Show the Way: Accelerating General Aviation Accident Investigations through LLMs and HFACS

Abstract

General Aviation (GA), with the highest accident and fatality rates in civil aviation, undergoes lengthy accident investigations that include site analysis, witness interviews, cause identification, and detailed reporting. These expert-driven processes, often extending for months or years, not only require extensive manpower but also delay vital accident prevention initiatives in GA. The advent of large language models (LLMs), with groundbreaking capabilities in understanding and generating complex text, offers a potential solution to these challenges. This study aims to conduct a General Aviation Accident Cause Automatic Prediction System (GA-ACAPS), which leverages witness narratives (established early in the investigation) through LLMs. The research utilizes 2250 GA accident reports from the National Transportation Safety Board (NTSB), employing the Human Factors Analysis and Classification System (HFACS) for structured accident causation predictions. Three preliminary experiments were conducted to compare the prediction performance of three different prompting methods before the formal experiment. The results from the preliminary experiments underscore that integrating witness narratives with basic accident information significantly boosts the performance of GA-ACAPS. This optimized prompt was thus implemented in the formal study. The formal experiment's findings demonstrate that GA-ACAPS is proficient in predicting unsafe acts and specific preconditions of unsafe acts like the physical environment and personal readiness. This study endorses the potential of GA-ACAPS to serve as a dependable tool for investigators, aiming to narrow down probable causes of accidents and thereby increase the efficiency of investigations. Moreover, the application of LLMs in GA accident analysis heralds a new era of innovative approaches and essential insights, contributing to the advancement of aviation safety.

Keywords: Large Language Models, HFACS, GA Accidents, Automated Accident Investigation System, Prompt.

DOI: 10.54941/ahfe1005197