Neural velocity observer trained with experimental data supporting stabilization of magnetically levitating sphere

This elaboration presents the method of obtaining experimental data for the neural network training process. A family of data and the initial conditions were obtained by the stabilization experiment of the levitating sphere and by a short-term control. On the basis of these data, a structure of the feedback network was developed. The nonlinear autoregresive neural network realizes the task of the velocity estimation of the levitating sphere. The structure of the neural network was assessed and then applied in real time to the stabilization task. The results of experiments were compared with two different methods of a velocity estimation.