Abstract: Knowledge graphs and large language models, as emerging technologies in recent years, have become a major research focus for fault diagnosis of Avionics equipment. Knowledge-graph-based fault diagnosis methods offer strong interpretability, but they mainly focus on entity prediction and relationship completion, exhibiting limited semantic generalization capability and an inability to produce diverse diagnostic results. In contrast, large language models possess strong semantic understanding and generation abilities, yet suffer from limited controllability and a tendency to generate hallucinations, which can reduce diagnostic accuracy. To improve the accuracy of fault diagnosis for Avionics equipment, this paper proposes a fusion-driven fault diagnosis method that integrates knowledge graphs and large language models, centered on the concept of “fault elements.” First, at the data layer, an “expert-guided and large-model-assisted” fault element extraction mechanism is constructed, in which a large language model is employed to clean raw fault data and extract fault elements. Based on the extracted fault elements and various entities, a structured knowledge graph is then constructed and visualized using Neo4j, providing intuitive reasoning paths for subsequent inference. Second, at the reasoning layer, causal chain modeling and multi-hop reasoning are performed on the constructed knowledge graph using the TuckER tensor decomposition model, while a large language model is introduced to conduct enhanced reasoning over semantically relevant subgraphs of the knowledge graph. Finally, at the decision layer, to improve the credibility of diagnostic conclusions, a consistency verification module is designed to cross-validate the reasoning results from the tensor decomposition model and the large language model based on subgraph inference. Diagnostic conclusions are output only when they meet a predefined threshold, thereby balancing accuracy and consistency. Case studies conducted on an aviation aircraft electronics equipment dataset demonstrate that the proposed method outperforms traditional approaches in terms of diagnostic accuracy, information richness, and professional consistency, and exhibits strong robustness and scalability. The results provide a novel technical pathway for constructing highly reliable and trustworthy intelligent diagnostic systems for Avionics equipment.

Key words: Fault diagnosis, Knowledge graphs, Knowledge reasoning, Large language models, Tensor decomposition, Knowledge-augmented inference.