The article discusses the use of expert systems with neural network knowledge bases for the process of helicopters turboshaft engine parameters at flight modes information monitoring. The method for determining the trend of aircraft gas turbine engine parameters based on neural networks, adapted to helicopters turboshaft engines that implement the dynamic knowledge base of an expert system during engine operation, has been further developed. A modified Jordan neural network with dynamic stack memory has been developed, which, through the use of dynamic stack memory, makes it possible to detect the appearance of a trend in the parameters of helicopters turboshaft engines at flight modes, increase the accuracy to 0.999 and reduce the trend recognition error to 0.056. A hybrid algorithm consisting of adaptive and genetic training of recurrent neural networks has been improved, adapted to Jordan modified neural network with dynamic stack memory, which made it possible to optimize its training process in relation to the problem of helicopters turboshaft engines at flight modes parameters trend recognizing. The modified method for determining the trend of parameters has been tested in the onboard neural network expert system for the integrated monitoring of helicopters turboshaft engines operational status. network, trend analysis Aircraft engine, expert system, neural network, operational status, modified Jordan neural
The process of monitoring and aircraft gas turbine engines (GTE) operation control is carried out by expert systems (ES) [1, 2]. The use of it allows taking into account a number of factors that contribute to the qualitative improvement of their functioning [3, 4]. The presence of close information interaction of the control system (aircraft GTE) with the environment using specially organized information communication channels; fundamental openness of the system in order to increase its intelligence and improve its own behavior. The presence of mechanisms for predicting changes in the environment and the behavior of the system; development a control system based on a multi-level hierarchical structure. It satisfies the following rule: as the rank of the hierarchy increases, the intelligence of the system increases and the requirements for its accuracy decrease, and vice versa; persistence of functioning in case of partial rupture of links or loss of control actions from higher levels of the hierarchy of the control system. Despite the weighty evidence base of the relevance of the use of ES in aircraft GTE on-board monitoring systems [5, 6], there are currently no such ES in relation to helicopters turboshaft engines (TE).
Thus, an urgent scientific and practical task is the development of an on-board ES for monitoring helicopters TE operational status, which, in the process of monitoring and their operation control, is able to fully control the parameters, analyze (simulate) the current situation with a predict of its development in the engine (information from sensors).
2022 Copyright for this paper by its authors.
One of the classic tasks of monitoring the parameters of helicopters TE is the disorder (determining the trend of controlled data). In the general case, trend analysis allows you to control the time series formed by the sequence of values of controlled indicators, and determine the presence of a trend: changes (disorder) in this series. The value of trend analysis in modern active ES is very high, as it allows you to identify defects at an early stage of their development (even if the values of the controlled parameters are within acceptable limits).
In connection with the foregoing, the purpose of this work is to develop and modify methods for determining the trend of controlled data (trend analysis) of helicopters TE, which will be part of monitoring helicopters TE operational status onboard ES.
Under the current conditions, along with the traditional classification of ES: static and dynamic, recently, in domestic and foreign periodicals, a class has been distinguished – active ES [7, 8]. From dynamic ES (real-time ES – RT ES), active ES differ in the participation of the human factor in the control loop. So, if in RT ES the share of a human operator in the decision-making process can be 30... 50 %, then in active ES this percentage of participation is reduced to a minimum of 5...10 % or completely eliminated [9].
It is quite obvious that with such an approach to the organization of active ES and its implementation in the management process, it is necessary to take into account a number of factors that contribute to the qualitative improvement of its functioning [10, 11]:
– presence of close information interaction of the control system with the environment using specially organized information communication channels;
– fundamental openness of the system in order to increase its intelligence and improve its own behavior;
– presence of mechanisms for predicting changes in the environment and the behavior of the system;
– construction of a control system based on a multilevel hierarchical structure that satisfies the following rule: as the rank of the hierarchy increases, the intelligence of the system increases and the requirements for its accuracy decrease, and vice versa;
– persistence of functioning in case of partial breakage of connections or loss of control actions from the highest levels of the hierarchy of the control system.
Thus, the developed ES should be easily reconfigurable (adaptable) to external changes, for which it requires the presence of the following subordinate levels [12]: training, self-organization (restructuring), predicting (recognition) of events (situations), working with event databases (databases) (DB) and knowledge bases (KB), decision making (DM); planning operations for the implementation of the formed solution, adaptation.
The listed levels form the strategic level of the active ES (fig. 1), the rest perform its tactical functions. The solver (logical output machine) of an active ES is complex. It is with known methods and knowledge (logic of predicates, semantic networks, frames, production output). The methods based on soft computing (fuzzy logic (FL), genetic algorithms (GA), neural networks (NN), cognitive networks (CN), probabilistic output (PO) – heuristics). The combination of methods and their extension with elements of soft computing increases the mobility of the computational process by the active ES solver and as a result, the quality of the decisions making with the connection to the planning of banks of algorithms and models. Having a powerful solver, the active ES is relatively easy to adapt to an external dynamic model, allowing you to set and solve direct, inverse and mixed problems. In the process of monitoring and operation control of helicopters turboshaft engines at flight modes, the ES is able to fully control the parameters, analyze (simulate) the current situation with a predict of its development in the engine (information from sensors).
Active ES knowledge bases store declarative and procedural knowledge. Procedural ones include conceptual knowledge bases (CKB): concepts in the form of formulas, dependencies, tables, procedures, etc. Declarative ones include expert knowledge bases (EKB) that are descriptive (qualitative) in nature. At the same time, CKB and EKB closely interact with each other, constantly checking for consistency (redundancy) of knowledge. In the process of interaction with the object and its own heterogeneous KB, the active ES performs training and self-training. Real-time scanning tests facts and knowledge. The new situation "forms" a precedent and is stored in the knowledge base. Elements of traditional modeling tools in the active ES carry out mathematical (simulation) modeling of the engine, as well as storage of a priori and a posteriori data in the active ES database (initial information and test results). Additional "flexibility" and mobility of the knowledge base in the active ES is achieved by pairing the artificial intelligence models and the mathematical model (MM) of the engine [11, 12].
Dynamic Knowledge Base Neural Neural Neural network network ... network
1 2 n
User Interface
...
...
During the operation of helicopter TE, the active ES connected to it allows real-time modeling, forecasting and evaluation of the efficiency of helicopter power plant.
It is assumed that х(t), t = 1…N is a sequence of discrete observations х(t) = f(t) + ξ(t) in the background of interference ξ(t) with zero mean and variance σ2. A set of polynomials are used as trend models according to [13]
j1 f t csjt s , j = 1…N;
s0 with unknown coefficients csj, where j – model type index.
With the current estimation, model (
j1 t s
f j t t f js t ; (
s0 s!
where Δt – time counted from the present moment of time. f js t – s-th derivative of the function
fj(t), the values of which are determined by the sliding x(t – N + 1), x(t – N + 2), ..., x(t) observations
sample of a constant volume N, which makes it possible to track the change in the coefficients csj of
the model (
The task of this work research is to construct a neural network detector (dynamic knowledge base
of an active ES), which allows, as a result of the observations x(t) processing, to establish the facts of
violation of the regularity and the time of these violations (trends) appearance.
(
– assessment of the belonging of two samples to the same general population (non-parametric methods that do not require a priori information, classical criteria: Hald-Abbe and its modifications [14]); – trend analysis of controlled parameters based on neural networks.
An important task in the process of analyzing experimental data, which reflects the recognition of the “appearance” of helicopters TE, is to determine the discord in the measured parameters of the time series, i.e. analysis of statistical characteristics of the results of registration of controlled parameters in order to determine their stationarity. The main task of trend analysis is to identify regularity in a sequence of data. The most complete description of trend detection methods is given at [14], among which the most common "classical" methods of trend analysis are: parametric, non-parametric and mixed methods. Of the parametric methods, according to [13], an integral criterion is used, which consists in the following sequence of operations:
– preliminary processing of the numerical series (measurement data) {Y1, ..., YN} is carried out in order to convert it to a form convenient for subsequent evaluation;
– logic and physics of the process is analyzed, which has a significant impact on both the choice of the type of approximating function and the determination of the limits of its parameters.
Preliminary processing of the initial number series within the time interval T t1,...,tN is aimed at reducing the influence of the random component ε(t) in the initial number series {Y1, ..., YN} (i.e., bringing it closer to the trend). The presentation of the information contained in the numerical series in such a way as to significantly reduce the difficulties of the analytical description of the trend.
The main methods for solving these problems are the procedures for smoothing and leveling the statistical series. In this case, the smoothing procedure is aimed at minimizing random deviations of points from some smooth curve of the assumed process trend. Smoothing is performed using polynomials that approximate groups of points measured during the experiment using the least squares method (LSM). Even in a simple linear version, the smoothing procedure is very effective in identifying a trend when superimposed on an empirical numerical series of random interference and measurement errors. If smoothing is aimed at the primary processing of a number series to eliminate random fluctuations and identify a trend, then alignment serves the purpose of a more convenient presentation of the original series while maintaining its values. In the simplest case, this procedure can be carried out by approximating the initial series of processed experimental points.
The choice as the criterion of optimality of the measure of the deviation of the points of the empirical series from the approximating function is carried out according to the expression (LSM): where Yj – points of the empirical series (measured values); η – approximating function; tj – time component; α1…αN – approximated points.
According to [13], the following functionals are used as one of the integral criteria for evaluating the trend:
N 2
Yj t j ,1,..., N min;
j1
N Yj Yn j
j1 Yn j
;
(
The application of this criterion (
Another integral criterion for evaluating the trend is a functional of the form [13]:
(
The application of criterion (
In the process of studying the trend by classical methods, it can be concluded that the "classic" integral criteria are very effective in express analysis, have accuracy, clarity and are able to determine with a high degree of certainty the moment the trend begins to appear. However, in order to apply these criteria in the onboard (expert) system for monitoring helicopters TE, it is necessary to develop an appropriate method.
Among the numerous methods of trend analysis, the following are noted [14, 15]: the linear filtering method, the Kalman filter, extrapolation methods, which are most simply implemented in a neural network basis, since they are based on smoothing and equalizing statistical series procedures.
Smoothing and equalization procedures can be implemented on the basis of neural networks in the form of two series-connected filters – low frequency (LF) and high frequency (HF), while in [13] their implementation based on recurrent neural networks was proposed. In this case, the low-pass filter “passes” the constant component fj(t) and filters the noise ξ(t), and the high-pass filter passes f js t and filters fj(t) and the noise ξ(t). Implementation of low-pass and high-pass filters based on recurrent neural networks is shown in fig. 3. These options differ in that they are implemented by the corresponding external filters. The structure of the external filter is shown in fig. 4.
It is known from the theory of neural networks [16, 17] those static architectures of neural networks are capable of approximating multidimensional, non-linear static functions. The identification of dynamic systems, on the other hand, requires a model with appropriate storage elements. Therefore, static full-sized neural networks should be extended with dynamic structures. One of the possibilities of dynamic extension is the addition of external filters that implement a dynamic model outside the network. Such neural networks with external dynamics include [18, 19]: – non-linear models with output feedback; – non-linear models with finite impulse response; – non-linear orthogonal models of basic functions.
During preprocessing, it is considered that the functions f(t) and ξ(t) are not correlated. It is required that the vector Outx(t) of the output values of the filter Outlx t , l 1, N , which is a response to an external influence, approaches the desired function of the useful signal: where F = (Fl) is some vector operator that describes the mapping of the set of useful signals into the output signals of the filter [20, 21].
As a measure of approximation Outx(t) to Ff(t) in the general case, one can choose the functional:
Out x t Ff t ;
J J Ff t Out x t ; where φ[•] – some measure of a vector function.
In the simplest case [20, 21] (fig. 2), the input signal is applied to a set of serially connected functional elements with a delay Z–1 (in synapses). Their input values are represented as signals Inx(t – kZ-1), k = 1, N with weights Wjk, forming a vector of estimates of useful signals x j t , on the basis of which, with the help of a network that implements the matrix of operators (Flj), a vector of output signals is formed:
Outlx t Flj Wjk Inx t kZ 1 . (
The task of filtering is to reproduce the useful signal against the background of noise and perform
the required transformation. To solve this problem, it is necessary to minimize the standard deviation
of the estimate of the useful signal xj(t) from the expected j – useful signal fj(t), characterizing the
corresponding useful result of the neural network filter, i.e. find:
min M f j t Wjk Inx t kZ 1 2 ; (
Wjk j k where M – mathematical expectation.
According to this criterion, classical filter adaptation algorithms can be implemented using a priori information about the useful signal and noise.
To solve the given problem, Professor Serhii Zhernakov [22] proposed to use a perceptron as a dynamic (recurrent) neural network that implements a low frequency filter; for high frequency filter – RBF (radial basis function) neural network. Ensemble neural network learning algorithm – complex back propagation.
A signal having N samples x = (x1, …, xn) can be approximated by a neural network with G neurons in the hidden layer by the following equations: – for perceptron: – for RBF:
G f t Wi0q W ihT
t ; i0 f t Wi0 Ri t,W ih ;
G
i0
(
In order to carry out helicopters TE parameters trend analysis at flight modes, that is, in real time, it is proposed to use a recurrent neural network with a dynamic stack memory, built on the basis of the Jordan neural network, the basis of which is a multilayer perceptron. Feedback is implemented by supplying to the input layer not only output data, but also network output signals with a delay of one or more cycles, which allows you to take into account the background history of the observed processes and accumulate information to develop the correct control strategy. Jordan’s modified neural network is obtained by adding a delay to the feedback signals of the hidden layer by several cycles, that is, by adding a dynamic stack memory to the layer [23, 24].
The outputs of the hidden layer c1, c2, …, ck are fed to the input neurons with weighting coefficients {wij}–t, where i – index of the neuron to which the signal is given (i = 1, 2, …, n); j – index of the output signal of the hidden layer neuron (j = 1, 2, ..., k); t – time delay index (t = 1, 2, ..., m). We will change the number of time delays from 1 to m. Thus, the Elman network is obtained at m = 1, and the multilayer perceptron is obtained at m = 0. A detailed examination of the architecture of the neural network (fig. 5) shows that the inverse of the hidden layer or the output of the network can be excluded by adding signals to the training sample feedback.
Jordan modified neural network is described by a system of recurrent equations: or in matrix form:
p N vj n 1 wj1iui n wcji yi n yi n 1 bj1 ;
i1 i1 x j n 1 F1 v j n 1; j = 1, 2, … , N;
N y j n 1 F2 wji2 xi n 1 bj2 ; j = 1, 2, … , M;
i1 Xn 1 F1 W1Un Wc Yn Yn 1 B1 ;
Y n 1 F2 W2Xn 1 B2 ;
where U(n) – vector of external input signals at time n; p – number of external network inputs; X(n + 1)
– vector of the output signals of the hidden layer at the moment of time (n + 1); N – number of signals
in the context layer. W(
The decision rule for a modified Jordan neural network with dynamic stack memory implementing low-pass and high-pass filters is as follows [13, 22]:
N f j1t f j t j1
2 t C; where the numerator of expression (17) means the accumulation of the sum of deviations of the controlled parameters (C – activation threshold (sensitivity) of the neural network; when C = 0 (normal operation mode), when α ≥ C (trend)).
Consider the process of transformation of the training sample to solve the problem of trend analysis of the time series of the gas temperature parameter in front of the compressor turbine (fig. 2a) using a modified Jordan neural network with a dynamic stack memory (table 1).
It is assumed that in modified Jordan neural network, the hidden layer contains three neurons, the output contains one neuron, the dynamic memory stack contains feedback signals of the hidden layer with a delay of two cycles. Since the number of neurons of the hidden layer with feedback to the input layer is three, the size of the input vector when training the modified Jordan neural network by remembering the previous output signal one step back will increase by three, by remembering two previous output signals – by six. We denote the input signals of the training sample that change during the transformation as x1, x2, x3, and the feedback signals as x4, x5, x6, x7, x8, x9. We transform the time series (for example, the values of the integral criterion, which are in the range from 1.375 h to 2.0 h) using the sliding window method (table 2). (12) (13) (14) (15) (16) (17)
Inputs x4, x5, x6 – fed output signals of the hidden layer with a delay of one cycle с1-1, c2-1, c3-1, inputs x7, x8, x9 – output signals of the hidden layer with a delay of two cycles с1-2, c2-2, c3-2. In fig. 6, the layer memory is represented as a stack of layer output signals y1, y2, …, yn, where n – size of the stack.
Dynamic memory Temporary
dela (-1) Temporary delay (-2) Temporary delay (-m) {wij}-1 {wij}-2 {wij}-m + + + 1 1 1
F2 c1 F2 F2 y1(n+1) yM(n+1) c2
yN(n+1) ck
Outputs u1(n) u2(n) up(n)
External outputs ...
y1 y2 yn y r o m e m r e y a L
Context layer
z-1 1
The algorithm and software implementation of the hybrid training method for recurrent neural networks were developed by the authors Hryhorii Bieliavskyi, Volodymyr Lila, Yevhenii Puchkov [25]. The result of their work is to obtain a parametric training model for recurrent neural networks that contains a wide range of known training algorithms (adaptive and genetic algorithms) and allows you to adjust the parameters for the best solution to the problem of time series analysis. In this work, a modification of the developed algorithm was carried out, which consists in clarifying the choice of the best individuals from the population, as well as introducing a correction factor into the crossing procedure, which takes into account the probability of a stochastic change in engine parameters under current operating conditions (table 3). To conduct an experiment, namely, training a neural network, committees of 10 modified Jordan neural networks were built using each algorithm. The number of neurons in the hidden layer is 9 (according to the Kolmogorov-Arnold-Hecht-Nielsen theorem). The criterion for stopping training of the modified Jordan neural network is the root mean square error with a value of 0.001, the step size ηk = 0.05, training epoch threshold – 1000. Examples from the training sample, which was formed according to fig. 2, a, were submitted by chance. Sigmoidal 1 ex 1 ex functions of the form f x
were used as the neuron activation function. Training the neural network with an adaptive algorithm (table 4) until the transition criterion to the genetic training method is reached Creation of a population of N–1 individuals. A neural network trained by an adaptive algorithm is added to the first population Individuals are crossed with the probability of choosing a pair Pc, while each pair produces S offspring. Determination of the genes of the descendant is made according to the expression: Gi D Random?0.5?Gia : Gi ; b where Gi – gene of the new chromosome; Ga and Gb – genes of parental chromosomes; i – serial number of the gene in the chromosome; Random – function that generates a uniform random value on the segment of real numbers [0; 1]; D – correction factor for introducing randomness into the value of the gene, the introduction of which is a modification of the existing algorithm, which makes it possible to take into account the “instant” change at helicopter flight engine parameter Selection of the best N individuals from the new population, giving the smallest recognition error, determined according to the expression: y0 where y0 – output signal reference value; y – output signal value when recognizing the reference value of the thermos-gas-dynamic parameter of the engine from the training sample with a given set of weight coefficients (output signal reference value in this problem is the value of engine thermos-gas-dynamic parameter in the absence of defects under normal operating conditions).
If the best representative of the individual corresponds to the given quality of training, the transition to step 9 is performed.
A mutation is carried out for individuals selected with a probability Pm. For each gene of the selected individual with probability Pg, the gene is mutated according to the expression:
Gi Gi Gi Km 2 Random 1; where Gi – chromosome gene; i – serial number of the gene in the chromosome; Random – function that generates a uniform random value on the segment of real numbers [0; 1]; Km – mutation coefficient (as a rule, Km∈[0; 1]).
If the best representative of the individual corresponds to the specified training quality, go to step 9, if not, return to step 4.
Finish Finding the initial values of the parameters: w0 – starting point, p0 – initial direction of movement, η0 – step Choosing the next vector from the training set and feeding it to the input of the neural network Determining the direction of movement pk according to the expression:
mink1,m pk g k i g ki ;
i1 where pk – direction of movement; g j – direction of the anti-gradient at the j-th iteration; βi – coefficient that determines the weight of the i-th gradient; m – number of memorized gradients; k – serial number of the current iteration.
Stop criteria computation – root mean square error
If the stop condition is met, go to step 6, if not, go to step 2 Step 1
According to the results of the comparison, the adaptive algorithm converges faster than the genetic and hybrid ones. The threshold of 0.001 for 1000 epochs has not been overcome by any gradient method. From this it follows that it is most expedient to use an adaptive method at the beginning of training, which quickly finds a solution with a root-mean-square error of 0.003, and then apply the genetic algorithm. In terms of the number of epochs, the genetic algorithm is in some cases faster than the hybrid one. But for this task, the time of one epoch of the genetic algorithm is much higher (200 milliseconds) than the average time of one epoch of the improved hybrid one (140 milliseconds). Therefore, we can conclude that the hybrid algorithm of the “improved adaptive + genetic” option is best suited for this task. Fig. 10 shows the diagram of the change in the mean b Number of epochs 550 640 400 1000 580 930 960 1000 980 1000 square error function according to the number of training epochs for "improved adaptive + genetic" algorithm.
The statistics of the approximation of the result to the reference value by the improved hybrid algorithm in the problem of trend recognition is shown in fig. 11, which shows the values of the output signal of the best chromosome in the current epoch. Experiments have shown that in order to achieve the best recognition result with the resulting set of weights, the crossover probability is Pc = 0.64, Pm = 0.01.
We will analyze the presence of a trend (fig. 2, a) in sections I and II using the neural network apparatus. The implementation of low-pass and high-pass filters based on a modified Jordan neural network with dynamic stack memory is shown in fig. 8. Analysis of the trend in the first section is shown in fig. 9. The determination of the trend of the neural network on the second characteristic area is shown in fig. 10. At the same time, one cell corresponds to half an hour of operation of the helicopter’s control system. It can be seen that the emergence of a trend is noticed by the neural network after the sixth cell. To complicate the process of trend recognition and to get as close to the real situation as possible, an obstacle is "superimposed" on the input signal identified by the neural network.
In the process of mathematical modeling on a modified Jordan neural network with a dynamic stack memory that implements recurrent filters, in comparison with the classic criteria for detecting the trend of parameters. For example, of the TV3-117 engine, the results shown in fig. 11, where 1 – neural network criterion (using a modified Jordan neural network with dynamic stack memory); 2 – neural network criterion (using an ensemble of neural networks consisting of a perceptron and an RBF network); 3 – scriterion; 4 – S'-criterion; 5 – Halden-Abbe r-criterion; 6 – modified r-criterion; 7 – u-criterion [14]. a a b b b
Figure 10: Determination of the trend using a modified Jordan neural network knowledge base: a – signal at the filter input; b – signal at the filter output
In the work, a comparative assessment of the effectiveness of trend analysis of neural network and classical criteria was carried out. A comparative study of the criteria was carried out on the basis of simulation modeling, which made it possible to check a wide range of changes in measurement errors and the intensity of trend manifestation. The value of the controlled parameter is equal to the sum of the deterministic basis and random normally distributed interference with variance ξ. The deterministic component is constant in the interval [0, t0], and then changes linearly with the rate a = tg(α) (1/s) (where α – intensity of the trend). During the simulation, the value of a varied in the range [0.01; 1]; and the value of ξ is in the range [0.001; 1]. During modeling, a sample variance calculated on the stationarity interval [0, t0] was used to adjust the mathematical model of the helicopter TE. Starting from the moment t0, the values of the criteria were calculated and the presence of a trend was checked. The effectiveness of the criteria was evaluated by the time of activation of the criteria from the beginning of the trend τ0 to the moment of time corresponding to the detection of the trend τlate [13, 22].
The results of numerical modeling (table 7) indicate the possibility of solving the problems of information monitoring of helicopter aircraft engines operational status, which allow, along with the classical criteria for detecting the trend of parameters, to apply qualitatively new neural network criteria that expand and complement the classical criteria. That increases the reliability of information at helicopters TE operational status monitoring and at the stages of decision-making.
Similar research was conducted using other architectures of neural recurrent networks as implementing low-pass and high-pass filters (table 8). From the table 8, it can be seen that Jordan’s modified neural network with dynamic stack memory is appropriate for use as an implementation of low-pass and high-pass filters for the purpose of solving the task of helicopters TE parameters trend analysis at flight modes, i.e., in real time.
The technique of helicopters TE parameters complex monitoring at flight modes in a neural network basis:
1. Obtaining a training sample in N modes of a normally operating engine at a real-time rate under normal operating conditions.
2. Obtaining a training sample in N engine modes with a parameter trend in real time at the current operating conditions.
3. Choice of neural network architecture. 4. Choice of training algorithms.
5. Training, testing and real-time trend recognition of engine parameters under current operating conditions.
6. Monitoring of helicopters TE parameters at flight modes by neural networks.
7. Adaptation of neural networks in the environment of an active expert system [27, 28].
Under the conditions of on-board implementation of the developed neural network method for trend analysis of helicopter turboshaft engine parameters, the expediency of using the 64-bit Intel Neural Compute Stick 2 neuro processor using a Python programming language (using the Keras open-source library) was proved in [29] (fig. 12).
Neuro processors of this series are widely used in modern digital automatic control systems, including in aviation. The presence of a multiplier-accumulator module in the core of this microprocessor makes it possible to increase the speed of calculating the algorithm by combining multiplication and addition operations with weighted summation in the neuron adder. 7. Conclusions
1. The neural network method for analyzing aircraft engine parameters trend has been further developed, which, through the use of a modified Jordan neural network with dynamic stack memory, makes it possible to detect the appearance of helicopters turboshaft engines parameters trend at flight modes, improve accuracy to 0.995 and reduce the trend recognition error to 0.056. 2. The hybrid algorithm (adaptive + genetic) for training recurrent neural networks, in particular, the Jordan neural network, has been improved, which, by refining the criterion for choosing the best individuals from the population, as well as introducing a correction factor into the crossing procedure. It takes into account the probability of a stochastic change in engine parameters in the current conditions, allowed to optimize the training of the modified Jordan neural network with dynamic stack memory to solve the task of helicopters turboshaft engines at flight modes parameters trend recognizing.
3. The results of numerical simulation indicate the possibility of solving the problems of integrated monitoring of helicopters turboshaft engines operational status at flight modes based on active expert systems [27, 28], which allow, along with the classical criteria for identification of parameters trend, to apply qualitatively new neural network criteria that expand and supplement the classical criteria. It increases the reliability of information up to 100 % in the control and diagnostics of helicopters turboshaft engines parameters and at the decision-making stages.
4. The technique developed and described in this work has been tested in the environment of active expert systems [27], has shown high efficiency in solving the problems of integrated monitoring of the operational status and operation management (parameters control, diagnostics, debugging and predicting) of helicopters turboshaft engines in flight modes.
5. The operation of aircraft gas turbine engines of the 5th–6th generations, including helicopters turboshaft engines, the complication of their technical systems and subsystems, as well as the increased requirements for flight safety, have led to the need to create intelligent systems capable of performing certain functions of a human expert. To assist in the search of the optimal solution, to issue advice and recommendations in real time in the process of integrated monitoring and operation management of helicopters turboshaft engines. 8. References [12] J. Raffort, C. Adam, L. C. Duy, M. Carrier, F. Lareyre, Application of Artificial Intelligence for Automatic Vascular Segmentation: Development of a Hybrid Method Combining Expert System with Deep Learning, European Journal of Vascular and Endovascular Surgery, vol. 63, issue 2 (2022) e42–e43. doi: 10.1016/j.ejvs.2021.12.033 [13] S. Vladov, L. Pylypenko, N. Tutova, I. Dieriabina, A. Yanitskyi, Control and diagnostics of TV3-117 aircraft engine technical state by analysis of its parameters trend, Visnyk of Kherson National Technical University, no. 1 (76) (2021) 141–154. doi: 10.35546/kntu20784481.2021.1.11 [14] A. Oseni, B. Dzwairo, Trend analysis and artificial neural networks forecasting for rainfall prediction, Environmental Economics, no. 7 (4-1) (2016) 149–160. doi: 10.21511/ee.07(41).2016.07 [15] F. Prado, M. C. Minutolo, W. Kristjanpoller, Forecasting based on an ensemble Autoregressive Moving Average – Adaptive neuro – Fuzzy inference system – Neural network – Genetic Algorithm Framework, Energy, vol. 197 (2020) 117159. doi: 10.1016/j.energy.2020.117159 [16] A. Y. Alanis, N. Arana-Daniel, C. Lopez-Franco, Artificial Neural Networks for Engineering
Applications, London, Academic Press (2019) 176 p. doi: 10.1016/C2018-0-01649-7 [17] S. Herzog, C. Tetzlaff, F. Worgotter, Evolving artificial neural networks with feedback, Neural
Networks, vol. 123 (2020) 153–162. doi: 10.1016/j.neunet.2019.12.004 [18] R. V. Williams-Garcia, S. Nicolis, Chaos, Solitons & Fractals, vol. 165, part 1 (2022) 112739.
doi: 10.1016/j.chaos.2022.112739 [19] H. Robinson, S. Pawar, A. Rasheed, O. San, Physics guided neural networks for modelling of non-linear dynamics, Neural Networks, vol. 154 (2022) 333–345. doi: 10.1016/j.neunet.2022.07.023 [20] X. Lai, S. Tong, G. Zhu, Adaptive fuzzy neural network-aided progressive Gaussian approximate filter for GPS/INS integration navigation, Measurement, vol. 200 (2022) 111641. doi: 10.1016/j.measurement.2022.111641 [21] Z. Yan, T. Guo, A. Zhao, Q. Kong, J. Zhou, Reliable exponential H∞ filtering for a class of switched reaction-diffusion neural networks, Applied Mathematics and Computation, vol. 414 (2022) 126661. doi: 10.1016/j.amc.2021.126661 [22] S. Zhernakov, R. Ravilov, Trend analysis of aircraft GTE parameters based on neural networks technology, Bulletin of USATU, vol. 15, no. 4 (44) (2011) 25–32. [23] A. Wysocki, M. Ławryńczuk, Jordan neural network for modelling and predictive control of dynamic systems, Conference: 2015 20th International Conference on Methods and Models in Automation and Robotics (MMAR), 24–27 August 2015. doi: 10.1109/MMAR.2015.7283862 [24] V. Lila, Ye. Puchkov, Methodology of training recurrent artificial neural network with dynamic stack memory, Software & Systems, no. 4 (108) (2014) 132–135. doi: 10.15827/0236235X.108.132-135 [25] H. Bieliavskyi, V. Lila, Ye. Puchkov, Algorithm and software implementation of the hybrid method training artificial neural network, Software products and systems, no. 4 (2012) 96–100. [26] S. Vladov, Y. Shmelov, R. Yakovliev, Methodology for Control of Helicopters Aircraft Engines Technical State in Flight Modes Using Neural Networks. The Fifth International Workshop on Computer Modeling and Intelligent Systems (CMIS-2022), May, 12, 2022, Zaporizhzhia, Ukraine. CEUR Workshop Proceedings (ISSN 1613-0073), vol. 3137 (2022) 108–125. doi: 10.32782/cmis/3137-10 [27] Y. Shmelov, S. Vladov, Y. Klimova, M. Kirukhina, Expert system for identification of the technical state of the aircraft engine TV3-117 in flight modes, System Analysis & Intelligent Computing : IEEE First International Conference on System Analysis & Intelligent Computing (SAIC), 08–12 October 2018 77–82. doi: 10.1109/SAIC.2018.8516864 [28] S. Vladov, Y. Shmelov, M. Petchenko, A Neuro-Fuzzy Expert System for the Control and Diagnostics of Helicopters Aircraft Engines Technical State. ICTERI 2021: ICT in Education, Research, and Industrial Applications, 28 September – 02 October 2021, Kherson, Ukraine.
CEUR Workshop Proceedings (ISSN 1613-0073), vol. 3013 (2021) 40–52. [29] S. Vladov, K. Kotliarov, S. Hrybanova, O. Husarova, L. Chyzhova, On-board information restoring method in case of failure of one of the sensors of the aircraft engine TV3-117 based on neural network technologies, Transactions of Kremenchuk Mykhailo Ostrohradskyi National University, issue 6/2019 (119) (2019) 91–98. doi: 10.30929/1995-0519.2019.6.91-98