Partially domain-adaptive rolling bearing fault diagnosis based on joint weighting and metric optimization

Abstract Partial-domain adaptive techniques are widely applied in cross-operational bearing fault diagnosis to address inconsistencies between source and target domain fault classes effectively. However, existing studies face challenges in feature alignment, including insufficient alignment of shared class fault features between the source and target domains, interference from outlier class samples in the source domain, and low-confidence pseudo-labels in the target domain. These issues hinder efficient alignment of shared class features, ultimately reducing fault diagnosis performance. This paper proposes a partial-domain deep migration model (JWMDA) that integrates joint weighting and metric optimization. To improve the alignment of shared class fault features, a joint metric combining covariance alignment (CORAL) and local maximum mean difference (LMMD) is developed. This metric complements adversarial training, reduces distributional differences between domains, and optimizes feature alignment. To address interference from outlier class samples in the source domain, class-level weights are employed to effectively mitigate the negative migration effects of these samples. Additionally, sample-level weights are introduced to reduce the negative migration effects of low-confidence pseudo-labels and boundary samples, enhancing the accuracy and robustness of shared class feature alignment. The proposed method is validated through experiments on both public and self-constructed bearing datasets. Experimental results demonstrate that the proposed method achieves higher diagnostic accuracy than existing partially domain-adaptive methods in cross-operational diagnostic tasks involving similar and different equipment.