A Study on Classifying Construction Disaster Cases in Report with CNN for Effective Management

This study proposes an efficient management direction for Korean construction disaster cases from KOSHA (Korea Occupational Safety & Health Agency) through a text data classification model based on CNN Algorithm. Five classes were defined to classify construction accidents: fall, electric shock, flying object, collapse, and narrowness. After the initial test, the classification accuracy of fall disasters was relatively high, while other types were mainly classified as fall disasters. These results showed that (1) specific accident-causing behavior, (2) similar sentence structure, and (3) complex accidents corresponding to complex types affect the model accuracy. Two improvement experiments were then conducted: (1) reclassification, and (2) elimination of complex accidents. With complex accidents eliminated, the classification performance improved by 185.7%. This result indicated that the multicollinearity of complex accidents was solved during elimination. In conclusion, this study suggests the necessity to manage complex accidents independently while preparing a system to describe the situation of future accidents in detail.