ADAPTIVE NEURAL NETWORK CONTROL : DESIGN, STABILITY AND PERFORMANCE ANALYSIS

Abstract

Recent years have seen a rapid development of neural network control techniques and their successful applications. However, from a theoretical viewpoint, many fundamental problems still remain in stability, robustness, and transient performance of neural network adaptive systems. This thesis investigates systematic design approaches for stable adaptive neural control of nonlinear dynamical systems. All the results obtained are derived through rigorous mathematical analysis, and control performance of the closed-loop neural control systems is quantified by mean square, L2 or L? error criteria analytically. In addition, the effects of controller design parameters, initial conditions and reference signals on system stability and transient behaviour are explicitly given for providing valuable insights into performance improvement and design trade-offs. The Lyapunov stability technique plays an important role in the design and stability analysis of neural-based adaptive control systems. This thesis presents three kinds of adaptive control designs using neural networks, i.e., regionally, semi-globally, and globally stable neural adaptive control, under different conditions and a priori knowledge for the studied plants. Without the requirements for (i) off-line training of neural networks, and (ii) persistent excitation (PE) conditions for system signals, stability of the overall neural network systems is rigorously proven by Lyapunov stability analysis.A major contribution of the thesis is the invention of an integral-type Lyapunov function for constructing adaptive NN controllers which completely solves the control singularity problem existing in adaptive feedback linearization method. By combining neural networks, integral-type Lyapunov function, and adaptive backstepping design, a stable neural controller is developed for strict-feedback nonlinear systems. The proposed techniques are also shown to be efficient in adaptive NN control for multivariable nonlinear systems in a triangular control form. Moreover, such a new kind of Lyapunov functions is successfully applied to the adaptive control problem of a class of nonlinearly parametrized plants. Both stability and transient behaviour of the developed adaptive systems are investigated, which are helpful for the evaluation of control performance and the choke of design parameters. In the last part of the thesis, adaptive control problem of general non-affine nonlinear systems is studied through multilayer neural network parametrization. When all the system states are available, an adaptive neural controller is firstly presented by state feedback. For the case of only system output being measurable, an output feedback neural control scheme is developed by utilizing a high-gain observer. To illustrate the effectiveness of the proposed controller, an example is provided through composition control for a continuously stirred tank reactor (CSTR) plant. Finally, conclusions of the thesis are made, and some suggestions are given for further research.