Leveraging HDBSCAN, LSTM and R-DTW for Proactive Detection and Collision Prediction in Maritime Traffic

Abstract

Detecting anomalies in Automatic Identification System (AIS) data is crucial for marine safety, especially with over 60,000 vessels navigating seaways at any moment. This study proposes an enhanced approach to AIS data analysis for detecting anomalous ship behaviours and predicting collisions in maritime environments. Unlike traditional methods that rely on static threshold-based rules or simpler clustering techniques, our approach integrates advanced machine learning methods like Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) and Long Short-Term Memory (LSTM) networks, along with Rhumb line approach Dynamic Time Warping (R-DTW) for improved trajectory similarity assessment and Closest Point of Approach (CPA) calculations. The study outperforms existing techniques by leveraging HDBSCAN’s ability to handle varying-density trajectory clusters, LSTM’s temporal sequence learning for more accurate movement predictions, and R-DTW’s adaptability in identifying anomalous route deviations. The method includes a robust AIS data preprocessing pipeline, the use of HDBSCAN for dynamically grouping complex maritime trajectories, and LSTM models trained using a sliding window approach to predict future ship movements. CPA computations are employed to assess collision risks based on predicted trajectories. The proposed method significantly enhances anomaly detection accuracy and collision prediction reliability over conventional approaches. This integrated and data-driven approach to anomaly detection and trajectory prediction provides a substantial improvement in maritime traffic management and collision avoidance, contributing to proactive maritime safety measures.