2156 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 25, NO.

4, OCTOBER 2010

Fault Detection and Classification in EHV

Transmission Line Based on Wavelet
Singular Entropy
Zhengyou He, Member, IEEE, Ling Fu, Sheng Lin, and Zhiqian Bo, Senior Member, IEEE

Abstract—A novel technique for fault detection and classifica- inconstancy, have been widely used in the fault detection and
tion in the extremely high-voltage transmission line using the fault classification by means of traveling-wave or high-frequency
transients is proposed in this paper. The novel technique, called transients.
wavelet singular entropy (WSE), incorporates the advantages of
the wavelet transform, singular value decomposition, and Shannon Based on fault transients, several algorithms have been re-
entropy. WSE is capable of being immune to the noise in the fault ported for fault detection and classification. For all of the pro-
transient and not being affected by the transient magnitude so it posed algorithms, how to extract the transients’ features from
can be used to extract features automatically from fault transients the original fault signal is the most important issue. Wavelet
and express the fault features intuitively and quantitatively even transform (WT), which is the perfect time-frequency localiza-
in the case of high-noise and low-magnitude fault transients. The
WSE-based fault detection is performed in this paper, which proves tion ability, has been chosen as an effective tool for analyzing
the availability and superiority of WSE technique in fault detec- the fault transients [2]–[6].
tion. A novel algorithm based on WSE is put forward for fault clas- Reference [4] proposed an effective feature extraction method
sification and it is verified to be effective and reliable under various using WT, [7] showed that WTs are well suited for the anal-
fault conditions, such as fault type, fault inception time, fault re- ysis of the nonstationary signals measured by the protection de-
sistance, and fault location. Therefore, the proposed WSE-based
fault detection and classification is feasible and has great potential vices, and [8] and [9] showed that the WT has the ability to
in practical applications. perform local analysis of relaying signals without losing the
time-frequency information. WT is used in [10] to capture the
Index Terms—Extremely high-voltage (EHV) transmission line,
fault classification, fault detection, singular value decomposition, high-frequency traveling waves for fault detection, classifica-
wavelet singular entropy, wavelet transform. tion, and phase selection of faults. Reference [11] used the dis-
crete wavelet transform (DWT) to design the fault classification
tool for the boundary protection of series-compensated trans-
I. INTRODUCTION mission lines, and [12] described the DWT-based technique in
detail and pointed out that DWT is an excellent online tool for
AULT detection and classification are two of the most im-
F portant tasks involved in transmission-line relaying [1].
They must be accomplished and as fast and accurate as possible
relaying applications. Wavelet multiresolution analysis (MRA)
is the computing algorithm used by DWT with the automati-
cally adjusted window to extract subband information from fault
to deenergize the system from the harmful faults and restore the transients [13], and it has been proved as an effective tool in
system after faults. analyzing fault transients. Reference [14] presented a method-
The traditional algorithms for fault detection and classifica- ology for fault classification based on MRA and [3] presented
tion, which are mostly based on steady-state components, have a method for disturbance detection and classification based on
difficulties in accelerating the protection speed and in escaping MRA. Wavelet modulus maxima (WMM) has been used in [15]
the impacts of many factors, such as fault type, fault resistance, and [16] to analyze the initial modal current traveling waves,
and fault inception time [1]–[5]. and an effective approach to fast and accurate fault detection
The fault-generated transient components, which contain and fault phase selection has been achieved.
abundant fault information and are immune to the system’s Although the WT performs well in the transient analysis and
some improvements have been achieved in fault detection and
Manuscript received January 15, 2009; revised June 13, 2009, August 19, classification by using WT, there are still several open problems
2009, October 12, 2009. First published August 23, 2010; current version to be solved. In many applications [4]–[16], WT is limited to
published September 22, 2010. This work was supported by the National
show several fancy pictures and its transformed results still con-
Natural Science Foundation of China (50877068) and in part by the Program
for New Century Excellent Talents in University: No. NCET-06-0799. Paper tain a large number of data which need further processing. This
no. TPWRD-00044-2009. greatly hinders the automated feature extraction in fault detec-
Z. He, L. Fu, and S. Lin are with the College of Electrical Engineering, South-
tion and classification.
west Jiaotong University, Chengdu 610031, China (e-mail: hezy@swjtu.cn).
Paper no. TPWRD-00044-2009. Therefore, combined techniques have already been used, such
Z. Bo is with the AREVA T&D—Automation and Information Systems, as WT with ANN [1], [17], [18], and WT with fuzzy logic
Stafford ST17 4LX, U.K. [19]. However, these techniques are dependent on huge sam-
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org. ples and trainings for knowledge representation, leading to an
Digital Object Identifier 10.1109/TPWRD.2010.2042624 excessively complicated job. Also, they cannot manage the un-
0885-8977/$26.00 © 2010 IEEE
HE et al.: FAULT DETECTION AND CLASSIFICATION IN EHV TRANSMISSION LINE 2157

certain factors in the transmission system which will influence be represented by its singular values. If matrix represents the
the reliability of fault detection and classification. time-frequency information of the fault transient, the matrix
Taking the aforementioned problems into account, this paper will represent the basic modal characteristics of . Therefore,
proposes a novel technique for fault detection and classifica- we use SVD to analyze the obtained WT coefficient matrix and
tion in extremely high-voltage (EHV) transmission line. This provide briefly numerical representation for the time–frequency
proposed methodology combines the techniques of WT, sin- distribution of the fault transient.
gular value decomposition (SVD) [20], and Shannon entropy
together; therefore, it is called wavelet singular entropy (WSE) C. Shannon’s Information Entropy
for the acronyms. Shannon’s entropy is an important uncertainty measure for
WSE can be used to extract features from fault transients evaluating structures and patterns of analyzed data. It is defined
quantitatively and automatically. It is immune to the noises and by Claude E. Shannon in 1948 as follows [23].
many other uncertain factors in the system. Further, it is inde- Let be a discrete random variable with
pendent on the magnitude and energy of the transients. The im- possible states. Let , whose values sat-
plementation of fault detection and classification based on WSE isfy the terms of and
is put forward and its efficiency is verified by virtue of the sim- as the probabilities associated with those states. The uncer-
ulation tests in this paper. tainty information of each possible state is

II. DEFINITION OF WAVELET SINGULAR ENTROPY

A. Wavelet Representation of Power System Transients (4)

The definition of continuous WT for a given signal with The Base-e logarithm will be used throughout this paper [21]
respect to a mother wavelet is given as follows [21]: [i.e., in (2)]. We may call the information content of
as self-information which is denoted as . As the is a
(1) random variable, it is not suitable for measuring the uncertainty
of the whole data. Therefore, the mathematical expectation of
where is the scale factor and is the translation factor. is defined as entropy which is denoted by
The coefficients of WT are defined by the following
inner product: (5)

(2) As described in [23], the quantity of has some inter-

esting properties and what we benefit from are as follows.
1) is zero if and only if all of the are zeros but one, and
Generally, WT consists of successive pairs of low- and high-
this nonzero one has the value unity. Intuitively, this means
pass filters. For each pair, the high-scale and low-frequency
that the result is certain since only one event occurs. is
components are called approximations, while the low-scale and
positive in any other situation.
high-frequency components are called details. The approxima-
2) For a given reaches its maximum and is equal to
tions and details form the WT-coefficient matrix that we need.
when all are equal. Intuitively, this is the most uncertain
B. Singular Value Decomposition (SVD) situation.
3) Any change to the equalization of the probabilities
According to the SVD, for any matrix , column- will increase the value of .
orthogonal matrix U, a transpose of an orthogonal matrix These three properties of the entropy qualify itself to measure
V, and a diagonal matrix consequentially exist, which the uncertainty of analyzed system X.
enable to be equivalently represented in the SVD form [22]
D. Definition of Wavelet Singular Entropy (WSE)
(3)
Let , which is a discrete sequence with samples, be the
where signal sequence to be analyzed as follows.
1) First, analyze the by WT, where the “db4” mother
wavelet and 4-scaled WT are chosen in the transformation.
Then, a 4 WT-coefficient matrix can be obtained by
means of (2).
2) Second, decompose the matrix with SVD in (4), and a
singular-value array can be obtained as ,
and its diagonal elements are called “sin- where is the rank of the diagonal matrix . The value
gular values” of matrix . The singular values are all nonneg- of may be very large and the value of as well as
ative and arranged in a nonincreasing order (i.e., its embodied information will decrease with the increase
). It is well known that various intrinsic algebraic of . In order to reduce the computing cost and keeping
properties of a matrix operator can be revealed by SVD and can the hypostasis of , the tiny singular values are neglected
2158 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 25, NO. 4, OCTOBER 2010

in our application and the original singular values are re-

placed by pieces of singular values
which must satisfy the stipulation of
%. Consequently, the value of should be different
under various situations, and these pieces of singular
values are called the effective singular values, which will
make the calculation of WSE become much faster and
more effective.
3) Third, in order to obtain the entropy of the singular-value
Fig. 1. Flowchart of the computation of WSE.
array, the probability associated with is defined as
follows:

(6)

4) Finally, the ordered WSE of is obtained by (5)

(7)

where is the number of effective singular values involved in

the WSE calculation.
According to the aforementioned definition, WSE is used to
Fig. 2. WSE values without noise when the order k = 4,8,16,32.
map the correlative wavelet space into independent linearity
space, and to indicate the uncertainty of the energy distribution
in the time–frequency domain with a high immunity to noise.
Due to its way of implementation, WSE is sensitive to the tran-
sients produced by the faults, and the fewer modes the transients
congregate to, the smaller the WSE is. Therefore, the proposed
WSE will be suitable and useful for measuring the uncertainty
and complexity of the analyzed signals, and will provide an intu-
itive and quantitative outcome for the fault diagnosis which can
be utilized to overcome the drawbacks in the previous method-
ologies.
Fig. 3. WSE values with noise SNR = 10 when the order k = 4,8,16,32.
III. APPLICATION OF WSE IN FAULT DETECTION

A. Validity of WSE in Fault Detection

A representative ideal signal , which is shown in the fol-
lowing subsection form, is taken as the investigated object:

(8)

Fig. 4. WSE values with noise SNR = 5 when the order k = 4,8,16,32.
where 50 Hz and it is the fundamental frequency;
Hz and correspond to the 3rd, 5th, and
7th harmonic frequency, respectively. That is, this 1.5-s-long instants can be calculated as shown in Figs. 2 –4. Figs. 2–4 show
signal contains one frequency component before 0.3 the values of ( 4, 8, 16, and 32) of under dif-
s, two frequency components during 0.3 s, and 1.1 s, ferent Gauss-random noise background: nonnoise 10
and four frequency components after 1.1 s. and 5, respectively.
Using the definition of WSE in Section II, we set the sampling As shown in Figs. 2–4, the value of varies with the
frequency to be 20 kHz, take the 100-sample-long sequence in a change of frequency components in : the greater number of
time window as the input of WSE, and move this time window modes of frequency components, the higher the values
by a step of 100 samples. The order of WSE is chosen as will be. And in each figure, there are two points of sudden in-
4, 8, 16, and 32. Consequently, the results of WSE at associated crements of the value at instants 0.3 s and 1.1
HE et al.: FAULT DETECTION AND CLASSIFICATION IN EHV TRANSMISSION LINE 2159

Fig. 5. Simplified model of the EHV transmission line.

s. These increments are associated with the sudden complica-

tion of frequency components. Accordingly, after the changing
duration, the values reduce to a lower level but are still
higher than before. Therefore, the change of frequency compo-
nents in signal can be perfectly detected by as well
as that in fault signals.
On the other hand, the noise background could affect the
values to some extent, which explains why the wave-
form in Fig. 2 is much smoother than that in Figs. 3 and 4. But
anyway, the changing trends of values in Figs. 3 and 4
are generally similar to those in Fig. 2, and the effect of noise
is too weak to affect the fault detection, which benefits from the
noise immunity of WT.
Furthermore, the order of WSE influences the results.
As shown in Fig. 1, the waveform of 4-ordered is much
lower than the other three, which means the signal informa-
tion extracted by 4-ordered is not as much as by other Fig. 6. WT and WSE of zero-sequence current. (a) Zero-sequence current
ones. In addition, the waveforms of 8-ordered , 16-or- signal s(t). (b) WT of s(t) at scale m = 2 . (c) WT of s(t) at scale m = 2 .
dered , and 32-ordered are nearly on the same (d) WT of s(t) at scale m = 2 . (e) WT of s(t) at scale m = 2 . (f) WT of
s(t) at scale m = 2 . (g) WSE of s(t).
level (i.e., the value tends to be constant when the order
reaches a certain quantity). This is because the embodied in-
formation of singular value will become weaker if becomes
larger. This explains why we chose effective singular values in calculated. The simulation results are shown in Figs. 6 and 7,
the application of WSE. We can conclude from this test that it from which we can see that the values of the current and
is proper to choose 8 in this paper, with enough precision the voltage have a sudden saltation at about 5 ms. A large
and acceptable computing cost. number of simulation tests have been carried out and the results
show that the WSE bears good capability for fault detection in
the EHV transmission line.
B. Fault Detection Tests in Simulations In order to prove the superiority of WSE, it is compared with
wavelet modulus maxima (WMM), which has recently been
A typical model of a 500-kV and 300-km EHV transmission widely used in fault detection [5], [6], [15], [16]. With the model
line with two power sources is established in EMTDC/PSCAD, in Fig. 5, the detection thresholds of WMM and WSE are set to
as shown in Fig. 5. The frequency-dependent(phase) mode is be 0.02 and 0.2, respectively. The comparison under different
chosen as the transmission-line model in order to obtain more conditions is shown in Table I.
accurate results during transient simulations since the feature The comparison with WMM and a couple of comparisons
that the different frequency component has for a different at- with other methodologies, such as WT-ANN [1], [17], [18] or
tenuation degree is incorporated into this model. Besides, the WT fuzzy [19], reveal that: 1) the threshold of WSE is higher
white-noise model has been included in this system and the than that of WMM, in which case the error of the detection re-
signal-to-noise ratio (SNR) is set to 40 according to the actual sult can be reduced and the adaptability of methodology can be
system, so all of the tests below with this system have taken improved; 2) in most conditions, faults can be detected more
the random noise into account. Under the normal condition, the rapidly by WSE than by other algorithms due to the intuitive,
positive-sequence parameters and zero-sequence parameters are nontraining, and nonfuzzy characteristics of WSE; 3) WSE is
/km, /km, S/km, more effective than others in the case of low-energy fault tran-
and /km, /km, sients, and this is because WSE takes all WT-scaled features
S/km, respectively. into account and considers their relative, instead of absolute,
Assume that a single-phase-to-ground fault occurs at 5 values as well; 4) failure in choosing the suitable scale of W.
ms. Set the sampling frequency to be 20 kHz, and the width time Other WT-involved methodologies tend to be affected by the
window and window step are both set as previously mentioned; signal magnitude while WSE is not. Therefore, WSE is better
the order of WSE is chosen as 8. Then, the zero-sequence and more applicable than most of the other previous methodolo-
current and voltage signals can be obtained and the WSE can be gies in terms of fault detection.
2160 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 25, NO. 4, OCTOBER 2010

TABLE II
EFFECTIVE SINGULAR VALUES OF THE PHASE-C FAULT TRANSIENT

Fig. 7. WT and WSE of the zero-sequence voltage. (a) Zero-sequence voltage

signal s(t). (b) WT of s(t) at scale m = 2 . (c) WT of s(t) at scale m = 2 .
(d) WT of s(t) at scale m = 2 . (e) WT of s(t) at scale m = 2 . (f) WT of values can extract the effective features from fault transient and
s(t) at scale m = 2 . (g) WSE of s(t).
express them in a simple way, and WSE can be used to discrim-
inate the faults intuitively and quantitatively. Therefore, WSE is
TABLE I a feasible technique for fault classification in EHV transmission
CONTRAST OF THE TWO METHODS APPLIED TO FAULT DETECTION: (R IS lines.
THE FAULT RESISTANCE; D IS THE FAULT LOCATION; AND “*” EXPRESSES
DISABILITY IN DETECTION)
B. Fault Classification Algorithm
It is known from the aforementioned investigations that the
WSE value is different between the faulty phase and sound
phase, and the WSE values of the faulty phase are much greater
than that of the sound phase. Also seen before, the uncertainty
in system and fault conditions would influence the WSE value
as well as the classification results. In order to remove the
effect of uncertainty and obtain reliable fault classification
IV. APPLICATION OF WSE IN FAULT CLASSIFICATION under various conditions, three indicators ( , and )
are introduced here to indicate the WSE value of each phase
A. Validity of WSE in Fault Classification
The system containing a 500-kV and 300-km transmission (9)
line in Fig. 5 is used for the investigation. Under different fault
types, the effective singular values of the phase-C fault transient
are described in Table II. In this table, the boldfaced data high- (10)
light the cases of the phase-C-involved fault.
It can be seen from Table II, in the case of the phase-C-in-
volved fault, that the eight effective singular values are relatively (11)
more average and their value is larger than other faults.
As seen, the values of phase-C-unconcerned faults are all where and indicate phase A, B, and C respectively;
smaller than 1 and that of the phase-C-involved are all greater denotes the -ordered WSE of the phase signal during
than 1.6 except for the “ac” fault. The value of the “ac” the first half cycle after fault inception. The WSE ratio, such
fault is 1.3774, which is smaller than other phase-C-involved as , indicates the relative differences
faults but is still much greater than other phase-C-unconcerned between faulty phase and sound phase, which is used to high-
faults. The reason rests with the effect of uncertainty in system light the difference of faulty phases. The summation of WSE
and fault condition, and this uncertainty will be dealt with in ratios is used to accumulate the differences of faulty phase
the following application. Generally speaking, effective singular under different ordered WSE. By virtue of this treatment, more
HE et al.: FAULT DETECTION AND CLASSIFICATION IN EHV TRANSMISSION LINE 2161

TABLE III
RESULTS OF FAULT CLASSIFICATION FOR VARIOUS FAULT TYPES (FAULT
INCEPTION TIME: A-PHASE VOLTAGE CROSS ZERO; FAULT LOCATION: 50 km
AWAY FROM THE BUS; TRANSITION RESISTANCES: 10
)

Fig. 8. Flowchart of the WSE-based fault classification algorithm. TABLE IV

RESULTS OF FAULT CLASSIFICATION FOR VARIOUS FAULT INCEPTION ANGLES
(FAULT LOCATION: 50 km AWAY FROM BUS; TRANSITION RESISTANCES: 0
)

TABLE V
RESULTS OF FAULT CLASSIFICATION FOR VARIOUS RESISTANCES (FAULT TIME:
A-PHASE VOLTAGE CROSS ZERO; FAULT LOCATION: 50 km AWAY FROM THE
BUS)

Fig. 9. WSE of different fault types.

accurate and reliable fault classification can be achieved. The

flowchart of WSE-based fault classification is shown in Fig. 8.
In Fig. 8, the values of threshold are determined ac-
cording to the system situation, generally in virtue of ,
and . When a three-phase-to-ground fault occurs, and
are set to be about the minimum and maximum, respectively, of
, and is set to be no less than . In this paper, we
set and for the simulation tests.
fault inception) as the input of WSE. The test results for var-
ious fault types, various fault inception angles, various fault re-
C. Fault Classification Tests in Simulation
sistance, and various fault locations are shown from Figs. 9–12
The system shown in Fig. 6 is used for simulation tests. The and from Tables III–VI.
current as well as the voltage transient of each phase, which is 1) Fault Classification for Various Fault Types: The cases of
measured in one end, is analyzed in the case of fault classifica- 11 types of faults are simulated and each case is repeated for 5
tion. Their performances are similar to each other. Therefore, times, and 55 sets of results are obtained.
we only illuminate the test results of voltage transient to verify The WSE of each phase in the case of various fault types
the algorithm. The sampling frequency is set to be 20 kHz and is shown in Fig. 9, where the abscissa is denoted as the WSE
the “db4” mother wavelet and 4-scaled WT are chosen, and we order and the ordinate is denoted as WSE values. As shown in
take the 200-sample-long sequence (i.e., half-cycle data after Fig. 9, the WSEs of the faulty phase are higher than that of the
2162 IEEE TRANSACTIONS ON POWER DELIVERY, VOL. 25, NO. 4, OCTOBER 2010

TABLE VI TABLE VII

RESULTS OF FAULT CLASSIFICATION FOR VARIOUS LOCATIONS (FROM SUMMARY OF THE RESULTS OBTAINED IN FAULT CLASSIFICATION: (t IS THE
THE BUS) (FAULT TIME: A-PHASE VOLTAGE CROSS ZERO; FAULT FAULT INCEPTION ANGLE OF A PHASE, IS THE FAULT RESISTANCE, AND
TRANSITION RESISTANCES: 0
) d IS THE FAULT LOCATION)

sound phase in general, which is in accordance with the afore-

mentioned theory and the proposed algorithm for fault classifi-
cation. The original values and outputs of the fault classification
TABLE VIII
are shown in Table III correspondingly. It can be seen from the COMPARISON WITH OTHER FAULT CLASSIFICATION METHODS (t IS
results that the WSE is feasible in the fault classification, and THE FAULT ANGLE OF A PHASE AND IS THE FAULT RESISTANCE; D IS
THE FAULT LOCATION)
various types of faults can be classified effectively based on the
proposed algorithm.
2) Fault Classification for Various Fault Inception Angles:
In this part, 73 cases with various fault inception angles ranging
from 0 to with a step of are investigated and all of the
results prove that the faults can be classified correctly by means
of the WSE-based algorithm.
The corresponding results of fault classification are shown in
Table IV and it is proven that the WSE-based fault classification
algorithm is effective even in the case of various fault inception 6) Comparison With Other Classification Methods: In this
angles. This is due to the advantages of WSE and the compar- part, further tests are conducted to perform the comparisons be-
ative ratios of the WSE between faulty phase and sound phase. tween WSE-based and other classification methods, such as the
In this case, the classification results will not be affected by the wavelet energy (WE) [10], [11] and WMM [15], [16]. The tests
transient magnitude. under different conditions show that the worst case for fault clas-
3) Fault Classification for Various Fault Resistances: In this sification is the low-transient condition. Some of the worst-case
part, 61 cases with various fault resistance ranging from 0 classification results are shown in Table VIII, where the fault
to 300 with a step of 5 are tested and all of the results are inception angle is 0. As seen in Table VIII, due to the very low
shown in Table V. The results also prove that the fault resistance transient magnitude, the accuracies of the classifications based
(i.e., fault transient magnitude) cannot affect the fault classifi- on WE and WMM are only 10% and 55%, respectively. How-
cation based on the WSE technique. ever, benefiting from its great antinoise property and low depen-
4) Fault Classification for Various Fault Locations: In this dence on transient absolute magnitude, the WSE-based classifi-
part, 61 cases with various fault locations ranging from 0 km to cation possesses the accuracy of 100%. This is the greatest supe-
300 km with a step of 5 km are tested and all of the sets of results riority and improvement of the proposed WSE-based method-
are shown in Table VI which proves that this fault classification ology compared to the other algorithms.
algorithm based on WSE is immune to fault locations.
5) Summary of Fault Classification Based on WSE: The sum- V. CONCLUSION
mary of all the fault classification tests is shown in Table VII. A new technique called WSE is proposed. The fault detection
It can be seen that the faults can be classified correctly, bene- based on WSE is studied, and a WSE-based fault classification
fiting from the proposed WSE-based algorithm, which proves algorithm is put forward as well. Conclusions supported by the
the reliability of the WSE-based algorithm and the validity of theory study and simulation tests are as follows.
the threshold selection in the application. 1) The WSE technique is proposed by combining WT with
This proposed fault classification algorithm is provided with SVD as well as the Shannon entropy. It provides a quanti-
high accuracy and is immune to different conditions, such as tative output which can act as an automatic feature extrac-
fault types, fault inception angles, fault resistance, and fault lo- tion technique in fault detection and classification. The test
cations, etc. Since the proposed algorithm only requires the first results prove that the WSE is sensitive to sudden changes in
half cycle of the postfault signal and its computing cost is similar transient signals, and is immune to noise. WSE can be used
to the methodology in [1], it has good real-time performance. to detect the faults under various situations even with the
HE et al.: FAULT DETECTION AND CLASSIFICATION IN EHV TRANSMISSION LINE 2163

low-magnitude transient where other methodologies en- [18] N. Zhang and M. Kezunovic, “Transmission line boundary protection
counter problems. using wavelet transform and neural network,” IEEE Trans. Power Del.,
vol. 22, no. 2, pp. 859–869, Apr. 2007.
2) WSE indicates the complexity of the analyzed transient [19] O. A. S. Youssef, “Combined fuzzy-logic wavelet-based fault classifi-
signal in the time–frequency domain and it can be used cation technique for power system relaying,” IEEE Trans. Power Del.,
to distinguish between the transients with different com- vol. 19, no. 2, pp. 582–589, Apr. 2004.
[20] V. C. Klema and A. J. Laub, “The singular value decomposition: Its
plexity intuitively and quantitatively. That is the reason computation and some applications,” IEEE Trans. Autom. Control, vol.
why WSE has been adopted to distinguish the faulty phase AC-25, no. 2, pp. 164–176, Apr. 1980.
from sound phase under different fault conditions. [21] Z. Y. He, X. Q. Chen, and G. M. Luo, “Wavelet entropy measure defini-
tion and its application for transmission line fault detection and identi-
3) The proposed algorithm for fault classification based on fication (part I: Definition and methodology),” in Proc. IEEE Int. Conf.
the WSE ratios between the faulty phase and sound phase Power System Technology, Oct. 2006, pp. 1–6.
is accurate and feasible. And it is an effective and robust [22] J. Cadzow, “SVD representation of unitarily invariant matrices,” IEEE
Trans. Acoust., Speech Signal Process., vol. 32, no. 3, pp. 512–516,
algorithm which is suitable in the case of various fault con- Jun. 1984.
ditions, so it has great potential for practical applications. [23] C. E. Shannon, “A mathematical theory of communication,” Bell Syst.
Tech. J., vol. 27, pp. 379–423, 623–656, Jul., Oct. 1948.
REFERENCES
[1] K. M. Silva, B. A. Souza, and N. S. D. Brito, “Fault detection and clas-
sification in transmission lines based on wavelet transform and ANN,”
IEEE Trans. Power Del., vol. 21, no. 4, pp. 2058–2063, Oct. 2006. Zhengyou He (M’10) was born in Sichuan Province,
[2] T. B. Littler and D. J. Morrow, “Wavelets for the analysis and compres- China, in 1970. He received the B.Sc. and M.Sc.
sion of power system disturbances,” IEEE Trans. Power Del., vol. 14, degrees from Chongqing University, Chongqing,
no. 2, pp. 358–364, Apr. 1999. China, in 1992 and 1995, respectively, and the Ph.D.
[3] A. M. Gaouda, M. M. A. Salama, M. R. Sultan, and A. Y. Chikhani, degree in electrical engineering from Southwest
“Power quality detection and classification using wavelet-multiresolu- Jiaotong University, Chengdu, in 2001.
tion signal decomposition,” IEEE Trans. Power Del., vol. 14, no. 4, pp. Since 2002, he has been a Professor with the Col-
1469–1476, Oct. 1999. lege of Electrical Engineering, Southwest Jiaotong
[4] X. Xu and M. Kezunovic, “Automated feature extraction from power University. His research interests are in the area of
system transients using wavelet transform,” in Proc. Int. Conf. Power signal processing and information theory and its ap-
System Technology, 2002, vol. 4, pp. 1994–1998. plication in electrical power systems, and the appli-
[5] C. L. Tu, W. L. Hwang, and J. Ho, “Analysis of singularities from cation of wavelet transforms in power system. Currently, he is the author or
modulus maxima of complex wavelets,” IEEE Trans. Inf. Theory, vol. coauthor of many journal papers and the owner of two invention patents.
51, no. 3, pp. 1049–1062, Mar. 2005.
[6] T. Le-Tien, H. Talhami, and D. T. Nguyen, “Simple algorithm for
wavelet maxima modulus extraction in time-scale representations,”
IEEE Electron. Lett., vol. 33, no. 5, pp. 370–371, Feb. 1997. Ling Fu is currently pursuing the Ph.D. degree in
[7] O. Chaari, M. Meunier, and F. Brouaye, “Wavelets: A new tool for the electrical engineering at Southwest Jiaotong Univer-
resonant grounded power distribution systems relaying,” IEEE Trans. sity, Chengdu, China.
Power Del., vol. 11, no. 3, pp. 1301–1308, Jul. 1996. Her research interests are in signal processing and
[8] O. A. S. Youssef, “New algorithm to phase selection based on wavelet information theory in electrical power systems.
transforms,” IEEE Trans. Power Del., vol. 17, no. 3, pp. 908–914, Jul.
2002.
[9] O. A. S. Youssef, “Fault classification based on wavelet transforms,” in
Proc. IEEE Transmission&Distribution Conf. Expo., Atlanta, GA, Oct.
28–Nov. 2, 2001, pp. 531–536.
[10] A. A. Hajjar, M. M. Mansour, and H. A. Talaat, “High-phase order
power transmission lines relaying approach based on the wavelet anal-
ysis of the fault generated traveling waves,” in Proc. 39th Int. Univ.
Power Eng. Conf., 2004, vol. 1, pp. 805–809.
[11] A. I. Megahed, A. M. Moussa, and A. E. Bayoumy, “Usage of wavelet Sheng Lin is currently pursuing the Ph.D. degree
transform in the protection of series-compensated transmission lines,” in electrical engineering at Southwest Jiaotong
IEEE Trans. Power Del., vol. 21, no. 3, pp. 1213–1221, Jul. 2006. University.
[12] O. A. S. Youssef, “Online applications of wavelet transforms to His research interests are in electrical power sys-
power system relaying,” IEEE Trans. Power Del., vol. 18, no. 4, pp. tems and automation.
1158–1165, Oct. 2003.
[13] S. G. Mallat, “A theory for multi-resolution signal decomposition: The
wavelet representation,” IEEE Trans. Pattern Anal. Mach. Intell., vol.
11, no. 7, pp. 674–693, Jul. 1989.
[14] D. Chanda, N. K. Kishore, and A. K. Sinha, “Application of wavelet
multiresolution analysis for classification of faults on transmission
lines,” in Proc. IEEE Conf. Convergent Technologies for Asia-Pacific
Region, vol. 4, pp. 1464–1469.
[15] J. D. Duan, B. H. Zhang, and H. X. Ha, “A novel approach to faulted-
phase selection using current traveling waves and wavelet analysis,” in Zhiqian Bo (SM’95) received the B.Sc. degree from
Proc. Int. Conf. Power System Technology, 2002, vol. 2, pp. 1146–1150. Northeastern University, Shenyang, China, in 1982
[16] Q. L. Su, X. Z. Dong, Z. Q. Bo, and F. Jiang, “New approach of fault and the Ph.D. degree from The Queen’s University
detection and fault phase selection based on initial current traveling of Belfast, Belfast, U.K., in 1988.
waves,” in Proc. IEEE Power Eng. Soc. Summer Meeting, 2002, vol. 1, From 1989 to 1997, he was with the Power Sys-
pp. 393–397. tems Group, University of Bath, Bath. Currently, he
[17] Z. Q. Bo, R. K. Aggarwal, A. T. Johns, H. Y. Li, Y. H. Song, R. L. King, is with AREVA T&D—Automation and Information
D. Novosel, and M. Kezunovic, “A new approach to phase selection Systems, Stafford, U.K., where he is responsible for
using fault generated high frequency noise and neural networks,” IEEE new technology developments. His main research in-
Trans. Power Del., vol. 12, no. 1, pp. 106–115, Jan. 1997. terests are power system protection and control.