Mechanical Fault Diagnosis of Power Transformer by GFCC Time-frequency Map of Acoustic Signal and Convolutional Neural Network

Abstract

To carefully describe the mechanical condition information from transformer acoustic signals and then identify its typical mechanical faults, the combination of gammatone filter cepstral coefficient (GFCC) time-frequency graph of acoustic signals and Convolution Neural Network is proposed in this paper when considered the excellent sound recognition ability of human auditory system. The gammatone filterbank is first used to decompose the acoustic signals frequency of power transformer with the abundant samples of GFCC time-frequency diagram obtained. Then the intrinsic features of transformer acoustic signal GFCC time-frequency diagram samples are extracted by AlexNet convolution neural network, which is used as input of classifier for recognition. The calculated results of acoustic signals under normal and typical mechanical faults of a 10 kV dry-type transformer have shown that the proposed mechanical fault model of transformer has good recognition effect with mechanical fault model of transformer. The accuracy can reach more than 98%. The research results can provide an important basis for sound diagnosis of typical mechanical faults of power transformers.