Anomaly Detection in Vessel Track Data

This thesis introduces novelty detection techniques that use a combination of Gaussian processes, extreme value theory and divergence measurement to identify anomalous behaviour in both streaming and batch marine data. The work is set in context by a review of current methodologies, identifying the limitations of current modelling processes within this domain. Marine data modelling is first improved by endowing the Gaussian process with the capacity to model both first order and second order dynamics; enhancing maritime data modelling through exploration of appropriate Gaussian process kernels. Gaussian processes are then used to forecast probable future vessel positions. The concept of combining the predictive uncertainty from the Gaussian process with extreme value distributions is then introduced. This provides a means of detecting anomalous vessel dynamics given the previously learnt model. The process is made amenable to online operation through adaption of the Gaussian process to sequential updates. The latter allows the model to be updated in an efficient online manner, after confirming that received data lies within the probability bounds of the model forecast behaviour. Finally a means of measuring distance between functions is introduced, which is used to identify communities of similar vessel types based on the underlying vessel dynamics. This is used to address the issue of vessel class (i.e. fishing vessel, cargo vessel etc.) misrepresentation through detection of anomalies between the inferred vessel class and the class broadcast by the vessel.