Оn some problems of neural network technologies in electric components diagnosing
The paper describes an idea of getting electric components diagnostic information and its transformation using the discrete Karhunen-Loeve expansion. Presorting of elements by their physical and technical states is proposed to operate with the MLP, self-organized and RBF- neural networks in the MATL...
Gespeichert in:
| Datum: | 2017 |
|---|---|
| Hauptverfasser: | , , , , , |
| Format: | Artikel |
| Sprache: | English |
| Veröffentlicht: |
Інститут проблем штучного інтелекту МОН України та НАН України
2017
|
| Schriftenreihe: | Штучний інтелект |
| Schlagworte: | |
| Online Zugang: | https://nasplib.isofts.kiev.ua/handle/123456789/162344 |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| Назва журналу: | Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| Zitieren: | Оn some problems of neural network technologies in electric components diagnosing / S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash // Штучний інтелект. — 2017. — № 3-4. — С. 95-104. — Бібліогр.: 10 назв. — англ. |
Institution
Digital Library of Periodicals of National Academy of Sciences of Ukraine| id |
nasplib_isofts_kiev_ua-123456789-162344 |
|---|---|
| record_format |
dspace |
| spelling |
nasplib_isofts_kiev_ua-123456789-1623442025-02-09T21:56:34Z Оn some problems of neural network technologies in electric components diagnosing Про деякі проблеми нейромережевих технологій в діагностиці електричних компонентів Telenyk, S.F. Savchuk, O.V. Pocrovskyi, E.O. Morgal, O.M. Krivenko, K.S. Latash, I.O. Програмно-технічні засоби інтелектуальних систем The paper describes an idea of getting electric components diagnostic information and its transformation using the discrete Karhunen-Loeve expansion. Presorting of elements by their physical and technical states is proposed to operate with the MLP, self-organized and RBF- neural networks in the MATLAB environment. The paper investigates the possibility of using neural network technologies for improving electric components diagnosing by integral effects for increasing reliability of complex technological systems. The statistical and individual classification and presorting of elements according to their physical and technical states for work with the use of neural network technologies is proposed. У статті описується ідея отримання діагностичної інформації про електричні компоненти та її перетворення за допомогою дискретного розкладання Карунена-Лоева. Запропоновано визначення фізичних та технічних станів елементів за допомогою MLP, самоорганізованих та RBF-нейронних мереж в середовищі MATLAB. У роботі досліджується можливість використання нейронних мережевих технологій для поліпшення діагностики електричних компонент за інтегральними ефектами для підвищення надійності складних технологічних систем. Запропоновано статистичну та індивідуальну класифікацію та сортування елементів відповідно до їх фізичних і технічних станів для роботи з використанням нейронних мережевих технологій. 2017 Article Оn some problems of neural network technologies in electric components diagnosing / S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash // Штучний інтелект. — 2017. — № 3-4. — С. 95-104. — Бібліогр.: 10 назв. — англ. 1561-5359 https://nasplib.isofts.kiev.ua/handle/123456789/162344 004.93 en Штучний інтелект application/pdf Інститут проблем штучного інтелекту МОН України та НАН України |
| institution |
Digital Library of Periodicals of National Academy of Sciences of Ukraine |
| collection |
DSpace DC |
| language |
English |
| topic |
Програмно-технічні засоби інтелектуальних систем Програмно-технічні засоби інтелектуальних систем |
| spellingShingle |
Програмно-технічні засоби інтелектуальних систем Програмно-технічні засоби інтелектуальних систем Telenyk, S.F. Savchuk, O.V. Pocrovskyi, E.O. Morgal, O.M. Krivenko, K.S. Latash, I.O. Оn some problems of neural network technologies in electric components diagnosing Штучний інтелект |
| description |
The paper describes an idea of getting electric components diagnostic information and its transformation using the discrete Karhunen-Loeve expansion. Presorting of elements by their physical and technical states is proposed to operate with the MLP, self-organized and RBF- neural networks in the MATLAB environment. The paper investigates the possibility of using neural network technologies for improving electric components diagnosing by integral effects for increasing reliability of complex technological systems. The statistical and individual classification and presorting of elements according to their physical and technical states for work with the use of neural network technologies is proposed. |
| format |
Article |
| author |
Telenyk, S.F. Savchuk, O.V. Pocrovskyi, E.O. Morgal, O.M. Krivenko, K.S. Latash, I.O. |
| author_facet |
Telenyk, S.F. Savchuk, O.V. Pocrovskyi, E.O. Morgal, O.M. Krivenko, K.S. Latash, I.O. |
| author_sort |
Telenyk, S.F. |
| title |
Оn some problems of neural network technologies in electric components diagnosing |
| title_short |
Оn some problems of neural network technologies in electric components diagnosing |
| title_full |
Оn some problems of neural network technologies in electric components diagnosing |
| title_fullStr |
Оn some problems of neural network technologies in electric components diagnosing |
| title_full_unstemmed |
Оn some problems of neural network technologies in electric components diagnosing |
| title_sort |
оn some problems of neural network technologies in electric components diagnosing |
| publisher |
Інститут проблем штучного інтелекту МОН України та НАН України |
| publishDate |
2017 |
| topic_facet |
Програмно-технічні засоби інтелектуальних систем |
| url |
https://nasplib.isofts.kiev.ua/handle/123456789/162344 |
| citation_txt |
Оn some problems of neural network technologies in electric components diagnosing / S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash // Штучний інтелект. — 2017. — № 3-4. — С. 95-104. — Бібліогр.: 10 назв. — англ. |
| series |
Штучний інтелект |
| work_keys_str_mv |
AT telenyksf onsomeproblemsofneuralnetworktechnologiesinelectriccomponentsdiagnosing AT savchukov onsomeproblemsofneuralnetworktechnologiesinelectriccomponentsdiagnosing AT pocrovskyieo onsomeproblemsofneuralnetworktechnologiesinelectriccomponentsdiagnosing AT morgalom onsomeproblemsofneuralnetworktechnologiesinelectriccomponentsdiagnosing AT krivenkoks onsomeproblemsofneuralnetworktechnologiesinelectriccomponentsdiagnosing AT latashio onsomeproblemsofneuralnetworktechnologiesinelectriccomponentsdiagnosing AT telenyksf prodeâkíproblemineiromereževihtehnologíivdíagnosticíelektričnihkomponentív AT savchukov prodeâkíproblemineiromereževihtehnologíivdíagnosticíelektričnihkomponentív AT pocrovskyieo prodeâkíproblemineiromereževihtehnologíivdíagnosticíelektričnihkomponentív AT morgalom prodeâkíproblemineiromereževihtehnologíivdíagnosticíelektričnihkomponentív AT krivenkoks prodeâkíproblemineiromereževihtehnologíivdíagnosticíelektričnihkomponentív AT latashio prodeâkíproblemineiromereževihtehnologíivdíagnosticíelektričnihkomponentív |
| first_indexed |
2025-12-01T04:45:12Z |
| last_indexed |
2025-12-01T04:45:12Z |
| _version_ |
1850279789786562560 |
| fulltext |
ISSN 1561-5359. Штучний інтелект, 2017, № 3-4
© S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash 95
УДК 004.93
S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash
Dept. of Automation and Control in Technical Systems NTUU “Igor Sikorskyi KPI” Kyiv, Ukraine
41, Polytechnic str., kor.18, fl. 528, Kyiv, 02056
ON SOME PROBLEMS OF NEURAL NETWORK
TECHNOLOGIES IN ELECTRIC COMPONENTS DIAGNOSING
С.Ф. Теленик, О.В. Савчук, Є.О. Покровський, О.М. Моргаль, К.С. Кривенко, І.О. Латаш
Кафедра автоматики та управління в технічних системах "КПІ ім. Ігоря Сікорського", Україна
вул. Політехнічна, буд. 41, кор. 18, кв. 528, м. Київ, 02056
ПРО ДЕЯКІ ПРОБЛЕМИ НЕЙРОМЕРЕЖЕВИХ ТЕХНОЛОГІЙ В
ДІАГНОСТИЦІ ЕЛЕКТРИЧНИХ КОМПОНЕНТІВ
The paper describes an idea of getting electric components diagnostic information and its transformation
using the discrete Karhunen-Loeve expansion. Presorting of elements by their physical and technical states is
proposed to operate with the MLP, self-organized and RBF- neural networks in the MATLAB
environment. The paper investigates the possibility of using neural network technologies for improving electric
components diagnosing by integral effects for increasing reliability of complex technological systems.
The statistical and individual classification and presorting of elements according to their physical and
technical states for work with the use of neural network technologies is proposed.
Key words: neural network technologies; diagnosing; integral effects; electric components.
У статті описується ідея отримання діагностичної інформації про електричні компоненти та її
перетворення за допомогою дискретного розкладання Карунена-Лоева. Запропоновано визначення
фізичних та технічних станів елементів за допомогою MLP, самоорганізованих та RBF-нейронних
мереж в середовищі MATLAB. У роботі досліджується можливість використання нейронних
мережевих технологій для поліпшення діагностики електричних компонент за інтегральними
ефектами для підвищення надійності складних технологічних систем. Запропоновано статистичну та
індивідуальну класифікацію та сортування елементів відповідно до їх фізичних і технічних станів для
роботи з використанням нейронних мережевих технологій.
Ключові слова: нейросетеві технології; діагностика; інтегральні ефекти; електричні
компоненти.
I. Introduction
Recently more definition is need in diagnosing physical condition of technical
products by which it is possible to determine some data about properties of physical
environment of the object being investigated. It depends on the type of hidden or overt
defects and performance degradation or various destructive processes (fatigue, wear,
corrosion, erosion, destruction, etc.) in electric components (EC). All these processes lead
to the possibility of changing technical condition that explains rapid development of
appropriate diagnosing methods of technologic forecasting by identifying relevant
characteristics of changes in their physical environment.
Thus, using the methods and means of technical diagnostics the relevant problems
may be solved:
1. To determine the type of technical state in which the object of diagnostics is located
that can be characterized as an appropriate check of its status relative to perform
working functions.
2. To trace or get the place of fault localization or determine the cause of transition into
the inactive status.
ISSN 1561-5359. Штучний інтелект, 2017, № 3-4
96 © S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash
3. To forecast changes in the technical state of the object to determine the causes of the
probability of such changes or to determine the length of time after which it may begin the
processes that lead to the unwanted changes of the product for its technical condition.
These problems are associated with using diagnostics during technical products
operation. But using diagnostics for various problems to improve technical level of modern
technology would also be appropriate to the stage of products development and their
replication. Especially, it is important for diagnosis and analysis of the changes causes in
the technical state with fixing the place and time of the unit malfunction and the type of
failure that caused appearance of the faults.
In modern production and operation of the most critical technologies there is a special
system failure analysis (SFA) which constantly helps monitoring failures and faults and
establishing their causes. Such an analysis is carried out to develop more sophisticated designs
of technical products and further improve processes with ability to avoid overt and hidden
defects. Thus, it appears improving quality and reliability of products during their operation.
Weighty importance in analysis of failures, malfunctions, defects, and breakages has
products operating conditions that lead to faults appearance. Thus, it takes place increasing
functional efficiency of product and its technical reliability and eliminating the possibility
of emergent or catastrophic situation. In the SFA systems failure analysis of technical
products in conjunction with other methods plays an effective feedback between various
stages of the product’s life cycle in order to improve quality and reliability of industrial
production at these stages.
The objective of the paper is to investigate the possibility of using the neural network
technologies for improving electric radio components diagnosing by the integral effects for
increasing reliability of complex computerized systems (CCS).
Comprehensive computerization of production processes, increasing requirements
for reliability and diagnosis of CCS need to transform their diagnostic software with
experiment informational technology to the neural network predictive analysis.
The direction of the problem solution is to test the possibility of using the neural
network technology to improve electro-physical methods of diagnosing EC by integral
effects to increase reliability of CCS.
II. Problems
The problems to be solved:
˗ usage of the electro-physical methods based on the integral physical effects of inertia
and non-linearity of electric components for getting primary diagnostic information;
˗ compression and intellectual analysis of a posteriori information using neural networks
for electric components being studied by their physical condition in the software
package MATLAB and its library Neural Network Toolbox. Informational
opportunity of electro-physical diagnosis methods for integral physical effects may be
explained by the Black Box method as a closed system. The Black Box term is
commonly defined as environment knowledge which can be obtained using only input
and output signals/information.
III. Methoodlogy for getting apriori diagnostic information about electric
components status
When utilizing the integrated methods for physical diagnosing and getting diagnostic
information some integrated physical effects are used that are observed at physical
environments of different nature [1-4]. These generally recognized effects include the
integral non-linearity effects, the effects of inertia, and the effects of fluctuations.
ISSN 1561-5359. Штучний інтелект, 2017, № 3-4
© S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash 97
Emergence of these effects is associated with workflows in the Object to be Diagnosed
(OD) and it has the same excitement origin. This means that for observability of the
integrated effects it is required the same energy activation sources that provide OD
working operation during its operation. If units are of electrical nature, it needs for
integrated methods of diagnosing to use electricity energy of the same level as for the
workflows. If there are mechanical products, they require mechanical energy, etc. In all
these cases there is no need to apply any additional conditions for operating personnel
protection in addition to implementation of safety rules during operation [5]. The main
advantage of the integrated diagnostic methods is their high efficiency due to the low
operational diagnosis lifetime with the reduced operational complexity and the operational
costs for the diagnosis process. Signals of diagnostic information more often have a kind of
the analogue signatures. Therefore, these methods are well suited for rapid diagnosis in
unfavourable conditions to identify not only the obvious defects that cause appearance of
the faults in OD but also to detect the hidden defects that trigger the sudden and gradual
failures with loss of the OD working condition state.
Non-linearity is the phenomenological property of the object physical environment
that appears in disproportionate and ambiguous nature of its response to activating the
physical quantity action depending on its changes meaning and direction. This property is
fundamental as in nature and in man-made products. It is caused by action of a large
number of direct and back-end connections between the dynamic and dissipative
subsystems of physical environments which may have very different thresholds of the
energy activation in solids.
Manifestation of non-linearity for intrinsic properties of OD physical environment is
nonlinear change of the favourable function at the physical layer or of the transfer function
at the analytical descriptive level. In this regard, all kinds of OD functional characteristics
have signs of nonlinearity (as significant and insignificant) relatively the object operating
functions and can be used to determine the OD technical condition.
But observation of non-linearity in the form of physical effects have two features.
First, there is no particular dimension (or metric scale) of nonlinearity as a physical quantity
that conveys a quantitative idea of the properties in this object. Secondly, in a quantitative
sense nonlinearity is quite multivalued because this property can’t be sufficiently identified
by one feature of functional characteristics. We have to use several features of nonlinearity
which will already make a range of the determining variables (degree polynomials,
derivatives of different orders, curvature of the first and highest orders, etc.).
With respect to the inertia property, relevant manifestation of inertia effects is
associated with the OD transition and impulse response characteristics. Their registration is
made by action of activating physical quantities at the OD entrance in the form of a pulse
jump or duble jump. In this case, mechanisms of energy conversion start acting in the
dissipative subsystems of the physical environment and mechanisms of energy dissipation
into the environment. In the language of physics, in the physical environment it takes place
transition from the adiabatic state to the isothermal state that affects the transient
characteristics of the active medium physical system. The transient integral characteristics
reflect local physical environment macrocharacteristics at time intervals. Thus, obvious
and hidden defects cause some transient or pulse characteristics changes. Measurement of
these characteristics is realised as the analogue signatures that reflect diagnostic
information about physical condition of the unit.
ISSN 1561-5359. Штучний інтелект, 2017, № 3-4
98 © S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash
Observability of the corresponding OD state at the transition process in action after
the energy jump when applying the test excitation is caused by the difference of the
character time scale change or the relaxation time in the dissipative excitation subsystems.
Therefore, it is possible to display and lead out certain dissipative processes at their
character time scales in the form of the constant time or relaxation time. In this case, the
individual parameters of the transition process are used that is fixed by the time you
change the membership function or the transfer function.
To determine the OD physical condition at its transient or pulse characteristics it is
also possible to compare these characteristics as the size continuum of physical quantity,
that varies over time, with the standard characteristics or the analogue signatures of the
same type. The fluctuation-dissipative processes occur in the thermodynamic systems of
any type. This gives rise to the stochastic temporal changes of physical quantities in the
form of fluctuations around the equilibrium state. These fluctuations generate the
stochastic signals-noises reflecting the random functions of the spontaneous oscillatory
processes in the OD physical environment.
IV. Getting primary diagnostic information
Consider a model in a black box which is defined as
),()()( tUgtY (4.1)
where U (t) - input function; Y (t) - output function; t - time.
For physical dynamic objects is the susceptibility function – the complex
environment function.The susceptibility function manifests integrated physical effects:
non-linearity, inertia (describes dynamics), fluctuations (own noise) [6]. Let us consider
the susceptibility function components:
)).(var())(()( ggconstg (4.2)
For physical objects the function g(.) is a complex function of environment
, (4.3)
where - the real part of the function; - the imaginary part; const- the fixed part;
var -the variable part.
Integral physical effects of nonlinearity, inertia, and fluctuations are used for getting
diagnostic information in carrying out electro-physical diagnostics. According to (4.3) and
using the Taylor conversion [1] for obtained EC characteristics it’s possible to describe,
say, integrated circuit by its dynamic resistance as the 1-st derivative of the nonlinear
Ampere-Volt characteristic:
(4.4)
The Volt – Farad characteristic (dynamic capacity):
(4.5)
The Ampere - Henry characteristic (dynamic inductance):
(4.6)
The general approach to hardware support of technical diagnostic methods is
sufficiently described in [1-4]. Synchronous and parallel development of electro-physical
methods allow to accumulate knowledge that determines the methodology of intellectual
diagnostic devices and sensors which are intellectual aids for accumulation of new
diagnostic knowledge.
To diagnose resistive structures by the inertia effect some options and features of
TTC are used. That is a resistor reaction as a complex object under influence of the electric
ISSN 1561-5359. Штучний інтелект, 2017, № 3-4
© S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash 99
jump with maximally non-destructive value [3]. Informational opportunity of electro-
physical methods by the integral non-linearity effect and a common approach to hardware
of technical diagnostic methods is sufficiently provided in [4].
V. Processing and compression of diagnostic information
Compression of primary diagnostic information about the state of ERC can be
accomplished by using the discrete Karhunen-Loeve expansion (DKLE) which is
expansion of the initial vectors ensemble by own vectors of the covariance matrix [6,7].
Termal transient characteristics (TTC) of different types of resistors from 10 Om to 1MOm
were examined and processed with using DKLE in number of 800 units.
Research of non-linearity was done for integrated microcircuit series K174XA11
(analog TDA2593) in number of 160 units on representative samples. Complex
dependencies of quadratic non-linearity by absolute value and phase which were gained
with using method of difference frequency separately for fit and defect probes. These
dependencies were transformed on cosine and sinus ) components and
calculated by DKLE.
Depending on an error resistor orthonormal space consists of two basic coordinates.
That means it is a plane. For circuits which space has three coordinates the number of
matrixes will increase to 5. The total number of vectors is 160 (32 samples for each
matrix). Items of the received orthonormal matrixes depicted in space are shown in Fig.1.
Fig.1. The canonical decomposition forms Fig. 2. Results of classification on the self-
of the orthonormal Euclidean space (three organized map: recognition accuracy for
dimensional X, Y, Z space) the “gridtop” topology with the“mandist”
distance
Research of the multidimensional information processing principles has allowed
selecting and justifying validity of the Karhunen-Loeve expantion as a mathematical tool for
processing diagnostic information of EC. It is proposed for practical realization to process
diagnostic information with integral physical effects at using modern neural network
technology (multilayer perceptron, Kohonen maps, radial-basis networks).
VI. Experimental results
For practical implementation of processing diagnostic information with integrated
physical effects it was proposed to use modern neural network technologies (multilayer
perceptron, Kohonen maps, radial-basis networks) [8]. For training MLP the following
algorithms in the MATLAB environment were selected: Bayesian regularization or
ISSN 1561-5359. Штучний інтелект, 2017, № 3-4
100 © S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash
function training based on the inverse error propagation using the Bayesian regularization;
the gradient descent back propagation method or the method of gradient descent; the
gradient descent method with adaptive back propagation; the Powell-Beale back
propagation or gradient method coupled with the Powell-Beale repetitions; the resilient
back propagation method or the method of inverse elastic distribution.
As the criterion for assessing the training accuracy it was used the international types
of error - MSE, MAE and others. For multilayer perceptron procedure training it was
obtained the best accuracy result for resistors (Table 1).
Table 1. Comparative results for resistors classification
Neural network type Obtained
classification accuracy, %
Training time, s
RBF-network 99.96 2
Kohonen map 84 2
MLP 98.75 39
Table 2. Results for classification resistors precision for multilayer perceptron
As clear from tables 1 and 4 the probabilistic radial basis neural networks have
higher classification accuracy than the neural network with radial basic elements and given
zero error. For classification of microcircuits using orthonormal space decomposition it is
enough to perform classification into two or three classes (Table 3).
Table 3. The best results for accuracy of the classification using multilayer perceptron
№ of
classes
Activation functions of the hidden layer
Error type, 10-7
kep,
unit
ttr,
s MSE SSE MSEREG MAE
MLP
2 1.4 4200 1.8 1 100 52 24
3 3 8000 3 2 100 62 33
Table 4. Comparing the best results for all training types of neural network for microcircuits
Neural network type
Classification
accuracy, %
Training
time, s
Classification
accuracy , %
Training time, s
Two classes
Three classes
RBF-network (pnn) 99.96 2 99,96 2
RBF-network
(newrbe)
82.81 2 70,83 2
MLP 99.23 24 99.05 33
Educational
function
Activation functions of hidden layers
logsig- logsig- logsig
Error type kep,
unit
ttr, s
MSE SSE MSEREG MAE
MLP 0.0000004 0.00092 0.00000034 0.00025 83 39
ISSN 1561-5359. Штучний інтелект, 2017, № 3-4
© S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash 101
Thus the best results again were shown by neural network with activation function of
each hidden layer. The data sample size is 64 samples. Further study was carried out on the
self-organized maps (Table 5).
Table 5. Dependence of the accuracy classification on the self-organized map setting parameters
Topology Hextop
Distance type between
adjacent clasters
Linkdist Dist Mandist Boxdist
Correctly classified
samples, %
83.33 83.5 61 70.67
Topology Gridtop
Distance type between
adjacent clasters
Linkdist Dist Mandist Boxdist
Correctly classified
samples, %
61 61 83.67 83.33
Number of training
periods
200
Average training time, s 2
The results in the Table 5 topology have the best performance with the “gridtop”
topology at using the distance between the “mandist” clusters. And moreover impact on the
classification accuracy of the step parameter used to estimate distance between the
neighboring clusters is not significant (Table 6).
Table 6. Influence of the step parameter on the classification accuracy
“Gridtop” topology, distance between adjacent “mandist”clusters
Step size 10 30 50 70 80 90 100
Probability of
correct
classification
0.61 0.84 0.837 0.597 0.603 61 0.837
Self-organized (Kohonen) network (excluding teachers) do not require submission
outputs on neural network training set. This algorithm is a neural network algorithm known
analogue K medium. At each step of the training input vector network served and sought
another neuron whose weight differs least from this vector. Found declared the winner
neuron and its weights vector W is updated as follows:
wi(k+1)= wi(k)+ (xn-wi(k))+awi(k)(1-||wi||
2), (6.1)
which parameter is responsible for setting the rate of learning and change its value in the
interval (0,1). All training vectors are processed one by one until they will fail to stabilize
or other condition stops. The applied technique traingd (gradient descent backpropagation)
and / or penalty functions. Recognition results on the Kohonen map are shown in Fig. 2.
For solving probability problems a special type of neural network PNN (Probabilistic
Neural Networks) is using. PNN network architecture is based on the architecture of radial
basis function network, but as a second layer uses so-called rival layer which calculates the
probability input vector belonging to a particular class and compares the vector of the class
whose probability of belonging to above.
ISSN 1561-5359. Штучний інтелект, 2017, № 3-4
102 © S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash
The main difference between RBF networks and conventional multilayer networks of
direct distribution is a function of the hidden layer neurons. In a conventional multi-layer
network, each neuron of the working layer implements a hyperplane in multidimensional
space and RBF-neuron implements a hypersphere. In problems where the formation of data
are close to the circular symmetry, it can reduce the number of neurons.
When MATLAB modeling the weight matrix of the first layer IW11 (net.IW) formed
using vectors from the training set input in the form of a matrix .When a new entry is
submitted, the unit calculates the proximity to the new vector of vectors of the
training set, then the calculated distance multiplied by bias and fed to the input function
activation radbas. The vector of the training set, which is closest to the entrance, to be
submitted in output vector . Number of close to 1.
Weight matrix of the second layer LW21 (net.LW) meets the connectivity matrix
built for this training sequence. This operation can be performed using M-function ind2vec,
which converts vector of targets in a matrix of connectivity . The product defines
the elements of a vector corresponding to each of the K classes. As a result the
competing activation function of second layer compet forms the output value equal to 1 for
most largest element of the vector and 0 otherwise. Thus, the network PNN performs
the input vector classification by K classes.
Classification and presorting of elements according to their physical and technical
states were performed on neural networks in the MATLAB environment [9,10]. For EC
presorting the probabilistic neural networks were used. In fact, the network is trained to
assess the probability density function. According to the Bayesian statistics to minimize
error the model with the greatest probability density is being chosen. The resulting data for
chips and resistors confirm the highest accuracy 99.96% and speed training time 2 hours at
using the probabilistic (pnn) RBF networks compared with the neural networks with the
radial- basis elements and given zero error (newrbe).
Conclusion
The observed identification features increase diagnostic possibilities of technical
methods for diagnostics electric components in exposing hidden defects, potential
instability and unplanned degradation processes. The discrete Karhunen-Loeve expansion
for information compression provides easy algorithm for neural network training and
classifying with usage neural networks in MATLAB environment.
For the further development of this problem area the fuzzy logical deduction on real
data may be proposed. This approach will allow to perform synthesis of the new type neural
network that will help to explore larger data samples with less time and more accuracy in the
overall problem of increasing reliability and diagnosis of complex infrastructures.
References
1. Сердюк Г.Б., Усатенко В.Г., Миненко А.П. Автоматизированные комплексы для электро-
физического диагностирования ИС и других изделий электронной техники по критериям качества
и надежности. К.:КПИ: НПП АМАТЕК, 1992.- C. 55-59.
2. Володарський Є.Т., Кухарчук В.В., Поджаренко В.О., Сердюк Г.Б. Метрологічне забезпечення
вимірювань і контролю. Навчальний посібник. – Вінниця: ВДТУ, 2001. – 219 с.
3. А.С.№ 1667118.СССР. Устройство для распознавания динамических характеристик объектов/
Сердюк Г.Б.,Cавчук Е.В., Усатенко В.Г., Попова Н.С., Шерстюк И.В., Свистунов А.Я. Приоритет
от 17.05.1988, БИ №28, 1991
4. Патент № 1056 на винахід (Україна). Спосіб відбраковки потенційно ненадійних та нестабіль-них
мікросхем / В.Г. Усатенко, Г.Б. Сердюк, С.Н. Сазонов. - G 01 R 31/27. - 1993.
ISSN 1561-5359. Штучний інтелект, 2017, № 3-4
© S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash 103
5. Теленик С.Ф., Ролік О.І., Ясочка М.В., Моргаль О.М. Нечітке оцінювання в задачах управління
рівнем обслуговування // В кн.: Наукові записки Українського науково-дослідного інституту
зв’язку. — 2011.— №2(18).— С. 29—42.
6. Castrillon-Candas J., Amaratunga K. – Fast estimation of continuous Karhunen-Loeve eigenfunctions
using wavelets. – IEEE: IEEE Signal Processing Society. – 2002. – pp.78-86.
7. Солодовщиков А.Ю., Платонов А.К. Исследование метода Карунена-Лоэва. – М.:РАН, 2006. – 29 с.
8. Telenyk S., Savchuk O., Pocrovskyi E., Morgal O., Krivenko K., Latash I. On Searching of Type and
Optimizing of Neural Network in Electric Component Diagnosis // Proceedings of the IV Congress on
«Automatic Control and Information Technology», (ICACIT 2017), Cracow, Poland, 14 - 16 of
December, 2017, pp.147-159.
9. Дияконів В., Круглов В. Математичні пакети розширення MATLAB. Спеціальний довідник. -
Спб.: Питер, 2001. - 480 с.
10. Yakimenko O. Engineering computations and modeling in MATLAB/Simulink. – Monterey, California:
American Institute of Aeronautics and Astronautics, INC., 2011. – 938 p.
References
1. Serdyuk G.B., Usatenko V.G., Minenko A.P. Avtomatizirovannyie kompleksyi dlya elektro-
fizicheskogo diagnostirovaniya IS i drugih izdeliy elektronnoy tehniki po kriteriyam kachestva i
nadezhnosti. K.:KPI: NPP AMATEK, 1992.- S. 55-59.
2. Volodarskyi Ye.T., Kukharchuk V.V., Podzharenko V.O., Serdiuk H.B. Metrolohichne zabezpechennia
vymiriuvan i kontroliu. Navchalnyi posibnyk. – Vinnytsia: VDTU, 2001. – 219 s.
3. A.S.# 1667118.SSSR. Ustroystvo dlya raspoznavaniya dinamicheskih harakteristik ob'ektov/
Serdyuk G.B.,Cavchuk E.V., Usatenko V.G., Popova N.S., Sherstyuk I.V., Svistunov A.Ya. Prioritet ot
17.05.1988, BI #28, 1991.
4. Patent № 1056 na vynakhid (Ukraina). Sposib vidbrakovky potentsiino nenadiinykh ta nestabil-nykh
mikroskhem / V.H. Usatenko, H.B. Serdiuk, S.N. Sazonov. - G 01 R 31/27. - 1993.
5. Telenyk S.F., Rolik O.I., Yasochka M.V., Morhal O.M. Nechitke otsiniuvannia v zadachakh
upravlinnia rivnem obsluhovuvannia // V kn.: Naukovi zapysky Ukrainskoho naukovo-doslidnoho
instytutu zviazku. — 2011.— №2(18).— S. 29—42.
6. Castrillon-Candas J., Amaratunga K. – Fast estimation of continuous Karhunen-Loeve eigenfunctions
using wavelets. – IEEE: IEEE Signal Processing Society. – 2002. – pp.78-86.
7. Solodovschikov A.Yu., Platonov A.K. Issledovanie metoda Karunena-Loeva. – M.:RAN, 2006. – 29 s.
8. Telenyk S., Savchuk O., Pocrovskyi E., Morgal O., Krivenko K., Latash I. On Searching of Type and
Optimizing of Neural Network in Electric Component Diagnosis // Proceedings of the IV Congress on
«Automatic Control and Information Technology», (ICACIT 2017), Cracow, Poland, 14 - 16 of
December, 2017, pp.147-159.
9. Dyiakoniv V., Kruhlov V. Matematychni pakety rozshyrennia MATLAB. Spetsialnyi dovidnyk. - Spb.:
Pyter, 2001. - 480 s.
10. Yakimenko O. Engineering computations and modeling in MATLAB/Simulink. – Monterey,
California: American Institute of Aeronautics and Astronautics, INC., 2011. – 938 p.
РЕЗЮМЕ
С.Ф. Теленик, О.В. Савчук, Є.О. Покровський, О.М. Моргаль,
К.С. Кривенко, І.О. Латаш
Про деякі проблеми нейромережевих технологій у діагностиці
електричних компонентів
У статті розглядається інформаційна можливість електофізичних методів діагностування
електричних компонентів за інтегральними ефектами інерційності та нелінійності, описується
ідея отримання діагностичної інформації про електричні компоненти.
Надається модель чорної скриньки, внутрішній зміст якої описується
комплексною функцією сприйнятливості. Використано перетворення Тейлора для
отримання моделі за інтегральним ефектом нелінійності для різних електричних
компонентів, таких як резистори, індуктивності, конденсатори, а також для
мікросхем у вигляді двохполюсника, що підключений до шини живлення.
ISSN 1561-5359. Штучний інтелект, 2017, № 3-4
104 © S.F. Telenyk, O.V. Savchuk, E.O. Pocrovskyi, O.M. Morgal, K.S. Krivenko, I.O. Latash
Інтелектуальний аналіз та перетворення діагностичної інформації про
електричні компоненти виконано за допомогою дискретного розкладання Карунена-
Лоева, що є розкладанням ансамбля початкових векторів за власними векторами
коваріаційної матриці. Експериментально доведено, що для електричних
компонентів цей простір складається з двох, трьох або шести компонентів в
залежності від похибки обчислень.
У роботі досліджується можливість використання нейронних мережевих
технологій для поліпшення діагностики електричних компонент за інтегральними
ефектами інерційності та нелінійності для підвищення надійності складних
технологічних систем. Запропоновано визначення фізичних та технічних станів
елементів за допомогою MLP, самоорганізованих та RBF-нейронних мереж, а також
наведений алгоритм моделювання вагових матриць в середовищі MATLAB.
Найкращий результат за точністю класифікації 99.96 % отриманий для RBF-
нейронних мереж.
Запропоновані статистична та індивідуальна класифікація та сортування
електричних компонентів відповідно до їх фізичних і технічних станів з
використанням нейронних мережевих технологій.
Надійшла до редакції 04.10.2017
|