Fault diagnosis method for different specification rolling bearings based on the method of federated model transfer learning

The issues of data privacy and security have gained attention gradually. With regard to the private vibration data of rolling bearings for each user exist the problems of isolated data island and non-sharing of data among different users. Meanwhile, the distributions of vibration data for different specification rolling bearings are of great difference, and there is lack of labeled data. These reasons lead to low diagnostic accuracy. For the above problems, a fault diagnosis frame for different specification rolling bearings based on federated model transfer learning was proposed. The vibration data of multiple users were processed using short-time Fourier transform to construct time-frequency spectrum datasets. The local model was trained by each user and the model parameters were uploaded to the server. The local model parameters transfer strategy of federated learning was improved by introducing a difference updating and parameter thinning algorithm. By using the federated averaging algorithm, the server aggregated local model parameters and updated the local model, then after iteration, a shared model for transfer learning was obtained. The server used layer-by-layer unfreezing strategy to reserve part of the shared model parameters and sent them to each user, and the users fine-tund the shared model using local data, then a personalized model was got to work for each user. Through the experimental verification, it is shown that the proposed method can realize the fault diagnosis of rolling bearings of different specifications in the case of isolated data island and lack of labels, it also has high accuracy and good generalization performance. © 2023 Chinese Vibration Engineering Society. All rights reserved.