Densely connected residual networks for cross-domain bearing fault diagnosis with insufficiently labeled samples

Abstract

In cross-domain bearing fault diagnosis, the lack of labeled samples in the target domain requires models to transfer knowledge from a labeled source domain. However, differences in operating conditions lead to domain drift between domains, and conventional global alignment methods may further introduce inter-class confusion, degrading diagnostic performance. In addition, in complex and variable application scenarios, it is difficult to ensure the availability of sufficient samples for training deep learning diagnostic models. To overcome these challenges, this paper proposes a residual network with a domain adaptation mechanism. First, the local maximum mean discrepancy (LMMD) is adopted as a metric to describe the feature variations between the source and target domains. Through feature distribution alignment of the same fault type in the relevant subdomains, the offset of the local domain is derived for the cross-domain cases. By introducing LMMD into a so-called domain adaptation layer, the unexpected feature difference for the same bearing fault type can be minimized in a high-dimensional feature space. Second, the residual network with dense connections is used as the backbone to represent the local and global features of the bearing faults completely. This network takes the RGB image transformed from the time-sequence signal as the input and strengthens global feature extraction through dense parallel connections. Finally, the effectiveness of the proposed method was validated on the Case Western Reserve University (CWRU) dataset and the Paderborn dataset. Experimental results have shown that the proposed network achieves high prediction accuracy in cross-domain bearing fault diagnosis.