Identification of hydraulic engineering construction accident causes using instruction tuning and structured prompting in large language models

Identifying the causes of accidents in hydraulic engineering construction is crucial for preventing recurrences and enhancing safety management. However, traditional methods relying on expert knowledge are often time-consuming, labor-intensive, and highly subjective. Although existing studies utilize text classification or information extraction, they frequently fail to provide granular cause analysis. Furthermore, general-purpose large language models (LLMs) struggle with domain-specific tasks, often suffering from hallucinations and unreliable outputs due to a lack of specialized knowledge. To address these limitations, this study introduces CausalE-LLM, a domain-adapted large language model designed for hydraulic engineering accident analysis. The model is developed through a systematic framework that integrates the construction of a specialized instruction data set, domain-specific fine-tuning, and structured prompt engineering to enhance causal reasoning capabilities. Comprehensive quantitative evaluations demonstrated that CausalE-LLM significantly outperforms mainstream LLMs and traditional deep learning baselines in terms of generation accuracy and semantic alignment with official investigation reports. Based on this model, an interactive system has been developed to assist managers in identifying accident causes efficiently. This research advances domain-specific accident causal reasoning in safety engineering, and it provides a robust and automated tool for improving the reliability and efficiency of safety management in hydraulic engineering.