Use of neural networks to identify and compensate for friction in precision, position controlled mechanisms

A special neural network topology has been developed that compensates for friction in precision, position controlled mechanisms. A major contribution is that knowledge of the friction´s form is used to determine the neural network´s structure. This unique approach solves network sizing and weight initializing problems. The friction model is used for feedforward decoupling of friction-induced torque. The neural network also explicitly incorporates inertia compensation and linear feedback control. Another contribution is a demonstration of the trajectory dependence of static friction compensation with a discrete time controller. The authors include both the theoretical formulation and practical implementation results for the control of a commercial DC motor having a significant amount of static friction.<>