Data-driven BN and DBN models prediction performance for ship collision risk assessment

Collision risk assessment plays a critical role in maritime transportation safety. While many studies have used Bayesian network (BN) and its derivatives (e.g., Dynamic Bayesian Networks (DBN)) for ship collision risk analysis, none of them have conducted a comparative study to enable the evaluation of their efficiency and effectiveness. This study aims to evaluate the performance of BN and DBN in collision risk assessment through a comparative analysis. Both models leverage geometric probability-derived parameters, including Distance to Closest Point of Approach (DCPA) and Time to Closest Point of Approach (TCPA), relative distance, relative bearing, and speed ratio, extracted from real-time Automatic Identification System (AIS) data. The BN model performs collision risk assessment based on instantaneous observations at each time slice, offering a snapshot analysis without historical context. In contrast, the DBN model incorporates temporal dependencies, integrating past observations over multiple time intervals, enabling a more holistic risk evaluation. Comparative analyses demonstrate that the DBN model provides more stable and reliable collision risk predictions with smooth transitions, making it particularly suitable for dynamic maritime environments. The study highlights the practical implications of incorporating temporal data into risk assessment models, enhancing maritime situational awareness and improving the reliability of collision avoidance support systems. The findings contribute to the advancement of intelligent risk assessment frameworks, supporting the safe and autonomous navigation of MASS in complex and high-traffic maritime conditions.