Dynamic Model-Assisted Transfer-Coupled Dictionary Learning Strategy for Bearing Cross-Domain Fault Diagnosis

Bearing cross-domain fault diagnosis (CDFD) remains challenging due to the significant domain discrepancy in different bearing data. A novel transfer-coupled dictionary learning (TCDL) strategy is proposed in this article to tackle the challenge in sparse representation space. First, a bearing vibration model is designed to produce the source domain samples. Second, a new TCDL method is developed, in which a dictionary learning-based maximum mean discrepancy (DLMMD) metric and a diffusion regularization term are designed. These two terms can be seen as the relaxation of correlation function terms in coupled dictionary learning (CDL), enabling it to adapt to transfer diagnosis. DLMMD projects sparse coefficients into specific high-dimensional space using an adaptive kernel bandwidth method, and then the discrepancy between two domains can be reduced by the iteration process. Besides, the sparse features from the two different domains are diffused to both sides using the diffusion regularization term, which further enhances the domain confusion of the model. The effectiveness of TCDL is verified by two cross-domain datasets. The results show that the TCDL strategy attains 98.43% and 97.12% classification accuracies in two cases, respectively, which also outperforms some cutting-edge methods.