Coupling Deep Models and Extreme Value Theory for Open Set Fault Diagnosis

Existing deep-learning-based fault diagnosis methods assume that all possible fault modes are available during training process, which is sometimes not consistent with real applications. Unknown fault types may occur in the testing phase due to the fact that it is impossible to collect all the fault modes in the training phase. Thus, in this paper, we introduce and define the open set fault diagnosis (OSFD), and handle this problem in both shared-domain and cross-domian scenarios. For shared-domain OSFD, an extreme-value-theory-based method is proposed to build a rejection model to detect samples from the unknown classes. For cross-domain OSFD, weighted domain adversarial neural networks is constructed to obtain domain-invariant features of the shared classes and separate samples of unknown classes by reweighting target samples. Learned features of source data are used to establish a rejection model, such that unknown samples in the target domain can be detected. Experimental results on the Case Western Reserve University dataset demonstrate the effectiveness of the proposed methods.