| dc.description.abstract | This thesis addresses the challenge of predicting vessel trajectories in confined waters, where the multi-dimensionality of the interaction scene introduces non-linearity in ship paths. The proposed approach utilizes machine learning models trained on processed Automatic Identification System (AIS) data to forecast future vessel trajectories based on observed states.
Until now, latitude, longitude, and other AIS features such as speed over ground (SOG), course over ground (COG) have often been used to train models for target latitude and longitude predictions. However, these models often struggle to capture the inherent non-linearity present in confined water scenarios. To overcome this limitation, the thesis proposes a novel data representation to better capture motion patterns.
The methodology starts with, processing the AIS data using a multi-step framework,
maximizing data trueness, retaining valuable information, and minimizing noise. Additionally, learnable features are engineered to improved the mapping of data to target trajectories. Three types of recurrent neural network (RNN) architectures, including Long-Short Term Memory (LSTM), Gated recurrent unit (GRU), and Convolutional Neural Network- Long short term memory (CNN-LSTM), are employed in the study. Two distinct approaches are used for training the models. The first approach uses traditional features such as latitude, longitude, COG, SOG, along with engineered features, mapping them to target latitude longitude coordinates, all in standardized form. The second approach, trains the models to predict relative distance and angle based on observed latitude, longitude and other features. The performance of different approaches and models is evaluated using the accuracy of median target predictions as the key performance indicator.
The results demonstrate that the second approach, which incorporates the novel Relative Displacement and Angle (RDA) data representation, significantly outperforms conventional data representation techniques. Across all models, the RDA approach exhibits approximately 65% improvement in accuracy compared to standard representation methods.The CNN-LSTM model achieves the lowest median deviation error of 0.021 nm for each time-step over a 5 minutes prediction horizon.
Though the RDA approach provides us with a better learnable representation, the models used in this thesis are conventional in nature, with LSTM and GRU being utilized in previous works. Incorporating the RDA approach with more sophisticated models has the potential to further enhance accuracy. Additionally, the thesis acknowledges the limitation of not considering encounter data, which can impact the accurate estimation of vessel intent/trajectories in encounter scenarios. | |
