Existing deep transfer learning methods assume that all labeled samples are correctly annotated. However, owing to factors such as human mistakes, measurement deviations, data transmission faults, and storage inaccuracies, it is unrealistic to accurately label all fault samples in actual industrial production. To solve this issue, a reliability confidence transfer learning framework (RCTLF) is proposed for crossdomain intelligent fault diagnosis (IFD) of electric motors in this study. Specifically, a reliability-aware evaluation mechanism is adopted to evaluate the reliability of each source sample. Meanwhile, a confidence estimation mechanism is utilized to assess the confidence level of each fault sample from the source and target domains. At last, a novel loss function is designed to train the constructed diagnostic model. To showcase the effectiveness of the proposed RCTLF in practical diagnostic scenarios, we conducted experiments on two electric motor datasets for IFD with noisy labels. The effectiveness of the proposed RCTLF is highlighted by comparisons with current advanced methods. The experiments confirm that it can still achieve satisfactory results, even when 80% of the source domain data is mislabeled.

Electric Motor,Transfer Learning,Noisy Label,fault diagnosis