Recurrent neural networks for identification of friction

Abstract

For high accuracy in mechanical control, the compensation of friction must be taken into account. When working at low velocities, the slip-stick phenomenon appears, introducing additive perturbation torque. Mechanical and robotics engineers use several models to simulate these torques. Here we select two dynamic friction laws (Dahl model and Reset-Integrator model), and we use Recurrent Neural Networks to model these dynamic systems in order to complete the a priori knowledge with what we don't know how to modelize. We use the canonical architecture to construct our network, and use a gradient based algorithm to train it. Results show that a general architecture for describing this family of friction laws can be obtained. This type of architecture may be used in regulation schemes.