A SAGE-GAT Hybrid Network With Dynamic Node Interaction for Fault Diagnosis of Civil Aircraft Landing Gear System

Fault diagnosis of civil aircraft landing gear systems is critically hindered by high-dimensional electromechanical coupling, which leads to confusing faults where distinct failures manifest similar sensor patterns. While graph neural networks (GNNs) can model system topology, standard approaches rely on static or local attention and fail to capture the higher order dynamic interactions essential for distinguishing such faults. We propose a novel SAGE-GAT Hybrid Network enhanced with a dynamic node interaction attention (DNIA) mechanism. This core innovation explicitly integrates domain-informed confusion nodes, allowing the model to dynamically recalibrate attention weights based on fault-specific topological signatures. The method is rigorously validated using a comprehensive evaluation framework comprising a high-fidelity simulation testbed driven by real flight data across multiple runway conditions and two real-world datasets from operational civil aircraft fleets. The proposed network demonstrates superior performance over state-of-the-art (SOTA) baselines, achieving high accuracy in disambiguating confusing faults while maintaining practical training efficiency, thereby offering a robust and applicable solution for aviation maintenance.