Cross-Attention Transformer-Based Domain Adaptation: A Novel Method for Fault Diagnosis of Rotating Machinery With High Generalizability and Alignment Capability

Transformer has demonstrated strong performance in processing sequential and multimodal signals, making the Transformer-based model a promising alternative to the convolutional neural network (CNN)-based model for developing a general fault diagnosis method. However, significant distribution differences between the source domain (S Domain) and the target domain (T Domain) often result in poor transfer performance in fault diagnosis. To overcome these limitations, a novel cross-attention transformer-based domain adaptation (CATDA) method is proposed. CATDA uses a two-layer Transformer architecture combined with a new source–target domain (S–T Domain) yielded by the cross-attention mechanism, which enhances the model’s capability to process global information and focus on the T Domain. In addition, the yield harmony maximum mean discrepancy (YHMMD) strategy effectively reduces the distributional discrepancy between the S Domain and T Domain, simplifying the alignment process by pulling in the mapping distance between the S Domain and S–T Domain first and improving diagnostic accuracy and transfer performance effectively. Experimental results on various rotating machinery datasets demonstrate that CATDA achieves superior accuracy in different diagnostic tasks, confirming its effectiveness and generalizability under complex working conditions. Code and models are available at https://github.com/CCSLab425/CATDA-Project.