DASAN-based cross-working-condition fault diagnosis for rolling bearings

Abstract

To address the insufficient discriminability problems of fault features extracted from rolling bearing data under complex operating conditions and low fault diagnosis accuracies caused by the lack of labeled data in the target domain, this study proposes a cross-operating-condition fault diagnosis method based on the Deep Adversarial Subdomain Adaptation Network (DASAN). In this method, the dynamic adaptive feature extraction module is adopted to perceive the context information of input samples and adaptively extract key feature information from samples. The hybrid pooling attention mechanism is utilized to generate complementary attention weights to enhance the fault feature discriminability. Finally, the Local Multi-kernel Maximum Mean Discrepancy (LMK-MMD) domain adaptation function is introduced into the subdomain adaptation module to improve the inter-class discriminability and intra-class compactness of fault categories, thereby solving the aforementioned problems of insufficient fault feature discriminability and the lack of labeled target domain data. The optimal configuration of the DASAN model was determined via parameter optimization experiments. Ablation experiments on the PU and BPS datasets verified the indispensability of each module. Comparative experiments demonstrated that the proposed model achieved an average accuracy exceeding 99% across all evaluated cross-working-condition fault diagnosis transfer tasks, exhibiting excellent diagnostic performance and strong robustness.