Abstract

In the thesis, motivated by the well-known universal approximation capability (input-output mapping) of the neural network (NNs), we have proposed adaptive NN controllers for a Rigid Link Flexible Joint (RLFJ) robot manipulator with unknown nonlinearities and piezoelectric actuator with unknown hysteresis, respectively. For a RLFJ robot manipulator, the dynamic model is decomposed into two different time scale models by using integral manifold method. The control torque consists of two terms: slow and fast terms for two time scale models. A composite NN-based control strategy is proposed for the position and velocity tracking of the manipulator. Two multilayer NNs are used to approximate two unknown nonlinear functions. These two NNs are tuned on-line without any off-line training. The stabilities of composite control system have been proved. The boundedness of NN weights and control signal of systems are guaranteed. Simulation results verify the developed control algorithms. The feedforward multilayer NN is also further investigated to approach the complicated nonlinear function in proposed hysteresis dynamics, which is described by Duhem model. An adaptive NN compensator is designed for unknown hysteresis in a piezoelectric actuator. A pre-inverse hysteresis function is well-structured and the effect of the actuator hysteresis is cancelled. The simulation results are also presented to show the effectiveness of the developed adaptive control scheme