WP 8100 [BAW]

Within work package 8100, the AIS data from WP 4300 are used. In order to prepare the AIS data for the subsequent calculation of the trajectories, the data are first separated for the different ship types and statistically evaluated separately for upstream and downstream navigation depending on the water level. From this, typical lanes and courses of action for the respective ship type can be derived. The associated probabilities of where the ship is can be determined, considering the respective traffic situation (e.g. encounter and / or overtaking) in a typified manner. This work step is carried out, if necessary, in coordination with experienced skippers.

A further approach is pursued with the machine learning method. For this purpose, an artificial neural network is trained with a large amount of recorded AIS data with the aim of predicting ship movements in a specific waterway section, i.e. depending on the location. In addition to the vessel motion data, water level data in particular are used as input for the neural network, since the vessel dynamic characteristics are significantly determined by the current and the available water depths. During the training of the neural network, the deviation of the predicted trajectories from the target trajectories given by the recorded AIS data is minimized. Machine learning will be used to determine likely trajectories for typed vessels in a manner analogous to the statistical analysis of AIS data. The aim of this second approach is to achieve a further improvement of the prognosis compared to the conventional statistical evaluations by taking into consideration unknown influencing factors, for example individual ship shapes (stern shape, bow shape, etc.) or influencing factors dependent on the shipmaster.

A third approach is pursued by the Chair of Dynamics and Control (SRS) at the University of Duisburg-Essen, which is based on statistical analysis and system identification. Anonymised (typed) AIS data are used for this purpose, with the aim of developing a location-independent behaviour model of skippers that is then also able to generate trajectories in waterway sections for which no recorded AIS data are available. Another aspect of this third approach is off-line forecasting, so that trajectories can be generated even in case of failures of the mobile phone connection to the BAW server.

In this work package, different methodological approaches for the prediction of trajectories are being developed in parallel, thus increasing the reliability and resilience of this elementarily important component for the automated control of inland vessels. In a final step, the trajectories determined with the different approaches have to be merged with decision and information fusion methods in order to generate predictions with maximum situational accuracy.