Hybrid controller with recurrent neural network for magnetic levitation system

Abstract

We propose a hybrid controller using a recurrent neural network (RNN) to control a levitated object in a magnetic levitation system. We describe a nonlinear dynamic model of the system and propose a computed force controller, based on feedback linearization, to control the position of the levitated object. To relax the requirement of the lumped uncertainty in the design of the computed force controller, an RNN functions as an uncertainty observer to adapt the lumped uncertainty on line. The computed force controller, the RNN uncertainty observer, and a compensated controller are embodied in a hybrid controller, which is based on Lyapunov stability. The computed force controller, with the RNN uncertainty observer, is the main tracking controller, and the compensated controller compensates the minimum approximation error of the RNN uncertainty observer. To ensure the convergence of the RNN, the adaptation law of the RNN is modified by using a projection algorithm. Experimental results illustrate the validity of the proposed control design for the magnetic levitation system.