Under the Convention on the International Regulations for Preventing Collisions at Sea(COLREGs), every vessel should determine the existence of a collision risk and take an action to avoid collision in ample time if there is a collision risk. However, the results of analysis on the maritime collision accidents over the past five years(2015-2019) revealed that 95% of all collision accidents were caused by the human factors, and 73% of them were look-out negligence. It shows that a sufficient assessment of the risk of collision and a collision avoidance of sailing ship are not properly performed. In order to solve this problem, this study aims to develop a new decision support methodology that helps the assessment of the collision risk and the collision avoidance for duty officer. The methodology consists of three steps: predicting the ship's trajectory, estimating the risk of collision between ships, and path finding for collision avoidance.
The first step proposed methodology of ship's trajectory prediction in coastal waters by the recurrent neural network model. Pre-processing processes such as classification, filtering, grouping, and trajectory interpolation were performed to ensure the accuracy and effectiveness of shipborne automatic identification system(AIS) data to be used in model development. From the preprocessed AIS data, each valid trajectory data was extracted, and the pattern was recognized by the spectral clustering method. In the clustering process, the similarity between the trajectories was measured by applying the LCSS distance measurement method in consideration of the characteristics of the ship trajectory data, and based on the results, the final clustered label was assigned to the each trajectory using the K-means clustering method. Each trajectory cluster was trained and predicted based on the LSTM and GRU models of the recurrent neural network. As a result of comparing the performance of the two models, the prediction accuracy of the LSTM model was better with a slightly different.
The second step proposed methodology to estimate the risk of collision between ships using support vector machine(SVM) and relevance vector machine(RVM) respectively, which are the supervised learning techniques of machine learning framework, and compared the estimation performance with each other. The data used as the input vector of the model was obtained by preprocessing the AIS data. The results of comparison showed that the RVM model showed higher accuracy than the SVM model, but the difference was small. However, the number of basis functions required for the SVM model was approximately 17 times higher than that of the RVM model, indicating that the computational complexity of the RVM model was lower, and the time required for training each model was also 250 times larger than that of the RVM model. It appears that RVM has excellent data learning efficiency.
Finally, model predictive control(MPC) method and particle swarm optimization(PSO) algorithm were applied to the study on the optimal path finding based on the ship's trajectory prediction and the estimation of collision risk between ships. In relation to the prediction, which is the main process of the MPC method, the trajectory of the target ship was predicted by the LSTM models of the recurrent neural network developed in Chapter 3 of this study. And the trajectory of the own ship was predicted by the ship dynamic model, which is the 3-degree of freedom(3 DOF) control model of horizontal motion. To estimate the risk of collision, the RVM model derived through Chapter 4 of this study was applied. In order to find the optimal path, propulsion commands and course offset commands were selected as variables with bound constranists in PSO algorithm. In addition, the line-of-sight(LOS) guidance was applied so that the planned straight path was continuously followed even if the trajectory was deviated due to the action to avoid collision. In order to verify the proposed method, an simulation was carried out for each encounter situation, such as the head-on situation, crossing situation, and overtaking defined in COLREGs. As a result, it was confirmed that the own ship complied with the steering and sailing rules of COLREGs and safely avoided the target ship.
The methodology proposed in this study is expected to effectively support the decision-making of navigators who are in danger of collision between ships and prevent collision accidents. It will also contribute to research related to finding the optimal path of unmanned ships to avoid collisions while complying with COLREGs.