An unsupervised transfer learning method based on SOCNN and FBNN and its application on bearing fault diagnosis

Abstract

In order to improve the ability of the transfer learning model to diagnose similar fault types of bearings in noisy environments, and reduce the dependence of the model on the labeled fault data, an unsupervised transfer learning method based on structure optimized convolutional neural networks and fast batch nuclear-norm maximization (UTL-SOCNN-FBNM) is designed for solving these problems. The method takes advantage of SOCNN to extract the fault features and classify the fault patterns. Meanwhile, the maximum mean difference (MMD) is designed to improve the distribution similarity between the source domain and the target domain. Furthermore, FBNM is utilized to improve the distinguishability and diversity of the target domain batch output matrix. Finally, this method has been applied into bearing fault diagnosis and compared with some various classic methods under different working conditions, the comparison results show that the proposed method obtains highest diagnostic accuracy in 29 out of 30 experimental groups and the highest mean diagnosis accuracies of 5 noise environments, which can accurately identify the fault types and hazard levels of bearings.