A rolling bearing fault diagnosis method based on cloud-edge collaboration federated transfer learning

Abstract To address the high cost of cloud communication in federated learning (FL) and the low diagnosis accuracy caused by the limited number of labeled rolling bearing samples with certain specifications, a new method is proposed. It utilizes cloud-edge collaboration federated transfer learning (CECFTL) for fault diagnosis in rolling bearings across different domains. A bearing fault diagnosis framework for multi-users is constructed by combining cloud-edge collaboration and FL. This is aimed at reducing the communication pressure of the cloud. First, the residual network is employed to capture the features of the vibration signal and construct a multi-user local model. Then, the shared model of the source domain built by cloud-edge collaboration is transmitted to the target domain users. An adaptive fine-tuning model transfer method is proposed, utilizing a personalized fine-tuning strategy to dynamically select the number of fine-tuning layers. Finally, the cross-domain fault diagnosis of rolling bearings under varying specifications is achieved while ensuring the privacy of multi-user island data. Through experimental validation, the proposed CECFTL saves an average of 69.1% of the time compared to the traditional FL. It effectively enhances fault diagnosis efficiency while reducing communication overhead and computational pressure in the cloud. Additionally, the accuracy of cross-domain fault diagnosis is improved by an average of 8.2%, demonstrating strong generalizability.