An unsupervised transfer learning approach for rolling bearing fault diagnosis based on dual pseudo-label screening

Abstract

Deep transfer learning is an effective method for unsupervised fault diagnosis of rolling bearings. In some works, the pseudo-label of target domain prediction is used to improve the ability of target domain prediction in transfer learning. However, its validity depends on the quality of pseudo-label generated by the network itself, which is easy to cause the misclassification of the samples. Aiming to this, a dual sample screening (DSS) method based on the information of predicted label changes is proposed in the article, and it is applied to the fault diagnosis of rolling bearings with variable working conditions. DSS combines pre-screening and real-time screening and uses the continuous output of prediction label change information in the training process to improve the network training. It owes to eliminating part of the target domain samples with prediction errors in the stage of network training with pseudo-label. The proposed method improves the stability of the pseudo-label involved in the training and alleviates the negative effects caused by the pseudo-label. The experimental results on Paderborn University dataset show that, compare with the deep transfer learning fault diagnosis method based on pseudo-label cross-entropy, the average diagnostic accuracy of the six transfer tasks using DSS is increased by 5.97%, which effectively improves the fault diagnosis accuracy of rolling bearings.