| MS TSA Meeting

Name of the Speaker: Mr. Shaik Mahammad Afshad (EE19S011)
Guide: Prof. Sarathi R
Online meeting link: https://meet.google.com/gyp-dhof-ngn
Date/Time: 28th March 2024 (Thursday), 11:00 AM
Title: Adoption of Compressive Sensing and Machine Learning Techniques for Classification of Incipient Discharges in Transformer Insulation.

Abstract

The present work investigated the acquisition and analysis of Ultra-High Frequency (UHF) signals emitted by various types of incipient discharge in transformer oil, including corona discharge, surface discharge, particle movement, and void discharge in solid material. Due to the wide frequency range of these signals, sampling the UHF signals according to the Nyquist rate for reconstruction generates a very high number of samples. It is not effective in terms of the resources required for developing an online monitoring system. Hence, it is essential to measure fewer samples from the UHF signals and also to recover them back reliably. Therefore, compressive sensing techniques are employed for signal reconstruction. Different methods of compressive sensing, including Convex, Non-Convex, Greedy, and Iterative Thresholding, were compared based on literature review. Orthogonal Matching Pursuit (OMP) has emerged as the optimal algorithm, achieving optimal reconstruction time and error at a compression ratio of 45%.

Using Fast Fourier Transform (FFT), reconstructed signals were compared with the originals, revealing similarities in the dominant frequencies. Leveraging these features, a Long Short-Term Memory (LSTM) machine learning model was employed for signal classification, outperforming SVM, Random Forest, Decision Tree, and KNN algorithms consistently. This study advances understanding of incipient/partial discharge detection and classification, highlighting the efficacy of innovative signal processing and machine learning approaches to explore complex engineering challenges in power systems.