| Title |
Predicting Construction Safety Accidents and Types Using AutoML-based Binary and Multi-class Classification Models |
| Authors |
전정호(Jeon, JungHo) ; 최수연(Choi, SuYeon) ; 윤성배(Yun, SeongBae) ; 용선진(Yong, SunJin) ; 허영기(Huh, YoungKi) |
| DOI |
https://doi.org/10.5659/JAIK.2024.40.9.247 |
| Keywords |
Construction Safety;Accident Prediction;Machine Learning;Binary Classification;Multi-class Classification;AutoML |
| Abstract |
In 2022, the construction industry accounted for nearly half of all fatal accidents across sectors in South Korea. This study aims to develop
binary like fatality and injury and multi-class such as fall and struck-by classification models using AutoML to predict the occurrence and
types of construction accidents, based on data from the Construction Safety Integrated Management System (CSI) database. The dataset,
consisting of 235,665 accident cases from January 2019 to February 2024, includes 54 types of information, with 18 influential accident
factors identified. Preprocessed data were trained and tested using AutoML to determine optimal algorithms and influencing factors. Accuracy,
precision, recall, and F1 score metrics were used for validation. The binary classification model for predicting fatalities and injuries,
developed using the Extra Trees (ET) algorithm, achieved the highest accuracy of 95.9% and an F1 score of 0.2771. For predicting accident
types, the multi-class classification model using the LightGBM (LGBM) algorithm recorded the highest accuracy of 57.4% and an F1 score
of 0.5503. Feature importance analysis revealed that the accident object was the most critical factor in both models. This research is expected
to enhance safety management performance by efficiently identifying the likelihood and types of construction accidents. |