Nonlinear system identification and control using a neural network approach

(In this thesis, the plant identification, state estimation based on the identifiedplant and also the design of a neuro-controller using multi-layer perceptrons(MLPs) for a complex system are presented. The quasi-linear system to becontrolled is both unstable and nonlinear. The complete nonlinear feedback controlsystem is designed without a priori information of the plant dynamics, using onlymeasured input/output data. The first design step is to combine a conventionalmethod of multivariable system identification with a dynamic multi-layerperceptron (MLP) to achieve a constructive method of system identification. Basedon the identified linear model of the system, states will be estimated and convertedto more appropriate state for control in the second design step. The class of quasi-linear nonlinear systems is assumed to operate nominally around an equilibriumpoint in the neighborhood of which a linearized model exists to represent thesystem, although normal operation is not limited to the linear region. The resultspresented here provide an accurate discrete-time nonlinear model, which is used inthe design of a nonlinear state estimator. The controller design is derived from aswitched-linear feedback controller from the estimated states using the identifiedlinearized model of the system around each suitable operating point, as a rolemodel for the neuro-controller in the initial phase. Finally, using the partially