AI4AirSafe: A Quintuple KG–LLM Framework for Causal Reasoning in Aviation Accidents

Civil aviation accident investigations demand the rigorous integration of heterogeneous safety-critical evidence, including narrative reports, flight operational data, maintenance records, and component histories, to reveal the causal mechanisms of safety breakdowns in complex socio-technical systems. Conventional manual and rule-based approaches are often constrained by limited scalability, subjectivity, and insufficient support for systematic safety knowledge reuse under time-critical conditions. This paper presents AI4AirSafe , a hybrid quintuple knowledge graph (KG)–large language model (LLM) framework for interpretable and safety-oriented causal reasoning in aviation accident analysis. The proposed quintuple schema extends traditional subject– predicate–object representations by incorporating explicit event and temporal dimensions, enabling structured modeling of accident evolution, hazard propagation, and safety barrier degradation. Domain safety information— covering unsafe acts, technical failures, organizational factors, environmental conditions, and resilience-related defenses—is embedded into the KG to capture multi-level safety semantics across human, machine, and operational contexts. Through structured subgraph retrieval and LLM-driven reasoning, AI4AirSafe reconstructs chronological causal chains, identifies critical safety weaknesses, and generates transparent explanations to support learning from accidents and safety improvement. Experiments on 63 official aviation accident reports, comprising over 4,000 structured safety knowledge units, experimental results demonstrate that the proposed framework outperforms representative baselines across four methodological paradigms, including transformer-based information extraction models (IE BLs), triple- based knowledge graph reasoning methods (KG BLs), temporal/event-centric knowledge graph models (TKG BLs), and hybrid KG–LLM architectures (KG–LLM BLs). Achieving an entity extraction F1-score of 0.89 and a causal reasoning F1-score of 0.74. The results indicate that integrating explicit safety-aware causal representations with adaptive generative reasoning substantially enhances analytical accuracy, transparency, and efficiency, providing a promising AI-enabled approach for strengthening safety analysis and resilience in civil aviation systems.