IEEE Transactions on Dielectrics and Electrical Insulation Vol. 17, No.

1; February 2010 149

A Generic Knowledge-based Approach

to the Analysis of Partial Discharge Data
S. Rudd, S. D. J. McArthur and M. D. Judd
University of Strathclyde
Institute for Energy and Environment
Glasgow, UK

ABSTRACT
Partial Discharge (PD) diagnosis is a recognized technique to detect defects within high
voltage insulation in power system equipment. A variety of methods exist to capture the
signals that are emitted during PD, and this paper focuses on the ultra high frequency
(UHF) and IEC 60270 techniques. Phase-resolved patterns can be constructed from the
PD data captured using either of these techniques and due to the individual signatures
that different defects generate, experts can examine the phase-resolved pattern to
classify the defect that created it. In recent years, knowledge regarding PD phenomena
and phase-resolved patterns has increased, providing an opportunity to employ a
knowledge-based system (KBS) to automate defect classification. Due to the consistent
physical nature of PD across different high voltage apparatus and the ability to
construct phase-resolved patterns from various sensors, the KBS offers a generic
approach to the analysis of PD by taking the phase-resolved pattern as its input and
identifying the physical PD processes associate with the pattern. This paper describes
the advances of this KBS, highlighting its generic application through the use of several
case studies, which present the diagnosis of defects captured through both the IEC
60270 and UHF techniques. This paper also demonstrates, in one of the case studies,
how a limitation of previous pattern recognition techniques can be overcome by
mimicking the approach of a PD expert when the pulses occur over the zero crossings
of the voltage waveform of the phase-resolved pattern.
Index Terms — Knowledge based systems, transformers, gas insulated substations,
condition monitoring, fault diagnosis, partial discharges, UHF devices, IEC 60270.

1. INTRODUCTION acoustic, thermal, chemical and more recently ultra high

frequency (UHF) monitoring. This paper will concentrate
THERE is a recognized need for condition monitoring on the UHF and electrical techniques and show how a
and automated diagnostic systems in the electrical utilities
knowledge-based system offers a generic approach to
[1], [2]. Such systems provide the ability to identify defects
diagnose defects from PD data captured through either
as they occur, allowing the scheduling of maintenance,
technique, by identifying descriptors of the phase-resolved
avoidance of equipment failure, optimizing the operation of
pattern that represent features that the expert would look
the equipment and increasing reliability. Automated
for. This builds on previous work by the authors [6], which
diagnostic systems can only be successful if there exists a
described the design of a knowledge-based system that is
way of capturing information that indicates the health of the
effective in diagnosing PD identified through UHF
apparatus. Analysis of partial discharge (PD) data provides
monitoring. This paper advances this work by describing
one such indication.
and demonstrating how the knowledge-based approach to
PD is a localized dielectric discharge that only partially PD diagnosis offers the potential of a novel generic system
bridges the insulation system between conductors, and that across different monitoring techniques and plant items.
may or may not occur adjacent to a conductor [3]. PD The novelty of the research demonstrates that by using
occurs at the site of a defect and can cause the production of the methodology proposed in [6], with phase-resolved data
light, sound, heat or a chemical reaction. The variety of and PD phenomena knowledge originally pertaining to
signals emitted during PD means that various methods exist UHF data, the knowledge-based approach has the potential
to detect their occurrence [1][4][5], ranging from electrical, to immediately diagnose faults in transformers and gas
insulated substations, through data from both IEC and UHF
Manuscript received on 23 December 2008, in final form 14 October 2009. sensors. This ability to diagnose defects using either type of

1070-9878/10/$25.00 © 2010 IEEE

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:5306 UTC from IE Xplore. Restricon aply.
150 S. Rudd et al.: A Generic Knowledge-based Approach to the Analysis of Partial Discharge Data
system, as well as on different plant items, offers the insulation on a 3D phase-resolved pattern, representing a one
potential of a generic system offering immediate decision second (50 cycle) snapshot of PD activity. This is achieved by
support to engineers. plotting each pulse, or in the case of the IEC data the peak
amplitude of the apparent charge, on a three-dimensional axis
1.1 UHF DETECTION OF PD consisting of the pulse’s relative amplitude, the cycle number
Originally developed in the 1980s for Gas Insulated on which the pulse appears and the phase position of the pulse
Substations (GIS) [7], UHF sensors were designed to capture the on the voltage cycle [12]. An example of a phase-resolved
electromagnetic waves in the UHF range (300 - 3000 MHz) pattern, which represents a protrusion defect, can be seen in
emitted by the PD current pulse. Now also used in power Figure 2. The pattern is in the form of a 50x64 matrix of
transformers, there exist three different types of UHF sensors to floating points that represent the PD activity in 50 consecutive
capture PD signals, shown in Figure 1. The first is the drain cycles across 64 phase windows of the voltage cycle; with the
valve sensor [8], which is inserted into the oil valve of the positive half cycle appearing first, between 0° and 180° and
transformer while it is still in operation. The second type of then the negative half cycle between 180° and 360°.
sensor is a window coupler [9], which is mounted on a special
dielectric window. This sensor is more suited for new
transformers due to the requirement of removing the transformer
from operation to install such a sensor. The third sensor is an
internal coupler built into the tank at manufacture [10].
UHF monitoring of PD data has become a recognized
technique [1] due to sensitivity and comparative immunity to
noise.

Figure 2. Phase-resolved pattern displaying one second snapshot of PD

activity within GIS.
Using data in this form, the knowledge-based system
described in [6] offers an automated approach to defect
classification and offers an explanation of the reasons for its
conclusion. This ability to offer a physical explanation for
the automatic classification sets it apart from machine
learning techniques [13], providing confidence to the user
of the reasons for diagnosing a particular type of PD. The
(c
knowledge-based system also provides a means of storing
Figure 1. Three UHF couplers for power transformers: (a) An internal
valuable expert knowledge regarding PD diagnosis.
coupler (courtesy of DMS Ltd), (b) an external dielectric window type, and The incremental knowledge-based approach to the
(c) oil valve probe type. analysis of PD data was originally created using knowledge
pertaining to UHF sensor data [6]. However, due to the
1.2 ELECTRICAL MEASUREMENT OF PD common physical nature of PD within high-voltage
The field of electrical monitoring of PDs is a mature area equipment [14] the knowledge-based system offers a
of measurement encapsulated in the IEC 60270 standard generic approach to classifying defects from any phase-
[3]. The IEC 60270 standard provides an electrical method resolved pattern created from IEC or UHF sensors. This
for measuring PDs. This is achieved by measuring small paper highlights the generic application by providing an
changes in the current using specialized equipment to example of a blind diagnosis from a set of IEC data,
acquire the apparent charge levels. performed using the knowledge already captured and
Electric measurements in the field often have to use implemented from UHF phase-resolved patterns. Due to the
ingenious variations of the sensor and measurement circuit consistent physical nature of PD across different high
to record the current pulse. For example, utilizing the high- voltage apparatus [14], the system has the potential to
voltage bushing (and tap) as the coupling [1]. diagnose defects in various high voltage equipment,
including power transformers, GIS and cables.
1.3 PHASE RESOLVED PATTERNS This paper will briefly discuss the implementation of the
To diagnose a fault from the PD data, captured by either knowledge-based system and then, by use of three case
IEC or UHF sensors [11], it first needs to be transformed into a studies, it will demonstrate the generic nature of this
generic workable format. One way of displaying the data is to approach by showing how defects can be recognized from
plot consecutive pulses generated by a defect present in the both the UHF and IEC data, in GIS and transformer models.

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:5306 UTC from IE Xplore. Restricon aply.
IEEE Transactions on Dielectrics and Electrical Insulation Vol. 17, No. 1; February 2010 151

2 OVERVIEW OF THE KNOWLEDGE-

BASED SYSTEM
Different types of defect can occur within the insulation
of high voltage equipment [6], [15], [16] and the
knowledge-based system is based on the knowledge that
each defect creates a different PD signature due to the
different physical processes occurring within the insulation
[15].
Previous research focused on the use of UHF sensors
[15][16] to detect seven PD defect classes that might
develop within insulation. Since then further defects have
been classified by the expert from phase-resolved patterns.
Therefore, the defects that can presently be identified by the
knowledge-based system are referred to as bad contact,
bouncing particle, floating electrode, floating component,
rolling particle, surface discharge and void [6], [15], [16].
A limited number of experts can look at the phase-
resolved pattern of the different defects and interpret the
signature. This knowledge regarding defect diagnosis was
captured using knowledge engineering techniques described
fully in [6]. Tacit knowledge is contextual information that Figure 3. UML activity diagram of the PD diagnosis.
the expert may or may not be aware of and is one of the strategies, forward chaining (data-driven) or backward chaining
hardest forms of knowledge to elicit. Knowledge elicitation (goal-directed). Drools [19] is an example of a forward chaining
techniques, such as case studies and interviews, were used inference engine and has been used to construct the knowledge-
to identify the tacit knowledge associated with diagnosing a based system described in this paper. Drools’ forward chaining
defect from a phase-resolved pattern and the PD engine provides a result by matching the precondition of the
phenomena. The resulting knowledge is represented in rules. To achieve this data driven approach, Drools uses the
various Unified Modeling Language (UML) models [17] to RETE algorithm [19] to traverse through the data until it results
allow easier validation by the expert prior to being in fewer matches and a conclusion can be reached. The rules are
implemented as rules in the knowledge-based system. stored in the five knowledge bases in production memory, while
The knowledge-based system is spilt into five stages in the facts that the inference engine matches are placed into the
order to mimic the process followed by an expert when working memory, where they can be further invoked to match
diagnosing a defect from the phase-resolved pattern, new rules.
explained fully in [6] and shown in Figure 3. Splitting the Semantic
knowledge into five separate knowledge bases allows easy Network
addition of further knowledge as it is gained, providing an Model
evolving knowledge base that can be easily altered or
updated if the knowledge proves inadequate or it can be
enhanced in future cases. It also provides varying levels of
explanation as the system proceeds with its diagnosis. The
expert has validated the knowledge incorporated into these IF Position in-between voltage and zero crossing
knowledge bases through experimental data and feedback.
OR
During the knowledge engineering task, semantic network Discharge on zero crossings
models were created to represent the expert’s knowledge. THEN Issue of space charge
These causal diagrams are directed graphs that show the
Figure 4. Conversion of semantic network model into “If condition Then
relationship between concepts. This knowledge, associated action” rule.
with each of the five stages, requires to be constructed as “If
condition Then action” rules and implemented into a rule- 3 CASE STUDY OF UHF DATA IN GIS
based system [18] to provide a classification. The rules are
To demonstrate the effectiveness and the generic nature of
assembled by converting the left-hand-side of the semantic
the knowledge-based system three case studies are shown.
network models into the condition and the right-hand side This first case study shows the diagnosis of a protrusion
into the action, shown in Figure 4. defect (fixed, sharp metallic objects on conductors), recorded
To determine which rules should be fired an inference during laboratory tests. A UHF sensor was employed to
engine is required [18]. The reasoning part of the inference capture the PD activity within a short section of 420 kV GIS
engine is accomplished by one of two fundamental busbar energized using a metalclad transformer.

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:5306 UTC from IE Xplore. Restricon aply.
152 S. Rudd et al.: A Generic Knowledge-based Approach to the Analysis of Partial Discharge Data
Most types of discharge tend to exhibit some characteristic 3.2 STAGE #2: MATCH ASSOCIATED PD
phase correlation with the 50 Hz voltage cycle, and the BEHAVIORS
protrusion, which is strictly a point to plane geometry, is the The second stage of the knowledge-based system is to
classic example of this. A protrusion with a high voltage (at 50 determine the PD behaviors, which led to the creation of the
Hz) applied to it results in a geometrical field pattern that is descriptors identified in stage #1. The inference engine achieves
greatest at the sharp tip. Since the field at the tip is proportional this by matching the preconditions in the behavior rule base with
to the instantaneous ac voltage, PD will tend to occur around the the descriptors that are in working memory. Once these behavior
voltage peaks at phase positions of 90° and 270°, as can be seen rules have been invoked their conclusions are also placed in
in Figure 2. When these pulses are captured by a measurement working memory for stage #3, while the output (see Table 1) is
system with phase-resolved capability, the patterns can be displayed in the “Behaviors” tab of the GUI shown in Figure 5.
passed into the knowledge-based system for classification.
Table 1. Descriptors to PD Behaviors.
The following sections show how a classification is derived Phase Descriptor
Descriptor Behaviors
for the protrusion example of Figure 2 by incrementing through Range Name
the five stages shown in Figure 3. Each stage is described in Whole Phase Asymmetrical 1. Conditions for PD inception are
detail here and subsequently two other case studies are provided cycle inception different for both polarities
symmetry
in sections 4 and 5 to show diagnosis, in less detail, of different
Whole Magnitude Positive more 1. Extract electrons from metal part
defects captured by different measurement systems on varying cycle symmetry than negative at high potential
plant items. Whole Density Positive more 1. Extract electrons from insulation
cycle symmetry sparse on positive phase
3.1 STAGE #1: SPLIT PHASE-RESOLVED
Whole Shape Asymmetrical 1. Defect is geometrically
PATTERN INTO DESCRIPTORS cycle symmetry asymmetrical
The input to stage #1 is a phase-resolved pattern (see Figure Positive Position Peaks 1. Minimal space charge present i.e.
2) created from the PD data in GIS captured from UHF sensors. half no memory effect beyond half cycle
From this, descriptors are determined using some of Gulski’s Positive Density Sparse 1. Defect experiencing inconsistent
statistical features [20] and other mathematical equations, which half electric field
2. Interference
describe different aspects of the pattern. These descriptors 3. Space charge with a long time
identified by experts during interviews are phase position, constant
magnitude, shape, inception symmetry, magnitude symmetry, Positive Range Broad 1. PD pulse phase influenced by
shape symmetry, density symmetry, pulse distribution, range, half local stored charge
density and magnitude consistency; all of which inform the 2. Many small discharge sites acting
simultaneously
expert about the PD behavior that is occurring within the 3. Charge can disperse easily
insulation. These descriptors are automatically identified through Positive Distribution Unbiased 1. No space charge
statistical analysis (see the left-hand side of Figure 5) and placed half
in working memory for utilization in later stages. Positive Shape Knife blade 1. Energetic discharge
half
Prior to identifying the descriptors of the phase-resolved
Positive Magnitude Medium 1. Pulses initiated in insulation
pattern, noise is automatically removed by dismissing small half
random pulses that lie outside the main phase regions in which Positive Magnitude Not constant 1. Locally stored charge
discharges are present; mimicking the expert’s process of noise half consistency magnitude 2. PD site is not confined to one
removal. For example, the pulses outside the knife blade in the region
positive half cycle of Figure 2 have been automatically stripped Negative Position Peaks 1. Minimal space charge present i.e.
by the system prior to calculating the shape of the pulses. half no memory effect beyond half cycle
However, it should be noted that if the pattern being analyzed is Negative Density Dense 1. Pulses at a conducting surface
half 2. No space charge
considered to be random, this process is not carried out;
otherwise valid data might be lost. Alternatively, the user can Negative Range Broad 1. PD pulse phase influenced by
half local stored charge
manually remove pulses below a noise threshold from the 2. Many small discharge sites acting
pattern by choosing an option in the systems graphical user simultaneously
interface (GUI). 3. Charge can disperse easily

If, in exceptional circumstances, excessive noise is Negative Distribution Unbiased 1. No space charge
half
encountered, resulting in disproportionately large noise pulses, it
Negative Shape Chopped sine 1. Voltage dependent (but not
is feasible that automated noise removal will not be possible to half proportional to it)
satisfactory degrees of accuracy. In these extreme situations, Negative Magnitude Small 1. Pulses at a very sharp tip
descriptor calculations may be distorted, however, if ultimately half 2. Pulses at a small site
necessary, the engineer has the ability to choose more accurate Negative Magnitude Not constant 1. Locally stored charge
descriptors from the GUI’s drop down menus and rerun stages half consistency magnitude 2. PD site is not confined to one
2-5 for a more representative diagnosis and better decision region
support.

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:5306 UTC from IE Xplore. Restricon aply.
IEEE Transactions on Dielectrics and Electrical Insulation Vol. 17, No. 1; February 2010 153

3.3 STAGE #3: MATCH ASSOCIATED DEFECT 3.5 STAGE #5: IDENTIFY PD LOCATION
CHARACTERISTICS The location of the defect is important when diagnosing
In order to understand where the classification was faults and stage #5 utilizes the expert’s knowledge to inform
derived from it is necessary to examine the defect the user about the subsystem within the equipment where the
characteristics associated with the PD behaviors identified defect identified in stage #4 exists. Previous research looked
in stage #2. These can be seen in Table 2. These defect at a “time of flight” technique to identify the location [9],
characteristics tell the user the physical presence of metal however, the knowledge-based system can also provide some
parts, or between which types of material the discharge is information to the user about the location of the defect. This
occurring. The output of stage #3 is displayed in the is achieved by invoking rules that identify the defect’s
“Characteristics” tab of the GUI shown in Figure 5. location from the recognized characteristics of the diagnosed
Table 2. PD Behaviors Relation to Defect Characteristics. defect. In this case study the knowledge has shown that due
Phase to the defect being a protrusion with the characteristic of the
Behaviors Characteristics
Range
metal part at high potential, it can therefore be concluded that
Whole Conditions for PD inception are different Metal part at
the protrusion can be found on the high voltage conductor.
cycle for both polarities AND minimal space high potential
charge present i.e. no memory effect This description would be displayed in the “Location” tab of
beyond half cycle AND extract electrons the GUI shown in Figure 5.
from insulation on positive phase Table 3. Classification of PD Source.
Whole Conditions for PD inception are different Discharge Characteristics Behaviors Classification
cycle for both polarities AND pulses at a between
conducting surface AND defect is conductor Metal or SF6 involved PROTRUSION
geometrically asymmetrical AND pulses surfaces, sharp
Involves gas to metal interface PROTRUSION
at a very sharp tip AND Minimal space tip
charge present i.e. no memory effect Metal part at high potential PROTRUSION
beyond half cycle AND PD pulse phase (on High Voltage
influenced by local stored charge Conductor)
Positive No space charge AND minimal space Metal or SF6 Metal part at high potential Extract SEVERE
half charge present i.e. no memory effect involved electrons from PROTRUSION
beyond half cycle metal part at (on High Voltage
high potential Conductor)
Positive Minimal space charge present i.e. no Involves gas to
half memory effect beyond half cycle AND metal interface Discharge between conductor PROTRUSION
space charge with a long time constant surfaces, sharp tip
AND conditions for PD inception are
different for both polarities
Positive Minimal space charge present i.e. no Involves gas to 3.6 KNOWLEDGE-BASED SYSTEM OUTPUT
half memory effect beyond half cycle AND metal interface The GUI of Figure 5 shows the output of the knowledge-
space charge with a long time constant
AND defect is geometrically
based system. To achieve the explanation associated with
asymmetrical the classification, the GUI is split into sections. The left-
hand side of the GUI shows the descriptors that have been
Negative No space charge AND minimal space Metal or SF6
half charge present i.e. no memory effect involved calculated from the phase-resolved pattern using statistical
beyond half cycle operators. These descriptors allow the user to see the
features that were extracted from the phase-resolved
3.4 STAGE #4: CLASSIFY DEFECT pattern, which the expert looks for when manually
Passing the phase-resolved pattern from Figure 2 into the diagnosing a defect from the phase-resolved pattern.
knowledge-based system results in the classification shown Initially, the right-hand side shows the 3D phase-resolved
in Table 3. As seen in the table, the system matches five pattern of the discharge pulses and the 2D mean and
different cases that suggest the defect is a protrusion. Each maximum graphs showing a straight on view of the average
of these cases are matched from rules that exist within the and maximum pulses calculated from the 3D pattern. All
working memory after the first three stages of diagnosis three of these graphs assist the expert when diagnosing a
have occurred. All five results, along with the evidence defect. The tabs across the top right-hand side of the GUI,
associated with each classification, can be viewed by the above the graphs, allow users with varying levels of
user to justify the final decision of the PD source. expertise to display the different stages of diagnosis
As shown in row four of Table 3, the severity of the according to their requirements. These tabs are ordered to
protrusion can also be recognized by the expert’s allow the engineer to delve deeper into the classification
knowledge. In this case, the severity of the protrusion was and so gain more confidence in the result. Examples of the
identified from the physical PD process of the electrons type of information that these tabs display have been shown
being extracted from a metal part at high potential, due to in the first five stages of this case study.
the descriptor of the magnitude in the positive half of the Although each of the five stages of the knowledge-based
voltage cycle being greater than the negative half. system are shown in the separate tabs across the right-hand

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:5306 UTC from IE Xplore. Restricon aply.
154 S. Rudd et al.: A Generic Knowledge-based Approach to the Analysis of Partial Discharge Data

Figure 5. GUI display of a protrusion defect in GIS captured through a UHF sensor.
side of the GUI, it was decided that also displaying the automatically adjusting the calculations accordingly leads
knowledge semantic network models, that were captured to an enhanced classification capability.
during knowledge engineering techniques, provides a The phase resolved pattern in Figure 6 represents PD
diagrammatic view of how the system reached its conclusion. from a solid dielectric sample captured through UHF
This would be shown in the “Flow” tab and is beneficial to sensors in GIS. This data was passed into the system for
explain the route the system took to reach the classification. diagnosis. As shown in the two 2D plots at the bottom of
Figure 6, the half cycles have been shifted before the
classification occurs. The system correctly classified the
4.0 SECOND CASE STUDY OF UHF defect as a surface discharge. The shaded circles in the
semantic network diagram of Figure 7 highlight the route
DATA IN GIS
A limitation of past pattern recognition systems that
diagnose PD defects exist when the discharge activity
occurs on the zero crossings relative to the voltage cycle Group 1
[21]. This problem arises from the statistics, such as skew Group 2
and kurtosis, which are used to help identify a specific Group 1
defect, being calculated on a per half cycle basis. When the
discharges occur across the zero crossings then the PD
activity is in fact occurring between the half cycles.
Conventional statistical algorithms do not recognize this
effect and therefore feed erroneous parameter values into
Figure 6. PD pattern displaying discharges occurring over
pattern recognition techniques. To overcome this limitation, the zero crossings of the voltage waveform.
the knowledge-based system implements the approach an
expert utilizes when the pulses are mainly occurring on the that the inference engine took through the knowledge-based
zero crossings. This is used to reproduce the way in which system to achieve this classification.
an expert views the pattern as being continuous over the 5.0 CASE STUDY OF IEC 60270 DATA
zero-crossing points of the ac waveform. PD data from a surface discharge defect within a power
When the expert looks at the pattern, the main phase transformer model under laboratory conditions was captured
bands in which discharges occur are first identified. In the by researchers at the University of Southampton. To capture
case of pulses occurring around the zero crossings (e.g. see this data, radio frequency current transducer sensors were
Figure 6), the expert views the pulses as two groups used with a bandwidth of 10 kHz – 200 MHz, along with a
spanning the boundaries of the positive and negative half conventional IEC 60270 PD detector. The defect itself was a
cycles. Hence the pulses on the right-hand side of the surface discharge, which is due to moisture ingress causing
pattern would be viewed as being in front of the pulses on pressboard to become semi-conducting.
the left-hand side (to become the positive half in the Raw PD data for a defect of a type not disclosed to the
system). This group (Group 1) would be compared with the authors was provided by the University of Southampton for
group of pulses in the middle of the pattern (Group 2, the “blind” diagnosis through the knowledge-based system. In
new negative half) to diagnose the defect. Incorporating this order to pass this “blind” set of data into the knowledge-
functionality into the knowledge-based system, and based system it was first converted into the specific data

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:5306 UTC from IE Xplore. Restricon aply.
IEEE Transactions on Dielectrics and Electrical Insulation Vol. 17, No. 1; February 2010 155

used in these buckets was the peak value (in volts) that
Descriptor Behavior Characteristic Classification
occurred in each consecutive 156 values i.e. 1 segment
(10,000/64).
4. Change the peak value from volts to pC
Since the calibration factor was 600 mV = 500 pC, the
peak value was transformed to pC by performing the
following calculation:
SegmentAmplitude = PeakValue/0.6 V x 500 pC
5. Find the relative amplitude
The relative amplitude was plotted as a percentage and
in terms of the Southampton data, 5 V (or 4167 pC by
using the equation in 4) corresponds to 100%
amplitude. In the knowledge-based system, the 100%
corresponds to a numerical value of 255 (8-bit
resolution). Therefore, the relative amplitude of the
value calculated in step 4 was found by:
RelativeAmplitude = (SegmentAmplitude/4167) x 255
6. Plot the relative amplitude against phase and cycle
number to form a 50x64 matrix
Inserting the relative amplitude at the phase and cycle
number of the matrix formed the phase-resolved
matrix.
Figure 7. Semantic network model of flow of knowledge. Once the data had been transformed into the 50x64 phase-
format required by the system. This was achieved by resolved matrix format it was passed through the knowledge-
plotting consecutive pulses generated by the defect on the based system for classification. The GUI of the knowledge-
3D phase-resolved pattern. The creation of the phase- based system is shown in Figure 8, highlighting the
resolved pattern was constructed using the Java programming descriptors that were extracted from the pattern. Although the
language, implementing the following process: “blind” data was captured using the conventional IEC 60270
technique, after blind diagnosis, the knowledge-based system
1. Split the raw data into 50 cycles
classified a surface discharge defect within the transformer
The raw file, consisting of 50 power cycles of PD data,
model. This classification was presented to the University of
contained 500,000 points, therefore, each cycle
Southampton, who confirmed that the data set was a surface
consisted of 10,000 points (500,000 points / 50 cycles).
discharge defect. This correct classification was derived
2. Remove noise and negative values
using knowledge that originally related to UHF diagnosis.
Due to the oscillating nature of the signal the negative
Validation of this classification demonstrated the potential
values were removed from the data. A noise level was
generic nature of the knowledge-based system.
then chosen to remove any noise from the signal.
Even though the PD measuring equipment and the plant
Displaying the raw data in the programming
item are different from the case study in section 4, and the
environment Matlab identified a noise level of 0.12 V
in this case. discharges are occurring on a per half cycle basis rather than
3. Construct the phase buckets across the zero crossings, the knowledge flow in the
The 10,000 points per cycle were split into 64 knowledge-based system that was utilized to correctly identify
segments to represent the 64 phase buckets. The value the surface discharge defect is also presented by Figure 7.

Figure 8. GUI display of a surface discharge in transformer insulation captured through IEC 60270 PD detector.

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:5306 UTC from IE Xplore. Restricon aply.
 156 S. Rudd et al.: A Generic Knowledge-based Approach to the Analysis of Partial Discharge Data
6 CONCLUSIONS [10] D. Templeton, H.Q. Li, K.F. Lee, J. Pearson, R. Brinzer, A. Reid and M.
Judd, “Sensitivity testing of a UHV power transformer monitoring system”,
The knowledge-based approach to the analysis of PD XVth Intern. Sympos. High Voltage Engineering, Ljubljana, Slovenia, T7-
data, described in this paper, automates the process of 617, 2007
diagnosing faults within high voltage equipment. Through [11] M. D. Judd, L. Yang, C.J. Bennoch and I.B.B Hunter, “UHF Diagnostic
Monitoring Techniques for Power Transformers”, Invited paper, XII EPRI
the use of a case study, this paper has demonstrated that by Substation Equipment Diagnostics Conf., New Orleans, February 15-18,
mimicking the way an expert views the data, limitations of 2004.
conventional pattern recognition techniques can be [12] J.S. Pearson, O. Farish, B.F. Hampton, M.D. Judd, D. Templeton, B.M. Pryor
and I.M. Welch, “Partial Discharge Diagnosis for Gas Insulated Substations”,
overcome when the discharges occur over the zero IEEE Trans. Dielectr. Electr. Insul., Vol. 2, pp. 893-905, 1995.
crossings of the voltage waveform. Grouping the discharges [13] S.D.J McArthur, S.M Strachan and G. Jahn, “The Design of a Multi-Agent
the way an expert would leads to the correct statistics being Transformer Condition Monitoring System”, IEEE Trans. Power Systems,
Vol. 19, pp. 1845-1852, 2004.
calculated prior to performing the classification. [14] J. Fuhr, B. Fruth, L. Niemeyer, D. Konigstein, M. Hassig and F. Gutfleisch,
Automated diagnostic systems in the past focused on “Generic Procedure for Classification of Aged Insulation System”, IEEE 3rd
employing machine learning techniques, offering no Intern. Conf. Properties and Applications of Dielectric Materials, Tokyo,
Japan, pp. 35-38, 1991.
explanation for their classification. Providing the user of the [15] G.P. Cleary and M.D. Judd, “An investigation of discharges in oil insulation
system with explanation at appropriate levels, through this using UHF PD detection,” 14th IEEE Int. Conf. Dielectr. Liquids, Graz,
knowledge-based system, will increase their confidence in Austria, pp. 341-344, 2002.
the result and provide decision justification. [16] G. P. Cleary and M.D. Judd, “UHF and current pulse measurements of partial
discharge activity in mineral oil”, IEE Proc. Sci., Measurement and
This research has shown, through the use of case studies, Technology, Vol. 153, pp. 47-54, 2006.
that by using the phase-resolved pattern as the input to the [17] G. Booch, The Unified Modelling Language User Guide, Second Ed,
knowledge-based system it is possible to classify defects Addison Wesley, 1998.
[18] B. G. Buchanan and E. H. Shortliffe, Rule-based expert systems, Addison-
from either IEC 60270 or UHF sensor data. The knowledge Wesley Publishing Company, United States, Chapter 2, 1985.
associated with PD behaviors is independent of the type of [19] Drools: http://www.jboss.org/drools/
equipment and so this knowledge can be used immediately [20] E. Gulski, Computer-aided recognition of partial discharges using statistical
tools, Delft University Press, 1991.
on various apparatus, including transformers and GIS, as [21] G. Berg and L.E. Lundgaard, “PD Signatures of Wedge Type Discharges in
highlighted in the case studies. Therefore, this knowledge- Transformer Insulation”, 14th IEEE Int. Conf. Dielectr. Liquids, Graz,
based approach could form the basis of a generic system for Austria, pp. 211-214, 2002.
the diagnosis of defects from a variety of sensory data
across several types of high voltage equipment. Susan Rudd received her B.Sc. (Hons) from the
University of Strathclyde in 2004. She is a Research
ACKNOWLEDGMENT Assistant in the Institute for Energy and Environment.
The authors would like to thank Dr Liwei Hao of the Her research interests include knowledge engineering,
intelligent system applications in power engineering and
University of Southampton for providing the IEC 60270 data. condition monitoring.
This work is funded through the EPSRC Supergen V, UK
Energy Infrastructure (AMPerES) grant in collaboration with
UK electricity network operators working under Ofgem's Professor Stephen McArthur (M’93-SM’07) received
Innovation Funding Incentive scheme. the B.Eng. (Hons) and Ph.D. degrees from the University
of Strathclyde in 1992 and 1996, respectively. He is a
REFERENCES Professor at and Co-director of the Institute for Energy
[1] R.E. James and Q. Su, Condition Assessment of High Voltage Insulation in and Environment, within the Department of Electronic
Power System Equipment, Institute of Engineering and Technology, London, and Electrical Engineering. He is Director of the British
United Kingdom, Chapter 8, 2008. Energy Advanced Diagnostics Centre and Co-director of
[2] Y. Han and Y.H. Song, “Condition Monitoring Techniques for Electrical the Strathclyde Centre for Applied Intelligent Systems.
Equipment – A literature Survey”, IEEE Trans. Power Delivery, Vol. 18, pp. He chairs the IEEE PES Working Group on Multi-agent Systems. His
4 –13, 2003. research interests include intelligent system applications in power,
[3] IEC 60270: High-voltage test techniques – Partial discharge measurements, condition monitoring, fault diagnosis and multi-agent systems.
Edition 3, 2000.
[4] I.J. Kemp, “Partial discharge plant-monitoring technology: Present and future
developments”, IEE Proc.–Sci., Measurement and Technology, Vol. 142, pp. Dr. Martin D. Judd (M’02-SM’04) is a Reader in the
4 –10, 1995. Institute for Energy and Environment at the University of
[5] A. Krivda, “Automated Recognition of Partial Discharges”, IEEE Trans. Strathclyde, where he manages the High Voltage
Dielectr. Electr. Insul., Vol. 2, pp. 796 – 821, 1995. Diagnostics Laboratory. He graduated from the
[6] S.M. Strachan, S. Rudd, S.D.J. McArthur, M.D. Judd, S. Meijer and E. University of Hull in 1985 with a B.Sc. (Hons.) degree in
Gulski, “Knowledge-Based Diagnosis of Partial Discharges in Power electronic engineering. His employment experience
Transformers”, IEEE Trans. Dielectr. Electr. Insul., Vol. 15, pp. 259-268, includes four years working for Marconi Electronic
2008. Devices Ltd followed by four years with EEV Ltd, both
[7] B.F. Hampton and R.J. Meats, “Diagnostic Measurements at UHF in Gas in Lincoln, England. Dr. Judd received the Ph.D. degree from the
Insulated Substations”, Proc. IEEE, Vol. 135, Pt C, No. 2, pp. 137-144, 1988. University of Strathclyde in 1996 for his research into the excitation of
[8] S. M. Markalous, Detection and Location of Partial Discharges in Power UHF signals by partial discharges in gas insulated switchgear. From 1999
Transformers using acoustic and electromagnetic signals, Sierke Verlag, to 2004 he was an EPSRC Advanced Research Fellow. He is an expert on
Cottingen, 2006. electromagnetic theory, propagation and measurement and leads a research
[9] M. .D Judd, L. Yang and I.B.B. Hunter “Partial discharge monitoring team specializing in high frequency diagnostic techniques for high voltage
for power transformers using UHF sensors Part 1: Sensors and signal electrical equipment and systems. He is a Chartered Engineer, a Member
interpretation”, IEEE Insul. Mag., Vol. 21, No. 2, pp. 5-14 2005 of the IET and a Fellow of the Higher Education Academy.

Authorized licensd use limted to: IE Xplore. Downlade on May 13,20 at 1:5306 UTC from IE Xplore. Restricon aply.