A Novel Federated Continual Learning-Based Intelligent Framework for Rolling Bearing Fault Diagnosis

Aiming at the challenges of data privacy protection and catastrophic forgetting under continual learning in fault diagnosis of rolling bearings, this study proposes an intelligent fault diagnosis method based on federated continual learning. First, a federated learning framework for multi-device collaboration is constructed. The improved FedAvg algorithm is used to achieve decentralized model training. A secure aggregation mechanism ensures that the raw data from each client does not need to be uploaded to a central server, thus safeguarding data privacy. Second, to address the dynamic introduction of new fault categories and the forgetting of old knowledge in continual learning scenarios, a dual-loss optimization strategy is designed by integrating the channel attention and spatial attention mechanisms to inhibit the interference of model parameter updates on existing knowledge, at the same time, an optimal model selection mechanism is introduced to retain the optimal model in the global rounds and combine with the continuous backpropagation algorithm to achieve the gradual reinforcement of old knowledge. Further, to adapt to the demand of time-frequency feature extraction of time-sequential vibration signals, the original one-dimensional vibration data are converted to time-frequency images by continuous wavelet transform to improve the model's sensitivity to fault modes. Based on the CWRU bearing fault dataset, the experiment simulates a multi-client decentralized data environment and dynamically expanding fault category scenarios, and the results show that the advantages of privacy protection, dynamic adaptability, and anti-forgetting ability are verified compared with the traditional federated learning and independent continual learning methods.