A graph attention network-based learning framework for automatic detection of abnormal vessel behaviors

Abstract

With the rapid expansion of maritime activities, the need to detect abnormal vessel behaviors using advanced data-driven methods has become increasingly critical for ensuring maritime safety and efficiency. Existing approaches often overlook the temporal dependencies and feature correlations in vessel behaviors. This limitation reduces their ability to capture the complexities of maritime operations. To address these challenges, we propose GAT-AD, a novel graph attention network-based framework for anomaly detection in vessel behavior. Our framework incorporates three key components: (1) a graph attention module that combines temporal and feature attention to capture sequential and feature dependencies, (2) an embedding layer to extract latent information from vessel data, enhancing representation learning, and (3) a joint detection module that calculates anomaly scores using both reconstruction-based and prediction-based techniques. We validate the effectiveness of GAT-AD through extensive experiments using real-world AIS data. Ablation studies highlight the contributions of individual components, while comparative experiments confirm that GAT-AD outperforms state-of-the-art baselines. Additionally, case studies highlight the framework’s ability to detect and explain different types of anomalies, further underscoring its practical applicability in real-world maritime scenarios.