MULTI-MODE MODEL IDENTIFICATION OF HELICOPTERS AIRCRAFT ENGINES IN FLIGHT MODES USING A MODIFIED GRADIENT ALGORITHM FOR TRAINING RADIAL-BASIC NEURAL NETWORKS

Abstract

This work is devoted to solving the applied problem of identification helicopters aircraft gas turbine engines in flight modes using their multi-mode models using the classical method – least squares method and the neural network method – by constructing a neural network in accordance with the initial data. The following methods are used: methods of probability theory and mathematical statistics, methods of neuroinformatics, methods of information systems theory and data processing. To achieve this goal and reduce the identification error of aircraft gas turbine engine multi-mode model, the use of radial-basis functions neural network with a modified gradient training algorithm is proposed, which consists in dynamically changing the structure of the neural network in the learning process, and to exclude situations when the parameters of the elements are close to each other. to a friend, the coefficient of mutual intersection of elements is introduced. When solving the applied problem of identification helicopters aircraft gas turbine engines, it was shown that the error in identifying a multi-mode model of helicopters aircraft gas turbine (using the example of the TV3-117 aircraft engine) using a perceptron when calculating individual engine parameters did not exceed 0.63 %; for radial-basis functions neural network – 0.74 %, for radial-basis functions neural network with a modified gradient learning algorithm – 0.47 %, while for the classical method (least squares method) it is about 1% in the considered the range of change of engine operating modes. Comparative analysis of neural network and classical identification methods under noise action shows that neural network methods are more robust to external disturbances: for a noise level σ = 0.025, the error in identifying parameters of an aircraft engine TV3-117 when using a perceptron increases from 0.63 to 0.84%; for radial-basis functions neural network – from 0.74 to 0.86 %; for radial basis functions neural network with a modified gradient learning algorithm – from 0.47 to 0.65 %, and for the least squares method – from 0.99 to 2.14 %.