Statistical identification guided open-set domain adaptation in fault diagnosis

As a critical module of prognostics and health management, fault diagnosis is important to enhance the reliability and safety of the machinery equipment. To improve the fault diagnosis performance in real applications, this paper focuses on the open-set domain adaptation (ODA) task, where the distribution discrepancy exists between the source and target domains, and both source and target label sets contain private classes not shared by the other domain. Previous methods suffer two shortcomings. First, existing weight criteria for feature alignment are mostly constructed by overconfident network predictions, which may be not reliable enough for unknown-class identification. Second, the threshold for unknown-class identification needs to be set manually. For this purpose, this paper proposes an extreme value theory (EVT) guided progressive adaptation method. EVT model is established to generate the open-set probability of target samples belonging to unknown classes, and then the open-set probability is exploited to down-weigh unknown-class target samples in domain adaptation. Moreover, target samples with highest open-set probability are used for training an extended label classifier to identify unknown-class samples, thereby no threshold parameter is required during the testing phase. Experimental results demonstrate that the proposed method outperforms state-of-the-art DA methods.