Ship risk prediction: A methodological study

Port state control (PSC) is a critical safeguard for maritime safety. The port states are facing heavy inspection workloads, but the limited manpower cannot handle the growing number of inspections. Prioritizing inspections of ships with the highest risk can effectively improve PSC efficiency, making ship risk prediction crucial. The PSC data includes inspection records of ships over time, constitutes an irregular time-series data due to the varying intervals between inspections. While current inspection methods mainly rely on statistics of PSC data, these methods are inadequate in capturing the dynamic variations nature of ship risk. To address this limitation, we propose a Ship Risk Long Short-Term Memory (SR-LSTM) model, which introduces a learnable time gate mechanism specifically designed to model irregular time-series patterns in ship risk prediction. Additionally, the risk score, a fusion label for ship risk assessment, is derived through a reversible weighted sum of two individual PSC labels, overcomes the limitations of using a single label for accurate quantification. In predicting ship risks, the proposed algorithm outperforms traditional methods in both relative and absolute metrics, demonstrated by its application to 81,660 inspection records from the Tokyo Memorandum of Understanding (MoU). This approach adopts a time-series perspective to predict dynamic ship risks, demonstrating the applicability of time-series models in maritime safety.