3-D Partial Discharge Patterns Recognition of Power Transformers Using Neural Networks
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
3-D Partial Discharge Patterns Recognition of Power
Transformers Using Neural Networks
Hung-Cheng Chen1, Po-Hung Chen2, and Chien-Ming Chou1
1
National Chin-Yi Institute of Technology,
Institute of Information and Electrical Energy,
Taiping, Taichung, 411, Taiwan, R.O.C.
{hcchen, s49312008}@chinyi.ncit.edu.tw
2
St. John’s University,
Department of Electrical Engineering,
Taipei, Taiwan, R.O.C.
phchen@mail.sju.edu.tw
Abstract. Partial discharge (PD) pattern recognition is an important tool in HV
insulation diagnosis. A PD pattern recognition approach of HV power
transformers based on a neural network is proposed in this paper. A commercial
PD detector is firstly used to measure the 3-D PD patterns of epoxy resin power
transformers. Then, two fractal features (fractal dimension and lacunarity)
extracted from the raw 3-D PD patterns are presented for the neural-
network-based (NN-based) recognition system. The system can quickly and
stably learn to categorize input patterns and permit adaptive processes to access
significant new information. To demonstrate the effectiveness of the proposed
method, the recognition ability is investigated on 150 sets of field tested PD
patterns of epoxy resin power transformers. Different types of PD within power
transformers are identified with rather encouraged results.
1 Introduction
Power transformers play a crucial role in operation of transmission and distribution
systems. A dielectric failure in a power transformer could result in unplanned outages
of power systems, which affects a large number of customers [1]. Therefore, it is of
great importance to detect incipient failures in power transformers as early as possible,
so that they can be switched safely and improve the reliability of the power systems.
Partial discharges phenomenon usually originates from insulation defects and is an
important symptom to detect incipient failures in power transformers. Classification of
different types of PDs is of importance for the diagnosis of the quality of HV power
transformers. PD behavior can be represented in various ways. Because of the
randomization of PD activity, one of the most popular representations is the
statistics-based φ-Q-N distribution, i.e., the PD pattern is described using a pulse count
N versus pulse height Q and phase angle φ diagram. Previous experimental results have
adequately demonstrated that φ-Q-N distributions are strongly dependent upon PD
J. Wang et al. (Eds.): ISNN 2006, LNCS 3972, pp. 1324 – 1331, 2006.
© Springer-Verlag Berlin Heidelberg 20063-D Partial Discharge Patterns Recognition of Power Transformers 1325 sources, therefore the 3-D patterns can be used to characterize insulation defects [2]. This provides the basis for pattern recognition techniques that can identify the different types of defects. The automated recognition of PD patterns has been widely studied recently. Various pattern recognition techniques have been proposed, including expert systems [3], fuzzy clustering [4], and neural networks (NNs) [5], [6]. The expert system and fuzzy approaches require human expertise, and have been successfully applied to this field. However, there are some difficulties in acquiring knowledge and in maintaining the database. NNs can directly acquire experience from the training data, and overcome some of the shortcomings of the expert system. However, the raw values of 3-D patterns were used with the NN for PD recognition in previous studies [7], the main drawbacks are that the structure of the NN has a great number of neurons with connections, and time-consuming in training. To improve the performance, two fractal features that extract relevant characteristics from the raw 3-D PD patterns are presented for the proposed NN-based classifier. It can quickly and stably learn to categorize input patterns and permit adaptive processes to access significant new information. To demonstrate the effectiveness of the proposed method, 150 sets of field-test PD patterns from HV epoxy resin power transformers are tested. Results of the studied cases show that different types of PD within power transformers are identified with rather encouraged results. 2 Fractal Features of 3-D PD Patterns for Recognition Purposes 2.1 Fractal Theory Fractals have been very successfully used to address the problem of modeling and to provide a description of naturally occurring phenomena and shapes, wherein conventional and existing mathematical methods were found to be inadequate. In recent years, this technique has attracted increased attention for classification of textures and objects present in images and natural scenes, and for modeling complex physical processes. In this theory, fractal dimensions are allowed to depict surface asperity of complicated geometric things. Therefore, it’s possible to study complex objects with simplified formulas and fewer parameters [8]. PD also is a natural phenomenon occurring in electrical insulation systems, which invariably contain tiny defects and non-uniformities, and gives rise to a variety of complex shapes and surfaces, both in a physical sense as well as in the shape of 3-D PD patterns acquired using digital PD detector. This complex nature of the PD pattern shapes and the ability of fractal geometry to model complex shapes, is the main reason which encouraged the authors to make an attempt to study its feasibility for PD pattern interpretation. 2.2 Calculation of Fractal Dimension While the definition of fractal dimension by self-similarity is straightforward, it is often difficult to estimate/compute for a given image data. However, a related measure of
1326 H.-C. Chen, P.-H. Chen, and C.-M. Chou
fractal dimension, the box dimension, can be computed more easily. In this work, the
method suggested by Voss [9] for the computation of fractal dimension D from the
image data has been followed. Let p(m,L) define the probability that there are m points
within a box of size L (i.e. cube of side L), which is centered about a point on the image
surface. P(m,L) is normalized, as below, for all L.
N
∑ p(m, L) = 1 (1)
m =1
where N is the number of possible points within the box. Let S be the number of image
points (i.e. pixels in an image). If one overlays the image with boxes of side L, then the
number of boxes with m points inside the box is (S/m)p(m,L). Therefore, the expected
total number of boxes needed to cover the whole image is
N S N 1
N ( L) = ∑ p (m, L) = S ∑ p(m, L) (2)
m =1 m m =1 m
Hence, if we let
N 1
N ( L) = ∑ p(m, L) (3)
m =1 m
this value is also proportional to L-D and the box dimension can be estimated by
calculating p(m,L) and N(L) for various values of L, and by doing a least square fit on
[log(L), - log(N(L))]. To estimate p(m,L), one must center the cube of size L around an
image point and count the number of neighboring points m, that fall within the cube.
Accumulating the occurrences of each number of neighboring points over the image
gives the frequency of occurrence of m. This is normalized to obtain p(m,L). Values of
L are chosen to be odd to simplify the centering process. Also, the centering and
counting activity is restricted to pixels having all their neighbors inside the image. This
obviously will leave out image portions of width = (L – 1)/2 on the borders. This
reduced image is then considered for the counting process. As is seen, large values of L
results in increased image areas from being excluded during the counting process,
thereby increasing uncertainty about counts near border areas of the image. This is one
of the sources of errors for the estimation of p(m,L) and thereby D. Additionally, the
computation time grows with the L value. Hence, L = 3, 5, 7, 9 and 11 were chosen for
this work.
2.3 Calculation of Lacunarity
Theoretically, ideal fractal could confirm to statistical similarity for all scales. In other
words, fractal dimensions are independent of scales. However, it has been observed that
fractal dimension alone is insufficient for purposes of discrimination, since two
differently appearing surfaces could have the same value of D. To overcome this,3-D Partial Discharge Patterns Recognition of Power Transformers 1327
Mandelbrot [l0] introduced the term called lacunarity Λ, which quantifies the denseness
of an image surface. Many definitions of this term have been proposed and the basic
idea in all these is to quantify the ‘gaps or lacunae’ present in a given surface. One of
the useful definitions of this term as suggested by Mandelbrot [l0] is
M 2 ( L) − ( M ( L)) 2
Λ (L ) = (4)
(( M ( L)) 2
where
N
M ( L) = ∑ m p(m, L) (5)
m =1
N
M 2 ( L) = ∑ m 2 p(m, L) (6)
m =1
From the definition, we can obtain the idea that lacunarity reflects the density of fractal
surfaces, namely the extent to which the density is. The smoother the surfaces, the less
the lacunarity Λ(L).
3 Discharge Experiments
In this paper, the tested object is an cast-resin HV power transformers that uses epoxy
resin for HV insulation. The rated voltage and capacity of the tested HV power
transformers are 12 kV and 2kVA, respectively. For testing purposes, four
experimental models of power transformers with artificial insulation defects were
purposely manufactured by an electrical manufacturer. The four PD models, including
no defect, HV corona discharge, low voltage (LV) coil PD, and high voltage (HV) coil
PD, are named Type I, II, III, and IV, respectively. In the testing process, all of the
measuring data are digitally converted in order to store them in the computer. Then, the
PD pattern classifier can automatically recognize the different defect types of the
testing objects.
The individual 3-D PD patterns (stored as a 256x256 matrix) are plotted. The x and y
axes correspond to the phase and amplitude (or charge), respectively. The matrix
elements correspond to the pulse count data (or the z axis of the 3-D pattern). An
example 3-D plot of the pattern from each one of the four types is given in Fig. 1. In
order to simplify the calculation of both fractal dimension and lacunarity, a real
gray-scaled image would be utilized instead of 3-D patterns. The amplitude values are
linearly mapped to the varying intensities of the white color (uniformly mapped to one
of the 16 gray colors in this work). This gray image is the input for computing the
fractal features.1328 H.-C. Chen, P.-H. Chen, and C.-M. Chou
(a) (b)
(c) (d)
Fig. 1. Four typical defect types of PD pattern. (a) No defect (Type I). (b) HV corona discharge
(Type II). (c) LV coil PD (Type III). (d) HV coil PD (Type IV).
Fig. 2. Sample plot of the set [log(L), -log(N(L)) Fig. 3. Sample plot of the variation of
for different value of box size L lacunarity with respect to box size L
Fig. 2 is a sample plot of the set [log(L), -log(N(L))] for the five chosen values of L
(computed for one of the pattern examples from Type III). A least square fit to this data
set is performed to obtain the fractal dimension D. The corresponding lacunarity is also
computed for each value of L. Fig. 3 shows its variation with respect to L. These two3-D Partial Discharge Patterns Recognition of Power Transformers 1329
fractal features are computed for all the available patterns. Fig. 4 is a plot of fractal
dimension and lacunarity of different discharge models. Lacunarity was found to be
maximum for all the pattern examples considered, at L = 3 and so, this L value was
chosen for convenience.
Fig. 4. Fractal dimension and lacunarity of different discharge models
4 Recognition Results and Discussion
Three neural network paradigms, back propagation network (BPN), probabilistic
neural network (PNN), and learning vector quantization (LVQ), are utilized to classify
PD pattern of the models. Four layers feed forward structure is used for the pattern
recognition system. Its topological structure is shown in Fig. 5. The neuron number of
its input is determined by the number of fractal features, viz., fractal dimension and
lacunarity. The neuron number of both hidden layers is 6. The neuron number of output
layer is determined by the number of patterns to be identified, which is 4 in this study.
To demonstrate the recognition ability, 150 sets of field test PD patterns are used to test
the proposed PD recognition system. The four defect models of 12-kV epoxy resin
power transformers include the no-defect, HV corona discharge, LV coil PD, and HV
coil PD, respectively. The NN-based PD recognition system randomly chooses 80
instances from the field test data as the training data set, and the rest of the instances of
the field test data are the testing data set. Table 1 shows the recognized results of the
proposed system with different input patterns. The recognition rates of the proposed
system are quite high with about 100%, 94% and 98% for BPN, PNN, and LVQ,
respectively. It is obvious that the NN-based PD recognition system has strong
generalized capability. The recognized results of the three neural networks are almost
of the same accuracy.
The field test data would unavoidably contain some noise and uncertainties which
originate in environmental noise, transducers, or human mistakes. To evaluate the fault1330 H.-C. Chen, P.-H. Chen, and C.-M. Chou
tolerance ability, total 150 sets of noise-contained testing data are generated by adding
±5% to ±30% of random, uniformly distributed, noise to the training data to take into
account the noise and uncertainties. The test results with different amounts of noise
added are also shown in Table 1 for the different neural networks. Usually, the
noise-contained data indeed degrade the recognition abilities in proportion to the
amounts of noise added. This table shows that all these neural networks rather
withstand remarkable tolerance to the noise contained in the data. The proposed
recognition systems show good tolerance to added noise, and have high accuracy rates
of 78%, 72% and 70% in extreme noise of 30%.
Type I
Fractal
Dimension
Type II
Type III
Lacunarity
Type IV
Input layer Two hidden layer Output layer
Fig. 5. Topology structure of NN-based pattern recognition system
Table 1. Recognized performances of different neural networks with various noises added
Recognition rate (%)
Proportion of noise
Back Propagation Network Probabilistic Neural network Learning Vector Quantization
(BPN) (PNN) (LVQ)
±0% 100% 94% 98%
±5% 92% 92% 98%
±10% 88% 90% 94%
±15% 86% 90% 94%
±20% 80% 88% 90%
±25% 80% 78% 78%
±30% 78% 72% 70%
5 Conclusions
A method to analyze a PD pattern and identify the type of discharge source is an
important tool for the diagnosis of HV insulation system. A NN-based PD pattern
recognition method for HV power transformers, that uses fractal features to highlight the
more detailed characteristics of the raw 3-D PD patterns, is proposed. This uses a
fractal theory to extract the fractal dimension and lacunarity from the raw 3-D PD
patterns. These fractal features are then applied to a neural network that performs the
classification. The recognition rates of the proposed system are quite high with about3-D Partial Discharge Patterns Recognition of Power Transformers 1331
100%, 94% and 98%, and 78%, 72% and 70% in extreme noise of 30%, for BPN, PNN,
and LVQ, respectively. The present experimental results indicate that this approach is
able to implement an efficient classification with a very high recognition rate.
Acknowledgments
The research was supported in part by the National Science Council of the Republic of
China, under Grant No. NSC93-2213-E-167-021.
Reference
1. Feinberg, R.: Modern Power Transformer Practice. John Wiley & Sons, New York (1979)
2. Krivda, A.: Automated Recognition of Partial Discharge. IEEE Trans. on Dielectrics and
Electrical Insulation 2 (1995) 796-821
3. Satish, L., Gururaj, B.I.: Application of Expert System to Partial Discharge Pattern
Recognition. in CIGRE Study Committee 33 Colloquium, Leningrad, Russia, (1991) Paper
GIGRE SC 33.91
4. Tomsovic, K., Tapper, M., Ingvarsson, T.T.: A Fuzzy Information Approach to Integrating
Different Transformer Diagnostic Methods. IEEE Trans. on Power Delivery 8 (1993)
1638-1643
5. Cho, K.B., Oh, J.Y.: An Overview of Application of Artificial Neural Network to Partial
Discharge Pattern Classification. Proc. of the 5th International Conference on Properties and
Applications of Dielectric Materials 1 (1997) 326-330
6. Zhang, H., Lee, W.J., Kwan, C., Ren, Z., Chen, H., Sheeley, J.: Artificial Neural Network
Based On-Line Partial Discharge Monitoring System for Motors. IEEE Conference of
Industrial and Commercial Power Systems Technical (2005) 125-132
7. Salama, M.M.A., Bartnikas, R.: Determination of Neural Network Topology for Partial
Discharge Pulse Pattern Recognition. IEEE Trans. on Neural Networks 13 (2002) 446-456
8. Satish, L. Zaengl, W.S.: Can Fractal Features Be Used for Recognizing 3-D Partial
Discharge Patterns. IEEE Trans. on Dielectrics and Electrical Insulation 3 (1995) 352-359
9. Voss, R.F.: Random Fractal: Characterization and Measurement. Plenum Press, New York
(1985)
10. Mandelbrot, B.B.: The Fractal Geometry of Nature. Freeman, New York (1983)You can also read
