A Study For Identification And Control Of Nonlinear Systems Using Neural Networks

In this paper, two typical multi-layer feedforward artificial neural networks- BP network and RBF network have been studied detailedly. Their learning and training rules have been analyzed profoundly and their abilities to approximate arbitrary nonlinear function have been testified and compared by the simulation. A new RBF neural network has been presented which uses a raised-cosine function as activation transfer function. It provides a wider generalization in comparison with Gaussian RBF neural networks by simulation as well as strong approximation ability, fast convergence, a rule to select the parameters of the networks.This paper uses multi-layer feedforward artificial neural networks (abbr. MLF NNs) to approximate the nonlinear dynamic inversion on the basis of analyzing principle and characteristics of nonlinear dynamic inversion and neural networks. The pseudo-linear system has been synthesized by classic PID control and self-adaptive control separately. For the latter, on the basis of analysis of the errors, a RBF NNs has been applied to approximate the forward model of the system and another dynamic neural network has been applied to compensate on line the errors from nonlinear dynamic inversion approximation and forward model identification. The training rule of the neural networks has been presented. The simulation results suggest that the control strategies can provide better robustness.The RBF NNs has been used to approximate the uncertainty in the process of an observer design for a class of nonlinear system. At the same time, the observers of nonlinear system with and without uncertainty have been analyzed and designed based on the input-output linearization theory. The training rule for the weights of the RBF NNs has been proposed in the paper based on the known system model and given Lyapunov function. The simulation shows effectivity of the designed observer.An indirect self-adaptive fuzzy-neural network controller (FNNC) has been proposed with its parameters and the structure tuned simultaneously by GA in virtue of the powerful optimization property of GA. The structure of the controller is based on the Radical Basis Function (RBF) neural network with Gaussian membership functions. The performance of the proposed FNNC is compared with a conventional fuzzy-PID controller and the simulation results show that the FNNC presents encouraging advantages.