Bearing Remaining Useful Life Prediction Using Federated Learning With Taylor-Expansion Network Pruning

Accurate prediction of bearing remaining useful life (RUL) is essential for machine health management. In existing data-driven prognostic methods, centralized data resources and deep neural networks (DNNs) are two requisites. However, conventional data aggregation may result in the privacy disclosure of equipment. Meanwhile, most DNN-based RUL predictors are overparameterized, making them hard to be deployed on edge devices. To deal with these shortcomings, a new bearing RUL prediction method based on the federated learning (FL) and Taylor-expansion network pruning, namely, RUL-FLTNP, is proposed in this article. In this method, a central server and multiple clients work together to train an RUL predictor. First, a multiscale convolutional neural network with a longish full connection in the first layer (LFMCNN) is designed as the RUL predictor. Specifically, LFMCNN contains three units, multiscale feature augmentation module (MFAM), deep feature extraction module (DFEM), and prediction module (PM), where the MFAM is used to extend shallow features from the data stored in each client. Next, the text Taylor-expansion pruning criterion is applied to the DFEM to delete unimportant network nodes, after which each client utilizes its local data to recover the pruned model. The server aggregates all rebirth models to a new global model using the federated averaging (FedAvg) algorithm. Network pruning and rebirth occur alternately in the model training process to produce a compact structure. Experimental results indicate that the proposed method provides a promising solution to prognostic problems in data privacy-preserving scenarios.