Signal Processing Seminar

Statistical and Array Signal Processing, Compressed Sensing, Tensor Analysis, Wireless Communications

Feiyu Wang

Additional information ...

Signal Processing Seminar

Graph Signal Processing, Data Science

Alberto Natali

Additional information ...

Signal Processing Seminar

Partial discharges recognition and (localization) in Gas Insulated Systems (GIS) using the cross wavelet transform.

Fabio Muñoz Muñoz
TU Delft

Partial Discharges (PD) detection is an essential tool for the diagnosis of high-voltage equipment because of their accuracy to detect and quantify defects and damages in the dielectric insulation, where the detection implies the measurement, acquisition, storage and processing of the PD phenomenon. Nowadays, the most widespread PD detection system is based on electrical measurements, in which the PD signals are acquired in the form of individual or series of electrical pulses.

In spite of the PD measurement has been exhaustively researched over the years, the separation of PD pulses from noise is one of the main challenges, especially in online applications. Noise contamination still one of the significant problems for PD measurements. Several studies have focused on the PD pulses separation and denoising techniques for PD measurements, in which the wavelet transform has been extensively used because is capable of locating time and frequency components allowing the analysis of aperiodic signals with irregular and transition features, such as the partial discharges. However, a major problem that most of these denoising techniques face is the ingress of external interferences having time-frequency characteristics similar to the partial discharge signals: periodic pulse-shaped interferences from power electronics, PD and corona discharges from the external power system, electrical pulses from switching operations, lightings, etc. This external noise can cause a false indication of PD activity, reducing the effectivity of the PD measurements as a diagnostic tool.

In PD measurement systems multiple signals can be simultaneously acquired for each PD event. Recording each signal through different sensors may provide extra useful information about the real nature of the waveform recorded. Tools like the correlation and trend analysis can provide the significance of relationships between the signals recorded. Nevertheless, these tools may not detect correlations if the signals are phase shifted; for instance, a phase shift of 180 ° between the signals may appear uncorrelated. The cross-correlation and the cross-spectral analysis can detect the phase shift, but only as average values and in stationary signals. For analysing aperiodic signals with irregular and transition features, the most suitable tool is the cross wavelet transform because it exposes regions with high common power and reveals the local relative phase between both signals.

In this presentation, we introduce the partial discharge measurements, the PD signals propagation in GIS, the cross wavelet as a tool to separate the PDs from the external disturbances, and some of the challenges that we are facing in the PD localization and detection in GIS.

Signal Processing Seminar

On Unlimited Sampling and Reconstruction: A New Way to Sense the Continuum

Ayush Bhandari

Almost all forms of data are captured using digital sensors or analog-to-digital converters (ADCs) which are inherently limited by dynamic range. Consequently, whenever a physical signal exceeds the maximum recordable voltage, the digital sensor saturates and results in clipped measurements. For example, a camera pointed towards the sun leads to an all-white photograph. Motivated by a variety of applications including scientific imaging, communication theory and digital sensing, a natural question that arises is: Can we capture a signal with arbitrary dynamic range?

In this work, we introduce the Unlimited Sensing framework which is a novel, non-linear sensing architecture that allows for recovery of an arbitrarily high dynamic range, continuous-time signal from its low dynamic range, digital measurements. Our work is based on a radically different ADC design, which allows for the ADC to reset rather than to saturate, thus producing modulo or folded samples.

In the first part of this talk, we discuss a recovery guarantee akin to Shannon’s sampling theorem which, remarkably, is independent of the maximum recordable ADC voltage. Our theory is complemented with a stable recovery algorithm. Moving further, we reinterpret the unlimited sensing framework as a generalized linear model and discuss the recovery of structured signals such as continuous-time sparse signals. This new sensing paradigm that is based on a co-design of hardware and algorithms leads to several interesting future research directions. On the theoretical front, a fundamental interplay of sampling theory and inverse problems raises new standalone questions. On the practical front, the benefits of a new way to sense the world (without dynamic range limitations) are clearly visible. We conclude this talk with a discussion on future directions and relevant applications.

Additional information ...

Signal Processing Seminar

Sparse Bayesian Learning: A Beamforming and Toeplitz Approximation Perspective

Bhaskar Rao

Sparse Bayesian Learning (SBL) methods that employ a Gaussian scale mixture prior have been successfully applied for solving the sparse signal recovery (SSR) problem. The SBL-EM based inference algorithm will be examined and interpreted using a beamforming framework. A contrast with the classical minimum power distortionless response (MPDR) beamformer will be drawn and the benefits highlighted. An interesting finding is the ability of SBL to deal with correlated sources. For a uniform linear array (ULA), the Toeplitz approximation property of SBL will be discussed and the potential benefits for a nested array demonstrated.

Speaker Biography

Bhaskar D. Rao received the B.Tech. degree in electronics and electrical communication engineering from the Indian Institute of Technology, Kharagpur, India, in 1979 and the M.S. and Ph.D. degrees from the University of Southern California, Los Angeles, in 1981 and 1983, respectively. Since 1983, he has been with the University of California at San Diego, La Jolla, where he is currently a Distinguished Professor in the Electrical and Computer Engineering department. He is the holder of the Ericsson endowed chair in Wireless Access Networks and was the Director of the Center for Wireless Communications (2008-2011). Prof. Rao’s interests are in the areas of digital signal processing, estimation theory, and optimization theory, with applications to digital communications, speech signal processing, and biomedical signal processing.

Prof. Rao was elected fellow of IEEE in 2000 for his contributions to the statistical analysis of subspace algorithms for harmonic retrieval. His work has received several paper awards; 2013 best paper award at the Fall 2013, IEEE Vehicular Technology Conference for the paper “Multicell Random Beamforming with CDF-based Scheduling: Exact Rate and Scaling Laws,” by Yichao Huang and Bhaskar D Rao, 2012 Signal Processing Society (SPS) best paper award for the paper “An Empirical Bayesian Strategy for Solving the Simultaneous Sparse Approximation Problem,” by David P. Wipf and Bhaskar D. Rao published in IEEE Transaction on Signal Processing, Volume: 55, No. 7, July 2007, 2008 Stephen O. Rice Prize paper award in the field of communication systems for the paper “Network Duality for Multiuser MIMO Beamforming Networks and Applications,” by B. Song, R. L. Cruz and B. D. Rao that appeared in the IEEE Transactions on Communications, Vol. 55, No. 3, March 2007, pp. 618 630. ( awards/rice.html), among others. Prof. Rao is also the recipient of the 2016 IEEE Signal Processing Society Technical Achievement Award.

Prof. Rao has been a member of the Statistical Signal and Array Processing technical committee, the Signal Processing Theory and Methods technical committee, the Communications technical committee of the IEEE Signal Processing Society and is currently chair of the Machine learning for Signal Processing technical committee.

Additional information ...

Signal Processing Seminar

Realizing the potential of precision medicine with mobile sensing and analytics

Emre Ertin
Ohio State University

Recent advances in wearable sensing and mobile computing have given researchers the ability to collect unprecedented amounts of data about everything from biology to behavior that can explain and improve people's health status. Day-to-day data from wearable sensors allows for better and more personalized decisions in regard to health care and management. Specifically, continuous monitoring of physiology and behavior can help us to assess disease risks, perform disease prevention and early detection of chronic conditions. However, there still exist a multitude of challenges to implement this vision of precision medicine. Wearable sensors provide large, noisy, complex data streams about the many facets of our life and health, but there is still a a gaping need for computational techniques that can transform sensor data into set of useful bio-markers readily interpretable by clinicians. This talk will describe our recent work in pairing rich probabilistic models with Bayesian methods to dramatically expand the scale and quality of physiological data we can obtain in the field while minimizing the burden to participants.

Signal Processing Seminar

Privacy-Preserving Distributed Optimization via subspace perturbation: A generalized convex optimization approach

Qiongxiu Li (Jane)
Audio Analysis Lab, Aalborg University

MSc ME Thesis Presentation

A Dynamic Zoom ADC for Audio Applications

Efraïm Eland

Audio ADCs used in high-fidelity portable audio and IoT are not only required to have high linearity and dynamic range (DR) but are also expected to be very energy efficient and occupy minimum silicon area. Zoom-ADCs combine a coarse asynchronous SAR with a fine Delta-Sigma Modulator (∆ΣM) to satisfy these requirements. Existing zoom ADC architectures are limited in terms of SQNR due to the need for the fine ADC to have some over-ranging. That, together with the leakage of the SAR ADC’s quantization noise, “fuzz,” into the audio band, puts a lower limit on the sampling frequency.
This thesis describes the design of a zoom-ADC for an audio bandwidth of 20kHz. Using a 4-level quantizer, instead of a conventional 1b quantizer, mitigates the adverse effects of over-ranging, making it possible to keep a very low sampling frequency. On top of that, it makes use of a simple, low power analog “fuzz” cancellation scheme to prevent the SAR quantization noise from leaking into the audio band.
The chip has been prototyped in a standard 160nm CMOS technology and consumes 339μW with 107.7dB DR and 105dB SNDR. Compared to state-of-the-art ADCs with a similar bandwidth, this work achieves a 2x lower OSR (fs = 2.5MHz), significantly improving the energy efficiency and achieving a Schreier FoM of 185.4dB.

MSc ME Thesis Presentation

Rail-to-rail input and output amplifier for ADC front-end applications.

Shubham Khandelwal

This work presents a unity-gain stable operational amplifier for an ADC front-end application. The op-amp focuses on delivering high linearity with low noise and offset while driving a switched capacitor load. To accomplish this the op-amp employs Current Spillover, Chopping and Gain-Boosting techniques. The op-amp achieves THD of -108 dB at 10kHz, offset of 2.7 µV and input noise density of 19.3 nV/√Hz while consuming 504 µW; resulting in an NEF of 12.28. The op-amp is fabricated in 0.16 µm CMOS technology and occupies 0.1 mm2 area.

MSc SS Thesis Presentation

Estimating the room impulse response

Gabriele Zacca

The response of a sound system in a room primarily varies with the room itself, the position of the loudspeakers and the listening position. The room boundaries cause reflections of the sound that can lead to undesired effects such as echoes, resonances or reverberation. Therefore the location of these large reflecting surfaces is important information for sound field estimation in a room.

This work focuses on exploiting the inherent information present in echoes measured by microphones, to infer the location of nearby reflecting surfaces. A built-in microphone array is used that is co-located with the loudspeaker. The loudspeaker probes the room by emitting a known signal. A signal model is proposed which provides a relationship between reflector locations and measured microphone signals.

The locations of reflections are estimated by fitting a sparse set of modeled reflections with measurements. We present two novelties with respect to prior art. First, the method is end-to-end where from raw microphone measurements it outputs an estimate of the location of reflectors. Where specifically for the compact uniform circular microphone array the symmetry is exploited to create an algorithm that is of reduced computational complexity. Secondly, the model is extended to include a loudspeaker model that is aware of the inherent directivity pattern of the loudspeaker.

The performance of the proposed localization method is compared in simulation to the existing state-of-the-art localization methods. Real world measurements are also used to validate the proposed loudspeaker model.

MSc SS Thesis Presentation

Atrial Fibrillation: Estimation of the local activation time in high-resolution mapping data

Bart Kölling

A common cardiac arrhythmia is atrial fibrillation, which is becoming more widespread worldwide. Currently there is some understanding about the mechanisms behind atrial fibrillation, however more insight into the conduction of the atrial tissue is desired.

Therefore, invasive mapping studies have been performed where an array of electrodes is used to record the electrical activity on the heart’s surface during open-chest surgery. The moment in time when the tissue under an electrode depolarizes, called the local activation time can be used to reconstruct the propagation pattern of the signal that triggers the tissue to contract.

In this thesis, the application of the cross-correlation for estimation of the local activation time of the atria is investigated. Specifically, the benefits of not only cross-correlating electrode pairs that are close, but also pairs that are far away are evaluated. A framework is constructed, based on a graph, that defines these higher order neighbouring pairs of electrodes.

This is compared to the golden standard of using the steepest deflection of an electrogram, as well as to other methods using the cross-correlation. Experiments are done on simulated electrograms where the true activation times are available, as well as on natural data recorded from patients. Finally some future research is proposed to investigate for which morphologies the proposed cross-correlation based methods may be most effective.

Additional information ...

MSc SS Thesis Presentation

Signal Modelling and Imaging of Low Field MRI

Sherine Brahma

MRI machines are devices that are used to non-invasively obtain images of the internal anatomy and physiological processes of the human body. It is safe to use as the patient is not exposed to any harmful radiation, and there are no known side effects. But such machines that are commercially available are very expensive. Due to this reason, it eludes access to a large portion of the population, particularly in developing countries.

This thesis investigates an inexpensive MRI machine that is based on a rotating inhomogeneous magnetic field map. Unlike conventional scanners, because of the rotating field, the signal model of this device has to account for it. The objective of this work is to examine the aforementioned model, and also to implement Krylov subspace-based reconstruction algorithms available in the IRTools package.

Additional information ...

Signal Processing Seminar

Delamination monitoring in composites with fibre Bragg grating sensors

Aydin Rajabzadeh

Additional information ...

MSc ES Thesis Presentation

N-Shot Training Methodology

Ninad Joshi

Traditional Artificial Neural Networks(ANNs)like CNNs have shown tremendous opportunities in various domains like autonomous cars, disease diagnosis, etc. Proven learning algorithms like backpropagation help ANNs in achieving higher accuracy. But there is a serious challenge with the increasing popularity of traditional ANNs is of energy consumption and computational complexity.

Spiking Neural Networks (SNNs) are considered to be next-generation neural networks that are capable of doing complex deep learning applications at fraction of energy that is needed in current deep learning applications because of its similarity to biological neurons. However, SNN is still not able to match the classification accuracy of ANNs which poses a big challenge for wide acceptance of SNN in various applications as traditional learning methods like backpropagation are not possible in SNN.

During training of a neural network the weight matrix is of the highest importance as it eventually decides the trajectory of learning. Currently, one existing solution is to just manually convert ANNs into SNNs to get weight matrix which doesnot focus on getting weight matrix from a small dataset and doesn’t consider spiking neuron parameters.

We aim to address this challenge by proposing a novel N-shot training methodology that is capable of providing a weight matrix for SNN and can give sufficient classification accuracy. The methodology not only provides the weight matrix but can do training with a very small dataset(up to 1 image per class) and still can give considerably higher accuracy. For a reduced MNIST dataset, the method can give an accuracy of 71.68% 10 images per class.

Additional information ...

MSc ME Thesis Presentation

A PLL-based eddy current displacement sensor for button applications

Matheus Ferreira Pimenta

This thesis presents an eddy current sensor (ECS) for button readout applications. The interface embeds the coil sensor in a digitally controlled oscillator (DCO) and uses a highly digital phase locked loop (PLL) to convert the displacement information into a digital output.
The sensor achieves more than 12bit effective resolution, which translates into an equivalent displacement resolution in excess of 10nm RMS. The interface consumes less than 235µA from a 1.8V supply, resulting in a very power efficient architecture.

MSc SS Thesis Presentation

Indoor localization using narrowband radios and switched antennas in indoor environment

Ye Cui

In this thesis, we explore the potential of indoor localization using Bluetooth narrowband radios. To start with, a data model according to the property of the conducted measurement data is developed. The conducted measurement data is radio channel measure- ments based on channel sounding technique. Then the data model is developed as a channel impulse response model and multipath signals are indicated by different time delays.

Delays are estimated after subspace estimation of the data covariance matrix. Smoothing techniques are employed to improve the covariance matrix estimate. To detect the rank of the subspace, two techniques are investigated, namely the MDL algorithm and the threshold method. New estimates for the thresholds are derived, valid for Hankel-structured data matrices. Experiments are conducted to investigate the performance and reliability of those two techniques, under different parameter values.

Next, we consider subspace-based super-resolution algorithm, in particular the MUSIC algorithm. The functionality of the MUSIC algorithm on narrowband radios measurements is tested and evaluated firstly by simulation experiments, which demonstrate the practicability of applying MUSIC algorithm on narrowband radios measurements. Then experiments are extended to the measurement data that conducted from real indoor environments, for the purpose of indoor localization realization using narrowband radios.

MSc SS Thesis Presentation

Radio astronomy image formation using Bayesian learning techniques

Yajie Tang

Radio astronomy image formation can be treated as a linear inverse problem. However, due to physical limitations, this inverse problem is ill-posed. To overcome the ill-posedness, side information should be involved. Based on the sparsity assumption of the sky image, we consider L1-regularization. We formulate the image formation problem as a L1-regularized weighted least square (WLS) problem and associate each variable with one regularization parameter. We use Bayesian learning to learn the regularization parameters from data by maximizing the posterior density. With the iterative update of the regularization parameters, the solution is updated until convergence of the regularization parameters. We involve a stopping rule based on the noise level to improve the computational eachciency and control the sparsity of the solution. We compare the performance of this Bayesian learning method with other existing imaging methods by simulations. Finally, we propose some future research directions in improving the performance of this Bayesian learning method.

Signal Processing Seminar

Analytical Full-Wave Free Induction Decay Signal Model for MRI

Patrick Fuchs

The derivation of the standard signal model in Magnetic Resonance Imaging (MRI) is based on a quasi-static electromagnetic field approximation and is essentially obtained through an application of the Biot-Savart law. Such an approach works fine for relatively low MR background fields (up till 1.5 T, say), but the model may lose its validity at higher static background fields, since the oscillation frequency of the electromagnetic radio-frequency fields is linearly related to the magnitude of this background field via the well-known Larmor equation. Consequently, an increase in the strength of the static background field leads to an increased Larmor frequency and the quasi-static field approximation may no longer be applicable.

In this presentation, We derive a signal model based on the full Maxwell system and no quasi-static field approximations are applied. We show that the measured signal consists of a direct term that relates the measured signal to the time-varying magnetization within the sample and a scattering term that is due to the dielectric contrast of the sample with respect to its surroundings (assuming no contrast in the permeability). Similar to the quasi-static case, the contribution of each term to the measured signal can be expressed in terms of so-called electric and magnetic receive fields, which basically act as frequency-dependent sensitivity functions. Furthermore, the scattering term is expressed in terms of the electric field strength inside the sample but this field is unknown in general. However, if the dielectric constitution of the sample is known then this field can be determined in principle. Additionally, for low-contrast samples, the Born approximation may be applied leading to an explicit analytic full-wave signal model in terms of the magnetization of the sample, its conductivity, and its permittivity. Finally, through simulations, we illustrate the receive field sensitivity functions for different measurement scenarios and show which terms in the signal model provide the largest contribution to the measured signal as a function of receiver location, frequency, and dielectric composition and size of the sample under test.

Additional information ...

Signal Processing Seminar

A Data Scientific Approach to Efficient Submillimeter Astronomical Spectroscopy

Akio Taniguchi
Nagoya University, Japan

Astronomical data have become huge, as a result of recent advances in wide-field and wide-band instruments. To efficiently extract astronomical signals from observations using these instruments, data scientific approaches are essential. In the (sub)millimeter waveband, spectroscopy with ground-based single-dish telescopes is the best method for surveying interstellar molecules and atoms. However, such observations are not efficient yet, because they always suffer from the intense and time-varying atmosphere of the Earth.

In this talk, I present a statistical method to remove the atmospheric emission from a large spectroscopic dataset by using its intrinsic frequency correlation or spectral shape. As an application, I introduce a recent development of frequency modulation (FM) spectroscopy, which is three times more efficient than a conventional method [1]. As a collaboration with TU Delft, I introduce another application of spectral-cleaning for an ultra-wide-band (UWB) spectrometer DESHIMA [2]. Grasping the UWB atmospheric characteristics by using our data analysis software [3], it removes atmospheric effects on an astronomical spectrum much better than a conventional method.

[1] Akio Taniguchi, Yoichi Tamura et al., "A new off-point-less observing method for millimeter and submillimeter spectroscopy with a frequency-modulating local oscillator (FMLO)", submitted to Publications of the Astronomical Society of Japan (2019)
[2] Akira Endo, Kenichi Karatsu, Yoichi Tamura, Tai Oshima, Akio Taniguchi, ..., Jochem J. A. Baselmans, "First light demonstration of the integrated superconducting spectrometer", Nature Astronomy (2019), Advanced Online Publication
[3] Akio Taniguchi, Tsuyoshi Ishida, "De:code - DESHIMA code for data analysis", DOI 10.5281/zenodo.3384216

Additional information ...

Signal Processing Seminar

Local Activation Time estimation in Atrial Electrograms

Bahareh Abdikivanani

The interpretation of unipolar electrograms is complicated by interference from nonlocal activities of neighbouring tissue. This happens due to the spatial blurring that is inherent to electrogram recordings. In this study, we aim to exploit the high-resolution multi-electrode recordings during atrial mapping to amplify local activities and suppress non-local activities in each of the electrograms. This will subsequently improve the annotation of local deflections and local activation times (LATs) of the electrograms.

According to electrophysiological models, electrogram array can be modelled as a spatial convolution of per cell transmembrane currents with an appropriate distance kernel, which depends on cells’ distances to the electrodes. By deconvolving the effect of the distance kernel from the electrogram array, we undo the blurring and estimate the underlying transmembrane currents as our desired local activities. However, deconvolution problems are typically highly ill-posed and result in unstable solutions.

To overcome this issue, we propose to use a regularization term that exploits the sparsity of the first-order time derivative of the electrograms. We also discuss, in summary, the required electrode array specifications including the spatial resolution and electrode diameter for an appropriate electrogram array recording and subsequent deconvolution.

We perform experiments on simulated two-dimensional tissues, as well as clinically recorded electrograms during paroxysmal atrial fibrillation. The results show that the proposed approach for deconvolution can efficiently amplify the local deflection in fractionated electrograms and attenuate nonlocal activities. This, in turn, improves the annotation of the true LAT in the fractionated electrogram.

Additional information ...