The construction industry has a higher rate of safety accidents than other industries due to difficulties in safety management due to the complex production structure and the variable and poor working environment resulting from on-site production. In ...
The construction industry has a higher rate of safety accidents than other industries due to difficulties in safety management due to the complex production structure and the variable and poor working environment resulting from on-site production. In particular, the rate of accidental deaths occurring at small and medium-sized construction sites due to causes such as lack of professional manpower accounts for approximately 50% of the total. The government proposed a systematic method to reduce industrial accident deaths, and in particular, risk assessment was introduced as an effort to prevent safety accidents under the Occupational Safety and Health Act, and has been legally mandated with several revisions. However, there are several obstacles to effectively performing risk assessment. To understand this, we reviewed previous research on the effect of risk assessment application and the current government's risk assessment support site. In addition, in order to improve the problems that impede risk assessment, research and review of smart technology-based risk assessment automation was conducted, and although most of it contributed to the dependency on expert supervision and complexity in the process, there is still an additional need for experts to identify hazards and risk factors and establish safety measures. We identified limitations that required help. To solve these limitations, we established a method to apply the GPT model and RAG model, which are LLMs for natural language understanding and generation.
Through previous research, it has been confirmed that RAG can effectively improve problems such as poor quality of domain knowledge-related content generation with LLM models including GPT, illusion of generation results, and limitations in generating up-to-date information. In this study, the RAG model was used in risk assessment. An automation plan was established for construction domain work, identifying hazards and risk factors and establishing safety measures. This process largely consisted of four stages. First, the relevant data needed to identify harmful and risk factors and establish safety measures were collected, and each database was built and stored and managed. Second, a filter and search process was performed to efficiently extract relevant information from the database. Third, GPT is used to create hazards and risk factors based on content extracted from the database. And, based on the safety data retrieved in the previous step, safety measures tailored to each hazard and risk factor are created. Lastly, considering the consistency of result generation and the convenience of application for safety managers, a UI that meets the work classification standards of the field was constructed.
Using the model constructed in this study, the effectiveness and efficiency of this model were verified through evaluation by a safety manager with experience in construction sites using a case study of reinforced concrete construction where accidents frequently occur at construction sites. The results show that the constructed model is more efficient than existing risk assessments and is helpful in identifying hazards and risk factors and establishing safety measures. Although the identification of hazards and risk factors is similar to that of an expert, it was confirmed that there is still a need to improve quality in establishing safety measures. The model proposed in this study has the academic significance of increasing the scalability of future research in related fields by providing a new approach that has not been studied much in the field of construction industry safety management, and the limitation of the lack of professional safety managers at small and medium-sized construction sites hinders effective risk assessment. There is practical significance in improving this.