Denoising Universal Domain Adaptation Method for PMSM Fault Diagnosis Under Category Inconsistency

Recently, the cross-domain fault diagnosis method based on domain adaptation (DA) has been developed to solve the domain inconsistency problem caused by the variable working conditions. However, most of the existing methods only consider the case where the fault categories of the source and target domains (TDs) are consistent and depend on the ideal setting that the source domain (SD) samples are correctly labeled. The fault category inconsistency and noisy samples may lead to the degradation of diagnostic performance. To cater to more practical application scenarios, this article proposes a novel denoising universal DA network (DUDAN) for fault diagnosis of permanent magnet synchronous motor (PMSM). First, the output divergence of multiple classifiers is dynamically optimized to filter out noisy samples with incorrect labels in the SD. Then, one-versus-all classifier learning and entropy minimization strategies are introduced to learn the minimum boundary distance between fault categories. Thus, DUDAN can achieve the alignment of samples in the shared label space, while the samples in the private label space are kept separately, and the fault category inconsistency is solved. Extensive experiments in various situations demonstrate the superiority of the proposed method.