Scan-Net: few-shot diagnosis of hydropower auxiliary bearings via Siamese mutual learning

Reliable operation of auxiliary equipment is critical for hydropower stations. However, data-driven diagnosis faces the “cold start” challenge due to fault sample scarcity in high-maintenance environments. We propose a few-shot diagnostic model, Siamese Cross-Attention Network (Scan-Net), combined with a transfer learning strategy. To extract discriminative features from limited data, we utilize Multi-Scale Synchrosqueezed Wavelet Transform (MSWT) for physically consistent time-frequency representations. Unlike simple concatenation, we design a dual-stream Siamese network with a bidirectional Cross-Attention mechanism that enables explicit inter-sensor feature interaction. We introduce a Deep Mutual Learning (DML) strategy with symmetric KL divergence constraints to align prediction distributions between dual branches, serving as self-supervised regularization to prevent overfitting. We establish a transfer pathway from public datasets to field equipment. Experiments show that Scan-Net achieves 96.50 % accuracy on the CWRU dataset under the 10-shot setting, and 94.43 % average accuracy in cross-load transfer. Pilot deployment at a large-scale hydropower station provides preliminary validation, with the system contributing to a reduction in routine inspection workload