A multi-adversarial open-set domain adaptation network for planetary gearbox fault diagnosis under varying operating conditions
Open-set fault diagnosis of planetary gearboxes under varying operating conditions remains a challenging issue in health monitoring because the target domain may contain fault categories which are unavailable in the labelled source domain. Most existing domain adaptation methods assume that the source and target domains share an identical label space. Under open-set conditions, this assumption may force unknown target samples to be aligned with known source classes, resulting in negative transfer and unreliable diagnostic results. To address this issue, a multi-adversarial open-set domain adaptation network, termed MA-OSDAN, is proposed for planetary gearbox fault diagnosis. The proposed framework jointly addresses transferable feature learning for shared known classes and decision boundary construction for unknown target samples. A global-local distribution alignment strategy is introduced to reduce both global distribution discrepancy and class-level local mismatch between domains. Conditional adversarial adaptation is further incorporated to exploit classifier prediction information during domain alignment, thereby enhancing the discriminability of transferable features. To suppress the interference of potential unknown target samples, an entropy-guided instance weighting mechanism is designed to reduce the contribution of highly uncertain samples during adversarial adaptation. Moreover, an extended classifier with an additional unknown-class output is constructed to strengthen the separation between known and unknown fault categories. Experiments are conducted on the public Case Western Reserve University bearing dataset, the Shandong University of Science and Technology planetary gearbox dataset, and a self-built planetary gearbox dataset. The results demonstrate that MA-OSDAN achieves better performance than representative open-set domain adaptation methods in both known-class recognition and unknown-fault recognition.