| Abstract [eng] |
This dissertation addresses the issue of evaluating ship collision risk and maritime traffic awareness by employing historical Automatic Identification System (AIS) data and deep learning techniques. Given the importance of maritime traffic safety, the study explores recurrent neural network architectures, optimises their hyperparameters, and applies feature engineering and coordinate transformations. The main goal is to predict vessel movement trajectories and assess collision risks. The proposed LSTM autoencoder model forecasts not absolute coordinates but their delta vectors, which are recursively added to the known vessel position. This approach enables more reliable multi-step predictions (up to 20 minutes ahead). The models are further enriched with vessel type and meteorological data, enhancing their accuracy and applicability across diverse scenarios. The dissertation presents a real-world case study, a 2021 vessel collision near Bornholm, demonstrating the practical application of the proposed methods in risk assessment. Statistical techniques, prediction and confidence intervals, elliptical (EPR), and conformal (CPR) prediction regions are employed to delineate potential ship locations. The probability of collision is estimated using the Jaccard index to evaluate the overlap of these regions. The findings indicate that deep recurrent neural networks can significantly contribute to improving maritime navigation safety and the prevention of dangerous situations at sea. |
