A methodology of large language model-driven dynamic Bayesian network for risk analysis of aircraft accidents

Aircraft accidents tend to generate tremendous personal casualties and economic losses. Therefore, effective risk analysis is essential to manage and mitigate the catastrophic consequences of aircraft accidents. Nevertheless, insufficient and heterogeneous data pose a challenge for risk analysis. In this paper, we propose a methodology of large language model (LLM)-driven dynamic Bayesian network (DBN) for risk analysis. First, according to the collected aircraft accident cases, LLM is applied to identify risk-influencing factors (RIFs) through the designed prompt words. To improve the retrieval efficiency and accuracy of retrieval results, a hybrid model integrating Best Matching 25 and BAAI general embedding is employed to enhance the capacity of context retrieval and generation. Second, to quantify dependency relationships of RIFs, fuzzy set theory and domain knowledge are utilized to determine their probability distributions. Subsequently, a DBN-based risk assessment model is developed to analyze dynamic causal evolution and predict time-varying failure probability. Eventually, sensitivity analysis is conducted to capture crucial RIFs, and corresponding risk prevention and control strategies for aircraft accidents are constructed. The study results demonstrate that the proposed methodology is helpful to address the data limitations in the field of risk analysis.

Keywords: Risk analysis, large language model, dynamic Bayesian network, aircraft accidents..