G.G. Kulikov – Dr. Sc. (Eng.), Professor, Department of Automated Control Systems, Ufa State Aviation Technical University
A.I. Abdulnagimov – Ph. D. (Eng.), Associate Professor, Department of Automated Control Systems, Ufa State Aviation Technical University
B.I. Badamshin – Ph. D. (Eng.), Engineer, JSC «Molniya» (Ufa)
In recent decades the intelligent technology based on neural network are widely used in investigation and development of the complex control systems for the gas turbine engines (GTE). There are different directions of research using neural networks (NN) in this field: identification, detection of engines’ operation modes, trend analysis, classification, state prediction, etc. Also the hardware-in-the-loop (HIL) simulation is already used in modelling and testing of automatic control systems by joint use of a real control system (such as FADEC) with a mathematical model of a gas turbine engine and its subsystems, for example, a gas generator, oil, fuel and other systems. However, the problem of adequacy and applicability of the GTE mathematical models (usually represented as piecewise-linear dynamic models) in HIL test-beds is still topical.
The research objective is the effectiveness increase of complex HIL simulation and testing of a real FADEC by using nonlinear dynamic mathematical models of gas turbines and its systems in the form of recurrent neural networks as a part of the HIL test-bed. The paper describes the complete process of modeling and experimental investigation – from designing of a GTE mathematical model in form of neural networks to its testing and debugging on the test-bed. The paper propose the engineering method for construction the recurrent neural network and identification a mathematical model of GTE on a real/flight data, describing the learning algorithm, the network structure and the hidden layer size.
The use the recurrent multilayer perceptron (NARX - nonlinear autoregressive neural network with external input) is proposed for de-signing of nonlinear dynamic model. The mathematical model in the form of NN with a feedback allows taking into account non-linear dynamic characteristics of an engine and ensures the structural-parametric adequate of an analytical GTE model. Two neural networks are trained for the start-up mode and for the land/flight modes, which are combined in one NN model during implementation on the HIL test-bed. One NN model simulates the whole set of GTE parameters. For the training of such networks the Bayesian regularization based on back propagation error algorithm is used, which modifies the values of the weighting factors and offsets in accordance with the algorithm of Levenberg-Markarth. To speed up the training with a large amount of data the procedures of parallel computing, map reduce and memory reduce are used.
The paper shows the structure of the gas turbine engine model, built using LabView System Design Software, and the scheme of its interaction with the HIL test-bed. The developed method of identification and modeling based on the recurrence neural networks considers the engine dynamic performance with compliance of structural and parametric adequacy of an analytical model and allows to get an adequate model of a gas turbine engine for any operational mode.
- Konczevich A.G., Epifanov S.V. Uchet faktora vremeni pri identifikaczii modeli GTD iskusstvennoj nejronnoj set’yu // Aviaczionno-kosmicheskaya texnika i texnologiya. 2006. № 10(36). S. 144−151.
- Zhernakov S.V. Identifikacziya xarakteristik GTD na osnove texnologii nejronny’x setej // Informaczionny’e texnologii. 2010. № 3. S. 39−47.
- Zhernakov S.V., Ravilov R.F. Trend-analiz parametrov aviaczionnogo GTD na osnove texnologii nejronny’x setej // Vestnik UGATU. 2011. T. 15. № 4 (44). S. 25−32.
- Kulikov G.G., Pogorelov G.I., Badamshin B.I., Abdulnagimov A.I. Metodika postroeniya nejrosetevoj modeli dvuxval’nogo GTD s soblyudeniem uslovij strukturnoj adekvatnosti // Aviaczionno-kosmicheskaya texnika i texnologiya. 2014. № 9 (116). S. 68−73.