A Domain-Specific Conversational AI Framework for Evidence-Grounded Aviation Safety Analysis Using NASA ASRS Data
编号:216
访问权限:仅限参会人
更新:2026-11-15 16:40:58 浏览:12次
Online
摘要
Aviation safety reports contain valuable information about operational events, but extracting relevant evidence from large collections of narratives remains challenging. This paper presents a domain-specific conversational artificial intelligence framework for evidence-grounded analysis of data from the National Aeronautics and Space Administration’s Aviation Safety Reporting System (ASRS). The framework integrates intent-based query routing, bounded multi-step planning, semantic incident retrieval, deterministic statistical tools, evidence verification, incremental data ingestion, report-volume monitoring, and problem–flight-phase co-occurrence analysis. Evaluation using 200 researcher-authored queries across eight intent categories achieved 82.00% classification accuracy and an 81.78% macro F1-score. All 200 evaluated planner executions remained within the six-step limit, with a mean of 4.59 steps. The evidence verifier correctly handled eight controlled test cases. In a different test using 50 queries for known target retrieval, the average reciprocal rank was 0.5941 and the Recall@10 was 70%. In a successful ingestion replay of 50 held-out documents, uniqueness and alignment were maintained. The searchable data repository contains a total of 7,240 documents with their corresponding 384-dimensional embeddings. While these results prove feasibility in terms of implementation, they do not constitute proof of safety efficacy or prediction capabilities.
关键词
Aviation safety, conversational AI, NASA ASRS, semantic retrieval, evidence verification, large language models, safety analytics.
稿件作者
Janani S
Thiagarajar college of engineering
Amirtha V
Thiagarajar college of engineering
Karthiyayini R
Thiagarajar college of engineering
Uma K V
Thiagarajar college of engineering
发表评论