A review of graph neural networks for rolling bearing fault diagnosis

Abstract In intelligent manufacturing and industrial operations, reliable condition monitoring of rolling bearings is crucial for sustaining equipment performance, ensuring operational safety, and enabling timely fault prognosis. Under complex operating conditions and non-stationary signal environments, conventional signal processing and deep learning techniques, which typically assume Euclidean spaces, often fail to characterize intricate multisource dependencies and graph topology, limiting advancements in intelligent diagnostics. Owing to their strong capability for modeling data on non-Euclidean domains, graph neural networks (GNNs) provide a graph-based learning framework for rolling bearing fault diagnosis. Addressing the lack of a domain-specific systematic review, this study synthesizes recent advances in GNN-based approaches, classifying existing methods into three modeling pathways, namely temporal graphs, feature graphs, and spatiotemporal graphs, and analyzing their graph construction logic, feature representation strategies, and task adaptability in detail. Using four public datasets and a multi-criteria evaluation, representative methods are systematically compared in terms of classification accuracy, robustness to noise, generalization across operating conditions, and interpretability. Building on a comprehensive synthesis of methodological developments and key challenges, this review establishes a unified modeling framework of practical value and, through integrated multi-metric analyses, proposes targeted research directions to advance theoretical understanding and promote the industrial deployment of GNN-based rolling bearing diagnostic systems.