2019 International Conference on High Voltage Engineering and Power System

October 1-4, 2019, Bali, Indonesia

Maximum Likelihood-based Technique for Accurate

Estimation of Time-delay between UHF Signals Radiated
from Partial Discharge Sources
Bhukya Anitha Chiranjib Koley
Department of Electrical Engineering Department of Electrical Engineering
National Institute of Technology Durgapur, National Institute of Technology Durgapur,
Durgapur, West Bengal, India Durgapur, West Bengal, India
[email protected] [email protected]

Abstract—One of the useful way for localization of partial and invisible light, and chemical changes in insulation
discharge (PD) sources inside any electrical substation is the use material, etc. [5]. The standard detection methods for the
of multiple ultra-high frequency (UHF) sensors. UHF signals partial discharge is the electrical method, where the test objective
radiated from any PD source can travel a longer distance and is directly connected through a coupling capacitor used for the
can be captured by the multiple UHF sensors placed around the sudden short current pulses and act as a voltage divider. Due to
substations. The localization of PD sources is done primarily by several limitations like sensitivity decreases with the increased
calculating the time delays between the impulsive signals capacitance of test object, some alternative methods are
received at the UHF sensors. proposed by various researchers in the past. These are acoustic
However, the time delay estimation is very challenging in the
method [6], optical method [7], and Radio-Frequency RF/UHF
presence of interference of other Electro-Magnetic signals. In
the present work, several time domain and frequency domain
measurement [8], techniques. In the acoustic method, the
based techniques have been investigated and applied on several acoustic sensor captures the acoustic signals generated during
PD signals captured from the different air-insulated substation the partial discharge.
in India. Various time-domain based techniques that have been The monitoring of radiofrequency signal in UHF (300
applied are the first peak (FP), cross-correlation (CC), and MHz-3 GHz) frequency range could be useful in the case of
cumulative energy (CE). The frequency-domain based monitoring entire air-insulated substation. For these RF
Generalized Cross Correlation (GCC) technique has also been signal has low attenuation and it travels a longer distance
applied with their different weighing functions such are the
through air insulating medium than the acoustic signal. Due
Phase Transform (PHAT), Smoothed Coherence Transform
(SCOT), Maximum Likelihood estimator and the Eckart Filter
to this, it is very helpful to apply to the energized system
(EF). From the experimental result, it is observed that the GCC without any direct contact with the test system.
with ML estimator provide the most suitable result across In the paper, the UHF method is applied for PD detection
different experimental conditions. by using four UHF sensors are placed at a different position
around the PD source [11, 12]. Thereafter the time delay
Keywords— High Voltage power apparatus, Partial Discharge
estimation techniques have been applied between the
localization, Substation, Ultra High-Frequency sensor, Time
captured signal for localization of detection of the source.
Delay Estimation, Generalized Cross-Correlation.
There are two different techniques for time delay estimation
I. INTRODUCTION are in time-domain based techniques and the other frequency
domain based techniques. The commonly applied time-
The electrical energy is one of the basic needs for the domain based techniques are first peak method (FP) [13, 14],
economic and social development of a nation. In present days cross-correlation method (CC) [14], and cumulated energy
it has become a part of our daily life and one cannot think of method (CE) [14]. Where the FP defined as when the signal
a world without electricity. Interruptions in power supply amplitude exceeds a specific threshold first time instant. The
cause damage to our present society. At present in India, CC method defined as the result of cross-correlation between
power interruptions are mainly caused by equipment failures. two signal is found as TDOA. In the CE method, the time
Historically, it has been observed that the failure in electrical difference between cumulated energy signals is determined
equipment is mainly due to in the insulation failures that by finding the ‘knee point’ where the change in the signal
apparatus [1]. Among the various causes of failures, one of amplitude is ‘sudden’. Although this is achieved based on
the main cause is the occurrence of the partial discharge on human judgment to locate the knee point manually. The
the insulation part [2]. According to IEC60270 standard [3], Generalized cross-correlation frequency-domain based
partial discharge is “a localized electrical discharge that techniques have been applied with its different weighting
partially bridges the insulation between conductors, which functions [15] are the Phase Transform (PHAT) [15,17], the
can or cannot occur adjacent to a conductor”. The successive Smoothed Coherence Transform (SCOT) [15, 16], the
occurrence of partial discharge is gradually degrading the Maximum Likelihood (ML) estimator [15, 16, 18], and the
insulation part and it gets affect other equipment or leads to a Eckart Filter (EF) [15, 16]. From the experiment results, it
found that the GCC with ML function gives accurate results.
total breakdown of the system [4]. Therefore, the measurement
of partial discharge can be a fruitful condition monitoring tool The rest of the paper organized as a section- II, is present
for taking protective action before system failure. about different applied delay techniques for captured UHF
signal from XLPE cable. In the section-III, presents about
Various physical outcomes of partial discharges are simulated and experimental results and discussions in section-
electrical discharge, generation of heat, emission of the acoustic IV and conclusion in section-V.
wave, electromagnetic, optical wave in the form of both visible

978-1-7281-2669-2/19/$31.00 ©2019 IEEE

XXX-X-XXXX-XXXX-X/XX/$XX.00 ©20XX IEEE
 II. METHODS kept at the same position for all 10 locations. After that, the
input voltage increases gradually up to 1.5kv applied at source
A. Experimental Setup
by using HV transformer across the source. When the voltage
A schematic diagram for partial discharge detection in the 1.5 kV applied across the source the partial discharge emits
instrumentation laboratory of NIT Durgapur is shown in Fig. the radio frequency signal. After that, the generated partial
2. The Fig. 1, shows the experimental setup in the laboratory discharge signal captured by using four sensors kept around
a discharge free transformer with 20kv, 100 mA, 50 Hz has the source. The same procedure has been adopted fig 4 shown
been connected across the coupling capacitor. A coupling the grid, source and sensor location for the remaining nine
capacitor of 1000 pf inside two capacitors connected in series locations. For each location, the experiment repeated to three
used as a voltage divider and it steps down the voltage across to four times and data has been stored for the corresponding
a source with 1000:1 ratio. Then the artificial partial discharge location. The fig: 3 shows the experiment in the open place.
signal is generated at a different combination of the electrode
by changing different gaps between the electrode. The
generated partial discharge signals are captured by using
multiple UHF sensors in the frequency ranges of 30-3000
MHz. These UHF sensors are connected through amplifier by
using RG6 connectors and coaxial cables to the digital storage
oscilloscope for further study. The UHF sensors are placed at
a different distance from each other for further analysis. The
radiated UHF signals travel at a speed of light velocity of 3x10
8 m/s in the air medium. Fig 1, shows an experimental setup
in the laboratory, and fig 2, shows the schematic diagram of
the experimental setup.

Fig 3: Experiment in the open place

Fig 1: Experimental setup in the laboratory

Fig 4: A Grid of 1m×1m in open place

C. Simulation Study
In order to validate the time delay algorithms with the
help of the PD like signals where an actual delay and the noise
level are known, simulated signals have been generated using
MATLAB software.
In the work, multiple sets of partial discharge signals have
been generated by using equation (1), [19]
( )= . ( )⁄ . ( )⁄
− 2 (1)
Fig 2: Schematic diagram of the experimental setup
Here G is the amplitude of PD pulses, ξ represents the
B. Data Collection attenuation coefficient, represents center oscillation
A experiment for collection of suitable set of data to frequency of the discharge spectrum, and is the starting
compare different time delay algorithm has been presented in moment of the pulse. To bring the particular partial discharge
open space, by dividing the 13 × 6 grid as shown in fig: 4, signal, the different parameters have been changed with the
in the grid each unit is taken as 1 m. The source location has help of mat-lab programs. To test the algorithm, the noise
been changed at 10 different locations within the selected grid level changed from 0N to 2N, the delay between each signal
and the four sensors kept at S1 (0,3), S2 (13,3), S3 (6.5,6) and taken as 200 micro sec. The exponential α and β have been
S4 (6.5,0) four different position within the grid. For location varied from −2 × 10 to −9 × 10 and −2 × 10 to
1, the source kept at S (2,4) and all sensors are at S1 (0,3), S2 −9 × 10 . When the α is low the β is taken as high, also
(13,3), S3 (6.5,6) and S4 (6.5,0), and these four sensors are when the β is low the α is taken as high. The value of taken
as 0.0300, the G value is 10, and ξ is taken as 1 in the applied to obtain the cross-power spectrum between two
equation. After generating different sets of signals in a groups of the signal. The cross-correlation function between
different combination, different delay techniques have been
the two groups of signals is represented by the below
applied. A basic theory of these techniques is presented in the
following section. equation.
( )= ( ) ∗ ( )
D. Theory for TDOA Methods ( ) (7)
( )= ( ) ∗
( ), ( ) is called a
In the section, the different time domain and frequency Where,
domain based time delay estimation techniques that have weighing function. The schematic block diagram of GCC is
been applied to the experimental and simulated PD signals shown in below fig:5.
are summarized.
1) Time Domain Techniques
Among the various time-domain based techniques, the
commonly applied delay estimation techniques are first peak
technique, cross-correlation, and cumulative energy methods.
These are presented the following subsection in a summarized
manner.
a. First Peak (FP) Method Fig: 5 The schematic block diagram of GCC
The first peak method defined as, the first signal peak that
The different weighing functions are selected to achieve
crosses some predetermined threshold value is considered as
suitable results. In the paper four different types of weighing
the first peak. In this paper, 15-25% of the highest peak is functions have been developed [15, 16], are discussed below.
chosen as the first peak. It is a very accurate method for
TDOA, but in some cases when the predetermined threshold a. The Smoothed Coherence Transform (SCOT)
misses the first peak in one of the signals, the result becomes When the cross-spectrum of ( ) ( )
highly erroneous. are not same in that case the ( ) is chosen to be ( ) =
b. Cross-Correlation Method (CC) ( )= , therefore, the weighing
( ) ( )
The cross-correlation applied on two signals, where function of equation (7).
one waveform considers a stationary position and other is
shifted toward the stationary one. Then resultant between ( ) = 1⁄ ( ) ( ) (8)
these two waveforms is then calculated. In the case of SCOT, the noises ( ) ( ) are
The signals received from the source presented in not same and exhibits some spreading. To avoid this problem,
equation 1, and 2 in the presence of noise at two spatially the PHAT transform given below.
separated sensors can be mathematically modeled as; b. The Phase Transform (PHAT)
= ( ) ( ) (2) The PHAT uses the weighing function as given in
= ( ) ( ) (3) equation (9).
1
Where ( ), ( ), ( ) are real, jointly ( )= (9)
stationary random processes. Signal ( ) is assumed to be ( )
uncorrelated with noise ( ) ( ), D is the time delay In this case, the uncorrelated noise ( ) = 0.
to be estimated with the help of (4). c. The Eckart Filter
( )= ( ) ( − ) (4) The weighing function for Eckart filter is assigned
where E denotes expectation. The argument that on the basis of signal to noise ratio, is given below equation
maximizes (4) provides an estimate of delay. (10).
( )
c. Cumulative Energy Method (CE) ( )= (10)
The CE method converts the time domain impulsive ( ) ( )
signal (V) to a cumulative energy signal (E), by using d. The ML Processor
equation (5). The equation (6) used for calculation in the In ML estimation the signal of (3) and (4) assume to
digital form. Where the resistance R=1 ohm, the cumulative be Gaussian and the weighting function is given below
energy up to sample number N. equation (11).
1 | ( )|
( )= (5) ( )= . (11)
( ) 1 − | ( )|

( )= ( ) (6) ( )
Where, ( ) =
( ) ( )
2) Frequency Domain Techniques
III. RESULTS AND DISCRETION
• Generalized Cross-Correlation
In the generalized cross-correlation (GCC) technique In this section typical nature of the PD waveform as
different weight frequency domain functions have been obtain from the experimental study (discussed in section-2A)
is presented first, thereafter similar simulated waveform measure the exact distance between the PD source and the
generated to verify the time delay algorithm by with known UHF sensors.
exact delay and noise level are presented. The performance
B. Simulated Waveform
of various time-delay algorithm as discussed in section-2D
are presented in the following sub-section. The similar kind of simulated waveform as obtained from
the experimental study under various condition have been
A. Experimental Result generated. A typical simulated waveform similar to the
A typical PD characteristic as obtained from the defected waveform as obtained from the experimental study. When the
XLPE cable with an application of 1.5kv from the PD free UHF sensor is at close position (within 3 meters), is shown in
transformer is shown in fig: 6, for a complete cycle of 50Hz. fig: 9. A PD waveform with a poor signal to noise ratio as
The fig:7 shows the captured partial discharge waveform generally obtained from a UHF sensor at the far end from the
from the UHF signal. The sharp spikes in the waveform are PD source as shown in fig: 10. A very extra polluted extreme
due to PD events. A zoom view for the first PD event of fig: case that may arise if the UHF sensor is at the very far end
6, is presented in fig: 7. A similar PD event captured by (greater than 100 meters) is shown in fig:11.
sensor kept at a distance of 13 m is shown in fig: 8.

Fig 9: Simulated UHF Signal at 3 m.

Fig 6: Captured complete PD signal

Fig 10: Simulated UHF Signal at far end

Fig 7: Zoom view for first PD event

Fig 11: Simulated UHF Signal at very far end (>100 m).

C. The performance of the time delay algorithm on the

simulated waveform
Fig 8: PD pulse at 13 m
The performance of the time delay algorithm on the
From fig: 7 and fig: 8 observed that as the distance simulated waveform with the known delay of 200 ns is
between PD source and UHF sensor increases the signal to presented in table-1, in two groups which are time-domain
noise ratio is decreases. However, from the experimental based and frequency-domain based techniques. The table also
study, it is difficult to control the noise level and also to shows the performance of all the techniques at a different
noise level as the Table-1 with the different noise level. It can
be seen that at 0N level all the techniques performing well. IV. CONCLUSION
However, with the increase of noise level, the error as In this paper, the performance of different time and
calculated using (12) increases. frequency domain based time-delay estimation technique
(%) have been evaluated on experimental signals collected using
− artificial PD sources in open space. To evaluate and more in-
= 100 × (12)
depth understanding of the experimental results, simulated
signals also been used by with different signal-to-noise ratio
The last row of the Table-1 shows the average error for and the different rise and fall times. The simulation results
the respective algorithm obtain by the respective column show that the performance of all the techniques falls heavily
from the average errors it can be seen that for time-domain with the increase in noise level. However, it is observed that
cross-correlation and in the frequency domain the ML the cumulative energy-based method performs well, even in
algorithm is performing well. However, it can be further the presence of high noise. The average error obtained in
observed that when the noise level is low both algorithm is delay estimation during the experimental study is
performing well. significantly high in comparison with simulation results due
D. Performance on Experiment Data to the presence of high noise. This is also due to measurement
errors involved in the measuring distances between source
A similar kind of table as presented in table one for the
experiment data is shown in table-2. However, in this case, and UHF sensors. From the experimental results, it can be
observed that the maximum likelihood-based technique
the noise level in the experimental study cannot be controlled.
performs best. However, the cumulative energy-based
Instead, the distance between the source and sensor has been
method remains in the second-best position.
varied. The first column of Table-2 indicates the different
location as discussed in section-2A and also shown in fig: 4. The future scope of work involves, investigation on
From the table-2 it can be observed the overall error across why the maximum likelihood-based techniques which did not
all the method in comparison with simulated study increases perform well in the noisy simulated signal but performed best
significantly. From the average error as indicated in the last on the signals collected from the experimental study.
row it can be seen that both cross-correlation and maximum
likelihood algorithm is performing well. ACKNOWLEDGMENT
TABLE I. ERROR (%) CHARACTERISTICS FROM SIMULATION STUDY
The authors gratefully acknowledge the SERB, DST,
GOVT.OF INDIA, for the support of this research work
Time Domain Frequency Domain under the grant (GRANT NO. EMR/2016/007840, Dated
Noise 02/08/2017).
Level
CC FP CE PHAT SCOT ECKART ML
REFERENCES
0N 0 0 0 -0.1 -0.1 -0.1 -0.01 [1] C. L. Wadhwa, “High Voltage Engineering (second edition)” 2007,
New Age International (P) Limited, publishers.
0.2 N 0 0. 0 -4.16 -4.13 -0.71 -0.05 [2] Power Systems “Sivaji Chakravorti Debangshu Dey Biswendu
Chatterjee” Recent Trends in the Condition Monitoring of
Transformers Theory, Implementation and Analysis
0.5 N 0 0 -0.2 -4.6 -6.2 15.3 -0.07
[3] IEC (3) International Standard 60270, “High-Voltage Test Techniques-
Partial Discharge Measurement”, International Electrotechnical
N -0. -3.7 -0.0 -19.0 -19.1 -26.3 -25.8
Commission (IEC), Geneva, Switzerland, 3rd Edition, 2000.
2N -17.7 -15.8 -4.6 -27.0 -24.0 -27.0 -20.2 [4] D. K. Mishra, B. Sarkar, C. Koley and N. K. Roy “Localization of
Partial Discharge source in High Voltage apparatus using multiple
UHF sensors”978-1-4678-6503-1/15/$31.00 © 2015 IEEE.
TABLE II. ERROR (%) CHARACTERISTICS FROM EXPERIMENTAL [5] E. Kuffel, W.S. Zaengl, J. Kuffel, “High Voltage Engineering
STUDY Fundamentals, Second edition” Reprinted 1986, Newnes.
[6] W. Sikorski and W. Ziomek, “Detection, Recognition and Location of
Time Domain Frequency Domain Partial Discharge Sources Using Acoustic Emission Method” available
Noise online at www.intechopen.com.
Level
CC FP CE PHAT SCOT ECKART ML [7] S. Biswas, C. Koley, B. Chatterjee, S. Chakravorti, “A Methodology
for Identification and Localization of Partial Discharge Sources using
1 19.6 30.5 27.9 35.2 25.7 37.5 8.6 Optical Sensors” IEEE Trans. Dielectr. Electr. Insul., Vol. 19, No. 1,
pp. 18-28, 2012.
2 36.5 32.8 31.4 35.5 41.9 32.2 32.8
[8] I. E. Portugués, P. J. Moore, I. A. Glover, C. Johnstone, R. H.
3 31.4 16.5 18.1 42.9 33.5 15.5 25.0 McKosky, M.B. Goff, and L. vander Zel, “RF-Based Partial Discharge
Early Warning System for Air-Insulated Substations”, IEEE Trans.
4 35.2 42.7 42.5 44.5 40.9 30.7 18.3 Power Delivery , Vol. 24, No. 1, pp. 20 – 29, 2009.
[9] D. K. Mishra, B. Sarkar, C. Koley and N. K. Roy “An Unsupervised
5 14.7 28 25.3 35.0 42.0 30.5 -5.2 Gaussian Mixer Model for Detection and Localization of Partial
Discharge Sources using RF Sensors” IEEE Transactions on
6 -3.1 9.4 3.8 45.8 50.5 34.7 18.2
Dielectrics and Electrical Insulation Vol. 24, No. 4; August 2017.
7 -0.1 -6.2 -5.8 45.8 -2.1 15.0 42.4 [10] Dipak Kumar Mishra, Chiranjib Koley and Nirmal Kumar Roy “A
Novel TDOA Estimation Method for UHF Sensor based PD
8 23.7 23.3 12.7 18.1 19.7 21.0 -5.0 Localization System in HV Power Apparatus”. 978-1-5090-3239-
6/17/$31.00©2017IEEE.
Overall
19.6 22.1 19.5 37.8 31.5 27.1 16.9 [11] Bhukya Anitha, Dipak Kumar Mishra and Chiranjib Koley
Average
“Localisation of Partial Discharge source through Delay calculation
 using UHF sensors” 2017 IEEE International WIE Conference on
Electrical and Computer Engineering (WIECON-ECE).
[12] Bhukya Anitha and Chiranjib Koley “Time Delay Estimation of Ultra-
High-Frequency Signals Radiated from Partial Discharge Sources”
2018 IEEE ASPCON.
[13] D. Dukanac, "Application of UHF method for partial discharge source
location in power transformers," in IEEE Transactions on Dielectrics
and Electrical Insulation, vol. 25, no. 6, pp. 2266-2278, Dec. 2018.
doi: 10.1109/TDEI.2018.006996
[14] H. H. Sinaga, B. T. Phung and T. R. Blackburn, “Partial Discharge
Localization in Transformers Using UHF Detection Method” IEEE
Transactions on Dielectrics and Electrical Insulation Vol. 19, No. 6;
December 2012.
[15] Charles h.knapp and g.clifford carter “the generalised correleation
method for estimation of time delay” IEEE Transactions on
Acoustics,Speech and Signal Processing, Vol. 24. No.4 August 1976.
[16] A. Hero and S. Schwartz, "A new generalized cross correlator,"
in IEEE Transactions on Acoustics, Speech, and Signal Processing,
vol. 33, no. 1, pp. 38-45, February 1985. doi:
10.1109/TASSP.1985.1164515
[17] Y. Tian, A. Tatematsu, K. Tanabe and K. Miyajima, "Development of
Locating System of Pulsed Electromagnetic Interference Source Based
on Advanced TDOA Estimation Method," in IEEE Transactions on
Electromagnetic Compatibility, vol. 56, no. 6, pp. 1326-1334, Dec.
2014.
[18] Chye Huat Peck and P. J. Moore, "A direction-finding technique for
wide-band impulsive noise source," in IEEE Transactions on
Electromagnetic Compatibility, vol. 43, no. 2, pp. 149-154, May 2001.
doi: 10.1109/15.925534
[19] H. Hou, G. Sheng and X. Jiang, "Robust Time Delay Estimation
Method for Locating UHF Signals of Partial Discharge in Substation,"
in IEEE Transactions on Power Delivery, vol. 28, no. 3, pp. 1960-
1968, July 2013. doi: 10.1109/TPWRD.2013.2262315.