Online recognition method of transformer partial discharge based on audio detection

Abstract

This study aims at the problem of low accuracy and poor real-time of transformer partial discharge defection. This paper uses ultrasonic sensors for contactless detection and proposes an Ultra-Lightweight Convolutional Neural Network (UL-CNN). The UL-CNN can extract audio features during partial discharges to achieve the online detection of the transformer. Even in the case of a small number of training samples, the accuracy can reach 98.6%, the online recognition rate is nearly nine times faster than that of MobileNet, and the recognition accuracy and real-time performance are better than those of the classic lightweight network MobileNet.