Federated adaptive edge node learning method for industrial equipment remaining useful life prediction

In distributed industrial equipment remaining useful life (RUL) prediction tasks, operational data are difficult to centrally share due to constraints such as data privacy protection and distributed deployment, which makes RUL prediction models trained on single-node data challenging to generalize across devices. Federated learning provides an effective solution for distributed RUL prediction by enabling multi-node collaborative modeling without sharing raw data. However, in real industrial settings, limitations in computational capabilities and communication conditions often lead to inconsistent participation levels among nodes in federated training. Moreover, differences in operating conditions and data distributions among nodes commonly exist, which can cause model aggregation bias and thereby degrade prediction performance across different nodes. To address these issues, this paper proposes a federated adaptive edge node learning method (FedAEN) for industrial equipment RUL prediction. The method incorporates coordinated design at both the node level and the model level: (1) At the node level, to handle inconsistent participation in training across nodes, similarity among model updates from different nodes is measured to guide adaptive weighted aggregation of the global model, thereby mitigating interference from low-consistency nodes on aggregation results; (2) At the model level, to address feature representation shift caused by heterogeneous data distributions among nodes, the RUL prediction model is split into a shared representation module and a local adaptive prediction module. During training, the variation of the shared representation module between consecutive federated rounds is constrained to stabilize the feature representation space, while the global model provides knowledge guidance for local training processes to enhance overall generalization performance without compromising node-specific predictive capabilities. Experiments on two typical industrial equipment datasets, namely cutting tools and bearings, demonstrate that the proposed method effectively improves the accuracy and stability of RUL prediction under multi-node distributed settings, which verifies its effectiveness in industrial applications.