Publishing house Radiotekhnika

"Publishing house Radiotekhnika":
scientific and technical literature.
Books and journals of publishing houses: IPRZHR, RS-PRESS, SCIENCE-PRESS

Тел.: +7 (495) 625-9241


Checking and diagnosis of complex failuries in gas-turbine engine control system with use of reccurent neural network


S. V. Zhernakov, S. V. Коbylev

In this paper authors propose the implementation of an active ES for complex monitoring and diagnosis of GTE’s control system based on recurrent neural networks (NN) on example of diagnosis and control of parameters of the acceleration automat of GTE. The aim of the problem is the investigation of algorithms for constructing the neural network classifier, discriminate the faulty state (defect) in the GTE and its regulator up to the node, as well as the development of an appropriate engineering techniques to achieve these objectives. The questions of appropriate formalization of the problem in neural basis,obtaining the source data (constructing a training sample), the choice of architecture, structure and learning algorithm of neural networks, decision-making procedures of the type and location of failure, results of performance evaluation and testing of the trained recurrence of neural networks are considered in this paper. The decision-making procedure is considered on example of diagnosis of the defect associated with a decrease in compressor efficiency and low pressure turbine at 5%. Main approvals of article are confirmed with experimental data, that reflected in the tables and illustrations. The following conclusions are formulated on the basis of research results: 1. Solution of the problem is the using of recurrent NN, where inputs are its current and delayed options, and the optimal size of the time window is chosen at range 8-12, while the number of neurons in the hidden layer is 75 – 90. 2. It was shown, that the smallest error NN training is provided using the conjugate gradient algorithm and quasi-Newton algorithm. 3. Analysis of the quality classification of the technical state of aviation GTE and its control using a recurrent neural network in this example did not exceed 1% of the test sample. 4. Neural network technology can be effectively used for solving the problems of diagnosing the technical state of aviation GTE and its regulator. They allow to work both with real data obtained for the individual and the reference (average) aircraft engine and the controller and the data calculated with the help of it’s mathematical model. The decision-making about the nature and location of a defect may be based on a comparison of this data. 5. In contrast to the classical methods of diagnosis, based on the calculation GTE’s parameters with using of elementwise nonlinear engine models, the implementation of the neural network method of diagnosis based on neural network learning on small training set with the checking the quality of obtained neural network model on the test sample. 6. The comparative analysis showed the possibility of solving the problems of diagnosing the state of GTE and controller with different architectures of neural networks (recurrent, temporary and Elman’s NN), and the method of least squares, providing high reliability of detection of defects, including double defects in different nodes of aircraft engine. 7. Checking the performance of neural networks in terms of the additive (random) disturbances showed a high quality of diagnosis of GTE and its regulator and robustness of this method in relation to the distortion of the input data. 8. Decision-making about location and reason of the defect in the real GTE and its controller can be based on a comparison of neuroclassifier’s output, which constructed on the basis of a recurrent perceptron, with plenty of precedents by using the nearest neighbor rule. Largest metric (the distance to the nearest precedent) The intensity and multiplicity of the defect may be judged by metric’s value, that’s equal to the distance to the nearest precedent.

© Издательство «РАДИОТЕХНИКА», 2004-2017            Тел.: (495) 625-9241                   Designed by [SWAP]Studio