Automatic Identification System (AIS) messages are a rich source of data for situational awareness of maritime surface vessels and identification of potentially anomalous vessels, however they are too voluminous and heterogeneously sequenced for a human operator and traditional compute algorithms to make sense of at a regional scale. While these challenges can limit the effectiveness of traditional AIS data processing methods to detect anomalous maritime surface vessel behavior, advanced machine learning methods, specifically Neural Networks, may prove more successful. Blue Ridge Envisioneering (BRE) proposes a multi-stage Recurrent Neural Network (RNN) that addresses these challenges in identifying anomalous ship behavior. Using AIS messages from U.S. coastal waters provided by the marinecadastre.gov team, we will train this network to detect anomalies and will characterize performance across a variety of geographic regions.
Benefit: Situational awareness of marine vessel activity is not only crucial U.S. Naval forces but also commercial shipping entities that need to monitor the activities of their own ship and the surrounding ships within the operating regions. The ability for shipping companies to detect anomalous behavior is applicable to both third-party vessels (from a safety perspective relative to their own fleet) and for monitoring of their own ships to track potentially dangerous or fraudulent events (ship has been hijacked, ship is engaged in illicit activities, etc.). The results of this effort represent a critical first step in determining the potential for neural networks to meet this need (commercially and militarily), and identifying the effort to transition to an operational capability
Keywords: anomaly detection, anomaly detection, Machine Learning, Artificial Intelligence, Recurrent Neural Networks, AIS, Model Generalization
