A Study on the Application of Deep Learning for Prevention of Major Industrial Accidents

Abstract

To prevent casualties caused by violations of safety obligations, the Serious Accidents Punishment Act has been enacted and enforced. Despite continuous efforts by the government and corporations to prevent accidents, industrial accidents have not significantly decreased, and the lack of safety managers and complex industrial site issues reduce the effectiveness of these measures. This study aims to propose a method for improving safety management efficiency and preventing serious accidents (particularly occupational fatal accidents) by analyzing accident images and predicting risk factors and safety measures using deep learning models. Image data is collected and refined to standardize and tokenize textual data related to work content, accident locations, workplace equipment, accident types, risk factors, and safety measures. A pre-trained language model, such as KoBERT, is utilized to embed the text, and a deep learning-based NLP model is designed to predict risk factors and propose safety measures. Data preprocessing, including the removal of irrelevant information and the enhancement of technical terms, is performed. The model is constructed with input data (e.g., work content, accident location, workplace equipment, accident type) and output data (e.g., risk factors, safety measures). Fine- tuning GPT-based models on safety-related data from industrial sites enables simultaneous learning for risk factor prediction and safety measure proposals. The results of the training data are reviewed by professional safety managers to evaluate accuracy and similarity. Based on the designed model, a ChatGPT-based QA system is developed to analyze user-provided text during operations, predict risk factors, and propose appropriate preventive measures for industrial applications. The system continuously incorporates user feedback and new data from industrial sites to improve or expand the deep learning-based NLP model through retraining. This research supports alternative personnel (e.g., project managers, supervisors) in identifying risk factors and implementing safety measures in industrial environments with insufficient safety managers. By enabling the prediction of serious accidents and efficient safety management, this system aims to prevent serious industrial accidents.