Accident Analysis for Port Terminals using Heterogeneous XAI-informed Machine Learning Approach

Abstract

Numerous trucks and container cargo are shifting in port areas, and special attention is required for safety management to prevent accidents. Particularly, in ‘port terminal’, the severity of the accident can be high in terms of human damage. Previous studies attempt to predict and detect port accidents, however, research on deriving major factors influencing port-accident severity and developing accident severity prediction models in port areas are rare. Therefore, to prepare safety management strategies, it is necessary to derive risk factors that affect the accident severity in each port area. This study developed an accident severity model using both statistical and machine learning techniques, including the Multilevel Bayesian logit model, Support Vector Machine (SVM), and Extreme Gradient Boosting (XGBoost). In data preparation, oversampling and clustering techniques are applied, which can augment data and separate accident risk situations to any characteristics of port accident data. As a consequence, port-accident severity models are developed with heterogeneous accident data. In this process, a severity model suitable for the characteristics of Korean ports was established by considering the heterogeneity of accident data. After developing the severity model, an explainable artificial intelligence (XAI) was used to find key variables affecting port-accident occurrence.