Federated Dynamic Graph Fusion Framework for Remaining Useful Life Prediction

Jun Wang, Qi Wang, Mingyi Li, Xiao Zhang, Dongxiao Yu

Abstract: In real scenarios, due to the limitation of data silos, it is difficult to perform collaborative training with sparse and heterogeneous data distributed in different factories. Besides, the complex operating conditions of equipment result in dynamic interactions among sensors, which diminishes the effectiveness of local models, and different sensor types and numbers between clients also hinder collaborative training. These problems are particularly significant in the remaining useful life (RUL) prediction task, which is essential for modern engineering to maintain high reliability. To overcome these challenges, we propose a Federated Dynamic Correlation-Aware (FDCA) framework, which generates global graph structure knowledge to enhance local model training and utilizes dynamic graphs to incorporate spatio-temporal features on the client side to improve prediction accuracy and robustness, maximizing the synergy between local data training and global knowledge inspiration. Extensive experiments conducted on four public sub-datasets show the advantages of our proposed FDCA over state-of-the-art baselines.