A neural network approach for generating derivative information from quantized position measurements

In the control of robot and other mechanized systems, there exists a need for the generation of first and higher order derivative information. Not much effort has been made to address this issue. Instead much of the work, especially in robot control, has relied on a priori knowledge of the system dynamics to determine the derivative information. The goal of this thesis is to propose a generalized methodology to determine first and higher order derivative information without any a priori knowledge of a system's dynamics. For illustration purposes the system that we will concentrate on is a flexible joint robot manipulator. In achieving our objective, we impose three restrictions on the resulting methodology. First, the acquisition of the derivative information required for the control of a robot manipulator must be done without assuming knowledge of the robot's joint position or its dynamics. Second, we must not use any additional hardware than is absolutely necessary. Thus, for our flexible manipulator system, we must only use a sensor to measure the position variable. The derivative information (velocity, acceleration and jerk) must be determined by the proposed methodology. Third, the derivative information must be determined in real time. We accomplish the objectives using Tap Delay Neural Networks (TDNN). We incorporate the proposed methodology into the control system of a flexible joint manipulator. Numerous simulation results show the success of our proposed scheme within robot control loops. Additionally, we compare the performance of a number of alternatives within this scheme with that using conventional linear filtering techniques. Again, we see the superiority of the proposed TDNN scheme over the linear filtering approac