IDENTIFICATION CONTROL SYSTEM BASED RECURRENT NEURAL NETWORKS FOR LARGE RADIO TELESCOPE SERVO DRIVE

The identification model control system using artiﬁcial neural networks to estimate the angle main mirror in azimuth moving of large radio telescope electrical servo drive is proposed. An architectural approach to design recurrent neural networks based on «Nonlinear Auto Regressive with Exogenous inputs – NARX models» in form of the state space is analyzed in this paper. It is convenient to application this design in the field of nonlinear prediction and modeling. The dynamic back-propagation algorithm through time, Bayesian regularization and Levenberg–Marquardt optimization algorithm are proposed for training purposes. These algorithms are extensions of the standard back-propagation algorithm and allow compute the gradients to converge much faster and better than the standard ones. During computer simulation, the performances of neural network identification model with different parameters are compared in the working of electric drives control systems of large radio telescope. Its performances depend on the value of the forecast horizon MPC-Laguerre regulator, the size of the layer, the training function and the length of tapped delays line of neural networks. In this work, designing, training, testing of neural networks identification model are implemented on Matlab/SIMULINK environment.

Authors: M. P. Belov, T. H. Phuong, N. C. Truong

Direction: Electrical Engineering

Keywords: Large radio telescope, system identification, recurrent neural networks, NARX models, servo drive system

View full article