Remote Assessment of High Voltage

Porcelain Insulators Using Radiated

Electromagnetic Field Signature

by

Ehsan Azordegan

A Thesis Submitted to the Faculty of Graduate Studies of

The University of Manitoba
in Partial Fulfilment of the Requirements for the Degree of
Doctor of Philosophy

Department of Electrical and Computer Engineering

University of Manitoba
Winnipeg, Manitoba, Canada

©November 2015 Ehsan Azordegan

 Abstract

A novel approach for inspecting the condition of porcelain insulators

based on statistical analysis of electromagnetic radiations of live
insulators is demonstrated in this thesis. Physical defects such as
puncture and contamination can degrade the insulators performance
and result in power outages, potentiating costs to utilities. Therefore,
condition assessment of line insulators has always been one of the
most important aspects of maintenance programs in power networks.
Realistic replicas of punctured and contaminated insulators were created
in the High Voltage Lab at University of Manitoba, following the
IEC standards. These defective insulators were tested under high
voltage stress while the electromagnetic radiations originated from
the partial discharge activities on the insulators were captured using
electromagnetic sensors. During the experimental part of this thesis,
a multitude of tests were conducted and resulted in measuring and
recording a total of 410,000 cycles of discharge activities. The
feature extraction algorithm, developed as part of this thesis, calculates
the statistical features of the phase resolved interpretation of partial
discharge (PD) pulses. The results of analyzing the extracted features
from the radiated signature of defective insulators indicate that the scale
and shape parameters of a two sided Weibull distribution function fit to
the recorded measurement entail distinct information about the source
of discharges that can be used to identify the source of defects. Based
on the library of features extracted from the recorded electromagnetic
radiations, a support vector machine (SVM) classifier, developed as
part of this thesis, can successfully classify the radiation signature of
punctured and contaminated insulators. Therefore, the main outcome
of this research was introducing a novel porcelain insulator inspection
technique that can remotely differentiate the defective punctured and
contaminated insulators using their electromagnetic radiation signature
in a laboratory environment. By utilizing the signature of common
discharge activities present in the recorded signature of all tested
insulators, a gating algorithm was developed which improved the
successful classification rate from the 51 % to 75%. The inspection
technique proposed in this research can eliminate the safety hazards
involved in the live maintenance of line insulators, lower the maintenance
costs, and improve the inspection efficiency considering the conventional
labour intensive live maintenance assessments.
 Acknowledgements

I would like to extend my utmost gratitude to the many people who

have helped along the way to make this research project possible and
to bring it to its completion. First, I would like to express my sincerest
appreciation to Dr. Behzad Kordi for providing me the opportunity to
take part in his research work. I am so deeply grateful for his help,
invaluable guidance, and support throughout this project and my entire
program of study.

I would like to thank the committee members Dr. Douglas Thomson,

Dr. Greg Bridges, and Dr. Yunhua Luo for their advice, guidance,
and support through the completion of the project. Without their
participation and input, the project could not have reached the level
of fruition it has achieved. I would also like to thank Dr. Chi Yung
Chung, my external PhD examiner, for his time to review my thesis and
make valuable suggestions and comments.

I would also like to acknowledge Dr. David Swatek, from Manitoba

Hydro, who provided technical guidance and support throughout this
research as well as accommodating the temporary test setup of the
project at Manitoba Hydro High Voltage Test Facility (HVTF). I also
want to thank Bill McDermid for his technical support, assistance,
and discussions that allowed me to better understand some of the
fundamental concepts of this work.

Furthermore, many thanks to Nathan Jacob who assisted with the

lab experiments at Manitoba Hydro High Voltage Test Facility and
accompanied me when running tests and conducting measurements as
well as numerous conversations we had discussing results that aided me
in solution development.

Also I would like to thank Daryl Hamelin and Cory Smit at the
University of Manitoba for their help with the experimental setup of
this project.

I would like to acknowledge financial support from Manitoba Hydro,

Mitacs Accelerate, Natural Science and Engineering Research Council
of Canada (NSERC), and the faculty of Graduate Studies, University of
Manitoba.

I must also express my very profound gratitude to my parents; my dear

mother Parvin, who always brightens my day with her positive energy,
and for my reassuring father Firouz, for making me find the strength to
continue my research, and to the rest of my family for the many thoughts
and words of encouragement that helped me get through challenging
times.

Lastly, I want to thank my friends for their unfailing encouragement

and support throughout my years of study, researching, and writing this
thesis. Thank you all for all the wise words, your efforts and everything
you did for me, I could not have done it without you.
Table of Contents

List of Tables viii

List of Figures ix

1 Synopsis 1
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Problem Statement . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.4 Research Outcomes and Contributions . . . . . . . . . . . . . . . . . . . 4

2 Outdoor Insulators 6
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
2.2 Insulator Types . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.3 Design and Manufacturing . . . . . . . . . . . . . . . . . . . . . . . . . . 8
2.4 Electrical and Mechanical Performance . . . . . . . . . . . . . . . . . . . 9
2.5 Standard Quality Tests . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10
2.6 Failures of Porcelain Insulators . . . . . . . . . . . . . . . . . . . . . . . 11
2.7 Contamination of Insulators . . . . . . . . . . . . . . . . . . . . . . . . . 12
2.8 Formation of Dry-Band Arcs . . . . . . . . . . . . . . . . . . . . . . . . . 14

- iii -
 TABLE OF CONTENTS

2.9 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3 Condition Assessment of Porcelain Insulators–A Literature Review 18

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
3.2 Detection of Defective Insulators . . . . . . . . . . . . . . . . . . . . . . 20
3.2.1 Close Contact Methods . . . . . . . . . . . . . . . . . . . . . . . . 20
3.2.2 Acoustic Detection Methods . . . . . . . . . . . . . . . . . . . . . 21
3.2.3 Ultraviolet and Infrared Detection Methods . . . . . . . . . . . 22
3.2.4 Leakage Current Analysis . . . . . . . . . . . . . . . . . . . . . . 24
3.2.4.1 The Effect of Humidity . . . . . . . . . . . . . . . . . . 25
3.2.4.2 Leakage Current Monitoring . . . . . . . . . . . . . . . 25
3.2.4.3 Frequency Domain Analysis of Leakage Current . . . 26
3.2.4.4 Phase Angle Analysis . . . . . . . . . . . . . . . . . . . 27
3.2.4.5 Time–Frequency Analysis . . . . . . . . . . . . . . . . . 28
3.2.5 Electric Field Measurement . . . . . . . . . . . . . . . . . . . . . 29
3.2.6 RF Radiation Detection Method . . . . . . . . . . . . . . . . . . 31
3.3 Classification of Partial Discharge . . . . . . . . . . . . . . . . . . . . . . 33
3.3.1 Distance Function Classifiers . . . . . . . . . . . . . . . . . . . . 34
3.3.1.1 Minimum Distance Classifier . . . . . . . . . . . . . . . 34
3.3.1.2 Polynomial Classifier . . . . . . . . . . . . . . . . . . . 35
3.3.2 Statistical Classifiers . . . . . . . . . . . . . . . . . . . . . . . . . 36
3.3.2.1 Bayes Classifier . . . . . . . . . . . . . . . . . . . . . . . 36
3.3.2.2 Recognition Rate Classifier . . . . . . . . . . . . . . . . 36
3.3.3 Support Vector Machines (SVM) . . . . . . . . . . . . . . . . . . 37
3.3.4 Power Spectrum Classifiers . . . . . . . . . . . . . . . . . . . . . 38

- iv -
 TABLE OF CONTENTS

3.3.5 Neural Network and Fuzzy Logic Based Classifiers . . . . . . . 39

3.3.6 Performance Assessment of Classifiers . . . . . . . . . . . . . . . 40
3.4 Data Processing Methodology of This Thesis . . . . . . . . . . . . . . . 40
3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 43

4 Experimental Setup 45
4.1 IEC Standards Referenced in this Research . . . . . . . . . . . . . . . . 45
4.1.1 IEC 60507: “Artificial pollution tests on high-voltage
insulators to be used on a.c. systems” . . . . . . . . . . . . . . . 46
4.1.2 IEC 60815: “Selection and dimensioning of high-voltage
insulators for polluted conditions” . . . . . . . . . . . . . . . . . 46
4.1.3 IEC 60383-1: “Insulators for overhead lines with a nominal
voltage above 1000 V” . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2 Lab Test Setup . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
4.2.1 High Voltage AC Transformers . . . . . . . . . . . . . . . . . . . 49
4.2.2 Divider . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51
4.2.3 D-dot Electric Field Sensor . . . . . . . . . . . . . . . . . . . . . 52
4.2.4 Current Transformer (CT) . . . . . . . . . . . . . . . . . . . . . . 53
4.2.5 Oscilloscope . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
4.2.6 Defective Porcelain Insulators . . . . . . . . . . . . . . . . . . . . 56
4.2.6.1 Contaminated Insulators . . . . . . . . . . . . . . . . . 56
4.2.6.2 Punctured Insulators . . . . . . . . . . . . . . . . . . . 60
4.2.6.3 Insulator Identification . . . . . . . . . . . . . . . . . . 65
4.3 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 66

-v-
 TABLE OF CONTENTS

5 Data Processing and Classification 67

5.1 Data Processing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
5.1.1 Weibull PDF Fitting on Phase Location of PD Pulses . . . . . 71
5.1.2 Weibull PDF Fitting on Amplitude of PD Pulses . . . . . . . . 73
5.1.3 Weibull PDF Fitting on Weighted Phase of PD Pulses . . . . . 74
5.1.4 Spectral Fractal Analysis . . . . . . . . . . . . . . . . . . . . . . . 75
5.2 Data Processing Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . 76
5.2.1 Weibull PDF Fitting on Phase Location of PD Pulses . . . . . 77
5.2.2 Weibull PDF Fitting on Amplitude of PD Pulses . . . . . . . . 80
5.2.3 Weibull PDF Fitting on Weighted Phase of PD Pulses . . . . . 84
5.2.4 Spectral Fractal Analysis . . . . . . . . . . . . . . . . . . . . . . . 88
5.3 Classification . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91
5.3.1 Classification Algorithm . . . . . . . . . . . . . . . . . . . . . . . 92
5.3.2 Fundamental 2D Classification of All Insulators . . . . . . . . 93
5.3.2.1 Puncture vs Contamination . . . . . . . . . . . . . . . 93
5.3.2.2 Puncture vs Normal . . . . . . . . . . . . . . . . . . . . 97
5.3.2.3 Contamination vs Normal . . . . . . . . . . . . . . . . 101
5.3.2.4 Discussion on the Results of 2D Classifications . . . . 105
5.3.3 Gating Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . 107
5.3.4 Analysis of Gated Classification Results . . . . . . . . . . . . . . 110
5.4 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 113

6 Conclusions and Future Work 114

6.1 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
6.2 Main Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 117

- vi -
 TABLE OF CONTENTS

6.3 Future Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 119

Appendix A Partial Discharges under AC Voltage 120

A.1 Definitions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 120
A.2 Partial discharge process in voids under AC conditions . . . . . . . . . 121
A.3 Equivalent electric circuit for voids inside insulating material . . . . . 125
A.4 Discharge waveform of metallic surfaced gap under AC supply . . . . 127
A.5 Discharge waveform of dielectric surfaced gap under AC supply . . . . 130

References 135

- vii -
List of Tables

4.1 Comparing available standards for site severity definitions based on

the ESDD level. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
4.2 Identification table for studied insulators . . . . . . . . . . . . . . . . . 65

5.1 Features with highest successful classification in Figure 5.20. . . . . . 95

5.2 Features with highest successful classification in Figure 5.23. . . . . . 97
5.3 Features with highest successful classification in Figure 5.26. . . . . . 101

- viii -
List of Figures

2.1 Cross section of a porcelain insulator. . . . . . . . . . . . . . . . . . . . 9

2.2 Formation of dry bands . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

3.1 Detection of defective insulators using the Buzz method. . . . . . . . . 21

3.2 Electric field measurement of porcelain insulators. . . . . . . . . . . . . 31
3.3 Hyperplane problem to illustrate SVM methodology. . . . . . . . . . . 37

4.1 Test arrangement in the High Voltage Lab. . . . . . . . . . . . . . . . . 49

4.2 230 kV potential transformer. . . . . . . . . . . . . . . . . . . . . . . . . 50
4.3 Prodyn’s ACD-S30 Sensor. . . . . . . . . . . . . . . . . . . . . . . . . . . 52
4.4 Recorded voltage of D-dot probe and the CT during a PD activity. . . 53
4.5 Current transformers manufactured by Magnelab/Bergoz. . . . . . . . 54
4.6 Contaminated insulator under the test; clean and healthy unit on top
and contaminated unit at the bottom. . . . . . . . . . . . . . . . . . . . 58
4.7 Measuring the ESDD level in the lab. . . . . . . . . . . . . . . . . . . . 59
4.8 Puncture test arrangement using Tensile Machine Instron DX300;
test setup on left, failed pin on top right, and custom built link and
resistance monitoring on bottom right. . . . . . . . . . . . . . . . . . . . 62

- ix -
 LIST OF FIGURES

4.9 Puncture test arrangement using IEC 60383-1; test setup on left,
insulator immersed in oil and under high voltage stress on top right,
development of puncture on middle right, and punctured insulator
ready to be tested for radiation signature on bottom right. . . . . . . . 64

5.1 View of the Graphical User Interface (GUI) developed in MATLAB

for performing signal processing. . . . . . . . . . . . . . . . . . . . . . . . 70
5.2 Identifying peak locations on the recorded partial discharge pulse. . . 70
5.3 Sample of recorded measurements in one cycle. . . . . . . . . . . . . . . 71
5.4 Fitting a two sided Weibull distribution to phase of PD pulses of
512 segments. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
5.5 Fitting a two sided Weibull distribution to the amplitude of PD pulses. 73
5.6 Fitting a two sided Weibull distribution to the weighted phase of PD
pulses. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 75
5.7 Parameters of spectral fractal curve fitting. . . . . . . . . . . . . . . . . 76
5.8 Fitted phase Weibull parameters - normal insulator; one blue point
(positive half cycle) and one red point (negative half cycle) represent
the extracted features from one of the 176 datasets during the phase
Weibull fitting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 78
5.9 Fitted phase Weibull parameters - punctured insulator; one blue point
(positive half cycle) and one red point (negative half cycle) represent
the extracted features from one of the 342 datasets during the phase
Weibull fitting. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79

-x-
 LIST OF FIGURES

5.10 Fitted phase Weibull parameters - contaminated insulator; one blue

point (positive half cycle) and one red point (negative half cycle)
represent the extracted features from one of the 288 datasets during
the phase Weibull fitting. . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.11 Fitted amplitude Weibull parameters - normal insulator; one blue
point (lower magnitude discharge) and one red point (higher
magnitude discharge) represent the extracted features from one of
the 176 datasets during the amplitude Weibull fitting. . . . . . . . . . 81
5.12 Fitted amplitude Weibull parameters - punctured insulator; one
blue point (lower magnitude discharge) and one red point (higher
magnitude discharge) represent the extracted features from one of
the 342 datasets during the amplitude Weibull fitting. . . . . . . . . . 82
5.13 Fitted amplitude Weibull parameters - contaminated insulator; one
blue point (lower magnitude discharge) and one red point (higher
magnitude discharge) represent the extracted features from one of
the 288 datasets during the amplitude Weibull fitting. . . . . . . . . . 83
5.14 Fitted weighted Weibull parameters - normal insulator; one blue
point (positive half cycle) and one red point (negative half cycle)
represent the extracted features from one of the 176 datasets during
the weighted phase Weibull fitting. . . . . . . . . . . . . . . . . . . . . . 85
5.15 Fitted weighted Weibull parameters - punctured insulator; one blue
point (positive half cycle) and one red point (negative half cycle)
represent the extracted features from one of the 342 datasets during
the weighted phase Weibull fitting. . . . . . . . . . . . . . . . . . . . . . 86

- xi -
 LIST OF FIGURES

5.16 Fitted weighted Weibull parameters - contaminated insulator; one

blue point (positive half cycle) and one red point (negative half cycle)
represent the extracted features from one of the 288 datasets during
the weighted phase Weibull fitting. . . . . . . . . . . . . . . . . . . . . . 87
5.17 Line parameters of the spectral fractal analysis - normal insulator;
each point represents the average of the fitted line parameters to 512
segments in each of the 176 datasets. . . . . . . . . . . . . . . . . . . . . 89
5.18 Line parameters of the spectral fractal analysis - punctured insulator;
each point represents the average of the fitted line parameters to 512
segments in each of the 342 datasets. . . . . . . . . . . . . . . . . . . . . 90
5.19 Line parameters of the spectral fractal analysis - contaminated
insulator; each point represents the average of the fitted line
parameters to 512 segments in each of the 288 datasets. . . . . . . . . 91
5.20 Successful classification rate of contaminated and punctured
insulators using various combinations of extracted features. . . . . . . 94
5.21 Sample of 2D classification of contaminated and punctured insulators
using Combination ID 92. . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.22 Zoomed highlighted area of Figure 5.21. . . . . . . . . . . . . . . . . . . 96
5.23 Successful classification rate of punctured and normal insulators using
various combinations of extracted features. . . . . . . . . . . . . . . . . 98
5.24 Sample of 2D classification of punctured and normal insulators using
Combination ID 66. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 99
5.25 Zoomed highlighted area of Figure 5.24. . . . . . . . . . . . . . . . . . . 100
5.26 Successful classification rate of contaminated and normal insulators
using various combinations of extracted features. . . . . . . . . . . . . . 102

- xii -
 LIST OF FIGURES

5.27 Sample of 2D classification of contaminated and normal insulators

using Combination ID 47. . . . . . . . . . . . . . . . . . . . . . . . . . . . 103
5.28 Zoomed highlighted area of Figure 5.27. . . . . . . . . . . . . . . . . . . 104
5.29 Normal discharge activities near the metal fitting on a punctured
insulator; corona camera on left and normal camera on right. . . . . . 106
5.30 Discharge through the puncture path of a punctured insulator; corona
camera on left and normal camera on right. . . . . . . . . . . . . . . . . 106
5.31 Successful classification rate of punctured and contaminated
insulators excluding the overlapped normal discharge activities using
various combinations of extracted features. . . . . . . . . . . . . . . . . 108
5.32 2D Classification of punctured and contaminated insulators using
Combination ID 92. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
5.33 Eliminated sets of features during the gating algorithm. . . . . . . . . 109
5.34 Four sample Weibull distribution functions. . . . . . . . . . . . . . . . . 112
5.35 Four clusters in 2D classification using Domain ID 92. . . . . . . . . . 112

A.1 Voids inside the insulation system . . . . . . . . . . . . . . . . . . . . . . 122

A.2 Circuit model for a void inside the insulation system . . . . . . . . . . 125
A.3 Voltage waveform across an ideal cavity with Er = 0 and Ea = 1.5 Eb . 128
A.4 Voltage waveform across an ideal cavity with Er = 0 and Ea = 3 Eb . . 129
A.5 Experimental test setup for studying the voltage waveform of
dielectric-metallic electrode . . . . . . . . . . . . . . . . . . . . . . . . . 130
A.6 Circuit model for the experimental test shown in Figure A.5 . . . . . . 131
A.7 Experimental test setup for studying the voltage waveform of
dielectric-metallic electrode . . . . . . . . . . . . . . . . . . . . . . . . . . 133

- xiii -
1

Synopsis

1.1 Introduction

With the emergence of new technologies and higher demand for the reliable
transmission of electric power, utilities are facing economical and technical challenges
to reduce the duration of line outages and keeping their lines energized that make
live line maintenance services a more desirable option. In any power network, the
condition assessment of line insulators is one of the most important maintenance
services. Physical defects, such as punctures and contamination, can degrade the
insulators performance. Line insulators should be checked regularly to ensure
maximum service reliability. However, due to the large number of insulators on
the transmission and distribution network, this is a daunting task. For example, in
North America there are more than 150 million porcelain insulators in service where
some are 50 years old [1].
A number of inspection methods ranging from close contact tools to acoustic,
ultraviolet, and infrared methods have been proposed by industrial companies and

-1-
 1.2 Problem Statement

scientists for insulators condition assessment. There has always been a trade-off
between the reliability of these inspection tools and how convenient they are for
linemen to use. A comparison of the successful rate of detection of faulty insulators
for different tools shows that the inspection methods which require close contact
to the insulators are far more accurate compared to remote inspection methods.
However, remote inspection methods can assess the insulators condition from a
distance and without the need to climb towers. Considering the importance of
avoiding an outage and the convenience and safety of using remote assessment
tools, the remote inspection tools are a better and more desirable solution to the
maintenance of line insulators. Moreover, the recent progress made in digital signal
processing has significantly increased the reliability and successful detection rate
of the remote inspection methods. Remote detection methods need to capture a
symptom of faulty insulators without making physical contact. These methods
will monitor the variations of certain parameters depending on the nature of their
detection algorithm. For example, acoustic methods monitor the sound wave
parameters originating from faults on the insulators or infrared detection tools focus
on the thermal distribution along a string of insulators. Each inspection method
has its own advantages and drawbacks which are discussed in Chapter 3.

1.2 Problem Statement

The motivation for this PhD thesis is the need of utility companies and the power
system industry to employ innovative methods for the remote detection of defective
porcelain insulators. The objective of the research was not only to detect the
abnormal signature of a defective insulator, but also to distinguish and classify

-2-
 1.3 Objectives

the source of the behaviour, a novel approach for condition assessment of outdoor
porcelain insulators. Therefore, the main focus of my research was to remotely
inspect and classify defective porcelain line insulators in a lab environment, taking
into consideration the two most common types of defects on porcelain insulators;
puncture and contamination [2, 3, 4].

1.3 Objectives

The main objective of this PhD thesis is to study and characterize the
electromagnetic radiation signature of partial discharges from defective insulators.
Defective insulators had to be made in the lab according to the available standards.
Due to the randomness of partial discharge behaviour, a time domain statistical
analysis platform was developed to separate the radiation signature of different
defects. The research was conducted inside a laboratory using a clean fog chamber
that was built to duplicate a realistic environment for testing the insulators.
At the High Voltage Laboratory of the University of Manitoba, punctured and
contaminated insulators were made and tested. Obtaining accurate and reliable
measurements was one of the most important aspects of this research. To analyse
the raw measurement data obtained in the lab, a MATLAB graphical user interface
(GUI) was developed to calculate statistical features of captured partial discharge
activities. A classifier was also developed to differentiate and classify the radiation
signature of contaminated and punctured insulators.

-3-
 1.4 Research Outcomes and Contributions

1.4 Research Outcomes and Contributions

The published literature related to the topic of this research have reported similar
efforts on classification of partial discharges captured from contaminated insulators
[5, 6, 7, 8]. However, most of the methods for capturing the partial discharges are
based on leakage current waveform taken from the ground lead. Also, the signal
processing and classification algorithm reported in each research is different, but
it mainly involves time-frequency algorithms. One of novel ideas presented in this
research is the use of electric field sensors to remotely capture the electromagnetic
radiations, a technique never used for PD detection of line insulators. Also, although
performing statistical analysis on PD pulses have been reported in the literature,
the mixture Weibull distribution developed as part of signal processing of this
research has not been used before. Moreover, research on PD signature of realistic
samples of punctured insulators (created in the High Voltage lab) have never been
reported in this extent before. Another advantage of this research compared to
the published literature is the partial discharge database developed throughout
the lab measurements. During this project, a multitude of tests were conducted
and resulted in an excessive amount of measured data for a meaningful statistical
analysis. A total of 410,000 cycles of radiated partial discharge activities from
normal (clean and healthy), punctured, and contaminated insulators were captured
and recorded. Further, methodologies and procedures were developed to perform
remote measurements of insulators electromagnetic radiation.
A MATLAB based GUI has been developed that performs statistical analysis on
recorded measurements and classifies recorded signatures. The classifier uses unique
statistical features (output of the GUI) to separate the signatures of punctured

-4-
 1.4 Research Outcomes and Contributions

and contaminated insulators, considering the baseline signatures of healthy and

clean insulators. Based on a number of classifications performed in various
feature domains, the shape and scale parameter of a fitted Weibull distribution
in the negative half cycle of the voltage is proved to have distinct information to
differentiate partial discharge activities on contaminated and punctured insulators.
This novel remote assessment tool utilizes unique statistical features of partial
discharge activities to differentiate the type of defects on porcelain insulators. The
results obtained from the measurements and analysis of samples tested in a high
voltage laboratory, indicate the successful use of feature domains to classify radiation
signatures. This innovative remote assessment tool, using statistical processing
kernels, can help utilities rank the condition of porcelain insulators in service and
determine the proper maintenance program, based on the remote assessment results.
The outcomes of this research have been published in a conference paper and
one journal paper [9, 10]. A second journal paper is in preparation that will be
submitted soon.

-5-
2

Outdoor Insulators

2.1 Introduction

Overhead transmission lines are responsible for delivering electric power from
generators to industrial and residential customers. The energized conductors on
the overhead transmission lines have to be attached to the tower while they are
electrically insulated from the grounded towers. The device that can hold the
weight of the conductors while it provides electrical insulation between the line
and the tower is called an insulator. Insulators are able to handle electrical and
mechanical stresses when the line is energized. Most of the insulators have outdoor
applications where they are subjected to environmental conditions such as moisture,
high temperature, contamination, and icing. To choose the right type of insulator
before the installation, the environmental conditions of the site has to be studied.

-6-
 2.2 Insulator Types

2.2 Insulator Types

Insulators are mainly composed of a dielectric material, electrodes, and end fittings.
Outdoor insulators can be classified either by the dielectric material that is used in
them or by their application. By application, they can be divided into four groups;
suspension, post, interphase, and bushing [11, 12]. In this research, we mainly focus
on suspension outdoor insulators. Insulators can also be classified by the type of
dielectric material that has been used in them. The three main dielectric materials
that are mostly used in outdoor insulators are polymer, glass, and porcelain. In
some documents, these insulators might be called by other names; for example,
porcelain insulators are also known as ceramic insulators or polymer insulators are
also known as non-ceramic or composite insulators. Polymer insulators have recently
become very popular in high voltage transmission line applications because of their
lower weight, easier installation, higher resistance to vandalism and their excellent
performance in polluted environments. Surveys also show that the maintenance
costs for non-ceramic insulators are lower [13]. The manufacturing technique for
these insulators has a minimum amount of waste and sheds can be produced using
a single mold. However, polymer insulators have not been in the service for a long
time and therefore, it is not clear how they would perform after being used for
more than 50 years, whereas toughened glass insulators have been in service since
1947. Around 10 million of them have been reported in service in North America.
Inspection of these insulators is quite easy. Any defect on the insulator string will
cause the glass shell to shatter and only the stub remains on the string. Although
simple detection of faulty glass insulators makes them a good option for 500 kV
transmission line, they are quite heavy and also a very good target to vandal.

-7-
 2.3 Design and Manufacturing

Porcelain insulators have been in the market since 1910. More than 150 million
of these insulators are now in service in North America. They are by far the
most commonly used outdoor insulator in service. Apart from the problems due to
vandalism or poor quality control from some manufacturers, they have served very
well in high voltage transmission systems for the past 50-60 years. For maintenance
purposes, these insulators have to be checked from a close range which involves
either taking them down from the line (random sampling) or climbing the tower to
attach the measurement probes to the cap ends. The weight of porcelain insulators
used to be a problem, but the newer generations do not suffer from the weight issue
anymore [14].
The main focus of this research is on porcelain suspension insulators since
Manitoba Hydro, the industrial sponsor of this research, has a large number of
these insulators currently in service.

2.3 Design and Manufacturing

Pin type insulators and cap-and-pin type insulators are the two most commonly used
suspension porcelain insulators in service. The main challenge in manufacturing
these insulators is the cement connection between the metal pin and porcelain.
Most of the failures of porcelain insulators are due to cracks and punctures caused
by expansion of the cement joint [15]. Depending on the line voltage, the porcelain
composition could be different. In North America, a combination of Quartz, ball
clays, and silica is used in lower voltages such as distribution networks. Alumina
on the other hand, which is high-strength porcelain, is used in manufacturing
of porcelain insulators in transmission line networks. Glazing is the last step in

-8-
 2.4 Electrical and Mechanical Performance

manufacturing porcelain insulators which adds more features such as high strength
and self-cleaning ability to the insulators. The type of cement used in gluing
the metal pin to porcelain is an important parameter in manufacturing porcelain
insulators. Portland cement is the most common type of cement that is used in
manufacturing suspension porcelain insulators. The cement should be designed to
handle the mechanical stress as well as the electrical stress. Figure 2.1 shows the
cross section of a typical porcelain suspension insulator [12].

Figure 2.1: Cross section of a porcelain insulator.

2.4 Electrical and Mechanical Performance

Electrical properties of insulators come from the dielectric that they are made of. In
order to study the electrical properties of porcelain insulators, dielectric properties of
porcelain should be studied. It should be noted that there are two dielectric property
measures; volume dielectric property and surface dielectric property. When there is
a puncture or void inside the porcelain, the dielectric strength of the insulator drops.
The electric field in that area will be magnified and partial discharge will take place

-9-
 2.5 Standard Quality Tests

in that region which may eventually cause a failure within the insulator (Appendix A
provides more details on this topic) [16]. Surface dielectric property relates to the
conductivity of the surface of the shells. Under humid and contaminated conditions,
the surface conductivity rises which may cause a discharge arc on the surface of the
shell. These failures are called surface flashovers which are the results of the air
breakdown around the shells [17, 18].
The mechanical performance of insulators on the overhead transmission lines is
as important as the electrical performance. The conductors will fall if the insulators
fail mechanically resulting in disastrous consequences. The mechanical properties of
insulators, such as internal attachment of the metal pins to the dielectric, are mainly
concerned during the manufacturing process. The electrical properties of porcelain
insulators will be the main focus of our work.

2.5 Standard Quality Tests

Before installing porcelain insulators, they have to pass standard tests to ensure their
good performance in service. In order to perform these tests in lab environment,
the conditions inside the lab should be similar to the actual outdoor environment.
Electrical (and even mechanical) sources and monitoring instruments are also needed
to perform the tests. In order to get consistent test results for a certain insulator in
different labs, standard tests have been developed. Standard tests can be divided
into two main categories; sample tests and type tests. For sample tests, random
samples of insulators are selected and tested to verify the quality of the material
and manufacturing process consistency. Type tests are usually performed on new
products, either changing the design or the materials. These tests are aimed to

- 10 -
 2.6 Failures of Porcelain Insulators

identify the main characteristics of the insulators for each unique design. For most
of the sample and type tests, it is more desirable to have both the mechanical and
electrical stress on the insulator to duplicate the real operating conditions.
Routine tests, which are one of the most important transmission line maintenance
tests, guarantee the performance of insulators during the service. Some of these tests
are performed at a higher voltage (ie. puncture tests are performed at 500 % of the
rated voltage)than transmission line’s rated voltage to guarantee the performance
of insulators under short overvoltage periods such as lightening, surges, or icing
[19]. Contamination tests will determine the performance of the insulators in heavy
contaminated environments. Contamination tests conducted inside a fog chamber
will produce more realistic results as they can resemble the environmental conditions
in real environment[20].

2.6 Failures of Porcelain Insulators

Porcelain insulators can fail due to a variety of reasons. Prior to the installation,
some insulators might get physically damaged during shipping and transportation.
Vandalism is another reason of porcelain insulators failure. However, glass insulators
are by far a better target when it comes to vandal. Gun shots break the surface of
porcelain shells which increases the chance of flashovers under wet conditions. Poor
manufacturing and low quality control are two other reasons of porcelain insulators
failures.
The expansion of Portland cement used in cap-and-pin insulators causes radial
cracks on insulator’s shell. The expansion mostly happens under wet condition
when the cement absorbs the moisture. Cement expansion decreases the strength

- 11 -
 2.7 Contamination of Insulators

of porcelain and increases the chance of surface cracks [21].

Punctures can develop inside the porcelain insulators when there is an
overvoltage due to lightning on the transmission line. Puncture paths mostly develop
inside the cap and in the area between the cement connections of cap and pin to
the porcelain. Cracks develop over time inside porcelain shells when the quality of
manufacturing is poor. During the manufacturing, there is a layer called ‘Bitumen
layer’ which is laid on top of the cement joints. When the quality of this layer is
poor, it will dry out over time and causes internal cracks inside the porcelain [21, 22].
Continuous corona is another reason of porcelain failure. Corona usually happens
between the cap and porcelain or pin and cement. In both cases, over time, the
strength of dielectric drops and it eventually breaks [23]. Contamination is another
reason of porcelain failures. Heavy contamination under moist conditions will
increases the leakage current which will ultimately result in hardware erosion. High
leakage current can also cause flashovers [5, 8, 24]. More details on contaminated
insulators will be presented in the following section.

2.7 Contamination of Insulators

Contaminated insulators under humid and moist conditions can cause contamination
flashover and eventually system outages. Some of the contaminants that can be
found on the polluted insulators are salt from the sea, dust and rubber particles from
the nearby highways, bird secretions, desert sand, and industrial pollution from the
factories. Type of contaminants near the sea is different from contaminants in desert
or near industrial sites. Contaminants are mostly salt (NaCl), Gypsum (CaSO4 )
and some type of sand (SiO2 ). The speed of accumulation of these contaminants

- 12 -
 2.7 Contamination of Insulators

on the surface of insulators depends on speed of the wind, direction of the wind,
type of insulators, and transmission lines orientation. Over time, the contaminants
will accumulate on the surface of the insulators, however, they also get washed off
by rain and heavy winds. The top surface of insulators is usually cleaner than
the bottom ribs. The bottom ribs on the vertical oriented insulators (known as I
formation) accumulate more contaminant compared to the horizontal and V oriented
insulators. The contaminants on the insulators will become conductive when they
are exposed to moisture. When the contamination layer becomes conductive, the
magnitude of the leakage current through the string increases [25]. Contamination
level is one of the key parameters in insulator design. Insulators should be tested
for performance under contamination prior to installation.
The main problem under heavy contamination conditions is the contamination
flashover [26]. Dry band arcs form over the wetted and contaminated surfaces,
increasing the total surface conductivity and eventually the chance of contamination
flashover. Details on formation of dry bands will be presented in next section. To
standardize the contamination level, the ESSD (equivalent salt deposit density)
parameter is defined. According to [20] , “ESDD is the amount of sodium
chloride that, when dissolved, gives the same conductance as that of the natural
deposit removed from a given surface of the insulator divided by the area of
this surface; generally expressed in mg/cm2 .” Therefore, ESDD can be used to
classify the contamination level of insulators. Although soluble contaminants have
a higher effect on the performance of insulators, in order to precisely classify the
contamination level, non-soluble contaminants should also be taken into account.
Non-Soluble Deposit Density (NSSD) is a measure for evaluating the level of non-
soluble contaminants like sand and dirt. To precisely study the contamination profile

- 13 -
 2.8 Formation of Dry-Band Arcs

of installation sites, ESDD and NSDD paramters should be taken into account [27].
However, the impact of NSSD is very small compared to ESDD.
According to IEC standards, contamination tests inside the lab can be conducted
in three ways; salt-fog tests, wet-contaminant tests, or clean fog test. For the salt fog
tests, the voltage is kept constant during the whole test and the string of insulators
is exposed to a salty fog condition where its salinity can be anywhere between 2.5
kg/m3 and 224 kg/m3 depending on the desired contamination severity. During
the wet-contaminant test, the insulators are sprayed with the polluted suspension
and the voltage is applied to the string shortly after the contamination process is
over, while the insulators are still wet. The voltage is then raised until the flashover
occurs. In case of a clean fog test, the insulators will first go under the contamination
process and will be put aside to dry. The string will be hung inside a clean fog
chamber and the voltage will be applied on the wetted contamination layer. This
technique resembles the actual environmental conditions much better than the other
test methods [27].

2.8 Formation of Dry-Band Arcs

One of the most important stages of contamination flashover is the formation of

dry-band arcs. There have been several studies to investigate the formation and
behavior of dry bands [28]. The leakage current starts flowing through the string
of insulators when the polluted surface gets exposed to moisture. The behavior
of this current depends on the how this conductive electrolyte behaves under high
voltage stress. The electrical properties of a common electrolyte solution, such as
NaCl solution, depend on the concentration of ions, ion charges, and the mobility of

- 14 -
 2.8 Formation of Dry-Band Arcs

the ions. The viscosity of water drops quickly when the temperature goes up which
speeds up the movement of water ions [14]. High electric field will also make the ions
move faster. Therefore, the conductivity of NaCl solution on the surface of insulators
under the high voltage tends to increase. The increase in the leakage current under
the high electric field stresses will increase the temperature and eventually will cause
the water on the surface of insulators to evaporate, forming high surface resistivity
areas known as Dry Bands. Dry bands will then experience recurring discharge
activities where the surface electrical properties of both the dried and wet parts
determine the magnitude of the leakage current [14, 29]. It should be noted that the
sinusoidal part of the leakage current has more contribution to the water evaporation
compared to the arcs and small bursts [14]. In order to investigate this behavior
in one of the published studies [29], pollution layers were deposited on strips of
glass while the high voltage was applied across the strip. The pollution layer was
wetted inside a clean fog chamber and the voltage was measured using coupling
capacitors attached to the structure. Figure 2.2 shows the steps of this observation.
After wetting the strip, the dry bands appear on the surface until one dry band
dominates (Figure 2.2.c). The arcs appear on this dry band and extend throughout
the strip until they completely bridge the strip (Figure 2.2.f). Not all the arcs can
extend to the whole strip and completely bridge the strip. Survival of the arcs from
extinguishing to propagate along the strip, requires a minimum electric field which
is calculated and formulated in [14].
The leakage current can be calculated knowing the voltage drop across the dry
bands and the resistance of the wet polluted layer in series with it. Not all of the
initial dry bands will survive to spark over. If the resistance of the band is too high,
the arcs will extinguish. Further wetting will drop the resistance of dry bands and

- 15 -
 2.9 Summary

(a) (b)

(c) (d)

(e) (f)
Figure 2.2: (a) Wetting the pollution layer
(b) Formation of dry bands
(c) Dominance of one dry band
(d) Sparks on the dry band
(e) Extension of arc on the surface
(f) Complete propagation of the arc along the strip (flashover)

eventually can increase the length of dry bands. When the surface contamination is
high, the resistance drops significantly after wetting and therefore the arcs start to
grow and bridge the insulation. In case of low surface contamination, the resistance
of dry bands is quite high and local arcs continue to happen but they will extinguish
eventually. The results of the study presented in [28] proves that dry contamination
is harmless to the insulators. However, light contamination under wet conditions
can cause continuous arcing and eventually insulation degradation and aging.

2.9 Summary

In this chapter, a brief description of transmission line insulators with their

application in power networks was presented. Various types of insulators as well as
advantages and drawbacks in using each type of insulator was explained. Because of
a large number of porcelain insulators that are currently in service in Manitoba and

- 16 -
 2.9 Summary

North America, this thesis focuses on condition assessment of porcelain insulators

only. A short background on failure mechanism of porcelain insulators was also
presented. Also the flashover mechanism due to surface contamination on the
porcelain insulators was briefly discussed.

- 17 -
3

Condition Assessment of Porcelain

Insulators–A Literature Review

3.1 Introduction

Enhancing reliability of power networks has recently become a very serious challenge
for power companies. Considering the continuous electricity demand mainly due to
manufacturing industries, network failures and power outages due to equipment
failure can be very costly for power utilities [30, 31]. On the other hand, all the
equipment installed in power networks have a limited life span and they need to be
replaced before they fail in service. That is why condition assessment of all high
voltage equipment, including high voltage insulators, is a very important regular
maintenance in power networks. In case of transmission line insulators, defective
units have to be identified and replaced before they fail. Among different types of
insulators, detecting defective toughened glass insulators is very easy as the glass
shell shatters when it becomes defective. In case of porcelain insulators however, the

- 18 -
 3.1 Introduction

defects cannot be identified simply by performing visual inspections. These defects

can be developed either before or after the installation. The defects developed prior
to installation are mostly caused during shipping, handling, and transportation.
Cracks, voids, and flaws inside porcelain shells are defects that are invisible to
visual monitoring and can cause failures after the insulators are installed [12]. Many
inspection methods to identify faulty porcelain insulators have been reported in the
literature. These methods can be classified based on their application. One way
to categorize these methods is based on the symptom they detect from a faulty
unit For example, thermal inspection methods identify the defective insulators by
analyzing the heat distribution along a string of insulators. In case of acoustic
detection methods, faulty units are detected based on the sound propagation and
reflections from the insulators. Cracks inside the shells change the reflection and
propagation of the incident sound wave.
Considering the electrical properties of defective porcelain insulators, they
have lower resistance, can undertake lower voltage, will let a higher current to
pass through them, and have a distorted electric field distribution along the
string. Detection methods which use these symptoms to identify a faulty unit are
categorized as electrical detection methods. The Buzz method, using DC resistance
or electric field meters, and leakage current measurement techniques are examples
of this method. These detection methods either need a close physical contact to
the insulators or they can remotely inspect them. Methods which need a physical
contact of the measuring instrument to the insulators are quite dangerous especially
when a live maintenance is needed. Although remote detection methods are very
convenient and safe for linemen to use, the remote detection methods that have been
proposed so far are not reliable enough to substitute the widely used close contact

- 19 -
 3.2 Detection of Defective Insulators

techniques such as DC impedance meters [4, 32].

3.2 Detection of Defective Insulators

There are different detections methods to identify defective insulators. Each method
has advantages and drawbacks. Depending on the application, one might do
better than the other. In this section, a brief summary of the most commonly
used inspection tools for condition assessment of porcelain line insulators will be
presented.

3.2.1 Close Contact Methods

One of the most commonly used techniques by linemen to identify faulty insulators
is inspecting each insulator shell by connecting a resistance measuring instrument
across the shell. The two popular instruments that are developed based on this
concept are the Buzz meter and the DC resistance meter. This method of inspection
is considered to be a quite low-tech detection tool for linemen to use. As shown
in Figure 3.1, in the Buzz method, two sharp pins on the inspection rod will be
connected across each shell. If the shell is not defective, a breakdown will occur
between the pins because of the high voltage across the pins and a Buzz noise can
be heard. But if the shell is damaged, there would be no significant voltage drop
across the pins and the breakdown will not happen. This method can only be used
for porcelain insulators as they have the metallic cap and pin that the rod can
connect to [12].
The resistance of the shells should ideally be very high if they are not
damaged [33]. When they become defective, the resistance drops and the current

- 20 -
 3.2 Detection of Defective Insulators

Figure 3.1: Detection of defective insulators using the Buzz method.

starts flowing through them. Another simple method to inspect porcelain insulators
is to connect a DC resistance meter to the string and measure the resistance of
each shell. Defective shells show a lower resistance on the meter compared to the
healthy shells. This method is also very effective in detecting internal conductive
paths inside the insulating material. However, using this inspection tool on live lines
is not as convenient as the remote detection methods. Moreover, the low resistance
can also be a result of high contamination on the surface of insulators and the DC
resistance meter is not able to determine the possible reasons of the low resistance
readings [34].

3.2.2 Acoustic Detection Methods

Detecting acoustic emissions from a faulty insulator is one of the inspection methods
than can be used to identify the defective insulators. Partial discharge activities on

- 21 -
 3.2 Detection of Defective Insulators

defective insulators make an audible noise as well as electromagnetic radiations in

the ultrasound band. These radiations can be captured by ultrasonic microphones.
When there are surface discharges on the insulators, electrons start to move more
rapidly which will ultimately create an acoustic wave. This acoustic wave travels
in the ultrasound band and its frequency is centred at 40 Khz [35]. Although
this method has been used in detection of faulty outdoor insulators, it has more
applications in detecting and locating the PD inside the transformers [36]. The main
difficulty with the acoustic PD detection on outdoor insulators is the relatively high
background noise levels near the live insulators in service.

3.2.3 Ultraviolet and Infrared Detection Methods

Partial discharges also emit light that can be captured and analyzed to assess
insulators conditions. There are commercially available UV cameras that can be
used for imaging and monitoring outdoor insulators. However, sun light and other
ultraviolet waves that exist during the day time will affect the performance of these
cameras. Because of these interferences during the day, inspection of insulators
using these cameras are mostly carried out during the night. Light emissions due to
partial discharge activities inside voids were the topic of one of the recent studies
[37]. It was shown that there is a good correlation between the light emissions and
the magnitude of partial discharge pulses. For the measurement setup, a 7 kV square
pulses was used as a source of electric stress and a wideband Photomultiplier tube
was used to capture the light emissions. The captured signals were fed to a 5 GS/sec
and 500 MHz oscilloscope and then transferred to the computer for further analysis.
Voids with different geometries were tested and the relation between the dimension

- 22 -
 3.2 Detection of Defective Insulators

of the voids and the corresponding light emissions was studied. In another study on
non-ceramic insulators [38], UV cameras were utilized to identify and locate different
types of corona. Corona can be present at the live end fittings especially when the
corona ring is not used. Newer UV cameras can detect the light emissions even
during the day. These cameras can detect certain frequency components that are
still below the frequency of natural daylight. They are capable of filtering daylight
emissions and also enhancing the image resolution.
Infrared cameras are capable of detecting the temperature distribution on the
insulators. Insulators temperature can rise due to corona, leakage current on the
surface, arcs on the dry-bands, and even punctures inside the insulation. In one of
the latest studies on detection of faulty insulators on live lines, Infrared cameras
were used to double check the results obtained from acoustic receivers [39]. The
pictures taken by the camera still need to be interpreted by trained personnel in
order to evaluate the condition of insulators. The camera is also very sensitive to
the environmental condition such as local weather, season, and illumination. In
order to have an accurate reading from the camera, the maximum and minimum
temperature should be entered manually into the camera considering the ambient
temperature. The pictures can also be transferred to a computer to detect high
temperature spots and further analysis.
In summary, Although the pictures taken by these cameras contain valuable
information about insulators condition, a successful assessment of their condition
requires additional testing and investigation.

- 23 -
 3.2 Detection of Defective Insulators

3.2.4 Leakage Current Analysis

Monitoring the leakage current waveform can reveal useful information about
insulators condition, especially in evaluating the contamination level of insulators.
One way to classify the contamination severity of porcelain insulators is using the
ESDD method which will be discussed in more detail in Section 4.2.6.1. However
in order to determine the contamination level of an insulator using ESDD method,
the unit has to be taken down to the lab for the measurements. In other words, this
method cannot be used for live line maintenance purposes. The alternative method
for measuring the contamination severity of insulators is monitoring the leakage
current. There are quite a few parameters in the leakage current pulse that can be
monitored in order to prevent a contamination flashover. Maximum value of the
leakage current, number of pulses and location of the highest peak of the pulses are
examples of these parameters [40, 41]. To properly study the leakage current pulses,
time domain and frequency domain characteristics of the waveform should be taken
into account. The leakage current can go through three stages before the actual
flashover takes place. These three stages are security, forecast, and danger stages.
The leakage current is much larger during the forecast and danger stages compared
to the security stage. However, the duration of the pulses are shorter in those stages.
If the possibility of a contamination flashover is detected in early stages, there would
be enough time to avoid it. Most of the research on leakage current monitoring is
focused on time domain analysis of the pulses. Some recent studies have been focused
on the frequency domain parameters and how monitoring other harmonics can lead
to an early diagnosis of a contamination flashover [42]. The effect of RTV (room
temperature vulcanization) silicone coating on the long term withstand performance

- 24 -
 3.2 Detection of Defective Insulators

of porcelain insulator been investigated in another study by monitoring the leakage

current [6]. The relation between the ESDD level of contaminated insulators and
the flashover voltage has also been studied while monitoring the leakage current [5].
Contamination along with the operating voltage and environmental conditions
are the three important factors in developing a flashover. Environmental factors are
mostly the temperature and relative humidity. However, the impact of humidity is
much higher than the temperature on the development of a flashover.

3.2.4.1 The Effect of Humidity

High ambient humidity and heavy surface contamination are the two most common
reasons for high leakage current. Properly distinguishing the source of this increase
in leakage current can be quite challenging. High leakage current due to high ambient
humidity is a false positive output when using this method. In [24], to properly
distinguish the source of high leakage current, two tests were conducted. During
the first test, already contaminated insulators were kept inside a fog chamber and
the humidity inside the chamber was varied in order to study the effect of humidity.
During the second experiment, the contamination level of insulators was changed
throughout the tests where the humidity of the chamber was kept constant. The
results of these measurements were used to distinguish the effect of humidity and
contamination level on the leakage current waveform, reducing the number of false
alarms in detecting contaminated insulators.

3.2.4.2 Leakage Current Monitoring

To duplicate the realistic environmental conditions during contamination tests, a fog

chamber is required. The leakage current waveform can be monitored through the

- 25 -
 3.2 Detection of Defective Insulators

ground wire that connects the insulators to the ground. To take samples from
the ground wire, two methods can be used. The current signal can be either
measured from the voltage drop across a resistor in series with the ground wire or
can be measured using current sensors such as Rogowski coils. The contamination
distribution on the surface of insulators in the service is a function of different
parameters such as wind direction, elevation, and etc. For the lab tests, however,
a uniform distribution of the contamination layer on the surface of insulators is
assumed. To make artificially contaminated insulators, a salty suspension will be
sprayed on the insulators. Details of this procedure will be discussed in Section
4.2.6.1.

3.2.4.3 Frequency Domain Analysis of Leakage Current

Frequency domain analysis of leakage current is basically monitoring the power of

different harmonics. It has been shown that by increasing the relative humidity
and contamination severity, the amplitude of the 60 Hz component of the leakage
current has the most significant increase. Also the amplitude of the third and fifth
harmonics change significantly when the contamination is high [43]. In another
study [7], the leakage current waveform is classified into six stages before the
occurrence of flashover. The classification is based on the frequency spectrum of
the leakage current waveform. The experiments showed that the insulators with
heavy contamination will ultimately have a flashover. On a heavy contaminated
insulator, the leakage current waveform was captured and monitored [7]. In time, the
leakage current waveform shows some distortions compared to the smooth sinusoidal
waveform that was captured at the beginning. After 12 minutes of testing the
heavy contaminated insulator, the leakage current waveform started experiencing

- 26 -
 3.2 Detection of Defective Insulators

all the six stages starting from a smooth and low amplitude sinusoidal waveform to
a distorted waveform with relatively high power harmonics [24]. It was reported that
in all cases that contamination flashover ultimately happened, the third harmonic
amplitude of the leakage current spectrum exceeded - 40 dBA. They also reported
that the information in the fifth and seventh harmonics cannot predict whether the
flashover is happening or not. This conclusion came from the observation of leakage
current power spectrum during the test where those two components remained
almost unchanged from the beginning of the test until the flashover occurred.

3.2.4.4 Phase Angle Analysis

Studying the time domain parameters of the leakage current waveform can reveal
important information about the contamination severity of the insulators. There
are numerous studies on the rise time, duration, amplitude, and other time-domain
parameters of the leakage current waveform [44, 45]. The phase angle between
the voltage and the current is one of the important time domain characteristic of
the leakage current waveform. During clean and dry conditions, applied voltage and
leakage current resemble an RC circuit where the current leads by almost 70 degrees.
This relation stays the same for the case of dry polluted insulators. But in case of
wet contaminated insulators, the phase difference is significantly lower (almost 30
degrees) and also the amplitude of the leakage current is higher [8]. The pollution
layer becomes conductive when it is exposed to the moisture and therefore the phase
angle decreases. Under heavy contamination the phase difference could be as low
as 5 degrees. This means that under wet contamination, the overall system acts
more like a resistive circuit. By just inspecting the phase difference, the clean dry
insulator can be distinguished from a wet contaminated insulator. But as soon as

- 27 -
 3.2 Detection of Defective Insulators

the phase angle drops due to wetting, it is very challenging to identify the reason of
further phase angle drop. In case of a dry polluted insulator, there are few stages
before the contamination flashover happens. The phase drop is significant when the
pollution layer gets exposed to the moisture, but during the development of short
dry-band arcs and the growth of those arcs to long arcs and even the flashover, the
drop is not that significant and therefore it is very hard to assess insulators condition
in those stages by just measuring the phase difference. The other difficulty in using
this method is differentiating humidity and the pollution as they both increase the
conductivity of the surface and result in a phase drop.

3.2.4.5 Time–Frequency Analysis

Wavelet is a time-frequency analysis tool that can be very useful in analyzing

the leakage current. The application of wavelet is not limited to the leakage
current and it can be used in a wide range of applications to analyze the captured
information. In [8], using a discrete wavelet transform (DWT) and multiresolution
signal decomposition (MRSD) techniques, the time-frequency characteristic of the
leakage current was studied. A standard single disc cap and pin porcelain insulator
was put under 11 kV rms, 50 Hz voltage source to assess the contamination level.
The insulator was hung vertically inside a 1.5m × 1.5m × 1.5m clean fog chamber.
The leakage current is sampled through a resistor in series with the ground wire.
The leakage current was then sampled and stored on a computer for further analysis.
The leakage current was decomposed into seven levels using wavelet functions. The
standard deviations and the mean of each decomposed signal in each level was
studied. The distortion ratio is defined as the ratio of standard deviation value of
detailed components # 3, 4 and 5 to fundamental component # 6. The distortion

- 28 -
 3.2 Detection of Defective Insulators

ratio increases considerably when the short arcs appear and it decreases when the
arcs start to grow into long arcs. Plots of standard deviation versus decomposition
level are presented in this work to explain the behavior of each detailed component
at different flashover stages. The fifth detailed component also changes quite
significantly during different stages which can be used towards stage identification.
In another recent study [46], a flat glass sheet was used to model an outdoor
insulator. This structure was proved to be a very good laboratory model for an
outdoor insulator. Two electrodes were mounted on a flat glass sheet to generate
electrical stress. The source is a 50 kVA/300 kV RMS, 50 Hz transformer. The
surface of the glass was sprayed with a salty solution and different salinities in order
to study various contamination levels. Frequency domain tools were applied on
the sampled data in order to study the spectrum behavior of recorded data. For
analyzing the voltage waveform, however, a continuous wavelet transform and a
complex Morlet wavelet transform were used. The behavior of wavelet coefficients
in different contamination levels was studied.

3.2.5 Electric Field Measurement

Electric field distribution along the insulators string is one of the important
parameters that should be considered during insulators design. High electric
field stress on the string can cause audible noise, partial discharge activities,
and eventually deterioration of the dielectric material. Studying the electric field
distribution around the insulators can avoid the occurrence of flashovers. It has
been shown in [14] that when the surface electric field component is sufficiently
low, the flashover will not occur. Electric field distribution along the insulators

- 29 -
 3.2 Detection of Defective Insulators

string can also be used to detect defects on each unit. Instead of solving the
complicated structure of insulators string to obtain an analytical solution, computer
aided simulation software can be utilized [47]. Many computer aided simulation
projects based on different numerical methods such as, finite element method,
finite difference method, or boundary element method have been conducted to
calculate the electric field distribution and electric potential graphs along the string
of insulators [48, 49, 50, 51]. The modelled electric fields have also been compared
to the measurement data taken from electric fields along the insulators string in
service. Figure 3.2 shows a test setup for measuring the electric field of energized
suspension insulators. The DL-1 electric field sensor mounted on the top of a rod
has to get close to the energized insulators and capture the fields [52]. The captured
data will be transferred to a data logger using a coax cable or a fiber optic link for
further analysis. The electric field probe slides along the string and measures the
RMS electric field. Some simple methods just use the electric field along the axis on
the string to make the decisions of the insulators condition. However more advanced
methods take into the account both the radial electric field component(Er ) and the
electric field component along the axis of the string (Ez ). The electric field method
has been reported to have a very good accuracy in detecting conductive defects
[52]. The electric field is low near the shells close to the low voltage conductor
and starts rising smoothly when it gets closer to the shells near the high voltage
conductor. Having a defect in one of the shells, causes a sudden drop in the electric
field value close to that defect. Electric field graphs from the string with defects
can be compared to the graphs with a string of all healthy insulators to locate the
faulty shells. Differences in the graphs are more obvious when the defects are closer
to the high voltage conductor. It is harder to identify the defects close to the low

- 30 -
 3.2 Detection of Defective Insulators

voltage conductor. It should be also noted that when a corona ring is used at the
high voltage side, it will be very hard to identify even the defects close to the high
voltage conductor [47].

Grou
nd c
onne
ction

Porcelain Insulators
Field Meter

High
Volta EZ
ge
Er

Dat
a to
Log
ger
Transformer

Figure 3.2: Electric field measurement of porcelain insulators.

3.2.6 RF Radiation Detection Method

Partial discharges result from fast movement of electrons. The movement of electrons
not only creates the leakage current, but it also creates electromagnetic waves that
will radiate. Rapid changes in the current due to the fast movement of electrons
extend the spectrum of this waveform to a higher frequency region (around 200 MHz
[53]). Using proper antennas designed for the appropriate frequency range, partial
discharge radiated pulses can be captured. In one study [54], radiations from a
spark gap breakdown were captured using disk-cone antennas. An optimization on

- 31 -
 3.2 Detection of Defective Insulators

the location of antennas for receiving the best radiation was also performed. The
sharp risetime of partial discharge pulses can be easily detected using high frequency
sampling and time domain analysis. In another study [55], antenna arrays were
mounted on a vehicle to make a mobile RF emission detector. The signals received
through the antennas were fed into a 2.5 GS/sec oscilloscope with a bandwidth
of 1 GHz. The antenna arrays have a flat frequency response between 100 MHz
and 3 GHz. The oscilloscope has a segmented memory feature which allows storing
multiple PD pulses without storing the dead time in between them. To analyze the
received signals, two diagnosis methods were used; Bearing calculation and RMS
time delay error. Bearing calculation method calculates the angle of the incident
radiation using wave propagation constant and arrival time delay. When there are
just two antennas for receiving the radiations, it is not possible to know if the angle
is above the antenna plane or below. This issue can be resolved by using four
antennas. RMS time delay method locates the source of radiation more accurately.
One of the four antennas was considered to be the reference antenna and the time
delay for a signal to travel from the source to each antenna was defined as,

1√
Tn = (xn − xs )2 + (yn − ys )2 + (zn − zs )2 (3.1)
c

where (xs , ys , zs ) are the coordinates of the source and (xn , yn , zn ) with n ∈ {1, 2, 3, 4}
are the coordinates of the antennas. Knowing the reference antenna, the delay time
differences can be calculated as

Tn1 = Tn − T1 (3.2)

- 32 -
 3.3 Classification of Partial Discharge

The error function is also defined as,

√
(T21 − τ21 )2 − (T31 − τ31 )2 − (T41 − τ41 )2
s = (3.3)
3

where τn1 is the measured time delay between the antenna array n and 1. By
sweeping the location of the source and minimizing the error function, the defective
unit can be located. As for the field measurement, the researcher took the mobile RF
emission detector to a reportedly high RF radiation area near an electrified railway
and a 132 kV transmission line. The source of radiations was unknown. The vehicle
was parked in two different positions for one hour at each position and gathered 371
samples at each location. Six imaginary sources of radiation were assumed to be at
the location of each high voltage string of insulators on the transmission line. The
source of radiation was identified by minimizing the error function.

3.3 Classification of Partial Discharge

Modern insulation systems can have different types of insulations such as solid,
liquid, or gas insulating materials. Partial discharge in any of type of these insulating
materials can degrade the insulation performance and eventually cause system
failures. Understanding the partial discharge mechanism helps developing inspection
tools to detect and locate the source of partial discharges, one of the essential stages
in insulation system monitoring. Signal processing techniques play a big role in
this stage by finding unique behaviors in time domain, frequency domain or time-
frequency domain parameters of the captured data. A lot of effort has gone into
classifying the partial discharges based on the behavior of one or more of these
parameters. Studies have shown [56, 57, 58] that because of the randomness of a

- 33 -
 3.3 Classification of Partial Discharge

partial discharge process, statistical analysis can be a great tool for classification of
this random phenomenon. Identifying the source of PD pulses can be summarized
into a two stage procedure; extracting the unique information from the captured
PD pulses followed by classification and recognition of the source. A survey on
partial discharge classification [59] has put to together a wide range of classifiers,
feature extraction techniques and pattern recognition methods. Classifiers can be
categorized based on the decision function that they use. In the rest of this chapter,
classifiers with different decision functions will be briefly discussed.

3.3.1 Distance Function Classifiers

These PD classifiers have a reference library for each class of PD sources. For
each unknown measurement, a distance function will be defined and the unknown
measurement will be assigned to one of the library reference classes based on the
decision made by the distance function [60]. Two examples of these classifiers will
be discussed here.

3.3.1.1 Minimum Distance Classifier

This classifier calculates the distance between the vector of unknown PD

measurements and the library references [61]. For example, if there are M possible
PD sources, then P1 , P2 , ⋯, PM are predefined vectors corresponding to each class.
If x is an unknown PD measurement, the distance of x from the library reference i
can be defined as,

1 N
di = ∑ ∣xj − mij ∣ i = 1, 2, ⋯, M (3.4)
N j=1

- 34 -
 3.3 Classification of Partial Discharge

where N is the size of vector x. The classifier finds the lowest di and assigns x to
the ith library.

3.3.1.2 Polynomial Classifier

This classifier converts the measurement vector x to a single score value using a
decision function [62]. The decision function is made out of sets of polynomials.
The simplest classifier in this group is the linear classifier. The decision function for
its ith class is calculated as

di = wi0 + wi1 x1 + wi2 x2 + ⋯ + wiN xN (3.5)

where wij represents the coefficients of the polynomial in the decision function and
x is the measurement vector. The matrix representation of this classifier would be

D =W ×X (3.6)

where W is the weighting matrix with M rows and N + 1 columns, D is the decision
matrix and X is the matrix with all the measurement vectors. If there were L events
captured during the tests, then x has N + 1 rows and L columns. Each column in D
represents the output of the decision function for each of the measurement vectors
in x. If the decision making function is based on the maximum likelihood function,
then in order to assign specific vector x to class i,

dij (x) > dkj (x) for all k ≠ i (3.7)

- 35 -
 3.3 Classification of Partial Discharge

3.3.2 Statistical Classifiers

These classifiers assess the statistical behavior of the recorded PD measurements

and classify the unknown measurement into one of the previously known classes.
Generally they can be divided into two groups; Bayes classifiers and recognition
rate classifiers [63].

3.3.2.1 Bayes Classifier

The decision function for this classifier is based on minimizing the probability
of assigning the unknown measurement vector to the wrong class based on the
maximum likelihood criteria [62]. The training sequence will produce the mean
vector and the covariance matrix. Having these statistical properties from the
training data files, a multivariate density function will be defined to calculate the
probability of the unknown measurement vector x belonging to each of reference
library classes.

3.3.2.2 Recognition Rate Classifier

In this algorithm, some feature vectors are built from the statistical properties of
sampled measurements of already known PD sources [64]. For every known PD
source, the mean, standard deviation and the 95% confidence interval are calculated
from several samples that have been taken from that source. Depending on the
algorithm, other statistical properties can also contribute to the feature vector.
Then, for each unknown measurement, each of these statistical properties will be
calculated and the algorithm checks if the calculated value lies in the 95% confidence
interval of any of the library classes. If it does, one score will be given to that class.

- 36 -
 3.3 Classification of Partial Discharge

All the parameters get checked with the same technique and at the end, the class
with the highest score will be assigned to the unknown measurement vector.

3.3.3 Support Vector Machines (SVM)

In one of the latest studies [65], previously recorded partial discharge measurements
from known sources were used to evaluate the capability of the SVM technique
to differentiate PD sources. The SVM method is based on statistical learning
theory. To understand the fundamental principles of this technique in differentiating
different sources, an optimal hyperplane problem is shown in Figure 3.3.

Optimal hyperplane
= Class 1
= Class 2
= Support Vector Point

Marginal plane

Figure 3.3: Hyperplane problem to illustrate SVM methodology.

A hyperplane is introduced between two sets of separable data classes with

margin m. The points on the marginal plane form the support vector and the
optimal plane can be driven from support vector points. The optimal hyperplane
will be iteratively calculated using training sequences. Support vector points will
be stored after each training sequence and the rest of the points will be discarded.

- 37 -
 3.3 Classification of Partial Discharge

When the training is over, the hyperplane can be applied on unknown data classes to
identify each class. Classification of the data classes that are not linearly separable
can be quite challenging. Therefore, mapping the input data into a higher dimension
space can be a very effective solution to those problems.
For intelligent PD classification algorithms, feature extraction is very critical.
Feature extraction involves extracting the most valuable and unique behavior of
the parameters from the available information and then mapping them into higher
dimensional spaces. Principle component analysis (PCA) is another tool that can be
used in conjunction with this method to optimize the mapping. PCA can maximize
the linear separation of the data points after mapping. PCA produces a matrix with
the same dimension of the input matrix where the first two columns in that matrix
contain more than 80% of the total variance of data [66].

3.3.4 Power Spectrum Classifiers

A number of unique features of partial discharge pulses can be found in their power
spectrum. No matter where and how the PD samples are taken, power spectrum
analysis can reveal distinct properties of the PD source. In one of the latest studies
[24], the PD pulses sampled from the leakage current were analyzed using the power
spectrum analysis using the amplitude of the Fast Fourier Transform. The goal in
this research was to develop a pre-warning stage for monitoring outdoor insulators
when the contamination severity is high. Power spectrum estimation of the PD
pulses is carried out using parametric auto-regressive (AR) model. Calculation of
the estimated AR model is published in [67]. To compare the effect of humidity on
the power spectrum, insulators with the same contamination level are placed in the

- 38 -
 3.3 Classification of Partial Discharge

fog chamber with different humidity levels. To study the power spectrum behavior,
plots of the amplitude of the Fast Fourier Transform for each case was developed.
Also the harmonic levels in each of those graphs were measured to relate harmonic
powers to the contamination severity.

3.3.5 Neural Network and Fuzzy Logic Based Classifiers

The concept of neural networks has been successfully applied to many different PD
classification problems. The ability to learn from the previous sets of measurements
to improve their performance is a remarkable feature of these classifiers. The
performance and methodology of the classifiers in this group could vary based on
the way the neural network has been designed. More details on different types of
this classifier can be found in [68, 69, 70].
The feature vectors extracted from the measurement data are usually not clear
enough to be used for differentiating PD sources. Fuzzy logic classifiers try to map
the feature vectors into membership values of fuzzy sets which are defined over the
fuzzy domain. The decision making process consists of three steps; fuzzification
of the feature vector, inference engine and defuzzification. Although the nature
of these classifiers is the same, some classifiers have slightly different classification
methodology. For example, in one study [71], the feature vector was first mapped
into the membership values and then the probability of that feature vector belonging
to each class was calculated. The class with the highest probability will be assigned
to that feature vector.

- 39 -
 3.4 Data Processing Methodology of This Thesis

3.3.6 Performance Assessment of Classifiers

One of the first steps after collecting the PD samples from a faulty unit is to
decide whether to use the phase information or not. Some researchers believe that
classification can be done without using the phase data where others find it easier
to do the classification using the phase information. It is almost impossible to
compare the performance of all the proposed methods in the published literature.
The lab conditions, the source of PD pulses, the assumption made in each case
study, the setup configuration and many other parameters are different in each
study. However, the rate of successful recognition of unknown PD pulses according
to what authors have claimed is the only assessment that we can discuss here. A
70%-80% success rate was reported in [61] using the minimum distance and neural
network classifier where gas insulated switch gears were studied. There have also
been some successful rate of recognition using statistical classifiers in [64, 72]. In [57],
a two parameter Weibull probability density function was fitted to the amplitude of
recorded measurements from PD activities with successful results. In another study
[73], the feature vector was extracted from the PD measurements using different
signal processing methods. Among all the classifiers that were studied during their
work, the minimum distance classifier had the highest successful rate of recognition.

3.4 Data Processing Methodology of This Thesis

Based on the publications discussed earlier in this chapter, the classification methods
using the phase resolved data are reported to have very successful classification rates.
These classifications locate the PD activity on the 360 degree phase domain and build
a comprehensive reference library for each defect. Besides the publications available

- 40 -
 3.4 Data Processing Methodology of This Thesis

in the literature on advantages of using phase resolved data, the leading technology
in the industry for partial discharge detection during commissioning tests of other
high voltage equipment, such as underground cables, is also based on phase resolved
analysis [74, 75].
Because of the random nature of partial discharge pulses, statistical analysis
of PD pulses reveals valuable information on the source of discharges. During the
literature review of this thesis, publications with high successful classification rates
either developed statistical classifiers or distance classifiers using extracted statistical
features of PD pulses. Another advantage of using statistical features versus time
domain pulse characteristics of PD pulses is avoiding misleading and inaccurate
classification of PD sources based on one dominant PD pulse. The statistical features
derived from analyzing many partial discharge activities can identify the trend of
discharges and better categorize the source of discharges.
After performing a preliminary round of measurements and analysis during the
initial stages of this research [10], a number of different probability distribution
functions were studied to find the best statistical function that can represent
the trend of discharges. There has been previous research providing physical
justifications for the use of Weibull statistics in the assessment of dielectric
breakdown [76]. Keeping in mind the application of Weibull distributions in
describing the dielectric breakdown process and the successful preliminary results
on using Weibull distributions for data processing of the captured radiations, a
Weibull probability distribution function was selected as the main statistical feature
to represent the partial discharge behaviors. Weibull probability density functions
are known for describing and analysing the trend of partial discharges in other
publications as well [57, 58, 77, 78]. However, the weighted phase resolved Weibull

- 41 -
 3.4 Data Processing Methodology of This Thesis

distribution (Section 5.1.3) introduced in this thesis has never been reported for
analysing and classifying the PD sources.
It should be noted that there is a broad range of parameters that can be extracted
from the recorded measurements to be used by the classification algorithm. During
the early stages of developing the data processing algorithm, there were many other
features that were calculated and extracted from recorded measurement in order
to enhance the classification. However, many of these extracted features, such
as higher statistical moments, concurrent discharge behaviors, etc. did not show
distinct information about the PD sources of defective insulators and therefore were
discarded from the classification algorithm.
A number of publications [79, 80] have reported successful classification of PD
pulses using the fractal dimension. Other publications have reported unique features
in the spectrum of partial discharge pulses [24, 67, 81]. Combining the two concepts,
a frequency domain fractal feature was included in the data processing feature
extraction algorithm. This parameter was described based on the transform fractal
feature introduced in [82]. More details on calculating the frequency domain fractal
feature is presented in Section 5.1.4. This feature along with the statistical features
extracted from the PD activities (Section 5.1) are the main outputs of the feature
extracting algorithm.
After finalizing the selection of features that will be extracted from the PD pulses,
a proper classifier is required to classify the type of insulators based on the extracted
features of PD pulses. After extracting the features from all the normal, punctured,
and contaminated insulators, it was clear that a linear classifier will not be able to
separate the source of defects in any of the feature domains (Section 5.3.1 explains
the classification challenges in more detail). The classifier not only should be able

- 42 -
 3.5 Summary

to calculate non-linear decision boundaries in the feature domains, it also should be

able to account for mislabelled observations and maximize the margin of the decision
boundary (also known as soft margin capability). Amongst the classifiers presented
in Section 3.3, the SVM classier was the best solution for this application since it
had both of the required capabilities. The SVM classifiers can map the observation
points to higher dimension feature domains where the observation points are linearly
separable and then return the linear decision boundary in the higher domain into
the original feature domain where the linear decision boundary will be translated
into a polynomial in the feature domain (also known as kernel trick). The SVM
classifiers which are basically non-linear distance classifier with the capability of
classifying overlapped data points using the soft margin algorithms, have also been
successfully used in other PD classification problems [56]. In many classification
problems, such as the PD classification in this thesis (Section 5.2), there are many
overlapping feature points in the feature domains where the SVM classifier can find
a maximum margin hyperplane and minimize the number of mislabelled data points
by taking advantage of the polynomial kernel mapping feature.

3.5 Summary

In this chapter, a summary of available inspection methods for assessment of

defective porcelain insulators was presented. The difficulties and advantages
involved in using each method was discussed. The data processing involved
in classification of defective insulators includes extracting features with distinct
information about the source of defect and training a classifier to properly
differentiate the type of defects. A selection of classifiers that have been reported

- 43 -
 3.5 Summary

in publications for PD pattern recognition was also presented. Moreover, the

data processing techniques and the classifier selected for classification of defective
insulators in this thesis were described.
In a one of the recent publications in this field [83], failed insulators in service
were taken to the lab where the performance of four different types of inspection
tools on detecting the failed units were investigated. The study reports that out
of 460 failed units tested by all four inspecting tools, RF tools and Ultrasound
tools properly identified 69% and 41 % of the failed unit respectively. However, the
infrared and ultraviolet tools did not identify any of the failed units.
Other publications using ultrasound techniques for detecting defects on porcelain
insulators have reported separable time and frequency features on captured
sound waves [84], however no classification was performed. Similar research on
contaminated glass insulators using captured ultrasound emissions concludes the
indications in time and frequency domain that can be used for early detection of
contaminated insulators, however no classification was reported [85].

- 44 -
4

Experimental Setup

This chapter describes the lab measurement setup and data analysis. The initial
measurements were performed in the High Voltage Lab at the University of
Manitoba. The test procedure developed in this thesis also involves a number of IEC
standards that will be discussed in this chapter. A MATLAB [86] based graphical
user interface (GUI) was developed to perform the data analysis, feature extraction,
and classification on the recorded measurements. Details on the signal processing
analysis performed within the GUI interface will be also presented in this chapter.

4.1 IEC Standards Referenced in this Research

The test procedures performed in our study are mainly adopted from IEC, IEEE,
and CIGRE standards. Although IEEE and CIGRE standards have been used as a
reference for some of the tests, the main reference library for most of the tests was
adopted from IEC standard libraries. The testing procedures in this research are
mostly based on three IEC standards that will be discussed in more details here.

- 45 -
 4.1 IEC Standards Referenced in this Research

4.1.1 IEC 60507: “Artificial pollution tests on high-voltage

insulators to be used on a.c. systems”

The IEC 60507 [20] is focused on artificial pollution tests on insulators. It addresses
two different ways of making contaminated insulators; salt fog method and solid
layer method. The salt fog method requires a salt fog chamber which was not
available in High Voltage Lab at the University of Manitoba. However, the solid
layer method does not require a salt fog chamber and the procedure only involves
depositing the specified suspension on the insulators. The details of these methods
have been explained thoroughly in the standard. One of the most important steps
during testing the artificially polluted insulators made by the solid layer method
is properly wetting the insulators. The standard explains two possible techniques;
wetting before and during the energization or wetting just after the energization.
The procedure of washing off the contaminants is also described. Reference tables
explaining the relationship between the salinity, volume conductivity and density
of the solution are also available in the standard. The correction factors for
temperature variations as well as all the equations for calculating Equivalent Salt
Deposit Density (ESDD) of contaminated insulators are also presented in the
standard. The IEC 60507 standard was mainly used in creating contaminated
insulators in the lab and calculating their contamination level (Section 4.2.6.1).

4.1.2 IEC 60815: “Selection and dimensioning of high-

voltage insulators for polluted conditions”

The IEC 60815 [27] is mainly focused on evaluating the contamination severity at
installation sites. The standard also provides guidelines and recommendations on

- 46 -
 4.1 IEC Standards Referenced in this Research

selecting proper insulators knowing the site contamination profile. For final selection
of the insulators, the insulators should be tested for contamination withstand. The
insulators with the best results will be selected for the installation. It should be noted
that there have been some inconsistencies in the relation between the ESDD and
site severity in different standard definitions. Table 4.1 shows different definitions
based on IEEE, IEC and CIGRE Standards [27, 87, 88].
Table 4.1: Comparing available standards for site severity definitions based on the
ESDD level.

ESDD (mg/cm2 )
Site Severity
IEC IEEE CIGRE
Very Light – 0 – 0.03 0.015 – 0.03
Light 0.03 – 0.06 0.03 – 0.06 0.03 – 0.06
Medium 0.1 – 0.2 0.06 – 0.1 0.06 – 0.12
Heavy 0.3 – 0.6 > 0.1 0.12 – 0.24
Very heavy – – 0.24 – 0.48

In our study, we followed the IEEE definition for labelling site severity levels [87].
Although labelling the site contamination severity has been reported differently in
these standards, the definition of ESDD is the same in all of them. The definition of
contamination severity and the standard procedures for measuring it was adopted
and used in this thesis (Section 4.2.6.1).

4.1.3 IEC 60383-1: “Insulators for overhead lines with a

nominal voltage above 1000 V”

The IEC 60383-1 [89] mainly focuses on performing standard tests on AC high
voltage porcelain and glass insulators. The details of different tests such as type
tests, sample tests, and routine tests are described and the verification criteria
are given. For electrical tests, the standard atmospheric conditions as well as

- 47 -
 4.2 Lab Test Setup

the correction factors for different atmospheric conditions are provided. Test
procedures for lightning impulse testing and the acceptance criteria are described.
Details on the puncture withstand test on porcelain insulators are also presented
in this standard which is particularly of interest to this research. According to
the puncture withstand test, the insulators should be immersed in a tank full
of liquid insulation material when the high voltage is rapidly applied across the
insulators. The procedure for electromechanical failing load test is also described.
The characteristics of the load, the distribution of voltage stress on the string
and the acceptance criteria are presented. Few other test procedures such as
thermal-mechanical performance test and temperature cycle tests are also available
in this standard. The IEC 60383-1 standard was mainly used in creating punctured
insulators in the lab based on the puncture withstand test described in the standard
(Section 4.2.6.2).

4.2 Lab Test Setup

Figure 4.1 shows the main equipment arrangement for the test setup. The test setup
consists of a high voltage source, a divider, an RF receiver, an oscilloscope, and the
string of insulators under test. When the HV source is energized, the insulators
inside the clean fog chamber will experience high voltage stresses, resulting is partial
discharge activities. The divider provides a synchronised waveform of the HV source
for phased resolved analysis. The radiation captured by the RF receiver and the
leakage current waveform measured by the CT are present at the oscilloscope input
terminals for sampling and storing. Copper braided wire was used for grounding
the insulators string. It should be noted that all the recorded measurements during

- 48 -
 4.2 Lab Test Setup

Figure 4.1: Test arrangement in the High Voltage Lab.

this research always included two insulators on the string while the condition of
the insulator attached to the high voltage bus was different depending on the test
scenario.

4.2.1 High Voltage AC Transformers

Two different high voltage sources were available in the Lab; a 45 kV oil filled
transformer and a 230 kV potential transformer.

 45 kV Oil Filled Transformer:

The initial measurements were performed using the 45 kV oil filled transformer

- 49 -
 4.2 Lab Test Setup

that is shown in Figure 4.1. To control the output voltage of this transformer,
a variac was placed in the circuit to adjust the output voltage. A calibration
was performed under no-load conditions and the reference calibration graphs were
developed. A multimeter connected to the primary input of the variac can precisely
read the input voltage of the variac. Using the readings on the multimeter and the
reference calibration graphs, the output voltage can be found at any time during
the test.

 230 kV Potential Transformer:

Figure 4.2: 230 kV potential transformer.

- 50 -
 4.2 Lab Test Setup

The 230 kV potential transformer, as shown is Figure 4.2, was the main power
supply for the lab measurements. Potential transformers are classified as instrument
transformers. The main application of these transformers is stepping down the
voltage for the measuring instruments to get a proper reading off the line voltage.
However, they can also be used as power sources to provide the power to the circuit.
Due to low power ratings on these units, caution should be taken before energizing
the circuit with these transformers. The short circuit characteristics as well as
the overvoltage ratings should be considered for each test. Same as the 45 kV
transformer, a variac was used to adjust the output voltage. A calibration has also
been performed on this unit to develop the primary-secondary reference calibration
graphs.
This transformer was also used when a higher voltage was required for
performing puncture tests on the insulators. Details on the puncture tests are
available in Section 4.2.6.2

4.2.2 Divider

To measure the output voltage of the transformer, a high voltage divider was
connected to the high voltage bus of the test set up as shown in Figure 4.1. The
capacitive divider used in this project is rated for 225 kV and is manufactured by
Ross Engineering [90].The voltage waveform from the divider was used for generating
the 60 Hz reference waveform during the phase resolved analysis.

- 51 -
 4.2 Lab Test Setup

4.2.3 D-dot Electric Field Sensor

The electromagnetic field sensors are passive devices can provide a voltage on their
output terminals as a replica of the incident fields. In case of D-dot sensors, the
voltage is proportional to the time derivative of the electric displacement field.
Figure 4.3 shows a Prodyn’s ACD-S30 passive D-dot sensor which was used in
this study which has a bandwidth of 1 GHz [91]. The output voltage of the D-dot
sensor is given by,

∂
V0 = R × Aeq × D × cos θ (4.1)
∂t

Figure 4.3: Prodyn’s ACD-S30 Sensor.

where, R is sensor’s load characteristic impedance (50 ohms in this case), Aeq
is the sensor equivalent area provided by the manufacturer, D is the magnitude of
electric displacement vector and θ is the angle between the normal vector of sensor
ground plane and the incident electric field, E.
To compare the voltage received at the terminal of the D-dot sensor and

- 52 -
 4.2 Lab Test Setup

the actual discharge current on the insulator, one of the spare channels on the
oscilloscope was used to capture the leakage current using a high frequency current
transformer (HFCT). As shown in Figure 4.4 the peak location of the waveform
detected by the D-dot probe is only 10 ns from the the peak of the waveform detected
by the CT. Since the focus of this experiment is on the relative phase location of the
discharge peaks, this difference in the peak location is negligible when performing
60 Hz phase resolved analysis (around 2 × 10−4 degrees).

Figure 4.4: Recorded voltage of D-dot probe and the CT during a PD activity.

4.2.4 Current Transformer (CT)

Current transformers (CT) were placed on the ground wire during the initial
phases of the project for validation purposes. It should be noted that any other
sensor/antenna can be used as a receiver for this purpose as long as its bandwidth

- 53 -
 4.2 Lab Test Setup

is broad enough and its transfer function is known. The bandwidth of the receiver
has to be broad enough to capture all the required high frequency components. PD
radiation spectrum is mostly below 200 MHz [92] and therefore receivers with a
bandwidth higher than of 200 MHz can be used for this application. The current
transformers used in this research were manufactured by Magnelab (previously
owned by Bergoz [93]). The CTs were used to compare the radiation signature
to the partial discharge activities in the ground wire. Also the CT signals were used
to generate the 60 Hz reference signal and compare with the divider phase reference
voltage during the phase resolved analysis. The CTs come in different sizes and
bandwidth. The CTs used in this research were CT-C1.0 (bandwidth of 500 MHz
and a one to one ratio) and CT-C0.5 (bandwidth of 200 MHz and half ratio). A
picture of these current transformers is shown in Figure 4.5.

Figure 4.5: Current transformers manufactured by Magnelab/Bergoz [93].

4.2.5 Oscilloscope

All the analog signals captured by the receivers have to be sampled and digitalized
before storing on digital memories. There are three key features that should be

- 54 -
 4.2 Lab Test Setup

considered in choosing the right oscilloscope for any application; the bandwidth,
sampling rate, and maximum storage points. For this research, a minimum
bandwidth of 1 GHz was required. Sampling rates could vary depending on the
number of channels in use and also the length of the time window. At smaller time
windows, the oscilloscope can operate at its maximum sampling rate. In case of
larger time frames, the sampling rate will be limited by the maximum storage point
in the memory. For example, consider an oscilloscope with a sampling rate of 4
GS/sec and an 8 million point memory storage capacity. When the time division is
set to 2 ms/div, the total screen covers 10×2 ms/div =20 ms of the recorded data.
Now if the oscilloscope samples this waveform at 4 GS/sec, the total points required
to save this information would be 80 million points which is larger the oscilloscope’s
built-in memory. Therefore, in this case, the oscilloscope will sample at 400 MS/sec
to be able to store the 8 million sample points in its memory. In case of smaller time
frames, although the built-in memory is capable of storing higher sampling rates,
we are limited by the maximum sampling rate specified by the manufacturer of the
unit. To summarize this discussion, When the time frame is large, we are limited
by oscilloscope’s memory size and when the time frame is small, we are limited by
oscilloscope’s maximum sampling rate. In this research, a 7104A Agilent digital
oscilloscope [94] is used for digitization and storage of captured electromagnetic
radiations. The oscilloscope has a bandwidth of 1 GHz and a maximum sampling
rate of 4 GS/sec and a 4 million point memory storage capacity. The oscilloscope
is set on the external rising edge trigger mode that stores the waveforms when it
detects a partial discharge pulse on the input channel assigned for triggering.

- 55 -
 4.2 Lab Test Setup

4.2.6 Defective Porcelain Insulators

The transformer energizes the porcelain insulators under test and the radiation
captured by the sensors is digitalized and stored on the oscilloscope. The string of
insulators is hung off a pi-shaped wooden structure as shown in Figure 4.1. The
string is connected to the ground from the top, and to the high voltage bus through
the bottom insulator. During all the measurements, the string had two insulators
where the insulator near the ground wire was always a normal, healthy, and clean
insulator. The other insulator, connected to the HV bus, was altered during the tests
to study the punctured, contaminated, and normal radiation signatures. There are
also various types of porcelain insulators available in the market. The insulators
used in this project were the class B, 15000 lbs NGK units [95] since Manitoba
Hydro is using a large number of these units in their system. The scope of this
project involved studying the radiation signature of contaminated and punctured
insulators. To duplicate these defective insulators, a lab procedure was developed.
Details on the procedures of making defective insulators in the lab is presented in
the next section.

4.2.6.1 Contaminated Insulators

To make contaminated insulators in the lab, we followed the IEC 60507 standard
[20] that was discussed in more details in Section 4.1.1. According to IEC 60507
and Table 4.1, the site contamination level can be categorized in four groups; very
light, light, moderate and heavy. Insulators studied in this research were made to
a heavy contamination level. In order to make a contaminated insulators with a

- 56 -
 4.2 Lab Test Setup

desired contamination level, the proper suspension should be made and sprayed on
the insulators. The conductivity of the suspension was calculated based on the IEC
standard where the relation between the Equivalent Salt Deposit Density (ESDD)
level and suspension’s conductivity is given [20]. To ensure uniformity and proper
dosage, nozzle should be kept 30 cm away and the insulator should be rotated 360
degrees. To make contaminated insulators with other ESDD levels, the conductivity
of the suspension should change accordingly.
The type of contaminants that can be deposited on the insulators in the field
depends on the area that they are installed. Contaminants can be classified as soluble
or non-soluble. Soluble contaminants have a much higher impact on insulators
performance. To standardize and classify the contamination level of insulators,
the measure of Equivalent Salt Deposit Density (ESDD) is used in this study. It
should be noted that in order to precisely classify the contamination level of outdoor
insulators, the Non-soluble Deposit Density (NSDD) level, representing the level of
non-soluble contaminants, must also be taken into account. However the impact of
NSSD is very small [96] and therefore it was not considered in this research.
The IEC procedure [20] was followed to make the proper suspension required to
be sprayed on the insulators. First, 40 grams of Kaolin was mixed into 1000 grams
of tap water. The conductivity of tap water in the lab was 357 µS/cm, measured by
Orion 3-Star Plus, a portable conductivity meter manufactured by Thermo Scientific
[97]. Eighty five grams of commercially-available salt was added to the solution to
raise the conductivity of the solution to 11.2 S/m. This solution was then sprayed on
the insulators in order to produce the desired ESDD level of 0.28 mg/cm2 . Figure 4.6
shows a string of two insulators where the bottom insulator has been sprayed with
the contaminated suspension.

- 57 -
 4.2 Lab Test Setup

Figure 4.6: Contaminated insulator under the test; clean and healthy unit on top
and contaminated unit at the bottom.

Depending on how the suspension is sprayed on the insulators and also the
condition of the insulator prior to spraying, the ESDD level of the insulator
can be different. Therefore, to ensure the ESDD level of the insulators after
spraying this suspension is close to the desired value of 0.28 mg/cm2 , different
scenarios were investigated. The effect of washing, drying, and cleaning prior
to applying the suspension, the angle of spray nozzle, and number of sprays
on the final ESDD level was studied. For each scenario, the ESDD level after
applying the suspension was calculated based on the method presented in IEC 60507.

Figure 4.7 shows the procedure of measuring the ESDD level of the contaminated
insulators. An area on the insulator was completely wiped off using previously
cleaned and dried cotton balls. After collecting all the dirt and contaminants, cotton

- 58 -
 4.2 Lab Test Setup

Figure 4.7: Measuring the ESDD level in the lab.

balls were immersed in distilled water to dissolve any contaminants absorbed by the
cotton balls and the conductivity of the solution was then measured. Knowing the
area that was cleaned with the cotton balls and also the volume of distilled water
that the cotton balls were dissolved in, the ESDD level of the polluted insulator was
calculated using [20]:

Sa = (5.7 × σ20 )
1.03
(4.2)

Sa × V
SDD = (4.3)
A

where SSD is the salt deposit density (mg/cm2 ), σ20 is the volume conductivity
of the suspension, Sa (kg/m3 ) is the salinity of the suspension, V is the volume of
distilled water, and A is the area that was wiped off by cotton balls. To imitate
the real pollution on the insulators, the insulators were wetted before and after the
energization according to the wetting procedure given in [20].

- 59 -
 4.2 Lab Test Setup

After trying different methods for applying the suspension on the insulators and
measuring the ESDD level of each sample, a procedure was developed and followed to
achieve the same ESDD level knowing the suspension conductivity. Also, throughout
the measurement, the ESDD of some samples were randomly checked to ensure the
contamination level is within the acceptable range.

4.2.6.2 Punctured Insulators

Internal punctures are one of the most common defects in porcelain insulators.
The formation of puncture paths inside the insulators is mostly due to the cement
expansion near the cap and pin joint. Low quality cement materials absorb the
moisture and expand over time, causing undesired puncture paths. Lightning can
also cause punctures in the cap. Punctured insulators are very hard to inspect
visually. There two routine tests which are described in IEC standard [89] that
were inspiring for making punctured insulators in the lab.

(1) Electromechanical Failing Load Test:

IEC has a standard testing procedure for performing tensile tests on porcelain
insulators. The IEC refers to this test as an electromechanical failing load test.
During the test, the string of porcelain insulators will be subjected to the AC power
source while the tensile load is applied to the string at the same time. According
to the IEC Standard 60383-1 [89], the tensile load should rapidly but smoothly
increase from 0 to 75% of the rated failing load value and then gradually increase
to 100%. One way to make punctured insulators is utilizing a similar procedure.
The idea is to put one porcelain insulator under an adjustable tensile load using

- 60 -
 4.2 Lab Test Setup

the tensile testing machine. A tensile machine is available in one of University

of Manitoba Structure Lab and can be controlled by a computer that programs
load characteristics. Increasing the tension will eventually develop punctures in
the insulator. One way to monitor the condition of the insulator is measuring the
resistance across the unit while it is under the tensile load. When the puncture
develops between the cap and pin, the resistance of the insulator should drop
significantly. Fluke 1507 was the Megger used for the resistance monitoring during
the tensile load test and can measure resistance up to 10 GΩ [98].
Insulators were taken to the University of Manitoba Structure lab and were placed
and secured inside the Instron 300 DX tensile tester. To fit the unit into the
tensile tester, special links were made out of steel plates, considering the maximum
mechanical withstand of the insulators. Figure 4.8 shows the test setup for the
puncture test using the tensile tester. During multiple tries on different insulators,
we noticed that the bottom metallic pin is the first part that fails the tensile test
(Figure 4.8, top right). Therefore, we were never able to apply enough force to
exceed the nominal mechanical strength of the insulators to introduce the puncture.
At this point we decided to make the puncture insulators using high voltage stresses
instead of mechanical forces.

- 61 -
 4.2 Lab Test Setup

Figure 4.8: Puncture test arrangement using Tensile Machine Instron DX300; test
setup on left, failed pin on top right, and custom built link and resistance monitoring
on bottom right.

(2) Puncture Withstand Test:

Section 15 of the IEC standard 60383-1 [89] explains the puncture withstand test
for testing porcelain insulators. According to the standard, the insulators should be
cleaned and dried before the test. Then the insulator should be immersed in a tank
filled with liquid insulating materials. The temperature of this liquid insulating
material should be at the room temperature. While immersing the insulator in
the tank, precautions should be taken to avoid forming air bubbles underneath the
shells. The power frequency voltage across the unit should then rapidly raise to the

- 62 -
 4.2 Lab Test Setup

withstand level specified by the manufacturer. If no puncture occurs while ramping

up the voltage to the withstand voltage, the insulator passes the test. If one is
interested in recording the puncture voltage, the voltage can be raised above the
specified withstand voltage limit until the puncture occurs. We decided to follow
this test procedure and apply higher voltage stresses across the insulators until a
puncture develops inside the insulation.
Figure 4.9 shows the test arrangement for making punctured insulators. When
the puncture path develops inside the insulator, the insulator can not withstand the
nominal puncture withstand voltage any more.

- 63 -
 4.2 Lab Test Setup

Figure 4.9: Puncture test arrangement using IEC 60383-1; test setup on left,
insulator immersed in oil and under high voltage stress on top right, development
of puncture on middle right, and punctured insulator ready to be tested for radiation
signature on bottom right.

- 64 -
 4.2 Lab Test Setup

4.2.6.3 Insulator Identification

As explained earlier in this Chapter, puncture and contamination are the two
types of defects that have been studied in this research. Also, strings of normal
healthy and clean insulators were energized and their electromagnetic radiations
were captured and analysed. Testing normal healthy and clean insulators provides
a baseline signature pattern that can be compared will defective units. Table 4.2
shows the eight insulators studied in this research showing the type of defect each
one had.

Table 4.2: Identification table for studied insulators

Total Number of Cycles Collected

Type of Defect Total Samples
of Captured PD Events Datasets
Clean/Normal 90,000 176 3
Contaminated 145,000 288 3
Punctured 175,000 342 2
Total 410,000 806 8

It should be noted that in case of testing the contaminated insulators, all three
sample insulators were first made into the desired ESDD level of 0.28 mg/cm2 , using
the procedures explained in the previous sections. When the contaminated insulator
was placed inside the fog chamber and one dataset of partial discharge radiation
was captured, the contaminated insulator was taken off the string and replaced with
another contaminated insulator that has never been tested inside the fog chamber.
The contaminated insulators that were taken off the string after being tested in
the fog chamber were then properly washed and cleaned. The clean and washed

- 65 -
 4.3 Summary

insulators went under the same contamination procedure to make new contaminated
insulators with the same desired ESDD levels. The reason for implementing the
cycle of washing, cleaning, and redepositing the contamination layer was to ensure
the contamination layer is not getting washed off after being in the fog chamber to
a long period of time.

4.3 Summary

In the chapter, the experimental setup and equipment employed in this thesis were
described. The test setup for collecting the measurements and the role of each
equipment in the test setup was explained. The lab procedures as well as the IEC
standards that were followed for making contaminated and punctured insulators
were also discussed.

- 66 -
5

Data Processing and Classification

A total of 410,000 cycles (806 datasets) of the recorded measurements were

analysed. This chapter presents the processing techniques employed in this thesis
and summarizes the results obtained from extracting the statistical features of
captured radiations. The classification algorithm for differentiating punctured and
contaminated insulators using their radiation signature is also presented.

5.1 Data Processing

The two most common types of defective porcelain insulators made in lab (details
were presented in Section 4.2.6), were placed under high voltage stresses (test setup
shown in Figure 4.1) to investigate their radiation signature. When the defective
insulators are placed under high voltage stresses, partial discharge activities initiate.
These discharges are basically the electrons bridging the insulation (internal in
case of puncture path and external in case of surface contamination) and can be
detected by monitoring the leakage current and radiation signature of the insulators.

- 67 -
 5.1 Data Processing

The focus of this thesis is to capture and process the radiation signature of these
discharges to find the relation between the radiation signature patterns and type of
defects. In this section, the signal processing techniques that were applied and
studied on the radiated signatures of recorded measurements will be presented.
Punctured insulators, contaminated insulators, and normal (healthy and clean)
insulators were studied and the radiation signature for each case was recorded and
compared with the others. The test setup for this measurement had two insulators
on the string connected to a 55 kV power frequency high voltage bus. The source
voltage of 55 kV was chosen based on the PD inception voltage when the string had
only normal insulators.
The oscilloscope was programmed to trigger when a discharge occurs and to
capture one 60 Hz cycle of data. One cycle of data that contains the partial discharge
data is referred to one “Segment” in this thesis. After capturing one segment, the
oscilloscope will be automatically re-armed and wait for the next partial discharge
to trigger and record it in the next segment. The oscilloscope is capable of recording
multiple segments. A total of 512 segments is referred to one “Dataset”. Due to
memory limitation on the oscilloscope, after capturing 512 segments (or one dataset)
the data had to be stored on the hard drive.
After collecting the data from the tested insulators, the information should
be processed to extract the unique features of radiated pulses. These features
will eventually help the classifier to identify the radiation signatures. Identifying
and extracting these features can be done in different domains. A wide range of
data processing techniques can be used to extract the unique features of radiated
pulses. The focus of our data processing techniques was on the phased resolved
statistical behaviour of partial discharges. Preliminary analysis performed on the

- 68 -
 5.1 Data Processing

captured radiation pulses showed distinct features in the pattern of the probability
density functions of the PD pulses. The signal possessing on the recorded data was
performed in MATLAB [86]. The recorded datasets captured by the oscilloscope
were loaded into MATLAB, where a graphical user interface (GUI) was developed to
perform the analysis. The GUI allows the user to select different datasets/segments
and perform a number of signal processing analysis on them. Figure 5.1 shows
a view of the GUI that was developed for this research. One of the first steps
prior to performing the data processing was to properly identify and locate the
location of partial discharge pulses. Figure 5.2 shows a sample of captured partial
discharge pulse in red while the peaks of the waveform are identified and shown
with blue points. However, this activity corresponds to only one discharge pulse
and the resonance following the initial discharge is due to the system response [99].
Therefore, only one peak location (maximum) will be extracted from this discharge
activity for data processing analysis.
In the remainder of this section, details of the signal processing techniques used
for extracting the features will be discussed.

- 69 -
 5.1 Data Processing

Figure 5.1: View of the Graphical User Interface (GUI) developed in MATLAB for
performing signal processing.

Figure 5.2: Identifying peak locations on the recorded partial discharge pulse.

- 70 -
 5.1 Data Processing

5.1.1 Weibull PDF Fitting on Phase Location of PD Pulses

One feature of the radiated PD pulses studied in this research is the phase location
of where the PD pulses happen. Figure 5.3 shows a sample of a recorded data of
one segment captured by the oscilloscope. In the example shown in Figure 5.3, two
PD pulses are present in the captured segment (one 60 Hz cycle); the first discharge
happens around phase location of 220 degrees and the second discharge happens at
the phase location of 20 degrees. The phase location of PD pulses on all segments
of each dataset are extracted and a two sided Weibull probability density function
(PDF) is fit to the extracted histogram of the phase location of the peak.
Sample of One Captured Segment
0.15
60 Hz Signal
Radiated PD Signal

0.1

0.05
Discharge Amp (V)

−0.05

−0.1

−0.15

−0.2
0 0.002 0.004 0.006 0.008 0.01 0.012 0.014 0.016 0.018
Time (seconds)

Figure 5.3: Sample of recorded measurements in one cycle.

A Weibull probability density function is given by

b x b−1 −( x )b
f (x∣ a, b) = ( ) e a (5.1)
a a

where:

- 71 -
 5.1 Data Processing

Weibull PDF Fitting − Punctured Insulator

0.016

0.014

0.012
Probability Density

0.01

0.008

0.006

0.004

0.002

0
0 50 100 150 200 250 300 350
Phase of Discharge Pulses

Figure 5.4: Fitting a two sided Weibull distribution to phase of PD pulses of

512 segments.

“a” is the shape parameter

“b” is the scale parameter

Figure 5.4 shows a Weibull fitting on one dataset of captured radiations. After
finding the Weibull fit at each dataset, the shape and scale parameters (a and b
in Equation 5.1) of the Weibull function at each side,and also the linkage between
the sides is calculated and stored. To fit a two sided Weibull distribution, two
separate Weibull distributions were fit into each side of the phase resolved graphs,
each having their own shape and scale parameters. The linkage is defined as a
normalized weighting function combining the two density functions when using the
maximum likelihood algorithm. In other words, the two sided distribution function
is a mixture PDF distribution where the linkage parameter (defined in this thesis)
is the mixture weight associated with each distribution.

- 72 -
 5.1 Data Processing

The shape (two values, one per side), scale (two values, one per side) and linkage
of the two sided Weibull density function for each dataset is then calculated and
recorded. The same procedure is repeated for all the recorded datasets.

5.1.2 Weibull PDF Fitting on Amplitude of PD Pulses

Another feature of the captured radiated PD pulses, is the amplitude of the

discharge. As shown in Figure 5.3, the amplitude the PD pulse is different in the
two events that are captured in the cycle. Therefore, the amplitude of the PD pulses
on all segments of each dataset are extracted and a two sided Weibull probability
density function (PDF) is fit to the extracted scatter plot, similar to the phase of
PD pulses. An example of Weibull amplitude curve fitting is presented in Figure
5.5
Amplitude Fitting
90

80

70

60
Probability Density

50

40

30

20

10

0
0.02 0.04 0.06 0.08 0.1 0.12 0.14
PD Amplitude

Figure 5.5: Fitting a two sided Weibull distribution to the amplitude of PD pulses.

The shape (two values, one per side), scale (two values, one per side) and linkage

- 73 -
 5.1 Data Processing

of the two sided Weibull density function for each dataset is then calculated and
recorded. The same procedure is repeated for all the recorded datasets.

5.1.3 Weibull PDF Fitting on Weighted Phase of PD Pulses

To simultaneously investigate both the amplitude and phase characteristics of the

PD pulses, a weighted phase plot was generated. In this representation, based on
the amplitude of the PD pulses, the number of discharges per each phase window
is weighted. The idea behind this representation is to study the statistical pattern
of the PD pulses in terms of phase location of PD peaks and the amplitude of PD
pulses simultaneously. Similar to previous Weibull fittings, a two sided Weibull
probability density function is fit to the calculated histograms and the shape (two
values, one per side), scale (two values, one per side) and linkage of the two sided
Weibull density function for each dataset is calculated and recorded. Figure 5.6
shows an example of the weighted phase fitting.

- 74 -
 5.1 Data Processing

Weighted Phase Fitting

0.025

0.02

0.015
Probability Density

0.01

0.005

0
0 50 100 150 200 250 300 350
Phase of Discharge Pulses

Figure 5.6: Fitting a two sided Weibull distribution to the weighted phase of PD
pulses.

5.1.4 Spectral Fractal Analysis

To study the signature of concurrent discharge activities on the radiated fields,

the concept of spectral fractal analysis was used. The power density spectrum
of the signal provides an estimate of the fluctuations of the signal vs frequency.
According to [82], if the underlying process has periodic (or quasi-periodic) behavior,
the power spectrum of the signal has equally spaced harmonics with significant power
at very low frequencies. During the initial rounds of measurements, in some of the
captured cycles, low amplitude concurrent discharges were observed. In order to
investigate these discharges within one cycle of captured radiations, the spectral
fractal exponent, which is the slope of the linear fit to the log log spectrum of the
signal in lower frequencies, was calculated [82].
Therefore, the power spectrum of each of the 512 segments in each dataset was

- 75 -
 5.2 Data Processing Results

calculated and a linear line in the range of 100 Hz to 10 kHz was fit to the log-
log spectrum. The line parameters of the linear fit for each dataset are stored and
averaged at the end of each run. Figure 5.7 shows an example of spectral fractal
parameters calculated for one segment of a dataset where the signal spectrum is
shown in blue and a linear curve fitting is shown in red.
−7

−8

−9
Log Scale Magnitude of Spectrum

−10

−11

−12

−13

−14

−15
1.5 2 2.5 3 3.5 4
Log Scale Frequency

Figure 5.7: Parameters of spectral fractal curve fitting.

5.2 Data Processing Results

The string of insulators containing different combinations of defective and healthy

units were placed under high voltage stresses to capture the electromagnetic
radiation off PD pulses. This section presents the signature statistical graphs
calculated based on the analysis described in Section 5.1 for the three types
of insulators that were studied in this research; normal (healthy and clean),
punctured, and contaminated insulators. The graphs presented in this section

- 76 -
 5.2 Data Processing Results

contain the statistical features of the captured electromagnetic radiation calculated

for each dataset of recorded measurements. From the total of 410,000 recorded
measurements, 90,000 cycles (176 datasets) of recorded radiations are from
normal insulators, 175,000 cycles (342 datasets) from punctured insulators, and
145,000 cycles (288 datasets) from contaminated insulators. This section presents
the statistical features resulted from the processing performed on the captured
measurements.

5.2.1 Weibull PDF Fitting on Phase Location of PD Pulses

As explained earlier in Section 5.1, a two sided Weibull distribution is fitted to

the phase location of discharge activities of each dataset (512 cycles) of captured
electromagnetic radiations. The extracted parameters of each Weibull fit include
the shape and scale parameter of positive half cycle (blue points), the shape and
scale parameter of negative half cycle (red points) and the linkage between the two
distributions. Figures 5.8, 5.9, 5.10 present the Weibull phase fit parameters in a 3D
representation, where the x axis is the shape parameter of the Weibull fit, the y axis
is the scale parameter of the Weibull fit, and the z axis is the linkage between the
two distributions for normal, punctured, and contaminated insulators respectively.

- 77 -
 5.2 Data Processing Results

Figure 5.8: Fitted phase Weibull parameters - normal insulator; one blue point
(positive half cycle) and one red point (negative half cycle) represent the extracted
features from one of the 176 datasets during the phase Weibull fitting.

- 78 -
 5.2 Data Processing Results

Figure 5.9: Fitted phase Weibull parameters - punctured insulator; one blue point
(positive half cycle) and one red point (negative half cycle) represent the extracted
features from one of the 342 datasets during the phase Weibull fitting.

Figure 5.10: Fitted phase Weibull parameters - contaminated insulator; one blue
point (positive half cycle) and one red point (negative half cycle) represent the
extracted features from one of the 288 datasets during the phase Weibull fitting.

- 79 -
 5.2 Data Processing Results

5.2.2 Weibull PDF Fitting on Amplitude of PD Pulses

As explained earlier in Section 5.1, a two sided Weibull distribution is fitted to the
amplitude of discharges of each dataset (512 cycles) of captured electromagnetic
radiation. The extracted parameters of each Weibull fit include the shape and
scale parameter of the lower magnitude discharge (blue points), the shape and scale
parameter of the higher magnitude discharge (red points) and the linkage between
the two distributions. Figures 5.11, 5.12, 5.13 present the Weibull amplitude fit
parameters in a 3D representation, where the x axis is the shape parameter of the
Weibull fit, the y axis is the scale parameter of the Weibull fit, and the z axis is
the linkage between the two distributions for normal, punctured, and contaminated
insulators respectively.

- 80 -
 5.2 Data Processing Results

Figure 5.11: Fitted amplitude Weibull parameters - normal insulator; one blue
point (lower magnitude discharge) and one red point (higher magnitude discharge)
represent the extracted features from one of the 176 datasets during the amplitude
Weibull fitting.

- 81 -
 5.2 Data Processing Results

Figure 5.12: Fitted amplitude Weibull parameters - punctured insulator; one blue
point (lower magnitude discharge) and one red point (higher magnitude discharge)
represent the extracted features from one of the 342 datasets during the amplitude
Weibull fitting.

- 82 -
 5.2 Data Processing Results

Figure 5.13: Fitted amplitude Weibull parameters - contaminated insulator; one blue
point (lower magnitude discharge) and one red point (higher magnitude discharge)
represent the extracted features from one of the 288 datasets during the amplitude
Weibull fitting.

- 83 -
 5.2 Data Processing Results

5.2.3 Weibull PDF Fitting on Weighted Phase of PD Pulses

As explained earlier in Section 5.1, a two sided Weibull distribution is fitted to the
weighted phase of discharges of each dataset (512 cycles) of captured electromagnetic
radiation. The extracted parameters of each weighted Weibull fit include the
shape and scale parameter of positive half cycle (blue points), the shape and
scale parameter of negative half cycle (red points) and the linkage between the
two distributions. Figures 5.14, 5.15, 5.16 present the weighted Weibull phase fit
parameters in a 3D representation, where the x axis is the shape parameter of the
weighted Weibull fit, the y axis is the scale parameter of the weighted Weibull fit,
and the z axis is the linkage between the two distributions for normal, punctured,
and contaminated insulators respectively.

- 84 -
 5.2 Data Processing Results

Figure 5.14: Fitted weighted Weibull parameters - normal insulator; one blue point
(positive half cycle) and one red point (negative half cycle) represent the extracted
features from one of the 176 datasets during the weighted phase Weibull fitting.

- 85 -
 5.2 Data Processing Results

Figure 5.15: Fitted weighted Weibull parameters - punctured insulator; one blue
point (positive half cycle) and one red point (negative half cycle) represent the
extracted features from one of the 342 datasets during the weighted phase Weibull
fitting.

- 86 -
 5.2 Data Processing Results

Figure 5.16: Fitted weighted Weibull parameters - contaminated insulator; one

blue point (positive half cycle) and one red point (negative half cycle) represent the
extracted features from one of the 288 datasets during the weighted phase Weibull
fitting.

- 87 -
 5.2 Data Processing Results

5.2.4 Spectral Fractal Analysis

As explained earlier in Section 5.1, the spectral fractal parameters for each segment
of the dataset is calculated as part of the signal processing on the recorded
electromagnetic radiations. The extracted parameters include the line parameters of
a linear fit to the log-log spectrum of each electromagnetic radiation waveform in the
range of 100 Hz to 10 kHz. The slope and Y-intercept of the fitted line in averaged
over the 512 segments in each dataset and ultimately two averaged parameters will
be extracted from each dataset. Figures 5.17,5.18,5.19 present the spectral fractal
line parameters where the x axis is the slope of the fitted line, the y axis is the
y-intercept parameter of the fitted line for normal, punctured, and contaminated
insulators respectively.

- 88 -
 5.2 Data Processing Results

Figure 5.17: Line parameters of the spectral fractal analysis - normal insulator; each
point represents the average of the fitted line parameters to 512 segments in each of
the 176 datasets.

- 89 -
 5.2 Data Processing Results

Figure 5.18: Line parameters of the spectral fractal analysis - punctured insulator;
each point represents the average of the fitted line parameters to 512 segments in each
of the 342 datasets.

- 90 -
 5.3 Classification

Figure 5.19: Line parameters of the spectral fractal analysis - contaminated

insulator; each point represents the average of the fitted line parameters to 512
segments in each of the 288 datasets.

5.3 Classification

As described in Section 5.1, a total of 17 features were extracted from the

electromagnetic radiation signature of insulators under test. These 17 features
include five parameters from the phase Weibull fit, five parameters from the
amplitude Weibull fit, five parameters from the weighted Weibull fit, and two
parameters from the spectral fractal analysis. To investigate the signature of
defective insulators within the extracted features, a 2D support vector machine
(SVM) classifier was developed. The SVM classifier targets to classify the type of

- 91 -
 5.3 Classification

insulators by searching distinct information in all 2D possible combinations of the

extracted features. Having 17 features, there are 136 unique possible combinations
(2 C17 = 136) for a 2D feature domain that the SVM classifier can look for distinct
features. The performance of the classifier is also measured by its ability to correctly
classify the recorded radiations. This section explains the algorithm for the SVM
classifier and classification performance.

5.3.1 Classification Algorithm

For each type of insulator, a matrix with 17 rows corresponding to the 17 features
was created. The number of columns on these matrices depends on the number
of observation datasets available for each type of defect. Each column represents
the statistical features from 512 cycles of recorded measurements. To identify the
features with the maximum variance of information suitable for classification, a
2D support vector machine (SVM) classifier is developed. By supplying known
observation sequences to the SVM algorithm, the classifier tries to find the best
polynomial which can separate the 2D group of data from each other in the specified
domain. The classifier uses support vectors si and weights ai and a bias value b to
find the optimized polynomial differentiating the two known categories. After finding
the best polynomial (k, the kernel function), any observation x can be passed on to
the SVM classifier where it calculates c using:

c = ∑ ai k(si , x) + b (5.2)
i

If c ⩾ 0, x will be classified in the first category and and if c ⩽ 0, it will be classified

in the second category.

- 92 -
 5.3 Classification

In the following section, using the extracted features and the classification
method explained earlier, every two types of insulators will be compared and
analysed to study unique features of each type of insulators.

5.3.2 Fundamental 2D Classification of All Insulators

Having the feature matrix for each type of insulators, the classifier was trained for all
possible 136 combinations of the 17 extracted features (2 C17 ) of recorded observation
datasets. In order to identify the features with maximum classification information,
during each 2D classification, four random samples from the first type and four
random samples from the second type of insulator were selected and a MATLAB
code was developed to classify these eight samples using 136 different classifiers. At
the end of this run, if a certain feature combination was able to correctly classify
three or four of the random samples correctly in both categories, that combination
was stored. This process is repeated for 10,000 times to evaluate the information in
various combination of the extracted features.

5.3.2.1 Puncture vs Contamination

Figure 5.20 shows the results of 10,000 classifications performed on measurement

datasets of punctured and contaminated insulators. The four most successful
combinations in differentiating the punctured and contaminated insulators are
shown in Table 5.1.
As shown in Table 5.1, the shape of fitted weighted Weibull in the negative cycle
appears in all the combinations. To better illustrate the classification process, a
classification using the combination ID 92 (see Table 5.1) is shown in Figure 5.21.
The highlighted area in Figure 5.21 is zoomed in and presented in Figure 5.22

- 93 -
 5.3 Classification

Figure 5.20: Successful classification rate of contaminated and punctured insulators

using various combinations of extracted features.

to illustrate the performance of the classifier in differentiating contaminated and

punctured insulators. Needless to say, many sets of features from both defects
overlap in the Figure 5.22 and it is not possible to distinguish between the two.
This proves that there are many common discharge activities that occur on both
punctured and contaminated insulators.
The red line in Figure 5.21 and Figure 5.22 is the polynomial representing the
decision function, separating the extracted features of punctured and contaminated
insulators in Domain ID 92.

- 94 -
 5.3 Classification

Table 5.1: Features with highest successful classification in Figure 5.20.

Combination Successful
Feature 1 Feature 2
ID Classification
Weighted Weibull : Weighted Weibull:
92 5100
Shape - Negative cycle Scale - Negative cycle
Weighted Weibull : Weighted Weibull:
93 5520
Shape - Negative cycle Linkage
Weighted Weibull : Amplitude Weibull:
94 5549
Shape - Negative cycle Shape - Positive Cycle
Weighted Weibull: Amplitude Weibull:
95 5006
Shape - Negative cycle Scale - Positive Cycle

Figure 5.21: Sample of 2D classification of contaminated and punctured insulators

using Combination ID 92.

- 95 -
 5.3 Classification

Figure 5.22: Zoomed highlighted area of Figure 5.21.

- 96 -
 5.3 Classification

5.3.2.2 Puncture vs Normal

Figure 5.23 shows the results of 10,000 classifications performed on measurement

datasets of punctured and normal insulators. The four most successful combinations
in differentiating the punctured and normal insulators are shown in Table 5.2.
No unique feature stands out in this classification for having a high successful
classification rate. To better illustrate an example of classification in this category,
classification of combination ID 66 is presented in Figure 5.24. The highlighted area
in Figure 5.24 is zoomed in and presented in Figure 5.25 to show the performance of
the classifier in differentiating between punctured and normal insulators. As it can
be concluded from the relatively lower successful classification rate in this category
(Figure 5.23), the features in these two categories overlap in various combinations
(Figure 5.25). Therefore, it is very difficult to find a polynomial that can completely
separate the features related to each type.

Table 5.2: Features with highest successful classification in Figure 5.23.

Combination Successful
Feature 1 Feature 2
ID Classification
Phase Weibull: Amplitude Weibull :
64 3293
Linkage Shape- Positive Cycle
Phase Weibull: Amplitude Weibull :
66 3355
Linkage Shape - Negative Cycle
Weighted Weibull : Amplitude Weibull :
110 3260
Linkage Scale - Positive Cycle
Amplitude Weibull:
124 3190 Amplitude Weibull - Linkage
Scale - Positive Cycle

The red line in Figure 5.24 and Figure 5.25 is the polynomial representing
the decision function, separating the extracted features of normal and punctured
insulators in Domain ID 66. The red circles in Figure 5.24 show four random

- 97 -
 5.3 Classification

Figure 5.23: Successful classification rate of punctured and normal insulators using
various combinations of extracted features.

puncture samples that were classified using the polynomial (in red) in this feature
domain. Three out of four random samples have been classified correctly while one
sample is classified incorrectly as normal. Similarly, the green circles in Figure 5.24
show four random normal samples that were classified using the polynomial (in
red) in this feature domain. Three out of four random samples have been classified
correctly while one sample is classified incorrectly as punctured.

- 98 -
 5.3 Classification

Figure 5.24: Sample of 2D classification of punctured and normal insulators using

Combination ID 66.

- 99 -
 5.3 Classification

Figure 5.25: Zoomed highlighted area of Figure 5.24.

- 100 -
 5.3 Classification

5.3.2.3 Contamination vs Normal

Figure 5.26 shows the results of 10,000 classifications performed on measurement

datasets of contaminated and normal insulators. The four most successful
combinations in differentiating the contaminated and normal insulators are shown
in Table 5.3. Similar to classification of punctured and normal insulators, no unique
feature stands out in this classification for having a high successful classification
rate. To better illustrate an example of classification in this category, classification
of combination ID 47 is presented in Figure 5.27. The highlighted area in Figure 5.27
is zoomed in and presented in Figure 5.28 to show the performance of the classifier in
differentiating between contaminated and normal insulators. As it can be concluded
from the relatively lower successful classification rate in this category (Figure 5.26),
the features in these two categories overlap in various combinations (Figure 5.28).
Therefore, it is very difficult to find a polynomial that can completely separate the
features related to each type.

Table 5.3: Features with highest successful classification in Figure 5.26.

Combination Successful
Feature 1 Feature 2
ID Classification
Phase Weibull: Amplitude Weibull:
12 2785
Shape - Positive Cycle Shape - Negative Cycle
Phase Weibull : Weighted Weibull:
47 3438
Scale - Negative cycle Shape - Positive cycle
Weighted Weibull : Amplitude Weibull:
75 2440
Shape - Positive cycle Shape - Positive Cycle
Weighted Weibull : Amplitude Weibull :
77 2571
Shape - Positive cycle Shape - Negative Cycle

- 101 -
 5.3 Classification

Figure 5.26: Successful classification rate of contaminated and normal insulators

using various combinations of extracted features.

- 102 -
 5.3 Classification

Figure 5.27: Sample of 2D classification of contaminated and normal insulators

using Combination ID 47.

- 103 -
 5.3 Classification

Figure 5.28: Zoomed highlighted area of Figure 5.27.

- 104 -
 5.3 Classification

5.3.2.4 Discussion on the Results of 2D Classifications

As presented in the previous section, many sets of features from all three types
of insulators overlap in many domains and it is not possible to find a decision
boundary that can successfully split the sets of features related to each defect even
when considering the concept of soft margin statistical classifiers [100].
The main reason for having so many overlapping datasets is in the nature of the
measurements. Partial discharge pulses are not solely originated from the defects on
the insulators. There are many other sources of partial discharges that are common
is all three different settings such as the PD sources within the transformer, corona,
and discharges due to high voltage stress around metal fittings of the insulators.
Since the HV source has a nominal rating of 230 kV, it is very unlikely that it can
be the source of PD around 55 kV, however there is no guarantee that the source is
PD free.
The classification results indicated that common discharge activities on all three
different test setups must be present, regardless of the type of defect on the
insulators. To investigate this common discharge activity, pictures with a corona
camera were taken from energized insulators in all three setups. Figure 5.29 shows a
normal discharge activity on a punctured insulator under high voltage stress. There
are many discharges around the metal fitting of the insulator. These discharge
activities are present even in the case of normal healthy insulators, contaminated
insulators and punctured insulators. These discharges are invisible to the naked eye
but can be seen using a corona camera. Figure 5.30 shows a discharge bridging
the puncture path in a punctured insulator differentiating the common discharge
activities from a puncture discharge.

- 105 -
 5.3 Classification

Figure 5.29: Normal discharge activities near the metal fitting on a punctured
insulator; corona camera on left and normal camera on right.

Figure 5.30: Discharge through the puncture path of a punctured insulator; corona
camera on left and normal camera on right.

According to the pictures taken by a corona camera and also the results presented
in the past sections, it was clear that a significant signature of normal discharge
activities will be present in the captured radiation of all insulators. Therefore, a
gating based algorithm was developed ([101],[102]) to filter out the common features
present in the discharge activities of normal insulators. The concept is to eliminate
the PD signatures that are present in the measurements recorded when only normal
insulators were on the tested string. The following section elaborates more on the

- 106 -
 5.3 Classification

gating algorithm and the results obtained after applying the gating algorithm.

5.3.3 Gating Algorithm

The measurement results show that normal discharge activities are quite dominant
in all types of studied insulators. Therefore, a gating algorithm was developed
to eliminate the signature of normal discharge activities from the captured
electromagnetic radiation of contaminated and punctured insulators. Examining
the successful rate of 2D classification of punctured and contaminated insulators
shown in Figure 5.20, the 2D Domain 92 (Domain parameters shown in Table 5.1)
was chosen to be the focus of the gating algorithm.
The gating algorithm discards any feature observation pair from the
contaminated or punctured insulator when they are in the vicinity of any feature
observation of normal insulators.The algorithm was implemented in the Domain 92
only and a new feature matrix was generated that excluded the overlaps of normal
insulators with contaminated and punctured insulators.
In other words, the normal discharge activities are treated like noise in the
classification algorithm. IEC 60270 [101] explains the use of gating algorithm for
noise cancellation applications as well as PD locating algorithms. OMICRON [75]
(one of the lead companies in building commercial PD detection systems) has also
implemented statistical gating algorithms in their PD acquisition system [103].
As shown in Figure 5.31, the successful classification rate has raised almost 50%
from 5100 successful classification in Table 5.1 to 7575 successful classification as
shown in Figure 5.31. The new classification shows unique information regarding
the type of defect in Domain 92.

- 107 -
 5.3 Classification

Figure 5.32 shows the 2D classification of between the punctured and

contaminated insulators when the overlapped normal discharge activities were
removed from the captured radiations of contaminated and punctured insulators.
Figure 5.33 shows the overlapped data in Domain 92 from normal, contaminated
and punctured insulators that were taken out by the gating algorithm.

Figure 5.31: Successful classification rate of punctured and contaminated insulators

excluding the overlapped normal discharge activities using various combinations of
extracted features.

- 108 -
 5.3 Classification

Figure 5.32: 2D Classification of punctured and contaminated insulators using

Combination ID 92.

Figure 5.33: Eliminated sets of features during the gating algorithm.

- 109 -
 5.3 Classification

5.3.4 Analysis of Gated Classification Results

Statistical processing on the captured electromagnetic radiation due to partial

discharge activities from normal, contaminated, and punctured insulators shows
that there are unique features that can separate the radiation signature of defective
units. The main challenge in separating the features of defective units is the common
normal partial discharge activities that are present on all three types of insulators.
As shown in Figure 5.29, the high voltage stress near the bottom metallic fitting
of porcelain insulators constantly causes discharge activities. Similar discharge
activities on normal and clean porcelain insulators have been also reported in other
publications [104]. The gating algorithm explained in the previous section eliminates
the observations that are suspected to be related to normal partial discharge
activities. After applying the gating algorithm, the 2D classification of punctured
and contaminated insulators had a significantly higher successful classification rate,
highlighting unique statistical features of each defect.
As shown in Figure 5.32, the signatures of punctured and contaminated insulators
in Domain 92 have separable features. To explain the mapping and features of
Domain 92, four sample Weibull distributions are presented in Figure 5.34. The
four Weibull distributions plots shown in Figure 5.34 represent the four highlighted
area shown in Figure 5.35.
Regions 1, 2, and 3 in Figure 5.35 correspond to puncture feature signatures,
representing discharge activities that have higher negative cycle phase occurrences
compared to Region 4 which relates to the signature of contaminated insulators.
The higher phase angles (closer to 270 degrees) on radiated signature of punctured
insulators indicate that the discharge requires higher electric stress to bridge the

- 110 -
 5.3 Classification

insulation. On the other hand, the discharge activities on the contaminated

insulators occur in lower phase angles, corresponding to lower required voltage.
When the recorded puncture observations shown in Figure 5.35 were traced back
to the measurement logbook, it was observed that the Region 1 points mostly relate
to punctured insulators that were tested right after they were made in the lab.
Similarly, the features in Region 3 mostly relate to the measurements that were
recorded in the last month of the lab work. By inspecting the sample distribution
plots shown in Figure 5.34, it can be concluded that the puncture path characteristics
change when the puncture channel has been under continuous electric stress. The
partial discharge activities on the early stages of puncture path (Region 1) are
mainly concentrated in higher negative cycle phase angles. As the puncture path
experiences continuous discharge activities, a wider phase range of discharges will
occur and the centre phase shifts to lower phase angles. This behaviour should be
related to the chemical reactions within the puncture track while the unit is kept
under electric stress which eventually changes the characteristic of the puncture
channel.
In conclusion, the statistical analysis performed on the radiation signature
of punctured and contaminated insulators indicate separable features in the
distribution of partial discharge activities. These features relate to the partial
discharge activities in the negative half cycle of the voltage.

- 111 -
 5.3 Classification

Figure 5.34: Four sample Weibull distribution functions.

Figure 5.35: Four clusters in 2D classification using Domain ID 92.

- 112 -
 5.4 Summary

5.4 Summary

The signal processing algorithms for extracting the features from the captured
electromagnetic radiations were explained in this chapter. Also scatter plots showing
the results of feature extracting algorithm on all three types of insulators under test
were presented. After extracting 17 features from each type of insulator, the SVM
classifier was programmed to find the best feature domain to separate the radiation
signatures. Difficulties in classifying the radiation signatures were improved by
introducing a gating algorithm, using the common discharge activities as a baseline
to unmask the unique features of each defect.
Prior to implementing the gating algorithm, the classifier’s successful
classification rate in differentiating contaminated and punctured insulators was
about 51 %. After implementing the gating algorithm, the successful classification
rate increased to about 75 %.

- 113 -
6

Conclusions and Future Work

6.1 Conclusions

Physical defects such as punctures and contamination can degrade the insulators
performance and result in power outages, potentiating costs to utilities. Therefore,
condition assessment of line insulators has always been one of the most important
aspects of maintenance programs in power networks. A novel approach for inspecting
the condition of porcelain insulators based on statistical analysis of electromagnetic
radiations of live insulators was proposed in this thesis. The methodology presented
in this thesis is based on recording the electromagnetic radiations from porcelain
insulators using a wideband electric field sensor (the D-dot sensor) and performing
statistical analysis and classification.
During the preliminary discussions with the Manitoba Hydro Field Testing
Group, contamination and punctures were identified as the two most common types
of defects on porcelain insulators. Punctured insulators were created in the lab
by applying high voltage stresses on the insulators submerged in oil, following the

- 114 -
 6.1 Conclusions

puncture withstand test specified in the IEC standard [89]. Contaminated insulators
were developed in the lab following the IEC standards [20] using the concept of
Equivalent Salt Deposit Density (ESDD). The contaminated insulators studied in
this research with the ESDD level of 0.28 mg/cm2 , can be categorized as heavily
contaminated insulators according to the contamination level definitions provided
in the IEEE standard 1313.2 [87]. To duplicate realistic environmental conditions,
a fog chamber was built and included in the measurement setup. The fog chamber
helped to control the level of humidity when the insulators were tested.
During the experimental part of this thesis, a multitude of tests were conducted
and resulted in a large amount of lab measurements. A total of 410,000 cycles of
discharge activities were measured and recorded. From the total of 410,000 recorded
measurements, 90,000 cycles (176 datasets) of recorded radiations originated from
normal insulators, 175,000 cycles (342 datasets) from punctured insulators, and
145,000 cycles (288 datasets) from contaminated insulators. To analyse the raw
measurement data obtained throughout testing, a MATLAB graphical user interface
(GUI) capable of performing various data processing techniques was developed. Due
to the random nature of partial discharges, the data processing was focused on
statistical features of the recorded measurements.
A total of 17 features were identified that were mainly focused on the phased
resolved pattern of discharges and were extracted from each dataset of recorded
measurements. Weibull distribution functions were fit to the histogram of the
raw data to extract the trends of discharge patterns of each type of insulators.
Even normal and healthy insulators under high voltage stresses experience partial
discharge activities, mainly near the bottom metallic fitting (see the corona camera
image shown in Figure 5.29). The signature of these partial discharge activities

- 115 -
 6.1 Conclusions

are not only present in the leakage current (ground wire current) but also in
the radiated electromagnetic fields (see Figure 4.4). To overcome this issue, a
gating algorithm was developed to eliminate the signature of normal, over-stressed
partial discharge activities. Also, a SVM classifier was developed to differentiate
the radiated signature of punctured and contaminated insulators from healthy
functioning insulators. The classification has shown promising detection rates,
confirming the classification ability of the novel remote assessment tool proposed in
this thesis. The statistical analysis of radiated signature of defective insulators shows
the partial discharge activities on the punctured insulators have higher negative cycle
phase occurrences compared to the contaminated insulators. The higher phase
angles (closer to 270 degrees) indicate that the discharges require more electric
stress to bridge the insulation in the case of punctured insulators. On the other
hand, the discharge activities on the contaminated insulators occur in lower phase
angles, indicating lower required voltages for discharges to occur. In conclusion, the
shape parameter of Weibull distribution fit to the weighted phase location of partial
discharges has unique information in the negative half cycle, which can be used for
the classification of punctured and contaminated insulators.

- 116 -
 6.2 Main Contributions

6.2 Main Contributions

The achievements and main contributions of this thesis are summarized as follows.

ˆ Introduction of a new porcelain insulator inspection technique that can

remotely identify normal, punctured, and contaminated porcelain insulators

using their electromagnetic radiation signature in a laboratory environment,

ˆ Demonstration of the ability and performance of the novel idea of utilizing

an electric field sensor to capture partial discharge radiations from defective

insulators. The use of D-Dot sensors in PD detection on line insulators has
not been reported in literature or industry,

ˆ Development of unique statistical feature extraction that can reveal

distinguishable information within each of the defective insulators studied in

the thesis,

ˆ Capturing and recording a total of 410,000 cycles of discharge activities during

the lab measurements that resulted in developing a meaningful population of

data to support the statistical feature extraction algorithm,

ˆ Development a 2D Support Vector Machine (SVM) classifier to differentiate

the signature of defects in various 2D feature domains and identifying the

feature domains with the maximum distinct information,

ˆ Overcoming the problem of overlapped sections of extracted features in the 2D

feature domains which led to identifying the source of a common signature;

normal discharge activities near the bottom pin of the insulators. Knowing
the source of common signature, a novel gating algorithm was developed based

- 117 -
 6.2 Main Contributions

on the extracted features of normal insulators. The performance of the SVM

classifier improved significantly after implementing the gating algorithm.

- 118 -
 6.3 Future Work

6.3 Future Work

The work described in this thesis is based on measurements recorded in a controlled

lab environment. The next steps for continuing this work involve data acquisition
from a string of insulators in service. To ensure the existence of a defective insulator
on the string, close contact methods (such as DC resistance meter or buzz method)
can be used. When the defective unit is identified on the string, a similar setup can
be used to record the electromagnetic radiations. Measurements from normal and
healthy insulators should also be taken to identify the baseline discharge activities
of normal insulators.
Field implementation of inspection techniques similar to the one presented in
this research can be quite challenging. The measurements results reported in
this research were obtained in a controlled lab environment with very low levels
of environmental noise. As part of furthering this research, noise cancellation
algorithms can be developed to improve the signal to noise ratio on the captured
radiations.
Furthermore, during the span of 6 months of lab measurements, the discharge
patterns of punctured insulators exhibited correlations with the age of puncture
tracks. The results indicate that the properties of the puncture track change when
it is exposed to continuous discharge activities. Studying the puncture track ageing
behaviour and investigating the effect of discharge activities on the overall insulation
properties of the porcelain insulator can be helpful in understanding the effect of
puncture track properties on the discharge signature, another development initiated
by this research.

- 119 -
Appendix A

Partial Discharges under AC

Voltage

A.1 Definitions

By definition, partial discharge is an electrical discharge that occurs over a portion

of the dielectric between two high voltage electrodes and does not bridge the whole
gap [105]. Partial discharges may occur in cavities within solid insulation, on the
surface of insulating material or around the sharp edges of high voltage electrodes.
The term corona is often used for the partial discharge around the sharp points
which manifests in gaseous media around conductors and remote from solid or liquid
insulation. “Corona (in air) is a luminous discharge due to ionization of the air
surrounding a conductor caused by a voltage gradient exceeding a certain critical
value” [88]. The term “Flashover” is used when a surface discharge bridges the
insulation. However, these terms are used interchangeably in high voltage articles
and textbooks. Other than the terms that were mentioned, terms such as “local

- 120 -
 A.2 Partial discharge process in voids under AC conditions

breakdown”, “sparking”, “streamer”, “partial breakdown” and “ swarming PD”

have been used in many papers to describe different types of partial discharge
phenomena concerning the discharge ionization mechanism or discharge location
[92]. In this report, the term “partial discharge” is defined based on IEC standards.
According to the IEC 60270[101], “partial discharge is a localized electrical discharge
that only partially bridges the insulation between conductors which can or cannot
occur adjacent to a conductor. Partial discharges are in general a consequence
of local electrical stress concentrations in the insulation or on the surface of the
insulation. Generally, such discharges appear as pulses having a duration of much
less than 1 picosecond. More continuous forms can, however, occur, such as the
so-called pulseless discharges in gaseous dielectrics. Partial discharges are often
accompanied by emission of sound, light, heat, and chemical reactions.”

A.2 Partial discharge process in voids under AC

conditions

One of the most common reasons for an insulator’s failure is the presence of voids
inside the insulation material. Depending on the type of insulation material and
also the manufacturing process, different shaped voids can develop inside insulation
systems. Having a void inside the insulation does not necessarily mean that the
insulation will fail over time; it highly depends on the level of partial discharge
activities inside that void under the normal voltage stress. In this section, the partial
discharge process inside voids will be discussed briefly. Voids inside the insulating
material are always under higher electrical stress compared to the surrounding

- 121 -
 A.2 Partial discharge process in voids under AC conditions

insulation. Consider a flat void as shown in figure A.1.(a):

High Voltage Electrode High Voltage Electrode

Spherical
Void Void

d1
d d
Є' d0
d2 Є'

High Voltage Electrode High Voltage Electrode

(a) (b)

Figure A.1: Voids inside insulation system.

(a) Rectangular void
(b) Spherical void

If E is the average electric field inside the insulation and Eb is the breakdown
voltage of the void, the electric field required to cause the breakdown inside the void
is given by [16]:

d1 + d2
E = Eb + Ed = Eb (1 + ) (A.1)
d′

Where Ed is the voltage drop across the dielectric layers surrounding the void.
In other words, in order for breakdown to occue, the void needs at least Eb across
its walls. Now to have Eb across the void, the electric field E has to be across the
whole gap. The electric field inside the void is a function of the void geometry. As
shown in Figure A.1.(b), most of the actual voids in the insulation systems have
approximately spherical shapes. The equation A.1 should be slightly modified in
order to properly formulate the required electric field for the gap breakdown of
spherical voids. Equation A.2 which is given by [106] can be used for the case of

- 122 -
 A.2 Partial discharge process in voids under AC conditions

spherical voids.
Eb d0 (1 + 2′ )
E= [d + ] (A.2)
d 3′

Where d0 is the total thickness of the dielectric and d is the diameter of the
void. There has been substantial work on calculating Eb for different geometries
of voids. It has been shown that Eb is a function of the product of pressure and
gap separation (pd) and also involves Paschen’s law. Measurements were done
for different geometries and the graphs are available to relate the required Eb
for specific pd in each case [107]. The breakdown voltage also depends on the
dielectric constant of the material inside the void, but in the most of practical
cases voids are filled by air. The breakdown mechanism of metallic surfaced gaps
under uniform electric field stresses can be interpreted by the Townsend discharge
process. The breakdown mechanism can be explained using Townsend coefficients
and the concept of electron avalanche [17]. However, this mechanism cannot be
easily extended to the breakdown inside the voids as they have dielectric walls
compared to the gap. For example, Townsend’s second coefficient which measures
the electron regeneration probability highly depends on the cathode and anode
materials. Moreover, the electric field inside the void is more distorted compared
to metallic surfaced gaps. Some studies have tried explaining the breakdown
mechanism using streamer discharge theory [18, 108]. According to this theory,
the high electric field stress will accelerate the electrons and these fast electrons will
give rise to other electrons on their way. Ions move at much slower speeds compared
to electrons, which is how the streamer forms inside the gap. When there are enough
high speed free electrons on the streamer, they bridge the gap and the breakdown
happens. When the streamer completely bridges the gap, a conductive plasma

- 123 -
 A.2 Partial discharge process in voids under AC conditions

channel is developed inside the gap and the spark can be observed. In summary, the
two types of discharge that can occur in metallic surfaced gaps are the Townsend’s
discharge and streamer discharge. The detected discharge resulting from Townsend’s
discharge is faster with lower amplitude compared to streamer’s discharge. On a
metallic surface gap, by increasing the voltage across the gap, Townsend’s discharge
will be observed first and as the voltage increases and the channel becomes ionized,
the streamer discharge can be observed. Now to extend these theories to discharge
mechanisms inside voids, we have to consider some differences:

ˆ The electric field distribution inside the void is not only dependent on the

voltage across the electrodes, but also on the surface charge distribution on
the dielectric boundaries.

ˆ When a discharge occurs on metallic surfaced gaps, the total charge will be

discharged into the metallic electrodes. However, due to the relatively high
resistance of the electrodes, in the dielectric surfaced gaps, there could be more
than one discharge happening at the same time to discharge all the electrons.

ˆ When a discharge happens on a specific point of the surface of a dielectric

surfaced gap, the charge distribution on other parts of the surface remains
unchanged. This is mainly due to the immobility of surface charges.

To investigate the discharge mechanism in voids inside dielectrics, some experiments

were done on a test setup that involved one metallic electrode and one dielectric
electrode [109]. In case of DC voltage supply, different discharge patterns were
recorded depending on the polarity of the dielectric electrode [110]. When the AC
voltage was applied, a combination of different patterns were observed and collected

- 124 -
 A.3 Equivalent electric circuit for voids inside insulating material

by changing the polarity of the DC voltage in the DC test setup. One certain
conclusion that was determined by these experiments was the fact that more than
one discharge is needed to entirely discharge the capacitance of the physical voids
inside the insulation. This means that the discharge amplitude in dielectric surfaced
gaps is lower compared to the metallic surfaced gaps.

A.3 Equivalent electric circuit for voids inside

insulating material

Knowing the electrical equivalent circuit of voids can be very helpful in

understanding the discharge parameters such as discharge waveform, rise time,
amplitude, etc. One of the first experiments that has been reported on this topic
consists of a very simple test setup with a capacitor in series with a metallic gap while
the oscilloscope is connected across the gap [111]. Four discharges were recorded per
cycle when the applied voltage across the gap was larger than the discharge onset
voltage. Based on this experiment, a very simple circuit model was proposed which
is shown in Figure A.2.

Dielectric Void
CS

CP
EO

CV Ea

Figure A.2: Circuit model for a void inside the insulation system

- 125 -
 A.3 Equivalent electric circuit for voids inside insulating material

In this circuit, CV represents the capacitance of void, CS represents the total

capacitance in series with the void and CP represents the remaining capacitance
which is in parallel with CV and CS in series. Although the circuit is very simplified
and totally neglects the surface leakage effect on the dielectric walls, it still can be
used to explain the basic discharge mechanism in voids. If Eb is the breakdown
voltage of the void, when there is enough voltage across the dielectric gap that
provides Eb across the void, the breakdown happens. When the breakdown occurs,
the voltage across the void drops significantly to a new voltage that is called Er .
The difference between these two voltages is ∆E(∆E = Eb − Er ). In order to charge
up the void capacitance again to the voltage Eb , charge ∆Q is needed to be drawn
from CP

∆ECV CS
∆Q = (A.3)
CV + C S

In most structure, CS ≪ CV ≪ CP , so the above equation can be simplified to

A.4:

∆Q = ∆E × CS (A.4)

So ∆Q is a good measure for discharge magnitude as it directly relates to ∆E.

Now we can calculate ∆W ,the average power dissipated in each discharge, using
the approximation given by [112]:

1
∆W = CV ∆E (Eb − ∆E) (A.5)
2

When the void is completely discharged (Er = 0 ), the discharge power will be

- 126 -
 A.4 Discharge waveform of metallic surfaced gap under AC supply

simplified into:

1
∆W = CV [∆E]
2
(A.6)
2

Discharge power is a very important parameter as it directly relates to the rate

of degradation in insulation systems. However in practical applications, CV and ∆E
can’t be found easily.

A.4 Discharge waveform of metallic surfaced gap

under AC supply

When a void is under AC voltage stress that is higher than its breakdown voltage,
the discharge will occur anytime that the voltage exceeds the breakdown voltage.
Because the AC signal repeats its pattern every cycle and the fact that discharge
happens at a very fast time frame compared to the applied voltage, studying just one
cycle would be sufficient. For simplicity let’s assume that there is a metallic gap with
breakdown voltage of Eb and the residual voltage is zero (Er = 0). The “apparent
voltage” is defined as the voltage that would appear across the void assuming the
breakdown voltage of the void to be infinity. Using this definition, the peak of
apparent voltage across the void using the circuit model in Figure A.2 would be:

E o CS
Ea = (A.7)
CS + CV

Now considering that the breakdown voltage is equal in both polarities of the
applied voltage and the time that discharge happens is negligible compared to the
60 Hz cycle, we are going to investigate the case where the applied voltage is 1.5

- 127 -
 A.4 Discharge waveform of metallic surfaced gap under AC supply

times the breakdown voltage. The gap we are studying here is an ideal gap with
metallic electrodes which means that the applied voltage is the apparent voltage
that appears on the gap. As it is shown in Figure A.3, when the voltage across the
gap reaches the breakdown voltage, the discharge happens and the voltage inside
the gap drops to Er which in this case is zero. The sudden drop of voltage across
CV will cause a sudden jump of the voltage across the CS . After the discharge
extinguishes, the voltage inside the gap starts increasing again but this time it will
not reach the Eb at the positive half cycle. During the negative half cycle however,
the voltage reaches the −Eb and that is when the second breakdown happens.

Ea

1.5Eb Apparent voltage

on the gap

Eb
Actual voltage
across the gap

-Eb

-1.5Eb

Figure A.3: Voltage waveform across an ideal cavity with Er = 0 and Ea = 1.5 Eb

It should be noted that the discharge time duration in the Figures was assumed
to be zero as it is very small compared to the 60 Hz cycle. Figure A.3 also tells
us that when apparent voltage is 1.5 times bigger than the breakdown voltage, four

- 128 -
 A.4 Discharge waveform of metallic surfaced gap under AC supply

discharges will happen in one cycle. Now when the apparent voltage is increased
by an integer number of the breakdown voltage, the number of discharges will be
multiplied by that number, for example Figure A.4 explains when the apparent
voltage is three times the breakdown voltage where twelve discharges will occur.

Ea

3Eb

Apparent voltage
on the gap

Actual voltage
across the gap

Eb

-Eb

-3Eb

Figure A.4: Voltage waveform across an ideal cavity with Er = 0 and Ea = 3 Eb

Although we assumed that breakdown voltage is equal in both polarities, in

actual experiments, it was shown that they are not equal and that is why the
discharge pattern is different for each half cycle. The residual voltage may also
vary in each polarity depending on the ionization inside the void. In order to model
the discharge pattern more appropriately, unequal residual voltage and breakdown
voltage in different polarities and even in the same polarity should be taken into
account.

- 129 -
 A.5 Discharge waveform of dielectric surfaced gap under AC supply

A.5 Discharge waveform of dielectric surfaced

gap under AC supply

Studying the discharge mechanism of dielectric surfaced gaps is more complicated

than metallic electrode gaps that was discussed in the previous section. The applied
voltage on the whole gap is the voltage that is across the gaseous gap in series with
the dielectric (CS and CV in series) therefore, the exact voltage on the gaseous gap
and the dielectric layer is not available. To study the discharge waveform for this
case, an experimental setup given by [113] was used which is shown in Figure A.5.

HV Metallic electrode

1 in Diameter
Adjustable
gaseous gap

Dielectric Layer

2 in Diameter

0.017 in
0.25 in

Grounded metallic electrode

Figure A.5: Experimental test setup for studying the voltage waveform of dielectric-
metallic electrode

The setup has two metallic electrodes where a dielectric layer is molded on the
ground electrode. The details of this process is available in [114] as they followed
specific procedure to avoid the development of voids and flaws between this layer
and the metallic electrode. The other metallic electrode can be adjusted to have the
desired gaseous gap size between the electrodes. In order to compare the results with

- 130 -
 A.5 Discharge waveform of dielectric surfaced gap under AC supply

the case where both electrodes were metallic, in some experiments, a Tin film layer
was used on top of the dielectric layer and the gap size was adjusted considering
the thickness of this layer. To measure the voltage in this setup, they put the whole
structure in series with a discharge free capacitor and then the whole combination
was placed in parallel with another discharge free capacitor. The equivalent circuit
for this new test setup is given in [113] and is shown in Figure A.6:

CS

CP CV

Cd

Figure A.6: Circuit model for the experimental test shown in Figure A.5

Where CV represents the gaseous gap between the dielectric layer and the
metallic electrode and was filled by Helium in this experiment. Cd represents the
capacitance of the dielectric layer molded on the ground electrode and Cs and CP are
the discharge free capacitors in series and parallel with the structure for measurement
purposes. Figure A.7 explains the shape of waveforms considering the discharges.
In this test setup the apparent voltage across the gaseous gap was assumed to be
3Eb . The actual voltage across the dielectric layer can be found by superposition
of two voltages. The first portion can be found by a voltage division between the
gap and the dielectric layer from the total voltage across the electrodes. The second

- 131 -
 A.5 Discharge waveform of dielectric surfaced gap under AC supply

portion is a result of the current due to discharges in the gaseous gap. As the gap is
in series with the dielectric layer, the ∆Q caused by the gap discharge is equal to:

∆Q = CV × ∆V1 = Cd × ∆V2 (A.8)

But because CV and Cd are unknown, the ratio (∆V1 /∆V2 ) is unknown as well.
∆V2 is the voltage that will appear across the dielectric layer due to discharges
happening inside the gap. For illustrative purposes ∆V2 was set to 31 Eb . Following
this assumption, the first portion of the voltage across the dielectric layer would
be equal to one third of the voltage across the gaseous gap. This means that the
apparent voltage across the dielectric layer is equal to Eb . Figure A.7(a) shows the
breakdown in the gaseous gap which is similar to Figure A.4 that was discussed
earlier. Discharges happening in the gaseous gap will result in a discharge current
that will change the actual voltage across the dielectric layer. This voltage was
named ∆V2 and Figure A.7(b) shows how the voltage across the dielectric layer
changes due to the discharge current. The total voltage across the electrodes can
be found by adding the breakdown voltage across the gaseous gap and the voltage
across the dielectric layer which is shown in Figure A.7(c). Although Figure A.7(c)
shows the waveform across the electrodes for a very ideal case, it has shown a good
analogy with some practical experiments [92]. It should be mentioned that in all
the waveforms shown in Figure A.7, the residual voltage was considered to be zero.
Also it was assumed that there is only one discharge happening at the time. In real
experiments however, there could be more than one discharge site at a time that
changes the timing between two consecutive discharges. To compare the dielectric
electrode with the metallic electrode, the Tin film layer was used on the dielectric

- 132 -
 A.5 Discharge waveform of dielectric surfaced gap under AC supply

Ea
3Eb
Voltage across the gap in
the absence of breakdown

Actual voltage across the gap

under breakdown conditions
(a) Eb

-Eb

2Eb Voltage across the dielectric layer

∆V2=1/3Eb in the absence of breakdown

Actual voltage across the dielectric

(b) Eb
layer under breakdown conditions

Eb-∆V2
2Eb
Actual voltage across the electodes
under breakdown conditions

(c) Eb

Eb-∆V2

Figure A.7: Experimental test setup for studying the voltage waveform of dielectric-
metallic electrode

- 133 -
 A.5 Discharge waveform of dielectric surfaced gap under AC supply

layer and the results are reported in [114]. There has also been some research on the
effect of different gases inside the gap on the discharge behavior and also pseudoglow
and glow discharge mechanism in the voids under certain conditions [115].

- 134 -
References

[1] E. Cherney, A. Baker, J. Kuffel, Z. Lodi, and D. S. G. Phillips, A.and Powell,

“Evaluation of and replacement strategies for aged high voltage porcelain
suspension-type insulators,” IEEE Transactions on Power Delivery, 2014. 1

[2] M. Farzaneh and W. Chisholm, Insulators for Icing and Polluted

Environments, ser. IEEE Press Series on Power Engineering. Wiley, 2009. 3

[3] E. Report, “Stress performance of porcelain and glass transmission line

insulators: Mechanical and electrical (m&e) and flashover testing results,”
Tech. Rep., 2014. 3

[4] E. Azordegan, “Group meeting communication with manitoba hydro field

testig group,” 2011. 3, 20

[5] A. Banik, S. Dalai, and B. Chatterjee, “Condition monitoring of overhead line

insulator by measuring surface leakage current,” in 2014 Annual IEEE India
Conference (INDICON), Dec 2014, pp. 1–5. 4, 12, 25

[6] Z. Jia, C. Chen, X. Wang, H. Lu, C. Yang, and T. Li, “Leakage current
analysis on rtv coated porcelain insulators during long term fog experiments,”

- 135 -
 REFERENCES

IEEE Transactions on Dielectrics and Electrical Insulation, vol. 21, no. 4, pp.
1547–1553, August 2014. 4, 25

[7] J. Li, W. Sima, C. Sun, and S. Sebo, “Use of leakage currents of insulators to
determine the stage characteristics of the flashover process and contamination
level prediction,” IEEE Transactions on Dielectrics and Electrical Insulation,
vol. 17, no. 2, pp. 490–501, 2010. 4, 26

[8] S. Chandrasekar, C. Kalaivanan, A. Cavallini, and G. Montanari,

“Investigations on leakage current and phase angle characteristics of porcelain
and polymeric insulator under contaminated conditions,” IEEE Transactions
on Dielectrics and Electrical Insulation, vol. 16, no. 2, pp. 574–583, 2009. 4,
12, 27, 28

[9] E. Azordegan, B. Kordi, and D. Swatek, “Radiated electromagnetic field

signature of faulty and polluted porcelain insulators,” in 2010 International
Conference on High Voltage Engineering and Application (ICHVE), Oct 2010,
pp. 449–452. 5

[10] E. Azordegan, D. R. Swatek, and B. Kordi, “Radiated electromagnetic field

signature of cracked and polluted porcelain insulators,” Gaodianya Jishu
Journal of High Voltage Engineering, vol. 37, no. 11, pp. 2649 – 2654, 2011.
5, 41

[11] “Ngk-locke catalog.” [Online]. Available: http://www.ngk.co.jp/english/

products/power/porcelain/suspension/catalogs.html 7

[12] R. Gorur, E. Cherney, and J. Burnham, Outdoor insulators. Ravi S. Gorur,

1999. 7, 9, 19, 20

- 136 -
 REFERENCES

[13] T. Zhao and R. Bernstorf, “Ageing tests of polymeric housing materials for
non-ceramic insulators,” IEEE Electrical Insulation Magazine, vol. 14, no. 2,
pp. 26 –33, march-april 1998. 7

[14] J. Looms, Insulators for high voltages. Peter Peregrinus Ltd, 1988, vol. 7. 8,
15, 29

[15] E. Cherney, “Cement growth failure of porcelain suspension insulators,” IEEE

Transactions on Power Apparatus and Systems, no. 8, pp. 2765–2774, 1983. 8

[16] T. Dakin and D. Berg, “Progress in dielectrics,” Theory of Gas Breakdowns,

vol. 4, pp. 167–168, 1962. 10, 122

[17] A. Von Hippel, Dielectrics and waves, 1966. 10, 123

[18] J. Devins, “The mechanism of the formation of discharges limited by series

dielectrics,” in 1961 Annual Report Conference on Electrical Insulation, 1961,
p. 97. 10, 123

[19] Insulators for overhead lines with a nominal voltage above 1000 V, Std. IEC
60 383-1, 1993. 11

[20] Artificial pollution tests on high-voltage insulators to be used on a.c. systems,

Std. IEC 60 507, 1991. 11, 13, 46, 56, 57, 59, 115

[21] E. Cherney and R. Hooton, “Cement growth failure mechanism in porcelain

suspension insulators,” Power Delivery, IEEE Transactions on, vol. 2, no. 1,
pp. 249–255, Jan 1987. 12

[22] A. Mishra, R. Gorur, and S. Venkataraman, “Evaluation of porcelain and

toughened glass suspension insulators removed from service,” Dielectrics and

- 137 -
 REFERENCES

Electrical Insulation, IEEE Transactions on, vol. 15, no. 2, pp. 467–475, April
2008. 12

[23] K. Morita, T. Imakoma, and K. Arakawa, “Long-term performance of ceramic

insulators,” NGK Publication, Tech. Rep. 12

[24] J. Li, C. Sun, and S. Sebo, “Humidity and contamination severity impact on
the leakage currents of porcelain insulators,” IET Generation, Transmission
& Distribution, vol. 5, no. 1, pp. 19–28, 2011. 12, 25, 27, 38, 42

[25] L. Grigsby, The electric power engineering handbook. CRC Pr I Llc, 2001. 13

[26] K. Naito, Y. Mizuno, and W. Naganawa, “A study on probabilistic assessment

of contamination flashover of high voltage insulator,” Power Delivery, IEEE
Transactions on, vol. 10, no. 3, pp. 1378–1384, Jul 1995. 13

[27] Selection and dimensioning of high-voltage insulators for polluted conditions,

Std. IEC 60 815, 2002. 14, 46, 47

[28] D. Williams, A. Haddad, A. Rowlands, H. Young, and R. Waters, “Formation

and characterization of dry bands in clean fog on polluted insulators,” IEEE
Transactions on Dielectrics and Electrical Insulation, vol. 6, no. 5, pp. 724–
731, 1999. 14, 16

[29] B. Hampton, “Flashover mechanism of polluted insulation,” Electronics and

Power, vol. 10, no. 4, p. 113, 1964. 15

[30] “Industrial energy use in canada - emerging trends - energy briefing

note,” 2008. [Online]. Available: https://www.neb-one.gc.ca/nrg/ntgrtd/
mrkt/archive/2010ndstrlnrgscnd/ndstrlnrgscnd-eng.html 18

- 138 -
 REFERENCES

[31] L. Lawton, M. Sullivan, K. V. Liere, and A. Katz, “A framework and review

of customer outage costs: Integration and analysis of electric utility outage
cost surveys,” Ernest Orlando Lawrence Berkeley National Laboratory, Tech.
Rep., November 2003. 18

[32] G. H. Vaillancourt, M. St-Jean, J. P. Bellerive, and C. Jean, “New live line

tester for porcelain suspension insulators on high-voltage power lines,” IEEE
Transactions on Power Delivery, vol. 9, pp. 208–219, 1994. 20

[33] Y. Cheng, C. Li, and X. Huang, “Study of corona discharge pattern on high
voltage transmission lines for inspecting faulty porcelain insulators,” IEEE
Transactions on Power Delivery, vol. 23, no. 2, pp. 945 –952, april 2008. 20

[34] M. Farzaneh, “Outdoor insulators: overview of in-service experience,

inspection practice and future challenges,” in IEEE Electrical Insulation
Conference, EIC 2009, May 2009, pp. 542–550. 21

[35] B. Qi, C. Li, Z. Hao, B. Geng, D. Xu, S. Liu, and C. Deng, “Partial discharge
detection for gis: A comparison between uhf and acoustic methods,” in
Conference Record of the 2010 IEEE International Symposium on Electrical
Insulation (ISEI). IEEE, 2010, pp. 1–5. 22

[36] M. Yaacob, M. Alsaedi, J. Rashed, A. Dakhil, and S. Atyah, “Review on partial

discharge detection techniques related to high voltage power equipment using
different sensors,” Photonic Sensors, vol. 4, no. 4, pp. 325–337, 2014. 22

[37] K. Wu, C. Pan, Y. Meng, C. Sun, M. Gao, K. Qin, and H. Long, “Study on the
characteristics of partial discharges in voids under square voltage by detecting

- 139 -
 REFERENCES

light emission intensity,” IEEE Transactions on Dielectrics and Electrical

Insulation, vol. 18, no. 5, pp. 1651–1657, 2011. 22

[38] F. Bologna, N. Mahatho, and D. Hoch, “Infra-red and ultra-violet imaging

techniques applied to the inspection of outdoor transmission voltage
insulators,” in IEEE AFRICON. 6th Africon Conference in Africa, vol. 2.
IEEE, 2002, pp. 593–598. 23

[39] H. Ha, S. Han, and J. Lee, “Fault detection on transmission lines using
a microphone array and an infrared thermal imaging camera,” IEEE
Transactions on Instrumentation and Measurement, no. 99, pp. 1–9, 2012.
23

[40] I. Ramirez-Vazquez and J. Fierro-Chavez, “Criteria for the diagnostic

of polluted ceramic insulators based on the leakage current monitoring
technique,” IEEE Conference on Electrical Insulation and Dielectric
Phenomena, pp. 715–718, 1999. 24

[41] G. Amaldo and F. Geraldo, “Leakage current monitoring of insulators exposed

to marine and industrial pollution,” IEEE International Symposium on
Electrical Insulation, pp. 271–274, 1996. 24

[42] T. Suda and S. Member, “Frequency characteristics of leakage current

waveforms of a string of suspension insulators,” IEEE Transactions on Power
Delivery, vol. 20, no. 1, pp. 481–488, 2005. 24

[43] J. Li, W. Sima, C. Sun, Z. Li, C. Yang, and J. Deng, “Frequency domain
spectrum of leakage currents during discharge detection for contamination
insulators [j],” High Voltage Engineering, vol. 3, 2008. 26

- 140 -
 REFERENCES

[44] S. Kumagai and N. Yoshimura, “Leakage current characterization for

estimating the conditions of ceramic and polymeric insulating surfaces,” IEEE
Transactions on Dielectrics and Electrical Insulation, vol. 11, no. 4, pp. 681–
690, 2004. 27

[45] H. Okubo and N. Hayakawa, “A novel technique for partial discharge and
breakdown investigation based on current pulse waveform analysis,” IEEE
Transactions on Dielectrics and Electrical Insulation, vol. 12, no. 4, pp. 736–
744, 2005. 27

[46] M. Douar, A. Mekhaldi, and M. Bouzidi, “Investigations on leakage current

and voltage waveforms for pollution level monitoring under wetted and
contaminated conditions,” IET Science, Measurement & Technology, vol. 5,
no. 2, pp. 67–75, 2011. 29

[47] V. Kontargyri, I. Gonos, I. Stathopulos, and A. Michaelides, “Measurement

and verification of the voltage distribution on high voltage insulators,”
Proceedings of the 12th Biennial IEEE Conference on Electromagnetic Field
Computation, pp. –, 2006. 30, 31

[48] J. Rasolonjanahary, L. Krahenbuhl, and A. Nicolas, “Computation of electric

fields and potential on polluted insulators using a boundary element method,”
IEEE Transactions on Magnetics, vol. 28, no. 2, pp. 1473–1476, 1992. 30

[49] T. Zhao and M. Comber, “Calculation of electric field and potential

distribution along nonceramic insulators considering the effects of conductors
and transmission towers,” IEEE Transactions on Power Delivery, vol. 15,
no. 1, pp. 313–318, 2000. 30

- 141 -
 REFERENCES

[50] V. Kontargyri, I. Gonos, I. Stathopulos, and A. Michaelides, “Calculation of

the electric field on an insulator string using the finite elements method,”
in Proceedings of the 38th International Universities Power Engineering
Conference (UPEC 2003), 2003, pp. 65–68. 30

[51] J. Du, Z. Peng, J. Li, S. Zhang, N. Li, and C. Fan, “Electric field calculation
and grading ring optimization for 1000 kv ac post porcelain insulator,” in 2013
IEEE International Conference on Solid Dielectrics (ICSD), June 2013, pp.
198–201. 30

[52] Y. Cheng and C. Li, “Online detecting composite insulators by two

dimensions electric field distribution,” in Conference Record of the 2006 IEEE
International Symposium on Electrical Insulation, june 2006, pp. 132 –135. 30

[53] X. Liu, D. Kasten, S. Sebo, R. Caldecott, D. Grosjean, and D. Schweickart,

“Partial discharge measurements-frequency related considerations,” in 2005
Annual Report Conference on Electrical Insulation and Dielectric Phenomena,
CEIDP ’05, oct. 2005, pp. 405 – 409. 31

[54] S. Xiao, P. Moore, M. Judd, and I. Portugues, “An investigation into

electromagnetic radiation due to partial discharges in high voltage equipment,”
in IEEE Power Engineering Society General Meeting. IEEE, 2007, pp. 1–7.
31

[55] P. Moore, I. Portugues, and I. Glover, “Remote diagnosis of overhead line

insulation defects,” in IEEE Power Engineering Society General Meeting.
IEEE, 2004, pp. 1831–1835. 32

- 142 -
 REFERENCES

[56] R. Sharkawy, R. Mangoubi, T. Abdel-Galil, M. Salama, and R. Bartnikas,

“SVM classification of contaminating particles in liquid dielectrics using
higher order statistics of electrical and acoustic pd measurements,” IEEE
Transactions on Dielectrics and Electrical Insulation, vol. 14, no. 3, pp. 669–
678, June 2007. 33, 43

[57] M. Cacciari, A. Contin, and G. Montanari, “Use of a mixed-weibull

distribution for the identification of pd phenomena [corrected version],” IEEE
Transactions on Dielectrics and Electrical Insulation, vol. 2, no. 6, pp. 1166–
1179, 1995. 33, 40, 41

[58] A. Contin, G. Montanari, and C. Ferraro, “Pd source recognition by weibull

processing of pulse height distributions,” IEEE Transactions on Dielectrics
and Electrical Insulation, vol. 7, no. 1, pp. 48–58, Feb 2000. 33, 41

[59] N. Sahoo, M. Salama, and R. Bartnikas, “Trends in partial discharge pattern

classification: a survey,” IEEE Transactions on Dielectrics and Electrical
Insulation, vol. 12, no. 2, pp. 248–264, 2005. 34

[60] R. González and J. Tou, Pattern recognition principles, 1974. 34

[61] H. Kranz, “Diagnosis of partial discharge signals using neural networks and
minimum distance classification,” IEEE Transactions on Electrical Insulation,
vol. 28, no. 6, pp. 1016–1024, 1993. 34, 40

[62] W. Ziomek, M. Reformat, and E. Kuffel, “Application of genetic algorithms

to pattern recognition of defects in gis,” IEEE Transactions on ielectrics and
Electrical Insulation, vol. 7, no. 2, pp. 161–168, 2000. 35, 36

- 143 -
 REFERENCES

[63] T. Anderson, An introduction to multivariate statistical analysis, ser.

Wiley series in probability and mathematical statistics: Probability and
mathematical statistics. Wiley, 1984. 36

[64] E. Gulski and A. Krivda, “Neural networks as a tool for recognition of partial
discharges,” IEEE Transactions on Electrical Insulation, vol. 28, no. 6, pp.
984–1001, 1993. 36, 40

[65] J. Hunter, L. Hao, P. Lewin, D. Evagorou, A. Kyprianou, and G. Georghiou,

“Comparison of two partial discharge classification methods,” in Conference
Record of the 2010 IEEE International Symposium on Electrical Insulation
(ISEI). IEEE, 2010, pp. 1–5. 37

[66] M. Rahman, R. Arora, and S. Srivastava, “Partial discharge classification

using principal component transformation,” in IEE Proceedings-Science,
Measurement and Technology, vol. 147, no. 1. IET, 2000, pp. 7–13. 38

[67] J. Li, C. Sun, W. Sima, and Q. Yang, “Stage pre-warning based on leakage
current characteristics before contamination flashover of porcelain and glass
insulators,” IET Generation, Transmission & Distribution, vol. 3, no. 7, pp.
605–615, 2009. 38, 42

[68] M. Salama and R. Bartnikas, “Determination of neural-network topology for

partial discharge pulse pattern recognition,” IEEE Transactions on Neural
Networks, vol. 13, no. 2, pp. 446–456, 2002. 39

[69] D. Evagorou, A. Kyprianou, P. Lewin, A. Stavrou, V. Efthymiou, A. Metaxas,

and G. Georghiou, “Feature extraction of partial discharge signals using

- 144 -
 REFERENCES

the wavelet packet transform and classification with a probabilistic neural

network,” IET Science, Measurement & Technology, vol. 4, no. 3, pp. 177–
192, 2010. 39

[70] D. Evagorou, A. Kyprianou, P. Lewin, A. Stavrou, V. Efthymiou, and

G. Georghiou, “Classification of partial discharge signals using probabilistic
neural network,” in IEEE International Conference on Solid Dielectrics,
ICSD’07. IEEE, 2007, pp. 609–615. 39

[71] A. Contin, A. Cavallini, G. Montanari, G. Pasini, and F. Puletti, “Digital

detection and fuzzy classification of partial discharge signals,” IEEE
Transactions on Dielectrics and Electrical Insulation, vol. 9, no. 3, pp. 335–
348, 2002. 39

[72] R. James and B. Phung, “Development of computer-based measurements and

their application to pd pattern analysis,” IEEE Transactions on Dielectrics
and Electrical Insulation, vol. 2, no. 5, pp. 838–856, 1995. 40

[73] T. Hucker and H. Krantz, “Requirements of automated pd diagnosis systems

for fault identification in noisy conditions,” IEEE Transactions on Dielectrics
and Electrical Insulation, vol. 2, no. 4, pp. 544–556, 1995. 40

[74] M. Fenger, M. Credland, and H. Sedding, “Recent experiences with ac hipot

& pd commissioning testing of xlpe cable systems rated 69kv and above,”
CIGRE, 2012. 41

[75] OMICRON, HIGHVOLT, and Kinectrics, “On-site testing

of high-voltage cable systems.” [Online]. Available: https:

- 145 -
 REFERENCES

//www.omicron.at/fileadmin/user upload/pdf/magazine/MPD 600 On-site

Testing of High-voltage Cable Systems 2011 issue1.pdf 41, 107

[76] L. Dissado, J. Fothergill, S. Wolfe, and R. Hill, “Weibull statistics in dielectric

breakdown; theoretical basis, applications and implications,” Electrical
Insulation, IEEE Transactions on, no. 3, pp. 227–233, 1984. 41

[77] T. Sipahutar, A. Kemma, N. Pattanadech, F. Pratomosiwi, Suwarno, and

M. Muhr, “Effect of test method and needle plane configuration on partial
discharge inception voltage measurement of mineral oil based on weibull
analysis,” Procedia Technology, vol. 11, no. 0, pp. 411 – 418, 2013, 4th
International Conference on Electrical Engineering and Informatics, {ICEEI}
2013. 41

[78] Z. Tang, W. Zhou, J. Yu, and C. Zhou, “Comparison of weibull distribution

and crow-amsaa model used in cable failure analysis,” in 2012 Spring Congress
on Engineering and Technology (S-CET), May 2012, pp. 1–4. 41

[79] F.-C. Gu, H.-C. Chang, F.-H. Chen, and C.-C. Kuo, “Partial discharge pattern
recognition of power cable joints using extension method with fractal feature
enhancement,” Expert Systems with Applications, vol. 39, no. 3, pp. 2804 –
2812, 2012. 42

[80] H. Li, J. Li, W. Li, X. Zhao, G. Wang, and M. Alim, “Fractal analysis of
side channels for breakdown structures in xlpe cable insulation,” Journal of
Materials Science: Materials in Electronics, vol. 24(5), pp. 1640–1643, 2013.
42

- 146 -
 REFERENCES

[81] V. Jeyabalan and S. Usa, “Frequency domain correlation technique for pd

location in transformer winding,” IEEE Transactions on Dielectrics and
Electrical Insulation, vol. 16, no. 4, pp. 1160–1167, August 2009. 42

[82] W. Kinsner, “A unified approach to fractal dimensions,” in Fourth IEEE

Conference on Cognitive Informatics (ICCI), Aug 2005, pp. 58–72. 42, 75

[83] R. Salustiano, R. Capelini, S. de Abreu, M. Martinez, I. Tavares, G. Ferraz,

and M. Romano, “Development of new methodology for insulators inspections
on aerial distribution lines based on partial discharge detection tools,” in High
Voltage Engineering and Application (ICHVE), 2014 International Conference
on, Sept 2014, pp. 1–4. 44

[84] C. Nyamupangedengu, L. Luhlanga, and T. Letlape, “Acoustic and hf

detection of defects on porcelain pin insulators,” in Power Engineering Society
Conference and Exposition in Africa, 2007. PowerAfrica ’07. IEEE, July 2007,
pp. 1–5. 44

[85] N. Algeelani and M. Piah, “Identification of acoustic signals of surface

discharges on glass insulator under different contamination levels,” in
Electrical, Control and Computer Engineering (INECCE), 2011 International
Conference on, June 2011, pp. 511–514. 44

[86] MATLAB, version 7.10.0 (R2010a). Natick, Massachusetts: The MathWorks

Inc., 2010. 45, 69

[87] IEEE Guide for the Application of Insulation Coordination, Std. 1313.2, 1999.
47, 115

- 147 -
 REFERENCES

[88] Standard Test Method for Detection and Measurement of Partial Discharge
(Corona) Pulses in Evaluation of Insulation Systems, Std. ASTM D1868, 2007.
47, 120

[89] Insulators for overhead lines with a nominal voltage above 1000 V, Part
1:ceramic or glass insulator units for AC. systems Definitions, test methods
and acceptance criteria, IEC International standard Std. IEC 60 383-1, 1993.
47, 60, 62, 115

[90] “http://www.rossengineeringcorp.com/products/measurement/hv-voltage-
dividers.html.” 51

[91] P. T. Inc., “Electric field sensors,” February 2012. [Online]. Available:

http://prodyntech.com/home/page/d dot sensor 52

[92] R. Bartnikas and E. McMahon, Engineering Dielectrics: Corona Measurement

and Interpretation. Astm Intl, 1979. 54, 121, 132

[93] C. Magnelab Current Transformers. [Online]. Available: http://www.gmw.

com/electric current/Magnelab/ct Specs.html 54

[94] K. InfiniiVision, “Agilent 7104a,” November 2014. [Online].

Available: http://www.keysight.com/en/pd-1293645-pn-DSO7104A/
oscilloscope-1-ghz-4-analog-channels?&cc=CA&lc=eng 55

[95] “Ngk-locke, inc.” January 2015. [Online]. Available: http://www.ngk-locke.

com/ 56

[96] G. Montoya, I. Ramirez, and J. Montoya, “Correlation among esdd, nsdd

- 148 -
 REFERENCES

and leakage current in distribution insulators,” IEE Proceedings-Generation,

Transmission and Distribution, vol. 151, no. 3, pp. 334 – 40, 2004/05/15. 57

[97] T. F. S. Inc., “Orion 3-star plus portable conductivity meter kit,” February
2012. [Online]. Available: http://www.thermoscientific.com/ecomm/servlet/
productsdetail 11152 L10920 80575 11958237 -1 57

[98] F. Corporation, “Fluke 1507/1503 insulation resistance testers,”

February 2012. [Online]. Available: http://www.fluke.com/fluke/caen/
Digital-Multimeters/Fluke-1507-1503.htm?PID=56007 61

[99] T. Czaszejko and S. Chen, “Ultra wide band response of partial discharge test
circuit,” in 2010 International Conference on High Voltage Engineering and
Application (ICHVE), Oct 2010, pp. 61–64. 69

[100] L. Wang, Support Vector Machines: Theory and Applications, ser. Studies in
Fuzziness and Soft Computing. Springer, 2005. 105

[101] High-voltage test techniques - Partial discharge measurements, Std. IEC

60 270, 2000. 106, 107, 121

[102] A. Kraetge, S. Hoek, M. Koch, and W. Koltunowicz, “Robust measurement,

monitoring and analysis of partial discharges in transformers and other
hv apparatus,” IEEE Transactions on Dielectrics and Electrical Insulation,
vol. 20, no. 6, pp. 2043–2051, December 2013. 106

[103] K. Rethmeier, W. Pichler, M. Krüger, R. Plath, C. Steineke, and O. Ried,

“Multi-channel pd measurements on transformers–a new approach for real-

- 149 -
 REFERENCES

time data evaluation,” in OMICRON Seminar on Diagnostic Measurement on

Power Transformers, 2008. 107

[104] A. Cavallini, S. Chandrasekar, G. C. Montanari, and F. Puletti, “Inferring

ceramic insulator pollution by an innovative approach resorting to pd
detection,” IEEE Transactions on Dielectrics and Electrical Insulation, vol. 14,
no. 1, pp. 23–29, Feb 2007. 110

[105] M. G. Danikas, “The definitions used for partial discharge phenomena,” IEEE
Transactions on Electrical Insulation, vol. 28, no. 6, pp. 1075–1081, 1993. 120

[106] J. Mason, “Discharge detection and measurements,” Proc. IEE, vol. 112, no. 7,
pp. 1407–1423, 1965. 122

[107] H. Hall and R. Russek, “Discharge inception and extinction in dielectric voids,”
Proceedings of the IEE-Part II: Power Engineering, vol. 101, no. 79, pp. 47–55,
1954. 123

[108] J. Meek and J. Craggs, Electrical Breakdown of Gases. Clarendon. Oxford,

1953. 123

[109] J. Birks and J. Schulman, Progress in dielectrics. Wiley, 1959, vol. 1. 124

[110] J. Mason, “Breakdown of insulation by discharges,” Proceedings of the IEE-

Part IIA: Insulating Materials, vol. 100, no. 3, pp. 149–158, 1953. 124

[111] A. Gemant and W. Von Philipoff, “Die funkenstrecke mit vorkondensator

[spark gaps with front capacitors],” Z. f. techn. Physik, vol. 13, no. 9, pp.
425–430, 1932. 125

- 150 -
 REFERENCES

[112] S. Whitehead, Dielectric Breakdown of solids. Oxford, 1953. 126

[113] R. Bartnikas and J. H. E. Levi, “Improved pulsed discharge rate measuring

apparatus for ionization discharge studies at low frequencies,” Review of
Scientific Instruments, vol. 37, no. 9, pp. 1245 –1251, sep 1966. 130, 131

[114] R. Bartnikas, “Some observations on the character of corona discharges in

short gap spaces,” IEEE Transactions on Electrical Insulation, vol. EI-6, no. 2,
pp. 63 –75, june 1971. 130, 134

[115] ——, “Note on discharges in helium under ac conditions,” Journal of Physics

D: Applied Physics, vol. 1, no. 5, p. 659, 1968. 134

- 151 -