Multiweight Adversarial Open-Set Domain Adaptation Network for Machinery Fault Diagnosis With Unknown Faults

Domain adaptation (DA) methods have proven successful in addressing the domain-shift challenge in rotating machinery fault diagnosis, and the basic tasks that the fault categories of source and target domains are identical have been well achieved. However, machine failures in the industry often unpredictably happen, which gives rise to a more challenging task called cross-domain open-set fault diagnosis (COFD). To tackle this task, a novel multiweight adversarial open-set DA network is proposed in this article. The proposed network uses the adversarial learning strategy to eliminate the marginal distribution discrepancy between source samples and shared-class target samples, thus ensuring that the generalization features across domains are learned. A weighted learning module combining the class-level with domain-level discriminative information is constructed to evaluate the similarity between target samples and the source classes, which adaptively assign larger weights for target shared classes and smaller weights for target private classes. An outlier classifier is established to perform pseudolabel learning on target samples, making the decision boundary between shared and outlier classes robust. Experiments on two cases with several open-set diagnostic tasks demonstrate that the proposed method is a potential tool for detecting new faults in mechanical devices.