SEAuAIS: A self-explainable autoencoder with AIS-based maritime anomaly detection

• A semi-supervised variational autoencoder for vessel route modeling based on AIS data with vessel trajectory reconstruction and vessel type classification. • Detection of anomalous vessel behavior, which generalizes across regions of interest, validated on two use cases: Search and rescue near Bornholm and a cable breach near Svalbard. • Explainable AI through self-learned archetypical vessel behaviors with kinematic constraints, enhancing situational awareness. Maritime anomaly detection systems aim to harness data from the automatic identification system (AIS) to enable early detection and prevention of imminent safety risks. Many existing systems are not interpretable and are also not evaluated in remote and sparsely traversed waters. This makes it challenging for operators to understand, adapt, and apply them effectively. This work aims to address both these gaps in the literature by introducing SEAuAIS : A novel, flexible anomaly detection framework based on the foundation of a self-explainable autoencoder model that incorporates physical constraints. We evaluate the model on two cases: anomalous vessel behavior following a cable breach in the remote waters around Svalbard and a search and rescue (SAR) emergency in a high-traffic zone near Bornholm Island. For both regions, we outperform state-of-the-art methods at significantly reduced compute time.