An Enhanced Hankel Matrix based Singular Value

Decomposition Method for Removing Noise from

Partial Discharge Signals
Kishonica JG Gayathri A Suganya Govindarajan Kannan Krithivasan
School of EEE School of EEE School of EEE School of Education
SASTRA Deemed University SASTRA Deemed University SASTRA Deemed University SASTRA Deemed University
Thanjavur. Thanjavur. Thanjavur. Thanjavur.
2019 11th International Conference on Advanced Computing (ICoAC) 978-1-7281-5286-8/20/$31.00 ©2020 IEEE 10.1109/ICoAC48765.2019.246869

Email: kkannan [email protected]

Abstract—Partial Discharges (PD) measurement has long noise. The magnitude and frequency spectrum of the inter-
been used as a test to evaluate insulation condition in electrical ference are considerably identical. These interferences cause
equipment. Different types of noise, such as white noise, random false predictions about the insulation level so the reliability
noise, and discrete spectral interference couples with the PD
signal during on line and/or onsite PD measurements. Because of as well as sensitivity of the PD data [2]. In this paper, a new
these interferences PD source separation becomes troublesome method named as Enhanced Singular Value Decomposition
process. In this present work, combination of Hankel Matrix (E-HSVD is introduced to de-noise the PD signal and com-
based Enhanced Singular Value Decomposition (E-HSVD) is pared with the other traditional techniques reported below in
proposed to remove the noise from PD signals. An adaptive the background literature.
spectral kurtosis is employed to select the optimal singular
component obtained by applying E-HSVD to the PD signal. II. BACKGROUND L ITERATURE
The proposed technique is applied on the PD signals using
simulated and PD signal measured at online onsite to exam- The major classification of filters used to suppress the
ine its performance under different noisy environments. The interference in PD signal is categorized as hardware based de-
evaluation metrics results confirm that E-HSVD has significant noising techniques and software based denoising techniques.
improvements in performance compared to existing state of the
art PD denoising techniques. A. Hardware Based Denoising Techniqus
Index Terms—Partial Discharge, Singular Value Decomposi-
tion, White Noise, Random Noise, Spectral Kurtosis It is widely used during online PD measurement for
the suppression of DSI. Basic hardware based methods are
I. I NTRODUCTION Filtering method, Differential System, and Pulse Polarity
Discrimination System . It is evident by the previously
The transformer is one of the most expensive equipment
reported literature that hardware based techniques suppress
in power system. In a transformer, the winding insula-
DSI and some external interferences effectively [3]. Hence, it
tion generally undergoes mechanical, electrical and thermal
is worth mentioning that this methodology may not suppress
stresses. In most of the cases, the transformer failures occur
white noise effectively.
because of insulation deterioration. As a point of fact, signals
appear when there is insulation breakdown (occurs because B. Software Based Denoising Techniqus
of gradual erosion) or premature damage in the windings. Feser proposed Fast Fourier Transform (FFT) based noise
As a consequence, evaluation of Partial Discharge (PD) is removal. Borsi reported advantages and disadvantages of us-
undoubtedly effective nondestructive tool to determine the ing adaptive filters to remove DSI from PD signals. Extensive
insulation level and it is necessary for taking corrective computation and processing time and prior knowledge of
measures such as maintenance and repair work etc. the noise and signal characteristics are the limitations in the
According to IEC 60270 standard, PD is an electrical adaptive filtering methods. The methods mentioned above
discharge (short electrical spark) which occurs by partially mostly eliminate DSI effectively and fails to remove white
bridging gap of the conducting electrodes. From the char- noise and pulse shaped interferences [4].
acteristics of PD signal the location and type of discharge Recently Wavelet Transform (WT) has been proposed as an
can be revealed [1]. Based on its principal characteristics, efficient tool to analyze irregular and non-stationary signal in
PD detection is ensured by the measurement of apparent timefrequency domain [5].Although Wavelet performs well,
charge, the rate of repetition and inception voltage and so on due to non linear nature of PD pulses the corresponding
neither when the measuring impedance (RC, RLC) is in series mother wavelet does not serve the purpose for majority of
connection with the test object nor by parallel connection of samples. In this paper, we propose a method for PD denoising
a coupling capacitor. that makes explicit reference to the behavior of dynamic
PD signals are typically small electrical pulses, char- linear system (we recall that the PD pulse is practically
acterized as non-periodic impulses with time duration of the impulse response of the electrical equipment) and has
10−9 to10−7 s. However, certain noises interfere with the a unique tuning parameter .
PD signal and distort the waveform which leads to mis-
categorization. The prime classification of PD interference
encountered during online PD measurements are Discrete
Spectral Interferences (DSI), white noise and pulse-shaped

978-1-7281-5286-8/19/$31.00 ©2019 IEEE 367

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 14:20:56 UTC from IEEE Xplore. Restrictions apply.
 III. H ANKEL M ATRIX BASED E NHANCED S INGULAR
VALUE D ECOMPOSITION
HLs = S ∗ S T (4)
The diagonal matrix contains the square root of the eigen
Singular Value Decomposition (SVD) has been widely
values of HLs and represented as
adopted for noise elimination in various fields such as acous-
tics, electronics, smart control and geophysical time series Σ = [diag(σ1 , σ2 , σ3 .....σR ), 0] (5)
and so on. However, its feasability in denoising PD signals
has not been considered so far. R = min(m, n) (6)
In this work, PD pulses are represented using Hankel
matrices. The procedure for deriving a Hankel matrix from These σ1 , σ2 , σ3 .....σR are called the Singular Values (SV)
a vector representing the PD pulse is known as embedding. of the matrix[S] and σ1 ≥ σ2 ≥ σ3 .....σR ≥ 0, which is
The matrix is filled by delaied replicas of the same vector. decided by whether m ≤ n or m ≥ n, while O is zero
A PD signal corrupted with noise can be represented by matrix and q = min(m, n). The rank of the matrix represents
the number of independent hidden variables in the data. The
Y (n) = x(n) + h(n) (1) dimension of HLs is (m*m). Building of Hankel matrix from
original time series is called embedding.
where the measured signal Y (n) is the superposition of the
This paper proposed an enhanced technique for data re-
noise free PD signal X(n) and the noise h(n).
construction. The principle of enhanced SVD (E-HSVD)
A. Principle of SVD reconstruction is based on PCA data reconstruction technique
Singular Value Decomposition (SVD) was introduced by [7]. The reconstruction has been given by,
Beltrami & Jordan in the 1870’s for real square matrices. Ȳ = Y ∗ V ∗ V T (7)
SVD is a numerical method, which decomposes the matrix
[s] of rank R into product of three matrices. This method is where Ȳ is the reconstructed test data matrix of dimension
usually represented as, m * n and V is the principal right singular vector matrix of
T the training data of dimension n * n.
S = [U ]m∗m [Σ]m∗n [V ]n∗n (2)
The squared envelope (SE) is calculated as absolute
where [U] and [v] are orthogonal matrices; [Σ] is a diagonal squared value of the analytical signal which is obtained by
matrix with non-negative entries. The columns of U & V are Hilbert transform applied to the real signal x(t).
called left & right Singular Vectors and the diagonal elements
are the Singular values. IV. K URTOSIS OF T HE S QUARED E NVELOPE S PECTRUM
The Singular value importance is merely depicted in its Kurtosis has been proven to be a very useful monitoring
magnitude. Especially, the square of Singular value is propor- parameter measuring the waveform peak and sensitive to the
tional to the variance of each Singular Vector. This property impulse characteristics of signal. Generally, kurtosis repre-
makes SVD more compatible for denoising. sents the characteristic of signals around its mean value and
For the signal Y = [Y (1), Y (2), · · · Y (N )], the Hankel characterizes the strength of transient impulses and is defined
matrix is constructed as follows: as the ratio of the fourth central moment of the signal to the
squared second central moment of the signal . It is expressed
⎡ ⎤
Y (1) Y (2) ··· Y (n) in the following equation,
⎢ Y (2) Y (3) ··· Y (n + 1)⎥ 
⎢ ⎥ E (x − μ)4
S=⎢ . .. .. .. ⎥ (3) kurtosis(x) = −3 (8)
⎣ .. . . . ⎦ σ4
Y (m) Y (m + 1) ··· Y (N ) (m∗n) In this paper, E-HSVD filter is proposed to remove the
external interferences and its effectiveness on noise reduction
where 1 ≤ n ≤ N . S is a Hankel matrix ; m is the embedding
is demonstrated by means of figures and numeric values. The
dimension which satisfies (N + 1 − m), where N represents
E-HSVD algorithm is given in Fig. 1.
total number of PD samples and m is the number of rows in
a matrix and n is columns [6]. V. N UMERICAL M ODELLING OF PD
PD signals are characterized as non-periodic electric im-
pulses. Location of PD, equipment size and the type of mea-
surement decide shape and characteristics of PD. External
interference is not possible as the high voltage laboratories
are shielded using faraday cage. However, noises from supply
system is possible to couple with PD data . Depending on the
measurement circuit, insulation failure etc. PD pulse models
differ. So, to evaluate the denoising performance, damped
exponential pulse (DEP) realized by RC circuit and damped
oscillatory pulse (DOP) realized by RLC circuit are numeri-
cally modeled by the following equation. The waveforms are
Fig. 1. Flowchart of E-HSVD expressed by the following mathematical equation

368

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 14:20:56 UTC from IEEE Xplore. Restrictions apply.
 −t −t
DOP = A ∗ sin(2Πfc )(exp − exp ) (9)
t1 t2
In the equation, A corresponds to pulse peak value, t1 , t2
are the pulse time constants and FC are the oscillatory
frequency of the DOP pulse. In particular white noise is
added to the simulated signal, to resemble actual PD signal
corrupted with noise. The simulated PD with noise is shown
in the Fig. 2.

Fig. 2. Simulated Damped Oscillatory Pulse corrupted by noise (PD with

Noise)

VI. PD M EASUREMENT FROM O NLINE – O NSITE

E XPERIMENT
In this work, PD measured from M/s. Shree Abirami
Engineering Works industries are also considered to verify
the proposed PD signal processing methods further. The PD
testing was performed by a 33kV /110kV three-phase trans-
former manufactured by Andrew Yule & Co. Ltd. A 750kV A
transformer was taken as a test specimen. The experiments Fig. 3. PD Signal measured at Online
were performed using the DTM model- WS Test Systems
analog PD meter. A Digital Storage Oscilloscope (DSO),
where Di denote the original PD signal and Ri denotes the
Keysight DSO1052B with 1GS/s bandwidth range was
denoised PD signal.
used to capture the PD signal. Fig. 3 shows the complete PD
MATLAB software is used here for simulation of signal
measurement setup.
and noise. The simulated white gaussian noise is added with
VII. P ERFORMANCE C OMPARISON simulated PD signal to create noisy signal . As illustrated
wavelet transform based on soft (WT-ST), hard (WT-HT)
This section validates the effectiveness of filtering effi- and Empirical Mode Decomposition (EMD) methods are
ciency of the GA based wavelet scheme by applying to both considered here for denoising simulated signals corrupted
real and simulated PD data corrupted by white noise. The with noise.
denoised results are compared with wavelet transform and Fig. 4 present the denoising results of the simulated PD
EMD based Wavelet transform respectively. The following signal. Fig. 4d shows the potential of the proposed E-HSVD
evaluation metrics are computed in the context of comparison method to restore the original PD signal. The results tabulated
of the conventional methods with the the proposed technique: in Table I without a fail show the superiority of the E-HSVD
Signal to Noise Ratio (SNR), Mean Square Error (MSE), method over the recent wavelet-based and EMD denoising
Root Mean Square Error (RMSE), Peak Signal to Noise Ratio methods. It is clearly apparent that E-HSVD achieves an
(PSNR). increase in the SNR, as it enhances the SNR to 11.542 dB
(denoised PD) and MSE reduced to 0.134. So, it is evident
max(Di ) that the proposed scheme preserves the information of PD
P SN R = 20 log10 (10)
1 N 2 signal. In contrast, there is a distortion in shape and amplitude
N i=1 |Di − Ri |
of the PD signal when other methods used.
N Fig. 4 (a-d) depicts the denoised PD signal using the
1 2 methods of WT-HT, WT-ST, EMD and E-HSVD. It is capable
M SE = (Di − Ri ) (11)
N i=1 of restoring the original signal with the least distortion with
MSE = 0.134. The numerical results and the Fig. 4d shows
N 2
i=1 |Ri | that proposed E-HSVD method, denoised signal effectively.
SN R = 10 log10 N 2
(12)
i=1 |Di − Ri | VIII. E XPERIMENTAL R ESULTS
 The high SNR and low MSE ensures the ability of pro-
1 N
RM SE = 
2 posed E-HSVD technique performance in denoising simu-
|Di − Ri | (13)
N i=1 lated signal. Next, to justify the proficiency of the E-HSVD

369

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 14:20:56 UTC from IEEE Xplore. Restrictions apply.
 Fig. 5. Denoising results of onsite measured PD signals a) WT-ST b) WT-
HT c) EMD d) E-HSVD

TABLE II
Fig. 4. Denoising results of simulated PD signals A) Noisy simulated PD C ASE STUDY- II: E VALUATION M ETRICS C OMPARISON - R ESULT
Signal a) WT-ST b) WT-HT c) EMD d) E-HSVD
Method MSE SNR CC RMSE PSNR
TABLE I WT-ST 0.405 9.821 0.94 0.25 10.135
C ASE STUDY- I: E VALUATION M ETRICS C OMPARISON - R ESULT WT-HT 0.324 10.421 0.96 0.13 11.214

Method MSE SNR CC RMSE PSNR EMD 0.218 11.823 0.95 0.108 12.985

WT-ST 0.354 6.928 0.94 0.357 8.782 E-HSVD 0.156 14.321 0.97 0.006 16.874

WT-HT 0.278 7.731 0.93 0.218 9.818

EMD 0.201 8.542 0.97 0.198 10.081
E-HSVD 0.134 10.955 0.97 0.16 11.542 laboratory data.

IX. C ONCLUSION
A novel algorithm for removing noise from PD signal
in the context of real time measurement, an actual PD signal based upon E-HSVD proposed in this paper. The performance
corrupted with noise is considered here. Fig. 5 shows the of E-HSVD is compared with some of the methods reported
denoised signals using WT-HT, WT-ST, EMD and E-HSVD. in literature such as wavelet (soft and hard thresholding) and
From the figure, it is evident that wavelet based on hard EMD thresholding using the evaluation metrics SNR, CC,
thresholding loses the signal originality. MSE and PSNR. The calibre of the proposed scheme was
The results of denoising demonstrate the efficacy of the tested on both simulated and real noisy PD data. Experimen-
proposed method in practical situations, where higher SNR tal and simulation results indicate that E-HSVD algorithm
is achieved by E-HSVD than both the conventional as well as performs more effectively to extract the PD pulse from strong
recent wavelet-based methods of denoising Fig. 5d confirms background noise.
the superiority of E-HSVD over the existing wavelet-based
methods . The SNR achieved by E-HSVD is 14.321, whereas ACKNOWLEDGMENT
the one obtained by wavelet is 10.421. So, it is evident from The authors thank the Management of SASTRA University
the results that E-HSVD performs better when compared to and DST-FIST (Sanction order ref: SR/FST/ETI-338/2013(C)
other traditional techniques for both simulated signals and dated10/09/2014) and (Sanction order ref SR / FST / MSI

370

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 14:20:56 UTC from IEEE Xplore. Restrictions apply.
– 107 / 2015(C)) for their motivation and financial support [3] V. Nagesh and B.I. Gururaj, “Automatic detection and elimination of
for creating the facilities at High Voltage Lab and Discrete periodic pulse shaped interferences in partial discharge measurements,”
IEE Proc. - Sci. Meas. Technol., vol. 141, no. 5, pp. 335 - 342, 1994.
Mathematics Lab of SASTRA, with which the experimental [4] U. Koepf and K. Feser, “Rejection of narrow-band noise and repeti-
works were carried out. The authors also wish to thank M/s tivepulses in on-site PD measurements,” IEEE Trans. Dielectr. Electr.
Sri Abirami Industries, Chennai for providing neccessary Insul., vol. 2 no. 6, pp. 1180-1191, 1995.
[5] R. Hussein, K. B. Shaban, and A. H. El-Hag, “Wavelet TransformWith
data. Histogram-Based Threshold Estimation for Online Partial Dis-charge
Signal Denoising,” IEEE Transactions on Instrumentation andMeasure-
R EFERENCES ment, vol. 64, no. 12, pp. 3601-3614, 2015.
[1] “High Voltage Test Techniques in Partial Discharge Measurements”, [6] G. Suganya, S. Jayalalitha, A. Cavallini, K. Kannan and J. Jaikanth,
IEC Std. 60270, 2000 ”Development of Hankel-SVD hybrid technique for multiple noise re-
[2] H. Zhang, T.R. Blackburn, B.T. Phung and D. Sen, “A novel wavelet moval from PD signature,” IET Science, Measurement & Technology,
transform technique for on-line partial discharge measurements part 1: vol. 13, no. 8, pp. 1075-1084, 2019.
WT de-noising algorithm,” IEEE Trans. Dielectr. Electr. Insul., vol. 14, [7] S. Mandal, B. Santhi, S. Sridhar, K. Vinolia, P. Swaminathan, ”Sensor
(2), pp. 3-14, 2007. fault detection in Nuclear Power Plant using statistical methods”,
NUCL ENG DES, vol. 324, pp. 103-110, Dec. 2017.

371

Authorized licensed use limited to: Auckland University of Technology. Downloaded on May 28,2020 at 14:20:56 UTC from IEEE Xplore. Restrictions apply.