Abstract

This thesis deals with on-line adaptive control of a class of dynamic systems preceded by backlash-like hysteresis nonlinearities via dynamic neural networks. A three layer recurrent neural network called the diagonal recurrent neural network (DRNN) is applied to construct the hysteresis inverse compensator (DRNNC) to remove the effect of hysteresis. An on-line learning algorithm called the dynamic back propagation (DBP) algorithm is developed to train the DRNN. Based on the cancellation of hysteresis effect, an adaptive tracking control architecture, which is constructed through the combination of sliding mode and Gaussian network (GNNC), is then proposed. The diagonal recurrent neural network compensator (DRNN) and Gaussian network controller (GNNC) are trained at the same time since DRNN requires fewer weights, and less training time, and still preserves the dynamic characteristics, which allow the DRNN model to be used for on-line application. The performance of this control structure is illustrated through simulations with example system.