The Dutch government has set an ambitious target of reducing CO2 emissions by 55% by 2030, primarily through the development of offshore wind farms in the North Sea. While this transition supports sustainability, it also reduces available space for maritime traffic, increasing th
...
The Dutch government has set an ambitious target of reducing CO2 emissions by 55% by 2030, primarily through the development of offshore wind farms in the North Sea. While this transition supports sustainability, it also reduces available space for maritime traffic, increasing the risk of incidents. In response, this research aims to improve the detection of anomalous behaviour in the North Sea using machine learning, assisting Dutch Coast Guard operators in their monitoring tasks.
The primary focus of this study is to detect anomalous cargo vessel behaviour, specifically drifting, which has safety implications, as evidenced by the Julietta D. incident in January 2022. The research uses Automatic Identification System (AIS) data to monitor vessel trajectories and apply a machine learning-based anomaly detection model. This model uses features extracted from ship motion (e.g., speed, rate of turn), spatial properties (e.g. presence in anchorage area) and metocean condition (e.g., wave height) to detect anomalous behavior. The Local Outlier Factor (LOF) algorithm is employed to identify local outliers in a two-dimensional embedding of vessel trips, which is created using a density-preserving dimension reduction technique called densMAP.
The model successfully detects anomalous vessel behaviour within 30 minutes of occurrence, demonstrating its potential for real-time monitoring. In a case study of the cargo vessel Julietta D. incident, the model identified the drifting behaviour of this vessel, validating its effectiveness. The model also shows promise in detecting other types of anomalous behaviour, such as sudden changes in speed, and presence in defined areas.
While the model performs well in detecting a drifting incident and certain anomalies, it does not account for global outliers or identify other types of anomalous behaviour. Additionally, due to the limitations of AIS data, heading information was not incorporated. The research contributes to the safety monitoring of maritime traffic by providing a scalable, interpretable, and operationally feasible machine learning approach for the detection of anomalous cargo vessel behaviour.
Future work should focus on integrating time components, and supervised learning with labeled incident data to further refine the model. Ultimately, this research offers significant potential for enhancing operational safety and supporting the Coast Guard in the monitoring of maritime traffic in the North Sea.
