Вы находитесь на странице: 1из 12

Electric Power Systems Research 125 (2015) 184–195

Contents lists available at ScienceDirect

Electric Power Systems Research


journal homepage: www.elsevier.com/locate/epsr

A new wavelet selection method for partial discharge denoising


Caio F.F.C. Cunha a,b , André T. Carvalho a,b , Mariane R. Petraglia a , Antonio C.S. Lima a,∗
a
Federal University of Rio de Janeiro, COPPE/UFRJ, Rio de Janeiro, RJ, Brazil
b
Electrical Energy Research Center, CEPEL/ELETROBRAS, Rio de Janeiro, RJ, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: Partial discharge signals are used to evaluate the insulation condition in several power devices. Their
Received 7 January 2015 measurements are often heavily contaminated by noise from different sources and thus the use of noise
Received in revised form 7 April 2015 reduction techniques is required. Wavelet shrinkage denoising methods are frequently employed and
Accepted 8 April 2015
several wavelet basis selection approaches have been discussed recently in the literature. This paper
Available online 25 May 2015
presents a new method for choosing the wavelet basis, comprising a novel approach for determining the
number of wavelet decomposition levels based on the energy spectral density of the signals, and a scale-
Keywords:
dependent algorithm for selecting the wavelet functions based on signal to noise ratios computed from
Partial discharge
Partial discharge measurement
the wavelet coefficients. At each decomposition stage, the predominant band of the partial discharge
Dielectric measurement pulse, called the signal band, is identified as the one that has the highest coefficient magnitude. The other
Wavelet transforms band is assumed to contain predominantly noise, and at each level the method selects the wavelet that
Signal Denoising maximizes the signal to noise ratio. The proposed approach is compared to the correlation based wavelet
Signal reconstruction selection (CBWS) and to the energy based wavelet selection (EBWS) methods for signals simulated and
obtained from high voltage equipment, showing better filtering performance for over 70% of the tested
signals and a smaller processing time.
© 2015 Elsevier B.V. All rights reserved.

1. Introduction proposed [2], denoising by wavelet shrinkage is the most frequently


employed approach to reduce wideband noise and discrete spec-
In power systems there is a high demand for diagnosing and tral interference [3–13]. For the extraction of pulse-shaped noise,
monitoring operating conditions of power equipment in an effi- additional denoising strategies have been successfully used after
cient and accurate way. In recent years this need has considerably the wavelet shrinkage, such as the ones based on back-propagation
grown, since any unexpected downtime results in very high costs. neural networks [11] and support vector machines [14]. To effec-
Furthermore, as power systems grow older, the number of aging tively suppress pulse-shaped noise, the wavelet shrinkage filtering
devices becomes substantial, leading to a need to increase their operation must minimize signal distortions to preserve as much as
lifetime, thereby reducing power outages and economic losses. possible the PD waveform [10]. In [15], linear discriminant analysis
The monitoring of partial discharges (PD) is an effective tool was applied to multi-scale fractal dimensions and energy parame-
for the evaluation of insulating system degradation in high volt- ters of ultra-high PD signals in order to recognize defects. In [16],
age equipment. A typical PD signal is often hampered by the spectral power analyses of PD pulses was employed to classify PD
presence of several noise sources. This contamination may over- sources and noise by means of a graphical representation. In [17],
whelm the PD signal and cause errors in the assessment of the modern methods for robust measurement and monitoring of high-
insulation conditions. In this scenario, the use of digital filtering voltage apparatus are described, while in [18] and [19] the designs
techniques is a critical step for the evaluation of PD signals. In par- of Hilbert fractal antenna and stacked Hilbert antenna array to
ticular, the wavelet transform (WT) has proved a powerful tool for increase the frequency bandwidth and sensitivity of the PD mea-
PD denoising, in view of its time-frequency localization proper- surement system are discussed.
ties [1]. Among several noise reduction algorithms that have been The discrete wavelet transform (DWT) decomposes discrete sig-
nals into a particular basis of the vector space of finite-energy
complex series l2 (Z). The wavelet basis of the DWT is determined
by the number of decomposition levels and by the wavelet filters
∗ Corresponding author. Tel.: +552139388599.
adopted in each level. The separation of PD pulses immersed in
E-mail addresses: caiocunha@cepel.br (C.F.F.C. Cunha), tomaz@cepel.br
(A.T. Carvalho), mariane@pads.ufrj.br (M.R. Petraglia), acsl@dee.ufrj.br
wideband noise in the WT domain is possible because the noise
(A.C.S. Lima). energy is uniformly distributed among the wavelet coefficients,

http://dx.doi.org/10.1016/j.epsr.2015.04.005
0378-7796/© 2015 Elsevier B.V. All rights reserved.
C.F.F.C. Cunha et al. / Electric Power Systems Research 125 (2015) 184–195 185

while the PD pulse energy is concentrated on a limited number


of coefficients with amplitudes above those of the correspond-
ing noise coefficients. This separation will be more efficient as the
wavelet basis better represents the PD pulses with the smallest
number of high amplitude coefficients [5]. Therefore, the problem
of finding the best wavelet basis for PD denoising can be formulated
as a search for the wavelet basis that maximizes the reference pulse
wavelet coefficients that are above the wideband noise level.
Despite the fundamental importance of determining an appro- Fig. 1. Bank structure for wavelet decomposition (analysis).
priate number of wavelet decomposition levels [6], very little
attention has been given to this issue in the literature. Most articles that of the input signal. These two signals are denoted as approxi-
have reported either the use of the maximum possible number or mation a1 and detail d 1 coefficients, respectively. In the next stage
the trial and error approach [4,5,7,9,10]. Regarding the selection of the coefficients a1 are applied to the low-pass and high-pass filters
the wavelet filters (associated to the wavelet function, also called followed by decimators, to generate the new approximation a2 and
wavelet mother), two important techniques have been proposed: detail d 2 coefficients. The decomposition process is repeated until
the correlation based wavelet selection (CBWS) method [9], which the final approximation aJ and detail d J coefficients are obtained.
searches the wavelet function that best correlates to the PD pulse The filter bank composed by the low-pass and high-pass filters,
and applies it in all decomposition levels, and the energy based which are also referred to as quadrature mirror filters (QMF), are
wavelet selection (EBWS) method [10], which, for each level, selects designed such that it is possible to obtain a perfect reconstruction
the wavelet filters that maximize the energy of the approximation of the original signal by applying the inverse digital wavelet trans-
coefficients located in the lower frequency subband. According to form (IDWT), which can be implemented by a synthesis filter bank
[10], EBWS provides better denoising results than CBWS. with factor of two interpolators (up-samplers) at each level.
In this work we propose a new approach for the selection of The denoising of a signal S by the wavelet shrinkage technique
the number of wavelet decomposition levels based on the energy can be summarized as follows:
spectral density (ESD) of a reference pulse, resulting in the min-
imum number of levels necessary to represent the pulse mainly
1. Determine the number of decomposition levels J and choose the
in the detail subbands. As the ESD is independent from the WT,
wavelet filters for each level j, where j = 1, 2, . . ., J; then perform
the number of decomposition levels can be calculated beforehand
the decomposition to obtain the DWT coefficients [21].
and provides essential information to the wavelet filter selection
2. Obtain the thresholded coefficient values, considering hard or
method. We also propose a new scale-dependent wavelet selec-
soft approaches.
tion method that identifies, at each level, the predominant band
3. Reconstruct the signal by applying the IDWT to the thresholded
of the PD signal and selects the wavelet filters that maximize the
coefficients to obtain the denoised signal.
signal to noise ratio (SNR) between the complementary decompo-
sition bands considering their peak coefficient values. The proposed
SNR based wavelet selection (SNRBWS) method tends to concen- The selection of the thresholding function and its parame-
trate the information of the PD signal on fewer coefficients with ters determines which wavelet coefficients will be associated to
higher values, producing better filtering results and reducing signal noise and hence eliminated. The thresholding procedure, known
distortions when compared to other techniques. as wavelet shrinkage, can be divided into three steps:
The remaining of this paper is organized as follows. Section 2
presents an overview of the wavelet-shrinkage denoising method, 1. Calculate the threshold value, which is frequently performed in
describing the most employed thresholding approaches. Section 3 one of the following four different rules [7]: rigrsure, sqtwolog,
describes the proposed method for the selection of the number of heursure and minimax.
wavelet decomposition levels, hereafter denoted as NWDLS. Sec- 2. Apply a multiplicative threshold rescaling method [7], according
tion 4 introduces the proposed wavelet basis selection technique, to one of the following procedures:
SNRBWS, as well as the previously proposed CBWS and EBWS a. sln: estimates the noise level based on the first detail
methods. Section 5 shows the partial discharge signals used in the coefficients d1 only, since in most cases they are mainly
denoising evaluations. The outputs are discussed in Section 6 and formed by noise, with a standard deviation that can be esti-
concluding remarks are given in Section 7. mated as
MAD|d1 |
= (1)
q
2. Wavelet-based denoising technique for PD detection
where MAD|d1 | is the median absolute deviation of the detail
The decomposition process of a measured PD signal S through a coefficients, and q may vary between 0.4 and 1, with typical
wavelet transform allows us to separate in different coefficients value equal to 0.6745. The combination of the threshold selec-
most of the existing noise and the signal of interest, making it tion rule sqtwolog with the multiplicative threshold rescaling
possible to mitigate or even eliminate the noise by the use of a sln corresponds to the well-known global thresholding rule
threshold shrinkage function. The dyadic digital wavelet transform [22] given by

MAD|d1 | 
is the most frequently employed transform in data denoising tech-
niques [5,6,20] due to its desirable time-frequency characteristics thr = 2 log(N) (2)
[1]. The practical implementation of a dyadic digital wavelet trans- 0.6745
form as an analysis filter bank is depicted in Fig. 1. In the first stage, b. mln: similar to sln, but the estimation of the noise variance
the PD signal S passes through a two-channel filter bank composed based on the coefficients of each level. The jth level noise
of a low-pass filter G(n), corresponding to the scale function, and variance is estimated by
a high-pass filter H(n), corresponding to the wavelet function, fol-
lowed by factor-of-two decimators (down-samplers). The outputs MAD|dj |
j = (3)
of the first stage are two signals whose lengths are equal to half of q
186 C.F.F.C. Cunha et al. / Electric Power Systems Research 125 (2015) 184–195

The use of the sqtwolog rule with multiplicative threshold the signal, and the vertical dashed line indicates the lowest fre-
rescaling given by quency component whose energy is greater than that indicated by
the horizontal line, which in this case is Fmin = 5.4932 MHz. Substi-
MAD|dj | 
thr j = 2 log(N) (4) tuting these values in (7) we found J = 8 decomposition levels, and
0.6745 by this procedure we ensured at least 85% of the pulse energy to be
is frequently employed since it tends to eliminate a smaller represented in detail subbands.
amount of the PD signal energy when compared to (2), and Besides providing a good wavelet decomposition through the
because the wavelet coefficients patterns of the PD pulse and selection of an appropriate number of decomposition levels, this
of the noise are level dependent [5]. method avoids the need for decomposing the signal in its maximum
3. Apply a threshold to the wavelet coefficients, which will deter- number of levels, which would be Jmax = 13 in the above example. As
mine how such coefficients will be attenuated or zeroed in order a result, large process time savings is obtained in both the wavelet
to eliminate or reduce the noise distributed among them. The selection and the denoising process.
most frequently used threshold functions are the hard and soft
functions [23,24]. The hard threshold function, given by 4. Wavelet filters selection

t, |t| ≥ thr In this section we briefly describe the previously proposed
fH (t, thr) = (5)
0, otherwise CBWS and EBWS methods [9,10]. Subsequently, we present a new
approach for selecting the mother wavelet of each stage of the PD
is commonly applied in denoising operations of PD signals,
signal decomposition, named SNRBWS, which has shown itself an
because it preserves the signal magnitude characteristics, which
efficient alternative to the CBWS and EBWS methods.
provides a better SNR [5,9]. Therefore, the hard function was
used in this work to compare the performance of the different
4.1. CBWS and EBWS methods
algorithms, rather than the soft function employed in [10].
In the CBWS procedure [9,10], the wavelet selection is made
3. Number of wavelet decomposition levels selection with the purpose of maximizing the correlation  between the
(NWDLS) ideal PD signal and the wavelet function chosen from a library of
wavelets. The selected wavelet is not scale-dependent, and hence
In the literature, a few procedures have been adopted for the the same mother wavelet is employed in all decomposition levels.
automatic selection of the number of wavelet decomposition levels The correlation is computed as
employed in the denoising of PD signals. In most cases [4,5,7,9,10], n−1
however, this selection is made at random or by verifying that, for (xi − x̄)(yi − ȳ)
=  i=0
n−1
(8)
a particular PD signal, a certain number of levels J is enough to n−1 2 2
(xi − x̄) (yi − ȳ)
provide good denoising results. Although pointed out in [6] that i=0 i=0
the number of levels must be related to the sampling rate and spec- where xi and yi represent the ith samples of the PD signal of inter-
trum of the PD pulse, its maximum possible value was employed. est and of the wavelet function, respectively, and the variables x
Considering that the signal length at the highest level of decompo- and y are their average values. The main difficulty associated to
sition must be at least equal to the length of the wavelet filter, the the CBWS method, however, is the determination of the appropri-
maximum number of decomposition levels is given by ate scale for the wavelet function in which the correlation with
  N
 the PD pulse should be calculated. In [10] the authors proposed a
J = fix log2 (6) heuristic determination of this scale, rescaling in time and align-
NW − 1
ing the wavelet function in order to coincide with the PD pulse in
where N is the signal length and NW is the length of the decom- their maximum and in their first point of zero crossing after their
position filter associated to the mother wavelet. However, such maximum. This approach was also adopted in this work.
method cannot be applied to a scale-dependent wavelet selection An alternative procedure to select scale-depended wavelet fil-
algorithm, which employs different wavelet filters in each level, ters is to apply the EBWS method presented in [10], which performs
since a priori knowledge of all wavelet functions would be nec- the search by an optimal wavelet at each level from a library of
essary to determine the number of decomposition levels. In this wavelet filters, by the maximization of the energy percentage in the
case, a wavelet independent maximum value could be calculated approximation aj , which leads to a reduction of the energy percent-
as Jmax = fix(log 2 N) [6]. age in d j . The energy percentage of the approximation coefficients
In this work we propose that the number of decomposition lev- Eaj is calculated by
els be selected based on the value of the lowest representative

frequency component present in the PD signal, obtained from its
k
(aj,k )2
energy spectral density (ESD), as follows: Eaj =  2
j  2
(9)
(aj,k ) + (di,k )
  F  k i=1 k
S
J = fix log2 (7) where j is the wavelet coefficient index. According to the results
Fmin
presented in [10], the EBWS method outperformed CBWS in both
where Fmin is determined by the lowest frequency component, noise filtering and runtime.
whose energy is greater than a certain percentage of the total sig-
nal energy. Such percentage is used with the purpose of discarding 4.2. SNRBWS method
the frequency components that do not contain significant energy.
If J exceeds the maximum allowable number of levels, then J = Jmax The proposed SNRBWS method stemmed from the fact that
must be adopted. the EBWS approach, always seeking to maximize the energy of
Consider, for example, the PD signal of length N =10,000 and the approximation coefficients, not necessarily concentrates the
sampling frequency FS =2500 MHz shown in Fig. 2a, and its energy energy of the PD signals in as few levels as possible, which is the
spectral density in Fig. 2b. The horizontal dashed line in Fig. 2b main goal of the wavelet filters selection for PD denoising [9]. For
indicates the level where the energy is 15% of the total energy of example, consider again the ESD of the PD pulse in Fig. 2. Until the
C.F.F.C. Cunha et al. / Electric Power Systems Research 125 (2015) 184–195 187

Fig. 2. Illustration of the NWDLS method: (a) PD signal, and (b) its energy spectral density (ESD). The vertical dashed line indicates the lowest frequency component whose
energy is larger than 15% of the total pulse energy. With an 8-level decomposition, at least 85% of the pulse energy is in the detail subbands.

6th level, where the energy of the pulse was still concentrated on Various wavelet filters suitable for signal decomposition are
the approximation bands, the EBWS method chose the best filters available in the literature [1,21]. Among the ones commonly
in order to maximize the energy corresponding to the pulse. How- applied to PD signals are the wavelets of Daubechies (db), Coiflets
ever, at level 7 the EBWS still tried to maximize the energy on the (coif) and Symlets (sym) families [9,10], because of properties such
approximation band, while the energy of the PD pulse was concen- as compactness, orthogonality, speed and symmetry, that are desir-
trated on the detail coefficients. Consequently, the filters selected able in the analysis of fast transient and irregular pulses as PD
from this level on were not the best ones to concentrate the wavelet signals [25,26]. The wavelet library employed in this work com-
coefficients energy, but rather to spread it. prised Daubechies wavelets of orders 1–40 (db1–db40), Symlets
To circumvent this difficulty, the proposed method identifies, at wavelets of orders 2–15 (sym2–sym15), and Coiflets wavelets of
each decomposition level, the predominant band of the PD pulse orders 1–5 (coif1–coif5).
(called the signal band) as the one that has the maximum absolute In summary, the selection process of the wavelet by the SNRBWS
coefficient value. The complementary band, which is assumed to algorithm is composed of the following steps:
contain less relevant information, is then identified as the noise
band. The wavelet filters for each level are then chosen in order 1. Create a library of wavelet function t,i , for each type of wavelet
to maximize the peak signal to noise ratio (SNR) between the two family (type t = 1, 2, ..., p) and selected wavelet orders (order i = 1,
complementary decomposition subbands. 2, ..., N); initialize j = 1;
Assuming that the band with the highest energy concentration 2. Perform a single-level decomposition of the PD signal S with each
corresponds to the band with the highest absolute coefficient value, wavelet of the library, generating the approximation aj and the
the SNRBWS method maximizes, at each level, the energy differ- detail d j coefficients;
ence between the adjacent approximation and detail subbands. By 3. Compare the peak values of aj and d j , assigning the coefficients
finding the wavelet filters that maximize the energy in those bands with the highest peak to signal and the others to noise, and
where supposedly the energy of the analyzed PD pulse is larger, calculate the SNR(t,i) defined by (10);
this method concentrates the pulse energy in a smaller number of 4. Find the maximum SNR(t,i) , for t = 1, 2, . . ., p and i = 1, 2, . . ., N, and
coefficients with larger amplitudes. Similar objective was applied save the corresponding wavelet T,I as the best for that decom-
in [9]. position level; thus, T and I correspond to the best values of t and
The proposed method is detailed next. At the jth stage, the i;
signal decomposition is performed for each filter pair of a can- 5. Repeat steps 2–4, with j = j + 1, until the maximum number of lev-
didate wavelet, generating the approximation aj and the detail els J is reached, taking the approximation coefficients obtained
coefficients d j . The signal coefficients are identified as aj or with the best wavelet T,I at level j − 1 as the signal S to the jth
d j , whichever has the highest magnitude. The remaining set of level.
coefficients is identified as noise. Summarizing:

if max(|dj |) ≥ max(|aj |) then signal = d j and noise = aj The flowchart describing the whole procedure of SNRBWS algo-
rithm is shown in Fig. 3, where the number of decomposition levels
else signal = aj and noise = d j J is obtained by the method NDWLS.
Since the wavelet filters selected by SNRBWS tend to concen-
The SNR is then calculated by
trate components with higher amplitudes in the frequency bands
SNR = max(|signal|)/ max(|noise|) (10) that contain most of the desired PD signal information, there is
a tendency that the reconstructed PD signals present less distor-
The wavelet filters with the highest SNR are then selected for this tion with better denoising results when compared to the other
stage. methods, particularly in low SNR conditions. A large number of
188 C.F.F.C. Cunha et al. / Electric Power Systems Research 125 (2015) 184–195

Fig. 3. Flowchart of SNRBWS.

experimental results confirming the above statement are presented originated, the calibration pulse will be shaped by the trans-
in Section 6. fer function of the measurement circuit, as would the real PD
As the calculation of the maxima of the complementary bands is pulse.
computationally less expensive than the calculation of their energy,
the SNRBWS method usually presents lower runtimes than does the
EBWS. In [27], we employed the SNR metric in dB. In the present 5. Partial discharge signals
work, however, we adopted Eq. (10) that, being less expensive than
the log equation, made the method even faster, while producing the In order to evaluate the performance of PD signals denoising,
same results. Simulation results presented in Section 6 confirm that experiments were carried out with both simulated and measured
the SNRBWS is faster than the other methods. PD signals.
A practical limitation of the three methods discussed above for
the wavelet basis selection may be the previous determination of a
typical PD pulse waveform, which can be difficult in very low SNR
conditions [11,14,16]. Usually, any PD pulse with high amplitude Table 1
Parameters for S3 –S5 .
can be used whenever available. When the noise level is too high,
a reference PD pulse waveform can be calculated by the average of Pulse  (␮s) f0 (kHz)
a set of normalized noisy pulses. When no PD pulse waveform is S3 2 1000
available, a calibration pulse can be adopted as a reference. Even S4 2.5 250
without carrying any information about how the PD signal was S5 2 500
C.F.F.C. Cunha et al. / Electric Power Systems Research 125 (2015) 184–195 189

Table 2
Parameters for S6 and S7 .

Pulse A1 A2 A3 A4 f1 (MHz) f2 (MHz) f3 (MHz) f4 (MHz)

S6 1.06 0.70 −1.06 −0.01 30 20 15 10


S7 −1.09 −1.00 −0.41 0.01 60 40 30 20

Table 3 Table 4
Sampling rates of simulated signals. Measured PD signals and sampling frequencies.

Signal S1 S2 S3 S4 S5 S6 S7 Designation Equipment Fs (MHz)

Sampling rate (MHz) 60 60 10 10 10 1000 1000 S8 Stator bar 250


S9 Circuit breaker 500
S10 Generator 500
5.1. Simulated PD Signals S11 GIS 1000
S12 Surge arrester 2500
S13 Transformer 2500
Seven signals proposed in the literature were employed. The S14 Current transformer 1000
signals S1 and S2 , adopted in [9], are given by

S1 = A(e−˛1 t − e−˛2 t ) (11)


−˛1 t −˛2 t
S2 = A(e cos(ωt − )) − e cos()) (12)

where A = 1.5, ˛1 = 1 ×106 s−1 , ˛2 = 1 ×107 s−1 , ω = 2 × 106 rad/s.


The signals S3 , S4 and S5 , proposed in [20], are defined as

⎪ t − t0
⎨ −
A(e  cos(2f0 (t − t0 ))), t ≥ t0
Sn = (13)

⎩ 0, t < t0

where A = 1.0 and t0 = 1 ␮s. The values of  and f0 are shown in Fig. 5. Impedance used for the transformer, generator, stator bar and breaker mea-
Table 1 for the signals S3 , S4 and S5 . Finally, signals S6 and S7 , surements.
proposed in [8], are

⎪ A sin(2f1 t), t1 ≤t < t2
⎨ 1 components that may come from the source. This filter can only be
Sn = e−ıt A2 sin(2f2 t), t2 ≤t < t3 (14) used offline or in laboratory measurements, since in on-line mea-


A3 sin(2f3 t) + A4 sin(2f4 t), t3 ≤t < t4 surements it would be impractical to insert a filter in the supply
branch that comes to equipment, both for security reasons as not
where ı = 7 ×106 , t1 = 200 ns, t2 = 300 ns, t3 = 500 ns, and t4 = 1000 ns. to interrupt the operation of the equipment under analysis. In this
The remaining parameters are presented in Table 2. The sampling work, an inductor of 3 H was employed to evaluate a generator sta-
rates of all simulated pulses are given in Table 3. tor bar and a circuit breaker. For the generator, GIS, surge arrester,
transformer and current transformer, the measurements were per-
5.2. Measured PD Signals formed on site, and the arrangement was that of Fig. 4, except for
the filter. The coupling device Z varied according to the equipment.
The PD measuring circuit presented in [25] and shown in Fig. 4 The list of analyzed equipment is shown in Table 4, along with the
was employed in our laboratory experiments. In this configuration, corresponding sampling rates.
the object under test, Ca , is placed in parallel with a coupling capac- In the transformer, generator, stator bar and circuit breaker
itor Ck in series with a coupling device Z, from which the PD is measurements, the impedance Zm , shown in Fig. 5, was employed,
detected and passed through a coaxial cable to the data acquisition where Cm , Rm and Lm are the measuring capacitance, resistance
system. A synchronization signal is obtained from the capacitive and inductance, respectively, which are accompanied by a gas cur-
voltage divider composed by capacitors C1 and C2 , so that the sys- rent suppressor at both circuit input and output in order to protect
tem can identify the main phase where each PD pulse occurs. The the measuring system. The impedance Zm corresponds to a high-
measurement circuit is fed by a high voltage source in series with pass filter and its component values are selected to provide a
an inductor that acts as a filter to reduce the high frequency noise cutoff frequency of about 1.2 kHz, which is sufficient to remove the

Fig. 4. PD measurement circuit used in laboratory.


190 C.F.F.C. Cunha et al. / Electric Power Systems Research 125 (2015) 184–195

Fig. 6. Photographs of evaluated generator, GIS and current transformer.

Table 5 6. Results
Selected wavelets for obtaining the reference signals.

Signal S8 S9 S10 S11 S12 S13 S14 The results of the SNRBWS method applied to the various PD
Levels 13 7 8 13 10 6 8 reference signals described in Section 5 are next compared to those
Wavelet coif5 sym11 coif2 db4 db10 sym12 coif3 obtained with the EBWS and CBWS methods.

6.1. Wavelet basis selection

The total number of decomposition levels for each reference sig-


60 Hz component and its possible harmonics.1 . For the gas insu- nal was obtained from (7), using the NWDLS method described in
lated switchgear (GIS) measurements, a capacitive strap was used Section 3. The frequency Fmin was calculated in order to concentrate
as a coupling capacitor to collect the PD signal, without external 90% of the energy of the pulses in the details subbands. Fig. 8 shows
impedance. On the evaluation of transformers, the coupling capac- the normalized ESD of each pulse and the decomposition bands
itor Ck is situated inside the bushing, and hence the impedance is generated with the calculated number of decomposition levels.
connected directly to its tap output [28]. For the surge arrester and Table 6 shows for each signal the optimal wavelets obtained
the current transformer, which have higher capacitances, a high with the CBWS, EBWS and SNRBWS methods in all decomposition
frequency current transformer connected to the ground cable of levels. Note that for the EBWS and SNRBWS algorithms a different
the object under test was employed as the sensor, dismissing the wavelet is obtained for each level, whereas for the CBWS algorithm
use of Ck . Photographs of the measured generator, GIS and cur- a single wavelet is selected for all levels.
rent transformer are shown in Fig. 6. The measurement system As originally stated in [9], the objective of the wavelet basis
bandwidth is determined by the characteristics of the measure- selection is to concentrate the energy of the pulses in as few levels as
ment equipment and cables [28]. Although the data were obtained possible in order to generate better denoising results. Fig. 9 presents
at different sampling rates, each signal was analyzed considering the energy distributions by levels produced by each method for all
10,000 samples. signals. Note that the SNRBWS approach yielded the largest energy
In order to obtain filtered reference pulses from the measured concentration for all the measured pulses, and for all the simu-
signals, we applied a procedure similar to the one employed in lated pulses except for S1 and S4 . For signals S1 and S4 the methods
[10]. The signals were filtered through the wavelet shrinkage algo- EBWS and CBWS, respectively, resulted in better energy concen-
rithm with the hard thresholding function, with threshold values tration. This fact confirms that no method can always produce the
obtained by the sqtwolog method and mln rescaling approach. The best energy concentration, as the result depends on the waveform
number of decomposition levels employed in this process was and on the initial time of each pulse.
selected by the NWDLS method described in Section 3, and the
wavelets were selected by the CBWS method applied directly to 6.2. Runtime comparison
the measured signals. Table 5 presents both parameters used to
denoise each of the acquired signals. The reference pulses selected Considering all decomposition levels and all signals, a total of
from the denoised signals, as well as the simulated pulses described 111 wavelet selections were performed by the SNRBWS and EBWS
in Section 5.1, are shown in Fig. 7. approaches, while only 14 wavelet selections were performed by
the CBWS technique. The total processing time required for each
method is shown in Table 7. The simulations were carried out on
a processor Intel Core2 Duo 2.4 GHz personal computer. Although
1
In our experiments, the following values were adopted: Cm = 10 ␮F, Lm = 820 ␮H the number of wavelet selections was considerably smaller in the
and Rm = 1 k . CBWS method, it should be noted that considering together all
C.F.F.C. Cunha et al. / Electric Power Systems Research 125 (2015) 184–195 191

Fig. 7. Reference pulses: S1 –S7 (simulated) and S8 –S14 (filtered from measured signals).

analyzed signals, both EBWS and SNRBWS approaches spent less decomposition levels, its difference to the CBWS processing time
than half of the time required by the CBWS, confirming the result was not as large as found in [10].
pointed in [10] that the CBWS method performs much slower The relative execution times of each method are shown in
than the multilevel techniques. However, as the processing time Fig. 10. It should be noted that for signal S11 the CBWS method
of EBWS (and also of SNRBWS) is proportional to the number of required less time than the other two, indicating that the processing

Fig. 8. Normalized ESD and decomposition subbands for signals S1 –S14 with the NWDLS method for 90% of the pulses energy decomposed in the detail subbands (see Fig. 2).
192 C.F.F.C. Cunha et al. / Electric Power Systems Research 125 (2015) 184–195

Table 6
Wavelets selected by: (A) CBWS, (B) EBWS and (C) SNRBWS.

PD Decomposition level

1 2 3 4 5 6 7 8 9 10 11 12 13

S1 A db3 db3 db3 db3 db3 db3 db3 db3 db3 db3 db3 db3 db3
B coif1 db2 db2 db5 db5 db27 db5 db3 db6 sym7 db3 db4 db4
C coif2 db1 db1 db1 db4 db1 db3 db1 db2 db1 db1 db4 sym4
S2 A sym6 sym6 sym6 sym6 sym6 sym6
B coif2 db9 sym4 db12 db2 db34
C coif2 db39 sym7 db11 db2 sym4
S3 A db3 db3 db3 db3
B sym5 sym10 db1 db2
C sym5 db40 coif2 db15
S4 A sym4 sym4 sym4 sym4 sym4 sym4 sym4 sym4 sym4 sym4 sym4 sym4 sym4
B db1 sym5 sym4 db2 db4 coif1 coif1 coif1 coif1 coif1 db2 coif1 coif1
C db1 db1 db8 sym7 sym5 db1 db1 db1 coif1 db1 db1 sym11 db1
S5 A db3 db3 db3 db3 db3 db3
B sym5 sym4 sym6 db3 db2 db2
C db1 sym13 coif2 db36 coif1 db2
S6 A db10 db10 db10 db10 db10 db10 db10 db10 db10
B sym5 coif2 coif2 db9 sym5 sym12 db35 db1 db3
C db37 db8 sym5 db39 sym11 db21 db28 db40 db2
S7 A sym13 sym13 sym13 sym13 sym13 sym13 sym13 sym13
B sym12 coif2 coif2 coif3 db9 db34 db1 db3
C coif1 db4 db3 sym7 db5 db37 db28 db6
S8 A coif3 coif3 coif3 coif3 coif3 coif3 coif3
B db2 db9 sym12 sym11 sym12 db4 db1
C coif5 db3 db5 sym12 db12 sym8 db1
S9 A sym11 sym11 sym11 sym11 sym11 sym11 sym11
B sym5 db7 db16 db15 db37 db37 db1
C sym4 db5 sym13 db12 db28 db22 db19
S1 0 A coif2 coif2 coif2 coif2 coif2 coif2 coif2 coif2
B coif2 coif2 coif2 coif2 coif3 coif2 db1 db1
C sym4 coif2 sym7 db5 sym6 db9 db13 db19
S1 1 A db3 db3 db3 db3 db3 db3 db3 db3 db3
B sym5 db6 db4 sym5 db3 db7 db39 db1 db5
C sym4 sym11 sym6 db6 coif3 db3 coif1 db4 db1
S1 2 A db10 db10 db10 db10 db10 db10 db10 db10
B sym5 db15 sym9 db40 db3 db37 coif1 db4
C db38 db22 sym12 db37 db9 db35 db39 db21
S1 3 A sym13 sym13 sym13 sym13 sym13 sym13
B db11 db13 db11 db12 db33 db1
C db33 db18 sym10 sym10 db32 sym12
S1 4 A coif4 coif4 coif4 coif4 coif4 coif4 coif4
B coif3 db36 db34 db36 db1 db1 db1
C coif3 db35 db35 db2 db39 sym13 db1

Table 7 The shrinkage wavelet filtering was performed with the hard
Total runtime required by each method to select the wavelet basis for all signals.
thresholding function, using the sqtwolog method with the mln scal-
Method CBWS EBWS SNRBWS ing to find the threshold values. To reconstruct the signals, only
Total runtime (s) 525.12 205.69 203.06 the decomposition bands in which the percentage energy of the
reference pulse was greater than 1% were used. This procedure
presented better denoising results than the simple sqtwolog-mln
approach, and was inspired by the visual band selection implemen-
time also depends on the pulse waveform, and that there is no guar- tation previously published in [29].
antee that any method always performs faster than the other ones. To evaluate the distortion introduced in the denoised signals, we
Table 7 shows that CBWS spent 155.29% more time than EBWS, and compared the resulting pulse waveforms with the reference ones
that EBWS spent 1.30% more time than SNRBWS. Although the exe- before the addition of noise. The parameters employed to quantify
cution times of SNRBWS and EBWS are similar, SNRBWS executed such distortion were the correlation coefficient (CC) and the mean
faster in 11 out of the 14 analyzed signals, as can be seen in Fig. 10. squared error (MSE), defined as

1 2
6.3. Denoising results N
MSE = x1 (i) − x2 (i) (15)
N
Noise is usually defined as an unwanted component of a mea- i=1
sured signal that has no relationship with the signal of interest
[2]. The most common sources of noise in PD measurements are N
x (i) x2 (i)
i=1 1
described in [6]. Wavelet shrinkage denoising results are affected CC =  N
(16)
not only by the waveform of the pulse and its relative initial time, N 2 2
x (i)
i=1 1
x (i)
i=1 1
but also by the stochastic noise sample added to the pulse. To pro-
duce more general results, for each SNR, 50 independent noise
signals were added to each PD pulse, and the average results of The best filtering results correspond to the largest CC and/or
the 50 simulations are presented. lowest MSE.
C.F.F.C. Cunha et al. / Electric Power Systems Research 125 (2015) 184–195 193

Fig. 9. Percent energy distributions among levels obtained by CBWS, EBWS and SNRBWS for signals S1 –S14 .

Fig. 10. Relative time of wavelet basis selection methods for signals S1 –S14 .

6.3.1. Wideband stochastic additive noise


To evaluate the filtering performance of each basis selection
method for each reference signal, 50 independent Gaussian white
noise signals were generated with five different SNR values (0.25,
0.5, 1, 2 and 4), and each of such signals was added to each of the
14 PD pulses, performing a total of 10,500 filtering simulations. We Fig. 11. Mean value of the correlation coefficient for each method and signal with
present next the mean results obtained for each signal and SNR additive Gaussian white noise: (a) SNR = 0.25, (b) SNR = 0.5, (c) SNR = 1, (d) SNR = 2,
(e) SNR = 4.
value.
Fig. 11 shows the mean values of CC. It should be noted that,
Table 8
especially in low SNR conditions (Fig. 11a and b), the SNRWBS
Number of signals, from S1 to S14 , for which the SNRBWS produced the best results
method produced higher CC for most of the signals. Fig. 12 presents for each SNR.
the resulting MSE mean values for all signals and SNR conditions.
SNR Best CC value Best MSE value
For better visual comparison, we show the logarithm of the mean
MSE. Again it can be seen that the basis produced by SNRBWS pro- 0.25 10 12
duced lower mean MSE values for all SNR values, particularly in low 0.5 11 11
1 11 11
SNR conditions (Fig. 12a and b). Table 8 displays, for each value of
2 12 12
SNR, the number of signals for which the basis selected by the pro- 4 10 10
posed method produced the best denoising results when compared
to CBWS and EBWS. Considering the universe of 10,500 filtering
simulations, Table 9 shows the percentage of best filtering results Fig. 13 illustrates the filtering results for signal S11 in pres-
obtained by each of the wavelet selection methods. Regarding both ence of additive Gaussian white noise with SNR = 0.25. Note that
CC and MSE parameters, the SNRBWS method presented the best the wavelet shrinkage filtering removes most of the wideband
performance in over 70% of the cases. noise using any of the three proposed basis selection methods.
194 C.F.F.C. Cunha et al. / Electric Power Systems Research 125 (2015) 184–195

Fig. 12. Logarithm of the MSE mean value for each method and signal with additive Fig. 14. (a) Mean value of the correlation coefficient for each method and signal
Gaussian white noise: (a) SNR = 0.25, (b) SNR = 0.5, (c) SNR = 1, (d) SNR = 2, (e) SNR = 4. with additive DSI noise of SNR = 1; (b) logarithm of the MSE mean value for each
method and signal with additive DSI noise of SNR = 1.
Table 9
Percentage of best results obtained by each method in the universe of 10,500
simulations.

Method CBWS EBWS SNRBWS

Best CC (%) 4 25 71
Best MSE (%) 8 19 73

However, this example clearly shows that the filtering operation


using the SNRBWS basis yields better preservation of the wave-
form and amplitude of the original pulse than does the CBWS. The
EBWS method, although performing better than CBWS in preser-
ving the pulse shape, kept some noise power where no signal was
expected.

6.3.2. Discrete spectral interference


Although the wavelet shrinkage processing is theoretically more
appropriate for wideband noise removal from pulsing signals, its
performance was here evaluated in the presence of discrete spectral

Fig. 15. Filtering results (in blue) for signal S8 in presence of DSI noise with SNR = 1
and Gaussian white noise with SNR = 1. The original reference signal is shown (in red)
for visual comparison: (a) noisy signal; (b) filtered signal using the CBWS basis; (c)
filtered signal using the EBWS basis; and (d) filtered signal using the SNRBWS basis.
(For interpretation of the references to color in this legend, the reader is referred to
the web version of the article.)

interference (DSI). As suggested in [6], the AM noise was generated


by

nAM = (c + m sin(2fm t)) sin(2fi t) (17)


i=1

with c = 1, m = 0.4, fm = 1 kHz and fi varying from 600 kHz to


1400 kHz in steps of 200 kHz. We added to the reference signals
the AM noise defined by (17) with SNR = 1 and a stochastic Gaussian
white noise with SNR = 1, and performed for each signal 50 filtering
simulations. The obtained CC and MSE mean values are shown in
Fig. 14. The SNRBWS method again outperformed CBWS and EBWS,
Fig. 13. Filtering results (in blue) for signal S11 in presence of additive Gaussian yielding higher CC in 10 out of 14 signals, and lower MSE in 11 out
white noise with SNR = 0.25. The original reference signal is shown (in red) in all of 14 signals.
graphs for visual comparison: (a) noisy signal; (b) filtered signal using the CBWS
Fig. 15 shows the filtering results of signal S8 in presence of addi-
basis; (c) filtered signal using the EBWS basis; (d) filtered signal using the SNRBWS
basis. (For interpretation of the references to color in this legend, the reader is tive AM noise with SNR = 1 and Gaussian white noise with SNR = 1.
referred to the web version of the article.) Again the wavelet shrinkage filtering process was able to remove
C.F.F.C. Cunha et al. / Electric Power Systems Research 125 (2015) 184–195 195

most of the wideband noise in all cases, but only the filtering with [5] X. Ma, C. Zhou, I. Kemp, Interpretation of wavelet analysis and its application
the SNRBWS basis preserved the waveform and the amplitude of in partial discharge detection, IEEE Trans. Dielectr. Electr. Insul. 9 (3) (2002)
446–457.
the original pulse. In this case the EBWS approach was unable to [6] X. Zhou, C. Zhou, I. Kemp, An improved methodology for application of wavelet
filter all the DSI noise, and consequently presented worse results transform to partial discharge measurement de-noising, IEEE Trans. Dielectr.
than did the other methods. Electr. Insul. 12 (3) (2005) 586–594.
[7] H. Zhang, T. Blackburn, B. Phung, D. Sen, A novel wavelet transform technique
for on-line partial discharge measurements: Part 1. WT de-noising algorithm,
7. Conclusions IEEE Trans. Dielectr. Electr. Insul. 14 (1) (2007) 3–14.
[8] A. Gaouda, A. El-Hag, T. Abdel-Galil, M. Salama, R. Bartnikas, On-line detection
and measurement of partial discharge signals in a noisy environment, IEEE
The proposed methodology for obtaining the best wavelet basis Trans. Dielectr. Electr. Insul. 15 (4) (2008) 1162–1173.
for PD denoising is composed of two parts: a new approach for [9] X. Ma, C. Zhou, I. Kemp, Automated wavelet selection and thresholding for PD
determining the wavelet decomposition level and a new scale- detection, IEEE Electr. Insul. Mag. 18 (2) (2002) 37–45.
[10] J. Li, T. Jiang, S. Grzybowski, C. Cheng, Scale dependent wavelet selection for
dependent wavelet selection method based on a peak SNR measure
de-noising of partial discharge detection, IEEE Trans. Dielectr. Electr. Insul. 17
between approximation and detail coefficients at each decomposi- (6) (2010) 1705–1714.
tion level. [11] D. Evagorou, et al., Feature extraction of partial discharge signals using the
wavelet packet transform and classification with a probabilistic neural net-
The proposed NWDLS method determines the minimum num-
work, IET Sci. Meas. Technol. 4 (3) (2010) 177–192.
ber of decomposition levels necessary to ensure the decomposition [12] J. Li, C. Cheng, T. Jiang, S. Gryzbowski, Wavelet de-noising of partial discharge
of a given percentage of the energy of a PD pulse into detail sub- signals based on genetic adaptive threshold estimation, IEEE Trans. Dielectr.
bands. As the number of levels is calculated from the pulse ESD and Electr. Insul. 19 (2) (2012) 543–549.
[13] L. Hongxia, Z. Xuefeng, A method of second wavelet transform automated
the sampling rate, the method is independent from the wavelet threshold for partial discharge signal extraction, Proc. Int. Conf. Digit. Manuf.
functions adopted for the basis, and thus can be used with any other Autom. (2011) 42–45.
wavelet selection method. Avoiding the use of the maximum num- [14] H. Mota, L. Rocha, T. Salles, F. Vasconcelos, Partial discharge signal denoising
with spatially adaptive wavelet thresholding and support vector machines,
ber of decomposition levels, the proposed method saves runtime Electric Power Syst. Res. 81 (2) (2011) 644–659.
on the selection of the wavelet functions, as well as on the wavelet [15] J. Li, et al., Recognition of ultra high frequency partial discharge signals using
shrinkage denoising processing. This contribution may fill a gap in multi-scale features, IEEE Trans. Dielectr. Electr. Insul. 19 (4) (2012) 1412–1420.
[16] J.A. Ardila-Rey, et al., Partial discharge and noise separation by means of
determining wavelet basis for PD denoising, since the selection of spectral-power clustering techniques, IEEE Trans. Dielectr. Electr. Insul. 20
WDL has not received as much attention in the literature as has the (2013) 1436–1443.
selection of the wavelet functions. [17] A. Kraetge, et al., Robust measurement, monitoring and analysis of partial dis-
charges in transformers and other HV apparatus, IEEE Trans. Dielectr. Electr.
The proposed SNRBWS method finds, at each level, the wavelet
Insul. 20 (2013) 2043–2051.
function that maximizes the difference between the peak ampli- [18] J. Li, et al., UHF stacked Hilbert antenna array for partial discharge detection,
tudes of complementary subbands. Simulation results indicated IEEE Trans. Antennas Propag. 61 (11) (2013) 5798–5801.
[19] J. Li, et al., Hilbert fractal antenna for UHF detection of partial discharges in
that the SNRBWS method outperformed the CBWS and the EBWS
transformers, IEEE Trans. Dielectr. Electr. Insul. 20 (6) (2013) 2017–2025.
methods in over 70% of the cases, producing greater concentration [20] S. Mortazavi, S. Shahrtash, Comparing denoising performance of DWT, WPT,
of the energy of the pulses in the wavelet domain, better denois- SWT and DT-CWT for partial discharge signals, in: Proc. 43rd Int. Universities
ing regarding the minimization of pulse distortions, and smaller Power Engineering Conf., September, 2008, pp. 1–6.
[21] S.G. Mallat, A theory for multiresolution signal decomposition: the wavelet
runtimes. representation, IEEE Trans. Pattern Anal. Mach. Intell. 11 (1989) 674–693.
The proposed NWDLS and the SNRBWS methods comprise [22] M. Misiti, Y. Misiti, G. Oppenheim, J. Poggi, Wavelet Toolbox Users Guide, vol.
therefore a more effective methodology to select the best wavelet 4, The MathWorks, Inc., Natick, 2012 (Chapter 5).
[23] D. Donoho, De-noising by soft-thresholding, IEEE Trans. Inf. Theory 4 (1995)
decomposition to denoise PD signals, providing better denoising 613–627.
results and lower computational load. Its application is expected [24] D. Donoho, I. Johnstone, Adapting to unknown smoothness via wavelet shrink-
to greatly improve on-site PD measurements and, consequently, age, J. Am. Stat. Assoc. 90 (1995) 1200–1224.
[25] IEC 60270: High Voltage Test Techniques: Partial Discharge Measurements, 3rd
on-line diagnosis of high voltage electrical equipment. ed., 2000.
[26] C.H. Kim, R. Aggarwal, Wavelet transforms in power systems: II. Examples
References of application to actual power system transients, Power Eng. J. 15 (4) (2001)
193–202.
[27] C.F.F.C. Cunha, et al., An improved scale dependent wavelet selection for data
[1] P.P. Vaidyanathan, Multirate Systems and Filter Banks, Prentice-Hall, 1993.
denoising of partial discharge measurement, Proc. IEEE Int. Conf. Solid Dielectr.
[2] I. Shim, J. Soraghan, W. Siew, Digital signal processing applied to the detection
(2013, June) 100–104.
of partial discharge: an overview, IEEE Electr. Insul. Mag. 16 (3) (2000) 6–12.
[28] H.P. Amorim, et al., On-site measurements of partial discharges through tap of
[3] I. Shin, J. Soraghan, W. Siew, Detection of PD utilizing digital signal processing
the bushings Brazilian experience in power transformers, Proc. IEEE Int. Conf.
methods: Part 3. Open-loop noise reduction, IEEE Electr. Insul. Mag. 17 (1)
Solid Dielectr. (2013, June) 1020–1023.
(2001) 6–13.
[29] L. Satish, B. Nazneen, Wavelet-based denoising of partial discharge signals
[4] S. Sriram, S. Nitin, K. Prabhu, M. Bastiaans, Signal denoising techniques for par-
buried in excessive noise and interference, IEEE Trans. Dielectr. Electr. Insul.
tial discharge measurements, IEEE Trans. Dielectr. Electr. Insul. 12 (6) (2005)
10 (2) (2003, April) 354–367.
1182–1191.

Вам также может понравиться