Domain adaptive bearing fault identification based on multiscale 

residual network

ZHAO Zhihong1,2, SUN Meiling1, DOU Guangjian1

(1.School of Information Science and Technology, Shijiazhuang Tiedao University, Shijiazhuang 050043, China; 2.State Key Lab of 

Mechanical Behavior and System Safety of Traffic Engineering Structures, Shijiazhuang Tiedao University, Shijiazhuang 050043, China)

Abstract: Aiming at the problem that the distribution of the original vibration signals of bearings under different working conditions was quite different and the fault feature extraction was insufficient, which led to the low accuracy of model diagnosis,a log correlation alignment (logCORAL) domain adaptive bearing fault diagnosis method based on multi-scale residual network(logCORAL-MsRN)was proposed. First of all, the original vibration signal of the bearing was pretreated and converted into a two-dimensional gray image. Then, the network structure of residual neural network ResNet50 was improved by using multi-scale residual blocks and dilated convolution,and a multi-scale residual network(MsRN)was designed to fully extract bearing fault feature, and avoid the gradient disappearance problem of deep network structure;a log correlation alignment (logCORAL) domain adaptation method was proposed to better align the distribution between domains. Finally, cross entropy loss and logCORAL loss were used as objective optimization functions to train the model, and comparative experiments and ablation experiments under variable working conditions were carried out on the open data set of Case Western Reserve University (CWRU). The research results show that the average accuracy of the logCORAL-MsRN method for bearing fault diagnosis under variable operating conditions is as high as 96.53%, and it is superior to other comparative methods. Namely, the feature extraction network MsRN can extract richer bearing fault information at different scales, and the domain adaptation method logCORAL can effectively align the feature distribution between the source and target domains, verifying the effectiveness and superiority of this method.