On The Modeling Of Ship Manoeuvring Motion By Using Support Vector Machines

Manoeuvrability is one of the most important hydrodynamic performances of ships, which has a close relationship with navigation safety. Manoeuvrability prediction is an important task at the ship design stage to guarantee good manoeuvrability of ship designed. For this purpose, International Maritime Organization (IMO) promulgated the interim standards and the standards for ship manoeuvrability in 1993 and 2002, respectively. In the standards, prediction of ship manoeuvrability at the ship design stage is explicitly required.At the ship design stage, four methods can be used to predict ship manoeuvrability, including the methods of database or empirical formula, free-running model tests, computer simulation based on mathematical model and Computational Fluid Dynamics (CFD) based numerical simulation. Among them, the third one is the most popular and effective. To use this method, accurately determining the hydrodynamic derivatives in the mathematical model is vital to the prediction accuracy. To determine the derivatives, four methods are available, i.e., the methods of database or empirical formula, captive model tests, theoretical and numerical calculation, and system identification combined with free-running model tests. With the development of measurement techniques and system identification techniques, the method of system identification combined with free-running model tests has renewedly attracted the attention of the ship hydrodynamics community and has a promising perspective in application.Nowadays, two mathematical models are usually adopted in predicting ship manoeuvrability. The first one is the Abkowitz model or the so-called whole-ship model, in which the hydrodynamic forces and moments acting on the hull, propeller and rudder are treated as a whole and expanded by Taylor series around the equilibrium point of forward motion with constant speed. The second one is the MMG model or the so-called modular model. This model decomposes the hydrodynamic forces and moments into three components, the one on hull, on rudder and on propeller, respectively, with the interactions between the hull, the rudder and the propeller being taken into account. Abkowitz model and MMG model are also called hydrodynamic models. Besides, another kind of model, i.e., the so-called response model, is frequently used in the study of ship manoeuvring and control. This model reflects the turning response of ship to rudder action and is mainly used in the autopilot design; it may also be used to predict simple ship manoeuvring motion.Several system identification techniques, such as least squares method, extended Kalman filter, ridge regression method, neural network method and spectral analysis method etc., are available for the modeling of ship manoeuvring motion. No matter which one is used, diminishing multicollinearity and resulted parameter drift during identification of the numerous hydrodynamic derivatives is the key problem to be solved to improve the modeling accuracy.In this thesis, a novel method of modern artificial intelligence, Support Vector Machines (SVM), is applied, for the first time, in the study of modeling of ship manoeuvring motion, including mechanism modeling and black-box modeling. At the stage of mechanism modeling, the hydrodynamic derivatives and manoeuvring indices in the mathematical models of manoeuvring motion are identified by using SVM, and the manoeuvring motions are predicted with the models. At the stage of black-box modeling, the input-output response characteristics of the nonlinear dynamic system of ship manoeuvring motion are identified by using SVM. The hydrodynamic forces and manoeuvrability parameters are predicted by using the SVM-based black-box model.At the stage of mechanism modeling, the SVM-based identification and its application are verified by simulation and model tests. At the stage of simulation verification, Least Squares Support Vector Machines (LS-SVM) based on linear kernel is adopted in the regression of the simulation data of the Tanker 210 000DWT and the Mariner-class ship. By using the regressive hydrodynamic derivatives in the linear hydrodynamic model and the Abkowitz model, the zigzag manoeuvres and turning circle tests are predicted. At the stage of test verification, the response models including the first-order linear, the first-order nonlinear, the second-order linear and the second-order nonlinear models are firstly used to analyse the data of free-running model tests conducted in the State Key Laboratory of Ocean Engineering at Shanghai Jiao Tong University by using LS-SVM, and the parameters in the models are identified. By using the regressive models, the zigzag and turning circle tests are predicted. Furthermore, test verification of SVM identification and its application is conducted by using the model test data of KVLCC1 and KVLCC2 models recommended by the International Towing Tank Conference (ITTC) Manoeuvring Committee for international comparative study. The model tests include free-running model tests and captive model tests. For the free-running model tests, response models and Abkowitz model are adopted. To moderate the drift effect of hydrodynamic derivatives, the Abkowitz model is simplified: The total speed of ship model is used as the non-dimensionalizing factor and the nonlinear terms concerning the surge speed are ignored. The coupling terms of sway speed and yaw rate in the equations of sway and yaw motion are substituted by the cross-flow models. By using the regressive model with the identified hydrodynamic derivatives, zigzag maoneuvres are predicted. The feasibility and validity of SVM method in modeling of ship manoeuvring motion by analyzing the data of free-running model tests are verified. For the captive model test, oblique towing test is taken as example and regression analysis is conducted by using the limited test data to obtain the mathematic models of sway force and yaw moment. The predited results at different test conditions are compared with the test results and the feasibility of SVM method in predicting the hydrodynamic forces by analyzing the data of captive model tests is preliminarily demonstrated.At the stage of mechanism modeling, several measures are taken to diminish the parameter drift. Firstly, applying the principal component analysis method, the linear terms of the hydrodynamic derivatives are decomposed and reconstructed to moderate the linear dependence between the linear terms and the other terms. Secondly, an additional excitation signal is designed to moderate the drift due to the dynamic cancellation effect of viscous forces while executing helm and the vanishment of inertial forces in transition stage. Thirdly, the input-output samples are expressed in a difference form to moderate the multicollinearity of the input variables in the regression model. The results of identification and prediction demonstrate the validity of the proposed measures.At the stage of black-box modeling, SVM based on Guass kernel is used for identification. With the rudder angle and the variables of manoeuvring motion as inputs and the hydrodynamic forces as outputs, the complicated nonlinear functions in the Abkowitz model are identified; and the surge force, sway force and yaw moment are predicted by using the functions identified. Taking turning test as example, with the rudder angle as inputs and the manoeuvrability parameters of turning circles as outputs, the input-output mappings are identified and the manoeuvrability parameters such as the advance, the transfer and the tactical diameter are predicted by using the identified mappings. Taking the identification and prediction of the heading angles as example, the learning abilities of SVM and classical BP neural networks are compared, from which the advantages of SVM over BP neural networks are verified. Taking the identification and prediction of thrust forces as example, the validity of SVM in filtering and smoothing data is verified. Taking the identification and prediction of rudder angle and variables of manoeuvring motion as example, the validity of SVM in reconstructing samples is verified.The study in this thesis has demonstated the validity of SVM method in modeling of ship manoeuvring motion. It provides a novel system identification approach to the modeling of ship manoeuvring motion, including mechanism modeling and black-box modeling. Moreover, it may provide new ideas to optimize the manoeuvrability-related model test design.