Neural Network Modeling of a 4-Pole 3-DoF Magnetic Levitation System based on NARX Architecture

3RD INTERNATIONAL EURASIAN CONFERENCE ON SCIENCE, ENGINEERING AND TECHNOLOGY, Ankara, Türkiye, 15 - 17 Aralık 2021, cilt.1, ss.883-886

Generating an accurate simulation for a noisy and highly non-linear system requires deep understanding of

mathematical modeling of uncertainties, noise and physical context. To address this bottleneck, in this work,

Nonlinear Autoregressive exogenous (NARX) recurrent neural network (RNN) modeling is proposed to

approximate the system behavior of a 4-pole 3-DoF magnetic levitation plant controlled via PID. The target

position of the levitator is given to the controller which generates control signals and drives the 4 poles of the

electromagnet to reach the target position. The RNN model is used to map the input target position to the

resulting position of the system, and to predict the levitator position for an unknown target position. To train

the model, the required data is collected by sampling given target positions and resulting levitator positions.

The RNN is trained with that data offline and evaluated by applying the same sequence of certain target

distance values to the real system and to the model and comparing both systems outputs. The results show that

the modelling based on the NARX method is adequate and can be used for simulation studies