Few-Shot Fault Diagnosis of On-Load Tap Changers Based on a Multi-Source Cross-Sensor Prototype Adaptation and Dual-Prototype Fusion Network

An on-load tap changer (OLTC) is a key electromechanical switching device for transformer on-load voltage regulation, and its mechanical condition is closely related to equipment reliability. Vibration signals are sensitive to sensor location, so the same fault may exhibit different responses at different locations, while only a few labeled fault samples are usually available at a new location, making cross-sensor diagnosis difficult. To address this issue, a Multi-Source Cross-Sensor Prototype Adaptation and Dual-Prototype Fusion Network (MPAF-Net) is proposed. It adopts episodic pretraining and, under multi-source conditions, constructs cross-sensor tasks using different source sensor locations, followed by class-prototype adaptation with a small number of labeled target samples. During diagnosis, target embedding and spectral prototypes are constructed separately, and their outputs are fused according to spectral discrimination confidence. The method is validated on three-sensor vibration data from an OLTC mechanical fault experimental platform. Results show average accuracies of 95.40% and 97.86% across six single-source and three multi-source cross-sensor tasks, respectively, with good performance under different target-label sample sizes. The method provides a new approach for OLTC cross-sensor fault diagnosis with limited labeled target samples.