hejazi2012
hejazi2012
Abstract— A new method for the on-line monitoring of trans- The requirement of safe and reliable operation of power
former winding radial deformations using ultra-wideband sensors transformers leads to study and development of several fault
is presented in this paper. A wideband signal is sent to a simplified detection and conditions monitoring methods. Each method
model of the transformer winding. The measurements of received
signals from the model winding demonstrate the sensitivity of the can be applied to a specific type of problem and has its own
proposed technique to the winding radial deformation. It is shown advantages and disadvantages [5], [6]. In recent years, several
that the deformation volume can be estimated using a proposed off-line methods such as Short Circuit test method (SC)
index. Also, the amount of the radial deformation is estimated [3], Low Voltage Impulse method (LVI) [7] and Frequency
by using a regression tree based on features extracted by the Response Analysis method (FRA) [5] for the detection of the
wavelet transform.
winding deformation have been proposed. In SC test method,
Index Terms— Condition monitoring, decision trees, discrete the short circuit reactance is measured while the transformer
wavelet transforms, nondestructive testing, power transformers, is off-line. In this method, the sensitivity of the reactance to
time of arrival estimation, ultra wideband antennas, ultra wide-
band radar. the winding displacement is very low, and the type and the
location of the mechanical damage in the winding cannot be
determined [5].
I. I NTRODUCTION
In the FRA method, experimental approaches of compar-
RX Disk no. 2
Disk no. 3
Fig. 1. Basic principle of monitoring by using UWB radar sensor. Disk no. 4
Disk no. 5
results. In this method, the scattering parameters are used as a Disk no. 6
fingerprint for the transformer winding [14]–[16]. This method
is also in the research phase and has not been used for any (a)
transformer.
In this paper, a new Nondestructive testing method of the
transformer winding damages has been developed using the
Deformation
high frequency electromagnetic waves analysis in the time
width
domain. The proposed method has been tested on a model
of transformer for different volumes of radial deformation.
The ultra-wideband (UWB) signals used for this method have
very high accuracy for fault detection and can give more
information about the type and location of the fault because
of special characteristics of UWB signals.
According to the U.S. Federal Communications Commis-
sion (FCC), a UWB signal is defined to have an absolute Sector 1 Sector 2 Sector 3
bandwidth of at least 500 MHz or a fractional (relative) (b)
bandwidth of larger than 20% [17]. Large bandwidths of UWB
signals bring many advantages such as penetration through Fig. 2. Modeling of radial deformation. (a) Disk numbering and deformation
obstacles, accurate position estimation, high-speed data trans- length and thickness. (b) Deformation width.
mission and low cost and low power transceiver designs [18].
The penetration capability of a UWB signal is due to its large
frequency spectrum that includes low frequency components designed as a solution for receiving the signals from the inside
as well as high frequency ones. This large spectrum also of transformer. The fundamental of the proposed method is
results in high time resolution, which improves ranging (i.e., based on UWB radar sensor. As shown in Fig. 1, the UWB
distance estimation) accuracy. From a radar viewpoint, short- radar generates short pulses and transmits them through the
pulse UWB techniques exhibit distinct advantages over more transmitting antenna (TX).
conventional radar approaches [19], [20]. Numerous applica- The signal propagates in an environment (transformer oil).
tions of short-pulse technology were developed for short-range When it meets target (transformer winding), a part of the elec-
radar sensing, subsurface sensing [21], metrology, communica- tromagnetic energy is reflected from the object and propagates
tions, and more recently, precision positioning [22] and radar back to the receiving antenna (RX). It has been shown by
imaging [23]. For active Radio Frequency (RF) tracking and simulations and measurements that the received waveforms
positioning applications, short-pulse UWB techniques offer are sensitive to winding axial displacements and radial defor-
distinct advantages in precision time-of-flight measurement, mations [27]–[29]. The proposed approach for interpreting the
multipath immunity for leading edge detection [i.e., first time received signals is based on comparison. The waveform of the
of arrival (TOA)] and low prime power requirements for received signals can be used as a fingerprint of the transformer.
extended-operation RF identification (RFID) tags [24]. In this The best way of comparison, like the FRA method, is time-
research, UWB pulses have been used for on-line monitoring based. The signal received from a healthy transformer can be
of transformer winding mechanical damages, which can be stored in a database. The transformer can be monitored on-
considered as a slowly moving or stationary target. line in any time. If all of the parameters in the test, e.g.
the antenna, transmitter and receiver characteristics and the
II. P ROPOSED M ONITORING M ETHOD U SING antenna location, remain constant, then every deviation in the
UWB S ENSORS received signal from the healthy fingerprint of the transformer
is only due to the mechanical changes in the environment of
High frequency antennas have been used to monitor the the wave propagation.
partial discharge in the transformer [25]. The Ultra High
Frequency (UHF) sensors of the partial discharge detection
system work only in the receiving mode [25], [26]. These III. M EASUREMENT M ETHOD
antennas can be placed inside or outside the transformer tank, A simplified model of transformer HV winding with the
through the dielectric windows, which have been recently ability of modeling the axial displacement and radial defor-
HEJAZI et al.: APPLICATION OF UWB SENSORS FOR ON-LINE MONITORING OF TRANSFORMER WINDING RADIAL DEFORMATIONS 1651
Disk Spacer
Diameter 60 cm 60 cm 180°
180° 0°
Thickness 2 cm 0.5 cm −20 dB −10 dB 0 dB
−20 dB −10 dB 0 dB
Number 6 6
z
TABLE II
S ECTORS D IMENSIONS 270° y 270° y 90°270°
(a) θ (b)
Deformation parameters Sector 1 Sector 2 Sector 3 90° 90°
Width (cm) 0.5 1 2
Thickness (cm) 2 2 2
0°
Area (cm2 ) 1 2 4 0°
0°
180°
180° −20 dB −10 dB 0 dB
Hub −20 dB −10 dB 0 dB
TX
PC
90° 90°
Fig. 3. Connection diagram of measurement set-up.
Disks have been made from Plexiglas sheets which are covered
0°
by a layer of copper. It should be mentioned that the high
frequency electromagnetic waves are totally reflected in the 270° (a) (b) 270°
metal-air interface, so it is not necessary to model the core 270° 90°
90° 90°
that is inside the windings. Disks are separated from each
other by spacers, which have equal thickness. Table I lists the
model dimensions.
One of these disks has been cut in sectors with different 180° 0° 180° 0°
−20 dB −10 dB 0 dB 180° −20 dB −10 dB 0 dB
dimensions. These sectors can be moved in radial direction as
shown in Fig. 2.
The amount of the deformation is characterized by the
deformation length. The thickness of each sector is equal 270° 270°
to the thickness of each disk. But their widths are differ- (c) (d)
ent, as listed in Table II, in order to model different radial
deformations. Fig. 5. Antenna vertical radiation pattern. (a) 3 GHz. (b) 4 GHz. (c) 5 GHz.
(d) 6 GHz.
In the oil-immerged power transformers, the oil is the
propagation medium. In this paper, the propagation medium is
considered to be air and the transformer tank is not modeled, are connected via Ethernet CAT-5 standard cable to a hub and
and the radial deformation of only one phase of the transformer a PC. The received signals are stored in a text file.
has been studied. It is assumed that there is not any high Two UWB antennas have been used for the measurements.
frequency source of electromagnetic waves in the transformer The antenna pattern is omni-directional. Fig. 4 shows the
except the transmitter. antenna azimuthal beam pattern, and Fig. 5 illustrates the
This method which is based on the comparison of results elevation beam pattern. For the azimuthal beam pattern, 0
in different time periods is not sensitive to the environmen- and 180 degrees represent the flat face of the antenna (“bore
tal change such as humidity and temperature. It is because sight”), and 90 and 270 degrees represent the edge of the
variation of humidity, in a maintenance period for an installed antenna. When two radios at the same elevation are rotated
transformer is negligible and it is possible to measure only in so the flat sides of the antennas face one another, due to the
a known winding temperature range. polarization of antennas, radio performance will be approxi-
A bi-static transmitter and receiver have been used for the mately 6 dB higher than the case, in which the antennas are
measurements as shown in Fig. 3. The transmitter and receiver edge-on.
1652 IEEE SENSORS JOURNAL, VOL. 12, NO. 6, JUNE 2012
Winding 1
model
Transmitted signal
0.5
d3 d3
d2
0
TX d1 RX
−0.5
d4 −1
0 0.5 1 1.5 2 2.5
Time (ns)
300 100
100 −100
0
−100 −300
−300 −500
10.8 11 11.2 11.4 11.6 11.8
−500 500
0 5 10 15 20 0 cm
Time (ns) 1 cm
(a) 300 2 cm
Normalized amplitude
500 3 cm
4 cm
100
Normalized amplitude
300
−100
100
0
−100 −300
−300 −500
6 8 10 12 14 16 18 20
–500 Time (ns)
0 5 10 15 20 T1
Time (ns) 100
(b)
60
Fig. 10. Two received pulses (a) before and (b) after synchronization.
20
TABLE IV
PARAMETERS OF M EASUREMENT S ET-U P
Parameter TC TI Ts Td d4 d3 d2 d1
Value 20 ns 50 ms 31.789 ps 5s 90 cm 50 cm 40 cm 60 cm
S1 TABLE V
S2 D IFFERENT C OMBINATIONS OF S ECTORS
S3
S4 Cross Section Deformation width (cm)× Cross Section
No. Deformation thickness (cm) Area (cm2 )
S1 1×2 2
S2 2×2 4
S3 1×2+2×2 6
(a)
S4 2×2+2×2+1×2 10
S5 S5 2×2+2×2+2×2 12
S6 S6 2×2+2×2+2×2+1×2 14
N W1 ( j )
40 j =T1 /Ts
TABLE VI
F EATURES AND T HEIR S YMBOLS E XTRACTED U SING WAVELET
15 0 −1
10 −1 −2
0 5 10 15 20 0 5 10 15 20
5
1 d6 d5
2
0 10 20 30 40 50 0 0
Deformed volume (cm3)
−1 −2
Fig. 14. Calculated MADC for second window versus deformed vol- −2 −4
ume (cm3 ). 0 5 10 15 20 0 5 10 15 20
d4 500 d3
10
an dn d2 d1 0 0
0 fs fs fs fs fs −10 −500
2(n + 1) 2n 8 4 2 0 5 10 15 20 0 5 10 15 20
500 d2 d1
Fig. 15. Frequency ranges cover for details and final approximation. 10
0 0
the calculated MADC for the second window (T1 = 12.076 ns
and T2 = 13.983 ns) versus deformed volume. −500 −10
0 5 10 15 20 0 5 10 15 20
A curve can be fitted to the calculated MADC for both
Time (ns) Time (ns)
windows. The results of the curve fitting are, as follows.
For the first window: Fig. 16. Typical waveform of detail and approximate components of received
signal for daubechies 16.
f (x) = 4.239e − 006 · x 5 + −0.0005416 · x 4 + 0.02353 · x 3
+ − 0.3995 · x 2 + 3.584 · x + 6.43 (6) estimate the amount of the radial deformation for an unknown
state of the winding.
and, for the second window:
f (x) = 9.006e − 006 · x 4 + −0.001053 · x 3 + 0.03569 · x 2 A. Feature Extraction Using Wavelet Transform
+0.06971 · x + 4.293. (7) A wavelet-based signal processing technique [32], [33] is an
effective tool for feature extraction. Some applications of the
The MADC indexes are ascending functions of the deformed technique have been reported for data compression and fault
volume and they can approximately estimate the severity of detection [34]. The wavelet transform is the breaking up of a
the radial deformation using fitted curves. Considering MADC signal into scaled and shifted versions of the mother wavelet
curves, presented in Figs. 13 and 14, it is obvious that they [35]. The wavelet function is localized in time and frequency
have not enough accuracy for finding the exact amount of yielding wavelet coefficients at different scales (levels). Any
radial deformations. But, they can be used to detect an existing discrete signal x[n] can be decomposed by using the wavelet
radial deformation. function and the wavelet coefficients, as follows:
j0
150 Sound
150 4 10
<FFT(d7)
<FFT(a7)
Deformed
|FFT(d7)|
|FFT(a7)| 100 100 2
5
50 50 0 Sound
Deformed
0 −2 0
0 0 5 10 15 0 5 10 15
0 3 6 9 12 15 0 3 6 9 12 15
150 10
150
<FFT(d6)
<FFT(d5)
|FFT(d5)|
|FFT(d6)|
100 100
5 5
50 50
0 0
0 3 6 9 12 15 0 3 6 9 12 15 0
0 5 10 15 0 5 10 15
200 10000 10 50
|FFT(d3)| 7500
|FFT(d4)|
150
25
<FFT(d3)
5
<FFT(d4)
100 5000
50 2500 0 0
0 0 −5 −25
0 3 6 9 12 15 0 3 6 9 12 15
−10 −50
10000 100 0 3 6 9 12 15 0 3 6 9 12 15
|FFT(d2)|
7500
|FFT(d1)|
75
<FFT(d2)
<FFT(d1)
5000 50 25 25
2500 25
0 0 0 0
0 3 6 9 12 15 0 3 6 9 12 15
Frequency (GHz) Frequency (GHz) 0 3 6 9 12 15 0 3 6 9 12 15
Frequency (GHz) Frequency (GHz)
Fig. 17. Magnitude of FFT of details and approximation of two different
received signal.
Fig. 18. Phase of FFT of details and approximation of two different received
signal.
where φ[n] is the scaling function, and φ[n] is the mother a small frequency range of the FFT magnitude is used. It
wavelet, j is the scale of decomposition, k is the shifting is obvious that, the magnitude and the phase of FFT differ
factor, a j 0,k are the approximation coefficients at a scale of significantly for different types of fault. Thus, the features are
s = 2 j 0 , d j,k are the detail coefficients at a scale of s = 2 j the magnitude, the phase and the power of FFT of details and
and N is equal to 2 j , where N is the number of x samples. approximation of received signals, which are listed with their
The DWT divides the given function into different frequency symbols in Table VI.
components based on a power of two divisions. More con-
cretely, if f s (in samples per second) is the sampling rate
used for capturing x, at the j ’th decomposition level, the B. Regression Tree
detail d j and the approximation a j coefficient contains the Classification and regression tree (CART) is extensively
information concerning the original signal components with implemented in machine fault diagnosis, for classification
the frequency bandwidth [ f s /2 j +1 , fs /2 j ] and [0, f s /2 j +1 ], or regression problems depending on the response variable,
respectively [35], [36]. Therefore, DWT carries out the filter- which is either categorical or numerical [31]. Regression tree
ing process shown in Fig. 15. Note that the filtering is not is applied to forecast the short-term load of the power system
ideal, a fact leading to a certain overlap between adjacent [38], [39].
frequency bands. The shape of the frequency response for A binary tree is developed for a regression problem with
these filters depends on the type and the order of the mother the repeated splits of subsets into two next subsets according
wavelet used in the analysis. to independent variables. The data must be as homogenous
Also, the type of wavelet function is important for the fault as possible with respect to the response variables, in order to
detection. Several wavelet functions such as daubechies 4, produce subsets of the data by the regression tree. The trees,
8, 16, 28, 32, symlet 2, 4, 8, coiflet 3 have been tested for which are produced by CART, consist of internal nodes and
a typical radial deformation. It is shown that daubechies 16 terminal nodes or leaf nodes. Each internal node is related to a
maximizes the fault index for the detection of the radial defor- decision function to show the next node, while each terminal
mation [37]. Therefore, daubechies 16 has been selected as the node is the output of a given input vector that leads to this
wavelet function to extract useful features. The approximate node [40].
a7 and details d1 , . . ., d7 for daubechies 16 mother wavelet are Both classification and regression trees built by CART
shown in Fig. 16. consist of the tree growing and pruning.
After the decomposition of signals, the extraction of useful Suppose L be a learning sample of size n, and it consists of
features is studied. The FFT is employed to feature extraction n couples of observations ( y1 , x 1 ), . . . , (yn , x n ), where x i =
from the DWT. The corresponding phase and magnitude of (x 1i , . . . , x di ) is a set of independent variables and yi ∈ R is a
FFT of the approximate a7 and details d1 , . . . , d7 of the response associated with x i [41]. In this paper, yi is a specific
received signals for two different states of winding are shown case of the radial deformation and x i is the feature vector for
in Figs. 17 and 18, respectively. For data reduction, only this case. To prepare the training set, the sector S2 has been
HEJAZI et al.: APPLICATION OF UWB SENSORS FOR ON-LINE MONITORING OF TRANSFORMER WINDING RADIAL DEFORMATIONS 1657
TABLE VII
C ART T EST R ESULTS 4. Training of RT
Estimated Absolute
Deformation (mm)
deformation (mm) error (mm)
2 2 0
4.5 4.5 0 RT training Is RT
7 7.5 0.5 training
complete?
9.5 10 0.5 N
12 12 0
14.5 14 0.5 Y
17 17.5 0.5
19.5 20 0.5
Trained RT (Important
features have been selected)
Javad Ebrahimi (S’10) was born in Isfahan, Iran, Reza Faraji-Dana received the B.Sc. degree (with
in 1986. He received the B.Sc. degree in electrical honors) from the University of Tehran, Tehran, Iran,
engineering from the University of Tabriz, Tabriz, in 1986, and the M.Sc. and Ph.D. degrees from
Iran, in 2008, and the M.S. degree (first class honors) the University of Waterloo, Waterloo, ON, Canada,
in electrical engineering from the Amirkabir Uni- in 1989 and 1993, respectively, all in electrical
versity of Technology (AUT), Tehran, Iran, in 2010. engineering.
He is currently pursuing the Ph.D. degree with the He was a Post-Doctoral Fellow with the University
Electrical Engineering Department, AUT. of Waterloo for one year. In 1994, he joined the
His current research interests include the analysis School of Electrical and Computer Engineering,
and control of power electronic converters, flexible University of Tehran, where he is currently a Profes-
alternative current transmission systems devices, dis- sor. He has been engaged in several academic and
tributed generation, and transformer monitoring. executive responsibilities, among which was his deanship of the Faculty of
Engineering for more than four years until 2002, when he was elected as
the University President by the university council. He was the President of
Gevork B. Gharehpetian (SM’08) received the the University of Tehran until December 2005. He is the author of several
Ph.D. degree (first class honors) in electrical engi- technical papers published in reputable international journals and refereed
neering from the University of Tehran, Tehran, Iran, conference proceedings.
in 1996. Prof. Faraji-Dana is an Associate Member of the Iran Academy of Sciences.
He was an Assistant Professor with the Amirkabir He was the Chairman of the IEEE-Iran Section from 2007 to 2009. He
University of Technology, Tehran, from 1997 to received the Institution of Electrical Engineers Marconi Premium Award in
2003, an Associate Professor from 2004 to 2007, and 1995.
has been a Professor since 2007. He was selected by
the ministry of higher education as the Distinguished
Professor of Iran and the Iranian Association of
Electrical and Electronics Engineers (IAEEE) as the
Distinguished Researcher of Iran. He is the author of more than 450 journal Gholamreza Moradi received the B.Sc. degree from
and conference papers. His current research interests include power systems, the University of Tehran, Tehran, Iran, the M.Sc.
transformers transients, and power electronics applications in power systems. degree from the Iran University of Science and
Prof. Gharehpetian was awarded the National Prize in 2008 and 2010, Technology, Tehran, and the Ph.D. degree from the
respectively. He is a Distinguished Member of IEEE and IAEEE, respectively, Amirkabir University of Technology (AUT), Tehran,
and a member of the Central Board of IAEEE. all in electrical engineering.
He was with Civil Aviation of Technology as a
Faculty Member from 1997 to 2006, then he joined
AUT, where he is an Associate Professor in the
Mohammad Mohammadi was born in Shiraz, Iran, Electrical Engineering Department. He is the co-
in 1977. He received the B.Sc. degree from Shiraz author of five books and numerous papers in his
University, Shiraz, in 2000, and the M.Sc. and research fields. His current research interests include microwave measure-
Ph.D. degrees from the Amirkabir University of ments, numerical measurements, radio frequency circuits, and system designs.
Technology (AUT), Tehran, Iran, in 2002 and 2007,
respectively, all in electrical engineering.
He is currently with Shiraz University. He is the
author of more than 40 journal and conference
papers. His current research interests include power
system security assessment, machine learning, and
power system dynamics.