Study On The Radial Basis Function Networks And Its Application To Ship Motion Control

Along with the development of shipping business, ships are becoming bigger, faster and more intelligent thus better performance of maneuver is demanded. To ensure safety and improve economy, it is necessary to adopt new control theories and techniques, and to research for better control strategies. The application of neural network techniques in ship control became an important research area in recent years. Radial basis function (RBF) network have unique adavantages in control applications due to its features of simple topological structure, quick convergence speed and no local minima. This thesis concentrates on studying the learning algorithms of RBF network and two novel learning algorithms are proposed. This thesis also discussed the applications of RBF networks in ship motion predictive control.To meet the demand of on-line application of neural networks in control system, a novel sequential learning algorithm of RBF network is proposed referred to as dynamic orthogonal structure adaptation (DOSA) algorithm. The self-adaptation ability of the resulting network was demonstrated in identification of systems with static and time-varying dynamics. The algorithm is also featured by small mumber of tuning parameters, explicit meaning of parameters, adaptive adjustment of parameters and its robustness to changes of parameters.Generalization capability is an important aspect of neural network capability. On the basis of PLS-based-RBF-network (RBF-PLS) algorithm, a novel two-stage RBF-PLS algorithm is proposed in this thesis. In the algorithm, the number of hidden units and the iterations of principal components extraction are fixed simultaneously. The combination of PLS regression and network structure simplification overcome the influence of multi-collinerity and ensure the good generalization performance of the resulting RBF network, and this is demonstrated by experiments of nonlinear function approximation and generalization.Aiming at the nonlinear and time delay characteristics of ship motion, also for application of control strategy, the RBF-network-based predictive control strategy is proposed. The strategy employ DOSA algorithm for sequential learning and multi-step prediction, and employ the proposed two-stage RBF-PLS algorithm to approximate and substitute the cost function minimization. Finally, the proposed strategy was applied in ship course tracking control simulation and the satisfying performances demonstrate the feasibility and effectiveness of the ship control strategy.