A Transfer Learning Method Using High Quality Pseudo Labels for Bearing Fault Diagnosis

Many supervised neural network frameworks work well only when the training data and the test data are independent and identically distributed for bearing fault diagnosis. In real industrial applications, the monitoring data follow different distributions owing to the changes of working conditions and data acquisition ways. These frameworks also require numerous labeled data for training, but labeling data are laborious, even labels often do not exist in many complex engineered systems. To address these problems, we proposed a novel transfer learning method that transfers knowledge across different distributed but related domains. The proposed method exploits the capabilities of multiple kernel variant of maximum mean discrepancy (MK-MMD) in measuring the marginal probability distribution discrepancy and pseudo label in calculating conditional probability distribution discrepancy. Considering the interference of pseudo-label noise, we develop an approach to filter out pseudo labels of low quality by an adaptive threshold and a making-decision-twice strategy. The performance of the proposed method is demonstrated with two bearing datasets. The comparison with the fixed threshold shows that the improved pseudo-label learning (IPLL) can resist data imbalance and raise prediction accuracy. The proposed method is validated by predicting the bearing health states of vibration signals under various working conditions and different acquisition ways. The comparative analysis results demonstrate its advantages over other transfer learning methods in terms of prediction accuracy, robustness, and convergence speed.