Open-Set Bearing Fault Diagnosis Under Variable Operating Conditions Using Mahalanobis-Weighted Adversarial Adaptation

Rolling bearings are key components in electric motor-driven energy equipment, such as motors, pumps, compressors, fans, and generator systems. Their unexpected failures may reduce energy conversion efficiency, increase maintenance costs, and threaten the operational reliability of energy systems. However, in practical condition monitoring, electric machinery usually operates under variable working conditions, and unknown fault modes that are absent from the source domain may appear in the target domain, which leads to negative transfer and unreliable diagnostic decisions for conventional unsupervised domain adaptation methods. To address this problem, this paper proposes an open-set bearing fault diagnosis method based on a Weighted Adversarial Adaptive Network (WAAN). Specifically, a Mahalanobis distance-based sample weighting strategy is introduced to measure the deviation between target-domain samples and source-domain known-class distributions, thereby suppressing the forced alignment of potential unknown samples during adversarial adaptation. In addition, a sliding-window quantile-based dynamic threshold mechanism is developed to adaptively identify unknown fault categories without relying on manually fixed thresholds. Experiments on the CWRU bearing dataset and a laboratory motor-bearing platform demonstrate that WAAN achieves a known-class classification accuracy of 99.72% and an unknown-class detection rate of 76.25% across six open-set transfer tasks. These results indicate the potential of the proposed method for open-set bearing fault diagnosis under the investigated operating conditions, while its generalization to other machinery and more challenging environments requires further validation.