Modelling and estimation of ship motions

Abstract

Roll, surge, sway, and yaw can occur in ships, boats, and submarines. Surge refers to the forward or backward movement of the vessel along its longitudinal axis, sway refers to side-to-side movement, roll refers to the tilting of the vessel from side to side, and yaw refers to the turning of the vessel around its vertical axis. Estimating these motions is important for a variety of reasons. In maneuvering, knowing the current and expected motion of the vessel can help the operator make informed decisions about how to move the vessel. In dynamic positioning, accurate motion estimations are necessary for maintaining the vessel's position using thrusters and other control systems. Motion estimations can stabilize the vessel or maintain the desired heading in control systems. For safety, knowing the vessel's motion can help the operator avoid hazardous situations and ensure the safety of the crew and passengers. It is also evaluated that the sea state and the direction of the wave can be estimated from the sensor data as a future study. Accordingly, with the heading change, fuel consumption can be optimized and a more stable maneuver can be predicted. Five different approaches are worked on and three of them were successful in this thesis. Roll, surge, sway, and yaw degree of freedoms, together with throttle references and rudder, are used as input and output. Training data generated by using these relations, three different models, namely, Iteratively Weighted Dynamic Transfer Function Model, Neural Network Model, and Deep Neural Network Supported Transfer Function Estimation with Disturbance Model are developed. All the models are used for seakeeping and maneuvering scenarios. Estimation results are validated with variable rudder and throttle commands to show model result generalization capability. All the scenarios are based on real-world applications. In addition to the simulation studies, different real-world tests were also done. The full-size sea experiments were carried out in Tuzla Bay using the pilot boat named Pilot 67 belonging to Istanbul Technical University. The boat was navigated by ITU personnel and Maritime Faculty students on two separate dates, when the weather conditions were 3-4 and 4-5 according to the beaufort chart. The maneuvers made during the experiment reflect the real situation with great success and meet the most difficult conditions because they are under the influence of the sea and they include complex controls other than the standard rudder maneuvers used in the literature. The obtained data was then used in deep learning assisted transfer function model diagnostic studies. The data collected on both days was split into two, half for training and half for validation. In addition, this model has been updated to allow two propellers and a single rudder input, unlike the previously designed model architecture, which is suitable for a single propeller and single rudder input, thus confirming the previously stated idea that the proposed method does not depend on the number of inputs and outputs. The presented work aims to consider the platform model from a different perspective, without having knowledge of the model, where the only required information is the required outputs and the excitation inputs. Consequently, the presented approaches have the adaptability to specify input - output variables to be arranged for the desired utilization and have the possibility to be specifically stretched out to boost the certainty and reliability of underwater, surface, and unmanned vehicle's control. Because of this the work presented here brings a new approach of usage of Deep NN in predicting ship motions. Studies about the planning of paths or avoidance of accidents can be conducted using the proposed methods. The Explanatory Notes to the Rules and Guidelines state, "The track width during the stop test at full astern should not exceed 15 ship lengths." (Germanischer Lloyd, 2012) This rule applies to ships over 100 meters in length. Considering this statement and having 51.5 m (Perez et al., 2006) as the ship length and a speed mediocre of five m/s during deceleration, about 155 seconds of prediction time is required. The results show that the proposed methods can be used for future emergency braking and accident avoidance scenarios.