中国安全科学学报 ›› 2024, Vol. 34 ›› Issue (8): 178-185.doi: 10.16265/j.cnki.issn1003-3033.2024.08.1901

• 安全工程技术 • 上一篇    下一篇

ChatSOS:基于大语言模型的安全工程知识问答系统

唐海洋(), 刘振翼, 陈东平, 初庆钊**()   

  1. 北京理工大学 机电学院,北京 100081
  • 收稿日期:2024-02-20 修回日期:2024-05-25 出版日期:2024-08-28
  • 通信作者:
    ** 初庆钊(1992—),男,山东烟台人,博士,副教授,主要从事危险物质安全分析等方面的研究。E-mail:
  • 作者简介:

    唐海洋 (2000—),男,湖南益阳人,硕士研究生,研究方向为大语言模型及相关应用场景等。E-mail:

    刘振翼, 副教授。

    陈东平, 副教授。

  • 基金资助:
    北京理工大学爆炸科学与技术国家重点实验室自主课题(ZDKT21-01); 北京理工大学科技创新项目(2022CX01028)

ChatSOS: large language model-based knowledge Q&A system for safety engineering

TANG Haiyang(), LIU Zhenyi, CHEN Dongping, CHU Qingzhao**()   

  1. School of Mechatronic Engineering, Beijing Institute of Technology, Beijing 100081, China
  • Received:2024-02-20 Revised:2024-05-25 Published:2024-08-28

摘要:

为解决大语言模型在安全工程领域应用时面临的语料库规模、输入处理能力和隐私性限制等问题,以2013—2023年间共117篇爆炸事故调查报告为基础构建向量数据库,利用大语言模型的生成式能力,进行提示工程,提出一个基于大语言模型的安全工程知识问答(Q&A)系统——ChatSOS;与ChatGPT大语言模型相比,ChatSOS能够通过整合外部知识库,使大语言模型根据用户的输入信息,从数据库中检索相关语料,并深入分析。结果表明:ChatSOS具备深入分析问题、自主分配任务的能力,能够详尽总结事故报告并提出建议;通过结合外部知识库解决基础大模型在安全工程领域语料不足和语料实时性不高的问题,避免了使用新数据集微调模型可能导致的模型性能下降等问题,提升了大语言模型在安全工程领域的应用能力。

关键词: ChatSOS, 大语言模型, 安全工程, 知识问答(Q&A)系统, 事故调查, 向量数据库

Abstract:

To address the limitations of large language models in safety engineering, such as the corpus size, input processing capabilities and privacy concerns, ChatSOS, a Q&A system based on large language models, was developed. Based on 117 explosion incident reports from 2013 to 2023, a vector database to enhance the system's capability was constructed. ChatSOS integrated prompt engineering and external knowledge base to retrieve and analyze relevant data from the database. Compared to ChatGPT, ChatSOS integrated the external knowledge base, so that the big language model could retrieve the relevant corpus from the database according to the user's input information and make in-depth analysis. The results show that ChatSOS excels in in-depth professional problem analysis, autonomous task allocation, and providing detailed summaries and recommendations based on incident reports. By combining with the external knowledge database, the limitations of the large language model's professional corpus in safety engineering are overcome, which prevents performance degradation associated with fine-tuning on new datasets, broadens the application of large language models in this field, and paves the way for future advancements in automation and intelligent systems.

Key words: chat safety oracles (ChatSOS), large language model, safety engineering, knowledge question answering (Q&A) system, accident investigation, vector database

中图分类号: