Machine Learning based Analysis on Maneuvering Data in Port for Automatic Berthing

Abstract

As interest in autonomous ships has increased in recent years, various research on collision avoidance, route optimization, and automatic berthing of ships are being actively conducted. Among such research topics, the docking process is one of the most sophisticated areas of research as it consists of complex maneuvering movements in a narrow port, and is greatly affected by environmental disturbances. Unfortunately, accidents still occur, especially during the berthing process and pilotage. Therefore, by making the berthing process automatic, it is widely acknowledged that the entire docking process can become more efficient. Also, automation will eliminate the need for human pilotage which has been the major cause of accidents. Generally, in order for a ship to berth, an experienced pilot, crew, and tugboat are required. However, since automatic berthing is performed without manpower and tugboat, automatic berthing guidance and control system needs to be highly accurate. In this thesis, three studies were conducted for automatic berthing; Pilot research was conducted in relation to manpower and two studies on trajectory analysis and berthing velocity were conducted for automatic berthing induction.

Piloting can be split into two types: sea piloting (navigation in fairways) and harbour piloting (maneuvering of the ship in the harbour area). In this study, ship berthing velocity data (harbour piloting) and trajectory data (sea piloting) were used as the main maneuvering data for automatic berthing. This is because the berthing velocity has the greatest influence on the amount of impact generated by contact between the ship and the dock facility, and trajectory data is the record of pilotage on which the automatic berthing will be based.

This study focused on optimal berthing velocity calculation based on ship berthing velocity. In order for the ship to dock safely, the appropriate berthing velocity should not be exceeded and the berthing energy should be smaller than the absorbed energy of the fender. This study intends to derive the allowable berthing velocity for different ship sizes by taking fender performance and berthing capability into account, and defines the extrapolated berthing velocity as the relative value calculated according to ship size and the berthing capability of the pier. The regression equation for allowable berthing velocity by ship size was derived by calculating the berthing energy for different fender performances. The study also focused on automatic berthing (Automation of the entire berthing process). However, since it is technically impossible for an off shore vessel to be completely automated, several other methods that aid such vessels are necessary. One way to achieve this is by using shore-based pilotage. A group of experienced pilots on the shore can provide quantitative analysis of data, and help a pilot in the process of automatic berthing. In this process, berthing velocity and berthing energy were employed as basic data, and the mean and standard deviation values for each pilot after preprocessing were used for analysis. As a result of using the agglomerative clustering algorithm, we grouped pilots into three types: cautious, efficient, and hazardous. The clustering algorithm was then applied to analyze the pattern of arrival trajectory data during sea piloting, and as a result, clustering showed the highest performance in spectral clustering. In operating the online control of the autonomous berthing, it is effective to use the predefined trajectory derived from trajectory planning and control input as a reference for tracking control obtained by the optimization problem. Using the trajectory data, a quaternary function through polynomial regression was set as the optimal trajectory (guideline), and the results were compared according to the presence or absence of the guideline through reinforcement learning. A basic study was conducted for automatic berthing to reduce human error through the presentation of allowable berthing velocity and clustering of pilot maneuvering types. In addition, through the analysis of entry patterns and learning of routes through guidelines, a crucial step that contributes to automatic berthing can be completed.