Recognition of Partial Discharge in High-Speed Train Cable Terminals Based on Few-Shot Learning: From PD Signals to PD Time Graph

High-speed train cable terminals are prone to insulation defects under complex operating conditions. However, traditional detection methods are limited by the scarcity of labeled data and the challenges in extracting features from non-stationary signals. To address this issue, this article proposes a novel semi-supervised diagnostic paradigm. First, to account for the time-shift characteristics of partial discharge (PD) signals, an adjacency matrix is constructed using dynamic time warping (DTW), converting PD signals into a spatiotemporal graph representation. Second, a fusion architecture combining a variational graph autoencoder (VGAE) and gated recurrent unit (GRU) is designed, where the GRU captures the dynamic evolution characteristics of PD signals. Furthermore, a K-nearest neighbor (KNN) temporal graph reconstruction strategy is proposed to enhance the model’s robustness. In addition, an auxiliary strategy is introduced, integrating the time-series generative adversarial network (TimeGAN)-based temporal augmentation technique, achieving an accuracy of 90.73% with only 10% labeled data. Finally, interpretability analysis using GNNExplainer verifies the effectiveness of the proposed method in detecting cable terminal insulation defects under conditions of limited data availability.