Adaptive Control For A Class Of MIMO Nonlinear Systems Based On Neural Networks

Considering the principles of combining theoretical study with engineering application and innovation with practicality, based on the state of the research of adaptive neural control theory, the existing main problems and the practical requirements of engineering applications, this dissertation deeply investigates the control theory for unknown nonlinear multi-input and multi-output systems, succeeds in constructing a systematic design structure, and provides new effective approaches.In this thesis, the problem of adaptive control for unknown affine MIMO nonlinear system is studied and proposes two adaptive controllers. The first one is adaptive RBF neural network controller based on Growing and Pruning algorithm for RBF (GP-RBF). The structure and parameters of RBF neural network are determined online with the algorithm of growing and pruning which is obtained through combining a proposed neural sensitivity concept with hyper-sphere clustering. When the error condition is satisfied, the weights of controller are further adjusted and obtained by the adaptive law based on Lyapunov theory, which guarantees the global stability and convergence of controller. Simulation results show that the fast tracking ability, perfect stability and convergence of the controller.The second controller is adaptive RBF neural network controller based on hybrid genetic algorithm (HGA-RBFNNC). The structure and parameters of RBF neural network are determined online by hybrid genetic algorithm which is the combination of the improved genetic algorithm with and the steepest decent algorithm. Then the weights of controller are further adjusted and obtained by the adaptive law based on Lyapunov theory, which guarantees the global stability and convergence of the system. Simulation results show that the fast tracking ability, perfect stability and convergence of the controller.The two controllers needn't experiences to define the structure and parameters of neural network. Not only online design the neural network don't need recognize the model of the plant but also the study of controller and the control of the system are simultaneous and the teacher signal is not needed, which avoids the shortage of the training data in the offline study. They all use the adaptive law based on Lyapunov theory, which guarantees the global stability and convergence of the system.The simulations of two-link robot manipulator and the experiment on the boiler temperature show that the proposed methods are effective, which realizes the adaptive tuning of the structure and parameters of RBF neural network and ensures the global stability and convergence of the system.