Partial Discharge Detection via Self-Supervised Graph Contrastive Clustering

Accurate detection of partial discharge (PD) is critical for ensuring the reliability and safety of high-voltage electrical equipment. This study addresses the challenge of distinguishing PD signals from various sources through unsupervised learning. Acoustic emission sensors were employed to collect PD pulse signals, which were analyzed using a novel cross-domain strategy to extract essential features. In this research, we introduce the self-supervised graph contrastive clustering (SGCC) method, combining graph networks with contrastive learning and residual connections to optimize self-supervised learning. This innovative approach enhances the learning of internode relationships and feature differentiation, effectively minimizing the risk of information homogenization. The temporal dynamic threshold negative sampling method accounts for temporal dynamics and diversity. In addition, we develop a feature contrast function to enhance feature independence and reduce information redundancy in high-dimensional embedding vectors. Clustering of PD pulses is efficiently executed using the Bisecting K-Means algorithm. Our experimental results demonstrate that the proposed features, along with the SGCC method, effectively segregate PD sources, thereby providing substantial support for the safety monitoring of high-voltage systems.