Study of Discharge Sound Diagnosis Using Convolutional Neural Networks

Abstract

External insulation aging and outside damage can cause discharge fault for electrical equipment. Using sound signals to diagnose a discharge fault emerged as a new method in recent years which has obvious advantages such as noncontact and low cost. To address the problem, the diagnosis of electrical discharge sounds was studied based on Convolutional Neural Networks (CNN). Discharge sound was stimulated by applying high voltage to experimental models. To simulate the working conditions of microphones, discharge sounds and various kinds of ambient noises were mixed and used as algorithm input. Feature extraction was executed by applying Mel Frequency Cepstral Coefficient (MFCC) to the sound signals. Lastly, CNN architecture was constructed and used to classify the sample. The main objective of this study is developing a method to recognize discharge sound accurately by using MFCC and CNN. The experimental results show that CNN can achieve 98.06% precision and 97.91% recall of discharge sound after MFCC and some other preprocessing of sound.