Physics-informed explainable graph neural network for trustworthy rolling bearing fault diagnosis

Abstract Robust and trustworthy fault diagnosis of rolling element bearings is paramount for preventing catastrophic failures and ensuring operational safety in complex industrial environments. While data-driven methods, particularly graph neural networks (GNNs), have shown promise, they often neglect the intrinsic physical mechanisms governing developed faults, specifically the deterministic fault-induced harmonics. Furthermore, the ‘black-box’ nature of these models compromises their trustworthiness in safety-critical applications. To address these challenges, this paper proposes a novel dual-path GNN (DP-GNN). We introduce a physics-data dual-driven graph construction strategy that explicitly embeds the harmonic signatures of developed faults as trustworthy physical priors alongside data-driven correlations. The architecture comprises a local path, employing a graph attention network to extract mechanism-consistent features along prior edges, and a global path, utilizing a graph convolutional network with self-attention to capture global topological dependencies. A cross-attention mechanism then adaptively fuses these complementary features. Extensive experiments on the Case Western Reserve University and Southeast University datasets demonstrate that DP-GNN achieves superior diagnostic accuracy and exhibits exceptional robustness against noise compared to state-of-the-art baselines. Moreover, post-hoc analysis using GNNExplainer confirms that the model’s decisions rely heavily on physical harmonic relationships, significantly enhancing its transparency and credibility for industrial deployment.