Predictive framework for damage severity in ship collisions involving fishing vessels in Korean waters: A data-driven Bayesian network approach

초록

Ship collision accidents pose significant maritime safety and economic risks. This study developed a data-driven Bayesian network framework to estimate vessel-level serious hull-damage probabilities in collisions involving fishing vessels in Korean waters. Six representative Bayesian network structures were compared under identical collision-event-level grouped validation, and a Bayesian information criterion (BIC)-optimized general Bayesian network (GBN) was selected. The selected GBN achieved a mean serious-class F1-score of 0.5651 and ROC-AUC of 0.7770. Mutual-information sensitivity analysis identified collision-induced change in velocity (Delta-V), striking/struck status, collision point, speed, and hull material as the five most informative variables. In two reconstructed collision applications, reduced-speed scenarios moved both fishing vessels into the lowest Delta-V state and reduced inferred serious-damage probability to approximately 0.107. The framework supports interpretable severity assessment, empirical BN structure selection, and scenario comparison under alternative operating conditions.

키워드