A Data-Driven Framework for Systemic Risk Identification in Aviation Systems Using Large Language Models and Graph Learning

With the increasing complexity of modern aviation systems, high non-linearity and dynamic interactions introduce uncertainties that are difficult to enumerate using conventional expert-driven approaches, limiting their ability to capture emergent systemic risks. To address this challenge, this study proposes a data-driven framework that unifies Large Language Models (LLMs), graph learning, and systemtheoretic perspectives to enable end-to-end risk identification from unstructured evidence. DeepSeek-R1671B(Liu et al., 2024) are employed to extract structured event sequences from 22,800 aviation accident reports in the NTSB database, which are aggregated into a multi-layer knowledge graph representing fault evolution. A Graph Convolutional Network (GCN) is then applied to learn topological dependencies and identify high-risk propagation paths. By bridging semantic extraction and structural inference, the proposed framework enables traceable discovery of latent interaction patterns and provides a scalable, data-driven complement to traditional safety analysis methods for systemic risk identification.

Keywords: Large Language Models, Risk Identification, Graph Learning, Aviation Systems.