ACFD: Heterogeneous Graph Neural Networks for Aircraft Cable Faulty Node Diagnosis

Abstract

The reliability of aircraft cables is crucial in aircraft manufacturing and operation, but their complex network structure and numerous test points make fault diagnosis extremely difficult. In aircraft cable networks, true and false faults are often confused with each other. Aircraft cable test equipment can only provide loop results without determining fault types, leading to low efficiency in manual fault detection, while signal reflection methods perform poorly in multi-stage cascade networks. Heterogeneous graph structures possess powerful representational capabilities, yet they are rarely applied to cable fault node classification tasks. In this study, the aircraft cable network is abstracted as a heterogeneous graph, and a novel aircraft cable fault diagnosis model named ACFD is proposed to improve diagnostic accuracy and efficiency. ACFD first employs the gated adaptive node aggregator (GANA) to preselect meta-path neighbors for node-level attention aggregation. Then, the semantic attention aggregator (SAA) and the topological attention aggregator (TAA) are adopted to learn node embeddings for faulty node classification. Experimental results demonstrate that the proposed ACFD outperforms the second-best baseline method by 8.39 and 8.97 percentage points in terms of Macro-F1 and Micro-F1, respectively.