Weighted asynchronous subdomain adaptation network for unsupervised fault diagnosis of bearing

Abstract

The research of unsupervised cross-domain fault diagnosis for bearing is of great significance. However, there are still some problems to be solved. For example, a single predictor may not enough to acquire accurate pseudo-labels in target domain. In addition, global feature alignment may cause different subdomains of source and target domains to be too close. Finally, due to the inaccuracy of pseudo-labels, the effect of local subdomain feature alignment is also very limited. To this end, this paper proposes a weighted asynchronous subdomain adaptation network. First, according to the domain adaptation loss, a weighted integrated adaptation mechanism is constructed to get more accurate target pseudo-labels. Secondly, an asynchronous joint distribution alignment strategy is utilized to address the above mentioned problem caused by considering global alignment and local alignment separately. The proposed network is applied to perform various unsupervised cross-domain fault diagnosis tasks, and the experiment results indicate its superior diagnostic performance.