| Ship manoeuvering mathematical model is not only one of the key technologies in ship-handling simulator,but also the significant simulation platform in ship motion control research.Non-parametric model directly learns the mapping between system input and output,which can avoid the problem of inaccuracy in model structure,and can theoretically establish the mathematical model of ship motion with higher precision.In the framework of data driven modeling,a novel non-parametric modeling method is proposed based on locally weighted learning(LWL)algorithm,and aims to provide the ship maneuvering mathematical model with high accuracy for ship-handling simulator and marine control engineering.LWL is one of the classical non-parametric learning methods,which has found its wide use in robot control system.However,ship motion differs from robot motion,be-cause the ship maneuvering motion has the characters of large inertia and high nonline-arity.In order to propose a modeling method,which can meet the need of ship-handling simulator,this dissertation conducts the research in three aspects,namely,algorithm application,distance metric optimization and computational complexity.The main con-tributions and results are summarized as follows:1.LWL algorithm application research.Both global optimal and locally optimal LWL are applied to ship maneuvering motion modeling,which is based on data driven idea and LWL algorithm.Motivated by the key idea of support vector machine(SVM),the last ship motion speed and acceleration signals are introduced to system input space to solve the one to many mapping and inseparability problems.Futhermore,the local approximation property of LWL algorithm circumvents the problem of ship motion nonlinearity.Numerical example,Delta Linda tug simulator and Abkowitz model of Mariner Class Vessel are studied.In numerical example,the fitting precision of locally optimal LWL is 40.1%higher than traditional global optimal LWL.The mathematical model of Delta Linda and Abkowitz model of Mariner Class Vessel also demonstrate the effec-tiveness of LWL.Simulation results are in good agreement with experimental data,ind-icating that the proposed scheme can effectively learn the characteristics of ship manoeuvering motion.2.Distance metric optimization research.The classical LWL which learns the dis-tance metric by gradient descent method may lead to locally optimal,over-learning or under-learning.A modified genetic algorithm is developed by redefining a fitness func-tion which would assign the maximum fitness to the optimal distance metric.The ge-netic algorithm searches the optimal solution parallelly and in multi-points,can solve the locally optimal problem in existing gradient descent method and eliminates the over-learning or under-learning in distance measure training.The advantage of fitness function compared with the traditional fitness function is qualitatively proved by use of schema theory.Experimental data from numerical example and the full scale trial of YUKUN are verified by simulation.By comparing the numerical example simulation with the traditional method,fitting precision is improved by 11.81%.The validity is also effectively verifyed by the full scale trial from YUKUN.Simulation studies show that this method can effectively learn the characteristics of ship maneuvering motion of YUKUN.3.Distance metric optimization research.Multi-innovation gradient iteration algo-rithm is proposed,which solves the problems of over-learning or under-learning and overlong training time during optimizing distance metric.The algorithm adopts the ob-jective function gradients of the past and present iteration to update the distance metric,so as to effectively reduce the training time.Simultaneously,many other studies are also conducted,such as the impact of the innovation length,distance metric initial value and convergence factor on algorithm convergence.It is proofed that estimation error of the proposed scheme can converge to zero.Abkowitz model of Mariner Class Vessel and full scale trials from YUKUN show the effectiveness of LWL.Simulation results demonstrate that training time is reduced by 95.26%,compared with the locally optimal LWL.The scheme can learn ship maneuvering motion and is an effective distance met-ric optimizing algorithm.4.Computational complexity research.In order to reduce the computational com-plexity of LWL algorithm,grid index subspace constructed algorithm is proposed to assign a subspace to each prediction point.LWL only need recall a subspace rather than the whole sample space,so as to reduce computational complexity of LWL algorithm.Due to small computational complexity,grid index algorithm can effectively reduce computational complexity of the whole model.The evaluation of computational com-plexity is measured in the following two aspects:calculation quantity and computer ex-ecution time.Abkowitz model of Mariner Class Vessel is taken as simulation example.Zigzag test simulation results show that computer execution time is reduced by 95.0%,compared with locally optimal LWL,and decreased by 34.80%,compared with k Nearest Neighboring algorithm.The algorithm can learn the characters of ship manoeuvering motion.It can also reduce the computational complexity of LWL,and ensure real-time performance of the model. |
