A real-time control of maglev system using neural networks and genetic algorithms

In this paper a novel method is proposed to control a magnetic levitation system (MAGLEV) based on neural networks and genetic algorithm. The output of the system is restricted due to manufacturing considerations. Structure of identifier is modeled using a multilayer feedforward neural network. Based on the proposed fitness function, genetic algorithm is used to optimally adjust parameters of the neural network. The proposed neuro-identifier system receives the levitated ball positions and system inputs in the previous time step and predicts the next ball position. The identifier is guaranteed to learn the rule that a large (small) system inputs will generate large (small) outputs. As long as the identifier is learned, the system can be controlled. Control signals are frequently updated within equal time intervals, where this proper control signals are calculated using the back propagation mechanism. Whenever an input is applied to the system, the controller starts calculating the next control signal. Simulations are performed in a Delphi 7 environment, where the system equations are solved using Runge-Kutta algorithm. The simulation results present the effectiveness of the proposed methods.