GraphFL: Graph Federated Learning for Fault Localization of Multirailway High-Speed Train Suspension Systems

High-speed trains (HSTs) have undergone significant development, leading to a substantial expansion of operational railways. Meanwhile, the adoption of large-scale sensor networks for monitoring trains and diagnostic models customized for railways amplify the complexity and the costs associated with maintenance. In this article, a novel graph federated learning (FL) framework, named GraphFL, is proposed for fault localization of multirailway HST suspension systems. More clearly, the multisensor network is first mapped into a graph with identical nodes and edge features, which is further optimized through graph spectral filtering (GSF) and node-level attention (NLA) mechanisms. Then, the fault components of the suspension system within a single railway are located via graph-level topology adaptive (GTA) technology and a postclassifier. In addition, hierarchical compression and distribution (HCD) is conducted for the parameters of each single railway model to reduce the communication cost during the multirailway FL process. Finally, the personalized model for each railway is tailored through railway-specific optimization. Focusing on the HST suspension systems across four railways, the experimental results show that the proposed GraphFL achieves superior accuracies in fault component location (94.05%) and health status identification (99.33%).