Rolling bearing life prediction method based on parameter decoupling personalized federated learning

Abstract In order to address the problems of isolated user data and significant distributional discrepancies across different working conditions, which often lead to lower prediction accuracy, a parameter-decoupled personalized federated learning (FL) method is developed for rolling bearing remaining useful life (RUL) prediction. Firstly, the bearing vibration signals are processed to extract time-domain statistical features, from which the first prediction time and the high-risk degradation time are identified, and multi-level degradation labels are generated. Secondly, the bearing vibration signals are converted into wavelet time–frequency spectrum diagram under different operating conditions, and an improved local SEResNet-ConvLSTM model is constructed for RUL prediction. Subsequently, the prediction model is decoupled into a shared representation layer and a personalized layer. The parameters of the shared representation layer are sent to the server for aggregation, whereas the personalized layer is kept on the local device and updated jointly with the new round parameters of the shared representation layer. Consequently, a personalized FL prediction model is established for different working conditions. Finally, experimental verification is carried out on two datasets. The presented method can achieve bearing RUL prediction across various operating conditions while protecting data privacy. The average prediction score is improved by 0.155 compared with the federated averaging algorithm.