Application of Narx and LSTM artificial neural networks in the problem of automatic control of a magnetic levitation system

When designing automated control systems for complex technical dynamic objects, a mathematical apparatus based on artificial neural networks is used, which has unique advantages: - the possibility of parallel computing; - finding previously unknown relationships between input and output sequences of digital signals; - providing more efficient control of nonlinear systems through the use of nonlinear activation functions. In addition, they sometimes remove the difficulties that arise when describing some problems in the form of analytical mathematical models. Neurocontrol is one of the promising areas, located at the intersection of theories of automatic control systems and artificial intelligence. This paper considers the problem of using neural network identification in the system of automated control of magnetic levitation (neural network controller). Method. The solution of the task is based on the methods of artificial intelligence - the well-known recurrent artificial neural networks NARX and LSTM. These networks were trained using the backpropagation algorithm and the Levenberg-Marquardt method, which have good convergence. When training artificial neural networks, it is necessary to take into account the effect of overfitting, which can lead to poor results. The MATLAB system was a tool for setting the architecture of neural networks, their construction, training and testing. Main results. The application of trained artificial neural networks to test data showed for this problem some advantage of the NARX network compared to LSTM. Moreover, the root mean squared error (RMSE) for the NARX network with 50 hidden layers for this problem is less than for the network with 100 hidden layers. Hence follows the recommendation to use the NARX artificial neural network in solving the problems of designing automated control systems for electromagnetic levitating objects. Practical significance. The results obtained can be applied in the design of automatic control systems for levitating objects. Currently, most of these systems are being developed for the transportation of goods for various purposes using the levitation effect.