ПОХИБКИ ПРОГНОЗУВАННЯ ЗНАЧЕННЯ ТЕМПЕРАТУРИ НЕЙРОННИМИ МЕРЕЖАМИ ЗА ІДЕАЛЬНИМ ПЕРЕХІДНИМ ПРОЦЕСОМ
Journal Title: Вимірювальна техніка та метрологія - Year 2017, Vol 1, Issue 78
Abstract
The present article describes the results of the study of the prediction error of temperature values using neural networks. In the introduction, the authors point out problems that arise (come up) during the measurement of high temperatures. The method proposed to solve these problems is neural networks application. At the very beginning the authors present a neural networks classification based on their architecture (feedforward neural networks, recurrent neural networks and completely linked neural networks were specially highlighted). Also mentioned previous researches where were made conclusions about the most relevant neural network architecture in case of temperature prediction problem using transition process. The studies described in the article are implemented in the MATLAB computing environment. An algorithm for creating and teaching neural networks was described. Sequences modeling for the neural network training, the functions using for neural network creation and studding, the formula for calculating the absolute error of temperature prediction were given. During the sequences creation, the measurement error was not taken into account, that is, the network studied on ideal sequences. The results of the study of dependence of the temperature value prediction error on the number of layers in the network, on the number of network inputs and on the number of sequences for training are presented. Investigation of the dependence of the temperature prediction error on the number of network inputs was carried out for two cases: when the time of transition process temperature measurement is the same and when the measurement time is different. In addition, the neural network was tested on sequences that coincided and did not coincide with the sequences on which the neural network studied. Each research was provided with drawings. At the end of the article the authors make conclusions about the most relevant neural network parameters (number of layers, number of inputs and the number of sequences for training neural network). Maximum temperature prediction error value was mentioned. Plans for further research were also outlined.
Authors and Affiliations
Ольга Лопатко, Ігор Микитин
Keywords
Related Articles
- EP ID EP280300
- DOI -
- Views 68
- Downloads 0
ОПТИМІЗАЦІЯ ПАРАМЕТРІВ ПРУЖНИХ ЕЛЕМЕНТІВ ТЕНЗОРЕЗИСТИВНИХ ВИМІРЮВАЧІВ ТИСКУ ЗА ДИНАМІЧНИМИ ХАРАКТЕРИСТИКАМИ
On based of analysis of dynamics of pressure piezoresistive sensors elastic elements the method of calculation and optimization of parameters of these elements from the point of view minimum of dynamic errors minimum at...
МІКРОЕЛЕКТРОМЕХАНІЧНІ СИСТЕМИ В СУЧАСНОМУ ПРИЛАДОБУДУВАННІ
The state of the microsystem technology as perspective direction of science and technique development is analyzed; description of the microelectromechanical systems basic parts and description of structural features and...
ВПЛИВ НЕВИЗНАЧЕНОСТЕЙ ФІЗИЧНИХ ВЛАСТИВОСТЕЙ ГАЗОПОДІБНИХ ЕНЕРГОНОСІЇВ НА ТОЧНІСТЬ ЇХ ОБЛІКУ
The analysis of the influence of uncertainties of gas physical properties on the uncertainty of the result of flow rate measurement by the method of differential pressure is made. The requirements to the precision of the...
ВИКОРИСТАННЯ ЗАСОБІВ МОВИ HTML ДЛЯ ВДОСКОНАЛЕННЯ НОРМАТИВНО-ТЕХНІЧНОГО ЗАБЕЗПЕЧЕННЯ ТУРИЗМУ
The informative reference system for perfection of the normatively technical providing of tourism is created. The mathematical model of optimization of subject of site is offered.
ПРОБЛЕМНО-ОРІЄНТОВАНИЙ АНАЛІЗАТОР ІМПЕДАНСУ НА ОСНОВІ СУЧАСНОЇ КОМПЛЕКТУЮЧОЇ БАЗИ
In the paper practical implementation of measuring system based on integral impedance converter AD5933is described. Experimental measurements results are presented. Prospects of presented impedance converter usage in por...