Rolling bearing fault diagnosis technology is one of the key techniques for diagnosing and maintaining rotating mechanical equipment. However, issues such as scarce labeled data and heterogeneity of data under different operating conditions make it challenging for traditional rolling bearing fault diagnosis methods to effectively perform cross-condition fault diagnosis tasks while ensuring data privacy. To address this, this paper proposes a multi-scenario rolling bearing fault diagnosis method based on prototype-based federated transfer learning. Firstly, local vibration data from clients are preprocessed using Short-Time Fourier Transform (STFT) to generate time-frequency spectrogram datasets. Secondly, clients initialize local models with their respective datasets and engage in federated learning with a central server for parameter aggregation and updates, utilizing prototypes as global knowledge to refine each client's local training and ultimately produce a shared model. Thirdly, clients employ parameter freezing strategies to locally fine-tune the shared model, correcting attention regions of local model parameters to obtain private models suitable for different operating conditions. Finally, through comparative case studies, the advantages of the proposed fault diagnosis method are demonstrated. Results indicate that while ensuring data privacy, the method enhances the accuracy and adaptability of fault diagnosis models.

Rolling bearing, Fault diagnosis, Data privacy, Federated learning, Transfer learning, Multi-scenario.