Cross-domain open-set fault diagnosis using prototype learning and extreme value theory

Abstract

The open-set cross-domain fault diagnosis scenario is similar to the practical industrial application. It assumes that a distribution discrepancy between the training (source) and testing (target) datasets exists, and the testing label set contains the training label set. The developed model transfers knowledge from the source domain to the share-class faults of the target domain and simultaneously rejects unknown fault samples. Previous studies have two shortcomings. First, feature alignment strategies only consider the alignment of the source and target-known distributions but ignore the segregation of the source and target-unknown distributions, resulting in unreasonable boundaries between known and unknown classes. Secondly, the similarity calculation for feature alignment relies on overconfident network predictions, resulting in biased similarities. To address these problems, this study employs two domain discriminators to clarify the boundaries between the known and unknown classes; one for aligning the source and target-known, and the other for segregating the source and target-unknown. In addition, an unbiased similarity calculation strategy using prototype learning and extreme value theory was proposed to quantify the contribution of each target sample to the domain adversarial training loss. Extensive experiments were conducted on a public bearing dataset and a laboratory gas-insulated switchgear vibration signals. The results outperformed those of other reported approaches, indicating that the proposed method is a promising model for practical industrial applications.