Autoencoder-Based Anomaly Detection for Safe Autonomous Ship Operations

The development of autonomous ships is advancing, but ensuring their safe operation remains a challenge. To aid safe operations, autonomous ships are expected to be monitored by humans in a remote operation center. A key challenge is ensuring that human operators remain alert and ready to take control of the system when necessary. Maritime traffic poses a potential hazard to autonomous vessels, and systems to aid the operator in identifying abnormal ship behavior in time should be in place. This study develops deep learning models that automatically detect anomalous ship behavior to aid human operators. A case study related to the remote operation center in Horten, Norway is conducted, where four various autoencoder architectures have been trained on historical Automatic Identification System data to detect maritime traffic anomalies in the Oslo fjord. The models are trained in an unsupervised manner, such that they are able to automatically identify anomalies, without the need for manual labelling. The results indicate that a recurrent autoencoder is the most promising architecture for decision support of remote operators, as it is able to identify a variety of anomalies, with fewer false positives.

Keywords: Autonomous ships, Remote operation centre, Machine learning, AI, Anomaly detection, Autoencoder, Maritime traffic, Maritime safety, Decision support.

Download PDF