Generalized open-set domain adaptation in mechanical fault diagnosis using multiple metric weighting learning network.

• A generalized open-set fault diagnosis (OSFD) scenario is defined. • A multiple metric weighting learning network is constructed to address two OSFD issues simultaneously. • Enhanced domain similarity measurement is designed to leverage class label information. • Ensemble uncertainty measurement is developed to reduce misclassification interference. • Twenty-seven OSFD diagnosis tasks are used to evaluate the proposed approach. The problem of practical open-set domain adaptation diagnosis has gained great attention considering unobserved fault categories in target domain. However, existing studies assume that the label space of source domain is a subset of target domain, ignoring that source domain may also contain private fault categories. This generalized open-set diagnosis issue is more challenging, making existing techniques less effective. To tackle this problem, a novel approach is proposed that focuses on addressing two open-set diagnosis issues simultaneously. A multiple metric weighting learning strategy with the integration of an enhanced domain similarity measurement and an ensemble uncertainty measurement is constructed to adaptively weight the importance of samples across domains. Then, weighted adversarial training with multiple metric weight functions is implemented to learn domain-invariant features by performing alignment across different distributions. As such, both unknown and known fault categories can be simultaneously and effectively recognized. Experiments on three bearing datasets are carried out. Results demonstrate the proposed approach can effectively deal with generalized open-set diagnosis tasks, outperforming existing diagnosis approaches. [ABSTRACT FROM AUTHOR]