Detecting abnormal ship states and joint risky behaviors based on an improved graph attention network

Abstract

The early detection of abnormal ship behaviors is crucial for maintaining maritime traffic order, combating illegal activities, and preventing maritime accidents. Current methods primarily focus on comparing current trajectory with historical patterns or detecting ship trajectory information to identify abnormal behaviors, which struggles to detect real-time ship abnormal states and joint risky behaviors between ships. To address these issues, an improved graph attention network was proposed for analyzing ship trajectory data from the Automatic Identification System. This approach designs two graph-coding structures: one is designed to capture the trajectory features of individual ships, while the other constructs the relationship features between ship trajectories. By enhancing the graph attention mechanism for each structure, the approach effectively identifies abnormal ship states and joint risky behaviors between ships. Experimental results demonstrate that the proposed method achieves high efficiency with detection accuracy of 95.64 % and inference time of 149.5(ms), thereby meeting the practical demands of detecting abnormal ship behaviors.