Diffusion model-assisted cross-domain fault diagnosis for rotating machinery under limited data

In industrial scenarios, cross-domain fault diagnosis faces the challenge of data scarcity due to the difficulty of data acquisition and the high cost of labeling. To overcome this issue, this paper proposes a diffusion model-assisted data generation method to enhance the model's cross-domain diagnostic capability by generating target domain data. Specifically, this paper establishes a diffusion model-assisted cross-domain fault diagnosis method, where a diffusion model is first constructed to augment the target domain data, and then a deep learning model is jointly trained using source domain data, a small amount of target domain data, and the generated target domain data to learn and transfer diagnostic knowledge. To align the global feature distributions, the maximum mean discrepancy loss is first employed to align the source domain data with both the target domain data and the generated target domain data. Additionally, a cross-domain triplet loss is established to achieve category alignment and separation, ensuring similar categories are aligned while different categories are distinguished. Finally, the deep consistency regularization is designed to enforce consistency across target domain data and its augmented versions, enhancing the model's robustness. Extensive experiments on two rotating machinery systems demonstrate the effectiveness of the proposed method in addressing limited-data cross-domain fault diagnosis, highlighting its potential for practical applications in intelligent health monitoring of rotating machinery.