Maritime Vessel State Prediction Using Machine Learning on AIS Data
Machine learning system for collision risk classification and trajectory prediction achieving 97.54% accuracy on 32.4M AIS records.
Duration
Fall 2025
Role
Developer
Institution
NTNU / Norwegian Electric Systems AS
Status
Completed
Technologies Used
Overview
This specialization project developed machine learning models for maritime vessel state prediction using AIS data from Sognefjorden, Norway. Two prediction tasks were addressed: collision risk classification (4 risk levels) and trajectory prediction across multiple time horizons from 30 seconds to 5 minutes. The Random Forest classifier achieved 97.54% test accuracy using 23 engineered features incorporating COLREGS encounter types and maritime domain knowledge.
Problem Statement
Maritime collision avoidance relies heavily on human judgment. By analyzing historical AIS data, we can develop predictive models to assist in early risk identification and improve maritime safety. This project was conducted in collaboration with Norwegian Electric Systems AS to explore integration potential with their RAVEN INS navigation systems.
Challenges & Solutions
| Challenge | Solution | Outcome |
|---|---|---|
| Large Dataset Processing | Developed efficient data preprocessing pipeline with chunked processing | Successfully processed 32.4M records from 1,065 vessels |
| Data Leakage Detection | Identified DCPA/TCPA features causing leakage and removed them | Built robust model using 23 kinematic and geometric features |
| Feature Engineering | Domain-specific features including COLREGS encounter types and relative motion | Achieved 97.54% classification accuracy without explicit CPA calculations |