Unsupervised Contrastive Learning-Based Single Domain Generalization Method for Intelligent Bearing Fault Diagnosis

Abstract

In the field of fault diagnosis (FD), an increasing number of domain generalization (DG) methods are being employed to address domain shift issues. The vast majority of these methods focus on learning domain-invariant features from multiple source domains, with very few considering the more realistic scenario of a single-source domain. Furthermore, there is a lack of work that achieves single-DG (SDG) through unsupervised means. Therefore, in this article, we introduce a data augmentation method for frequency-domain signals called multi-amplitude random spectrum (MARS), which randomly adjusts the amplitude of each point in the spectrum to generate multiple pseudo-target domain samples from a single source domain sample. Then, we combine MARS with unsupervised contrastive learning to bring the pseudo target domain samples closer to the source domain samples in the feature space, which enables generalization to unknown target domains since the pseudo target domain samples contain potentially true target domain samples as much as possible. Unsupervised SDG intelligent FD can thus be achieved. Extensive experiments on three datasets demonstrate effectiveness of the proposed method. The code is available at https://github.com/WuQiangXDU/UCL-SDG.