=Paper=
{{Paper
|id=Vol-3664/paper9
|storemode=property
|title=Neuro-Fuzzy Methods for Detecting Sensor Failures in Helicopters Turboshaft Engines
|pdfUrl=https://ceur-ws.org/Vol-3664/paper9.pdf
|volume=Vol-3664
|authors=Victoria Vysotska,Serhii Vladov,Ruslan Yakovliev,Alexey Yurko,Andrii Voronin
|dblpUrl=https://dblp.org/rec/conf/colins/VysotskaVYYV24
}}
==Neuro-Fuzzy Methods for Detecting Sensor Failures in Helicopters Turboshaft Engines==
https://ceur-ws.org/Vol-3664/paper9.pdf
Neuro-Fuzzy Methods for Detecting Sensor Failures in
Helicopters Turboshaft Engines
Victoria Vysotska1, Serhii Vladov2, Ruslan Yakovliev2, Alexey Yurko3 and Andrii Voronin4
1 Lviv Polytechnic National University, Stepan Bandera Street 12, Lviv, 79013, Ukraine
2 Kremenchuk Flight College of Kharkiv National University of Internal Affairs, Peremohy Street 17/6, Kremenchuk,
39605, Ukraine
3 Kremenchuk Mykhailo Ostrohradskyi National University, University Street 20, Kremenchuk, 39600, Ukraine
4 Ivan Kozhedub Kharkiv National Air Force University, Sumska Street 77/79, Kharkiv, 61023, Ukraine
Abstract
This work is devoted to the development of a neural network method for detecting failures of sensors
of helicopters turboshaft engines under on-board operation conditions. The proposed method is based
on the use of the ANFIS neuro-fuzzy network with a modified hybrid method for its training. A
modification of the hybrid method of training the neuro-fuzzy network ANFIS is proposed, which,
through the use of the Adam method as a gradient-based optimization algorithm, as well as the adaptive
k-means clustering method for optimizing the shape of fuzzy membership functions, allowed reducing
the number of training epochs from 400 to 50 to obtain the minimum the standard deviation of the
training error is 2.646 · 10–4 using the Gaussian membership function. An evolutionary system of fuzzy
rules has been developed to determine the gas temperature in front of the compressor turbine sensor
failure, the compressor defect, and the failure of the free turbine rotor speed sensor failure. The
proposed system can be extended by adding new fuzzy rules in order to detect and identify other failures
of sensors and components of helicopters turboshaft engines. An experiment was carried out, which
consists of computer modeling of the gradual failure of a gas temperature sensor in front of a
compressor turbine. The results of a comparative analysis of traditional and neural network methods
for detecting failures in helicopters turboshaft engines sensors showed that the maximum errors of the
first and second types when using neural network methods did not exceed 0.78 and 0.52 %, while for
traditional methods they reached 2.48 and 1.91 %.
Keywords
Helicopters turboshaft engines, neuro-fuzzy network ANFIS, sensor failure, gas temperature in front of
the compressor turbine, mathematical model, error, training1
1. Introduction
The movement control of modern helicopters has to be ensured under conditions of significant
and varied uncertainties in the values of their parameters and characteristics, flight modes, and
environmental influences. In addition, during the flight, various emergency situations may arise,
in particular, engine failures and structural damage [1].
Some of these failures and damage have a direct impact on the dynamic characteristics of the
helicopter as a control object. At the same time, it is extremely difficult to foresee all possible
failures and their combinations in advance. From the above it follows that the situation in which
the helicopter finds itself at any given moment in time can change in a significant and
unpredictable way.
In this regard, it seems appropriate from a management point of view to interpret possible
sudden changes in the dynamic properties of helicopters turboshaft engines (TE) due to failures
and damage as another class of uncertainty factors, the countering of which is assigned to
adaptation mechanisms. They must provide fault-tolerant control, that is, control that is able to
COLINS-2024: 8th International Conference on Computational Linguistics and Intelligent Systems, April 12–13, 2024,
Lviv, Ukraine
victoria.a.vysotska@lpnu.ua (V. Vysotska); serhii.vladov@univd.edu.ua (S. Vladov); director.klk.hnuvs@gmail.com
(R. Yakovliev); yurkoalexe@gmail.com (A. Yurko); n_voronina77@ukr.net (A. Voronin)
0000-0001-6417-3689 (V. Vysotska); 0000-0001-8009-5254 (S. Vladov); 0000-0002-3788-2583 (R. Yakovliev);
0000-0002-8244-2376 (A. Yurko); 0009-0007-6448-5878 (A. Voronin)
© 2024 Copyright for this paper by its authors.
Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
ceur-ws.org
Workshop ISSN 1613-0073
Proceedings
adapt to changes in the dynamics of the control object generated by failures or damage, providing
acceptable quality of control [2, 3].
With this approach, the task of providing fault-tolerant control is divided into two parts. The
first of them is related to the reconfiguration of helicopters TE control algorithms when a failure
situation occurs. But in addition to reconfiguration, it is necessary to simultaneously solve the
task of identifying a failure situation, its nature and source of occurrence. Helicopters TE sensors
failures are a serious problem, since the information received from them is used to control the
movement of the helicopter.
The use of classical failure detection methods to solve the task under consideration is
associated with a number of difficulties caused by the nonlinearity of the models, inaccuracies in
measuring the outputs of the control object, and the large amount of data used. In addition,
classical methods work satisfactorily only for sufficiently large values of the signal-to-noise ratio,
and also have high computational complexity. It is also significant that the use of classical
identification methods usually involves linearization and significant simplification of the system
model, which does not always correspond to the nature of the task being solved [4, 5].
Neural network methods [6] are one of the promising approaches to providing fault-tolerant
control [7, 8]. Neural network tools can overcome many of these disadvantages [9, 10]. In
particular, as the available research results show, neuro-fuzzy networks can provide an effective
solution to identification tasks [11, 12].
Based on the above, an urgent scientific and practical task is the development of neuro-fuzzy
methods for helicopters TE sensors failures detecting, since through timely detection of failures
it is possible to prevent their development and the occurrence of emergency situations. In
addition, a neural network failure classifier can help reduce helicopter downtime by more
accurately and quickly diagnosing failures, as well as complement traditional monitoring and
diagnostic methods, increasing their accuracy and efficiency.
2. Related works
The contemporary digital control system for the helicopters TE manages engine operation across
all modes, maintaining stability during transitions and averting emergencies (Fig. 1). Comprising
three key components – a parameter measurement control unit, an onboard monitoring and
diagnostic system, and an automatic control system [13] – it guarantees smooth and safe engine
performance.
Helicopter
Executive
turboshaft Sensors
mechanisms
engine
Control system for
measured parameters
System for control,
monitoring and Automatic control
system
diagnostic
Figure 1: The overarching framework of the digital control system designed for helicopters TE [13]
At present, failures within specified thresholds are detected over time, triggering a failure
assessment when these limits are exceeded. In instances of measuring channel failure, the last
dependable parameter value is utilized to restore lost data [14, 15]. However, this method proves
ineffective in cases of gradual or intermittent failures, especially during engine transitions,
leading to low accuracy in recovered data [16, 17]. Addressing this challenge necessitates
augmenting traditional monitoring and diagnostic techniques for helicopters TE with intelligent
methods, which exhibit superior efficacy across all operational modes [18]. Among these
methods, hybrid intelligent algorithms, combining various intelligent techniques, alongside
neural networks and fuzzy logic algorithms, present promising avenues [19]. Consequently, the
objective of this research is to develop an intelligent system utilizing a neural network
mathematical model in tandem with a neuro-fuzzy method to address this issue [20].
In connection with the above, the purpose of the article is to develop a neural network failure
classifier for helicopters TE. The following questions will be considered in the work:
1. Selection of neural network architecture. Within the framework of this task, various neural
network architectures will be considered and the choice of the most suitable one for the task of
failure classification will be justified.
2. Training a neural network classifier. As part of this task, methods and algorithms for
training a neural network classifier will be described.
3. Evaluating the effectiveness of the neural network classifier. As part of this task, the results
of assessing the effectiveness of a neural network classifier on test data will be presented.
The construction of such a model can be considered as the introduction of analytical
redundancy of critical elements. As with the introduction of physical redundancy, the location of
a faulty sensor is determined using a voting scheme. With this approach, the consequences of a
failure situation can be countered by replacing the readings of a faulty sensor element with the
output of its model.
3. Methods and materials
Traditionally, failure detection includes two main stages: identifying an abnormal situation, as
well as determining its location and symptoms [21, 22]. The implementation of these stages can
be interpreted as a sequential solution to the task of identifying a dynamic system and classifying
the signs of a failure situation. The paper proposes a failure detection algorithm that combines
solutions to these two tasks using neural network methods using neuro-fuzzy networks.
The implementation of the first stage is a typical task of monitoring a control object and
measuring its outputs. A decision on the occurrence of a failure is made by comparing the current
and predicted phase states of the dynamic system. If deviations reach a certain level, then a
solution to the task of classifying failure signs is required. To obtain predicted phase states, a
solution to the task of identifying the control object is required.
The neural network model makes it possible to assess the state of the control object at each
moment in time, therefore in the proposed algorithm it is used at the stage of identifying an
emergency situation for both groups of failures. Each of the failure groups has its own impact on
the dynamics of helicopters TE, therefore, the proposed algorithm uses classification methods
specific to each of the failure groups.
To address the aforementioned issue, an intelligent system can be employed, utilizing the FDI
(Fault Detection and Identification) technique. This method relies on a neural network
mathematical model of the engine in conjunction with a neuro-fuzzy classifier [20, 23]. By
implementing this proposed intelligent system, it becomes feasible to detect and categorize
abnormal operational states of a helicopters TE, as well as anomalies in measurement channels
and actuators, all within onboard conditions (Fig. 2).
Neural
Ym
network S1
model Neuro-fuzzy
U ε failure S2
...
Helicopter classifier
turboshaft Y Sn
engine
Figure 2: Helicopters TE control and diagnostic system configuration
Helicopters TE mathematical model plays the role of a reference model as part of the on-board
control and diagnostic system. Comparing the calculated data of the mathematical model with the
data of the measuring channels allows you to track changes in the controlled object. In addition,
this model can be used to restore data in a failed measuring channel. A mathematical model must
have a number of qualities, the most important of which are the following [24]: the model
describes the non-stationary nature of the work processes of a helicopters TE (thus, the use of a
dynamic model is necessary); the structure of the mathematical model of the helicopters TE
provides the practical possibility of its functioning in combination with mathematical models of
helicopter other elements.
The mathematical model, described in [25], is tailored for determining the specific fuel
consumption of helicopters TE installed in helicopters, using the TV3-117 engine of the Mi-8MTV
helicopter as an illustrative example. According to the insights provided in [23], the specific fuel
consumption of such engines depends on factors like air intake in the combustion chamber,
specific engine output, and the ratio of fuel to air consumption within the chamber, with the
choice of aviation fuel exerting a direct influence.
The computation of helicopters TE thermogas-dynamic parameters, encompassing air intake
in the combustion chamber, specific engine output, and the fuel-to-air consumption ratio within
the chamber, is accomplished through a neural network model specifically designed for
helicopters TE. This model, developed by the author and elaborated upon in [26, 27], is utilized
(Fig. 3). The author performs the creation and setup of a trial version of the helicopters
mathematical model using the Neural Network Toolbox, an extension package within the
MATLAB environment.
Figure 3: A fragment of the mathematical model for helicopters TE within the Matlab/Simulink
environment, where 11 thermogas-dynamic parameters governing the engine's operational
dynamics are calculated [26, 27]
One promising avenue in this domain involves crafting a mathematical model grounded in
neural networks, renowned for their capacity to train and generalize accumulated knowledge.
This feature facilitates the adjustment of model parameters to suit the characteristics of specific
engines, leveraging data derived from both bench and flight tests. Recurrent neural networks,
such as Elman networks, Jordan networks, and multilayer perceptron with general feedback
(NARX), fulfill these requirements for the mathematical model [28, 29].
Despite the advantages of recurrent neural networks like Elman, Jordan, and NARX networks,
they also pose certain drawbacks. Chief among them is the challenge of training and fine-tuning
network parameters, particularly when confronted with vast datasets or intricate nonlinear
relations between input and output data. Achieving an acceptable level of model performance
under such circumstances may demand significant time and computational resources. Moreover,
recurrent neural networks may encounter issues like decay and gradient explosion when trained
on lengthy sequences of data, leading to difficulties in consistent training and a decline in the
network's generalization ability to new data.
To address these limitations, ongoing research is pivoting towards the adoption of neuro-fuzzy
networks. These systems amalgamate the strengths of neural networks and fuzzy logic, rendering
them more adaptable and flexible across diverse datasets and conditions. Neuro-fuzzy systems
possess the capability to autonomously adapt to variations in input data and environmental
factors, making them well-suited for modeling complex systems like engines, which contend with
highly variable operating conditions [30, 31].
Approaches grounded in neuro-fuzzy networks enable the incorporation of uncertainty and
fuzziness in data, a crucial aspect when dealing with real-world data susceptible to noise and
errors. This adaptive capacity allows neuro-fuzzy networks to more effectively accommodate
diverse operating conditions, furnishing more precise predictions and engine control.
Hence, the shift towards utilizing neuro-fuzzy networks represents a promising trajectory for
advancing the mathematical modeling of helicopters TE, facilitating more efficient and accurate
control and monitoring of their operations.
Fig. 4 shows the structure of a neural network model of a helicopter TE construct on a five-
layer feed-forward network, in contrast to [20, 23], where it was proposed to use a multilayer
recurrent perceptron (NARX). The adaptive neuro-fuzzy network (inference system) ANFIS
(Adaptive Network-based Fuzzy Inference System) is a hybrid multilayer artificial neural
network of a special structure without feedback [32]. The values of the inputs, outputs and
synaptic weights of the hybrid neural network are real numbers on the interval [0, 1]. The
adaptive network ANFIS in its functions is analogous to a fuzzy inference system [33]. The ANFIS
network uses a hybrid training algorithm. Neurons in the ANFIS network have different
structures and purposes, corresponding to the fuzzy inference system and implementing the
main stages of its operation [34]:
• Fuzzification (introduction of fuzziness) using membership functions of input variables –
the first layer of neurons of the network (layer 1);
• Aggregation (determining the degree of truth of conditions) by processing a base of fuzzy
linguistic rules – the second layer of neurons in the network (layer 2);
• Activation (determining the degrees of truth of statements) by normalizing the activation
levels of fuzzy rules – the third layer of neurons in the network (layer 3);
• Accumulation (combination of degrees of truth) using membership functions of output
variables – the fourth layer of neurons in the network (layer 4);
• Defuzzification (transition to clarity) with obtaining a clear value of the output variable –
the fifth layer of neurons in the network (layer 5).
The first adaptive layer of the ANFIS network contains neurons that calculate the values of the
membership functions of input variables μi(GT) and μj(nTC), where GT and nTC are input variables,
i = 1, 2 and j = 3, 4. The adaptability of the layer is achieved by selecting type of membership
functions of input variables.
The second fixed layer of the ANFIS network contains neurons that calculate the products of
the values of the membership functions obtained on the first layer:
wi = i ( GT ) j ( nTC ) , (1)
where wi is the network synaptic weights.
The third fixed layer of the ANFIS network contains neurons that calculate normalized
activation levels of fuzzy rules:
wi
waverage _ i = . (2)
w1 + w2 + w3 + w4
The fourth adaptive layer of the ANFIS network contains neurons that calculate the values of
the membership functions of the output variables, as well as the product of the values of synaptic
weights and membership functions:
waverage _ i i = waverage _ i i ( GT , nTC , i , i , i ) , (3)
where i is the output variables membership functions values, αi, βi, γi are the parameters of the
membership functions. The adaptability of the layer is achieved by selecting the type of
membership functions of the output variables.
The fifth fixed layer of the ANFIS network contains a neuron that calculates the sum of the
products of the values of the membership functions of the output variables and synaptic weights
Qi = waverage _ i i .
μ1
GT (t) wi waverage_i
μ11(GT )
П N
μ12(nTC )
μ2
μ21(GT )
П N Σ Qi
μ22(nTC )
Q1 -TG (t)
μ3 Q2 - nTC (t)
μ31(GT ) Q3 - nFT (t)
nTC (t-1) П N
μ32(nTC )
GT (t)
nTC (t-1)
Figure 4: Helicopters TE neural network model structure in the form of adaptive neuro-fuzzy
network (inference system) ANFIS
As an algorithm for training the adaptive neuro-fuzzy network ANFIS, an algorithm consisting
of two stages is proposed [35]:
• first stage (algorithm direct course): we set the initial values of the parameters of the first
adaptive layer, perform calculations on the second and third layers, determine the parameters of
the fourth adaptive layer and calculate the value of the error function. If the value of the error
function is within acceptable limits, then training of the adaptive neuro-fuzzy network ANFIS is
completed, otherwise we proceed to the second stage;
• second stage (reverse algorithm): using the backpropagation method, we refine the
parameters of the first adaptive layer.
At the same time, to adjust the parameters of the neuro-fuzzy network ANFIS, instead of the
least square’s method, it is proposed to use a more effective optimization algorithm based on
gradients, for example, the Adam method [36]. To optimize the shape of fuzzy membership
functions, it is proposed to use an adaptive training method, for example, the k-means clustering
method. Thus, the use of a gradient-based optimization algorithm allows you to more accurately
adjust the parameters of the ANFIS neuro-fuzzy network, and the use of the k-means clustering
method allows you to reduce its training time.
For the mathematical description of the proposed modifications of the hybrid algorithm for
training the neuro-fuzzy network ANFIS, the following notations are introduced: x is the vector
of input data (GT, nTC), y is the vector of output data, w is the vector of weights of the neuro-fuzzy
network, μi is the fuzzy membership function of the i-th rule, fi is the output function of the i-th
rule, N is the number of rules, α is the training parameter, η is the regularization parameter. The
proposed modification of the hybrid algorithm for training the neuro-fuzzy network ANFIS
consists of two stages:
1. Gradient-based optimization algorithm:
1.1. Initialization of weights of the neuro-fuzzy network ANFIS w.
1.2. Calculation of the gradient of the loss function L(w) by weights w.
1.3. Update weights w:
w = w − L ( w) , (4)
where w represents the parameters of a model that we're optimizing; α is the training rate, a small
positive scalar that determines the step size in each iteration; ∇L(w) is the gradient of the loss
function L(w) with respect to the parameters w. The gradient points in the direction of the
steepest increase of the function.
1.4. Repeat steps 1.1–1.3 until stopping criterion is reached.
2. Adaptive teaching method:
2.1. Clustering training data into N clusters using the k-means method.
2.2. For each cluster i, the centroid of the cluster ci is determined, and the parameters of the
fuzzy membership function μi are also initialized.
2.3. Training of the ANFIS neuro-fuzzy network with fixed parameters of fuzzy membership
functions.
2.4. Update parameters of fuzzy membership functions:
x − ci 2
i ( x ) = exp − , (5)
2 2
where ci is the i-th center (centroid) in the feature space; σ is the smoothing parameter that
controls the width of the Gaussian function curve; x − ci is the square of the distance between
2
the input vector x and the center ci.
2.5. Repeat steps 2.3–2.4 until stopping criterion is reached.
The root mean square error can be used as the loss function L(w):
(1 N
)
L ( w) = yi − y i ,
2
(6)
2 i =1
where y i is the output of the ANFIS neuro-fuzzy network for the i-th example.
To prevent overfitting, Tikhonov regularization can be used:
M 2
1 N
(
i i 2 )
2
L ( w) = y − y + wj , (7)
2 i =1 j =1
where M is the number of weights of the ANFIS neuro-fuzzy network; η is the regularization
coefficient, which controls the importance of regularization in relation to the error.
4. Experiment
Table 1 presents a segment of the expert knowledge matrix for the neuro-fuzzy network ANFIS
designed for helicopters TE. The general fuzzy rule with serial number k has the form: IF GT(t)
and nTC(t – 1) THEN {1, 2, 3}