An Adaptive Multi-Adversarial Transfer Network With a Complementary Transferable Metric Mechanism for Diagnosing Unknown Fault

Bearing fault diagnosis is vital for ensuring the stable operation of machines. Conventional domain adaptation methods for fault diagnosis under variable operational conditions typically presume that the label spaces of the source and target domains (TDs) are identical. Nevertheless, this presumption is challenging to fulfill in real-world applications. Emerging unknown fault classes may appear in the TD. The existing methods have difficulty in suppressing the interference of unknown-class samples in domain adaptation because of the significant domain shifts and label space differences across domains, leading to negative transfer. To tackle the above problems, this study presents an adaptive multi-adversarial transfer network with a complementary transferable metric mechanism. First, the complementary transferable metric mechanism is designed to dynamically assess the similarity between the target samples and the known classes of the source domain (SD), overcoming the drawback of low reliability of a single metric. Second, interdomain selectable feature alignment is achieved by assigning discriminative weights to the target samples during adversarial domain adaptation. Finally, the unknown fault separable adversarial learning module is constructed to build an accurate hyperplane between the known and unknown classes, thereby achieving effective differentiation of fault classes. In addition, the feature extractor with channel attention is constructed to decrease the interference of complex background noise during feature extraction. The experimental results on public and real-world industrial datasets demonstrate that the presented approach can effectively distinguish known health states, identify new unknown fault, and significantly outperform comparison methods in fault diagnosis performance.