Classifier Discrepancy Guided Soft-Weight Adaptation Network for Machinery Fault Diagnosis Under Domain and Category Shift

Cross-domain diagnosis approaches based on transfer learning have received considerable attention in the past few years. Conventional closed-set diagnosis tasks, where the fault modes of source and target machines are consistent, are effectively addressed by existing methods. However, the closed-set assumption is difficult to hold in actual applications because different machines usually have different fault modes in real industries. To handle the fault diagnosis problem with category shift, this study proposes a novel classifier discrepancy-guided soft-weight adaptation network (CDSAN). The network incorporates two different state classifiers to extract domain-invariant discriminative features from different domains. An outlier classifier is constructed to perform pseudolabel learning on target samples, effectively separating shared-class and unknown-class samples between domains. On the basis of the entropy of the class likelihood, a sample-wise soft-weight term, which can adaptatively measure the similarity of target samples attributed to shared label space, is constructed. After that, a weighted classifier discrepancy loss is designed to bridge the domain shift and to achieve known-class alignment across domains. The experimental results conducted on two real datasets with diverse diagnostic tasks demonstrate that the proposed network offers a promising solution for fault diagnosis problems in the presence of domain and category shift.