Kishan Supreet Alguri
Graduate Research Assistant at University of Utah
I am a 4th year PhD student in applied signal processing and machine learning at University of Utah under Dr. Joel Harley. We work on structural health monitoring and complex wave predictions using learning based frameworks (machine learning, dictionary learning and deep learning).
Apart from my research I like compressive sensing, computer vision and Bayesian statistics and their respective applications.
When I am not doing research, I love to drive to new places, participate in trivia, read about world history, play volleyball and build new chess strategies.
SKILLS
Knowledgeable in
Signal, Audio and Image Processing
Machine Learning
Deep Learning
Compressive Sensing
Sparse Coding
Bayesian Statistics
Estimation Detection Theory
Probabilistic Graphical Modelling
Guided Wave Simulations
Programming Languages
C
C++
Python (Pytorch, Tensorflow and Keras)
Matlab
R Programming
EDUCATION
University of Utah
PhD Candidate Electrical and Computer Engineering 2014 to..
Thesis: Predicting complex wave propagation using dictionary and machine learning techniques.
Courses :
Advanced Signal Processing, Advanced Signal Processing II , Adaptive Filters, Random Processes, Digital Image Processing, Computer Architecture, Estimation Detection theory, Probabilistic Modeling, Machine Learning, Computer VisionAcheivements:
-Graduate research scholarship.
-Travel grant from SPIE conf.
Activities:- General Secretary of IEEE signal processing Society 2015-16
- General Secretary of Indian Students Association (ISA) 2016-17
GITAM university
Bachelor's Degree, Electronics and Communications Engineering 2008 - 2012
GPA - 3.85/4.0
Thesis: Performance of OFDM (Orthogonal Frequency Division Multiplexing) in various wireless channels.
Relevant Courses:
Digital Signal Processing
Control Systems
Random Processes
Digital Logic Design
Acheivements:
-Awarded ‘Best Robotic Design’ at IIT Bhuvaneshwar techfest.
-ERICSSON India, Professional Telecom Certified Student
-Nominated for Merit Scholarship.
Activities:-Editor IETE (Institute for Electronic and Telecommunication Engineers)
EXPERIENCE
Mitsubishi Electric Research Laboratories (MERL)
Boston (MA), USA
Research Intern
May 2018 – Aug 2018
o Project: Exploiting synthetic data to learn deep neural networks.
o Developed transfer learning framework for semantic scene segmentation.
o Combined Cycle-GANs and FCNs (fully convolutional networks) with segmentation loss forunpaired domain translation.University of Florida
Gainesville (FL), USA
Visiting Research Student
Jan 2018 – Aug 2019
o Acquiring data using laser Doppler vibrometer to use in my thesis.
o Working on transfer learning techniques to predict guided waves in complex structures.
o Writing my PhD dissertation.
University of Utah
Salt Lake City (UT), USA
Graduate Research Assistant
Aug 2014 – Present
o Developed transfer learning framework in video (wavefield) reconstruction using very few pixel(sensor) values.
o Worked on acoustic source localization using transfer learning models.
o Currently working on physics inspired neural networks for wavefield predictions.
o Used autoencoders, CNNs, dictionary learning and compressive sensing.Los Alamos National Laboratory (LANL)
Los Alamos (NM), USA
Applied Machine Learning Research Intern
Jun 2017 – Aug 2017
Project: Forecasting and anomaly detection of seismic waves using Machine learning techniques
o Used many data cleansing techniques to make an understanding of the data
o Developed a machine learning model using gradient boosted trees for forecasting.
o Developed anomaly detection model using auto-encoders with time-frequency seismic data.Wipro Technologies
Bangalore (KA), India
Project Engineer
Oct 2012 – Jul 2014
o Yeti is an Ink Printer drivers project which is developed in MS UniDriver framework.
o Worked on development and customization of PCL3 Ink printer drivers using C++, Win32, MFC, and DotNet.
o Developed many tools and dialogue based windows applications for increasing the efficiency in the project using MFC and DotNet.Tata Consultancy Services (TCS)
Hyderabad (TS), India
Assistant systems engineer
Jun 2012 – Sep 2012
o Trained and worked on Mainframes.
o Gained extensive knowledge on JCL, CICS and Cobol languages.
Electronic Corporation Of India Limited (ECIL)
Hyderabad (TS), India
Research Intern
Jun 2011 – Aug 2011
o Worked on RADAR detection systems.
o Also learned about acoustic source localization techniques.
Bharat Heavy Electricals Limited (BHEL)
Visakhapatnam (AP), India
Summer Intern
Jun 2010 – Aug 2010
o Worked on Bimetallic strips and Various Electronic devices.
o Also learnt about the manufacturing of Synchronous motor, induction motor, hydro-generators, DC machines.
Projects
Line follower cum Obstacle detector and Object Displacer.
January 2011-March 2011
- Writing code and dumping on to ATMEGA16 processor
- Designing the Robot
- Won the best design award
Performance of OFDM(Orthogonal Frequency Division Multiplexing) in various wireless channels.
January 2012-June 2012
- Simulated and observed the performance of OFDM in various fading channels like AWGN and Rayleigh channels
- Calculated Bit Error Rates for each channel
- Performed analysis and calculated BER using Matlab
Application of Principle Component Analysis (PCA) on a database of gait dynamics in Parkinson's disease
- Application of PCA on gait dynamics database of neurodegenerative diseases.
- We were given a data set of gait dynamic recordings on 3 groups of patients. viz a)parkinson's disease b) Huntington's disease and a group of controlled healthy person.
- Did feature extraction and classification of the diseases with the given datasets with PCA.
Cabinet File Creator (CabApp)- A user friendly interface for Cabinet file creation
March 2014 to May 2014
- Developed a dialogue based windows application which has a very user friendly front end design.
- This application is used to create the cabinet files that are used in installation packages.
- Received awards and appreciations for this particular tool.
Localization of Active Sources in Cylindrical and Hollow Structures.
August 2014 - December 2014
- Developed a model for cylindrical structures, which considers the multipath reflections into account.
- Implemented three localization algorithms,namely
- delay-and-sum technique.
- Time of arrival technique.
- Time difference of arrival technique
- Did an extensive comparison on the results.
IIR Adaptive Line Enhancer
August 2014 - December 2014
- Implemented two different algorithms to simulate adaptive line enhancer. Did an extensive comparison on the results.
- Extended the algorithms to speech processing. Induced a wideband FM noise into a MP3 song and adaptively removed it using Adaptive Line Enhancer.
Rapid Imaging in Cannula Microscope using Orthogonal Matching Pursuit
- Reformulated the microendoscopy image reconstruction linear, inverse problem as a compressive sensing problem in which we applied orthogonal matching pursuit to recover unscrambled microscopy images.
- Achieved up to 70 times improvement in reconstruction time with visually smoother and less noisy image reconstruction compared with the state-of-the-art reconstruction algorithm, the direct binary search.
- In future work, we will further refine the computational cannula microscope with video for real-time imaging.
K-SVD: An Algorithm for Designing Over complete Dictionaries for Sparse Representation
Feb 2015 - May 2015
- Implemented K-SVD algorithm for dictionary learning for single scale images.
- Compared the results of K-SVD adaptive dictionary and DCT dictionary using signal to noise ratio and execution time as metrics.
- Trying to extend the algorithm to multi-scale images.
Acoustic Source localization Using Statistical Array Processing Techniques
- Developed a wave propagation model using Dictionary learning.
- Obtained experimental observations on an aluminum plate.
- Implemented array processing algorithms to detect sources with experimental observations and proposed model.
Damage detection in structures using Machine Learning techniques
- Extracted more than 20 features in 2 different domains to perform SVM.
- Introduced a novel dictionary learning framework to detect damages in structures.
- Did extensive comparison of both the methods. Dictionary learning gave us an accuracy of 98% compared to 77% using SVM.
Android Sensor Based Road Condition Estimation using Machine Learning
- Extracted about 20 features from accelerometer data from android sensor package from my samsung note 5. For data acquisition I mounted my mobile phone and a dash cam on to my car and drove on different road conditions.
- I used unsupervised learning methods with K-Means clustering to classify different types of road conditions using the accelerometer data. I could classify good, bad, very bad road types with an accuracy of about 93%.
- This particular project is self made and I believe has very good application for google maps. Instead of giving the shortest route, your maps could also give you a better quality road route.
Publications
c.1) Fast Imaging in Cannula Microscope using Orthogonal Matching Pursuit
K. Supreet Alguri, Ahmad Zoubi, G. Kim, V. J. Mathews, Rajesh Menon, Joel B. Harley
IEEE Signal Processing 07/14/2015
c.2) Consolidating Guided Wave Simulations with Experiments: A Dictionary Learning Approach
K. Supreet Alguri and Joel B. Harley
SPIE Smart Structures\NDE 03/24/2016
c.3) Guided Wave Reconstruction in Complex Geometries with a Dictionary Learning Framework (Accepted)
K. Supreet Alguri and Joel B. Harley
Acoustic Society of America 05/20/2016
c.4) Merging Models and Data: Predictive Modeling for Guided Waves (Accepted)
Joel B. Harley, K. Supreet Alguri, Alexander Douglas
Acoustic Society of America 05/20/2016
c.5) Robust Baseline Subtraction for Ultrasonic Full Wavefield Analysis
K. Supreet Alguri, Jennifer E. Michaels and Joel B. Harley
Quantitative Non-Destructive Evaluation (QNDE) 07/20/2016
c.6) Model-driven, Wavefield Baseline Subtraction for Damage Visualization using Dictionary Learning
K. Supreet Alguri, Chen Ciang Chia and Joel B. Harley
International Workshop for Structural Health Monitoring (IWSHM) 03/20/2017
c.7) Overcoming Complexities: Damage Detection using Dictionary Learning Framework
K. Supreet Alguri, Joseph Melville, Chris Deemer and Joel B. Harley
Quantitative Non-Destructive Evaluation (QNDE) 07/17/2017
c.8) Structural Damage Detection using Deep Learning of Ultrasonic Guided Waves
Joseph Melville, K. Supreet Alguri, Chris Deemer and Joel B. Harley
Quantitative Non-Destructive Evaluation (QNDE) 07/17/2017
c.9) Learning Data-Driven Models for Ultrasonic Guided Wave Nondestructive Testing
Joel B. Harley, K. Supreet Alguri, and Soroosh Sabeti
European Federation for Non-Destructive Testing (ECNDT) 06/11/2018
c.10) Identifying Bio-Mechanical Wrist Impairments with Machine Learning: A Feasibility Study
Jennifer A. Nichols, K. Supreet Alguri, and Joel B. Harley
American Society of Bio-mechanics (ABS) 08/11/2018
c.11) Transfer Learning of Ultrasonic Guided Waves using Autoencoders: A preliminary study (Accepted)
K. Supreet Alguri, and Joel B. Harley
Quantitative Non-Destructive Evaluation (QNDE) 16/04/2018
J.1) Baseline-Free Guided Wave Damage Detection with Surrogate Data and Dictionary Learning
K. Supreet Alguri, Joseph Melville, and Joel B. Harley
Special Issue: Compressive Sensing in Acoustics, Journal of Acoustic Society of America (JASA) 05/01/2018
Cool things I get to do in my research
Fast imaging using Compressive sensing
Comparison of reconstruction images. Top row (a)-(d) and bottom row (e)-(h) each illustrate different images from the three-bead and the dense cluster. Each column shows the reference images, direct binary search (DBS) reconstruction results, orthogonal matching pursuit(OMP) reconstructions without compression, and OMP reconstructions with compression, from left to right.
Waves in an Aluminum plate
Guided wave propagation in aluminum plates
Guided Wave Reconstructions using Dictionary Learning
Comparison of reconstruction images. First column represents the data we learn and second column white squares represents the data we measure (~0.2% of full data). Third column represents the reconstruction of data from column 2 using just 0.2% of its data with an accuracy of above 90%.
Behind the scenes
Experimental system to obtain beautiful wavefields.
Acoustic Source Localization
Brightest shade of yellow corresponds to the estimated source the black rectangle shows the actual location of the source.
Damage Visualization using Dictionary Learning
(a) Wavefield with damage. The damage signature is not clearly visible. (b) Direct subtraction of baseline wavefield from the wavefield with damage. The damage wavefield is barely visible and incident waves are not totally suppressed. (c) Reconstructed wavefield using dictionary learning framework subtracted from the wavefield with damage. The damage wavefield is easily visible and incident waves are suppressed.
Damage Detection using Deep Neural Networks
This network was trained using data from plate A, using input images of ten randomly sampled signals at a time, in 10,000 training iterations. The input image is a stimulus of ten signals randomly selected from the remaining undamaged versions of plates B, C, and D. The 16 and 32 feature maps are those produced by this stimulus to the trained network.
Visualizing Convolutional Layers for Damage Detection
A visualization of the first convolutional layer for the network trained on plate A, including the 16 filters and the resulting 16 feature maps for an undamaged and damaged input image each. The feature distributions plot the average intensities of the undamaged (red) and damaged (blue) feature maps.
Damage Detection using Dictionary Learning Framework
Figure (a) illustrates the accuracy versus number of measurements for direct baseline comparison, sparse wavenumber analysis and dictionary learning framework. Figure (b) illustrates the gap between the two clusters versus number of measurements. Figure (c) illustrates the average correlation coefficient versus number of measurements
Reconstruction of a Test Structure using Dictionary Learning
Figure shows a time signal from an Aluminum structure, the lighter colored signal is the reconstructed signal using dictionary learning. The dictionary was learnt using data from a steel structure which is very different to the test structure.
CONTACT ME!
LET'S CHAT
Copyright 2015