Date of Award
Doctor of Philosophy (PhD)
Dr. Hoda A. ElMaraghy
This thesis is devoted to investigating adaptive neurocontrol of nonlinear systems with uncertain or unknown dynamic models. Novel theoretical synthesis and analysis of neurocontrol systems have been conducted, and applied to the control of flexible joint robots with experimental tests. The contributions of this thesis fall into the following three areas: (1) neural networks, (2) adaptive neurocontrol and (3) control of flexible joint robots.
The aim of my research in the neural network area is to search for fast and global convergent learning algorithms with reduced computation burden. The localized neural networks with competitive lateral inhibitory cells were introduced. The developed extended Kalman filtering algorithm with UD factorization can make the localized polynomial networks and localized pi-sigma networks possess fast learning convergence and less computation. The multi-step localized adaptive learning algorithm was derived for RBF networks which leads to about 10 fold improvement in the speed of learning convergence. New neural network models of nonlinear systems were introduced to facilitate neurocontroller design.
In the adaptive neurocontrol area, theoretical issues of the existing backprop-based adaptive neurocontrol schemes were first clarified. Then new direct and indirect adaptive neurocontrol schemes, with better performance, were proposed. It is noticed that the system stability of many existing neurocontrol schemes cannot be proved. In addition, few stability-based adaptive neurocontrol schemes are available and can only be applied to feedback linearizable nonlinear systems. The thesis provides two major contributions to the stability-based adaptive neurocontrol approach. The first contribution is extending the classical self-tuning control methodologies for linear systems to the self-tuning neurocontrol of nonlinear systems by using localized neural networks. This extension greatly enriches the neurocontrol algorithms with guaranteed system stability. The second contribution is proposing the variable index control approach, which is of great significance in the control field, and applying it to derive new stable robust adaptive neurocontrol schemes. Those new schemes possess inherent robustness to system model uncertainty, which is not required to satisfy any matching condition. They do not impose any growth condition and infinite differentiability assumption on the system nonlinearity. They can also be applied to nonlinear systems which are not feedback-linearizable.
As applications and extensions of the above theory, three different robust adaptive neurocontrol schemes for general flexible joint robots were derived with proven system stability. All three schemes are able to incorporate a priori information about the robot dynamics into the neurocontroller design to simplify the neural network design. No acceleration and jerk signals are required in these control laws. Moreover, arbitrary joint stiffness is allowed in the control algorithms.
To demonstrate the feasibility of the proposed learning algorithms and adaptive neurocontrol schemes, intensive computer simulations were conducted based on different nonlinear systems and functions. Different types of adaptive tracking problems and regulation problems were considered. Furthermore, the proposed adaptive neurocontrol schemes were experimentally tested using an existing experimental flexible joint robot. Both the simulation and experimental results confirm the practicability of the proposed schemes.
The thesis concludes that the neurocontrol approach, along with the development of neural computers and large scale parallel distributed processors, is capable of solving the complex control problem of nonlinear systems with uncertain or unknown dynamic models.
Liang, Feng, "Adaptive Neurocontrol and Its Application to Robots" (1995). Open Access Dissertations and Theses. Paper 1769.