Open-set fault diagnosis of rolling bearings via unknown class perception with a multiple-adversarial network

Abstract Open-set domain adaptation techniques have been widely applied to cross-condition bearing fault diagnosis by mitigating distribution shifts between the source and target domains for known fault categories, thereby improving the alignment of shared classes and enabling the identification of unknown faults in the target domain. However, existing methods still encounter difficulties in both shared class recognition and unknown class detection. These challenges primarily arise from the interference of unknown target samples during domain alignment, which leads to negative transfer, and from insufficient classification capability on target domain data. To address these issues, this study proposes an unknown class-aware multi-adversarial classification network. Specifically, a hybrid adversarial architecture is designed to suppress negative transfer and achieve precise domain alignment. A non-adversarial domain discriminator is employed to assign weights to target domain samples, enabling the model to focus on learning shared class features and reducing the impact of unknown class interference. In addition, a local-global collaborative alignment framework is constructed by combining a sample-weighted adversarial domain discriminator with an unweighted adversarial discriminator that enforces global distribution consistency. This strategy significantly enhances the robustness of cross-domain feature mappings and the transferability of the model. Furthermore, to improve classification performance, a collaborative decision mechanism is developed, consisting of dual shared-class classifiers and an unknown-class classifier. Through logical decision-making, this mechanism improves classification accuracy for known classes while enhancing the detection capability for unknown classes, thereby increasing diagnostic performance under open-set conditions. Experimental results on public and self-constructed bearing datasets demonstrate that the proposed method achieves superior diagnostic accuracy and transfer performance compared to existing open-set fault diagnosis approaches across multiple cross-condition diagnostic tasks.