Experimental comparative studies on neural network controllers for DC motor micromaneuvering

The performance attributes of a structure-free and a model-based neural network (NN) controller for DC motor micromaneuvering purposes are compared in experimental studies. The former NN has a generic structure independent of the friction model, where its input vector consists of the time history of the motor angular shaft velocity within a time window. The NN provides the nonlinear control mapping to the supplied motor input through adjustment of its weights using the sign gradient descent algorithm. The model-based NN has a predetermined structure that depends on the utilized friction model. This NN provides a feedforward term which compensates for the inherent friction while rejecting noise via an additional linear velocity error feedback term. Application of both NN-based controllers on a DC motor system reveals that the model-based NN has a superior performance measured in terms of its: (a) convergence, (b) implementation computational requirements, and (c) integral square error after the weights´ convergence