Investigation of Compressive Sensing and Machine Learning Techniques for Classification of Incipient Discharges in Transformer Insulation.

Present study deals with the acquisition and analysis of different types of incipient discharges in transformer by adopting Ultra-High Frequency (UHF) technique. The Nyquist rate sampling method generates a large number of samples, making it inefficient for developing an online monitoring system. To reduce this, compressive sensing techniques are employed for signal compression and reconstruction. Various compressive sensing methods, including Convex, Non-Convex, Greedy, and Iterative Thresholding, were compared. Orthogonal Matching Pursuit (OMP) was found to be the optimal algorithm, achieving optimal reconstruction time and error at a compression ratio of 45%. The reconstructed signals were compared with the originals using Fast Fourier Transform (FFT), revealing similarities in dominant frequencies. A Long Short-Term Memory (LSTM) machine learning model was used for signal classification, consistently outperforming other algorithms. This study enhances understanding of incipient/partial discharge detection and classification, highlighting the effectiveness of innovative signal processing and machine learning approaches in power system engineering. [ABSTRACT FROM AUTHOR]