Research on ship abnormal behavior recognition based on spatiotemporal dependency

In modern maritime traffic safety supervision and maritime safety science research, recognizing abnormal ship behavior is a critical component. Current methods for identifying such behaviors often neglect the spatiotemporal dependency, making it challenging to handle complex, multidimensional data. In this study, ship trajectories were first meshed using quadtrees and Delaunay triangulation profiles to make the trajectories more centralized. After that, use the sliding window to split the trajectory into multiple graph data objects. Then, graph attention network (GAT) is used to capture spatial dependencies Subsequently, GAT-GRU model is applied to capture temporal dependencies. Ultimately, the total inference score is used to determine whether anomalous behavior occurs, and the average inference score for individual features is used to determine the type of anomalous behavior. This approach integrates spatial and temporal information, assesses the impact of each feature on abnormal behavior, enhances the accuracy of abnormal behavior recognition, and increases the interpretability of anomaly detection, providing more precise and reliable information for maritime supervision.