Advances and Applications in Discrete Mathematics
Volume 19, Issue 4, Pages 359 - 371
(October 2018) http://dx.doi.org/10.17654/AADMOct2018_359_371 |
|
TRAINING OF NEURAL NETWORK BASED PWM CONTROLLERS
N. N. Kucyi, A. V. Lukyanov, S. K. Kargapol’cev and Tikhii
|
Abstract: One can see the growing interest in the use of pulse elements formed on the basis of artificial neural networks (ANNs), in particular, to systems with pulse-width modulation (PWM), in the synthesis of automatic control systems (AСSs). This raises the problem of parametric optimization of synaptic weights of the neural network with respect to the values of the selected quality criterion of AСS. In addition, a disturbance in the form of a change in the volumetric efficiency processed by the extruder electric drive can occur. Of course, such perturbations have a direct impact on the rate of pulling of cable products for the entire length of the extrusion line, but due to the presence of other disturbances, coming from the electric drive haul-off, maintenance of the set pulling speed not always provide the desired insulation thickness of the finished cable product with the required accuracy. To solve the problem of parametric optimization of neural network controller of AСS for thick plastic insulation, we propose an algorithm for neural network training (NNT), which allows to solve this problem using different quality criteria. Parameters of NNT are presented, and the methods used in forming the initial simplex algorithm itself are briefly described. Next, we have solved the problem of parametric optimization with respect to the quadratic integral criterion. The optimal adjustment of neural network controllers using different activation functions of neurons and their associated transients has been obtained. A comparative analysis of the results was carried out. It is shown that the use of logistic activation function allows to cope with the appearing disturbances efficiently. |
Keywords and phrases: artificial neural network, cascade control system, latitude-but-width modulation, the training of the neural network, Nelder-Mead method.
|
|
Number of Downloads: 350 | Number of Views: 4726 |
|