Neural control of non-stationary objects based on detection methods of adaptation discomfiture

DOI №______

Abstract

The article deals with the methods of automatic control adapted to the non-stationary behavior of the object. It is noted that in the process of functioning possible random destructive effects that change the characteristics of system elements. These processes can sometimes be foreseen, but it's always difficult to describe precisely. The conclusion is drawn on the importance of the development of methods of automatic control, adapted to the unsteady behavior of the object. An approach with constant adaptation of the neural network regulator in which the neural network is adapted to changes in the dynamic characteristics of the object is investigated. In the scheme of constant adaptation used neural networks of direct distribution. It is noted that when changing the dynamic properties of the object of management requires adaptation of both the neural network regulator and the neural network identifier. The method can be applied to the case of smooth non-stationary, that is, when changes in object parameters occur gradually and slowly. It is concluded that the method of constant adaptation is rather uneconomical, since the learning algorithm of both neural networks is constantly on. It is suggested to use the method of adaptation to detect the disorder. The difference between the method is the presence of a block of detection of changes in the properties of the control object. The task of determining the dissociation is solved using the algorithm of cumulative sums, and for the reliability of the detection of dissociation, a pair of triggers is taken within the calculated average delay time. As a parameter that is well defined for changing the parameters of the control object, the variance of the identification error is used. After detecting the disruption, the necessary data is collected and the neural network identifier is trained outside the control circuit. An algorithm is proposed that allows to dynamically form the training set for the reliability of the neural network approximation of an unknown function, which presupposes the behavior of the object of management. This training set is used when setting the NMO outside the control loop.

Keywords: automatic control, neural network regulator, adaptation, disruption, identification errors.

References
1. Werbos P. J. Beyond regression: New tools for prediction and analysis in the behavioral sciences. PhD Thesis, Harvard University, Cambridge, MA. – 1974.
2. Narendra K.S., Parthasarathy K. Identification and control of dynamical system using neural networks // IEEE Trans. Neural Networks. – 1990. – Vol.1. – ʋ1. – P. 4-27.
3. Круглов В.В. Нечеткая логика и искусственные нейронные сети / В.В.Круглов, М.И.Дли, Р.Ю. Голунов. М., 2001.
4. Орловський І.А. Використання нейроконтролерів в електромеханічних системах / Орловський І.А., Горбань Є.І. // Радіоелектроніка, інформатика, управління. – 2014. – № 1. – С. 177-188.
5. Головко В. А. Нейронные сети: обучение, организация и применение. Нейрокомпьютеры и их применение. М.: ИПРЖР, 2001.
6. Комашинский В. И., Смирнов Д. А. Нейронные сети и их применение в системах управления и связи. М.: Горячая линия–Телеком, 2003.
7. Кузнецов Б.И. Синтез нейросетевого регулятора NARMA-L2 CONTROLLER для системы наведения и стабилизации. / Кузнецов Б.И., Василец Т.Е., Варфоломеев А.А. // Електротехніка і Електромеханіка. – 2011. – № 4. – С.41-46.
8. Стеклов В.К., Беркман Л.Н., Карпенко Н.Ф. Многокритериальная оптимизация системы управления телекоммуникационными сетями // Зв’язок. – 1999. – № 6. – С. 13 – 16. 9. Стеклов В.К., Беркман Л.Н. Проектування телекомунікаційних мереж Підруч. для студ. вищ. навч. закл. за напрямком «Телекомунікації». — К.: Техніка, 2002. — 792 с.