A novel partial discharge pattern recognition for GIS with unbalanced sample based on conditional variational autoencoder

Abstract

Deep learning has been widely used in gas insulated switchgear (GIS) partial discharge (PD) pattern recognition with its powerful ability to automatically extract features. As a typical data-driven model, the diagnostic performance of deep learning methods will decrease with the reduction of the training sample size, resulting in lower recognition rates for minority class samples. In actual situation, the probability of failure of various insulation defects in GIS is not the same, so that the samples of various types of defects obtained in practical applications cannot maintain a balanced distribution. To solve the problem of class imbalance among training samples, this paper proposes an improved variational autoencoder (VAE) to augment the minority class fault samples. The PD signal is first converted into a time-spectrogram with high time-frequency resolution by S transform. Then use conditional AE to perform directed augmentation on unbalanced class samples. Finally, the obtained balanced data set is input into convolutional neural network to complete the fault diagnosis of PD. Experiments show that the proposed method achieves excellent recognition results under various unbalanced conditions, indicating that it has strong tolerance and high generalization ability for unbalanced samples.