Recognition of partial discharge patterns of GIS based on CWGAN-div and Mi-CNN

Abstract

In order to solve the constraints of the limited number and uneven distribution of samples on the performance of the deep learning model in the identification of partial discharge patterns of GIS （gas-insulated switchgear）， a CWGAN-div（conditional Wassertein generative adversarial network-divergence） model is proposed to guide the generation of multi-class partial discharge patterns， which overcomes the instability of the original GAN （generative adversarial network） training， enhances the sample data， and reduces the average imbalance rate from 11.01 to 3.03. Then after using five kinds of classifiers for the comparative experiments before and after sample enhancement， the F1mean value of each classifier has been improved by more than 3.7% after sample enhancement. In the experiment， the Mi-CNN （multi-input-convolutional neural networks） model proposed in this paper can use the PRPD （phase resolved partial discharge） spectrum of ultra-high frequency method and ultrasonic method at the same time， and its final F1mean value reaches 95.8%.

Keywords

• gis （gas-insulated switchgear）
• partial discharge
• gan （generative adversarial network）
• convolution neural network