A multi-evidence based open-set domain generalization method for bearing fault diagnosis

In recent years, open-set fault diagnosis (OSFD) methods have been widely used, but most existing methods do not consider domain distribution differences. In practical applications, operating conditions are always changing, and data from some scenarios cannot be obtained in advance. To address these issues, this paper proposes a multi-evidence based open-set domain generalization method. The extracted features are decomposed into domain-invariant features and domain-related features. Cosine-margin discriminative constraints are imposed on the domain-invariant features to enlarge inter-class separation and reduce intra-class variation. For unknown fault identification, a multi-evidence decision mechanism is constructed using energy score, prototype distance, reconstruction error, and prototype margin. Since these four types of evidence are on different scales, they are first standardized and calibrated to a unified scale, and then fused using a maximization strategy. Experimental results on the CWRU, HUST and PU datasets demonstrate the effectiveness of the proposed method.