Research on transfer learning method for fault diagnosis with enhanced cross-domain generalization

With the advancement of intelligent industrial equipment, data-driven fault diagnosis technologies have been widely applied. However, in practical scenarios, the inconsistency of data distribution caused by varying working conditions and equipment differences, along with the scarcity of fault samples, significantly reduces the generalization ability of traditional diagnostic models. Transfer learning provides an effective solution to these problems by transferring knowledge across domains. This paper focuses on two core challenges: poor cross-domain generalization ability and difficulty in fusing multi-source heterogeneous data. A Multi-source Domain Adversarial Adaptation Network (MDAAN) is proposed. This model combines Maximum Mean Discrepancy (MMD) and adversarial training to align the feature distributions of source and target domains, and adaptively integrates classification results from multiple source domains through a weighted ensemble strategy to enhance adaptability to unknown working conditions. Through theoretical analysis and literature validation on two typical cases—wind turbine blade icing diagnosis and cross-device bearing fault diagnosis—the proposed method demonstrates significant improvement in diagnostic accuracy (approximately 30%) and robustness (accuracy >90%). This work provides a theoretical reference and methodological support for the engineering application of transfer learning in mechanical fault diagnosis.