deepak Somayajula
Date : 08/05/2017
Venue : C 241 MMCR, EE
Abstract: Distributed Energy Resources (DERs) on the distribution grid can cause many power quality and reliability problems like voltage sags, swells, real and reactive power imbalances. Such active and reactive imbalances can be compensated with the help of grid-tied shunt and series filters which will act as power conditioners. The series and shunt active filters are back to back inverters which can compensate for voltage sags/swells and active/reactive power imbalances. However, both the filters need active and reactive power support from an additional source apart from the dc-link capacitor. It is observed that the ultra-capacitor(UCAP) energy storage integration is ideally suited for providing good active power support for the series filter which compensates for voltage sags and swells. And UCAP integration helps the shunt active filter in providing active/reactive power support to the distribution grid to handle the intermittencies due to renewable energy sources. A brief discussion on the benefits of installing grid-tied solar panel level inverters will also be presented.
Speaker bio: Deepak Somayajula received his BS degree in Electrical Engineering from Pondicherry University (India) in 2005. He worked as a software engineer from 2005 to 2007. He received his MS and PhD from Missouri S & T in 2009 and 2014 respectively where his research focus was on hardware integration of UCAP based energy storage into the distribution grid. In 2014 he started working as a Postdoctoral Research Associate at UNC – Charlotte in collaboration with SineWatts Inc. on Tier 0 of Department of Energy’s prestigious SunShot Incubator Program. In 2015 he started working with SineWatts Inc. on the Tier 1 of the SunShot Incubator program where he is currently working on the development/field deployment of grid-tied solar panel level inverters.

Go top

Title : The Riesz Transform - A New Tool for Spectro-Temporal Analysis of Speech Signals
Speaker : Jitendra Kumar Dhiman
Advisor : Prof. Chandra Sekhar Seelamantula
Date : 05/05/2017
Venue : C 241 MMCR, EE
Abstract: Speech signals feature a rich time-varying spectral content which makes their analysis a challenging problem in signal processing. Developing methods for accurate speech analysis has direct impact on applications such as speech synthesis, speaker recognition, speech recognition, and voice morphing etc. A widely used tool to visualize the time-varying spectral content is 2-D spectrogram. By making observations on structured 2-D patterns in the spectrograms, we propose modeling of them using 2-D amplitude-modulated and frequency-modulated (AM-FM) sinusoids. In contrast to existing temporal/spectral methods for speech analysis, the proposed modeling allows spectro-temporal analysis of speech. We use Riesz transform, a 2-D extension of the Hilbert transform, for demodulation of narrow-band spectrograms. Interestingly, the 2-D AM and FM components obtained as a result of demodulation have potential benefits for speech analysis. From the speech production prospective, the AM and FM components correspond to the vocal tract smooth envelope and excitation signal, respectively. Utilizing this insight, we will demonstrate the applicability of the proposed modeling for applications such as voiced/unvoiced separation, pitch tracking, speech synthesis, and de-noising.
Speaker bio: Jitendra Kumar Dhiman received his B.Tech. degree in Electronics and Telecommunication Engineering from the Institution of Electronics and Telecommunication Engineering, Delhi, India, in 2010 and M.Tech degree in Signal processing from Indian Institute of Technology Hyderabad, India, in 2013. He is currently working toward the PhD degree in the Department of Electrical Engineering at Indian Institute of Science, Bangalore, India. His research interests include speech and audio signal processing.

Go top

Title : Get Your Next Glaucoma Diagnosis on a Smartphone
Speaker : Harish Kumar J. R
Advisor : Prof. Chandra Sekhar Seelamantula
Date : 21/04/2017
Venue : C 241 MMCR, EE
Abstract: We have developed a reliable and fully automated method for segmentation and outlining of the optic disc and optic cup using fundus images with relevant parameter for glaucoma prescreening. The segmentation is based on the notion of active disc, which comprises a pair of concentric discs as the template. The active disc is made to evolve from a normalized matched filtering based automatic initialization towards the boundary of the optic disc by minimizing a local disc energy function. Optimization is achieved using accelerated gradient descent, and Green's theorem. The initialization used for optic disc is also used to outline the optic cup region. We use the circular active disc to perform coarse segmentation and an elliptical active disc for fine segmentation. After segmentation of the optic disc and optic cup, we calculate the cup-to-disc ratio from the segmented optic disc and cup. The cup-to-disc ratio value is compared against the existing international classification of diseases rules to finally assist in diagnosing the progression of glaucoma by categorizing the condition as normal, mild, moderate, or severely glaucomatous. We have validated our glaucoma prescreening technique on publicly available as well as locally obtained fundus image databases. The algorithm performance is compared vis-a-vis clinician outlining as the reference and for quantitative comparison, we have used Jaccard and Dice similarity measures. The tool is Java-based, repeatable, easy to use, provides quantitative analysis, and takes only few seconds per image. The software implementation can be used alongside desktops, laptops, and handheld fundus cameras. In addition, in keeping with the contemporary trend of developing smartphone-based eyecare solutions, we have developed iOS and Android-based Apps for real-time implementation of the proposed method.
Speaker bio: Harish Kumar J. R, received the B.E. degree from the Adichunchanagiri Institute of Technology, Kuvempu University, India, in 1998, with a specialization in Electrical and Electronics Engineering, and the M. Tech. degree from the Department of Electronics and Communication Engineering, Sri Jayachamarajendra College of Engineering, Mysuru, India in 2004. From then he is serving as an Assistant Professor in the Department of Electrical and Electronics Engineering, Manipal Institute of Technology, Manipal University, India. He is currently pursuing the Ph.D. degree (under QIP) with the Spectrum Laboratory, Department of Electrical Engineering, Indian Institute of Science, Bangalore. His research interest includes signal/image processing, medical imaging for health care applications, and bio-medical image analysis for automated disease diagnosis.

Go top

Title : New strategies for online signature verification based on Dynamic Time Warping Algorithm
Speaker : Dr. Suresh Sundaram
Date : 13/04/2017
Venue : C 241 MMCR, EE
Abstract: In recent times, owing to security reasons, the authentication of a person has become the need of the hour. A number of biometric traits have been considered for identification of a person based on their physical or behavioral characteristics. The handwritten signature is one such biometric that has been well accepted for authentication. In this talk, we first present an overview of the literature on systems pertaining to `online signature verification'. Thereafter, we discuss on our recent work based on extensions of Dynamic Time Warping (DTW). DTW is a popular matching algorithm that is used to find the similarity between two temporal sequences of varying lengths. The proposed systems are based on extraction of additional information from the warping path - obtained as a by-product from the DTW algorithm.
Speaker bio: Suresh Sundaram received the Ph.D. degree from the Department of Electrical Engineering, I. I. f Sc. in 2012. He was a Research Consultant with Hewlett Packard Research Labs, Bengaluru, from Oct 2012 to June 2013. Since July 2013, he has been serving as an Assistant Professor with the Department of Electronics and Electrical Engineering, IIT Guwahati. His research interests include handwriting recognition, biometrics, and document analysis.

Go top

Title : Polymeric Insulators for High Voltage Transmission Line
Speaker : Alok Ranjan Varma
Advisor : Dr. Subba Reddy B
Date : 31/03/2017
Venue : C 241 MMCR, EE
Abstract: In modern era of upcoming industrialization, with the high consumption of electric energy it is very much important to achieve the required demand in power sector for the uninterrupted operation. The continuous availability of electric power is a major factor to fulfill the requirement, which can be achieved by the efficient, safe and reliable power transmission system supported by an effective insulation system. The insulation in over head transmission line (OHTL) is provided by porcelain and glass insulators conventionally but now a days recent advancement in materials lead to polymeric or non ceramic insulating materials. These polymeric insulators are composed of Polydimethylsiloxane (PDMS) as base polymer with different fillers like silica or Alumina trihydrate (ATH) to improve their material properties. There are several advantages of using polymeric insulators like easy manufacturing process, light weight, better mechanical properties, better hydrophobicity, better short term pollution performance etc., but major disadvantage is the penetration of moisture, also being organic in nature it is sensitive to environment and degrades with time by itself. In the present work, efforts are made to understand the behaviour of polymeric insulators material degradation under different environmental conditions (including acid rain condition) and their electrical behaviour against tracking and erosion using Inclined plane tracking method and accelerated aging studies using rotating wheel and dip test arrangement. Some preliminary results are discussed.
Speaker bio: Alok Ranjan Verma born in Aligarh, Uttar Pradesh, India in 1991. He received his B.Tech (Electrical Engineering) from Aligarh Muslim University, Aligarh, Uttar Pradesh and M.E. (Electrical Engineering) from Indian Institute of Science, Bangalore, Karnataka in 2012 and 2014 respectively. He is currently working towards his Ph.D. in High Voltage Engineering from Indian Institute of Science, Bangalore , Karnataka. His areas of interest includes, High Voltage Engineering, Polymeric Insulators for Outdoor Applications, , Computational Electromagnetism, Numerical Techniques in Electrostatics, Lightning Induced disturbances, Over-voltages in power systems.

Go top

Title : Time Scales in Control of Wind Energy Systems
Speaker : Prof. D. Subbaram Naidu
Date : 23/03/2017
Venue : C 241 MMCR, EE
Abstract: An overview of the author's journey of research experiences in the field of Singular Perturbations and Time Scales (SPaTS) in Control Theory and Applications (CTA) from Indian Institute of Technology (IIT), Kharagpur to University of Minnesota is presented. The SPaTS methodologies focus on the analysis of decoupling of high-order dynamical systems with slow and fast phenomena and the synthesis (design) of controllers for slow and fast subsystems. The research covers both theory and applications to a wide spectrum of fields in engineering such as aerospace, electrical, mechanical, and in sciences such as biology and ecology with particular emphasis to wind energy conversion systems.
Speaker bio: Desineni “Subbaram” Naidu received MTech and PhD degrees in Electrical Engineering (Control Systems Engineering), from Indian Institute of Technology (IIT), Kharagpur. Dr. Naidu taught, visited and/or conducted research at IIT; Guidance and Control Division at NASA Langley Research Center; Old Domain University; Measurement and Control Engineering Research Center at Idaho State University; Center of Excellence in Advanced Flight Research at United States (US) Air Force Research Laboratory; Center of Excellence for Ships and Ocean Structures at Norwegian University of Science and Technology; Measurement and Control Laboratory at Swiss Federal Institute of Technology; Nantong University, China; the University of Western Australia in Perth, Center for Industrial and Applied Mathematics at the University of South Australia in Adelaide; Jiangsu College of Information Technology, Jiangsu, China; Center for Applied and Interdisciplinary Mathematics at East China Normal University, Shanghai, China; Institute of Systems Science, Academy of Mathematics and Systems Science, Chinese Academy of Sciences, Beijing, China; Shanghai Jiao-Tong University, Shanghai, China. Since August 2014, Professor Naidu has been with University of Minnesota Duluth as Minnesota Power Jack Rowe Endowed Chair. Professor Naidu received twice the Senior National Research Council Associateship award from the US National Academy of Sciences, and is an elected (Life) Fellow of the Institute of Electrical and Electronic Engineers (IEEE), USA and an elected Fellow of the World Innovation Foundation, UK. He has over 200 publications including 8 books. He has been on the editorial boards of several journals including the IEEE Transactions on Automatic Control and Optimal Control: Applications and Methods.

Go top

Title : Modeling, Control and Analysis of Wound Rotor Induction Machines
Speaker : Ramu Nair
Advisor : Prof. G. Narayanan
Date : 17/03/2017
Venue : C 241 MMCR, EE
Abstract: Wound Rotor Induction Machines (WRIM) are widely employed in wind energy conversion systems and in high power drives. Most of control techniques used in drives, are model based and hence machine parameter dependent. Therefore, to achieve better performance and control, accurate estimates of motor parameters are necessary. In this work, a method for studying parameter variation in WRIM is suggested and validated through experimentation.

Rotor position is inevitably required in the closed loop control of induction machines. A rotor position estimation technique based on Model Reference Adaptive Control will be presented with experimental results.

In wind energy conversion systems, the stator side is directly connected to grid, while the rotor side of WRIM is fed from a controlled power converter. Direct connection of WRIM stator to grid results in transfer of grid-voltage disturbances on to rotor side. A grid-voltage sag manifests as an over-voltage in rotor windings. A simulation study on the same is presented.

Speaker bio: Ramu Nair received the B.Tech. degree in Electrical and Electronics Engineering from Mar Athanasius College of Engineering, Ernakulam, India, in 2009 and M.Tech. degree in Energy Systems Engineering from Indian Institute of Technology Bombay, Mumbai, India, in 2012. He is currently working toward the Ph.D. degree in the Department of Electrical Engineering at Indian Institute of Science, Bangalore, India. His research interests include power electronics, drives and control systems.

Go top

Title : Understanding the role of brain oscillations in cortical processing
Speaker : Dr. Supratim Ray
Date : 10/03/2017
Venue : C 241 MMCR, EE
Abstract: Brain signals often show oscillations at different frequencies, which are tightly coupled to different behavioral states. We are interested in a high-frequency oscillation called “gamma” (30-80 Hz), which is modulated by high-level cognitive processes such as attention, memory, and meditation. In the first part of my talk, I will discuss some characteristics of gamma oscillations, in particular how varying the color, size and contrast of the stimulus can modulate gamma oscillations, and how these oscillations can be disrupted by introducing discontinuities in the stimulus. In the second part of the talk, I will discuss signal-processing techniques that are used to study some properties of gamma rhythm, such as its duration.
Speaker bio: Dr. Supratim Ray received a B.Tech in Electrical Engineering from IIT Kanpur and a PhD in Biomedical Engineering from the Johns Hopkins University. His postdoctoral training was in the department of Neurobiology at Harvard Medical School. He joined the Center for Neuroscience in June 2011 and is an Associate Faculty in the Electrical Engineering Department since 2012.

His lab studies the mechanisms of attention, i.e., our ability to focus on behaviorally interesting and relevant stimuli while ignoring others. In particular, he is interested in particular brain rhythms thought to be associated with higher order cognitive functions such as attention.

Go top

Title : Stability Analysis of Laurent Systems
Speaker : Dr. Chirayu D. Athalye.
Date : 02/03/2017
Venue : C 241 MMCR, EE
Abstract: goo.gl/eEtqUi
Speaker bio: Dr. Chirayu Athalye did bachelor's in Electrical Engineering from SP College in Mumbai University. He received MTech and PhD degrees from IIT-Bombay. His research areas of interest are dynamical systems, multidimensional systems, infinite dimensional systems, optimal control, multiagent systems, stability analysis, applied and numerical linear algebra, convex analysis and optimization, LMI, and matrix completion problem.

Go top

Title : Bipedal robots: Bridging the gap between theory and experiment
Speaker : Dr. ShiShir N.Y. Kolathaya
Date : 20/02/2017
Venue : C 241 MMCR, EE
Abstract: Natural selection has enabled us to adapt to our environments and achieve complex tasks with relative ease, especially in the area of legged locomotion. These abilities have not yet been translated to bipedal robots, despite the use of complex models, computing power, and novel actuators and sensors. This gap between simulated and observed behavior gets wider with more dynamic tasks like running. Therefore, this talk focuses on a mathematical framework that formalizes the process of implementation in real world systems, i.e., that bridges the divide between theory and experiment. Specifically, the notion of input-to-state stability (ISS) is applied for the construction of robust controllers for a class of hybrid systems that characterize bipedal robots. By treating uncertainties (modeling, model parameter, measurement), or functions of uncertainties as inputs to a system, the talk will describe how to reduce this to a form amenable for input-to-state stability analysis. With this analysis, robust controllers are realized, with the goal of realizing dynamic locomotion behaviors like walking and running, thereby bridging the gap between theory and experiment. This will be demonstrated on multiple robotic platforms including a humanoid robot and running robot.
Speaker bio: Dr. Shishir is a Postdoctoral scholar working for AMBER Lab in the California Insitute of Technology. He received his PhD degree in Mechanical Engineering (2016) from the Georgia Institute of Technology, M.S. degree in Electrical Engineering (2012) from Texas A & M University, and B. Tech degree in Electrical & Electronics Engineering (2008) from the National Institute of Technology Karnataka, Surathkal, India. Prior to pursuing his Master's degree, he also worked for two years as a power supply designer in Tejas Networks Ltd., Bangalore. Shishir has been an integral part of AMBER Lab for more than six years working with Dr. Aaron Ames across three different institutions from 2011-2017. He is interested in nonlinear control, dynamical systems, hybrid dynamical systems, robotics, and particularly in bipedal locomotion.

Go top

Title : Applications of Fourier transform for computing the signed Euclidean distance function and its gradient density function
Speaker : Karthik Gurumoorthy
Date : 17/02/2017
Venue : C 241 MMCR, EE
Abstract: In this presentation, I will present a fast convolution-based technique for computing an approximate, signed Euclidean distance function. The solution stems from first solving for a scalar field in a linear differential equation and then deriving the solution by taking a negative logarithm. The linear formalism results in a closed form solution expressible as discrete convolution and hence efficiently computable using the fast Fourier transform. Computing the winding number and topological degree aid in determining the sign of the distance function whose computations can also be performed via fast convolutions. The complex wave representation (CWR) converts unsigned 2D distance transforms into their corresponding wave functions. Here, the distance transform appears as the phase of a wave function. I will demonstrate using the higher-order stationary phase approximation the convergence of the normalized power spectrum (squared magnitude of the Fourier transform) of the wave function to the density function of the distance transform gradients as a parameter approaches zero. In colloquial terms, spatial frequencies are gradient histogram bins. Some applications of the density of the orientations, known as HOGs (histogram of oriented gradients) include human and object detection and sketch based image retrieval. Towards the end of my talk, I will provide a unified representation from which both the distance function and its gradient density function can be simultaneously retrieved.
Speaker bio: Karthik Gurumoorthy graduated with a dual masters degree in Mathematics and in Computer Science in 2009 and 2010 respectively and earned a doctorate degree in Computer Science in 2011 from the University of Florida, Gainesville. He continued at the same institution for a year in the capacity of a post-doctoral researcher and later joined GE Global Research, Bangalore as a Research Scientist in 2012 pursuing research in the field of medical image analysis. After completing a year and 3 months at GE, he accepted an AIRBUS post-doctoral fellowship position at International Center for Theoretical Sciences, Tata Institute of Fundamental Research (ICTS-TIFR), Bangalore where he conducted research in data assimilation and filtering theory for over a year and 6 months. He currently works at Amazon Development Center, Bangalore as a Machine Learning Scientist in the core Machine Learning Group and is also an Associated Faculty at ICTS-TIFR. He has worked on a wide gamut of problems covering domains like signal processing, computer vision, machine learning, density estimation, filtering theory and image compression and is motivated by problems which are mathematical in nature.

Go top

Title : A Primer on Blockchain Technologies
Speaker : N S Amarnath
Date : 03/02/2017
Venue : C 241 MMCR, EE
Abstract: This talk will cover, briefly, the history of blockchain technology and provide a brief description and rationale for some of the more popular blockchain technologies in the market today. A blockchain is a distributed, secure ledger of transactions which are easy to read, and hard to write or modify. Each block in a blockchain represents a collection of transactions.

The focus of the talk will be on blockchains for contracts and ledgers, with an example from power contracts. Along the way, some encryption algorithms and techniques will be covered, explaining how blockchains work, and why they are considered secure. At the end of the talk, a sample application on Chain, an open source implementation of blockchains, will be presented.

Keywords: Blockchains, contracts

Speaker bio: Amarnath has had a versatile career over 30 years where he has managed people, technology and large, critical programs. He has worked in areas ranging from computer vision, software engineering tools, online advertising, payment services and enterprise mobility. He’s worked in senior positions in organisations like Amazon Web Services, Yahoo! and Samsung Electronics. His last corporate position was Sr. VP in Samsung Electronics. His strengths are in using technology, especially large, distributed, secure services to achieve business goals. He is an alumnus of dept. EE, IISc, having obtained both BE and ME degrees from there. LinkedIn profile is https://in.linkedin.com/in/n-s-amarnath-72b1ba4”

Go top

Title : Ancillary from renewables
Speaker : Sukumar Mishra
Date : 31/01/2017
Venue : C 241 MMCR, EE
Abstract: With the enhanced controllability of Renewable energy sources (RES) in the domain of microgrid and distribution networks, several tasks can be accomplished over the conventional constant active and reactive power (P-Q) dispatch. While the RES are mostly inertia-less sources, modifications in the control loop can be made for attaining the services which are predominant at the transmission level. Such services include voltage and frequency regulation, response based services such as the short term frequency and voltage response, and addressing power quality issues at a distribution level. The RES can be controlled to operate the corresponding inverter in the grid forming, grid feeding and the grid supporting modes. In the isolated mode of operation, when there are no rotating generators (diesel generator), the RES is responsible for the voltage and frequency regulation. This can be achieved using a centralized or a decentralized approach. The decentralized approach is more popularly known as the droop based approach. Droop control enables the proper power sharing as per their droop setting and power rating. Further, a secondary control aids in the efficient voltage and frequency regulation of microgrids in the absence of DG. Even in the presence of a DG, a modification in the droop control strategy to mimic an inertial control to work in consensus with the conventional DG can be achieved providing the inertial support to existing DG. The droop controlled inverters with an inherent synchronization loop are generally known as grid forming inverters and a complete set of the droop and the synchronization control can enable of seamless mode transition between the isolated and the grid connected modes. The grid feeding RES behave mostly as constant power sources. Such controllers, when enabled with inertial and voltage control in the grid connected modes can aid in the ancillary services such as the provision of frequency and voltage response. In addition to the aforementioned ancillary services, the RES in the grid supporting mode can provide enhanced power quality control by modification in the current controller using the sequence components and virtual impedance. Further, single phase RES can be used for three phase balancing, by properly choosing current injection in response to the voltage unbalances.
Speaker bio: Dr. Sukumar Mishra is a Professor in the Department of Electrical Engineering at the Indian Institute of Technology Delhi. His interest lies in the field of Power Systems, Power Quality Studies, and Renewable Energy. He has published over 100 research articles (inclu ding papers in international journals, conferences and book chapters). He is currently holding the position of Vice Chair of Intelligent System Subcommittee of Power and Energy society of IEEE. He is a recipient of the INSA medal for young scientist, the INAE young engineer award, and the INAE silver jubilee young engineer award. He is also a Fellow of IET (UK), NASI (India), INAE (India) and IETE (India). He is working as the NTPC Chair professor and has previously worked as the Power Grid Chair professor. He is also serving as an Independent Director of the Cross Border Power Transmission Company Ltd. and Industry Academic Distinguish Professor. He is currently serving as an Associate Editor for the IET Generation, Transmission & Distribution journal.

Go top

Title : Discriminative Pose-Free Descriptors for Face and Object Matching
Speaker : Soubhik Sanyal
Advisor: Dr. Soma Biswas
Date : 27/01/2017
Venue : C 241 MMCR, EE
Abstract: Matching faces and objects across pose is a very important area of research in the field of computer vision with many applications. For example, in surveillance setting, the face of a person captured by the overhead cameras may be in any uncontrolled pose and resolution as opposed to the frontal image under high resolution that is typically captured during enrolment. For object matching, the images captured during testing can be taken from a different viewpoint compared to the images stored in the database which again requires comparing objects present in different poses. In this talk, we will discuss about a discriminative pose-free descriptor (DPF-SPR) which can be used to match faces/objects across pose variations.
Speaker bio: Soubhik Sanyal completed his Bachelor’s from Jadavpur University, Kolkata in the year 2013. He is currently pursuing his M.Sc(Engg.) from Dept. of Electrical Engineering, Indian Institute of Science, Bangalore. His research interests are in computer vision, machine learning and image processing.

Go top

Title : Polymeric Insulators for High Voltage Transmission Line
Speaker : Alok Ranjan Verma
Advisor: Dr. Subba Reddy B
Date :20/01/2017
Venue : C 241 MMCR, EE
Abstract: In modern era of upcoming industrialization, with the high consumption of electric energy it is very much important to achieve the required demand in power sector for the uninterrupted operation. The continuous availability of electric power is a major factor to fulfill the requirement, which can be achieved by the efficient, safe and reliable power transmission system supported by an effective insulation system. The insulation in over head transmission line (OHTL) is provided by porcelain and glass insulators conventionally but now a days recent advancement in materials lead to polymeric or non ceramic insulating materials. These polymeric insulators are composed of Polydimethylsiloxane (PDMS) as base polymer with different fillers like silica or Alumina trihydrate (ATH) to improve their material properties. There are several advantages of using polymeric insulators like easy manufacturing process, light weight, better mechanical properties, better hydrophobicity, better short term pollution performance etc., but major disadvantage is the penetration of moisture, also being organic in nature it is sensitive to environment and degrades with time by itself. In the present work, efforts are made to understand the behaviour of polymeric insulators material degradation under different environmental conditions (including acid rain condition) and their electrical behaviour against tracking and erosion using Inclined plane tracking method and accelerated aging studies using rotating wheel and dip test arrangement. Some preliminary results are discussed.
Speaker bio: Alok Ranjan Verma born in Aligarh, Uttar Pradesh, India in 1991. He received his B.Tech (Electrical Engineering) from Aligarh Muslim University, Aligarh, Uttar Pradesh and M.E. (Electrical Engineering) from Indian Institute of Science, Bangalore, Karnataka in 2012 and 2014 respectively. He is currently working towards his Ph.D. in High Voltage Engineering from Indian Institute of Science, Bangalore , Karnataka. His areas of interest includes, High Voltage Engineering, Polymeric Insulators for Outdoor Applications, , Computational Electromagnetism, Numerical Techniques in Electrostatics, Lightning Induced disturbances, Over-voltages in power systems.

Go top

Title : Convolution Neural Network(CNN) applications in computer vision, tracking, and rotation invariant classification
Speaker : Prof. Deepak Mishra
Date :13/01/2017
Venue : C 241 MMCR, EE
Abstract: This lecture is aimed to discuss two recent work in application of CNN for computer vision tracking and rotation invariant classification. In the first part of talk I will discuss the CNN application to fast robust visual tracking. In the year 2015 at visual object tracking (VOT) challenge 2015 MultiDomain Network (MDNet) tracker stood as the first tracker in most of the real world challenges and this novel application of CNN/deep learning motivated us to attempt some improvements in the existing MDNet and We had modified the fine tuning part of the guided MDNet by reducing the number of samples used for the fine tuning. This had helped us in decreasing the time to a good extent. In second part I will discuss the work on Convolutional neural networks for rotational invariance classification. are one of the most widely applied deep learning architectures. They extract deep, hierarchical features from the input image, which are robust to scale changes and small distortions in the input, but are sensitive to rotations. We put forward an architecture which provides rotational invariant classification, even when it is trained only with data of single orientation. The proposed idea is then applied to three tasks, namely: handwritten digit classification, captcha recognition and texture classification. Moreover, along with rotational invariant classification, without any additional computational complexity, the proposed architecture is able to determine the approximate orientation of the object in the image.
Speaker bio: Prof. Deepak Mishra received his B.E., in Electrical Engg. (2000) and M Tech in Instrumentation (2003) from Devi Ahilya University Indore, Dr. Mishra pursued his PhD at IIT Kanpur (2007) in the Electrical Engg. Department. His Thesis title was “Novel Biologically Inspired Neural Network Models”. Later He joined as a postdoc researcher at University of Louisville, KY, USA in the field of signal processing and system neuroscience. After a brief stint 2009-2010 as a senior software engg at CMC limited Hyderabad. He opted to work as a academic faculty at Indian Institute of Space Science and Technology Trivandrum in 2010 and continued to work as Associate Professor in the department of Avionics. He is responsible for both research and teaching UG and PG students moreover he was coordinator for Mtech program in Digital Signal Processing and developed a Virtual Reality center of excellence during his stay at IIST. He was also awarded Young Scientist award from System Society of India for his research work in the year 2012. His research interest includes Neural networks and Machine learning, Computer vision and Graphics, Image and Video processing. He has published research papers both in International and National Journal of repute and presented his research work in various international and national conferences.

Go top

Title : Cross-scale predictive dictionaries
Speaker : Vishwanath Saragadam
Date :10/01/2017
Venue : C 241 MMCR, EE
Abstract: Sparse representation, where a signal is represented using a linear combination of a few basis elements presents a promising framework for various image processing and computer vision tasks such as denoising, compression, and recognition. Sparse representation is also used for compressive sensing, where a signal is recovered from far fewer measurements than the signal dimension. Though analytical bases for sparse representation for arbitrary signals is hard to find, over-complete bases or dictionaries, which have more basis elements than the signal dimension are good alternatives. However, for high dimensional signals such as videos, very large dictionaries are needed which take significant time for computation. We propose a novel signal model, based on sparse representations, that captures cross-scale features, particularly for visual signals. We show that cross-scale predictive model enables faster solutions to sparse approximation problems. This is achieved by first solving the sparse approximation problem for the downsampled signal and using the non-zero indices of the solution to constrain the non-zero entries at the original resolution. The speedups obtained are especially compelling for high-dimensional signals that require large dictionaries to provide precise sparse approximations. We demonstrate speedups in the order of 10 âˆ’ 20Ã— for denoising and up to 15Ã— speed-ups for compressive sensing of images and videos.
Speaker bio: Vishwanath is a third year PhD student in the ECE department at Carnegie Mellon University, Pittsburgh. He holds B.Tech (Hons.) From Indian Institute of Technology Madras in electrical engineering. Vishwanath works on compressive sensing and computational photography. His work is directed at reducing sensing and computational load of various cameras through novel signal models and algorithm design.

Go top

Title : High Speed Solutions for Power System Operation
Speaker : R. Gnanavignesh
Advisor: Dr. U. Jayachandra Shenoy
Date :06/01/2017
Venue : C 241 MMCR, EE
Abstract: With increasing dependency on electricity for day to day activities, reliability of the electric power system has become imperative. The talk starts with a brief discussion on blackouts and an intuitive explanation of power system security. The role of Supervisory Control and Data Acquisition/Energy Management System (SCADA/EMS) will be highlighted. Subsequently, the challenges in operating the system in a secure manner will be looked upon. Finally, the on-going work regarding speeding up of power flow - an essential component of SCADA/EMS will be presented.
Speaker bio: R.Gnanavignesh completed his Master's from National Institute of Technology, Tiruchirapalli. He is currently pursuing his PhD from the Dept. of Electrical Engineering, Indian Institute of Science, Bangalore. His research interests are Power System Analysis, Dynamics and Control, High performance computing for power system applications.

Go top

Title : A Deconvolution Based Approach for Feature Enhancement in Cryo Scanning Transmission Electron Tomograms
Speaker : Barnali Waugh
Date : 11/01/2017
Venue : C 241 MMCR, EE
Abstract: Electron tomography is a technique of choice for 3D imaging of subcellular objects. Recently it was demonstrated that scanning transmission electron microscopy (STEM) could be combined with cryo tomography to image whole unstained bacteria and human tissue culture cells, providing fine contrast and detail. However, tomograms contain only partial information about the specimen. This leads to characteristic distortions in the 3D reconstruction. Prior knowledge of the image formation mechanism, which is particularly simple in STEM permit amelioration of these distortions. In this talk I will describe a deconvolution based feature enhancement technique for cryo STEM tomograms and a few case studies with gold nano beads and biological specimens.
Speaker bio:

Go top

Title : REAL VIRTUALITY: SPACE TRAVELERS
Speaker :Caecilia Charbonnier
Date :16/01/2017
Venue : C 241 MMCR, EE
Abstract: Real Virtuality is a multi-user immersive platform combining a 3D environment – which can be seen and heard through a VR headset – with a real life stage set. Users are tracked by a motion capture system allowing them to see their own bodies and move physically in the virtual environment. The experience offered by Real Virtuality brings a once in a lifetime experience. Unlike other static position VR systems, Real Virtuality allows users to become immersed in a VR scene by walking, running, interacting with physical objects and meeting other people. Because user’s movement exactly match their avatars movements in the 3D environment and are streamed to the users with very low latency, there is no discomfort or interface required. The bodies of the visitors become the interface. This ground breaking technology is issued from fundamental research undertaken in the last four years by the Artanim Foundation. It is the only solution available today offering a “matrix-like” degree of immersion over a large area, up to hundreds of square meters. This work was selected among the three finalists of the Immersive Realities (AR/VR) contest at SIGGRAPH Los Angeles 2015, awarding the best augmented/virtual reality experiences possible using today’s technologies. It was selected at the Sundance film festival and presented at Cannes Festival in 2016. This project was also awarded the Laval Virtual 2016 award in the category “3D game and entertainment”. VIDEO LINKS: “Space Travelers” https://www.youtube.com/watch?v=kxW5H4xj1kA
Speaker bio: Caecilia Charbonnier obtained a Master of Advanced Studies (MAS) in Computer Graphics at EPFL and a PhD degree in Computer Science at MIRALab - University of Geneva. She is currently President and Research Director at Artanim. Her work focus on the interdisciplinary use of motion capture from 3D animation, VR applications, live performances to movement science, orthopedics and sports medicine. (Paralinguistics/Cognitive Load) and in 2015 (Non-nativeness detection). She has published over 700 papers and has been granted 17 U.S. patents.

Go top

Title : Glottal source modeling in text to speech synthesis
Speaker :Achuth Rao M V
Advisor: Dr. Prasanta Kumar Ghosh
Date : 30/12/2016
Venue : C 241 MMCR, EE
Abstract: One of the major factor which causes a deterioration in speech quality in text to speech synthesis is the use of a simple delta pulse signal to generate the excitation of voiced speech. There are several methods proposed to model the glottal source of speech like Liljencrants Fant and Rosenberg models. But these models has limitations such as non-convex estimation methods and higher synthesis time. We propose a new glottal models based on incomplete beta function. The objective scores shows that the proposed model is better than the Liljencrants Fant model and has few number of parameters.
Speaker bio: Achuth Rao MV received his BE degree in Electronics and Communication Engineering from RV college of Engineering Bangalore. He is currently a PhD student in Dept. Of Electrical Engineering at IISc. His research interests broadly include Speech processing, Pattern recognition and bio-medical signal processing.

Go top

Title : Switched Reluctance Machine Drive
Speaker :Syed Shahjahan Ahmad
Advisor: Prof. G. Narayanan
Date :23/12/2016
Venue : C 241 MMCR, EE
Abstract: High speed electrical machines find applications in power generation, gas compressors and precision machining among others. The switched reluctance machine (SRM) is a potential candidate for high speed turbo-alternator, given its robust rotor with no permanent magnets or conductors. However, the machine is highly non-linear and challenging to control. In this presentation, the basic working of the SRM and its challenges will be described. Existing methods for current control along with proposed methods will be discussed and compared. Theoretical and experimental studies carried out on SRM in generating mode will be presented.
Speaker bio: Syed Shahjahan Ahmad received the B.E degree in Electrical Engineering from Indian Institute of Engineering Science and Technology, Shibpur in 2012, and M.E. degree in Electrical Engineering from Indian Institute of Science, Bangalore in 2014. He is currently pursuing his Ph.D. at the Indian Institute of Science, Bangalore, in the Department of Electrical ngineering. His research interests include design and control of switched reluctance machines, high speed electric drives, power electronic converters, and modelling and control of power electronic systems.

Go top

Title : Motors for Electric Vehicles - Part II
Speaker :Prof. K. Ragavan, IIT-Gandhinagar
Advisor: Motors for Electric Vehicles - Part II
Date :15/12/2016
Venue : HV Seminar Hall
Abstract: This is continuation of the talk that was given on Dec 6, 2016. In Part-I, basics of vehicle load requirement were discussed. Further, it is well known that it is possible to alter the terminal characteristics of a motor by altering the supply characteristics. However, few aspects are to be considered while selecting a motor for electric vehicles. Those are: power density, maintenance, extended speed range, high performance, etc. With regard to power density, the preferred choice is a Permanent Magnet motor (popularly called brushless DC motor). Details about making this choice will be discussed.
Speaker bio: Dr Ragavan K, Associate Professor, Electrical Engineering, IIT Gandhinagar did his B.Tech. in Electrical and Electronics Engineering (1993-1997) from Pondicherry Engineering College, Pondicherry and his M.E. in Power Electronics and Drives (2000-2002) from College of Engineering Guindy, Chennai. Later, he received his Ph.D. degree from Department of Electrical Engineering, Indian Institute of Science, in 2006. After completing Ph.D., he worked in GE India Technology Centre Private Limited, Bangalore as Research IP Professional and Research Engineer (Jan. 2007 – Feb. 2009). Then, he joined Indian Institute of Technology Gandhinagar (Electrical Engineering) in May 2009. His research interests are condition monitoring of transformer, design of rotating electrical machines and drives for electric vehicle applications.

Go top

Title : Adaptive Sampling Pattern Design Methods for MR Imaging
Speaker :Chennakeshava K
Advisor: Prof. K Rajgopal
Date :16/12/2016
Venue : C 241 MMCR, EE
Abstract: Magnetic Resonance Imaging is a non-invasive and non-ionizing medical imaging modality having multiple utilities. The scan time involved with MRI is higher and there is necessity to reduce it to avoid the discomfort caused to the patients undergoing the scan. One of the methods to reduce the scan time has been to develop suitable sampling patterns. Sampling patterns represent the indices from which the k-space data is collected. The talk discusses the development of adaptive sampling patterns which consider the k-space behavior, by formulation of a Knapsack problem. A cost function indicative of the energy captured by the sampling patterns is defined, and comparison of reconstruction metrics at several undersampling ratios of different sampling patterns will be shown. A brief discussion explaining the unique patterns of k-space energy distribution in MR Images, and its utility will be discussed.
Speaker bio: Chennakeshava completed his Bachelor’s from Sri Jayachamarajendra College of Engineering, Mysore, worked with Robert Bosch Engineering Solutions, Bangalore and is currently pursuing his M.Sc (Engg.) from the Dept. of Electrical Engineering, Indian Institute of Science, Bangalore. His research interests include Medical Imaging, Machine Learning, and Image Processing.

Go top

Title : Making Sense of Real life observations in the Indian Grid and Waveform Relaxation Method for in-situ testing of Power System/Power Electronic Controllers
Speaker : Prof. Anil Kulkarni
Date : 09/12/2016
Venue : C 241 MMCR, EE
Abstract: a) Making Sense of Real life observations in the Indian Grid: Classical power system dynamic phenomena like unstable swings, loss of synchronism and system frequency changes have been studied extensively in the past. The advent of synchronized wide-area measurement systems has made real-life observations of system wide dynamics easily accessible to the power system community. While many of these observations are a “mere” confirmation of what has been known in theory, some do require more than a cursory examination. In this talk, some interesting real-life observations will be presented and related issues will be discussed.

b) Waveform Relaxation Method for in-situ testing of Power System/Power Electronic Controllers: An alternative to real-time simulation for hardware-in-the-loop testing is presented. This involves system simulation, not necessarily done in real time, and real-time playback of the simulated output to the controller under test. The time-stamped controller output is stored and subsequently fed as an input to the simulation. This whole process is done iteratively as in the Waveform Relaxation method, till the waveforms converge. This method can be used for testing multiple and dispersed controller hardware and the associated communication equipment, e.g., wide-area measurement based control and system protection schemes. It also has the potential to be an alternative to real-time simulators which are expensive when large systems have to be simulated. The basic scheme and potential applications are discussed.

Speaker bio: Prof. Anil Kulkarni received B.E. degree in Electrical Engineering in 1992 from the University of Roorkee.Received M.E in 1994, and Ph.D degree in Electrical Engineering in 1997, from IISc Bangalore. He is currently working in Department of Electrical Engineering, IIT bombay. HIs research Interests are Power System Dynamics, Flexible AC Transmission Systems, HVDC Transmission Systems.

Go top

Title : Standards and Their Impact on Future R&D in Electric Motors, Drives and Control
Speaker :Prof. Krishnan Ramu
Date :09/12/2016
Venue : C 241 MMCR, EE
Abstract: On going efforts in international and national standards agencies in the area of line operated electric motors and converter operated electric motors (also known as variable speed motor drives or simply as motor drives) have many impacts. They draw the attention of a handful of engineers in the world and, surprisingly, it is an understatement. The standards have immense impact on many practical aspects such as efficiency, testing and selection and application in the industries, offices and as well as in homes. Consider only efficiency standards, usually known as minimum energy performance standards (MEPS), and say their enforcement which results in very significant energy savings and hence in a reduction in operational costs to the consumer and industries.An inexorable march of the standards towards higher efficiencies requires continuous research and development efforts to deliver them. Higher efficiency electric motors come at the expense of additional cost due to redesign with much more materials than that of the lower efficiency motors. The presentation highlights the International Electro technical Commission’s (IEC) standards on electric motors both in force and evolving ones and their impact with concrete examples on the incremental initial investment in adoption of higher efficiency standard motors and their payback period, resulting energy savings, and life cycle operational cost savings. The benefits of higher efficiency standards are demonstrated. The key to all these developments is the basic and applied research in electric machines, converters and control. Newer opportunities to face these challenges are in the design and development of newer electric motors as well asknown class of machines in the newer configurations, converters and their operation, and finally in the software control integration of the motor and converter for high efficiency operation. Some possible directions are also identified here. Note that the R&D efforts lead the entrepreneurship that may follow when these areas of research result in innovation with cost effective ways to realize the motors, converters and control can be put to production and into the practice.
Speaker bio: R. Krishnan is a professor of electrical and computer engineering at Virginia Tech, Blacksburg, VA. He is currently the director of the Center for Rapid Transit Systems in linear and rotating motor drives. His research interests are in electric motor drives, electric machines, power electronics and control. Krishnan is a recipient of best paper prize awards from IEEE Industry Applications Society’s Industrial Drives committee (5 awards) and Electric Machines committee (1 award). In addition, he received the first prize from IEEE Transactions on Industry Applications for his paper and the 2007 Best Paper Award from IEEE Industrial Electronics Magazine. His co-edited book Control in Power Electronics won the best book award from Ministry of Education and Sport, Poland, in 2003. He was awarded IEEE Industrial Electronics Society’s Dr. Eugene-Mittelmann Achievement Award for Outstanding Technical Contributions to the field of Industrial Electronics in 2003. Krishnan is a Fellow of the IEEE and a Distinguished Lecturer of IEEE Industrial Electronics Society.

Go top

Title : Motors for Electric Vehicles
Speaker :Prof. K. Ragavan
Date :06/12/2016
Venue : HV Seminar Hall, EE
Abstract: Pollutants from engine powered vehicles can be eliminated with the use of motor driven vehicles. The motor that was preferred for vehicle applications is now being replaced. This change has become possible due to developments in semiconductor devices. With the use of high energy density permanent magnets, the power density of the motors can be improved. For extending the speed-range of the motor beyond base-speed, the air-gap flux has to be reduced. Is it possible to achieve such flux weakening operation with permanent magnetmotors? Reluctance motors are preferred due to its simple constructional features. The torque produced by reluctance motors has higher ripple and produces more acoustic noise. Instead of dissipating kinetic energy by mechanical brake, it can be converted to electrical form and stored in the batteries (regenerative braking). All these aspects will be discussed.
Speaker bio: Dr Ragavan K, Associate Professor, Electrical Engineering, IIT Gandhinagardid his B.Tech. in Electrical and Electronics Engineering (1993-1997) from Pondicherry Engineering College, Pondicherry and his M.E. in Power Electronics and Drives (2000-2002) from College of Engineering Guindy, Chennai. Later, he received his Ph.D. degree from Department of Electrical Engineering, Indian Institute of Science, in 2006.After completing Ph.D., he worked in GE India Technology Centre Private Limited, Bangalore as Research IP Professional and Research Engineer (Jan. 2007 – Feb. 2009). Then, he joined Indian Institute of Technology Gandhinagar (Electrical Engineering) in May 2009. His research interests are condition monitoring of transformer, design of rotating electrical machines and drives for electric vehicle applications.

Go top

Title :Advanced Approaches to Renewable Energy Integrated Power System Modeling and Energy Management
Speaker :Dr. Sudipta Ghosh, Shiva Nadra University, Delhi
Date :28/11/2016
Venue : C 241 MMCR, EE
Abstract: Modern power grid is more complex and shows unforeseen dynamics. Rapid control decisions have to be taken on the basis of multiple contingency evaluations using limited computational resources. Such analyses are time consuming, exceeds computational limits and difficult to accomplish in faster time frames. Model order reduction (MOR) is one such tool in control engineering which can simplify system formulations while retaining phenomena of interest. This work shows 1) a new approach for coherency identification that captures dynamic behavior of the power grid, 2) a new methodology for scalable power system coherency grouping based on mathematical (BT & Krylov subspace) MOR of larger power grid including online PSS tuning methodology. Further the work explains reduced order modeling of wind turbines & wind farms. The work also describes 1) a new wind farm control framework for inertial and primary frequency response for a high wind integrated power system, 2) an energy function based optimal control strategy for output stabilization of integrated DFIG-flywheel energy storage architecture to avoid voltage and power fluctuations, 3) a new dynamic reactive power estimation based coordinated control of grid integrated DFIGs to improve network stability. Further a real-time model reduction of large power grid into an equivalent network while preserving low and high frequency behavior of the original system will be presented.
Speaker bio: Sudipta Ghosh received his Ph.D. degree from Indian Institute of Technology, Delhi in 2013 in Power Systems. From 2013-2014, he worked as Assistant Professor at IIT, Dhanbad. In 2014 he came to USA as a Research fellow at the University of North Carolina (UNCC), at Charlotte. He was a lead researcher for the Hybrid real time simulator based advance modeling of the Southern California Edison grid. He also worked as a lead researcher for the NSF funded project on power system on new methods of optimal control designs integrating renewable energy systems research. Next year, he was appointed as Graduate Associate Faculty at UNCC. In this context he was assisting PhD and Master’s students and also working on a NSF project. He has just joined Shiv Nadar University as an Assistant Professor. He has 7 years teaching/research experience and 3 years industrial experiences. He is a member of the IEEE and IEEE PES. He was in Student Program Committee for NAPS (North American Power Symposium) 2015. His publication record consists of nine journals and seven conference papers (total 292 google scholar citations). He received the National Scholarship in 1995 and POSCO power system Award from Power Grid in 2014.

Go top

Title :Principles and Design of a system for Academic Information Retrieval based on Human-Machine Dialog
Speaker :Prof. Hiroya Fujisaki
Date :24/11/2016
Venue : C 241 MMCR, EE
Abstract: This talk describes the outcome of a successful national project led by Fujisaki under the “Research-for-the-Future” program, The system is based on the following three original features: (1) Use of “Key Concepts” in information retrieval (including processing of polysemy, homonymy, and unknown words), (b) Dialogue Management based on both User and System Modeling (by introducing a novel type of interacting automaton), and (c) Optimization of Retrieval Performance through Relevance Score Estimation.
Speaker bio: HIROYA FUJISAKI is a Professor Emeritus at the University of Tokyo and a Professor in the Department of Applied Electronics at the Science University of Tokyo. In 1991, he retired as a Professor of Electrical Engineering from the Univ. of Tokya and took on his current appointments. His research in speech and language processing covers a broad range of topics, including production, perception, acquisition, impairment, analysis, synthesis, coding, and recognition. His work on mathematical and physical modeling has led to the development of models of language use, models for perceptual processes in speech identification and discrimination and models for the process of fundamental frequency control in speech. He also developed a model of road traffic flow which has been applied to road traffic control since the 1970s. Among his many honors, Professor Fujisaki is an Honorary Member of the Acoustical Society of Japan (ASJ) and a Fellow of the Acoustical Society of America (ASA) and of the Institute of Electronics, Information and Communication Engineers (IEICE). He has been honored by the Mayor of Tokyo as a "Person of Merit in Science and Technology", and has received The Third Millennium Medal from the Institute of Electrical and Electronics Engineers (IEEE).

Go top

Title : Problems and Prospects of spoken Language Processing
Speaker : Prof. Hiroya Fujisaki
Date :24/11/2016
Venue : C 241 MMCR, EE
Abstract: Instead of the conventional distinction between “Speech” and “Language”, Fujisaki introduced the concepts of “Spoken Language” as contrasted to “Written Language”, pointing out that speech contains certain linguistic information that is not in the written language. He also made clear that what has been traditionally called “Natural Language Processing” is actually “Written Language Processing”, and defined the field of “Spoken Language Processing,” dealing with both the aspects of speech as a signal and its aspects as a code. This talk describes the rationale that led to this conceptual turn, and shows the progresses, unsolved problems and future prospects of the field.
Speaker bio: HIROYA FUJISAKI is a Professor Emeritus at the University of Tokyo and a Professor in the Department of Applied Electronics at the Science University of Tokyo. In 1991, he retired as a Professor of Electrical Engineering from the Univ. of Tokya and took on his current appointments. His research in speech and language processing covers a broad range of topics, including production, perception, acquisition, impairment, analysis, synthesis, coding, and recognition. His work on mathematical and physical modeling has led to the development of models of language use, models for perceptual processes in speech identification and discrimination and models for the process of fundamental frequency control in speech. He also developed a model of road traffic flow which has been applied to road traffic control since the 1970s. Among his many honors, Professor Fujisaki is an Honorary Member of the Acoustical Society of Japan (ASJ) and a Fellow of the Acoustical Society of America (ASA) and of the Institute of Electronics, Information and Communication Engineers (IEICE). He has been honored by the Mayor of Tokyo as a "Person of Merit in Science and Technology", and has received The Third Millennium Medal from the Institute of Electrical and Electronics Engineers (IEEE).

Go top

Title : Microgrids-Operation and Control Issues
Speaker : Dibakar Das
Advisor: Dr U Jayachandra Shenoy & Dr Gurunath Gurrala
Date : 25/11/2016
Venue : C 241 MMCR, EE
Abstract: With the increased popularity of non-conventional energy sources like wind, solar, etc, the conventional electrical grid has undergone some major transformations in the past few years. This talk discusses one such transformation, the microgrids. A microgrid is a collection of distributed sources along with loads which can operate with the main grid as well as in the absence of grid. This talk discusses some of the recent developments in the field and some of the major control challenges. The operating modes of the microgrids will be discussed in detail. Finally, the concept of seamless transfer will be introduced and a linear quadratic regulator theory based seamless transfer algorithm will be briefly discussed.
Speaker bio: Dibakar Das completed his Bachelor’s from National Institute of Technology, Durgapur in the year 2014. He is currently pursuing his M.Sc (Engg.) from Dept. of Electrical Engineering, Indian Institute of Science, Bangalore. His research interests are power electronics, renewable integration and microgrids.

Go top

Title :Brain and Health
Speaker :Prof. A. G. Ramakrishnan
Date :18/11/2016
Venue : C 241 MMCR, EE
Abstract: The talk will start with an interesting demo of a crystal ball, that can receive the intentions of the holder and make linear and circular movements. It will also cover the largely unknown fact of the intimate connection between oxygen shortage and cancer (Nobel Prize work of Otto). By demonstrating that we have conscious control over our blood flow (biofeedback), the strong connection between one’s thoughts and health will be made clear. How we can make use of the miraculous mechanisms of the body for self-healing. How to prevent cancer with simple antiangiogenic fruits & vegetables? How our genes are NOT our fate and DNA is NOT our identity (epigenetics). The 4 golden rules for a disease-free life. How we can we make use of the non-visual photoreceptors in our retina. The final aim of the talk is to establish that it is possible for anyone to accelerate towards great positive health, by following rather simple steps of eating, breathing and sleeping: more than what to eat, focus on how to eat. The speaker himself is a demonstration of the ideas he is forwarding: After being dependent on medicine for keeping his hypertension under control for over nine years, he has been able to bring back normal blood pressure by following some of these steps and is completely free of his medicine for the past six months, with no diet control !
Speaker bio: A G Ramakrishnan obtained his Ph D in Biomedical Engineering from IIT Madras. For his work on nerve conduction in leprosy, he received the Sir Andrew Watt Kay Young Researcher’s Award from the Royal College of Physicians and Surgeons, Glasgow. He has collaborated and jointly published with Padmashri Dr. H R Nagendra (Director of Sri Vivekananda Yoga Anusandhana Samsthana), Prof. B N Gangadhar, Director of NIMHANS and Dr. S Suresh, Director of Fetal Care Research Foundation. He has graduated 25 research students so far. He was the President of Biomedical Engineering Society of India. Blind students are using over 600 Braille books in Tamil, converted from printed books using his OCR, Mozhi Vallaan. This work earned him the Manthan Award 2014 - South Asia and Asia Pacific in the category e-inclusion and accessibility. He has developed unrestricted vocabulary, handwritten word recognition system in Tamil, for which he received Prof. M Anandakrishnan award from INFITT in 2013. He has also developed Thirukkural and Madhura Vaachaka - good quality text to speech conversion software for Tamil and Kannada, used by blind students, for which he received the Manthan Award 2015 in the e-education category. He conceived of Linguistic Data Consortium for Indian Languages, currently managed by CIIL, Mysore.

Go top

Title :Large-scale Sensor Network Localization
Speaker :Rajat Sanyal
Advisor: Dr. Kunal Narayan Chaudhury
Date :11/11/2016
Venue : C 241 MMCR, EE
Abstract: Recent developments in wireless communications and electromechanics have proliferated the deployment of wireless sensor networks. While such networks are typically used to monitor different physical quantities over a region, they are also used in surveillance and disaster management. A fundamental computational problem in this regard is to estimate the distribution of the entire sensor network from the available inter-sensor distances (estimated using local communication links). In this talk, I will give an overview of our recent work on large-scale sensor network localization.
Speaker bio: Rajat Sanyal (GM’ 16) received the B.Tech. in electronics and communication engineering from National Institute of Technology, Durgapur, WB, India, in 2014. He is currently pursuing his M.Sc. (Engg.) in electrical engineering from Indian Institute of Science, Bangalore, KA, India. He is working on large-scale sensor network localization problem. His research interests broadly include convex optimization, wireless communication and machine learning.

Go top

Title : Hybrid grid-tie inverters for back-up power applications
Speaker : Venkatramanan D
Advisor: Prof. Vinod John
Date : 04/11/2016
Venue : C 241 MMCR, EE
Abstract: Energy demand increases continually. Renewable energy based distributed generation (DG) systems are gaining popularity today as they address the growing energy demand. Particularly, solar photo-voltaics (PV) based grid-tie inverters (GTI) are available in the market for a range of power levels. However, the GTI system, by design, would function only in the presence of power grid and would remain idle in the event of a power outage. In this work, a hybrid GTI configuration is presented, which by applying appropriate control, would function even in the absence of grid and cater to local power needs, thus providing back-up power while accessing renewable energy.
For realizing such a hybrid GTI functionality, a flexible power converter system is required. Hence, focus is laid on design of power converters, where a procedure suggested that is both simple and state-of-the-art. Details of the hardware platform developed in laboratory are presented that can cater to a variety of power conversion applications. Experimental results are presented that illustrate the high performance that is achieved with the power converter.

Go top

Title : Stability of Electric Gird: Challenges and Solutions
Speaker : Ajit Kumar
Advisor: Dr. G. Gurrala and Dr. I. Sen
Date : 28/10/2016
Venue : C 241 MMCR, EE
Abstract: Power systems are large complex systems which are highly nonlinear and high order system. Security of a power system is affected by characteristics of the physical systems. It is mainly affected by integrated generation, transmission and distribution. Different form of instability includes rotor angle, voltage and frequency instability.

In this talk, we will focus on rotor angle instability of electric grid. This form of instability is mainly caused by inter-area and local modes of oscillations among generators. Inter-are modes are of low frequency and involve large geographical areas, limiting the transmission capacity of tie-line between two large areas. Since power equipment’s are manufactured to operate in a specified range, voltage regulation (AVR) is required for system operation. Since AVR action affects local and inter-area modes damping. Traditionally, damping controllers are used to enhance the damping performance of these modes.

We will present damping controller design using local measurements. Furthermore, a nonlinear AVR design is discussed based on differential geometric theory. We will show that the nonlinear AVR damp inter-area modes better than conventional AVRs, paving the way for more power transfer across areas.

Go top

Title : Dictionary Learning for Matching Data Under Cross-Modal and Privileged Information Scenarios
Speaker : Devraj Mandal
Advisor: Dr. Soma Biswas
Date : 21/10/2016
Venue : C 241 MMCR, EE
Abstract: Cross-modal recognition and matching with hidden information are important challenging problems in the field of computer vision. The cross-modal scenario deals with matching across different modalities and need to take care of the large variations present across and within each modality. The hidden information scenario deals with the situation that all the information available during training may not be available during the testing stage and hence algorithms need to leverage the extra information from the training stage itself. Though separate algorithms have been designed to specifically handle the two situations efficiently, there is a lack of single joint framework which is able to handle the two problems concurrently. We show that for multi-modal data, either one of the above situations may arise if one modality is absent during testing. Here, we propose a novel joint framework which can handle both these scenarios seamlessly with applications to matching multi-modal data. The proposed approach jointly uses the data from the two modalities to build a canonical representation which encompasses the information from both the modalities. We explore three different types of canonical representation for different types of data. The algorithm computes dictionaries for the data from both the modalities and the canonical representation, such that, the transformed sparse coefficients of both the modalities are equal to that of the anonical representation. The sparse coefficients are finally matched using a metric learning algorithm. Extensive experiments on different datasets, involving RGBD, text-image, audio-image data show the effectiveness of the proposed framework.
Speaker bio: Devraj Mandal received the B. Tech degree in Electronics & Communication Engineering from West Bengal University of Technology, Kolkata, in 2011 and the M. Tech degree from Jadavpur University, Kolkata, in 2014. He is currently a doctoral student in the Department of Electrical Engineering, Indian Institute of Science, Bangalore, India. His research interests are in image processing, computer vision, and pattern recognition.

Go top

Title :Detection of significant transitions and estimation of glottal closure instants in a speech signal
Speaker :K V Vijay Girish
Advisor: Prof. A G Ramakrishnan
Date :14/10.2016
Venue : C 241 MMCR, EE
Abstract: A unsupervised acoustic-phonetics knowledge based approach is used to detect transitions between broad phonetic classes in a speech signal. This has applications such as landmark detection and segmentation. A rule-based approach using relative thresholds learnt from a small development set is devised to detect transitions of silence to non-silence, sonorant to non-sonorant and vice-versa. This approach does not require significant training data for determining the parameters of the proposed approach. When tested on the entire TIMIT database for clean speech, 93.6% of the detected transitions are within a tolerance of 20 ms from the hand-labeled boundaries. The proposed method is also tested on the test set of the TIMIT database for robustness with respect to white, babble and Schroeder noise, and about 90% of the detected transitions are within a tolerance of 20 ms at a SNR of 5 dB. As another part of my work, I have proposed subband analysis of linear prediction residual (LPR) to estimate the Glottal Closure Instants (GCIs). It is evaluated using 6 different databases and compared with 3 state-of-the-art LPR based methods. The proposed method is comparable to the best of the LPR based techniques for clean and noisy speech.
Speaker bio: K V Vijay Girish graduated from National Institute of Technology Karnataka, Surathkal in 2008 with a B.Tech in Electrical and Electronics Engineering. He joined Dept. of Electrical Engineering, Indian Institute of Science, Bangalore, India in 2010 and is pursuing PhD in the field of Machine Listening since then as a research student. His research interests include Machine Listening, Audio Source Separation, Speech Signal Processing, Audio and Speech Analysis, Image Processing and Pattern Recognition.

Go top

Title :Industrial Plasma Technology: An Overview
Speaker :Anusuya Bhattacharyya
Advisor: Prof. B. S. Rajanikanth
Date :7/10/2016
Venue : C 241 MMCR, EE
Abstract: Plasmas make up more than 99% of the visible matter in the universe, and it mainly consists of positive ions, electrons and neutral particles. Whereas natural plasmas have been the object of scientific studies right from the 17th century, the twentieth century has witnessed rapid progress in the development, diagnostics and applications of plasma. This talk will include an introduction to low temperature plasma technology and various types of plasma discharges such as pulsed corona discharge, dielectric barrier discharge, surface discharge, glow discharge (low pressure discharge) and their properties. The latter part of the talk will discuss some case studies where these technologies have contributed, namely pollution control from diesel engines, wastewater treatment, biological applications and surface treatment of materials.
Speaker bio: Anusuya Bhattacharyya obtained her PhD from the Department of Electrical Engineering, Indian Institute of Science, Bangalore, India. Her research interests include application of electric discharges for pollution control in cascade with adsorbent/ catalytic materials and other plasma technology

Go top

Title :Photovoltaic Energy Conversion Systems
Speaker :Pallavi Bharadwaj
Advisor: Prof. Vinod John
Date : 23/09/2016
Venue : C 241 MMCR, EE
Abstract: This talk covers several aspects of photovoltaic (PV) energy conversion, namely: irradiation measurement, static and dynamic modelling, and characterization of PV panels. The development of an electronic PV panel output characterisation hardware setup that offers the advantage of both static and dynamic panel measurements is presented. After system modelling, a case study on the effect of storage on the cost of grid-connected PV systems is analysed. Based on this study, given a grid outage scenario, a method to choose between grid-tied and dual mode PV systems is presented.
Speaker bio: Pallavi Bharadwaj is working towards her PhD degree in the department of Electrical Engineering at the Indian Institute of Science, Bangalore, India. Her research interests include development and control of power electronic systems for renewable energy applications and grid integration.

Go top

Title :A Fast Approximation of the Bilateral Filter
Speaker : Sanjay Ghosh
Advisor: Dr. Kunal Narayan Chaudhury
Date :16/09/2016
Venue : C 241 MMCR, EE
Abstract: The bilateral filter is an edge-preserving smoother that has applications in image processing, computer vision, computer graphics, and computational photography. The direct implementation of the bilateral filter requires O(w^2) operations, where w is the width of the spatial kernel. In this talk, we will discuss a fast approximation of the bilateral filter which can cut down the complexity to O(1), without appreciably compromising the filtering quality.
Speaker bio: Sanjay Ghosh is working toward his PhD degree in electrical engineering at the Indian Institute of Science, Bangalore, India. His research interests broadly include inverse problems in imaging, computational photography, and computer vision.

Go top

Title : Voltage Stability Analysis in Power Systems
Speaker : A Santosh Kumar
Advisor: Prof D. Thukaram
Date :09/09/2016
Venue : C 241 MMCR, EE
Abstract: This presentation gives a brief look at Voltage stability Analysis in power systems. Voltage instability has led to severe grid failures in recent past. Voltage instability is a phenomenon usually observed in heavily loaded systems. Maintaining a stability margin is very crucial for the system in case of any disturbance, to study this voltage stability analysis is done. This presentation will cover the following points briefly:
• Voltage stability causes and effects.
• Methods for voltage stability analysis: PV curves, VQ curves, Thevenin equivalent based and L-Index.
• Ways in which voltage collapse can be mitigated: VAR compensation.
• Using synchrophasors (PMU) for stability analysis.
• Issues with renewables integration from voltage stability point of view.
Speaker bio: Santosh is doing MSc(Engg) under Prof D. Thukaram at Dept. of Electrical Engineering, IISc Bengaluru. His research work is in the field of power systems voltage stability.

Go top

Title : Braids of partitions
Speaker : Dr B Ravi Kiran
Date : 02/09/2016
Venue : C 241 MMCR, EE
Abstract: In this talk we focus on the problem of extracting an optimal partition from a hierarchy of partitions by dynamic programming. We look at conditions under which the dynamic programming (DP) gives an optimal solution, firstly by defining the conditions on the energy define over the partial partitions of the subset of the space, and secondly describing the partial ordering between partitions necessary to preserve the DP substructure. The talk further identifies various possible braids in literature and how this structure relaxes the segmentation problem. We shall show demonstrative examples of optimal cuts in the context of image segmentation. Given that there could be many solutions possible, we impose unique solutions and we define the necessary conditions for its existence. This uniqueness induces an ordering relation between partitions, and in this case a lattice structure (family of partitions with unique extremal elements). In this talk we shall also briefly review decision trees, decision forests and decision jungles and the creation of the partially ordered partitions of the feature spaces when creating these classifiers.
Speaker bio: B Ravi Kiran has finished a PhD in computer science and mathematical morphology from Université Paris-Est, A3SI-LIGM in Oct 2014. Following this he worked in hyperspectral image processing for tumor detection at Mines ParisTech as Post-doctoral researcher in the European project Helicoid. Currently he is a Postdoctoral researcher at the DATA lab in ENS Paris, France, working on unsupervised time series anomaly detection.

Go top

Title : Recurrent Neural Network and its Applications in Sequence Predictions
Speaker : Vidyadhar Upadhya
Advisor: Prof. P. S. Sastry
Date : 02/09/2016
Venue : C 241 MMCR, EE
Abstract: Numerous learning tasks which deal with sequential data cannot be modeled using standard feed forward neural networks because they assume that the training samples are independent. In such tasks, the dynamics of sequential data should be modeled explicitly. The recurrent neural networks (RNNs), however, can model the sequential data well by retaining a state that represents information from an arbitrarily long context window. In this talk, I will discuss briefly about different RNN architectures, issues involved in training them and some existing solutions. In addition, I will also present an application of RNN in medical diagnostics. Specifically, diseases such as pneumonia are characterized by abnormalities in respiratory signal, which is a physiological time series. In this context, I will explain the extraction of the respiratory signal, from videos acquired through a regular camera, using RNN.
Speaker bio: Vidyadhar received his MS degree in Electrical Engineering from the Indian Institute of Technology Madras. He is currently a PhD student in the Dept. of Electrical Engineering at IISc. His research interests broadly include Pattern Recognition, Deep Learning.

Go top

Title : Image Restoration using Inverse Correlation based Roughness Minimization
Speaker : Mr. Sanjay Viswanath
Advisor: Dr. Muthuvel Arigovindan
Date : 26/08/2016
Venue : C 241 MMCR, EE
Abstract: Image Restoration using regularization is highly popular topic of research in image processing. While the focal point of innovation has been general image priors as regularization functions, dictionary learning algorithms have shown the advantage of using training samples to build better priors for specific classes of images. However these techniques are generally based on sparsity prior and are computationally expensive. Inverse Correlation (IC) is proposed as a novel regularization method which incorporates learning in Total Variation(TV) prior through an IC matrix. The IC matrix is built from derivatives computed at each pixel of training images and consequently adapts to the structure for that class of images. The IC matrix is then used to formulate a convex non-quadratic IC cost function which retains the simplicity of derivative filter based general regularization schemes. We also propose a reconditioned gradient descent algorithm to minimize the IC cost function. The simulation results show that the IC regularization can adapt to training samples and yield better performance than general priors like TV.
Speaker bio: Sanjay V. received his M Tech degree in Electronics and Communication Engineering (Signal Processing) from Indian Institute of Technology Guwahati. He is currently a PhD student in the Dept. of Electrical Engineering at IISc. His research interests broadly include Image Processing, Compressive Sensing, and Pattern Recognition.

Go top

Title : Microgrid Energy Manager - A Platform for Plug-and-Play Management of AC/DC/Hybrid Microgrids
Speaker : Mr. Ashray G Manur
Date : 25/08/2016
Venue : C 241 MMCR, EE
Abstract: With the emergence of microgrids, deploying small-scale (micro) grids in homes, buildings and communities for reliable electricity access is becoming a viable option. This research introduces the concept of grid management at a home/building level called homegrids, managed by Homegrid Energy Manager (HEM) which works towards meeting local constraints and requirements. A group of homegrids are managed by a Microgrid Energy Manager (MEM) which takes care of overall grid management at a community level. Both HEM and MEM adapt a Plug-and-Play model and work with existing storage systems, inverters, appliances and other loads/energy sources. Capable of working with AC/DC or hybrid microgrids, MEM/HEM use a multi-layer approach to provide real-time control and intelligent management of energy sources and loads using embedded sensing and computing, wireless networks, internet-of-things and cloud-computing technologies.
Speaker bio: Ashray Manur is a PhD student at University of Wisconsin-Madison and a member of Wisconsin Electric Machines and Power Electronics Consortium (WEMPEC). He is advised by Prof. Giri Venkataramanan. Ashray’s research interests include microgrid management, cyber physical energy systems and IoT/cloud computing technologies for energy systems.

Go top

Title : Manipulation and Interrogation of Matter at the Small Scale: A Control Systems Perspective
Speaker : Dr. Murti V. Salapaka
Date :19/08/2016
Venue : C 241 MMCR, EE
Abstract: New temporal and spatial regimes of exploration enabled by nanoscience and nanotechnology have led to significant insights into fundamental processes that govern dynamics at the small scale of matter including bio-matter at the molecular scale. These abilities were enabled by breakthroughs in instrumentation that had to overcome fundamental sources of uncertainty such as thermal noise. In this talk, the primary challenges to nanoscale interrogation and manipulation will be presented in a systems perspective. Here, solution methodologies enabled by a modern control approach will be highlighted. With the exploration of biological processes at the molecular and cellular scale using nano-interrogation tools, it has become evident that evolution has endowed biology with remarkable machinery to perform and achieve precise functionality at the small scale in the presence of a highly uncertain environment. Understanding these bio-molecular systems, apart from providing key insights into biology and the related therapeutic impact, holds the promise for strategies to engineer material and systems at the small scale. Recent efforts into probing and understanding transport at the molecular scale and key proteins that provide structural integrity will be detailed to showcase the power of a control systems perspectives.
Speaker bio: Murti V. Salapaka was born in Andhra Pradesh, India, in 1969. He received the B.Tech. degree in Mechanical Engineering from the Indian Institute of Technology, Madras. He received the M.S. and Ph.D. degrees in Mechanical engineering from the University of California, Santa Barbara, in 1991, 1993, and 1997, respectively. From 1997-2007, he was with the Electrical Engineering Department at Iowa State University, From 2007 to 2010, he was Associate Professor at University of Minnesota (UMN), Twin-Cities, where he currently holds the Vincentine Hermes-Luh Chair in Electrical Engineering. He is the Director of Graduate Studies in the Electrical and Computer Engineering Department at UMN. Dr. Salapaka was the recipient of the 1997 National Science Foundation CAREER Award, and the 2001 Iowa State University Young Engineering Faculty Research Award. His research interests are in control and systems theory, nanotechnology and molecular biology. His research is supported by numerous grants form National Science Foundation, Google, and ARPA-E.

Go top

Title : Power Divider
Speaker : Dr.Sairaj Dhople
Date : 05/08/2016
Venue : C 241 MMCR, EE
Abstract: This talk presents analytical closed-form expressions that uncover the contributions of nodal active- and reactive-power injections to the active- and reactive-power flows on transmission lines in an AC electrical network. Paying due homage to current- and voltage-divider laws that are similar in spirit, we baptize these as the power divider laws. Derived from a circuit-theoretic examination of AC power-flow expressions, the constitution of the power divider laws reflects the topology and voltage profile of the network. We demonstrate the utility of the power divider laws to the analysis and control of power networks by highlighting applications to transmission-network allocation, transmission-loss allocation, tracing the flow of power, and identifying feasible injections while respecting line active-power flow set points.
Speaker bio: Sairaj Dhople received the B.S., M.S., and Ph.D. degrees in electrical engineering from the University of Illinois, Urbana-Champaign, IL, USA, in 2007, 2009, and 2012, respectively. Currently, he is an Assistant Professor in the Department of Electrical and Computer Engineering at the University of Minnesota, Minneapolis, MN, USA, where he is affiliated with the Power and Energy Systems Research Group. His research interests include modeling, analysis, and control of power electronics and power systems with a focus on renewable integration. Sairaj received the National Science Foundation CAREER award in 2015, and he currently serves as an Associate Editor for the IEEE Transactions on Energy Conversion.

Go top

Title : Bayesian Nonparameric Modeling of Temporal Coherence for Entity-driven Video Analytics.
Speaker : Dr. Adway Mitra
Date : 05/08/2016
Venue : C 241 MMCR, EE
Abstract: Due to the advent of video-sharing sites like Youtube, online user-generated video content is increasing very rapidly. To simplify search of meaningful information from such huge volume of content, Computer Vision researchers have started to work on problems like Video Summarization and Scene Discovery from videos. People understand videos based on high-level semantic concepts. But most of the current research in video analytics makes use of low level features and descriptors, which may not have semantic interpretation. We have aimed to fill in this gap, by modeling implicit structural information about videos, such as spatio-temporal properties. We have represented videos as a collection of semantic visual concepts which we call “entities”, such as persons in a movie.To aid these tasks, we have attempted to model the important property of “temporal coherence”, which means that adjacent frames are likely to have similar visual features, and contain the same set of entities. Bayesian nonparametrics is a natural way of modeling all these, but they have also given rise to the need for new models and algorithms.
A tracklet is a spatiotemporal fragment of a video a set of spatial regions in a short sequence of consecutive frames, each of which enclose a particular entity. We first attempt to find a representation of tracklets to aid tracking entities across videos using region descriptors like Covariance Matrices of spatial features making use of temporal coherence. Next, we move to modeling temporal coherence at a semantic level. Each tracklet is associated to an entity. Spatio-temporally close but nonoverlapping tracklets are likely to belong to the same entity, while tracklets that overlap in time can never belong to the same entity. The aim is to cluster the tracklets based on the entities associated with them, with the goal of discovering the entities in the video along with all their occurrences. We represented an entity with a mixture component, and proposed a temporally coherent version of Chinese Restaurant Process (TCCRP) that can encode the constraints easily. TCCRP shows excellent performance on person discovery from TV series videos. We also discuss semantic video summarization, based on entity discovery. Next, we consider entity-driven temporal segmentation of the video into scenes, where each scene is modeled as a sparse distribution over entities. We propose EntScene: a generative model for videos based on entities and scenes, and also an inference algorithm based on Dynamically Blocked Gibbs Sampling. We experimentally demonstrate significant improvements in terms of segmentation and scene discovery compared to alternative inference algorithms.
We briefly turn to modeling temporal coherence in hierarchically grouped sequential data, such as word-tokens grouped into sentences, paragraphs, documents etc in a text corpus. We attempt Bayesian modeling for such data, with application to multi-layer segmentation of a set of news transcripts- into broad categories (like politics, sports etc) and individual stories. We consider a Markovian and explicit-duration (semi-Markov) approach for this purpose, and provide an efficient inference algorithm for both.
Speaker bio: Adway Mitra has completed his PhD in Machine Learning and Computer Vision from CSA Department, Indian Institute of Science. Before that, he earned a Masters' Degree in Computer Science from the same department, and Bachelors' Degree from Jadavpur University, Kolkata. He is currently working on modeling complex spatio-temporal processes such as climatic processes as a postdoctoral fellow in International Center for Theoretical Sciences, Bangalore.

Go top

Title : HV SiC Power Devices – A Panacea for Enabling MV Power Conversion Application
Speaker : Dr.Subhashish Bhattacharya
Date : 22/07/2016
Venue : C 241 MMCR, EE
Abstract:The advent of WBG (SiC and GaN) devices is poised to revolutionize the power electronics applications – both in the low power and low voltage applications, as well as the Medium Voltage (MV) and High Voltage (HV) applications at high power levels. This talk outlines opportunities for HV SiC devices for MV Power Converters and utility applications and the challenges to apply these HV SiC devices successfully. The talk will focus on SiC devices based power electronics applications with SiC device voltage ranges from 1200 V to 1700 V MOSFETs and JBS diodes, through HV 10 kV - 15 kV MOSFETs, JBS diodes, and 15 kV SiC IGBTs. The talk will develop understanding of the high frequency switching characteristics of these SiC devices and their potential application areas. The potential and challenges of the HV 10-15 kV devices to enable MV power conversion systems, including the large market space of MV motor drives will be discussed. The utility applications area of FACTS and VSC based HVDC and in particular MVDC systems can be enabled by these HV SiC devices. Challenges in adopting these HV SiC devices for MV power conversion in terms of magnetics, capacitors, insulation materials, etc. will be discussed.
Speaker bio:Subhashish Bhattacharya received his B.E. (Hons), M.E. and PhD degrees in Electrical Engineering from University of Roorkee, India in 1986, Indian Institute of Science (IISc), Bangalore, India in 1988, and University of Wisconsin-Madison in 2003, respectively. He worked in the FACTS (Flexible AC Transmission Systems) and Power Quality group at Westinghouse R&D Center in Pittsburgh which later became part of Siemens Power Transmission & Distribution, from 1998 to 2005. He joined the Department of Electrical and Computer Engineering at North Carolina State University (NCSU) in August 2005, where he is the ABB Term Professor, and also a founding faculty member of NSF ERC FREEDM systems center (www.freedm.ncsu.edu), Advanced Transportation Energy Center [ATEC] (www.atec.ncsu.edu) and the newly established DOE initiative on WBG based Manufacturing Innovation Institute – PowerAmerica - at NCSU. His research interests are Solid-State Transformers, MV power converters, FACTS, Utility applications of power electronics and power quality issues; high-frequency magnetics, active filters, and application of new power semiconductor devices such as SiC for converter topologies.

Go top

Title : Height Information Extraction from Speech Data
Speaker : Dr. Deepu Vijayasenan
Date : 22/07/2016
Venue : C 241 MMCR, EE
Abstract: Physical parameter estimation from human speech is a very important task that finds its application in forensics. Eg., such systems would be able to derive some information about anonymous callers. In this talk, height estimation of a person from speech data is addressed. Support Vector regression models are derived from various different features.. Conventional short term spectral features such as Mel filter bank energies as well as prosodic features like formant locations, harmonic features and their statistics are explored.
Speaker bio: He is an assistant professor at NITK, Surathkal since Jan 1, 2013. Prior to that he spent two years at University of Saarbreucken, Saarland as Post doctoral researcher. He did his PhD at EPFL, Lausanne. His main interests include machine learning and speech signal processing.

Go top

Title : Advanced Power Electronics and Motor Drives for Transportation Electrification: Opportunities and Challenges
Speaker : Dr. Sheldon S. Williamson
Date : 15/07/2016
Venue : C 241 MMCR, EE
Abstract: This presentation will largely showcase the launch and current status of UOIT’s new research program on transportation electrification and electric energy storage systems. The new, one-of-a-kind Advanced Storage Systems and Electric Transportation (ASSET) research facility will also be introduced. UOIT’s novel research initiatives within the newly established Smart Transportation Electrification and Energy Research (STEER) group has the potential of providing a significant link for future progress with regards to efficiency and performance improvement of electric transportation and personal e-mobility vehicles.
The main areas of research focus within the STEER group includes:
• Advanced electric energy storage systems: Lithium-ion battery management; improvement of battery cycle life; increasing driving range of electric vehicles (EVs); high-power ultracapacitor power management for electric mass transit vehicles – buses, trains, and trams.
• Smart EV charging infrastructures: Fast charging of EVs; charging from renewable energy (solar-photovoltaics); integration with the smart grid.
• Mass transit electrification: Fast charging of buses, trains, and trams using wireless power transfer; powering buses and trams using high-power ultracapacitors, including new power management strategies; electric machines and motor drives.
• E-mobility: Electric bikes (e-bikes), e-golf carts, UOIT campus vehicle electrification plans. The above mentioned research initiatives will be briefly described in the presentation and industry-specific projects within the STEER group will be highlighted. The high-level goals of the presentation will be focused on advanced power electronics solutions for EV traction batteries and ultracapacitors as well as plugged and wireless charging/inductivepower transfer (IPT) technologies. Novel motor drive technologies and controller designs for high-voltage DC power on board electric mass transit buses, trains, and trams will also be presented. UOIT’s new Canada Research Chair (CRC) program incl udes several novel initiatives in the areas of transportation electrification, and is built upon the expertise and know-how of the STEER group in a number of promising interdisciplinary areas related to power electronics and motor drives solutions for e-mobility and e-transportation.
Speaker bio: Sheldon S. Williamson received his Bachelors of Engineering (B.E.) degree in Electrical Engineering with high distinction from University of Mumbai, Mumbai, India, in 1999. He received the Masters of Science (M.S.) degree in 2002, and the Doctor of Philosophy (Ph.D.) degree (with Honors) in 2006, both in Electrical Engineering, from the Illinois Institute of Technology, Chicago, IL, specializing in automotive power electronics and motor drives, at the Grainger Power Electronics and Motor Drives Laboratory. From June 2006 to May 2011, Dr.Williamson held a Tenure-track Assistant Professor position in the Department of Electrical and Computer Engineering, at Concordia University, in Montreal, Canada. Also, from June 2011 to June 2014, Dr. Williamson held a tenured Associate Professor position at Concordia University. Currently, Dr. Williamson is an Associate Professor at the Smart Transportation Electrification and Energy Research (STEER) group, within the Department of Electrical, Computer, and Software Engineering, at the University of Ontario-Institute of Technology (UOIT), in Oshawa, Ontario, Canada. He also holds the prestigious NSERC Canada Research Chair position in Electric Energy Storage Systems for Transportation Electrification, at UOIT, since Sept. 2015. His main research interests include advanced power electronics and motor drives for transportation electrification, electric energy storage systems, and electric propulsion. Dr. Williamson has offered numerous conference tutorials, lectures, and short courses in the areas of electric transportation, electric energy storage systems, as well as automotive power electronics, and motor drives. He is the principal author/co-author of over 150 journal and conference papers. He is also the author/co-author of several books and b ook chapters on electric transportation and energy storage systems. He has been selected as the General Chair for the IEEE International Conference on Industrial Technology, to be held in Toronto, Ontario, in May 2017. In addition, Dr. Williamson has also served on the technical program committees of several IEEE conferences in the past. Dr. Williamson is the beneficiary of numerous awards and recognitions. He was the recipient of the prestigious “paper of the year” award, for the year 2006, in the field of Automotive Power Electronics, from the IEEE Vehicular Technology Society (IEEE VTS). In addition, he has also received several “best paper” awards for papers he has co-authored with his graduate research students in major IEEE conferences. He was awarded the prest igious Sigma Xi/IIT Award for Excellence in University Research, for the academic year 2005-2006. In 2006, he also received the “Best Research Student” award, Ph.D. category, within the ECE Department, at the Illinois Institute of Technology, in Chicago, IL. Dr. Williamson is a Senior Member of the IEEE and is a Member of the IEEE Power Electronics Society (IEEE PELS) and the IEEE Industrial Electronics Society (IEEE IES). He also currently serves as a Distinguished Lecturer of the IEEE Vehicular Technology Society (VTS). He is an Associate Editor for the IEEE Transactions on Industrial Electronics, IEEE Transactions on Power Electronics, IEEE Transactions on Transportation Electrification, and the IEEE Journal of Emerging and Selected Topics in Power Electronics. He is a Member of the prestigious IEEE Transportation Technologies.

Go top

Title : Multimodal behavioral informatics for health applications
Speaker : Prof. Shrikanth (Shri) Narayanan
Date : 07/07/2016
Venue : C 241 MMCR, EE
Abstract: The confluence of sensing, communication and computing technologies is allowing capture and access to data, in diverse forms and modalities, in ways that were unimaginable even a few years ago. These include data that afford the analysis and interpretation of multimodal cues of verbal and non-verbal human behavior to facilitate human behavioral research and its translational applications in health. These data not only carry crucial information about a person’s intent, identity and trait but also underlying attitudes, emotions and other mental state constructs. Automatically capturing these cues, although vastly challenging, offers the promise of not just efficient data processing but in tools for discovery that enable hitherto unimagined scientific insights, and means for supporting diagnostics and interventions. Recent computational approaches that have leveraged judicious use of both data and knowledge have yielded significant advances in this regards, for example in deriving rich, context-aware information from multimodal signal sources including human speech, language, and videos of behavior. These are even complemented and integrated with data about human brain and body physiology. This talk will focus on some of the advances and challenges in gathering such data and creating algorithms for machine processing of such cues. It will highlight some of our ongoing efforts in Behavioral Signal Processing (BSP)—technology and algorithms for quantitatively and objectively understanding typical, atypical and distressed human behavior—with a specific focus on communicative, affective and social behavior. The talk will illustrate Behavioral Informatics applications of these techniques that contribute to quantifying higher-level, often subjectively described, human behavior in a domain-sensitive fashion. Examples will be drawn from mental health and well being realms such as Autism Spectrum Disorders, Couple therapy, Depression and Addiction counseling.
Speaker bio: Shrikanth (Shri) Narayanan is Andrew J. Viterbi Professor of Engineering at the University of Southern California, where he is Professor of Electrical Engineering, and jointly in Computer Science, Linguistics, Psychology, Neuroscience and Pediatrics, and Director of the Ming Hsieh Institute. Prior to USC he was with AT&T Bell Labs and AT&T Research. His research focuses on human-centered information processing and communication technologies. He is a Fellow of the Acoustical Society of America, IEEE, and the American Association for the Advancement of Science (AAAS). Shri Narayanan is Editor in Chief for IEEE Journal on Selected Topics in Signal Processing, an Editor for the Computer, Speech and Language Journal and an Associate Editor for the IEEE Transactions on Affective Computing, the Journal of Acoustical Society of America, and the APISPA Transactions on Signal and Information Processing having previously served an Associate Editor for the IEEE Transactions of Speech and Audio Processing (2000-2004), the IEEE Signal Processing Magazine (2005-2008), the IEEE Transactions on Signal and Information Processing over Networks (2014-2015) and the IEEE Transactions on Multimedia (2008-2012). He is a recipient of several honors including the 2015 Engineers Council’s Distinguished Educator Award, the 2005 and 2009 Best Transactions Paper awards from the IEEE Signal Processing Society and serving as its Distinguished Lecturer for 2010-11, and as an ISCA Distinguished Lecturer for 2015-16. With his students, he has received a number of best paper awards including a 2014 Ten-year Technical Impact Award from ACM ICMI and Interspeech Challenges in 2009 (Emotion classification), 2011 (Speaker state classification), 2012 (Speaker trait classification), 2013 (Paralinguistics/Social Signals), 2014 (Paralinguistics/Cognitive Load) and in 2015 (Non-nativeness detection). He has published over 700 papers and has been granted 17 U.S. patents.

Go top

Title : Enhancing clinical voice assessment with smartphone-based ambulatory voice monitoring
Speaker : Dr Daryush Mehta
Date : 01/07/2016
Venue : C 241 MMCR, EE
Abstract: An estimated 30% of the adult U.S. population suffers from a voice disorder at some point in their lives and often experience significant communication disabilities with far-reaching social, professional, and personal consequences. Most voice disorders are chronic or recurring conditions and result from inefficient and/or abusive patterns of vocal behavior, termed vocal hyperfunction. Thus, an ongoing clinical research goal is the prevention, diagnosis, and treatment of vocal hyperfunction through noninvasive, long-term monitoring of an individual’s daily voice use. During this talk, I will present my work investigating vocal hyperfunction in voice patients and matched healthy controls using smartphone-based ambulatory voice monitoring. Voice use and vocal function measures were derived from neck-surface acceleration recordings using vocal dose theory and novel impedance-based acoustic modeling to yield glottal airflow estimates. Results indicate that the clinical treatment of vocal hyperfunction would be improved by the ability to unobtrusively monitor and quantify detrimental voice use and simultaneously provide real-time biofeedback, thereby facilitating the learning of healthier vocal behaviors. Future research aims to enhance clinical voice assessment through integrating innovations in wearable sensor technology and laryngeal endoscopic imaging.
Speaker bio: Daryush Mehta is Assistant Biomedical Engineer at the Center for Laryngeal Surgery and Voice Rehabilitation at the Massachusetts General Hospital (MGH), Instructor in Surgery at Harvard Medical School, and Adjunct Assistant Professor at the MGH Institute of Health Professions. Daryush received his PhD in Speech and Hearing Bioscience and Technology from the Harvard–MIT Division of Health Sciences and Technology (2010), Master's degree in Electrical Engineering and Computer Science from MIT (2006), and Bachelor's degree in Electrical Engineering from University of Florida (2003). Daryush’s research interests include high-speed video imaging of vocal vibration, speech signal processing, and clinical voice disorder assessment. Read all about his work at http://web.mit.edu/dmehta/www.

Go top

Title : Detection and analysis of whispered speech
Speaker : Nisha Meenakshi G
Advisor:Dr. Prasanta Kumar Ghosh
Date : 24/06/2016
Venue : C 241 MMCR, EE
Abstract:Whispering is an indispensable form of communication that emerges in private conversations as well as in pathological situations. In conditions like laryngectomy and neurogenic disorders such as Vocal chord Paresis and Spasmodic Disphonia one or both of the vocal chords are affected, leading to the patient’s voice becoming a breathy and rough whisper. Thus, the conversion of whispered speech to neutral speech would be potentially beneficial for these patients. In this talk, two key components of such a conversion- the detection and analysis of whispered speech will be discussed. Whisper activity detection (WAD) detects whispered speech in the presence of noise. In this regard, a novel feature, that utilizes the long-term logarithmic energy variation (LTLEV) of the sub-band signal is proposed. It is found to be effective in distinguishing noisy whispered speech from noise, under eight different noise conditions and four different signal-to-noise ratios (SNR). In the second part of the talk, the discriminative analysis of the voiced and unvoiced consonants in whispered speech space, across six different Indian languages, will be discussed. Some key findings of this study include a reduction in the whispered consonant space, reduced discrimination between whispered voiced and unvoiced phonemes and some interesting language specific findings.
Whispering is an indispensable form of communication that emerges in private conversations as well as in pathological situations. In conditions like laryngectomy and neurogenic disorders such as Vocal chord Paresis and Spasmodic Disphonia one or both of the vocal chords are affected, leading to the patient’s voice becoming a breathy and rough whisper. Thus, the conversion of whispered speech to neutral speech would be potentially beneficial for these patients. In this talk, two key components of such a conversion- the detection and analysis of whispered speech will be discussed. Whisper activity detection (WAD) detects whispered speech in the presence of noise. In this regard, a novel feature, that utilizes the long-term logarithmic energy variation (LTLEV) of the sub-band signal is proposed. It is found to be effective in distinguishing noisy whispered speech from noise, under eight different noise conditions and four different signal-to-noise ratios (SNR). In the second part of the talk, the discriminative analysis of the voiced and unvoiced consonants in whispered speech space, across six different Indian languages, will be discussed. Some key findings of this study include a reduction in the whispered consonant space, reduced discrimination between whispered voiced and unvoiced phonemes and some interesting language specific findings.

Go top

Title : Synchrophasor Technology Applications to Power Stability Monitoring and Control
Speaker : Prof. S C Srivastava
Date : 20/06/2016
Venue : C 241 MMCR, EE
Abstract: Increased size and complexity of the electric power system networks pose various operating challenges in maintaining continuity, quality and security of the system. Several events of blackouts in recent past, worldwide, have necessitated the use of more intelligent and automated systems for online monitoring, protection and control of the power systems. Wide Area Monitoring, and Control System (WAMCS), employing synchrophasor technology, are being increasingly used in power system networks, which enhances system observability and facilitates use of several online decision tools for monitoring and control of the power system stability and security. This technology has been tried at pilot level and now being deployed in large number in Indian power networks.This talk will discuss some of the concepts of the synchrophasor measurements and its applications, specifically relevant to the system stability monitoring and control, along with an overview of plans for its deployment in the Indian power grid.
Speaker bio: Prof. Srivastava received B.Tech. degree in Electrical Engineering in 1976 from Institute of Technology, Banaras Hindu University, and Ph.D. from Indian Institute of Technology (IIT) Delhi. He worked at Engineers India Limited New Delhi, a consultancy organization, during Nov.1976-Nov.1988 in its Project Engineering and Engineering Technology Development divisions. Since November 1988, he is a faculty member in the Department of Electrical Engineering at IIT Kanpur, where he became ‘Professor’ in Dec. 1995. He also served as Head of Electrical Engineering Department during Jan. 2000 to Dec. 2002, Dean of Research and Development during Jan. 2005 to Jan. 2008, and Deputy Director during Sep. 2011 to Sep. 2014 at IIT Kanpur. During August 2008-July 2009, he was as a ‘Visiting Research Professor’ in the ECE Department at Mississippi State University, USA and also as a Faculty member at Asian Institute of Technology, Bangkok, Thailand during 1996-97, on leave from IIT Kanpur. He also held ‘P.K. Kelkar Chair Professor’ position at IIT Kanpur, and now holding ‘Ministry of Labour and Employment Chair Professor’ position. He has supervised 26 Ph.D. and 60 Masters theses in the Power Systems area. He has published about 300 papers in refereed journals and conference proceedings. His research interests include Power System Stability and Security Analysis, Synchrophasor Applications, Power System Restructuring and AC/DC Microgrid. He is a Fellow of the Indian National Academy of Engineering (INAE), Institution of Engineers (India) & IETE (India), and Senior member of the IEEE.

Go top

Title : Acoustic based speech rate estimation using data-driven approaches
Speaker : Mr.Chiranjeevi Yarra
Advisor : Dr. Prasanta Kumar Ghosh
Date : 17/06/2016, 4 pm
Venue : C 241 MMCR, EE
Abstract: Acoustic feature based speech (syllable) rate estimation is important problem in automatic speech recognition (ASR), computer assisted language learning (CALL) and fluency analysis. A typical solution for the problem consists of two stages. The first stage involves computing a short-time feature contour such that most of the peaks of the contour correspond to the syllabic nuclei. In the second stage, the peaks corresponding to the syllable nuclei are detected. In this work, we address both the problems separately in a data-driven manner. For the first problem, temporal correlation selected sub-band correlation (TCSSBC) is often used as a feature contour for the speech rate estimation in which correlation within and across a few selected sub-band energies are computed. Instead of a fixed set of sub-bands, we learn them using a dictionary learning approach. Similarly, instead of the energy contours, we use the activation profile from the learned dictionary elements. We found that the peaks detected from the activation profiles significantly improve the speech rate estimation when combined with the traditional TCSSBC approach using a proposed peak-merging strategy. For the second problem, rule based approaches are often used for detecting the peaks of the feature contour. Instead of the peak detection, we perform a mode-shape classification, which is formulated as a supervised binary classification problem – mode-shapes representing the syllabic nuclei as one class and remaining as the other. We use the temporal correlation and selected sub-band correlation (TCSSBC) feature contour and the mode-shapes in the TCSSBC feature contour are converted into a set of feature vectors using an interpolation technique. A support vector machine classifier is used for the classification. Experiments are performed separately using Switchboard, TIMIT and CTIMIT corpora and the proposed methods outperform the best of the existing methods.

Go top

Title : Emerging trends in Computational Imaging and Displays
Speaker : Dr. Ram Narayanswamy
Date : 15/06/2016, 10.30Am
Venue : C 241 MMCR, EE
Abstract: In this talk we will explore how the evolution of photography and personal media has reached an inflection point in its evolution. Recent developments of imaging systems have created rich immersive data sets that allow for new forms of media and interaction with our digital environment. We will give a glimpse into research at Intel Labs that explores new computational imaging capabilities in the area of multi-camera systems and interactive visual experiences. We will highlight new imaging technologies that made their way from research into recent Intel products, and what capabilities these offer to the researchers and developers. In closing, we will give an outlook on where we see this trend going and the opportunities this offers for research in computational imaging.
Speaker bio: Ram Narayanswamy is currently part of Intel’s Computational Imaging Lab. He started his career at NASA Langley Research Center in the Visual Image Processing lab working on imaging system design and optimization. There he co-authored a paper titled “Characterizing digital image acquisition devices”, better known today as the “slanted-edge test” a de facto standard to measure camera MTF. He was one of the early members of CDM Optics, a Boulder start-up which pioneered Wavefront Coding and the field of computational imaging. Upon CDM’s acquisition by OmniVision Technologies, Ram led the effort to productize Wavefront Coding for the mobile phone segment. Later, he joined Aptina Imaging where he helped bring the world’s first performance 720p reflowable cameras modules with molded glass optics to market – a complete-camera that ships in tape and reel! While at Aptina, Ram also led their effort in Array cameras. He is a Program Chairman’s for The Optics Society’s Annual Congress on Imaging and Applied Optics to be held in Heidelberg, Germany July 2016. Ram has a PhD from the University of Colorado-Boulder and BS from the National Institute of Technology – Trichy. He has over 40 publications and loves this golden age of imaging and looks forward to ushering the platinum age.

Go top

Title: Networked and distributed CPS: control under network constraints and networked transportation systems
Speaker: Dr. Pavan Tallapragada
Date: 14/06/2016, 4 pm
Venue: C 241 MMCR, EE
Abstracts: With various enabling technologies, large scale distributed and pervasive control systems are becoming a reality, such as wireless automation networks in large industries, building HVAC systems, distributed smart energy grids, next generation cars and networked transportation systems. The complexity of such systems necessitates an integrated approach to the design of control, communication and computing components. The recognition of this fact has given rise to the paradigm of Cyber Physical Systems (CPS) in recent years.
A common problem in large scale networked and distributed systems is that of constrained or limited resources such as energy/power, communication and computational capabilities. In this presentation, I will talk about my work on opportunistic state-triggered control, which seeks to design controllers under communication constraints.
I will also talk about a CPS domain, namely networked transportation systems. Emerging technologies such as networked and computer controlled vehicles offer the opportunity to design novel systems of transportation and traffic control. Such systems have the potential to hugely improve safety, travel ease, travel times and energy consumption. I will illustrate this with the example of a hierarchical-distributed system for coordination of intersection traffic.
Speaker Bio : Pavan Tallapragada is a postdoc since 2014 at the University of California, San Diego. He received B.E. in 2005 from SGGSIE&T, SRTMU, Nanded and MSc. (Engg.) in 2007 from the Indian Institue of Science, Bangalore, both in Instrumentation. He received Ph.D. in Mechanical Engineering from the University of Maryland, College Park in 2013. He worked briefly as a Junior Research Fellow at the National Institute of Advanced Studies, Bangalore in 2007 and as a Technical Specialist at Panacea Medical Technologies, Bangalore in 2008.
His research interests include networked and distributed control systems and Cyber Physical Systems (CPS). Specifically, he is interested in topics such as control under communication constraints, privacy in CPS and networked transportation systems.

Go top

Title : A QR Decomposition Approach to Factor Modelling
Speaker : Dr. Bharath Bhikkaji
Date : 14/06/2016, 11.30 am
Venue : C 241 MMCR, EE
Abstracts : An observed K-dimensional series {y_{n}, n=1,....N} is expressed in terms of a lower p- dimensional latent series called factors f_{n} and random noise epsilon_{n}. The equation, y_{n}=Qf_{n}+ epsilon_{n} is taken to relate the factors with the observation. The goal is to determine the dimension of the factors, p, the factor loading matrix, Q, and the factors f_{n}. Here, it is assumed that the noise co-variance is positive definite and is allowed to be correlated with the factors. This paper proposes the use of QR decomposition instead of the standard Eigenvalue Decomposition (EVD) for determining the model order p and the loading matrix Q. Estimation of the model order p is formulated as a Numerical Rank determination problem. Rank Revealing QR (RRQR) decomposition is used for estimating the loading matrix Q. The asymptotic performances of the estimates of p, Q and f_{n} are analyzed by letting K, N tend to infinity. The asymptotic rates, and empirical results, suggests that the proposed technique is both computationally efficient and accurate.
Speaker Bio : Dr. Bharath Bhikkaji is currently working as an Assistant Professor in the Dept. of Elec, Engg in IIT Madras. He received his Ph.D. degree in Signal Processing from the Uppasla University, Uppsala, Sweden, in the year 2004, and M.E. degree from Electrical Communication Engineering, Indian Institute of Science, Bangalore. He worked as a Research Academic at the School of Electrical Engineering and Computer Science, University of Newcastle, Newcastle, Australia from 2004-2008. His research interests include System Identification, Robust Control and Active noise and Vibration control of Flexible structures.

Go top

Title : Understanding the Perception and Impact of Social Signals
Speaker : Dr Tanaya Guha
Date : 14/06/2016, 10 am
Venue : C 241 MMCR, EE
Abstracts : Understanding the diverse behavioral and social patterns that exist around us can improve our social experience and interaction in many ways. With this broader objective in mind, we focus on two important but different problems relevant to social signal processing. The first problem involves social perception and its impact on media. Here, we attempt to quantify a very subjective and often not-so-well-defined concept of gender representation and bias. Starting with content analyzing popular Hollywood movies, we show how gender representation can be objectively measured from multimodal cues. In the next part, we focus on the production and perception of social signals in autism. Our goal is to understand how behavioral signals, such as facial expressions, produced by children with autism are perceived by healthy observers. Methodologies, approaches, results and challenges related to both of these problems will be discussed.
Speaker Bio : Tanaya Guha is currently an Assistant Professor in the department of Electrical Engineering at IIT Kanpur. She is also a part of the computer vision group at IITK. Prior to joining IITK, she was a postdoctoral fellow at the Signal Analysis and Interpretation Lab (SAIL), University of Southern California, Los Angeles from 2013 to 2015. She has received her PhD in Electrical and Computer Engineering from the University of British Columbia, Vancouver in 2013. She was a recipient of Mensa Canada Woodhams memorial scholarship, Google Anita Borg memorial scholarship and Amazon Grace Hopper celebration scholarship. Her current research interests include human emotion and behavior analysis, multimodal signal processing, and image analysis.

Go top

Title : Introduction to Smart Grids
Speaker : Prof. S A Khaparde
Date : 03/06/2016, 4pm
Venue : C 241 MMCR, EE
Abstract: Smart grids are evolving for the last two decades. Many countries, including India, are investing large funds in this technology. The talk will introduce the fundamentals of smart grids. Pilot project in India will be discussed. The research opportunities will be identified at the end of the talk.
Speaker bio: Prof. Shrikrishna A. Khaparde received the Ph.D. degree from IIT, Kharagpur in 1981. He is a Professor with the Department of Electrical Engineering, IIT Bombay. He was awarded the DSK Energy Award in 2009, for outstanding contribution in energy sector, by The Institution of Engineers (India), Pune Centre.
He is a Consultant to Maharashtra Electricity Regulatory Commission (MERC), Indian Energy Exchange, Power Grid Corporation of India Ltd., etc. He has coauthored the book "Computational Methods for Large Sparse Power Systems Analysis: An Object Oriented Approach" (Norwell, MA, USA: Kluwer, 2001). He has published over 200 research papers in leading journals and conferences. His research interests include restructured power systems, distributed generation, renewable energy policies, and Common Information Model (CIM) implementation in India.
Prof. Khaparde is a member of advisory committees to Maharashtra Electricity Regulatory Commission and Indian Energy Exchange. He is an Editor of the International Journal of Emerging Electrical Power Systems. He is a member of IEC TC57 for working groups 13 and 16, representing India. He is a BIS LITD-10 Committee Member and Chair of the Working Group (WG3) on CIM.

Go top

Title : Face Recognition in Unconstrained Environment
Speaker : Mr. Sivaram Prasad Mudunuri
Advisor : Dr.Soma Biswas
Date : 27/05/2016 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Unconstrained face recognition is an important area of research in the field of computer vision and pattern recognition. The increasing use of surveillance cameras for addressing security concerns has led to increased demand for robust face recognition systems. The images captured by the surveillance cameras usually have poor resolution, uncontrolled pose and illumination conditions which makes the task of recognizing these faces extremely challenging. Significant attention has been devoted to addressing one or more of the different challenges like poor illumination, non-frontal pose, expression, etc. But addressing all these challenges together is essential in many real world applications. In this presentation, an overview on recent state-of-the-art approaches and our algorithm of recognizing low resolution faces with variations in pose and illumination will be discussed. To be specific, our work assumes that, the faces are already detected and cropped from the given image. Though the motivation behind the proposed algorithm is to match the facial images captured under real outdoor surveillance scenario, we are able to match the faces captured under controlled environment and limited surveillance quality cameras only.

Go top

Title : Innovative Applications from Analysis of Frequency Response of Transformer Winding
Speaker : Dr. Saurav Pramanik
Date : 25/05/2016 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract:Power transformers are vital links in an electrical network, whose normal working is paramount for uninterrupted operation of a power system. In the post-deregulated era, it emerges that condition monitoring of HV apparatus followed by meaningful diagnosis, is the ideal way to avoid unplanned outages, optimize assets, reduce operating costs and thus operate the plant efficiently. Over the years, many powerful tools have become available for assessing the status of transformers. Among them, frequency response measurement (also called FRA), is the most-sensitive to detect mechanical damage occurring to the windings and core. In spite of its existence for many years, FRA continues to remain as a monitoring tool. It has lot of untapped potential and inherent abilities but the relevant issues have not received the attention it deserves, and research groups working in these areas are also few. So, research efforts are essential to address these issues. In this talk the speaker will present few innovative methods to extract information embedded in the frequency/time domain response of the transformer winding and utilizes them for suggesting simple, but, yet impressive solutions to a few tasks which have until now been thought difficult, if not impossible, to resolve. Two such applications are (i) Investigate indirect measurement of the series capacitance of a transformer winding using the measured response, and (ii) A new diagnostic approach for fault detection in transformer winding using tank current measurement. Details will be presented in the seminar. In the end, speaker will also highlight his postdoctoral research on energy efficient transformers carried out at ABB Corporate Research, Sweden.
Speaker bio: Dr. Saurav Pramanik is currently working as a visiting assistant professor in the Department of Electrical Engineering, IIT Kharagpur. Earlier He had been working as a postdoctoral scientist in Corporate Research ABB in completed his Masters and Ph.D. from the Department of Electrical Engineering, IISc., Bangalore in 2010 and 2013 respectively. He has a B.E. degree in Electrical Engineering from Jadavpur University, Kolkata in 2006. His research area mainly includes monitoring and diagnostics of power transformers. He has also worked on problems related to core magnetics in power transformer with the objective to reduce the core loss, as well as, some other stray losses in transformer winding. He was conferred national and international awards for his research contribution.

Go top

Title : Research Activities in High Voltage Laboratory @IITM
Speaker : Prof. R. Sarathi, Dept of Electrical Engineering, IIT Madras
Date : 25/05/2016 11.30 am
Venue : HV Lab, Seminar Hall

Go top

Title : A matrix free iterative reconstruction algorithm for Photoacoustic Tomography
Speaker : Rejesh N.A
Advisor : Dr. Muthuvel Arigovindan
Date : 20/05/2016 (4.00 pm - 5.00 pm)
Venue : MMCR (C241)
Abstract: Photoacoustic tomography (PAT) is an emerging biomedical imaging modality that shows great potential for preclinical research and clinical practice. PAT is a hybrid technique based on optical excitation and ultrasonic detection resulting from absorption of light. Absorption of photons by biomolecules thermoelastically induces pressures waves, which are received by acoustic detectors to form images. Because ultrasound scatters much less than light in tissue, PAT generates high-resolution images in both the optical ballistic and diffusive regimes. Existing approaches to image reconstruction in PAT are computationally intensive, and/or cannot effectively mitigate the effects of measurement data incompleteness and noise. In this presentation, an iterative reconstruction algorithm for PAT images, which is both computationally and memory efficient will be presented.

Go top

Title: Graph Signal Processing (Lecture Video)
Speaker: Prof. K. R. Ramakrishnan
Venue: C 241 MMCR, EE
Date and time: 13th May 2016, 4:00pm - 5.00pm
Abstract: In applications such as social, energy, transportation, sensor,and neuronal networks, high-dimensional data naturally reside on the vertices of weighted graphs. The emerging field of Graph signal processing draws the concepts from traditional DSP to process such signals on graphs. In this seminar we discuss ways to define graph spectral domains, which are the analogs to the classical frequency domain, and highlight the importance of incorporating the irregular structures of graph data domains when processing signals on graphs. We also generalize fundamental operations such as filtering, translation, modulation, dilation, and down sampling to the graph setting.

Go top

Title: Google little box challenge
Speaker: Dr. Rajesh Ghosh
Venue: EE 217 (MMCR)
Date and time: 6th May 2016, 4:00pm - 5.00pm
Abstract: Inverters convert DC from solar/battery into AC at home, automobiles, UPS drives and many other applications. Normal inverters are big in size (picnic cooler). Smaller inverters enable more solar-powered homes, more efficient distributed grids, compact automobiles and low cost, help bring electricity to the most remote parts of the planet. Little Box Challenge was a worldwide open competition to build world's first smallest 2kW inverter at $1,000,000 prize. It was jointly sponsored by Google and IEEE. From Schneider Electric, we participated into this contest, and secured the second position. Our achieved power density is 100W/cubic inch, which is just double the required power density set by Google. In this presentation, detailed technical challenges and how we address them will be discussed in detail. Also, the approaches used by others will be discussed.
Speaker bio: Rajesh Ghosh received the Ph. D degree in Electrical Engineering from The Indian Institute of Science Bangalore in 2007. He joined APC by Schneider Electric, Bangalore in 2007 as an Electrical Engineer. Currently he is a staff Electrical Engineer in the IT business of Schneider Electric. His interests include high-frequency power conversion and digital control. Prior to joining APC he worked with CESC Ltd. Calcutta, India and GE Global Research center Bangalore, India. He has published many journal and conference papers in IEEE and holds many US patents. His research interest includes High-frequency power conversion, application of wide band gap devices in future power converters, digital control and soft switching technique.

Go top

Title: Machine Vision Applications: Challenges and Opportunities
Speaker: Dr. Dinesh Ramegowda
Venue: EE 217 (MMCR)
Date and time: 15th Apr 2016, 4:00pm - 5.00pm
Abstract: Man from the beginning of time, tried to automate things for comfort, accuracy, precision and speed. Technology advanced from manual to mechanical and then from mechanical to automatic. Vision based applications are the products of the future. Today's high speed, complex manufacturing systems require the development of automation technologies that can be efficiently integrated into the systems and used in manufacture floors. Machine vision systems integrate electronic components with software systems to imitate a variety of human functions. This presentation gives overview on the recent developments in the area of Machine Vision with few case studies. The presentation also highlights the challenges and opportunities in the area of Machine Vision. The research and development work at Samsung Electro-Mechanics will be highlighted towards, the end of presentation.
Speaker bio: Dinesh is a Chief Engineer with Samsung Electro-Mechanics, India. Prior to joining Samsung, he worked for companies Amazon, Honeywell, HCL Technologies. He was a post doctoral researcher at Seoul National University, Seoul. His areas of research interest includes Image processing, Pattern Recognition, Computer Vision and Document Image Analysis. He has over 60 publications to his credit at both International and National Journals and conferences. He also has 12 awarded patents to his credit. He is serving in the editorial board of several journals. He has been identified as reviewer for several International Journals/conferences. He has served as program committee member for several National/International Conferences. He is presently guiding 5 Ph.D scholars. He has successfully guided many academic and research projects. He is a life member of IEEE, Computer Society of India and Society of Statistics, Compute and Applications, International Association of Engineers, Hong Kong.

Go top

Title: CP-mtML: Coupled Projection multi-task Metric Learning for Large Scale Face Retrieval
Speaker: Dr. Gaurav Sharma
Venue: EE 217 (MMCR)
Date and time: 13th Apr 2016, 11:00am
Abstract: I will present our recent work about a novel Coupled Projection multi-task Metric Learning (CP-mtML) method for large scale face retrieval. In contrast to previous works which were limited to low dimensional features and small datasets, the proposed method scales to large datasets with high dimensional face descriptors. It utilises pairwise (dis-)similarity constraints as supervision and hence does not require exhaustive class annotation for every training image. While, traditionally, multi-task learning methods have been validated on same dataset but different tasks, the method is tested on the more challenging setting with heterogeneous datasets and different tasks. Empirical validation on multiple face image datasets of different facial traits, e.g. identity, age and expression support the method. The experiments clearly demonstrate the scalability and improved performance of the proposed method on the tasks of identity and age based face image retrieval compared to competitive existing methods, on the standard datasets and with the presence of a million distractor face images.
Speaker bio: Gaurav Sharma is currently an Assistant Professor at the Department of Computer Science and Engineering of the Indian Institute of Technology Kanpur. He obtained his PhD in 2012 under the joint supervision of Cordelia Schmid (LEAR, INRIA Grenoble) and Frederic Jurie (GREYC, CNRS UMR 6072, University of Caen) and an Integrated Master of Technology in Mathematics and Computing in 2008 from the Indian Institute of Technology Delhi (IIT Delhi), India. He was, prior to his PhD, a Senior Engineer at the Technology Planning Group of Samsung R&D India and, after his PhD, a Researcher in the Exploratory Research Group of Technicolor R&I, France. and a Postdoctoral Fellow at the Max Planck Institute for Informatics, Germany. His research interests are Computer Vision and Machine Learning.

Go top

Title: Overview of deep learning in Echo’s speech recognizer
Speaker: Dr Sivaram Garimella
Venue: EE 217 (MMCR)
Date and time : 11th Apr 2016, 5.00pm
Abstract: Amazon Echo's high accuracy speech recognition system uses deep neural network (DNN) acoustic models. Such models are trained using in-house distributed Stochastic Gradient Descent (SGD) training framework. This talk provides an overview of speech recognition focusing on DNN based acoustic modeling and our distributed DNN training for building production models.
Speaker bio: Sri Garimella has been working as a machine learning scientist in Amazon Echo speech recognition team since September 2012. He has relocated to India and currently based in Amazon, Bangalore. He has obtained PhD from the Johns Hopkins University, Center for Language and Speech Processing, Baltimore in 2012, and Master of Engineering in Signal Processing from the Indian Institute of Science, Bangalore in 2006. His research interests include deep learning, statistical machine learning, and speech & speaker recognition. He has several publications in peer-reviewed journals and conferences in this area.

Go top

Title: Computational Pathology: Unlocking Tissue Content in Precision Medicine
Speaker: Dr. Chukka Srinivas
Venue: EE 217 (MMCR)
Date and time: 5th Apr 2016,11:30am - 1.00pm
Abstract: Traditionally, pathologists detect, stage and grade cancer in patient tissue biopsies based on a microscopic manual evaluation of small set of cellular and morphological features in Hematoxylin and Eosin (H&E) stained tissue slides. This information along with an assessment of protein overexpression in the tumor, semi-quantitatively scored from multiple immunohistochemical (IHC) stained slides, is used to subtype the cancer for prognostic and precise treatment selection and planning.
Digital pathology refers to the technologies for digitization of tissue whole slides and image analysis algorithms for automated slide interpretation and precise quantification. So far, the image analysis algorithm development has been focused on generating consistent interpretation and reproducible slide scores, accounting for the inherent challenges of wide biological and staining variability in a clinical setting, but primarily limited to mimicking the manual interpretation process of the pathologists.
There is ever increasing medical evidence that while there is potentially a large amount of prognostic information for a given patient in the tissue, today this information is being analyzed separately from a clinical standpoint without its holistic integration into a single comparative prognostic dataset. Computational pathology is a data-driven pipeline, based on statistical and machine learning methods, for systematic extraction of multi-dimensional information in the digitized tissue slides at multiple scales, statistically combine these features and directly correlating against patient outcome, which is given in terms of patient survival or a response to a drug treatment, to discover discriminant features for prognostication and prediction to a drug response.
In this talk, we propose a computational pathology framework and showcase an end to end application to a specific example of prediction of risk of recurrence in early stage breast cancer patients. Supervised learning based fully automated image analysis algorithms are used to analyze H&E and multiple IHC whole slides and extract an exhaustive set of image features. Based on extracted image features and clinical outcome for a patient cohort, a L1-regularized logistic regression based prognostic model is constructed. We show here there are morphological, relational and co-expression image features, which are significantly associated with patient overall survival and could therefore be used to improve prognosis and guide follow-up treatment.
Speaker bio: Chukka Srinivas is a Scientific Fellow in Roche Tissue Diagnostics working on image analysis and machine learning solutions for cancer tissue diagnostics. He got his Bachelors in Electronics from Osmania University in ’83, Masters from the Indian Institute of Science in ‘85 and PhD from the Southern Methodist University, Dallas in 1990. He has worked before at GE Global Research, HP, Teradyne, Polaroid and Hologic. As an imaging researcher, has worked on variety of topics in signal and image processing, computer vision and machine learning topics for medical, semiconductor, aerial and photographic applications and contributed to the development of multiple products.

Go top

Title : Studies on polymer insulators subjected to electrical and environmental stresses
Speaker : Shakthi Prasad D
Advisor : Dr. Subba Reddy B. and Prof. Rajanikanth B. S.
Date : 04/03/2016 (4.00 pm - 5.00 pm)
Venue : MMCR (C241)
Abstract: Corona is an unavoidable phenomenon in the high voltage power transmission system. It is a proven fact that the polymeric insulators subjected to continuous corona lead to severe degradation like surface rapture, hydroxylation, oxidation etc., Further; moisture has a positive influence on the corona activity. In the present study, a methodology has been developed to evaluate the corona performance of the silicone rubber housing material with simultaneous application of cold fog. Physico-chemical analyses like Fourier transform spectroscopy, scanning electron microscopy, contact angle measurements etc., were carried out on the treated samples. Experimental investigations conducted prove that moisture assists in accelerating the corona induced degradation on the polymeric insulators.

Go top

Title: High Power Factor High-Current Variable-Voltage Rectifiers
Speaker: Dr.Jitendra Solanki
Venue: EE 217 (MMCR)
Date and time:26th Feb 2016,15:30 hrs
Abstract: High-current variable-voltage (HCVV) rectifiers are used in the metal and chemical industries. Typical power ratings vary from tens of kW to hundreds of MW. Even with the advancement of the power factor correction rectifiers, accepted choices of high-current AC to DC converters remain 6, 12 or 24-pulse thyristor/diode rectifiers, because of high reliability, efficiency and availability of suitable semiconductor devices. The main issues with these rectifiers are poor input power factor, high current harmonic distortion, high-maintenance cost, high weight and large volume. The seminar highlights some of these issues and explains the remedial measures taken. Apart from this, the talk also (briefly) covers some of the other projects in fields of high-power high-frequency resonant inverter and switch-mode power-supplies.
Speaker bio: Jitendra Solanki received his B.-Tech. degree in electrical engineering from G.B. Pant University of Agriculture and Technology, Pantnagar and M.-Tech. degree in power electronics electrical machines and drives from IIT Delhi, New Delhi in 2004 and 2006, respectively. He received doctorate degree from University of Paderborn, Germany in 2015. Before joining Ph.D., from 2006-2009, he was working with GE Global Research, Bangalore. From 2009-2010, he was with Philips Research Asia, Bangalore. Since Nov. 2014, he is working with Corporate Research, Robert Bosch (SEA) Pte. Ltd., Singapore. Dr. Solanki is a recipient of the ‘innovative student project award' from the Indian National Academy of Engineering and the 'ISTE-L&T second best project award' from Indian Society of Technical Education for his work during master’s degree. His research interests include high-power rectifiers, power-quality compensators and DC-DC converters.

Go top

Title: Exploiting latent reliability information for classification tasks
Speaker: Mr. Naveen Kumar
Venue: EE 217 (MMCR)
Date and time: 4th Feb 2016,11:30 hrs
Abstract: Despite significant advances in machine learning techniques, we find that certain pattern recognition problems are intrinsically more difficult. For example, in a classification task certain classes may be more ``noise-like" or difficult to model compared to others. The challenge in these tasks often lies in adequately modeling the variability in observations within a given data set. In theory, we would be interested to isolate and learn from only those aspects of this variability that are useful for the classification task. This information could be either in the form of knowledge of certain useful features or training samples that are more informative than others. In this work, I shall use the attribute “reliable” to refer to such aspects that are informative in the context of a pattern recognition task. In practice, this reliability information is usually latent and must be jointly estimated during learning. I shall propose techniques to account for and exploit this inherent heterogeneity in reliability associated with samples during training of classification models. Reliability is modeled as a latent factor that governs the dependence between observed features and the corresponding annotated class label. In addition to augmenting classification models, the reliability scores also lend themselves to interpretation in other contexts, such as for multiple annotators in a crowd-sourcing task.
Speaker bio: Naveen Kumar received his B.Tech. in Instrumentation Engineering from the Indian Institute of Technology, Kharagpur. He is currently an Electrical Engineering Ph.D. candidate at the University of Southern California, Los Angeles. His research focuses on models for exploiting latent reliability information associated with features or labels in pattern recognition tasks. His broad interests lie in statistical signal processing, speech and audio processing, machine learning and robust multimodal recognition. He is a recipient of the USC Dean's Fellowship and was part of the USC team that won the Interspeech-2012 Computational Paralinguistics Challenge.

Go top

Title: On Risk-Sensitive Reinforcement Learning: Algorithms, Analysis and Applications
Speaker: Dr. L.A. Prashanth
Venue: EE 217 (MMCR)
Date and time: 22nd January 2016, 4:00-5:00 pm
Abstract:
In many sequential decision-making problems, one may want to manage risk by minimizing some measure of variability in rewards in addition to maximizing a standard criterion. Variance related risk measures are among the most common risk-sensitive criteria in finance and operations research. While the theory of risk-sensitive Markov decision processes (MDPs) is relatively well-understood and that we know many of such problems are computationally intractable, not much work has been done to solverisk-sensitive MDPs in a typical reinforcement learning (RL) setting.
In this talk. I will describe a few important steps that I took to approximately solve risk-sensitive MDPs - both discounted and average reward. For each formulation, I will first define a measure of variability for a policy, which in turn gives us a set of risk-sensitive criteria to optimize. For each of these criteria, I derive a formula for computing its gradient and then devise actor-critic algorithms that operate on three timescales - a temporal difference (TD) critic on the fastest timescale, a policy gradient (actor) on the intermediate timescale, and a dual ascent for Lagrange multipliers on the slowest timescale. In the discounted setting, I will point out the difficulty in estimating the gradient of the variance of the return and then present a simultaneous perturbation approach to alleviate this problem. The average setting, on the other hand, allows for an actor update using compatible features to estimate the gradient of the variance.
The analysis of the aforementioned risk-sensitive RL algorithms involves statistical aspects of the popular TD algorithm with function approximation and I will present concentration bounds that I derived in a recent work for the latter algorithm. These bounds help in establishing the convergence of risk-sensitive RL algorithms using the ordinary differential equations (ODE) method to locally risk-sensitive optimal policies.
Finally, I will demonstrate the usefulness of the risk-sensitive RL algorithms in a traffic signal control application. In particular, the empirical results show that risk-sensitive RL algorithms exhibit lower variance in the delay experienced by road users, as compared to corresponding risk-neutral RL variants.
Speaker bio: Prashanth L.A. is currently a postdoctoral researcher at the Institute for Systems Research, University of Maryland - College Park. Prior to this, he was a postdoctoral researcher at INRIA Lille - Team SequeL from 2012 to 2014. From 2002 to 2009, he was with Texas Instruments (India) Pvt Ltd, Bangalore, India. He received his Masters and Ph.D degrees in Computer Science and Automation from Indian Institute of Science, in 2008 and 2013, respectively. He was awarded the third prize for his Ph.D. dissertation, by the IEEE Intelligent Transportation Systems Society (ITSS). He is the coauthor of a book entitled `Stochastic Recursive Algorithms for Optimization: Simultaneous Perturbation Methods', published by Springer in 2013.His research interests are in reinforcement learning, stochastic optimization and multi-armed bandits, with applications in transportation systems, wireless networks and recommendation systems.

Go top

Title: Signal Processing and Machine Learning for Carnatic Music
Speaker: Prof. Hema Murthy
Venue: EE 217 (MMCR)
Date and time: 21st January 2016, 3:30 pm-5:00 pm
Abstract:Carnatic Music is based on a particular genre of Indian Music belonging to the South India. Similar to HindustaniMusic, Carnatic Music is also based on the oral tradition and, therefore hardly any music annotated accurately. The purpose of the research is to develop signal processing and machine learning algorithms that can be used to characterise and organise Carnatic Music to enhance listener experience. In Carnatic Music, a composition is based on a “r¯a ga” or melody. Each melody is based on a scale. But a scale is inadequate to represent the melody completely. Owing to the extensive use of gamak¯as two melodies may be identical in terms of scale but completely different in terms of gamak¯as. R¯a gas are characterised by signature phrases. The phrase is essentially a melodic motif. Therefore the first task was to characterise and identify the melodic motifsthat can characterise “r¯a gas”. This is not trivial since a “r¯a ga” is defined with respect to the tonic (shadja) of the performer (vocal or instrumental). Any meaningful analysis would first require that we first normalise the music with respect to the tonic. A number of different signal processing and dictionary based algorithms are developed for tonic identification from the compositions. Success rates between 90-100% have been obtained for both Carnatic music and Hindustani music. To understand what a motif is, a motif database was first created by musicians. Since motifs can be characterised by pitch contours, the basic idea in motif recognition was to arrive at a set of basic primitives, in terms of which signature phrases of a r¯a ga can be defined. Owing to the differences in schools of music, it was observed that a motif is a T-F trajectory and should not be quantised further. Using a musician in the loop, a variant of the rough longest common subsequence algorithm is developed to query motifs defined by musicians. Next automatic identification of motifs is attempted. In Carnatic music compositions are rich in motifs. Using tonic normalised spectral features, a composition is first segmented in lines. Longest common subsequence set is used to determine the common motifs across compositions. A cohort set of phrases for every r¯aga is also determined. R¯aga verification is performed using the motifs of a r¯aga and its cohorts. In Carnatic Music, percussion has a very important role. There are 108 talas and each tala can be characterised by different nadais. The number of strokes used by a musician vary from a mere 10 to about 41. Every stroke is made up of an onset, attack and decay. Considering the mridagam stroke as an AM-FM signal, the waveform is segmented at the stroke level using syllable-based segmentation algorithms. As there can be variants in the strokes HMM-based models are built for every stroke and recognised. Language models derived from the training data are used to reduce the error in transcription. Finally, Carnatic music concert recordings are continuous recordings. Using applauses as a cue, the concert recordings are segmented into items. Using informal reviews available from the web, the items that make up a concert are archived.
Speaker bio: Prof. Hema Murthy is a Professor at the Department of Computer Science and Engineering, Indian Institute of Technology (IIT), Madras. She obtained a Masters from McMaster, Canada in 1985 and PhD from IIT Madras in 1992. She held visiting positions at SRI International, Menlo Park, TIFR Bombay. Her research interests include Speech and music processing, speaker and language recognition, network traffic analysis, machine learning, data mining, etc. She has supervised 68 Masters theses and 42 research students. She is well known for her contributions to group-delay analysis in speech. She received the IBM Faculty award in 2006, Prof. Rais Ahmed Memorial Lecture Award from the Acoustical Society of India in 2012, and GE innovation award in 2012.

Go top

Title: Developments in Power Sector – An overview of System Analysis, Planning and Operation
Speaker: Prof. Thukaram
Venue: EE 217 (MMCR)
Date and time: 1st January 2016, 4:00-5:00 pm
Abstract: Electric Energy is the most versatile form of energy. It can be generated in large quantities, transmitted over long distances and utilized efficiently. It has flexibility and most amenable to effective control. The per capita consumption of energy has become the symbol of a country’s progress. The first DC generating station at Pearl Street, New York city was commissioned by Thomas Alva Edison in 1882. In India Hydro power stations were commissioned at Darjeeling, West Bengal in 1898, and at Shivasamudram, Karnataka in 1902. Since that time the developments in power generation, transmission and distribution have been phenomenal. Present day Power Systems are growing in size and complexity with interconnections of regional grids, introduction of EHV/UHV AC, High Voltage Direct Current (HVDC) transmission systems and also the induction of sophisticated control devices such as Flexible AC Transmission Systems (FACTS). Worldwide there exists a serious concern with dwindling fuel reserves and the potential impact of conventional energy systems on the environment. The increasing rate of depletion of conventional energy sources, escalation in electric energy costs associated with fossil and nuclear fuels, and enhanced public awareness of potential environmental impacts of conventional energy systems has created an increased interest in the development and utilization of alternate Renewable sources, such as wind and solar energy Recent trends in the Electric Power utility industry in developed and developing countries have been towards increased un-bundling of the services provided by the utilities. Indian Power Systems are also under going the restructuring process. The planning, operation and control of these systems pose complex technical challenges. Researchers are working to address these challenges. The lecture covers some of the milestones in the development of the Indian Power Sector and also provides a brief overview of planning and operational studies intended to prevent catastrophic accidents and major blackouts.

Go top

Title: Power System Operation and some Stability issues
Speaker: Prof. Indraneel Sen
Venue: EE 217 (MMCR)
Date and time: 18th December 2015, 3:30-4:30 pm
Abstract: Modern power systems are interconnected, non linear, time varying systems. Such a system which is continuously subjected to large and small impacts must provide acceptable quality of power at all times. The system must be able to meet the continually changing load demands for active and reactive powers to maintain the voltage and frequency within reasonable limits. The system must operate with a high degree of reliability without major interruptions. This requires three distinct areas of operation and planning. 1. Power system planning 2. Operation planning and 3. Real time operation In spite of considerable efforts in planning and operation Blackouts do happen. The proposed talk will address some of these issues and include, * A historical perspective of worldwide development of power system including in India * The Indian power system operational and control structure * The reasons blackouts occur * Analyses of major blackouts in India * Power system stability and its categorizations * The concept of Damping and Synchronizing restoring torques * The missing link of ‘damping torque’ during planning * The concept of Power System Stabilizers * Some of the design methodologies of Power System Stabilizers developed in our lab. The Future?

Go top

Title : Hand segmentation using 3D pointclouds for Hand gesture recognition
Speaker : Mr. Shome Subhra Das
Advisor : Prof. K. R. Ramakrishnan
Date : 11/12/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Hand gesture recognition is an important part of Human Computer Interaction(HCI)systems. Hand gesture estimation is a challenging problem due to complexity in articulations of the hand, self occlusions and variation in hand shapes. Due to availability of cheap depth sensors research on hand gesture recognition has received a great boost. Reliable identification of the hand region is an important step for hand gesture recognition. We propose a method for hand region segmentation with accurate wrist detection from depth data. This method converts depth data to 3D pointcloud to use techniques in 3D. Area estimation, 3D edge recognition & anthropometric calculations are used to accurately determine the location of wrist thereby enabling correct segmentation of hand region. The method is not restricted to frontal views & works even for deformed hand shapes & with small objects in hand. Preliminary experiments with real hand depth data in various poses, orientations & with objects demonstrate the accuracy and robustness of our hand segmentation system.

Go top

Title : Fixed Frequency Static Phase Converter For Single-Phase Power Grids
Speaker : Mr.Anil Kumar Adapa
Advisor : Prof. Vinod John
Date : 4/12/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract:Single phase grids are commonly used to power the communities with low power demand and sparsely located load centers due to low cost feature. Single phase microgrid is a first choice to provide electricity to the aforementioned places with no access to the electric grid. In either case, usage of three-phase motors as well as generators is limited, which calls for double conversion using power electronic converters.
An attractive feature of grid interactive static phase converters is the ability to deliver a fraction of total load power directly from the source. This talk presents a reduced switch, fixed frequency static phase converter that not only reduces the size and cost of the system but also results in better efficiency. This presentation covers (1) control strategy for the proposed static phase converter with active front end converter that leads to improved dc bus utilization and (2) a simple inverter control to provide balanced three-phase voltage. The presentation ends with simulation results of powering three-phase loads from a single-phase grid as well injecting power from a three-phase machine to single-phase grid using the static phase converter which endorses the controller design.

Go top

Title : Better Electromagnetic Shielding using Polymer Composites
Speaker : Joseph Vimal Vas
Advisor : Dr. M. Joy Thomas
Date : 20/11/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Electromagnetic waves are used routinely for a variety of applications to improve the quality of life. This has led to an increased electromagnetic pollution in the environment. These EM waves can adversely interact with many sensitive electronic devices leading to temporary or permanent damage of these devices or even loss of life. The interference thus caused by electromagnetic fields is extremely difficult to identify and control especially in the automobile sector where the use of on board electronics is continuously on the rise. Other industries are also heavily relying on sensitive electronic circuitry for various applications. Reducing the ambient fields by smart design of electronics is no longer a viable option and electromagnetic shielding has become very important in the design and fabrication of sensitive electronic systems. Traditionally used metallic shields may be less attractive because of their large metal content and lack of flexibility. The advent of conducting polymers has brought an interesting alternative for electromagnetic shielding materials. Intrinsically Conducting Polymers and conducting polymer composites have very low metal content and being polymeric they can be easily processed. Another problem is the estimation of the shielding effectiveness in anisotropic materials. Theoretical estimation of the shielding effectiveness becomes difficult especially when the characteristics vary with the frequency of interest. ASTM D 4935 standard provides a method for measuring the shielding effectiveness of thin samples in the 30 MHz – 1.5 GHz frequency range. In my work, the conductivity and shielding effectiveness of silicone rubber filled with carbon nanofibers (CNF) and Multiwalled carbon nanotubes (MWCNT) are studied. A test jig as per ASTM D 4935 has been developed for the measurement of shielding effectiveness. The filler loading required to turn the polymer conducting, depends on the percolation threshold. For fillers with large aspect ratio, the percolation threshold is very low which was seen for both the fillers. The conductivity beyond the percolation threshold is majorly governed by the electron tunnelling between the filler particles and this puts a restriction on the conductivity. In order to overcome this problem, the fillers were functionalised with nano- Ag fibers to create wafers of extremely high conductivity. The polymer was then adsorbed into the wafer matrix to achieve the required mechanical properties. The talk will also present some of the theoretical studies carried out to understand the influence of the type of fillers and their orientation inside the polymer matrix on the shielding effectiveness.

Go top

Title : Processing RGB-D Data for Entertainment Systems
Speaker : Suraj K
Advisor : Prof. K. R. Ramakrishnan
Date : 13/11/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Thanks to the advances in research in display systems, various new forms of entertainment systems are now being created. Unlike conventional devices such as televisions that provide minimal realistic experience to the viewer along with no scope for any engagement, these systems create an "immersive" experience, a more realistic experience. Further, these systems also allow the user to be more engaging, the user "walk through" the scene. However, creating content and processing data to facilitate the usage of such systems is still an open problem. Video streams are captured from standard cameras that are coupled with depth sensors, which is also called RGB-D Video. In this talk, we discuss about two such entertainment systems, the Virtual Reality and Free Viewpoint Television. We discuss about the challenges posed when RGB-D video are used to create contents for these systems. Finally, we discuss about an application of RGB-D data i.e., synthesizing a virtual video using RGB-D video streams captured from multiple cameras.

Go top

Title: Automatic Video Mashup Techniques for Multimedia Systems
Speaker: Dr. Mukesh Saini
Venue: EE 217 (MMCR)
Date and time: 9th November 2015, 3:30-4:30 pm
Abstract: In a video mashup system, multiple videos are combined to produce a single representative video. The output video can be built based on different criteria depending on the application requirement. I will present three mashup techniques in this talk. In the first technique, we combine different mobile recorded videos of live performances to produce MTV style videos for broadcasting and sharing. The shot transitions are learned from professionally edited videos. In the second approach, we create video mashups for security operators for dynamic load sharing in a multi-camera and multi-operator scenario. Finally, I will present a technique to create low resolution mashup from single high definition video for surveillance through handheld devices.
Speaker bio: Dr. Mukesh Saini is Postdoctoral Associate at New York University in Abu Dhab.

Go top

Title: SMART GRIDS: An Overview
Speaker: Prof. P. S. Nagendra Rao
Venue: EE 217 (MMCR)
Date and time: 6th November 2015, 4:00-5:00 pm
Abstract: The concept of Smart grids has been attracting the interest of not only many stake holders in the Electrical Power Industry but also of the general public , in the recent past. The aim of this talk is to give a broad view of the various issues in the development of such systems. The core idea of smart grids is about exploiting the phenomenal progress the in communication and computer technologies, to modernise all aspects of the power systems. In this talk we first look at the evolution of the concept of Smart Grids in the context of the history of application of computer and communication technologies in Power Systems and understand the contemporary vision for Smart Grids. Various technologies- hardware, software and materials that are expected to help in building the smart grids of the future are introduced and their novelty as well as potential are assessed. We look at the efforts in India in adopting the Smart Grids technologies and also some experiences gained from a pilot project-the Pondicherry Smart Grid Project. Some key smart grid technologies such as Demand Response/Consumer participation, Self healing technologies, Cyber security, micro grids etc are assessed in the Indian context. Finally, I share my ‘crystal ball’ view of the Power Systems of the future.

Go top

Title:Transmission Power Control Using CTMC Based Interference
Speaker: K. Srilatha
Advisor: Prof. Lawrence Jenkins and Prof. P.S. Sastry
Venue: EE 217 (MMCR)
Date and time: 30th October 2015, 4:00-5:00 pm
Abstract:Transmission power control (TPC) considers the fixing of the transmission power level. TPC is an important problem in energy constrained networks like wireless sensor networks. The existing methods of power control vary the power level for every node or use the same power level (uniform) across the nodes. We consider uniform power control with carrier sense multiple access (CSMA) protocol. Interference signals from nodes add to the desired signal at the receiver and affect the performance indices such as bit error rate (BER). Hence, characterization of interference is crucial. The conventional models of interference either consider transmissions as Poisson distributed or approximate interference with Gaussian noise. In our work, we model the interference from nodes using continuous time Markov chain. The stationary probabilities of the Markov chain gives the probability of interference from the interfering transmitters. The model is verified using simulations.

Go top

Title: Study of Propagation of Ultra-Wide Band Electromagnetic Pulse through a Dispersive Soil Medium for the Detection of Buried Unexploded Ordnance
Speaker: Vijayakumar Solaiselvam
Advisor: Dr. Joy Thomas
Venue: EE 217 (MMCR)
Date and time: 22nd October 2015, 4:00-5:00 pm
Abstract: Buried Unexploded ordnance (UXO) poses a serious humanitarian problem. At Present, there is no single equipment/system available to detect the unexploded buried ordnance (UXO), which fully satisfies the United Nations Mine Action Standard (UNMAS) requirements for safe removal and disposal of such ordnance. Due to the attenuation characteristics of electromagnetic wave in moist soil, it is difficult to detect UXO using low power electromagnetic waves. So there is a need for using pulsed Ultra Wide band (UWB) high power electromagnetic waves to detect the buried UXO. Objective of this thesis is to carry out detailed parametric study of the high Power EM wave landmine detection system.

Go top

Title: A class-specific speech enhancement for phoneme recognition
Speaker: Nazreen P.M
Advisor: Prof. A. G. Ramakrishnan
Venue: EE 217 (MMCR)
Date and time: 16th October 2015, 4:00-5:00 pm
Abstract: The performance of speech recognizer degrades significantly in the presence of noise. Noise distorts the spectrum of speech and hence degrades the performance. In this work, we analyze the impact of dictionary based speech enhancement on phoneme recognition. The improvement in recognition performance after enhancement with class-specific dictionaries is examined over that with a class-independent dictionary. All experiments in this study are performed in a speaker independent manner with speech data from TIMIT corpus and noise samples from NOISEX-92 database. Using KSVD, the following four types of dictionaries have been learned: class-independent, manner-of-articulation-class, place-of articulation-class and 39 phoneme-class. Initially, a set of labels are obtained by recognizing the speech enhanced using the class-independent dictionary. Using these approximate labels, the corresponding class-specific dictionaries are used to enhance each frame of the original noisy data, and this enhanced speech is recognized further. The results show that in most cases, the 39 phoneme-class based enhancement outperforms the mannerof-articulation and place-of-articulation class based enhancement, in terms of the recognition accuracy.

Go top

Title: Electrically Triggered Thyristor based Solid State Crowbar
Speaker: Subhash Joshi T G
Advisor: Prof. Vinod John
Venue: EE 217 (MMCR)
Date and time: 9th October 2015, 4.00 PM
Abstract: High Voltage Power Supplies (HVPS) are used in many application like (i) Radar (ii) X-rays (iii) Plasma applications (iv) Electrostatic precipitator (v) Corona generators (vii) Particle accelerators etc. In most of these applications the HVPS feeds power to Micro Wave (MW) tubes. The cost of MW tubes are multiples of the cost of the HVPS. Hence MW tubes are to be protected from various fault conditions, like over voltage, short circuit or any other event which trigger protection circuits.
Crowbar are mainly used as a protection for MW tubes. Conventionally it is build with ignitron or thyratron or spark gap. Ignitron and thyratron are restricted now a days due to use of mercury gas where spark gap demand frequent maintenance even after each crowbar operation. The research is now to replace these devices with semiconductor switch called solid state crowbar (SSC).
The presentation is on the development of Solid State Crowbar having high reliability, fast operation, wide operational voltage range, high peak pulse power rating and low cost. The presentation details various research outcomes made during the development of Solid State Crowbar.

Go top

Title: Developments in Power Sector – An overview of System Analysis, Planning and Operation
Speaker:Prof. D. Thukaram
Venue: EE 217 (MMCR)
Date and time: 18th September (Friday) 2015, 4.00 PM
Abstract: Electric Energy is the most versatile form of energy. It can be generated in large quantities, transmitted over long distances and utilized efficiently. It has flexibility and most amenable to effective control. The per capita consumption of energy has become the symbol of a country’s progress. The first DC generating station at Pearl Street, New York city was commissioned by Thomas Alva Edison in 1882. In India Hydro power stations were commissioned at Darjeeling, West Bengal in 1898, and at Shivasamudram, Karnataka in 1902. Since that time the developments in power generation, transmission and distribution have been phenomenal. Present day Power Systems are growing in size and complexity with interconnections of regional grids, introduction of EHV/UHV AC, High Voltage Direct Current (HVDC) transmission systems and also the induction of sophisticated control devices such as Flexible AC Transmission Systems (FACTS). Worldwide there exists a serious concern with dwindling fuel reserves and the potential impact of conventional energy systems on the environment. The increasing rate of depletion of conventional energy sources, escalation in electric energy costs associated with fossil and nuclear fuels, and enhanced public awareness of potential environmental impacts of conventional energy systems has created an increased interest in the development and utilization of alternate Renewable sources, such as wind and solar energy Recent trends in the Electric Power utility industry in developed and developing countries have been towards increased un-bundling of the services provided by the utilities. Indian Power Systems are also under going the restructuring process. The planning, operation and control of these systems pose complex technical challenges. Researchers are working to address these challenges. The lecture covers some of the milestones in the development of the Indian Power Sector and also provides a brief overview of planning and operational studies intended to prevent catastrophic accidents and major blackouts.

Go top

Title : Dual Comparison One Cycle Control for single phase AC-DC power converters
Speaker : Nimesh V
Advisor : Prof. Vinod John
Date : 11/09/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Classical control and one cycle control of current are popular methods used to generate PWM pulses in active rectifiers for ac-dc power conversion. One cycle control has lower control complexity and can be implemented using simple analog circuits and digital circuits, when compared with the classical approach. However, it also suffers from distortion in input current when the converter is lightly loaded and steady state dc offset in input current under all load conditions. In this paper a new dual comparison one cycle control is proposed which overcomes the above limitations. The proposed control strategy makes use of two comparators which compares sensed input current and inverted sensed input current with a saw-tooth carrier to generate gating signals for switching devices.

Go top

Title : Algorithms for Learning Sparsifying Transform
Speaker : Mr. Subhadip Mukherjee
Advisor : Prof. Chandra Sekhar Seelamantula
Date : 4/09/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: The sparsity of signals and images in a suitable transform domain or in a dictionary has been exploited extensively in numerous signal processing applications. Recently, learning sparsifying synthesis dictionary has received considerable attention in applications, such as denoising, image inpainting, super-resolution etc. However, the idea of learning a sparsifying transform has been relatively less explored. In this talk, we will address the problem of learning a well-conditioned sparsifying transform adapted to the data and develop efficient algorithms for solving the problem. We will illustrate the computational advantage of learning a transform over learning a synthesis dictionary and show that the approach holds promise for efficient signal sparsification and denoising, using experimental results on synthesized and real signals.

Go top

Title : Control of Ultracapacitor based Bidirectional DC-DC Converters for Ride Through Applications
Speaker : K. Saichand
Advisor : Prof. Vinod John
Date : 21/08/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Ultracapacitor based bidirectional DC/DC converters find applications in hybrid electric vehicles, traction and transport systems, power-quality, and in micro-grids for energy backup provision. Ultracapacitor based DC/DC converters essentially handle critical loads in case of momentary power main failures and for addressing peak power demands. The conventional control techniques which ensure seam-less mode transition for the Ultracapacitor (UC) based bidirectional DC/DC converters such as duty-ratio based control or unified control strategy are well established for current control based applications. But these control techniques does not allow different control structures for both charging and discharging modes which is crucial for the UC based ride-through applications where the charging and discharging times are critical. To address these issues, a separate switch control strategy using PWM block method is proposed which not only ensures seam-less transition between charging and discharging control modes, but also identifies control modes accurately which is key to the separate switch control. The advantages of such a switch control using PWM block method compared against conventional controls is elucidated. The advantage of using PWM block method as a mode change logic is also presented. The proposed control strategy has been verified in simulations and experimentation and the mode identification algorithm proposed is found to work well.

Go top

Title: Self-organizing Power-electronic Systems
Speaker: Prof. Sairaj Dhople, University of Minnesota, Minneapolis
Venue: EE 217 (MMCR)
Date and time: 18/08/2015, 4.00 PM
Abstract: Next-generation power systems are expected to be sustainable in composition, distributed in operation, and resilient to extenuating weather conditions. A compelling framework to seek these goals is provided by low-inertia microgrids. These are a heterogenous collection of renewable-energy resources and energy-storage devices that are interfaced to an AC electrical distribution system through power-electronic inverters. In this talk, we focus on islanded microgrids that are controlled and operated independently from the bulk power system. We introduce a control method called Virtual Oscillator Control for synchronizing and regulating a collection of islanded power-electronic inverters without communication. The premise of virtual oscillator control is to program power-electronic inverters to emulate the dynamics of Lienard-type nonlinear oscillators. A system with virtual oscillator control is self-organizing in that the inverters synchronize their AC outputs, share the load, and collectively maintain voltage and frequency within regulatory limits without any supervisory control. A stable power system emerges innately by design, and the only form of communication is that provided by the physical electrical network that couples the inverters (oscillators). The proposed technique is developed using concepts from nonlinear control theory and experimental results are presented to validate the concept. The system-theoretic methods that will be outlined in this talk are relevant to the broad domain of synchronization phenomena in complex networks of coupled nonlinear oscillator circuits; a pervasive research topic in various scientific disciplines including neuroscience, physics, systems biology, social networks, and engineering.
Speaker bio: Sairaj Dhople received the B.S., M.S., and Ph.D. degrees in electrical engineering, in 2007, 2009, and 2012, respectively, from the University of Illinois, Urbana-Champaign. He is currently an Assistant Professor in the Department of Electrical and Computer Engineering at the University of Minnesota in Minneapolis, where he is affiliated with the Power and Energy Systems research group. Dr. Dhople received the NSF CAREER Award in 2015. He currently serves as an Associate Editor for the IEEE TRANSACTIONS ON ENERGY CONVERSION. His research interests include modeling, analysis, and control of power electronics and power systems with a focus on renewable integration.

Go top

Title: Indian Buffet Process for vision applications
Speaker: Suresh Kirthi K
Advisor: Prof. K R Ramakrishnan
Venue: MMCR, EE
Date & Time: 14/08/2015, 4 pm
Abstract: Indian Buffet Process(IBP) is a Bayesian Non-parametric prior used for latent feature analysis. It tries to model a case where an observation is modeled to be composed of unlimited latent features. The most popular and reasonably computable model of IBP is Linear Gaussian model, where the latent features are estimated using random sampling and variational inference under Gaussian assumption.
As an application to computer vision we have tried to explore the possibility of using IBP for some of the vision problems like: scene recognition, object recognition and canonical views.

Go top

Title: High Frequency response of windings.
Speaker: Santosh J
Advisor: Prof. Udaya Kumar
Venue: MMCR, EE
Date & Time: 07/08/2015, 4 pm
Abstract: Power transformer forms one of most critical and expensive equipment in the electric power network. About 29% of its failure is traced on to the electrical disturbances and the consequential insulation failures in the winding and bushings. Classically, the electric stress distribution on the winding under normal operating voltages, as well as, the switching and lightning surges is analyzed through distributed circuit models.
The transformer in a gas insulated substation is subjected to Very Fast Transient Overvoltages (VFTO). These surges can have a rise time less than 20 ns and significant frequency ranging up to 100 MHz. The associated wavelengths are much smaller than the actual length of the winding thereby the use of distributed circuit model or its variant is put into a serious question.
Similarly, the Partial Discharges (PD) occurring in the insulating system involves a local current pulse of rise time less than few nanoseconds. The propagation of such current pulses in the winding would invoke a similar question.
Apart from the above two, even the penetration of very fast rising lightning surge through an ultra high voltage class power transformer is found to be inadequately simulated by the routinely employed distributed circuit model.
Transformers are subjected to various stresses, our concern being electrical stresses due to very fast transient overvoltages (VFTOs) in transformer connected to GIS and Partial discharge (PD) propagation in windings.
In order to primarily address the above issue, the present work is taken up. The goal of the present work is to investigate on to the propagation characteristics of VFTO and possibly the PD pulses through the representative winding configuration and quantify the resulting stress levels for the former. Subsequently, it aims to draw the upper frequency limit for the variants of the distributed circuit model for the transformer winding.
To study the propagation characteristics, time-domain thin-wire formulation of the electric field integral equation will be employed. Suitable modification to accommodate the cross section of the conductor is planned. The in-house code can be switched to a quasi-static domain and this will be employed to derive the upper frequency limit for the distributed circuit modeling approach.

Go top

Title: Multiview Video + Depth Data Acquisition, Compression and Application.
Speaker: Hemanth S
Advisor: Prof. K. R. Ramakrishnan
Venue: MMCR, EE
Date & Time: 31/07/2015, 4 pm
Abstract: Advances in display and camera technology have enabled 3D scene communication, stereoscopic displays are an example. But for display systems such as Auto-stereoscopic displays necessitates multi-camera scene acquisition along with depth information for each view. Depth information can be used to generate views from virtual viewpoints at the decoder. In this talk we present our work on Multiview Plus Depth data acquisition and compression.

Go top

Title: Adjusted Load Flow Solutions in Complementarity Framework
Speaker: Shantanu Chakrabarty
Advisor: Prof. P S Nagendra Rao
Venue: MMCR, EE
Date & Time: 24/07/2015, 4 pm
Abstract: Under Normal Conditions, electrical transmission networks work in their steady state and the calculations required to determine the characteristics of this state is termed as load flow or power flow.The solution is expected to provide information on the voltage angles and magnitudes,power flows across the individual transmission lines and losses.The network is represented by linear, bilateral and lumped elements but due to power and voltage constraints, the problem is non-linear.The solutions of the load flows are adjusted to incorporate the generator reactive power constraints, on load tap changing transformers,phase shifters, area interchange control and provision can also be made to include the constraints due to FACTS devices.This research work is concerned with the incorporation of complementarity framework to model the constraints due to devices mentioned above and incorporate them in the conventional load flow in matrix vector as opposed to solving it as an optimization problem.

Go top

Title: Integrating Dynamic Data for Predictive Operations in Power Systems
Speaker: Prof. Le Xie, Texas A&M University, USA.
Venue: MMCR, EE
Date & Time: 15th July 2015, 4 pm
Abstract: This talk concerns the handling and utilization of streaming data (such as synchrophasors and smart meters) for enhancing power system real-time physical and market operations. The first part of the talk analyzes the dimensionality of the phasor measurement unit (PMU) data under both normal and abnormal conditions. We observe that the underlying dimensionality is extremely low despite the high dimensions of the raw PMU data. Justification of this observation is proposed using linear dynamical systems theory. A novel early anomaly detection algorithm based on the switch of core subspace at the occurrence of an event is proposed. The second part of the talk presents our work of quantifying benefits of incorporating look-ahead dispatch with responsive demand from Electric Reliability Council of Texas (ERCOT) data. Demand elasticity at ERCOT is estimated, and the market price behavior with price responsive demand is analyzed. This talk concludes with several open interdisciplinary opportunities that would synergistically contribute towards a low-carbon smart grid.
Speaker bio: Le Xie is an Associate Professor (Sep 2015) in the Department of Electrical and Computer Engineering at Texas A&M University. He received B.E. in Electrical Engineering from Tsinghua Universityin 2004, S.M. in Engineering Sciences from Harvard in 2005, and Ph.D. from Carnegie Mellon in 2009. His industry experience includes internshipsat ISO-New England and Edison Mission Energy Marketing and Trading. His research interest includes modeling and controlin data-rich large-scale systems, grid integration of low-carbon energy resources, and electricity markets. Dr. Xie received the National Science Foundation CAREER Award,and theDepartment of EnergyRalph E. Powe Junior Faculty Enhancement Award. He was Texas A&M Engineering Select Young Fellow in 2013. He is an Editor of IEEE Transactions on Smart Grid, and the founding chair of IEEE Power and Energy SocietyWorking Group on Big Data Analytics for Grid Operations. He and his students received the Best Paper awards at North American Power Symposium and IEEE SmartGridComm2012. He is the founding faculty advisor of A&M Energy Club.

Go top

Title: Exact Phase Retrieval for Certain Classes of Signals
Speaker: Basty Ajay Shenoy
Advisor: Prof. Chandrasekhar Seelamantula
Venue: MMCR, EE
Date & Time: 03/07/2015, 4 pm
Abstract: The Fourier transform (spectrum) of signals are complex functions and are characterized by their magnitude and phase spectra. Phase retrieval is the reconstruction of the phase spectrum from the measurements of the magnitude spectrum of a signal. Such problems are encountered in imaging modalities such as frequency-domain optical coherence tomography (FDOCT), quantitative phase microscopy, digital holography, etc., where only the magnitudes of the wavefront can be detected by the sensors. The phase retrieval problem is a priori ill-posed, since an infinite number of signals can have the same magnitude spectrum. Typical phase retrieval techniques rely on certain prior knowledge about the signal, such as its support or sparsity, to reconstruct the signal. We consider two classes of signals, one being the two-dimensional parametric signals, the other being signals in shift-invariant spaces. In each of these cases, we develop phase retrieval methods that guarantee exact reconstruction when the measurements are noiseless. Further, we show applicability of the proposed methods for FDOCT image reconstruction.

Go top

Title: Discovering compressing serial episodes from event sequences
Speaker: Ibrahim A
Advisor: Prof. P. S. Sastry
Venue: MMCR, EE
Date & Time: 19/06/2015, 4 pm
Abstract: Frequent pattern mining is an important subfield of data mining with lots of important applications. It is the process of extracting interesting patterns in the data, where a pattern could signify, depending upon the type of the data, some local inherent structure or dependencies among certain variables or attributes of the data. Even though, there are many efficient algorithms for discovering frequent patterns, the number of frequent patterns could be very huge to make any use of these patterns. In this talk, we look at the problem of selecting a small subset of non-redundant episode patterns from a sequence data, that best represents the data. We propose a scheme for finding a small set of non redundant serial episodes, that summarizes and represents a given data sequence, using the Minimum Description Length (MDL) principle. The effectiveness of the selected serial episodes by our method over other summarization based schemes is shown with respect to three criteria: the interpretability of the patterns with respect to the data, the compression achieved while encoding the data using the selected serial episodes and the accuracy achieved in classification when the selected serial episodes are used as features.
In the second part of the talk, we propose a similarity measure to compare two sets of patterns. We prove that the similarity measure satisfies the conditions of a kernel. The measure is highly useful to find the degree of similarity between different datasets, which can be represented using sets of summarizing patterns. We show the effectiveness of our kernel based similarity measure in detecting changes in streams and also in classification.

Go top

Title : Control and Design of Power Converters for Renewable Energy Systems
Speaker : Abhijit K
Advisor : Prof. Vinod John
Date : 12/06/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Renewable energy sources such as photovoltaic, wind are increasingly used to reduce the dependence on conventional energy sources. Renewable energy sources normally require a power converter to convert their energy into standardized form of electricity. Power converters form the interface between the source and the grid or load. They must be designed and controlled to ensure very good power quality, efficiency and reliability. In this work, two power converter topologies with transformers are developed for renewable energy systems such as photovoltaic (PV) system with low input dc voltage. Control and modulation methods are proposed to improve power quality and efficiency in these power converter topologies.
Control of grid connected power converters requires grid synchronization. This is achieved by using phase-locked loops (PLLs). PLLs can have impact on the power quality of the power converters when the grid voltage contains harmonics, unbalance and dc offsets. Detailed analysis and designs of low-complexity PLLs such as synchronous reference frame PLL (SRF- PLL) and second-order generalized integrator (SOGI) based single-phase PLL is discussed. The proposed designs ensure that the control references generated using the PLLs will have minimal distortion and dc offsets to satisfy grid interconnection standards such as IEEE 1547-2003.

Go top

Title : Low-switching-frequency pulse width modulation techniques for high-power and high-speed induction motor drives
Speaker : Avanish Tripathi
Advisor : Prof. G. Narayanan
Date : 29/05/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Inverters based on semiconductor switches are employed now-a-days, to convert fixed-amplitude DC voltage into variable-amplitude variable-frequency AC voltage for electric motor drives and other applications. Pulse width modulation (PWM) is well known technique for operating these inverters. Sine-triangle PWM, space-vector PWM, selective-harmonic-elimination PWM and optimal PWM are some of the common PWM techniques used.
Depending upon the ratio of switching frequency and fundamental frequency (Pulse number, P), PWM techniques can be broadly divided into two streams i.e. high-switching-frequency PWM (P>21) and low-switching-frequency (LSF) PWM (P<21). LSF PWM techniques are employed for high-power electric motor drives (> 1 MW) due to constraints on switching losses. Further, due to limitation on maximum switching frequency of semiconductor switches, high speed drives (>10,000 rpm) also employ LSF PWM techniques. Due to low pulse number, the output PWM voltage contains dominating lower-order voltage harmonics. These lower-order voltage harmonics arise two major issues in the motor drives namely high total-harmonic-distortion (THD) in line current and high pulsating torque. An optimal PWM is proposed to minimize THD in line-current for complete range of operation considering a set of pulse numbers. Further, based on the optimal PWM, a hybrid PWM is proposed with maximum switching frequency of 250 Hz. The proposed PWM is compared with sine-triangle PWM with 450 Hz switching frequency. Another, optimal PWM to minimize pulsating torque is proposed for a set of pulse numbers. The proposed optimal PWMs are validated through simulations and experiments.

Go top

Title : Scattered Data Approximation by Regular Grid Smoothing
Speaker : Bibin Francis
Advisor : Dr. Muthuvel Arigovindan
Date : 22/05/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Scattered data approximation refers to techniques which estimates the value of the underlying function from the given non-uniform measurements. Non-uniform sampling introduces rapid variations in sampling density throughout the imaging region with severe under sampling in certain regions and oversampling in some other regions. So the scattered data approximation technique must take care of these variations in sampling densities and the measurement noise to produce meaningful results. In the present work, we have adopted a variational framework in which the reconstruction problem is posed as an unconstrained optimization problem. The weighted least square term in the problem measures the accuracy of the fit and a regularization term measures the smoothness of the reconstructed function. The solution of the above optimization approach involves solving a sparse system of linear equation which can be effectively implemented by using digital filtering.

Go top

Title : Localization of mechanical damages in a transformer winding
Speaker : Pritam Mukherjee
Advisor : Prof. L. Satish
Date : 15/05/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Localization of mechanical damage viz. displacement or deformation in a transformer winding poses many challenges. The measured frequency response is undoubtedly the most sensitive means of detecting it. Even though it is well-known that the frequency response has embedded in it relevant information pertaining to the damage, there exists no generalized method to extract and utilize it, say e.g., draw inference about location of the damage or its severity. With this motivation, a method has been developed which, at present, is able to locate any radial displacement based on terminal measurements. Theoretical formulation and experimental verification will be presented.

Go top

Title : Machine Listening: Making sense of sounds in the environment
Speaker : K V Vijay Girish
Advisor : Prof. A G Ramakrishnan
Date : 08/05/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Machine listening is the latest research area in the field of audio. Getting useful information about sounds in the environment is the first step in machine listening. A dictionary learning based algorithm is proposed to classify a sample audio signal as one amongst a finite set of different environmental audio sources. Cosine similarity measure is used to select the atoms during dictionary learning. Based on three objective measures proposed, namely, signal to distortion ratio, the number of non-zero weights and the sum of weights, a frame-wise source classification accuracy of 98.2% is obtained for twelve different sources.

Go top

Title : Correlation between Corona and RIV in Substation Hardware
Speaker : Debasish Nath
Advisor : Prof. Udaya Kumar
Date : 01/05/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Transmission of Electric Power from Generating stations to the load centers are done at very high voltages. Due to the inherent geometry, the line and substation hardware of EHV and UHV class can generate high Electric fields. This results in local ionization of air leading to a visual discharge called Corona. In addition, it also produces a hissing noise and leads to problematic Electromagnetic Interference. Extensive work has been carried out on this problem, however, limited to measured currents or assumed induced current pulse as the basis for any theoretical calculation. However, the phenomena is more complicated as the electron avalanche is in space surrounding the conductor and it produces its own electromagnetic field. The coupling of this field to the conductor is in different proportions and has different frequency characteristics. The quantification of this requires adequate representation of the associated dynamic electric fields, retardation effects and coupling mechanism. These have be aimed in the present work. This talk will elaborate on the field due to avalanche at different distances along with its indirect validation, coupling to the conductor and the difference in the frequency spectrum of the current and the actual field.

Go top

Title : Distributed Target Tracking in Camera Networks
Speaker : Shiva Kumar K A
Advisor : Prof. K R Ramakrishnan and Dr G N Rathna
Date : 10/04/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Due to the availability of low-cost cameras and communication, camera networks are being used in applications such as wide-area surveillance, smart homes, disaster response, etc. One way to analyse the videos in camera networks is the centralized method. Here, videos from all cameras are sent to a central server. But centralized method has the following disadvantages: High communication bandwidth, high memory and computational requirement at the central server. Also, if the central server fails then the whole network fails. Distributed processing circumvents these issues. In distributed processing, there would be no central server, cameras carry out local processing and communicate with immediate neighbours to iteratively improve the results. In this talk we will discuss about distributed target tracking in camera networks using the principles of sigma point information filters and average consensus algorithm.

Go top

Title : Numerical problems in the simulation of electromagnetic flow meter
Speaker : Sethupathy S
Advisor : Prof. Udaya Kumar
Date : 27/03/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: The numerical evaluation of the sensitivity of magnetic flowmeter is a best possible choice especially for liquid metal flow measurements. When classical Galerkin's finite element formulation is employed for this, it is known to introduce numerical oscillations at high flow rates. The magnetic field produced by the flow induced currents circulate within the pipe and it can be associated with the numerical problem. To overcome this, modified methods like stream-line upwind Petrov-Galerkin (SUPG) schemes are generally suggested in the allied areas like fluid dynamics, in which a similar dominance of advective (curl or circulation) component occurs over diffusion (divergence) component. However the direct application of SUPG scheme does not solve the problem completely.

Go top

Title : An Alternating Minimization Approach for Sparse Blind Deconvolution
Speaker : Aniruddha Adiga
Advisor : Prof. Chandrasekhar Seelamantula
Date : 20/03/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Deconvolution is a class of inverse problems that involves estimating a signal from the observation which is a noisy version of the original signal convolved with known system response. It finds applications in image deblurring, microscopy, speech analysis, astronomy, biomedical imaging to name a few. The talk will give an overview of deconvolution and ill-posedness associated with the problem. We will then discuss about blind deconvolution, which involves estimating both the original signal as well as the system response based on the noisy observations. In particular, we will consider a class of signals which can be modelled as a convolution of system response and point-source excitations. Although ill-posed, we use structures pertaining to the underlying signals to make the problem tractable. The cost function used for estimating the signals from observations incorporates the data fidelity term and sparsity constraint on excitation. However, joint-optimization of the cost function is a non-convex problem and hence hard to solve. We develop an alternating minimization algorithm that iteratively estimates one of the signals by keeping the other fixed. We also discuss about the convergence of the algorithm. We show application of the technique for epoch extraction in speech signals and compare results with sparse linear prediction and electroglottograph.

Go top

Title : Grid Outage and Its Prevention
Speaker : Ajit Kumar
Advisor : Prof. Indraneel Sen
Date : 13/03/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Widespread Power Outages is a major concern of power grid. Great blackouts are global phenomenon affecting millions of lives. In this talk, a brief survey of blackout will be presented followed by introduction of power system. Improving grid stability suffer from curse of dimensionality. Thereafter, a method to enhance power system stability will be presented using local measurements. In the second part of the talk, a nonlinear voltage controller designed using differential geometric theory will be discussed. Here, we will demonstrate that voltage regulation is decoupled with rotor oscillations.

Go top

Title : Modeling and Applications of Noise in KINECT
Speaker : Avishek Chatterjee
Advisor : Prof. Venu Madhav Govindu
Date : 06/03/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: 3D scanning is used extensively in many computer vision applications, e.g. human-computer interface (HCI), virtual reality, game programming, industrial monitoring, archeology, etc. Although applications such as HCI or gaming generally demand speed rather than precision, the accuracy of the 3D reconstruction is of crucial importance for many other tasks including archeology or quality monitoring in industrial production. The recent commercial availability of inexpensive 3D cameras like Kinect has opened up new possibilities for 3D scanning and shape reconstruction. 3D scans obtained from Kinect, are easy to use, but are affected by significant amounts of noise. This talk is devoted to a study of the intrinsic noise characteristics of such depth maps, i.e. the standard deviation of noise in estimated depth varies quadratically with the distance of the object from the depth camera. We validate this theoretical model against empirical observations and demonstrate the utility of this noise model in three popular applications: depth map denoising, volumetric scan merging for 3D modeling, We also integrate this noise model in a 3D reconstruction pipeline.

Go top

Title : Integrated CM Filter for Single-Phase and Three-Phase PWM Rectifiers
Speaker : Mohammad Hassan Hedayati
Advisor : Prof. Vinod John
Date : 27/02/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: Use of IGBT base grid connected power converters are increasing in applications such as power factor correction (PFC); adjustable speed drives (ASD), PWM rectifiers, battery chargers, active filters, etc. This is due to high efficiency, ease of control, low THD, etc. of the power converters. As the technology of semiconductor devices advances more, the power converters become more efficient. Faster devices can be switched at higher frequency and more efficiently, this makes the power converters even more attractive. However, fast turn-on and turn-off of the IGBT generate high dv/dt, which excites parasitic capacitances in the circuit. This leads to injection of narrow peaky current to ground that contains excitations at high frequencies of the spectrum. This causes problems such as shaft induced voltage, bearing currents, and high EMI/EMC noise level. Standards specify the limits of voltage disturbance on the mains by power converters for industrial, commercial, and domestic applications. This work, focuses on studying of different filter topology for single phase, parallel single phase and three phase grid connected power converter with reduced EMI and ground leakage current, different PWM methods and their impacts on EMI and leakage current, active and passive damping of the resonance oscillations in common mode (CM) and in the differential mode (DM) are investigated.

Go top

Title : Action Recognition in Videos using Causality Descriptors
Speaker : Sanath Narayan
Advisor : Prof. K. R. Ramakrishnan
Date : 30/01/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: An action is typically composed of different parts of the object moving in particular sequences. The presence of different motions (represented as a 1D histogram) has been used in the traditional bag-of-words (BoW) approach for recognizing actions. However the interactions among the motions also form a crucial part of an action. Different object-parts have varying degrees of interactions with the other parts during an action cycle. It is these interactions we want to quantify in order to bring in additional information about the actions. In this talk, I will discuss our causality based approach for quantifying the interactions to aid action classification. Granger causality is used to compute the cause and effect relationships for pairs of motion trajectories of a video. A 2D histogram descriptor for the video is constructed using these pairwise measures. Our method of obtaining pairwise measures for videos is also applicable for large datasets. We have conducted experiments on challenging action recognition databases such as HMDB51 and UCF50 to show that our causality descriptor helps in encoding additional information regarding the actions and performs on par with the state-of-the art approaches. Due to the complementary nature, a further increase in performance can be observed by combining our approach with state-of-the-art approaches.

Go top

Title : Analysis of Grid Connected Doubly Fed Induction Machine – Accurate steady state equivalent model and its impact on Voltage Stability of Power Grid
Speaker : V Seshadri Sravan Kumar
Advisor : Prof. D Thukaram
Date : 23/01/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: There is a continuous increase in the installed capacity of wind generators due to lack of sufficient generation to meet ever increasing load in some countries and search for renewable energy sources in the rest. This increase in penetration requires power engineers to tackle multifold challenges related to operational and stability aspects of the grid. Increasingly popular among the class of wind turbine generating units are the variable speed generators which use Doubly Fed Induction Machines (DFIM). In this talk, we re look at various modeling aspects of DFIM and discuss in brief some key observations. In particular, the short comings of existing steady state equivalent circuit of DFIM is emphasized. Further a proposed steady state equivalent circuit of DFIM which accurately represents its behavior is presented. The last part of the talk discusses in brief some aspects of voltage stability of power grid with penetration of wind generators.

Go top

Title : Finite-Rate-of-Innovation Signal Sampling and Reconstruction
Speaker : Mr. Satish M
Advisor : Prof. Chandra Sekhar Seelamantula
Date : 16/01/2015 (4.00 pm - 5.00 pm)
Venue : MMCR (C 241)
Abstract: In the last 5 decades there has been a rapid development in the field of digital computers and digital signal processors. With this development, the problem of representing a larger class of continuous-time signals by discrete measurements always attracted researchers from the signal processing community and other engineering/scientific fields. Among various discretization methods, Shannonâ€™s sampling theorem for banlimited signals is well known and applied widely in practical scenarios with the help of bandlimiting filters. Since Shannon, many researchers worked to represent a larger class of non-bandlimited signals. In this talk, we focus on a particular class of signals called finite-rate-of-innovation (FRI) signals which can be represented by finite number of parameters per unit time-interval. We discuss the basics of Shannonâ€™s sampling framework and then the topics of sampling and reconstruction methods for FRI signals. We show applications in ultrasound imaging and frequency-domain optical coherence tomography (FDOCT).

Go top

Title: Personalized Feedback in Sensor-enhanced Social Media Systems
Name of the Speaker: Prof. Mohan Kankanhalli
Venue: EE 303
Date & Time: 10:00 hrs 12th June Thursday 2014
Abstract: This talk will first start with a very brief overview of the SeSaMe (Sensor-enhanced Social Media Centre) at NUS. Then preliminary work from two SeSaMe research projects will be presented.
The first is about real-time photography assistance. To assist photographers in taking better photos, camera devices have intelligent features, such as auto-focus and face detection, etc. But capturing high-quality photos still remains a challenge for amateur users who lack skills and training. While post-processing can improve poorly captured images, we propose a photographic assistance system where we can utilize social media to learn photography rules. The system can then provide real-time suggestions based on learning from social media which will enable novice users to capture better quality photos.
The second project is about tweeting cameras for situation awareness. Motivated by the growing popularity of social media, we propose a multi-layer tweeting cameras framework to broadcast the events-of-interest, where various levels of semantic concepts are automatically detected and tweeted like humans. We define a unified Probabilistic Spatio-Temporal data structure to represent the low-level camera-based information, where an array of operators and analysis functions are employed to extract the mid-level concepts. We also have developed a graphical concept representation method, namely concept based image (C-Mage), as an intuitive data visualization tool. We have done preliminary testing of our approach on two real-world data-sets: New York City traffic feeds and NUS canteen feeds. The studies show promising results about the effectiveness of the proposed framework.
Some of the open problems in this area will also be highlighted.
Speaker bio: Mohan Kankanhalli is a Professor at the Department of Computer Science of the National University of Singapore. He is also the Vice Provost for Graduate Education at NUS. Before becoming the Vice Provost in 2014, he was the Associate Provost (Graduate Education) during 2011-2013. Earlier, he was the Vice-Dean for Academic Affairs & Graduate Studies at the NUS School of Computing during 2008-2010 and Vice-Dean for Research during 2001-2007.
Mohan obtained his BTech from IIT Kharagpur and MS & PhD from the Rensselaer Polytechnic Institute.
His current research interests are in Multimedia Systems (content processing, retrieval) and Multimedia Security (surveillance and privacy). He also directs Singapore's National Research Foundation funded Centre for "Sensor-enhanced Social Media" (sesame.comp.nus.edu.sg).
Mohan is actively involved in organizing of many major conferences in the area of Multimedia. He was the Technical Program Co-Chair for ICMR 2014. He was the Director of Conferences for ACM SIG Multimedia during 2009-2013. He is on the editorial boards of several journals including the ACM Transactions on Multimedia, Springer Multimedia Systems Journal, Pattern Recognition Journal and Multimedia Tools & Applications Journal. He is a Fellow of IEEE.

Events and Seminars Archieve

Events

Effective Communication Seminar Series

Other seminars list

Thesis Colloquium/Defence List

Details of Seminars/Colloquium/Defence

Sl.No.	Name of the student	Date of Colloquium	Title of the Thesis
57	Praveen Kumar Pokala	06/08/2021	Robust Nonconvex Penalties for Solving Sparse Linear Inverse Problems and Applications to Computational Imaging. Guide: Prof. Chandra Sekhar Seelamantula
56	Sanjay Viswanath	07/06/2021	Spacially Adaptive Regularization for Image Restoration Guide : Dr. Muthuvel Arigovindan
55	Rejesh R A	13/05/2021	Novel Regularized Image Reconstruction Methods for Sparse Photoacoustic Tomography. Guide: Dr. Muthuvel Arigovindan
54	Shome Subhra Das	26/04/2021	Techniques for estimating the direction of pointing gestures using depth images in the presence of orientation and distance variations from the depth sensor. Guide: --
53	Indla Rajitha Sai Priyamvada	09/02/2021	Analysis and Enhancement of Stability of Power Systems with Grid-scale Photovoltaic Power Plants. Guide : Dr. Sarasij Das
52	Aravind Illa	18/12/2020	Acoustic-Articulatory Mapping: Analysis and Improvements with Neural Network Learning Paradigms Guide : Prof. Prasanta Kumar Ghosh
51	Harish Kumar J R	20/11/2020	Shape-constrained Biomedical Image Segmentation and Applications. Guide : Prof. Chandra Sekhar Seelamantula
50	Rupam Pal	20/11/2020	Investigation on the Influence of soil's electrical parameters on the lightning stroke-current evolution and the fields in close range. Guide : Prof. Udaya Kumar
49	R Gnanavignesh	19/10/2020	Parallel Computing Techniques for High Speed Power System Solutions. Guide : Dr. U J Shenoy
48	Achuth Rao M V	16/10/2020	Probabilistic source-filter model of speech, Guide : Prof. Prasanta Kumar Ghosh
47	Vidyadhar Upadhya	07/08/2020	Efficient Algorithms for Learning Restricted Boltzmann Machines. Guide : Prof. P S Sastry
46	Vinod V	30/07/2020	Protection Schemes for maintaining the Coordination Time Interval between the relays in Micro-grid. Guide : Dr. U. J. Shenoy
45	Purvi Agrawal	03/07/2020	Neural Representation Learning for Speech and Audio Signals. Guide : Dr. Sriram Ganapathy
44	Madhavaraj A	26/06/2020	Strategies for Handling Large Vocabulary and Data Sparsity Problems for Tamil Speech Recognition Guide : Prof. A. G. Ramakrishnan
43	Shimul Kumar Dam	10/12/2019	Power Electronic Converters for Condition Monitoring and Voltage Equalization of Batteries Guide : Prof. Vinod John
42	Anirban Pal	10/12/2019	Unidirectional High-Frequency-Link DC-Three Phase AC Conversion: Topology, Modulation and Converter Design. Guide : Dr. Kaushik Basu
41	Sunil Rudresh	05/12/2019	Sampling of Structured Signals -- Algorithms and Imaging Applications. Guide: Prof. Chandra Sekhar Seelamantula
40	Ram Krishna Pandey	16/09/2019	Knowledge-driven training of deep models for better reconstruction and recognition. Guide : Prof. A G Ramakrishnan
39	Nazreen	09/07/2019	Class-specific and noise-specific speech enhancement approaches. Guide : Prof. A G Ramakrishnan
38	Chiranjeevi Yarra	08/07/2019	Pronunciation assessment and semi-supervised feedback prediction for spoken English tutoring. Guide : Dr. Prassanta Kumar Ghosh
37	Dipanwita Sinha M	05/07/2019	Estimating the variation of NOx in diesel exhaust treated with discharge-plasma/ozone injection: a case study on modelling with ANN & Dimensional-Analysis. Guide : Prof. B S Rajanikanth
36	Nisha Meenakshi	25/04/2019	Analysis of whispered speech and its conversion to neutral speech. Guide : Dr. Prassanta Kumar Ghosh
35	Syed Shahjahan Ahmad	16/04/2019	Modeling, Characterization, Control and Design of Switched Reluctance Machines. Guide : Prof. G. Narayanan
34	Bibin Francis	19/02/2019	Fast and Robust Biomedical Image Reconstruction from Nonuniform Samples. Guide : Dr. Muthuvel Arigovindan
33	Anand Sharma	18/02/2019	Hindi OnlineHandwritten Character Recognition. Guide : Prof. A G Ramakrishnan
32	Pallavi Bharadwaj	30/01/2019	Modelling, Optimisation and Control of Photovoltaic Energy Conversion Systems. Guide : Prof. Vinod John
31	Mr. Shiva Kumar H R	28/01/2019	Attention-Feedback and Representations in OCR Guide : Prof. A G Ramakrishnan
30	Sanjay Ghosh	09.01/2018	Kernel-Based Image Filtering: Fast Algorithms and Applications. Guide: Dr. Kunal Narayan Chaudhury
29	Suraj K	12/12/2018	Algorithms for Processing RGBD Images and Videos for Depth-Based 3D Video Systems. Guide : Dr. Soma Biswas
28	Venkatramanan D	30/11/2018	Modeling, Analysis and Control of Reconfigurable Battery/Grid Tied Solar Photovoltaic Inverter. Guide : Prof. Vinod John
27	Apeksha Madhukar	28/11/2018	Stationary diesel exhaust treatment by blending discharge plasma/ozone with industry wastes: a study on abatement of NOx and THC. Guide : Prof. B.S. Rajanikanth
26	Alok Ranjan Verma	23/08/2018	Degradation Studies on Polymeric Insulators used for EHV and UHV Transmission. Dr. Subba Reddy B
25	Vijayakumar Solaiselvam	20/08/2018	Detection and Imaging of Buried Landmines Using Ultra Wide Band. Guide: Dr. Joy Thomas M
24	Sivaram Prasad Mudunuri	17/08/2018	Face Recognition in Unconstrained Environment. Guide: Dr.Soma Biswas
23	Nisha Meenakshi	13/08/2018	Analysis of whispered speech and its conversion to neutral speech. Guide: Prasant Kumar Ghosh
22	Debasish Nath	10/08/2018	Total electric field due to an electron avalanche and its coupling to the transmission line conductors . Guide: Prof. Udaya Kumar
21	P. Sidharthan	31/07/2018	Performance Improvement of a High Voltage Power Converter for Microwave Power Module. Guides : Prof. G. Narayanan, Dr. S. K. Datta, MTRDC, DRDO
20	Narasimhan R S	09/07/2018	Target Detection and Tracking under Non-ideal Conditions in Airborne Radars Guide: Prof. P S Sastry
19	Santosh J	18/05/2018	Upper Frequency Bound for Circuit Based Models of Coils. Guide: Prof. Udaya Kumar
18	Sk. Mohammadul Haque	06/03/2018	Denoising and Refinement Methods for 3D Reconstruction. Guide: Prof. Venu Madhav Govindu
17	Saichand K	26/02/2018	Modeling and Control of Ultracapacitor based Bidirectional DC-DC Converter Systems. Guide: Prof. Vinod John
16	Shiva Kumar K A	29/01/2018	Distributed Target Tracking in Camera Networks. Guides: Prof. K R Ramakrishnan (Retd) and Dr. G N Rathna
15	Anil Kumar Adapa	14/11/2017	Analysis, Control and Application of Active Phase Converters for Single-Phase Power Grids
14	Subhadip Mukherjee	13/09/2017	Sparsity Driven Solutions to Linear and Quadratic Inverse Problems Guide :Prof. Chandra Sekhar Seelamantula
13	Hemanth Kumar. S	28/08/2017	On Applications of 3D-Warping and An Analysis of a RANSAC Heuristic Guide: Dr. G. N. Rathna
12	Ajit Kumar	17/07/2017	Enhancement of Small Signal Stability of Power Systems: Novel Approaches. Guide : Dr. Indraneel Sen & Dr. Gurunath Gurrala
11	Debasish Nath	17/07/2017	Total electric field due to an electron avalanche and its coupling to the transmission line conductors . Guide : Prof. Udaya Kumar
10	Nimesh V	06/07/217	Dual Comparison One Cycle Control for Grid Connected Converters. Guide: Prof. Vinod John
9	Subhash Joshi	22/05/2017	Power Electronic Technologies for Medium and High Power High Voltage Power Supplies Guide: Prof. Vinod John
8	Shakthi Prasad D	22/05/2017	Investigations on the Corona induced degradation on Polymeric Insulating Samples Guides: Dr. Subba Reddy B. and Prof. B. S. Rajanikanth
7	Vijay H. Bhosale	26/10/2016	Developmental Studies On Ultra-Wide Band Type High Power Electromagnetic Radiating System Guide:
6	K V Vijay Girish	28/09/2016	Sparse representation and acoustic-phonetics knowledge for speech analysis Guide : Prof. A G Ramakrishnan
5	Basty Ajay Shenoy	15/09/2016	Phase Retrieval and Hilbert Integral Equations -- Thinking Beyond Minimum Phase Guide: Prof. Chandra Sekhar Seelamantula
4	Satish Mulleti	25/08/2016	Sub-Nyquist Sampling of Structured Signals and Applications to Imaging Guide: Prof. Chandra Sekhar Seelamantula
3	Aniruddha Adiga	18/08/2016	Sparsity Motivated Auditory Wavelet Representation and Blind Deconvolution Guide: Prof. Chandra Sekhar Seelamantula
2	Shantanu Chakrabarty	29/07/2016	Algorithms for Adjusted Load Flow Solutions using the Complementarity Principle Guide: Prof. P S Nagendra Rao
1	Sunil Kumar P.R.	31/05/2016	Design and Analysis of Real-time Message Scheduling under FlexRay Protocol

Sl.No.	Name of the student	Date of Defence	Title of the Thesis
67	Shome Subhra Das	17/11/2021	Techniques for estimating the direction of pointing gestures using depth images in the presence of orientation and distance variations from the depth sensor.
66	Indla Rajitha Sai Priyamvada	14/10/2021	Analysis and Enhancement of Stability of Power Systems with Utility-scale Photovoltaic Power Plants Guide: Dr. Sarasij Das
65	Aravind Illa	29/09/2021	Acoustic-Articulatory Mapping: Analysis and Improvements with Neural Network Learning Paradigms. Guide: Prof. Prasanta Kumar Ghosh
64	Achuth Rao M V	28/06/2021	Probabilistic source-filter model of speech . Guide : Prof. Prasanta Kumar Ghosh
63	R. Gnanavignesh	12/03/2021	Parallel Computing Techniques for High Speed Power System Solutions. Guide : Dr. U J Shenoy
62	Ramu Nair	08/03/2021	Emulation of wind turbine and sensorless control of doubly-fed induction generator for wind energy application. Guide : Prof.G. Narayanan
61	Purvi Agrawal	28/01/2021	Neural Representation Learning of Speech and Audio Signals Guide: Dr. Sriram Ganapathy
60	Vinod V	14/12/2020	Protection Schemes for maintaining the Coordination Time Interval between the relays in Micro-grids. Guide : Dr U Jayachandra Shenoy
59	Madhavaraj A	08/12/2020	Strategies for Handling Large Vocabulary and Data Sparsity Problems for Tamil Speech Recognition. Guide : Prof. A. G. Ramakrishnan
58	Sunil Rudresh	29/09/2020	Sampling of Structured Signals -- Algorithms and Imaging Applications. Guide : Prof. Chandra Sekhar Seelamantula
57	Syed Shahjahan Ahmad	07/09/2020	Modeling, Characterization, Control and Design of Switched Reluctance Machines. Guide : Prog. G Narayanan
56	Shimul Kumar Dam	28/07/2020	Power Electronic Converters for Condition Monitoring and Voltage Equalization of Batteries. Guide : Prof. Vinod John
55	Anirban Pal	27/07/2020	Unidirectional High-Frequency-Link DC to Three Phase AC Conversion: Topology, Modulation and Converter Design. Guide : Dr. Kaushik Basu
54	Chiranjeevi Yarra	20/04/2020	Pronunciation assessment and semi-supervised feedback prediction for spoken English tutoring. Guide : Prof. Prasanta Kumar Ghosh
53	Ram Krishna Pandey	13/02/2020	Knowledge-driven training of deep models for better reconstruction and recognition. Guide : Prof. A G Ramakrishnan
52	Nazreen P M	29/12/2019	Class-specific and noise-specific speech enhancement approaches. Guide : Prof. A G Ramakrishnan
51	Venkatramanan D	12/12/2019	Modeling, Analysis, and Control of Reconfigurable Battery/Grid-Tied Solar Photo-Voltaic Inverter. Guide Prof. Vinod John
50	G Nisha Meenakshi	22/10/2019	Analysis of whispered speech and its conversion to neutral speech. Guide : Prasanta Kumar Ghosh
49	Pallavi Bharadwaj	10/10/2019	Modelling, Optimisation and Control of Photovoltaic Energy Conversion Systems. Guide : Prof. Vinod John
48	Bibin Francis	29/08/2019	Fast and Robust Biomedical Image Reconstruction from Nonuniform Samples. Guide : Dr. Muthuvel Arigovindan
47	Sivaram Prasad Mudunuri	27/06/2019	Face Recognition in Unconstrained Environment, Guide : Dr. Soma Biswas
46	Shiva Kumar H R	24/06/2019,	Attention Feedback and Representations in OCR. Guide : Prof. A G Ramakrishnan
45	Anand Sharma	20/06/2019	Hindi Online Handwritten Character Recognition. Guide : Prof. A G Ramakrishnan
44	Suraj K	29/04/2019	Algorithms for Processing RGBD Images and Videos for Depth-Based 3D Video Systems. Guide : Dr. Soma Biswas
43	Debasish Nath	27/02/2019	Total Electric field due to a singel electron avalanche and its coupling to transmission line conductor. Guide : Udaya Kumar
42	Suresh Kirthi K	21/02/2019	Multisource Subnetwork Level Transfer in Deep CNNs Using Bank of Weight Filters. Guide :P S Sastry and K R Ramakrishnan
41	Apeksha Madhukar	27/02/2019	Stationary diesel exhaust treatment by blending discharge plasma/ ozone with industry wastes: a study on abatement of NOx and THC Guide : B S Rajanikanth
40	Subhash Joshi TG	07/12/2018	Power Electronic Technologies for Medium and High Power High Voltage Power Supplies. Guide : Prof. Vinod John
39	Alok Ranjan Verma	05/12/2018	Degradation Studies on Polymeric Insulators used for EHV and UHV Transmission Guide : Dr. Subba Reddy B
38	Narasimhan R S	26/11/2018	Target Detection and Tracking under Non-ideal Conditions in Airborne Radars: Guide: Prof. P S Sastry
37	Saichand K	08.11.2018	Modeling, Analysis and Control ofUltracapacitor based Bidirectional DC-DC Converter Systems. Guide : Prof. Vinod John
36	Santosh J	26/10/2018	Upper frequency bound on circuit based models for transformer windings. Guide : Prof. Udaya Kumar
35	Subhadip Mukherjee	26/09/2018	Sparsity Driven Solutions to Linear and Quadratic Inverse Problems. Guide: Chandra Sekhar Seelamantula
34	Sk. Mohammadul Haque	26/09/2018	Denoising and Refinement Methods for 3D Reconstruction. Guide: Prof. Venu Madhav Govindu
33	Nimesh V	11/06/2018	Dual Comparison One Cycle Control for Grid Connected Converters Guide : Prof. Vinod John
32	Anil Kumar Adapa	04/06/2018	Analysis,Control and Applications of Active Phase Converters for Single-Phase Power Grids. Guide : Prof. Vinod John
31	Shiva Kumar K A	11/05/2018	Distributed Target Tracking in Camera Networks. Guides: Prof. K R Ramakrishnan (Retd) and Dr. G N Rathna
30	Shantanu Chakrabarty	27/04/2018	Algorithms for Adjusted Load Flow Solutions using the Complementarity Principle Guide: Dr. Sarasij Das
29	Ajit Kumar	18/04/2018	Enhancement of Small Signal Stability in Power Systems: Novel Approaches. Guides: Prof. Indraneel Sen and Dr. Gurunath Gurrala
28	Shantanu Chakrabarty	27/04/2018	Algorithms for Adjusted Load Flow Solutions using the Complementarity Principle. Guide : Dr.Sarasij Das
27	Hemanth Kumar. S	27/03/2018	On Applications of3D-Warping and An Analysis of a RANSAC Heuristic. Guide : Dr. G. N. Rathna
26	Shakthi Prasad D	02/03/2018	Investigations on the Corona induced degradation on Polymeric Insulating Samples Guides: Dr. Subba Reddy B. and Prof. B. S. Rajanikanth
25	Basty Ajay Shenoy	17/01/2018	Phase Retrieval And Hilbert Integral Equations- Beyond Minimum-Phase. Guide: Prof. Chandra Sekhar Seelamantula
24	Bhosale Vijay Hiralal	26/12/2017	Developmental Studies On Ultra-wide Band Type High Power Electromagnetic Radiating System Guide: Dr. Joy Thomas M
23	Aniruddha Adiga	30/11/2017	Sparsity motivated auditory wavelet representation and blind deconvolution
22	Avanish Tripathi	11/10/2017	Low Switching Frequency Pulse Width Modulation for Induction Motor Drives
21	K. Venkata Vijay Girish	25/09/2017	Speech and Noise Analysis using Sparse Representation and Acoustic-Phonetics Knowledge. Guide : Prof. A G Ramakrishnan
20	M. Satish	11/09/2017	Sub-Nyquist Sampling and Super-Resolution Imaging. Guide : Chandra Sekhar Seelamantula
19	P.R.Sunil Kumar	04/09/2017	Design and Analysis of Real-time Message Scheduling under FlexRay Protocol. Guide: Prof. P S Sastry
18	Joseph Vimal Vas	11/08/2017	Electromagnetic Properties of Carbon based Polymer Nanocomposites for Shielding, Chaffing and Camouflage Applications. Guide : Dr. Joy Thomas M
17	Sethupathy S.	28/04/2017	Stable Galerkin Finite Element Formulation for the Simulation of Electromagnetic Flowmeter Guide: Prof. Udaya Kumar
16	Anirudh Guha	10/03/2017	Dead-time induced oscillations in voltage source inverter-fed induction motor drives Guide: G. Narayanan
15	Sachin Srivastava	05/12/2016	Behavior of Distance Relay Characteristics on Interconnecting Lines Fed from Wind Farms Guide: Dr. U. J. Shenoy
14	S P Manivannan	07/10/2016	DSA Image Registration and Respiratory Motion Tracking using Probabilistic graphical Models. Guide: Prof. K. R. Ramakrishnan
13	Avishek Chatterjee	27/06/2016	Geometric Calibration and Shape Refinement for 3D Reconstruction, Guide: Venu Madhav Govindu
12	Tukaram Moger	20/06/2016	Reactive Power Planning and Operation of Power Systems with Wind Farms for Voltage Stability Improvement, Guide:Prof. D. Thukaram
11	V.Seshadri Sravan Kumar	03/06/2016	Modeling and Analysis of Grid Connected Variable Speed Wind Generators, Guide:
10	A. Ibrahim	27/05/2016	Effective Characterization of Sequence Data through Frequent Episodes, Guide: Prof. P. S. Sastry
9	Mr. Pritam Mukherjee	26/05/2016	,A Novel Generalized Framework to Diagnose True Radial and Axial Displacements in an Actual Transformer Winding Guide: Prof. L. Satish
8	Mr.Manoj K.Mandlik	23/04/2015	Moisture Aided Degradation of Oil Impregnated Paper Insulation in Power Transformers, Guide:
7	Abhijit K	18/04/2016	Design and Control of Power Converters for Renewable Energy Systems, Guide: Prof. Vinod John
6	Binoj Kumar A.C	17/03/2016	Experimental Studies on Acoustic Noise Emitted by Induction Motor Drives Operated with Different Pulse Width Modulation Schemes, Guide:Prof. G. Narayanan
5	Ms.Anusuya Bhattacharyya	20/11/2015	Discharge plasma based NOx abatement in engine exhaust assisted by industry wastes: A parametric evaluation with diesel fuels and corona electrodes, Guide: Prof. B.S. Rajanikanth
4	Ms.N.S.Jyothi	29/06/2015	Thermal and Electrical Degradation of Resin Impregnated Insulation for High Voltage Transformer Bushings, Guide:
3	Mr.A.P.Prathosh	24/06/2015	Temporal Processing for event-based Speech Analysis with Focus on stop consonants, Guide: Prof. A. G. Ramakrishnan
2	Mr.V. S. S. Pavan Kumar Hari	28/03/2015	Space-Vector-Based Pulse Width Modulation Strategies to Reduce Pulsating Torque in Induction Motor Drives, Guide: Prof. G. Narayanan
1	Mrs.Sunitha K	25/07/2014	Coupling of Electromagnetic Fields from Intentional High Power Electromagnetic Sources with a Buried Cable and an Airborne Vehicle in Flight, Guide: