Turbofan engine diagnostics neuron network size optimization method which takes into account overlaerning effect
DOI:
https://doi.org/10.18372/2306-1472.42.1813Keywords:
amount of neurons, diagnostics, gas turbine engine, neuron network, optimization of structureAbstract
The article is devoted to the problem of gas turbine engine (GTE) technical state class automatic recognition with operation parameters by neuron networks. The one of main problems for creation the neuron networks is determination of their optimal structures size (amount of layers in network and count of neurons in each layer).The method of neuron network size optimization intended for classification of GTE technical state is considered in the article. Optimization is cared out with taking into account of overlearning effect possibility when a learning network loses property of generalization and begins strictly describing educational data set. To determinate a moment when overlearning effect is appeared in learning neuron network the method of three data sets is used. The method is based on the comparison of recognition quality parameters changes which were calculated during recognition of educational and control data sets. As the moment when network overlearning effect is appeared the moment when control data set recognition quality begins deteriorating but educational data set recognition quality continues still improving is used. To determinate this moment learning process periodically is terminated and simulation of network with education and control data sets is fulfilled. The optimization of two-, three- and four-layer networks is conducted and some results of optimization are shown. Also the extended educational set is created and shown. The set describes 16 GTE technical state classes and each class is represented with 200 points (200 possible technical state class realizations) instead of 20 points using in the former articles. It was done to increase representativeness of data set.In the article the algorithm of optimization is considered and some results which were obtained with it are shown. The results of experiments were analyzed to determinate most optimal neuron network structure. This structure provides most high-quality GTE technical state classification and high level of network generalization.
References
Кучер О.Г. Розпізнавання технічного стану авіаційних ГТД з використанням нейронних мереж / О.Г. Кучер, О.С. Якушенко, В.Ю. Сухоруков // Вісник двигунобудування. – Запоріжжя: ЗНТУ, ХАІ, МоторСіч, 2002. − №1. – С. 101–106.
Якушенко О.С. Нейронні мережі для діагностування газотурбінних двигунів/ О.С. Якушенко // Вісник НАУ. – 2004. – № 2 (20). – С. 67–71.
Многокритериальное диагностирование проточной части авиационных газотурбинных двигателей/ А.Г. Кучер, С.А. Дмитриев, А.В. Попов, А.С. Якушенко // Вісник двигунобудування. – Запоріжжя: ОАО Мотор-Січ, 2009. – № 3. – С. 166−171.
Первинна оптимізація архітектури нейронної мережі/ А.Г. Кучер, О.С. Якушенко, В.М. Степаненко, В.Є. Мільцов // Вісник двигунобудування. – Запоріжжя: ОАО Мотор-Січ. – 2009. – № 3. – С. 166−171.
Нейросетевое распознавание технического состояния авиационных газотурбинных двигателей/ А.М. Пашаев, П.Ш. Абдуллаев, А.С. Якушенко, А.Д. Мирзоев // Изв. Нац. акад. наук Азербайджана. Информатика и проблемы управления. – 2008. – № 3 (28). – С. 3–7.
Медведев В.С. Нейронные сети. Matlab 6 / В.С. Медведев, В.Г. Потемкин. / под общ. ред. В.Г. Потемкина. − М.: ДИАЛОГ-МИФИ, 2002. − 496 с.
Двигатель ПС-90А. Руководство по технической эксплуатации. в 3 кн. – Пермь: ПНПО "Авиадвигатель", 1990.
