| Author: | Liu, Qingli |
| Title: | LLM-driven life-cycle accident investigation in general aviation : a multi-stage framework for causation graph construction |
| Advisors: | Li, Fan (AAE) Xu, Gangyan (AAE) |
| Degree: | Ph.D. |
| Year: | 2025 |
| Subject: | Aircraft accidents -- Investigation Private flying Natural language processing (Computer science) Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Aeronautical and Aviation Engineering |
| Pages: | xvi, 178 pages : color illustrations |
| Language: | English |
| Abstract: | General aviation (GA) plays a vital role in the low-altitude economy but continues to suffer from disproportionately high accident rates. Accident investigation, the foundation of aviation safety, remains expert-dependent and slow, with its core challenge lying in reconstructing the causal sequence of accidents rather than identifying a single root cause. Recent advances in large language models (LLMs) offer new opportunities to automate this process through natural language understanding and multi-step reasoning, yet their application is limited by hallucination, unstable reasoning, and poor adaptability across investigation stages. This thesis proposes an LLM-driven life-cycle framework for GA accident investigation, aligned with the preliminary, detailed, and post-investigation stages of real-world practice. To reduce hallucinations and enhance reasoning performance, the framework integrates three complementary methods: (1) domain-informed Chain-of-Thought (HFACS-CoT) prompting for preliminary investigations, guiding LLMs to reliably identify and classify human errors from pilot narratives; (2) a Dynamic Aviation Risk Field–guided workflow (DARF-ACR) for detailed investigations, enabling structured reasoning to reconstruct multi-event causal chains; and (3) random-parameter bivariate probit modeling with heterogeneity in means, coupled with a structured knowledge base and LLM semantic retrieval, for post-investigation analyses, augmenting causal graphs with statistically validated exogenous risk factors. The framework is validated on a self-constructed dataset derived from the U.S. National Transportation Safety Board (NTSB). Results demonstrate that: (1) HFACS-CoT prompting significantly improves the accuracy of human error inference, achieving higher F1 scores than baseline prompts; (2) the DARF-ACR workflow produces more complete and temporally consistent causal graphs, outperforming existing methods in both node identification and edge reconstruction; and (3) the integration of statistically validated risk factors through econometric modeling and LLM retrieval, further verified through a case study, enriches causal graphs with contextually grounded explanations, extending the analysis from how accidents occurred to why they were more likely under specific conditions. Overall, this thesis advances the theory and practice of GA accident investigation by presenting an innovative LLM-based framework for life-cycle causation analysis. It shifts the focus from static root-cause identification to dynamic causal graph construction, offers a paradigm that can accelerate investigations and improve decision support, and demonstrates a transferable methodology for applying trustworthy and explainable AI in other safety-critical domains. |
| Rights: | All rights reserved |
| Access: | open access |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/14332