Neural Network-based Robust Control And Filtering For Nonlinear Systems

This thesis is concerned with the problem of robust control and nonlinear filtering for two kinds of complex nonlinear systems with the help of neural networks, i. e., stochastic complex nonlinear systems with Markov jump parameters and deterministic complex systems of greenhouse climate control, which falls into two major parts.In the first part, based on the robust control and filtering of a class of nonlinear Markov jump systems (MJSs), the steady-state behavior and the transient performance of nonlinear MJSs with partially known transition jump rates are further analyzed and discussed.First of all, the analysis and synthesis problem of performance robustness is studied for a class of nonlinear MJSs with uncertainties and external disturbances with the aid of intelligent control method. Incorporating neural network with robust control, the nonlinearities are initially approximated by multilayer feedback neural networks. Subsequently, the neural networks are piecewisely interpolated to generate a linear differential inclusion (LDI) model. Then, substituting the neural network with the representation of LDI into nonlinear systems, the resulting closed-loop control systems or the closed-loop error systems are robustly stochastic stable and satisfy the H∞performance index for all admissible uncertainties, the external disturbances and approximation errors of the networks.Subsequently, in real physical systems, not all the transition probabilities of the jumps are easy to measure, and even part of the elements in the desired transition rate matrix is not available. Therefore, it is necessary to study the steady-state behavior for more general nonlinear MJSs with partially known transition probabilities, i. e., the robust H∞control and the robust H∞filtering. Considering the uncertainties produced by modeling errors and the situation involving unknown statistic characteristics of the processor noise and the measurement noise, the robust H∞controller and filter design methodology is presented. Based on the Lyapunov theory, sufficient conditions for the existence of the desired controller and filter are derived.Then, in terms of practical engineering, transient performance of nonlinear MJSs is further studied, i. e., in a fixed time interval, the trajectories of the system stay within a given bound for all admissible uncertainties, the norm bounded external disturbances and approximation errors of the networks. In both situation of partially known transition probabilities and complete access to transition probabilities, the robust finite-time stabilization and filtering criteria for the underlying systems is discussed. What is more, the detailed design proposal is given for optimal finite-time controller and optimal finite-time filter.In the second part, with the view of improving the control accuracy of greenhouse climate systems, the relevant key technique has been further explored from two aspects:one is the adaptive capability of control policy, the other is the accuracy of filter.Firstly, in order to improve the adaptive capability of control policy, a general framework of robust adaptive neural network-based controller design for greenhouse climate system with time-varying external disturbances is presented, in which not only the stability, convergence property and the robustness are discussed, but also the robust adaptive control method, which combines the well-known feedback linearization with radial basis function neural networks, makes great sense to the control quality of greenhouse climate system.Secondly, from filter point of view and assuming the process noises and sensor noises are uncorrelated white Gaussian random process, the extended Kalman filter (EKF) is proposed to estimate the greenhouse climate system. The simulation results show that the control accuracy of EKF observer-controller combination is much better than that of controller alone.Thirdly, the ability of the unscented Kalman filter (UKF) to accurately estimate nonlinearities makes it attractive for implementation on greenhouse climate control systems. The unscented transformation coupled with certain parts of the classic Kalman filter, provides a more accurate method than the EKF for estimating the greenhouse states and filtering out the noises. What's more, the UKF is far easier to implement because it does not involve any linearization steps.Finally, to reduce the influence of filtering ability caused by Gaussian assumption of real noises and further improve the control accuracy, the nonlinear particle filter is derived in an attempt to solve the state estimation problem of the greenhouse climate control systems with non-Gaussian process and measurement noises, which solves the trouble controlling of nonlinear and non-Gaussian noises in industrial process control for a long time.