A physics-informed graph attention network with sparse temporal block selection for aircraft hydraulic system fault diagnosis

Abstract

Precise fault diagnosis of aircraft hydraulic systems is critical to ensuring flight safety. However, conventional methods are often topology-agnostic, failing to leverage physical constraints, while standard graph-based models struggle with the computational burden and information dilution inherent in ultra-long time-series. This paper proposes the Topology-Aware Block-Sparse Graph Attention Network (TABS-GATNet), a physics-informed hierarchical framework that incorporates an efficient sparse temporal block selection mechanism. The framework embeds physical knowledge by constructing a spatio-temporal graph that explicitly mirrors the system's component topology and dynamic connections. To address the challenge of long sequences, TABS-GATNet deploys an attention-guided block-sparse selection mechanism. It formulates a novel block-scoring criterion by fusing global and local attention statistics to adaptively identify the most salient temporal blocks. The model then adopts a coarse-to-fine strategy, performing fine-grained feature extraction exclusively on these blocks and integrating the resulting features with the coarse representation through a gated mechanism for final diagnosis. Evaluated on a high-fidelity simulation dataset of an aircraft hydraulic system, the proposed TABS-GATNet achieves 99.01% diagnostic accuracy, demonstrating significant improvements over conventional benchmarks. Furthermore, ablation studies and attention visualizations validate the efficacy of the sparse mechanism and confirm that the model's diagnostic decisions align with the system's physical fault mechanisms, establishing a trustworthy and effective solution.