Dynamic evolution of maritime accidents: Comparative analysis through data-driven Bayesian Networks

Maritime accident research has primarily focused on characteristics and risk analysis, which often overlooks the evolution of the associated risk patterns over time. This study aims to investigate the dynamic changes in maritime accidents from 2012 to 2021 by employing a data-driven Bayesian Network (BN) model and conducting a systematic dynamic pattern comparison. It presents two-stage models for two databases and five models against different timeframes to capture the evolving characteristics of global maritime accidents. Furthermore, within the context of the accident investigation, this study pioneers the analysis of the effectiveness of two network structures, namely a layered BN model and a Tree-Augmented Naive Bayesian (TAN) network, in terms of the accuracy of predicting the accident severity. The key findings regarding the changes in maritime accidents in the past decade include: (1) a significant rise in maritime risks linked to large ships (30.8%), port areas (11.67%), anchoring (11.82%), and manoeuvering operations (3.8%); (2) a connection between poor anchoring practices on fishing boats and ‘overboard’ accidents, and between inadequate equipment on tankers or chemical ships and ‘fire/explosion’ accidents; (3) the TAN model's superior performance in forecasting accident severity compared to the layered BN model; and (4) the probability of ‘very serious’ accidents in terms of ship-related factors is 74.7%, which is for the layered BN network, significantly lower than the TAN network's 99.4%. This study reveals shifts in accident patterns over time and underscores the importance of continuous monitoring and analysis for effective safety and risk management.