Abstract: The traditional classifier method has problems of weak generalization ability and low recognition rate when diagnosing transformer faults, so the paper proposes a transformer acoustic feature extraction and fault identification method based on Mel Frequency Cepstral Coefficient(MFCC)and Convolutional Neural Networks(CNN). First, acoustic signals of transformers in different operating states are collected by digital microphones, and after the preprocess their MFCC features are calculated as static features and then further processed to obtain ΔMFCC features as well as ΔΔMFCC features as dynamic features. Second, the convolutional neural network model is introduced as the classifier, and static features and the fused features of the three are used respectively as the data set for training. Third, training results of the two models are analyzed, and the system is validated with experiments in other distribution rooms. The experimental results show that this method can effectively identify the normal working state, the overload state and the discharge fault of transformers based on their working sound. Besides, the introduction of dynamic features can increase the identification of the model to a certain extent.

    Key words: transformer; acoustic signals; fault diagnosis; Mel frequency cepstral coefficient; convolutional neural network; dynamic feature