Scalable AI Clusters Maritime Trajectories, Detects Anomalies

A new framework, STCAD, has been developed to address the challenge of analyzing massive datasets of maritime vessel trajectories, specifically those derived from Automatic Identification System (AIS) archives. This system is designed for unsupervised clustering and anomaly detection, capable of handling terabyte-scale data. The core of STCAD involves encoding variable-length trajectories using a custom BERT-based model, which is trained through masked token modeling to understand navigational patterns. Following encoding, the system employs CURE hierarchical clustering to group similar trajectories, producing physically interpretable clusters without needing a predefined number of groups. An intrinsic anomaly detection mechanism then identifies irregular navigation patterns by assessing reconstruction loss and assigning "noise" to outliers. The framework's effectiveness was demonstrated on a national-scale AIS dataset, successfully identifying stable trajectory clusters and clearly distinguishing between normal and anomalous vessel activities.