Abstract
Artificial neural networks (ANNs) have been successfully applied in various areas ranging from signal processing to automatic control. This success is mostly due to the fact that neural networks are equipped with a remarkable learning capability such that a desired input-output mapping can be discovered through training by examples or by on-line adaptation with stable adaptive laws. Although many types of neural networks have been studied and reported, there are still many valuable topics needed to be further investigated. The conventional radial-basis-function (RBF) neural networks adjust only output weights in order to simplify the controller structure. However, this simplification will decrease the approximation ability of NN system. For the multi-layer perceptron neural network, it adapts both of the weights of input-to-hidden layer and hidden-to-output layer, and the approximation ability is highly increased in this MLP-based NN. However, those added adaptation laws for tuning the weights of input-to-hidden layer markedly increase the complexity of the whole adaptive NN system. The aim of this dissertation is to construct an advanced adaptation scheme for upgrading the approximation accuracy of RBF-based neural network. Therefore, we propose an enhanced adaptive RBFN control methodology, in which the stable adaptations of variance parameters is involved to enhance the controller performance and simultaneously the complexity of the controller for the whole system isn’t increased too much. Since almost all of the reported NN-based controller design are only applicable on the minimum phase cases, in order to cope with the non-minimum phase cases, we introduced a generalized controller structure based on the notion of synthesizing a stable compound system. Based on this generalization, we can extend the proposed enhanced RBFN system to deal with the output tracking control for a class of nonlinear systems with non-minimum phase. In addition, the stability assurance is highly important in these on-line control applications. Therefore, in this dissertation, the direct-type adaptive-control architecture is employed, in which the proposed enhanced adaptive RBFN system combined with sliding mode control is used to on-line re-construct the desired control law so that the problem of controller singularity can be avoided. Not only stability of the whole adaptive neural-network-based system is guaranteed but also the influence of the re-construction error and external disturbance can be well compensated. Furthermore, the practical realization is also an important issue on the application of the neural-network-based controllers. Not only through computer simulations, we also realize the hardware implementation on a practical experimental pole-cart system to verify the effectiveness of the proposed schemes.