Empirical verification of a transformer voiceprint fault diagnosis method based on convolutional neutral network-long-short term memory and Mel gammatone cepstral coefficient features

Introduction Transformers are core equipment in power grids. Their malfunctions may cause widespread power outages or even grid paralysis. Accurate diagnosis is of vital importance. Methods Aiming at the problem of insufficient accuracy of traditional voiceprint diagnosis techniques under complex working conditions, this paper proposes a transformer voiceprint fault diagnosis method that integrates CNN and LSTM. Through the series fusion of MFCC and GFCC and Fisher criterion screening, the MGCC characteristic parameters that take into account both accuracy and noise resistance are constructed for model input. Empirical tests were carried out on the voiceprint signals of three types of working conditions: normal transformer, loose winding and loose core. Results The results show that the fault recognition rate of this method for normal working conditions is 88%, the recognition rate for loose winding working conditions is 93%, and the recognition rate for loose core working conditions is 98%. Discussion Studies show that the transformer voiceprint fault diagnosis method based on CNN-LSTM network has high diagnostic accuracy and can meet the requirements of practical applications.