Machine Learning for Enhanced Maritime Situation Awareness: Leveraging Historical AIS Data for Ship Trajectory Prediction

Abstract

In this thesis, methods to support high level situation awareness in ship navigators through appropriate automation are investigated. Situation awareness relates to the perception of the environment (level 1), comprehension of the situation (level 2), and projection of future dynamics (level 3). Ship navigators likely conduct mental simulations of future ship traffic (level 3 projections), that facilitate proactive collision avoidance actions. Such actions may include minor speed and/or heading alterations that can prevent future close-encounter situations from arising, enhancing the overall safety of maritime operations.

Currently, there is limited automation support for level 3 projections, where the most common approaches utilize linear predictions based on constant speed and course values. Such approaches, however, are not capable of predicting more complex ship behavior. Ship navigators likely facilitate such predictions by developing models for level 3 situation awareness through experience. It is, therefore, suggested in this thesis to develop methods that emulate the development of high level human situation awareness. This is facilitated by leveraging machine learning, where navigational experience is artificially represented by historical AIS data.

First, methods are developed to emulate human situation awareness by developing categorization functions. In this manner, historical ship behavior is categorized to reflect distinct patterns. To facilitate this, machine learning is leveraged to generate meaningful representations of historical AIS trajectories, and discover clusters of specific behavior. Second, methods are developed to facilitate pattern matching of an observed trajectory segment to clusters of historical ship behavior. Finally, the research in this thesis presents methods to predict future ship behavior with respect to a given cluster. Such predictions are, furthermore, on a scale intended to support proactive collision avoidance actions.

Two main approaches are used to facilitate these functions. The first utilizes eigendecomposition-based approaches via locally extracted AIS trajectory segments. Anomaly detection is also facilitated via this approach in support of the outlined functions. The second utilizes deep learning-based approaches applied to regionally extracted trajectories. Both approaches are found to be successful in discovering clusters of specific ship behavior in relevant data sets, classifying a trajectory segment to a given cluster or clusters, as well as predicting the future behavior. Furthermore, the local ship behavior techniques can be trained to facilitate live predictions. The deep learning-based techniques, however, require significantly more training time. These models will, therefore, need to be pre-trained. Once trained, however, the deep learning models will facilitate almost instantaneous predictions.