Identification of Electrical Anomalies in Co-Phase Power Supply System

The Co-Phase Power Supply (CPPS) system is prone to IGBT faults and train operation-induced abnormal behavior in CPPS device, which could lead to misdiagnosis and render traditional diagnostic methods ineffective. Moreover, the lack of practical data hinders the application of data-driven approaches. This paper proposes a feature extractor-attribute learner framework for diagnosing CPPS device anomalies and train-grid coupling anomalies (TGCA) in the CPPS system. By simultaneously collecting three electrical signals, the method facilitates a comprehensive analysis of abnormal operation states. The TimesNet-based feature extractor is designed to accurately distinguish the stable periodic features of inverter IGBT faults from the transient or quasi-steady-state features of other CPPS device anomalies. The zero-shot learning method is adopted to address the issue of missing composite operation state samples. Experimental results demonstrate that our method achieves an identification accuracy of 99.08% for seen types and 96.50% for unseen types. The engineering cases confirmed that the method achieves an identification accuracy exceeding 98.8% for seen types and effectively prevents misdiagnosis of inverter anomalies.