Transformer-based Anomaly Detection in Marine Settings — 54p — Amandus Reimer, Julius Thunstrom, Nickolas Weir, Oswald Adohinzin

A significant portion of worldwide trade is carried out by vessels operating the oceans and seas. According to the World Economic Forum, about 90% of all traded goods are shipped by sea. Monitoring the travelpaths and finding unconventional behavior as early as possible is a large part of ensuring the safety of trade across the globe, and is what this project is investigating. 

In this project we attempt to identify anomalies in marine traffic in real time, based on AIS messages. To do this we are utilizing two transformer architectures; one decoder-only transformer and one encoder-decoder transformer, referred to as the autoencoder. The decoder-only model is used to predict the future trajectory which is then fed to the autoencoder to perform the actual anomaly detection. 

To perform the anomaly detection we apply statistical algorithms, for example isolation forests, to the latent space generated by the encoder part of the transformer. We also evaluate doing anomaly detection based on the reconstruction error of the autoencoder. That is, if the error between the original sequence and the reconstructed one surpasses a certain threshold the sequence is considered anomalous. 

Results indicate these methods can detect some outliers but further research is required to reduce the number of false positives and to improve recall. 

This project is created in collaboration with AI Sweden, Dakota State University and the Swedish Defence University.

Chalmers University of Technology
Mark Spanier