Engineering material failure analysis report generation based on QWen and Llama2

作者：Sijie Chang, Meng Wan, Jiaxiang Wang, Hao Du, Pufen Zhang, Peng Shi · 发表于：Results in Engineering · 年份：2025 · DOI：10.1016/j.rineng.2025.104532 · 被引用次数：11 · 研究领域：Advanced Decision-Making Techniques、Software Engineering Research、Rough Sets and Fuzzy Logic

• To the best of our knowledge, this paper represents the inaugural effort to use pre-trained fine-tuning and prompt engineering based on LLMs to generate failure cases for specific domain tasks. • Use GPT4-based and manual evaluation-based methods to judge the quality of reports generated by LLMs. • Simultaneously, QWen, which supports Chinese, and Llama2, which does not support Chinese, are employed to generate failure cases. Various questioning methodologies are implemented from the perspectives of linguistic differences and prompt engineering techniques, followed by a comparative analysis of the results. Engineering material failure carries risks to the public and personal safety and causes economic loss. Failure analysis report describes the causes of failure and gives the improvement strategy, but requiring a lot of manpower of experts. The powerful text understanding and generation capabilities of large language models make automatic analysis possible. This paper presents an automatic report generation method using large language models, including training on only a few samples and instruction fine-tuning method (prompt engineering). The experiments are executed on two popular large language model bases: QWen and Llama2, involving 127 failure reports in 9 fields such as petroleum and petrochemical, nuclear and thermal power, ocean engineering, energy engineering, infrastructure, aerospace, transportation, equipment manufacturing, and water conservancy. Using different ...