Intelligent bearing fault diagnosis method based on a domain aligned clustering network

Abstract

In recent years, intelligent fault diagnosis algorithms based on domain adaptation have provided a feasible solution to the problem of diagnosing performance degradation caused by different data distributions and a lack of target labels. However, most of the existing domain adaptation fault diagnosis algorithms are highly dependent on the label space and prior knowledge of the source and target domain, which greatly limits their application in practical scenarios. In this paper, faced with the circumstances that fault information and label space of the target mechanical device are completely unknown, a novel intelligent diagnostic method based on universal domain adaptation—the domain aligned clustering network (DACN)—is proposed. On the one hand, the number of clusters is determined by domain clustering analysis, and the public class and private class samples in both domains are identified. On the other hand, in order to achieve high accuracy of model identification on common class samples, this paper introduces the contrast domain difference and realizes class alignment between different domains by maximizing the inter-class difference and minimizing the intra-class difference. Finally, the effectiveness of the DACN is verified on the bearing datasets from Case Western Reserve University and Paderborn University. A one-dimensional class gradient activation map is calculated to explain the performance of the fault diagnosis model.