Artificial neural network based control on nonlinear systems with application to robotic manipulators

Abstract

This dissertation presents a novel approach to the control of nonlinear dynamic systems with an application to robotic manipulator control. A new nonlinear computational model which is commonly known as an Artificial Neural Network (ANN) is utilized to adaptively estimate the system's nonlinear dynamics. The ANN model utilized has a multilayer feedforward architecture and makes use of a standard backpropagation algorithm for adaptive training of the network weights. An adaptive system identification and control scheme based on ANN models is proposed. A convergence analysis of the adaptive weight estimation is given and the bounds on the output tracking errors are studied based on the functional approximation properties of ANN models. The proposed architecture combined with a servo feedback controller is applied to the trajectory tracking control of rigid robotic manipulators. The scheme basically resembles the feedforward control structure of robotic manipulators, except that the manipulator's inverse dynamics model is replaced with generic ANN models, one per joint. Each ANN model adaptively approximates the corresponding joint's inverse dynamics through repeated number of trials for a specific trajectory following task. In this sense, the method can be considered as a trajectory learning controller. A robustness analysis is also given based on the error dynamics equation. The overall controller architecture is simulated and tested for trajectory tracking tasks of a manipulator. After only a few trial runs satisfactory trajectory tracking performance is observed. The controller is then tested for its adaptation to sudden changes in the manipulator dynamics and shown to perform well under disturbances of parametric variations. An experimental study of the generalization properties of the proposed scheme is also given. The simulation tests clearly demonstrate the efficient use of the proposed controller architecture for tracking control problems.