Multivariate anomalous behaviour detection for inland ships via dynamic motion pattern matching

In congested inland waterways, anomalous ship behaviours foreshadow collisions, delays, and regulatory breaches, yet what constitutes ‘normal’ behaviour is dynamically blurred by overlapping and intersecting traffic patterns that vary from bend to bend—rendering static threshold-driven approaches ineffective. We propose a self-learning-based anomaly detection framework that automatically adapts to evolving traffic patterns and incorporates multivariate behaviour interactions into anomaly evaluation. First, a ship trajectory representation learning model extracts latent features from trajectory data, enabling unsupervised discovery of regional motion patterns. Second, a real-time pattern matching mechanism is introduced to support the adaptive pattern monitoring. Third, a pattern-specific Motion–Spatial regression detection (M-SRD) model is proposed that models the motion-spatial parameter correlations and quantifies the anomaly scores of movement behaviours. Validations using Automatic Identification System (AIS) data from Yangtze River demonstrate two key advancements: 1) Identification of seven regional motion patterns, achieving a 32.01% improvement in clustering purity (SC = 0.808) over benchmarks; 2) Detections with high precision in pinpointing anomalous behaviours, with majority detected anomalies occurring within a 0–5 step offset relative to expert-annotated ground truth, and strong detection effectiveness in high-traffic scenarios. The framework provides maritime traffic managers with a robust, data-driven tool for proactive risk assessment and decision support.