A similarity-weighted adversarial network with auxiliary discriminator for open-set fault diagnosis

Abstract In industrial practice, faults in rotating machinery, particularly bearings, can cause substantial economic losses. To mitigate the negative transfer that arises when diagnosis is performed across operating conditions with mismatched label spaces, which prevents recognition of unknown classes in the target domain, this paper proposes the similarity-weighted domain adversarial neural network (SW-DANN) for open-set fault diagnosis. The method is based on the classic DANN framework and employs class-specific domain discriminators for fine-grained alignment of the shared classes. An auxiliary domain discriminator estimates sample-wise similarity for target instances, and these estimates are normalized and used as weights to modulate the classifier-side objective so that unknown-like target samples contribute less to learning. In addition, adversarial training between the feature extractor and an extended classifier that includes an explicit unknown node yields a sharper decision boundary that separates known and unknown fault modes. Experiments on the Case Western Reserve University (CWRU) dataset and a self-built automated guided vehicle (AGV) motor-bearing dataset show consistently superior recognition accuracy over competitive baselines, with H scores of 91.25% on the CWRU dataset and 73.92% on the AGV dataset. Visualizations of the feature space and confusion matrices indicate that the SW-DANN aligns shared classes while isolating unknowns, improving robustness to label-space mismatch in cross-domain diagnosis. These results support the SW-DANN as an effective and practically deployable approach for open-set cross-domain bearing fault diagnosis.