Identification of Partial Discharge Defects Based on Deep Learning Method

Since repetitive partial discharge (PD) leads to insulation breakdown, it is one of the most critical defects that affect operation life of electrical equipment. In this paper, four kinds of PD defects are identified with deep learning (DL) method according to the current waveforms. A modified IEC-60270 experiment platform with ultra-high frequency (UHF) and current probe is built to acquire PD current waveforms and their corresponding detecting pulse current and UHF pulse signal. Fourier transform, principle component analysis, and t-distributed stochastic neighbor embedding methods are applied to visualize the data set, which proves the feasibility of classifying the PD current waveform. Two basic parts of this DL framework are sparse autoencoder layer and softmax layer, the former extracting features of the input signal and the latter operating as the classifier. Hyper-parameters of the network such as sparsity, activation function, number of hidden nodes, and network depth were discussed. The final classifying accuracy of the proposed method is up to 99.7%, that is much better than the traditional identifying method. A comprehensive blind test is designed to prove the general validity and robustness of the proposed model.