A GNN-GAN-based fault diagnosis method for partial discharge signals in substation GIS equipment

Abstract

Directly constructing a classification model based on raw partial discharge (PD) signal samples suffers from sample scarcity, resulting in low accuracy in diagnosing PD signal faults in substation GIS equipment. Therefore, this paper studies a PD signal fault diagnosis method for substation GIS equipment based on GNN-GAN. Ultrahigh-frequency sensors were employed to capture partial discharge signals. After wavelet denoising, segmentation, and feature extraction, a graph structure was constructed, followed by GNN-based extraction of 64-dimensional deep features. A GAN was designed to generate synthetic samples for dataset augmentation. A model was developed with GNN features as inputs and a fully connected network as the discriminator, trained using cross-entropy loss and the Adam optimizer, to output the most probable fault type. Experiments show that this method can effectively diagnose PD signal faults in substation GIS equipment. Compared with the comparative method, the diagnostic results have a higher overlap with the true values, and the average recall rate reaches 0.912, far exceeding the highest level of 0.744 of the comparative method, significantly improving the accuracy of PD signal fault diagnosis in substation GIS equipment.