Exploiting underutilized heterogeneous data: A multi-source domain adaptive siamese network for few-shot bearing...

Industrial bearing fault diagnosis is constrained by the scarcity of labeled data. Although abundant data is available from diverse equipment, its heterogeneity in sampling parameters, operational conditions, and label granularity hinders direct use by existing methods. To bridge this gap, we propose a Multi-source Domain Adaptive Siamese Network (MDASN) to exploit such heterogeneous data effectively. It integrates multi-domain sample alignment for unified preprocessing, a Bidirectional Importance Attention mechanism to dynamically localize discriminative fault characteristics, and a detail division strategy for fine-grained feature learning. Evaluations on five bearing datasets show that MDASN significantly improves diagnostic accuracy, with gains of 26% on laboratory data and 24% on industrial data. Using only one sample per class, it achieves 86.74% accuracy in a challenging wind turbine bearing task, demonstrating strong potential for robust, data-efficient bearing health monitoring.