A Fault Diagnosis Method Based on Unsupervised Dynamic Domain Adaptation for Rolling Bearings

Existing unsupervised learning-based methods rely heavily on lots of test bed collection data when constructing source domains, and it is difficult to effectively optimize the alignment of marginal and conditional distributions between domains. This study constructs an unsupervised dynamic domain adaptive network (UDDAN) using simulation data as the source domain. A loss function dynamic adjustment mechanism with an adaptive factor is designed to flexibly balance contribution weights of two distribution alignment in the optimization objective according to the model training state. A simulation-generated bearing fault dataset with complete fault information and sufficient labels is used to establish the source domain, which significantly reduces the consumption of experimental resources. The results show that UDDAN can effectively match the deep feature distributions of simulated and experimental data, thereby improving the accuracy of unsupervised cross-domain fault diagnosis.