PD Acoustic Signals Classification of GIS Defects Using GST and GoogLeNet Transfer Learning

Jiahe Zhu,Zhetao Zhang,Zhengnan Sun,Hongtao Li,Hanyan Xiao,Qiaogen Zhang

DOI: https://doi.org/10.1109/iceeps62542.2024.10693155

2024-01-01

Abstract:The acoustic signals generated by partial discharge (PD) within GIS contain valuable information regarding the insulation status of the equipment. However, conventional manual extraction of acoustic features is susceptible to subjective uncertainty and often yields poor recognition results. Therefore, this paper proposes a PD defect classification approach based on generalized S transform (GST) and transfer learning using GoogLeNet model. Firstly, the acoustic signals of three typical defects, namely metal tips, floating potentials, and epoxy bubbles, were collected utilizing a built-in optical microphone on GIS experimental platform. Then GST was applied to conduct time-frequency analysis of these signals, generating time-frequency spectra for each sample and thereby constructing a dataset of acoustic signals. Finally, the pre-trained GoogLeNet model was employed for transfer learning on this dataset to implement the recognition of acoustic signals and the classification of PD defects. The results indicate that this method effectively captures the time-frequency characteristics of PD acoustic signals and accurately identifies defect sources, achieving an average identification accuracy of 98.9%. The proposed classification approach for PD defects not only addresses the limitations of conventional manual methods but also offers insights into feature extraction and fault classification for other equipment.