Scale adaptive subdomain matching network for bearing fault diagnosis

Abstract

The wide application of transfer learning technology can effectively solve the problem of the difference between data collection and actual application equipment of traditional intelligent fault diagnosis methods in the practical application process. However, the difference in subdomain space and the serious imbalance of data samples in the process of simultaneous transfer restricts the deep transfer learning technology to the engineering application of high-precision diagnosis. In order to solve the problem of subdomain matching with different subspaces and unbalanced data samples, in this paper we study the subdomain adaptive method and propose a scale adaptive subdomain matching (SASM) method. The SASM method divides the global feature space according to the sample labels, and features with the same label will be divided into the same sub-feature space. Using the edge distribution of the sample and the category weight of the label, the SASM method can effectively optimize the feature distribution of the same subdomain and the weight distribution of different subdomains. Based on the establishment of a clearer internal structure of features, the field adaptation effect is improved, and the matching ability is enhanced when the sample is unevenly distributed. At the same time, the SASM network (SASMN) method for unsupervised bearing fault diagnosis is constructed and validated by experiments. The results indicate that SASMN can effectively optimize the subdomain adaptive effect, and the diagnostic accuracy of the target domain data set is significantly higher than the other three currently popular domain adaptive fault diagnosis methods.