Enhanced identification of equipment failures from descriptive accident reports using language generative model

Purpose – Equipment failure is a critical factor in construction accidents, often leading to severe consequences. Therefore, this study addresses two significant gaps in construction safety research: (1) effectively using historical data to investigate equipment failure and (2) understanding the classification of equipment failure according to Occupational Safety and Health Administration (OSHA) standards. Design/methodology/approach – Our research utilized a multi-stage methodology. We curated data from the OSHA database, distinguishing accidents involving equipment failures. Then we developed a framework using generative artificial intelligence (AI) and large language models (LLMs) to minimize manual processing. This framework employed a two-step prompting strategy: (1) classifying narratives that describe equipment failures and (2) analyzing these cases to extract specific failure details (e.g. names, types, categories). To ensure accuracy, we conducted a manual analysis of a subset of reports to establish ground truth and tested two different LLMs within our approach, comparing their performance against this ground truth. Findings – The tested LLMs demonstrated 95% accuracy in determining if narratives describe equipment failures and 73% accuracy in extracting equipment names, enabling automated categorical identifications. These findings highlight LLMs' promising identification accuracy compared to manual methods. Research limitations/implications – The research's focus on equipment data not only validates the research framework but also highlights its potential for broader application across various accident categories beyond construction, extending into any domain with accessible accident narratives. Given that such data are essential for regulatory bodies like OSHA, the framework's adoption could significantly enhance safety analysis and reporting, contributing to more robust safety protocols industry-wide. Practical implications – Using the developed approach, the research enables us to use accident narratives, a reliable source of accident data, in accident analysis. It provides deeper insights than traditional data types, enabling a more detailed understanding of accidents at an unprecedented level. This enhanced understanding can significantly inform and improve worker safety training, education and safety policies, with the potential for broader applications across various safety-critical domains. Originality/value – This research presents a novel approach to analyzing construction accident reports using AI and LLMs, significantly reducing manual processing time while maintaining high accuracy. By identifying equipment failures more efficiently, our work lays the groundwork for developing targeted safety protocols, contributing to overall safety improvements in construction practices and advancing data-driven analysis processes.