349 – 358,
2011
Competitive Learning for Self Organizing Maps used in
Classification of Partial Discharge
Rubén JaramilloVacio
1, 2
, Alberto OchoaZezzatti
3
, Julio Ponce
4
1
Comisión Federal de ElectricidadLaboratorio de Pruebas a Equipos y Materiales (LAPEM).
2
Centro de Innovación Aplicada en Tecnologías Competitivas (CIATEC).
3
Universidad Autónoma de Ciudad Juárez.
4
Universidad Autónoma de Aguascalientes
ruben.jaramillo@cfe.gob.mx
Abstract. This paper different competitive learning algorithms for Self
Organizing Map (SOM) are experimentally examined, the characterization of
the obtainable results in terms of quality of SOM. The competitive learning
algorithms showed to SOM algorithm are Winnertakesall, Frequency
Sensitive Competitive Learning and Rival Penalized Competitive Learning. As
a case study: the performance in classification of partial discharge on power
cables.
1. Introduction
Competitive learning is an efficient tool for Self Organizing Maps, widely applied in
variety of signal processing problems such as classification, data compression, etc.
In the field of data analysis two terms frequently encountered are supervised and
unsupervised clustering methodologies. While supervised methods mostly deal with
training classifiers for known symptoms, unsupervised clustering provides exploratory
techniques for finding hidden patterns in data. With the huge volumes of data being
generated from the different systems everyday, what makes a system intelligent is its
ability to analyze the data for efficient decisionmaking based on known or new cluster
discovery. The partial discharge (PD) is a common phenomenon which occurs in
insulation of high voltage, this definition is given in [1]. In general, the partial
discharges are in consequence of local stress in the insulation or on the surface of the
insulation.
The typical competitive learning algorithm kmeans (or called hard cmeans)
clustering is a batch algorithm for designing a vector quantizer, which is a mapping of
input vectors to one of c predetermined codevectors (also called codebooks) [2].
Fuzzy cmeans (FCM) clustering is a fuzzy extension of hard cmeans clustering. The
FCM and its varieties have been widely studied and applied in various areas [35].
During the last fifteen years there have been developed new advanced algorithms that
eliminate the ”dead units” problem and perform clustering without predicting the exact
cluster number, as for example: the frequency competitive algorithm (FSCL) [6], the
350 JaramilloVacio Rubén, OchoaZezzatti Alberto, Ponce Julio
incremental kmeans algorithm, the rival penalizing competitive algorithm (RPCL)
[7].
We evaluate the performance of algorithms in which competitive learning is applied
of partial discharge dataset, quantization error, topological error and time in seconds
per training epoch. The result from classification of PD shows that Winnertakesall
(WTA) has better performance than Frequency Sensitive Competitive Learning
(FSCL) and Rival Penalized Competitive Learning (RPCL).
Table 1 show a concentration of researchers who worked on the feature extraction,
recognition and classification of PD, as well as the different artificial intelligent tools
and the constraint to utilize these methods.
Table 1. Classification and diagnosis in pd using data mining tools.
Authors Tool and Objective Constraints
Mozroua et al [8]
Kravida [9]
Tool: Supervised Neural Networks.
Objective: Recognition between
different sources formed of cylindrical
cavities.
Recognition of different
sources in the same sample
Kim et al
[10]
Tool: FuzzyNeural Networks.
Objective: Comparison between Back
Propagation Neural Network and
FuzzyNeural Networks
Performance in the case of
multiple discharges and
including defects and
noises.
RiCheng et al [11] Tool: Particle Swarm Optimization
Objective: Localization of PD in the
power transformer.
On site application should
improve performance.
Chang et al [12] Tool: Self Organizing Map (SOM).
Objective: PD pattern recognition and
classification.
Quality and optimization
structure of SOM.
Fadilah Ab Aziz et
al[13]
Tool: Support Vector Machine (SVM).
Objective: Feature Selection and PD
classification.
SVM is not reliable for
small dataset.
Hirose et al [14] Tool: Decision Tree
Objective: Feature Extraction and PD
classification
The allocation rules are
sensitive to small
perturbations in the dataset
(Instability)
This discovered knowledge then forms the basis of two separate decision support
systems for the condition assessment/defect classification of these respective plant
items. In this paper is shown a comparative of competitive learning algorithms to
classify measured PD activities into underlaying insulation defects or source that
generate PD’s using Self Organizing Maps (SOM). Multidimensional scaling (MDS)
is a nonlinear feature extraction technique [15]; it aims to represent a
multidimensional dataset in two or three dimensions such that the distance matrix in
the original kdimensional feature space is preserved as faithfully as possible in the
projected space. The SOM, or Kohonen Map [16], can also be used for nonlinear
feature extraction. It should be emphasized that the goal here is not to find an optimal
clustering for the data but to get good insight into the cluster structure of the data for
data mining purposes. Therefore, the clustering method must be fast, robust, and
visually efficient.
Competitive Learning for Self Organizing Maps 351
2. Partial discharge: concepts
Partial discharges occur wherever the electrical field is higher than the breakdown
field of an insulating medium: Air: 27 kV/cm (1 bar), SF6: 360 kV/cm (4 bar),
Polymers: 4000 kV/cm
They are generally divided into three different groups because of their different
origins:
 Corona Discharges – Occurs in gases or liquids caused by concentrated electric
fields at any sharp points on the electrodes.
 Internal Discharges – Occurs inside a cavity that is surrounded completely by
insulation material; might be in the form of voids (e.g. dried out regions in oil
impregnated papercables).
 Surface Discharges – Occurs on the surface of an electrical insulation where the
tangential field is high e.g. end windings of stator windings.
CORE
INSULATION
GROUND
PD
B
R
E
A
K
D
O
W
N
EVOLUTION
Fig. 1. Example of damage in polymeric power cable from the PD in a cavity to breakdown.
In general, the partial discharges are in consequence of local stress in the insulation
or on the surface of the insulation. This phenomenon has a damaging effect on the
equipments, for example transformers, power cables, switchgears, and others (see
Figure 1). The first approach in a diagnosis is selecting the different features to
classify measured PD activities into underlying insulation defects or source that
generate PD’s. The partial discharge measurement is a typical nondestructive test and
it can be used to judge the insulation performance at the beginning of the service time
taking into account the reduction of the performance during the service time by
ageing, whereby the ageing depends on numerous parameters like electrical stress,
thermal stress and mechanical stress. In particular for solid insulation like XLPE on
power cables where a complete breakdown seriously damages the test object the
partial discharge measurement is a tool for quality assessment. The charge that a PD
generates in a cavity is called the physical charge and the portion of the cavity surface
that the PD affects is called the discharge area. E
applied
is the applied electric field and
q
physical
is the physical charge [17].
The pulse repetition rate n is given by the number of partial discharge pulses
recorded in a selected time interval and the duration of this time interval. The recorded
pulses should be above a certain limit, depending on the measuring system as well as
on the noise level during the measurement. The pulse repetition frequency N is the
352 JaramilloVacio Rubén, OchoaZezzatti Alberto, Ponce Julio
number of partial discharge pulses per second in the case of equidistant pulses.
Furthermore, the phase angle  and the time of occurrence t
i
are information on the
partial discharge pulse in relation to the phase angle or time of the applied voltage
with period T. For PD diagnosis test, is very important to classify measured PD
activities, since PD is a stochastic process, namely, the occurrence of PD depends on
many factors, such as temperature, pressure, applied voltage and test duration;
moreover PD signals contain noise and interference [18]. Therefore, the test engineer
is responsible for choosing proper methods to diagnosis for the given problem. In
order to choose the features, it is important to know the different source of PD; an
alternative is though pattern recognition. This task can be challenging, nevertheless,
features selection has been widely used in other field, such as data mining [19] and
pattern recognition using neural networks [810]. This research only shows test on
laboratory without environment noise source, and it is a condition that does not
represent the conditions on site, Markalous [20] presented the noise levels on site
based on previous experiences.
The phase resolved analysis investigates the PD pattern in relation to the variable
frequency AC cycle. The voltage phase angle is divided into small equal windows.
The analysis aims to calculate the integrated parameters for each phase window and to
plot them against the phase position ().
 (q
m
−): the peak discharge magnitude for each phase window plotted against,
where q
m
is peak discharge magnitude.
3. Self Organizing Map (SOM)
3.1 Winner takes all
The Self Organizing Map developed by Kohonen, is the most popular neural
network models. The SOM algorithm [15,21] is based on unsupervised competitive
learning called winner – takes – all, which means that the training in entirely data
driven and that the neurons of the map compete with each other.
Supervised algorithms [8, 9] like multilayered perceptron, required that the target
values for each data vector are known, but the SOM does not have this limitation. The
SOM is a neural network model that implements a characteristics nonlinear mapping
from the highdimensional space of input signal onto a typically 2dimensional grid of
neurons. The SOM is a twolayer neural network that consists of an input layer in a
line and an output layer constructed of neurons in a twodimensional grid.
The neighborhood relation of neuron i, an ndimensional weight vector w is
associated; n is the dimension of input vector. At each training step, an input vector x
is randomly selected and the Euclidean distances between x and w are computed. The
image of the input vector and the SOM grid is thus defined as the nearest unit w
ik
and
bestmatching unit (BMU) whose weight vector is closest to the x [21]:
Competitive Learning for Self Organizing Maps 353
( ) ( )
2
,
i ik k
k
D x w w x = ÷
¿
(2)
The weight vectors in the bestmatching unit and its neighbors on the grid are
moved towards the input vector according the following rule:
( ) ( )
( )
,
ij j ij
ij j ij
ij
w c i x w
w x w to i c
w to i c
o o
o
A = ÷
A = ÷ =
A =
(3)
where c denote the neighborhood kernel around the bestmatching unit and α is the
learning rated and δ is the neighborhood function.
The number of panels in the SOM is according the A x B neurons, the Umatrix
representation is a matrix U ((2A1) x (2B1)) dimensional [22]. The selection of the
distance criterion depends on application. In this paper, Euclidean distance is used
because it is widely worn with SOM [23].
It is complicated to measure the quality of an SOM. Resolution and topology
preservation are generally used to measure SOM quality [24]. There are many ways to
measure them. The quantization error (qe) is calculated to measure the quality of the
map. The quantization error qe is the average distance between each data vector and
its BMU, measuring map resolution. The topological error te is the proportion of all
data vectors for which first and second BMUs are adjacent units, otherwise this is
regarded as violation of topology and thus penalized by increasing the error value.
3.2 The frequency sensitive competitive algorithm
The kmeans algorithm has also the “dead units” problem, which means that if a
centre is inappropriately chosen, it may never be updated, thus it may never represent
a class.
To solve the “dead units” problem it has been introduced the so called “frequency
sensitive competitive learning” algorithm (FSCL) [25] or competitive algorithm “with
conscience”. Each centre counts the number of times when it has won the competition
and reduces its learning rate consequently. If a center has won too often “it feels
guilty” and it pulls itself out of the competition. The FSCL algorithm is an extension
of kmeans algorithm, obtained by modifying relation (2) according to the following
one:
( ) ( ) argmin 1,...,
i i
j x n c n i N ¸ = ÷ = (4)
where n is the inputs, N represents the centres numbers, the relative winning frequency
¸
i of
the centre c
i
defined as:
354 JaramilloVacio Rubén, OchoaZezzatti Alberto, Ponce Julio
1
i
i n
i
i
s
s
¸
=
=
¿
(6)
where s
i
is the number of times when the centre c
i
was declared winner in the past. So
the centers that have won the competition during the past have a reduced chance to
win again, proportional with their frequency term¸. After selecting out the winner, the
FSCL algorithm updates the winner with next equation:
( ) ( ) ( ) ( ) 1
i i i
c n c n x n c n q ( + = ÷ ÷
¸ ¸
(7)
q is the learning rate, in the same way as the kmeans algorithm, and meanwhile
adjusting the corresponding s
i
with the following relation:
( ) ( ) 1 1
i i
s n s n + = + (8)
3.3 The rival penalized competitive learning algorithm
The rival penalized competitive algorithm (RPCL) [25] performs appropriate
clustering without knowing the clusters number. It determines not only the winning
centre j but also the second winning center r, named rival
( ) ( ) argmin , 1,...,
i i
r x n c n i N i j ¸ = ÷ = = (9)
The second winning centre will move away its centre from the input with a ratio
learning rate. All the other centres vectors will not change. So the
learning law can be synthesized in the following relation:
( )
( ) ( ) ( )
( ) ( ) ( )
( )
1
i i
i i i
i
c n x n c n if i j
c n c n x n c n if i j
c n if i j and i r
q

¦ ( + ÷ =
¸ ¸
¦
¦
( + = ÷ ÷ =
´
¸ ¸
¦
= =
¦
¹
(10)
If the learning speed q is chosen much greater than , with at least one order of
magnitude, the number of the output data classes will be automatically found. In other
words, suppose that the number of classes is unknown and the centres number N is
greater than the clusters number, than the centres vectors will converge towards the
centroids of the input data classes. The RPCL will move away the rival, in each
iteration, converging much faster than the kmeans and the FSCL algorithms.
Competitive Learning for Self Organizing Maps 355
4. Analysis of PD Data
PD measurements for power cables are generated and recorded through laboratory
tests. Corona was produced with a point to hemisphere configuration: needle at high
voltage and hemispherical cup at ground. Surface discharge XLPE cable with no stress
relief termination applied to the two ends. High voltage was applied to the cable inner
conductor and the cable sheath was grounded, this produces discharges along the outer
insulation surface at the cable ends. Internal discharge was used a power cable with a
fault due to electrical treeing. Were considered the pattern characteristic of univariate
phaseresolved distributions as inputs, the magnitude of PD is the most important
input as it shows the level of danger, for this reason the input in the SOM the raw data
is the peak discharge magnitude for each phase window plotted against (qm − ).
Figure 2 shows the conceptual diagram training. In the cases analyzed, the original
dataset is 1 million of items, was used a neurons array of 10×10 cells to extract
features. As it is well known, in fact, a too small number of neurons per class could be
not sufficient to represent the variability of the samples to be classified, while a too
large number in general makes the net too much specialized on the samples belonging
to the training set and consequently reduces its generalization capability. Moreover a
too large number of neuron per class implies a long training time and a possible
underutilization of some of the neural units.
DATA
NORMALIZATION/
PREPROCESSING
QUALITY
MEASURES FOR
SOM
SOM
(TRAINING)
UNSUPERVISED MODULE
DATAWAREHOUSE
(KNOWLEDGE)
FAULT DIAGNOSIS/
DETECTION
Fig. 2. The component interaction between SOM
In table 2 are shower the parameters for training to each competitive learning
algorithm. The coefficient ¸ has been dynamically changed during the training.
356 JaramilloVacio Rubén, OchoaZezzatti Alberto, Ponce Julio
0 20 40 60 80 100
9
9.5
10
10.5
Quantization Error per Training Epoch (Surface Discharge)
WTA
FSCL
RPCL
0 20 40 60 80 100
0.96
0.97
0.98
0.99
1
Topological Error per Training Epoch (Surface Discharge)
WTA
FSCL
RPCL
0 20 40 60 80 100
5.4
5.6
5.8
6
6.2
6.4
Quantization Error per Training Epoch (Internal Discharge)
WTA
FSCL
RPCL
0 20 40 60 80 100
0.96
0.97
0.98
0.99
1
Topological Error per Training Epoch (Internal Discharge)
WTA
FSCL
RPCL
0 20 40 60 80 100
0.7
0.75
0.8
0.85
0.9
0.95
1
Quantization Error per Training Epoch (Corona Discharge)
WTA
FSCL
RPCL
0 20 40 60 80 100
0.94
0.95
0.96
0.97
0.98
0.99
1
Topological Error per Training Epoch (Corona Discharge)
WTA
FSCL
RPCL
Table 2 Parameter for training
WTA FSCL RPCL
Epoch 100 100 100
q 0.1 0.1 0.05
 0.01 0.01 0.01
Fig. 3. Quantization and Topological error per Training Epoch (Surface Discharge)
Fig. 4. Quantization and Topological error per Training Epoch (Internal Discharge)
Fig. 5. Quantization and Topological error per Training Epoch (Corona Discharge)
In figure 3, 4 and 5 is showed the performance of the competitive learning algorithms
to different PD source.
Competitive Learning for Self Organizing Maps 357
We evaluate the performance of algorithms in which competitive learning is applied
of partial discharge dataset, quantization error, topological error and time in seconds
per training epoch are showed in table 3, and it is observer that the WTA works with
less error and less time of training, but the FSCL is not always satisfactory because the
training time is very long. The RCPL have longest training time but is the algorithm
with more error.
Table 3 Performance of training in SOM
WTA FSCL RPCL
Surface Discharge
q
e
9.0 9.8 9.1
t
e
0.985 1 0.97
time 849 seconds 1160 seconds 1226 seconds
Internal Discharge
q
e
5.5 5.9 5.4
t
e
0.98 0.99 0.98
time 173 seconds 222 seconds 241 seconds
Corona Discharge
q
e
0.78 0.85 0.75
t
e
0.99 0.98 0.95
time 889 seconds 1191 seconds 1362 seconds
Conclusion
PD patterns recognition and classification require an understanding of the traits
commonly associated with the different source and relationship between observed PD
activity and responsible defect sources. This paper shows the performance of SOM
using different competitive learning algorithms to classify measured PD activities into
underlaying insulation defects or source that generate PD’s, its showed that WTA is
the better algorithm with less error and training time, but its overall performance are
not always satisfactory, being alternative in accord at the performance FSCL or RPCL
algorithms.
References
[1] IEC 60270 Ed. 2. Highvoltage test techniques  Partial discharge measurements.
(2000). 15 – 16.
[2] Pollard D. Quantization and the method of kmeans. IEEE Trans. Information Theory
(1982). 28 – 199
[3] Bezdek JC. Pattern Recognition with Fuzzy Objective Function Algorithms. Plenum
Press; 1981.
[4] Yang MS. A survey of fuzzy clustering, Math Comput Model 1993;18:1.16.
[5] Höppner F, Klawonn F, Kruse R, Runkler T. Fuzzy Cluster Analysis: Methods for
Classification Data Analysis and Image Recognition. New York: Wiley; 1999.
358 JaramilloVacio Rubén, OchoaZezzatti Alberto, Ponce Julio
[6] X. Wang, H. Liu, J. Lu, and T. Yohagy, “Combining recurrent neural networks with
selforganizing maps for channel equalization,” IEEE Trans. on Communications,
vol. E85E, pp. 2227–2235, 2002.
[7] L. Xu, A. Krzyzak, and A. E. Oja, “Rival penalized competitive learning for clustering
analysis. rbf net and curve detection” IEEE Trans. on Neural Networks, no. 4, pp.
636–64, 1993.
[8] Mazroua, A. PD pattern recognition with neural netwoks using the multilayer
perception technique. IEEE Transactions on Electrical Insulation, 28, (1993). 1082
1089.
[9] Krivda, A. Automated Recognition of Partial Discharge. IEEE Transactions on
Dielectrics and Electrical Insulation, 28, (1995). 796821.
[10] 10 Kim, J. Choi, W. Oh, S. Park, K. Grzybowski, S. Partial Discharge Pattern
Recognition Using FuzzyNeural Networks (FNNs) Algorithm. IEEE International
Power Modulators and High Voltage Conference. (2008). 272275.
[11] RiCheng, L. Kai, B. Chun, D. ShaoYu, L. GouZheng, X. Study on Partial Discharge
Localization by Ultrasonic Measuring in Power Transformer Based on Particle
Swarm Optimization. International Conference on High Voltage Engineering and
Application. (2008). 600603.
[12] Chang, W. Yang, H. Application of Self Organizing Map Approach to Partial Discharge
Pattern Recognition of CastResin Current Transformers. WSEAS Transaction on
Computer Research. Issue 3. Volume 3. (2008). 142151.
[13] Fadilah Ab Aziz, N. Hao, L. Lewin, P. Analysis of Partial Discharge Measurement
Data Using a Support Vector Machine. 5th Student Conference on Research and
Development. (2007). 16.
[14] Hirose, H. Hikita, M. Ohtsuka, S. Tsuru, S. Ichimaru, J. Diagnosis of Electric Power
Apparatus using the Decision Tree Method. IEEE Transactions on Dielectrics and
Electrical Insulation, 15, (2008). 12521260.
[15] M. Kantardzic, Data Clustering, Theory, Algorithms and Methods. ASASIAM, 2007.
53 – 57.
[16] Vesanto, J. Data Exploration Process Based on the Self Organizing Map. Doctoral
Thesis in Computer Science. Helsinki University of Technology. (2002).
[17] Forssén, C. Modelling of cavity partial discharges at variable applied frequency.
Sweden: Doctoral Thesis in Electrical Systems. KTH Electrical Engineering. (2008).
[18] Edin, H. Partial discharge studies with variable frequency of the applied voltage.
Sweden: Doctoral Thesis in Electrical Systems. KTH Electrical Engineering. (2001).
[19] K. Lai, B. P. Descriptive Data Mining of Partial Discharge using Decision Tree with
genetic algorithms. AUPEC. (2008).
[20] Markalous, S. Detection and location of Partial Discharges in Power Transformers
using acoustic and electromagnetic signals. Stuttgart University: PhD Thesis. (2006).
[21] Kohonen T. Engineering Applications of Self Organizing Map. Proceedings of the
IEEE. (1996).
[22] RubioSánchez, M. Nuevos Métodos para Análisis Visual de Mapas Auto
organizativos. PhD Thesis. Madrid Politechnic University. (2004).
[23] Vesanto J., Alhoniemi E., Clustering of the Self Organizing Map. IEEE Transactions
on Neural Networks , Vol. 11, Nr. 3, pages 10821089. (2000).
[24] Pölzlbauer, G. Survey and Comparison of Quality Measures for DelfOrganizing Maps.
Proceedings of the Fifth Workshop on Data Analysis (2004). 6782.
[25] C. Ahalt, A. K. Krishnamurthy, P. Chen, and D. E. Melton "Competitive learning
algorithms for vector quantization", Neural Networks, vol. 3, no. 3, pp. 277290.
(1990).
SVM is not reliable for small dataset. or Kohonen Map [16]. Objective: PD pattern recognition and classification. The result from classification of PD shows that Winnertakesall (WTA) has better performance than Frequency Sensitive Competitive Learning (FSCL) and Rival Penalized Competitive Learning (RPCL). and visually efficient. Quality and optimization structure of SOM. can also be used for nonlinear feature extraction. it aims to represent a multidimensional dataset in two or three dimensions such that the distance matrix in the original kdimensional feature space is preserved as faithfully as possible in the projected space. Ponce Julio incremental kmeans algorithm. Tool: Support Vector Machine (SVM). Table 1 show a concentration of researchers who worked on the feature extraction. the rival penalizing competitive algorithm (RPCL) [7]. quantization error. Table 1. Classification and diagnosis in pd using data mining tools. Multidimensional scaling (MDS) is a nonlinear feature extraction technique [15]. Objective: Recognition between different sources formed of cylindrical cavities. Authors Mozroua et al [8] Kravida [9] Tool and Objective Tool: Supervised Neural Networks. On site application should improve performance. Objective: Feature Selection and PD classification. We evaluate the performance of algorithms in which competitive learning is applied of partial discharge dataset. Therefore.350 JaramilloVacio Rubén. the clustering method must be fast. recognition and classification of PD. Objective: Comparison between Back Propagation Neural Network and FuzzyNeural Networks Tool: Particle Swarm Optimization Objective: Localization of PD in the power transformer. It should be emphasized that the goal here is not to find an optimal clustering for the data but to get good insight into the cluster structure of the data for data mining purposes. The allocation rules are sensitive to small perturbations in the dataset (Instability) Chang et al [12] Fadilah Ab Aziz et al[13] Hirose et al [14] This discovered knowledge then forms the basis of two separate decision support systems for the condition assessment/defect classification of these respective plant items. The SOM. as well as the different artificial intelligent tools and the constraint to utilize these methods. topological error and time in seconds per training epoch. robust. In this paper is shown a comparative of competitive learning algorithms to classify measured PD activities into underlaying insulation defects or source that generate PD’s using Self Organizing Maps (SOM). Tool: Self Organizing Map (SOM). Tool: FuzzyNeural Networks. . OchoaZezzatti Alberto. Tool: Decision Tree Objective: Feature Extraction and PD classification Constraints Recognition of different sources in the same sample Kim et al [10] RiCheng et al [11] Performance in the case of multiple discharges and including defects and noises.
switchgears. the partial discharges are in consequence of local stress in the insulation or on the surface of the insulation. This phenomenon has a damaging effect on the equipments. The recorded pulses should be above a certain limit. might be in the form of voids (e. Eapplied is the applied electric field and qphysical is the physical charge [17]. power cables. Example of damage in polymeric power cable from the PD in a cavity to breakdown. In general. The charge that a PD generates in a cavity is called the physical charge and the portion of the cavity surface that the PD affects is called the discharge area. Surface Discharges – Occurs on the surface of an electrical insulation where the tangential field is high e. Internal Discharges – Occurs inside a cavity that is surrounded completely by insulation material.Competitive Learning for Self Organizing Maps 351 2. thermal stress and mechanical stress. CORE INSULATION BREAKDOWN PD EVOLUTION GROUND Fig. whereby the ageing depends on numerous parameters like electrical stress. for example transformers.g. The pulse repetition frequency N is the . The first approach in a diagnosis is selecting the different features to classify measured PD activities into underlying insulation defects or source that generate PD’s. dried out regions in oil impregnated papercables). Polymers: 4000 kV/cm They are generally divided into three different groups because of their different origins: Corona Discharges – Occurs in gases or liquids caused by concentrated electric fields at any sharp points on the electrodes. The pulse repetition rate n is given by the number of partial discharge pulses recorded in a selected time interval and the duration of this time interval. The partial discharge measurement is a typical nondestructive test and it can be used to judge the insulation performance at the beginning of the service time taking into account the reduction of the performance during the service time by ageing. end windings of stator windings. 1.g. Partial discharge: concepts Partial discharges occur wherever the electrical field is higher than the breakdown field of an insulating medium: Air: 27 kV/cm (1 bar). depending on the measuring system as well as on the noise level during the measurement. SF6: 360 kV/cm (4 bar). and others (see Figure 1). In particular for solid insulation like XLPE on power cables where a complete breakdown seriously damages the test object the partial discharge measurement is a tool for quality assessment.
an input vector x is randomly selected and the Euclidean distances between x and w are computed. an ndimensional weight vector w is associated. Therefore. Supervised algorithms [8. For PD diagnosis test.352 JaramilloVacio Rubén. features selection has been widely used in other field. OchoaZezzatti Alberto. In order to choose the features. but the SOM does not have this limitation. The SOM algorithm [15. it is important to know the different source of PD. Ponce Julio number of partial discharge pulses per second in the case of equidistant pulses. the test engineer is responsible for choosing proper methods to diagnosis for the given problem. 9] like multilayered perceptron. The voltage phase angle is divided into small equal windows. pressure. the phase angle and the time of occurrence ti are information on the partial discharge pulse in relation to the phase angle or time of the applied voltage with period T. is very important to classify measured PD activities. is the most popular neural network models. the occurrence of PD depends on many factors. applied voltage and test duration. n is the dimension of input vector. such as data mining [19] and pattern recognition using neural networks [810]. The neighborhood relation of neuron i. The analysis aims to calculate the integrated parameters for each phase window and to plot them against the phase position (). 3. moreover PD signals contain noise and interference [18]. nevertheless.21] is based on unsupervised competitive learning called winner – takes – all. which means that the training in entirely datadriven and that the neurons of the map compete with each other. The image of the input vector and the SOM grid is thus defined as the nearest unit wik and bestmatching unit (BMU) whose weight vector is closest to the x [21]: . At each training step. Furthermore. The SOM is a neural network model that implements a characteristics nonlinear mapping from the highdimensional space of input signal onto a typically 2dimensional grid of neurons. This research only shows test on laboratory without environment noise source. Markalous [20] presented the noise levels on site based on previous experiences. The SOM is a twolayer neural network that consists of an input layer in a line and an output layer constructed of neurons in a twodimensional grid. where qm is peak discharge magnitude. and it is a condition that does not represent the conditions on site. The phase resolved analysis investigates the PD pattern in relation to the variable frequency AC cycle. Self Organizing Map (SOM) 3. required that the target values for each data vector are known. an alternative is though pattern recognition. since PD is a stochastic process.1 Winner takes all The Self Organizing Map developed by Kohonen. such as temperature. (qm −): the peak discharge magnitude for each phase window plotted against. This task can be challenging. namely.
3. Each centre counts the number of times when it has won the competition and reduces its learning rate consequently.. The selection of the distance criterion depends on application. In this paper. The FSCL algorithm is an extension of kmeans algorithm.. it may never be updated. which means that if a centre is inappropriately chosen. N where n is the inputs. The topological error te is the proportion of all data vectors for which first and second BMUs are adjacent units.2 The frequency sensitive competitive algorithm The kmeans algorithm has also the “dead units” problem.. The quantization error qe is the average distance between each data vector and its BMU. the relative winning frequency i of the centre ci defined as: . There are many ways to measure them. The number of panels in the SOM is according the A x B neurons. The quantization error (qe) is calculated to measure the quality of the map. Resolution and topology preservation are generally used to measure SOM quality [24]. Euclidean distance is used because it is widely worn with SOM [23]. i x j wij wij x j wij to i c wij to ic (3) where c denote the neighborhood kernel around the bestmatching unit and α is the learning rated and δ is the neighborhood function. wi w k ik xk 2 (2) The weight vectors in the bestmatching unit and its neighbors on the grid are moved towards the input vector according the following rule: wij c. otherwise this is regarded as violation of topology and thus penalized by increasing the error value. measuring map resolution. thus it may never represent a class. To solve the “dead units” problem it has been introduced the so called “frequency sensitive competitive learning” algorithm (FSCL) [25] or competitive algorithm “with conscience”.. obtained by modifying relation (2) according to the following one: (4) j arg min i x n ci n i 1. It is complicated to measure the quality of an SOM. N represents the centres numbers. the Umatrix representation is a matrix U ((2A1) x (2B1)) dimensional [22]. If a center has won too often “it feels guilty” and it pulls itself out of the competition.Competitive Learning for Self Organizing Maps 353 D x.
3 The rival penalized competitive learning algorithm The rival penalized competitive algorithm (RPCL) [25] performs appropriate clustering without knowing the clusters number. .354 JaramilloVacio Rubén. In other words. than the centres vectors will converge towards the centroids of the input data classes. After selecting out the winner. suppose that the number of classes is unknown and the centres number N is greater than the clusters number. It determines not only the winning centre j but also the second winning center r.. So the centers that have won the competition during the past have a reduced chance to win again. the number of the output data classes will be automatically found. in each iteration. converging much faster than the kmeans and the FSCL algorithms.. the FSCL algorithm updates the winner with next equation: ci n 1 ci n x n ci n (7) is the learning rate. in the same way as the kmeans algorithm. So the learning law can be synthesized in the following relation: ci n x n ci n if i j ci n 1 ci n x n ci n if i j ci n if i j and i r (10) If the learning speed is chosen much greater than . Ponce Julio i si s i 1 n (6) i where si is the number of times when the centre ci was declared winner in the past. OchoaZezzatti Alberto.. with at least one order of magnitude.. i 1. and meanwhile adjusting the corresponding si with the following relation: si n 1 si n 1 (8) 3. proportional with their frequency term. named rival r arg min i x n ci n . N i j (9) The second winning centre will move away its centre from the input with a ratio learning rate. All the other centres vectors will not change. The RPCL will move away the rival.
2. Corona was produced with a point to hemisphere configuration: needle at high voltage and hemispherical cup at ground. was used a neurons array of 1010 cells to extract features. The component interaction between SOM In table 2 are shower the parameters for training to each competitive learning algorithm. for this reason the input in the SOM the raw data is the peak discharge magnitude for each phase window plotted against (qm − ). . Moreover a too large number of neuron per class implies a long training time and a possible underutilization of some of the neural units. a too small number of neurons per class could be not sufficient to represent the variability of the samples to be classified. As it is well known. The coefficient has been dynamically changed during the training. QUALITY MEASURES FOR SOM DATA UNSUPERVISED MODULE NORMALIZATION/ PREPROCESSING SOM (TRAINING) DATAWAREHOUSE (KNOWLEDGE) FAULT DIAGNOSIS/ DETECTION Fig. Surface discharge XLPE cable with no stress relief termination applied to the two ends. High voltage was applied to the cable inner conductor and the cable sheath was grounded. In the cases analyzed. Were considered the pattern characteristic of univariate phaseresolved distributions as inputs. while a too large number in general makes the net too much specialized on the samples belonging to the training set and consequently reduces its generalization capability. Internal discharge was used a power cable with a fault due to electrical treeing. this produces discharges along the outer insulation surface at the cable ends. in fact. Analysis of PD Data PD measurements for power cables are generated and recorded through laboratory tests. Figure 2 shows the conceptual diagram training.Competitive Learning for Self Organizing Maps 355 4. the original dataset is 1 million of items. the magnitude of PD is the most important input as it shows the level of danger.
99 0.9 0.356 JaramilloVacio Rubén. 4 and 5 is showed the performance of the competitive learning algorithms to different PD source.97 20 40 60 80 100 0.8 5.01 Quantization Error per Training Epoch (Surface Discharge) 10.4 6.96 0 20 40 60 80 100 Fig. 4.96 0.01 FSCL 100 0. Quantization and Topological error per Training Epoch (Corona Discharge) In figure 3. Ponce Julio Table 2 Parameter for training Epoch WTA 100 0.98 0. Quantization and Topological error per Training Epoch (Internal Discharge) Quantization Error per Training Epoch (Corona Discharge) 1 0.99 0.95 0.5 Topological Error per Training Epoch (Surface Discharge) 1 WTA FSCL RPCL 0.1 0.98 5.8 0.75 0. 5. OchoaZezzatti Alberto.98 9.5 0.97 9 0 20 40 60 80 100 0.7 0 20 40 60 80 100 1 Topological Error per Training Epoch (Corona Discharge) WTA FSCL RPCL 0.97 0.96 0 20 40 60 80 WTA FSCL RPCL 100 Fig.1 0.2 6 WTA FSCL RPCL Topological Error per Training Epoch (Internal Discharge) 1 0. 3.94 0 20 40 60 80 100 WTA FSCL RPCL Fig.99 10 WTA FSCL RPCL 0.05 0.85 0. .95 0.6 5.01 RPCL 100 0.4 0 0. Quantization and Topological error per Training Epoch (Surface Discharge) Quantization Error per Training Epoch (Internal Discharge) 6.
1981. IEEE Trans. and it is observer that the WTA works with less error and less time of training.16.97 1226 seconds 5.5 5.4 0.78 0. topological error and time in seconds per training epoch are showed in table 3.99 173 seconds 222 seconds Corona Discharge 0. Math Comput Model 1993. Pattern Recognition with Fuzzy Objective Function Algorithms.Partial discharge measurements. 28 – 199 Bezdek JC. Pollard D.8 0.Competitive Learning for Self Organizing Maps 357 We evaluate the performance of algorithms in which competitive learning is applied of partial discharge dataset. but the FSCL is not always satisfactory because the training time is very long.99 0. its showed that WTA is the better algorithm with less error and training time. Information Theory (1982). New York: Wiley. (2000). This paper shows the performance of SOM using different competitive learning algorithms to classify measured PD activities into underlaying insulation defects or source that generate PD’s. 15 – 16. but its overall performance are not always satisfactory.9 0. Table 3 Performance of training in SOM qe te time qe te time qe te time WTA FSCL Surface Discharge 9. A survey of fuzzy clustering. Klawonn F. Kruse R. quantization error. . 1999.98 241 seconds 0. References [1] [2] [3] [4] [5] IEC 60270 Ed. Yang MS. being alternative in accord at the performance FSCL or RPCL algorithms.0 9.85 0.75 0. Plenum Press. Quantization and the method of kmeans.18:1. The RCPL have longest training time but is the algorithm with more error. Runkler T. Fuzzy Cluster Analysis: Methods for Classification Data Analysis and Image Recognition.95 1362 seconds Conclusion PD patterns recognition and classification require an understanding of the traits commonly associated with the different source and relationship between observed PD activity and responsible defect sources.1 0. Höppner F.98 0.985 1 849 seconds 1160 seconds Internal Discharge 5. 2. Highvoltage test techniques .98 889 seconds 1191 seconds RPCL 9.
X. Analysis of Partial Discharge Measurement Data Using a Support Vector Machine. Hao. Mazroua. 600603. Automated Recognition of Partial Discharge.. Stuttgart University: PhD Thesis. H. WSEAS Transaction on Computer Research. Partial discharge studies with variable frequency of the applied voltage. Vesanto J. A. no. Nr. 53 – 57. Lu. Data Clustering. Oh. A. Proceedings of the IEEE. PD pattern recognition with neural netwoks using the multilayer perception technique. KTH Electrical Engineering. and A. Application of Self Organizing Map Approach to Partial Discharge Pattern Recognition of CastResin Current Transformers. (2008). Ahalt. Modelling of cavity partial discharges at variable applied frequency. K. IEEE Transactions on Neural Networks . Liu.. Kohonen T. Volume 3. no. D. 10821089. GouZheng. vol. Lai. K. L. S. 3. P. (2004). RiCheng. and D. Krishnamurthy. Krivda. AUPEC. (2001). J. Lewin. (2007). Yohagy. Edin. Grzybowski. A. 796821. J. (2008). 6782. C. IEEE Transactions on Dielectrics and Electrical Insulation. Pölzlbauer. ASASIAM. Alhoniemi E. rbf net and curve detection” IEEE Trans. (2006). Engineering Applications of Self Organizing Map. Chen. Hirose. Sweden: Doctoral Thesis in Electrical Systems. 12521260. C. Melton "Competitive learning algorithms for vector quantization". [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20] [21] [22] [23] [24] [25] . Kantardzic. Chun. E. 2227–2235. (2002). pp. Vol. “Combining recurrent neural networks with selforganizing maps for channel equalization. B. Madrid Politechnic University. 4. W.” IEEE Trans. E85E. 2007. L. 28. 1993. Kai. IEEE International Power Modulators and High Voltage Conference. (2008). KTH Electrical Engineering. Ohtsuka. S. Detection and location of Partial Discharges in Power Transformers using acoustic and electromagnetic signals. Tsuru. 15. Hikita. J. Ponce Julio [6] X. Issue 3. Vesanto. Descriptive Data Mining of Partial Discharge using Decision Tree with genetic algorithms. Neural Networks. (1995). (1996). L. P. Study on Partial Discharge Localization by Ultrasonic Measuring in Power Transformer Based on Particle Swarm Optimization. G. pp. 277290. H. and T. M. IEEE Transactions on Electrical Insulation. P. Park. 11. OchoaZezzatti Alberto. (2008). S. Helsinki University of Technology. Markalous. pages 10821089. S. S. 3. International Conference on High Voltage Engineering and Application. on Neural Networks. L.358 JaramilloVacio Rubén. pp. 28. N. Krzyzak. “Rival penalized competitive learning for clustering analysis. J. 5th Student Conference on Research and Development. Data Exploration Process Based on the Self Organizing Map. IEEE Transactions on Dielectrics and Electrical Insulation. Xu. 2002. B. (2008). K. H. Theory. 10 Kim. A. (2008). PhD Thesis. Proceedings of the Fifth Workshop on Data Analysis (2004). M. Oja. E. Partial Discharge Pattern Recognition Using FuzzyNeural Networks (FNNs) Algorithm. 3. Choi. Fadilah Ab Aziz. Diagnosis of Electric Power Apparatus using the Decision Tree Method. 272275. Chang. Doctoral Thesis in Computer Science. ShaoYu. (1990). Ichimaru. RubioSánchez. (2000). Wang. 142151. H. Survey and Comparison of Quality Measures for DelfOrganizing Maps. Nuevos Métodos para Análisis Visual de Mapas Autoorganizativos. Sweden: Doctoral Thesis in Electrical Systems. (1993). M. 16. Clustering of the Self Organizing Map. vol. Forssén. on Communications. 636–64. W. Yang. Algorithms and Methods.