Knowledge-Based Diagnosis of Partial Discharges in Power Transformers

7/29/2019 Knowledge-Based Diagnosis of Partial Discharges in Power Transformers

1/10

IEEE Transactions on Dielectrics and Electrical Insulation Vol. 15, No. 1; February 2008

1070-9878/08/$25.00 2008 IEEE

259

Knowledge-Based Diagnosis of Partial Dischargesin Power Transformers

S. M. Strachan, S. Rudd, S. D. J. McArthur, M. D. JuddUniversity of Strathclyde

Institute for Energy and Environment

Glasgow, UK

and S. Meijer and E. GulskiDelft University of Technology

HV Components and Power Systems

Delft, The Netherlands

ABSTRACTThe abstraction of meaningful diagnostic information from raw condition monitoring

data in domains where diagnostic expertise and knowledge is limited and constantly

evolving presents a significant research challenge. Expert diagnosis and location ofpartial discharges in High Voltage electrical plant is one such domain. This paper

describes the functionality of a knowledge-based decision support system capable of

providing engineers with a comprehensive diagnosis of the defects responsible for

partial discharge activity detected in oil-filled power transformers. Plant data captured

from partial discharge (PD) sensors can be processed to generate phase-resolved

partial discharge (PRPD) patterns. This paper proposes a means of abstracting the

salient features characterizing the observed PRPD patterns. Captured knowledge

describing the visual interpretation of these patterns can be applied for defect diagnosis

and location. The knowledge-based PRPD pattern interpretation system can support

on-line plant condition assessment and defect diagnosis by presenting a comprehensive

diagnosis of PD activity detected and classification of the defect source. The paper also

discusses how the system justifies its diagnosis of the PD activity to offer the expert

greater confidence in the result, a feature generally absent in black-box patternrecognition techniques. The incremental approach exhibited by the system reflects that

of a PD experts visual interpretation of the PRPD pattern. The paper describes how

this functional system design has evolved from the approach taken by PD experts to the

visual interpretation of PRPD patterns.

Index Terms Transformers, partial discharges, knowledge-based system,

classification, diagnosis, condition monitoring, UHF sensors.

1 INTRODUCTION

WITH a design life of around 20-35 years (increasing to

50 years in practice with appropriate maintenance) [1], large

power transformers at transmission voltages are widely

regarded as reliable items of electrical plant. The crucial role

performed by transformers in maintaining a secure and

continuous delivery of electricity to consumers, demands

their safe and reliable operation. While significantly more

reliable in comparison to many other network components, it

is the potential impact of a transformer malfunction that

drives the requirement for effective condition monitoring of

these commercially and technically valuable network assets.

Transformer failures may result in loss of supplies to large

Manuscript received on 21March 2007, in final form 10 August 2007.

populations of customers, while the threat of explosions and

fire causing irreparable damage to adjacent plant and injury

to personnel pose a risk to safety and the environment with

potentially serious legal and regulatory implications.

Therefore, while transformers may have a generally reliablereputation in the operation of the network, the welfare of

these critical network assets remains at the forefront of

utilities asset management responsibilities and obligations.

Condition monitoring technology captures operational plant

data characterizing its vital signs, enabling subsequent

on/off-line assessment of plant condition and performance.

Numerous on-line and off-line condition monitoring

technologies operating on a variety of observables are

available for the detection of incipient failures within

electrical power transformers [2].


2/10

S. M. Strachan et al.: Knowledge-Based Diagnosis of Partial Discharges in Power Transformers260However, the novel nature of some condition monitoring

and sensor technologies means that plant experts can find

themselves presented with unfamiliar data sets with the

potential to provide fresh insights into previously

uncharacterized areas of plant behavior and associated

phenomena. As this knowledge grows and evolves, a

mechanism is required to harness, store and leverage the

knowledge and understanding required to map these

condition monitoring data sets to distinct observations in the

plants behavior and its failure mechanisms. Partial discharge

is one such phenomenon that is the subject of ongoing

research focusing on its behavior, cause and effects, and

methods of detection and diagnosis [3],[4].

Partial discharge (PD) activity is a prominent symptom of

the service ageing and degradation of transformer insulation.

Historically, monitoring of this phenomenon has been limited

and as such a lot remains unknown of its causes, breakdown

mechanisms, and exactly what influences its behavior. The

development of an extensible knowledge based system will

allow the incorporation of new knowledge of previously

unknown PD behavior and new defect types as it is learned(either through empirical analysis or practical experience).

What is known, is that the dielectric properties of plant

insulation may be severely impaired if subjected to consistent

PD activity over prolonged periods of time (e.g. initiated by

transient over-voltage or incipient weaknesses in insulation,

etc.). This may ultimately lead to complete failure where PD

activity remains undetected and untreated. Insulation integrity

is therefore regarded as a critical aspect of a transformers

condition. As such, suitable safeguards are required to

monitor transformer plant for the presence of potentially

damaging PD activity.

Historically, research into partial discharges has focusedon the nature of these discharges occurring within solid and

gas insulation systems [5],[6]. Investigations into the partial

discharge phenomenon in liquid dielectrics (such as the oil

insulation of a transformer) are more difficult and as a

consequence are generally less well understood within the

PD research community as a whole. Limited knowledge of

the nature of partial discharges generated in the oil

insulation of power plant with complex geometry, such as

oil-filled power transformers, prompted some empirical

research within the University of Strathclyde [7]. Partial

discharge data captured using UHF sensor technology was

analyzed by University of Strathclyde researchers in anattempt to understand the physical process underpinning

dielectric breakdown within large oil-filled power

transformers and the PD behavior arising from particular

defect geometries [7][8]. This research sought to enhance

the general PD communitys knowledge-base of PD

behavior associated with this transformer plant and oil

insulating medium.

This paper proposes that as our knowledge base evolves

and matures, and the linkage between the cause and effect

of the PD phenomenon becomes more established and

better understood, this domain knowledge can be retained

and embedded within a knowledge-based system

facilitating future PD defect diagnosis and location. A

knowledge-based decision support system is proposed as a

means of providing a practical diagnostic explanation of

PD behavior and defect geometry to end-users (e.g.

maintenance personnel, asset managers, PD experts, plant

manufacturers).

The knowledge-based system proposed in this paper

mimics a partial discharge experts approach to the

interpretation of phase-resolved partial discharge (PRPD)

patterns for the purpose of diagnosing the nature, severity

and cause of detected PD activity. The system uses

statistical parameters to characterize the key feature (or

descriptors) of the PRPD pattern, which a PD expert would

typically look for when performing their diagnosis. These

descriptors are then used to explain in physical terms the

behavior associated with the observed partial discharge

activity. Understanding the PD behavior enables distinct

aspects of the physical geometry and properties of the

responsible defect source to be deduced and subsequentlymatched to the diagnostic rules of a knowledge base.

Ultimately, knowledge of the defect site geometry can be

used to classify and subsequently locate the PD defect

source.

This knowledge-based system contains knowledge

derived from multiple experts which readily maintainable

and generally applicable to any H.V. oil filled transformer.

Another benefit of the knowledge-based approach is the

provision of a meaningful diagnostic explanation

accompanying the systems defect classification. This

increases the users confidence in the system output.

The paper describes the structured approach taken to theknowledge capture required to populate the respective rule-

bases with the appropriate areas of domain knowledge.

Furthermore, the paper proposes that the modular system

design provides an extensible framework, which can be

readily augmented with fresh knowledge (e.g. concerning

new types of PD, different plant types, different insulating

mediums) as it evolves and matures through continued

research and on-site experience. In addition, this system will

assist experts in the recognition and understanding of new PD

sources or different manifestations of existing ones.

2PARTIALDISCHARGEACTIVITYINOIL-

FILLEDPOWERTRANSFORMERS

A partial discharge occurs in a high voltage (HV)

transformer when the electric field strength exceeds the

dielectric breakdown strength of the insulating medium in a

localized area. This results in an electrical discharge that

only partially bridges the insulation between conductors.

High electron energies can break molecular bonds in

adjacent solid or liquid dielectrics resulting in changes to

the chemical properties of the insulating material and its

subsequent ablation.


3/10

IEEE Transactions on Dielectrics and Electrical Insulation Vol. 15, No. 1; February 2008 261

Partial discharge activity may be categorized by the

defect source responsible and its locale [4]. The range of

defect classes for which the nature of the discharge

behaviour was observed in the programme of research

reported in [7] include:

Bad Contact microdischarges between conducting

surfaces separated by a thin layer of oil, e.g. between

the threads of loose nuts and bolts.

Floating Component small fixed conducting particle

at floating potential that experiences a localised PD

that does not bridge the gap to the system conductors.

Suspended Particle small, free-moving conducting

objects or debris within the insulating oil that can

discharge to the system conductors.

Rolling Particle conducting particles lying on a

conductive surface until influenced by the electric field

causing them to roll or bounce around.

Protrusion fixed, sharp metallic protrusions on HV

conductors.

Surface Discharge at pressboard surface due to

moisture ingress or material damage.

Floating Electrodes large components such as stress

shields that have become detached from the chamber,

acquiring a floating potential that results in capacitive

sparking.

These defect classes represent those targeted by the

knowledge-based system described in this paper.

3UHFDETECTIONOFPARTIALDISCHARGESANDPHASE-RESOLVED

PATTERNREPRESENTATION

Figure1. Phase-resolved partial discharge pattern.

Recent research has demonstrated the efficacy of

employing ultra-high frequency (UHF) detection of PD in

the monitoring of transformers [9],[10]. When a PD occurs

inside a transformer, an electromagnetic wave is radiated

that resonates in the tank and can be detected using UHF

sensors. Signals from the sensors are typically filtered,

amplified and converted to a simple pulse by a detector

circuit before being digitized. A phase signal derived from

the power frequency waveform provides additional

reference information for the digitized PD data. Each PD

pulse recorded is associated with a particular time and

point-on-wave. The amplitude of the displayed pulses is

proportional to the intensity of the UHF signal. A typical

display is shown in Figure 1, in which PD pulses occur in

bursts on the rising and falling portions of the high voltage

sinusoid of the power frequency waveform.

Three main measurements are used to quantitatively

describe the degradation of insulation resulting from partial

discharge activity, these are:

Discharge pulse magnitude trend - indicative of the

rate of dielectric deterioration.

Discharge pulse repetition rate/number of pulses

indicative of defect severity, where periods of

sustained activity are considered more damaging

than periods of intermittent activity.

Discharge pulse phase angle - indicative of the

defects ignition and extinction conditions in relation

to the electric field.These quantities are inherent in the PRPD pattern and

collectively provide a suitable means of characterizing

partial discharge activity for diagnostic interpretation and

defect classification [11].

A PRPD pattern may be constructed using data derived

from many different types of sensor and measurement

systems. Therefore, using the PRPD pattern as the basis of

the system diagnosis means that the system could in

principle support different monitoring technologies.

4 KNOWLEDGE-BASEDSUPPORT

FORPARTIALDISCHARGEDEFECTDIAGNOSISINTRANSFORMERS

Historically, PD in transformers has not been monitored

in service. Utilities have relied principally on laboratory

based Dissolved Gas Analysis (DGA) to establish the

extent and nature of the PD occurring within plant. While

this monitoring technology has provided significant benefit

in the detection of PDs within transformer plant, it has its

limitations in diagnosing the nature, cause, severity and

location of detected PD activity. The lack of understanding

of the PD process associated with dielectric breakdown, has

historically prompted the need for a less knowledge

intensive and more data driven approach to PD defectclassification [12]. On the other hand, the power

transformer in general is a complex system due to the

amount of different dielectrics in- and outside the

transformer, all of them having their own failure

mechanisms. In order to detect these failures in advance,

several types of diagnostics are available and should be

applied during the service life of a power transformer. In

particular, as shown in [3] performing PD diagnosis may

contribute to detect 21% of problems related to transformer

vessel and to 14% of problems related to HV bushings.


4/10

S. M. Strachan et al.: Knowledge-Based Diagnosis of Partial Discharges in Power Transformers262However as on-line monitoring of in-service plant has

become more practical and viable, PD experts are presented

with an opportunity to gain new insights into the physics of

the partial discharge process, and subsequently improve

their understanding of this phenomenon.

While previous research has demonstrated the application

of machine learning pattern recognition techniques as a

useful means of defect classification [13]-[15], some

constraints are associated with these approaches, including:

the trained classifiers remain specific to the transformer

from which the initial training data set is derived; historical

data sets are required to train the classifier; a classifier will

only recognize defects on which it has been trained and as

such must be retrained (or train new networks) to recognize

new defects. More significantly however, the nature of

these black-box algorithms is such that no practical

explanation may be provided as justification for the output

classification presented by the system. Machine learning

techniques remain a useful means of defect classification,

despite their limited explainability. As such, this paper does

not discount the data driven approach, but insteadrecognizes the potential role of a knowledge-based system

in providing decision support to plant experts. Providing a

detailed diagnosis of the detected PD activity reassuringly

offers the expert an explanation of the health and condition

of transformer plant, and provides a degree of confidence in

the system output, generally absent in use of black-box

pattern recognition techniques.

PD diagnostic knowledge can be captured from empirical

and practical engineering experience, which relates measured

partial discharge quantities to observed defects and the

resulting breakdown processes, taking into account external

influencing conditions such as plant geometry, propagationeffects, attenuation, interference, etc. This knowledge can be

implemented as heuristic rules to support engineers in the

diagnosis of partial discharge activity when assessing the

condition of transformer insulation. A CIGRE Task Force

(15.11/33.03.02) involving some key contributors within the

PD community has recognized the potential for knowledge-

based decision support in this area [16].

The expert interpretation of the PRPD patterns suggests

an emerging area of expertise which if suitably harnessed

by means of a knowledge (rule) based expert system, offers

a means of automating PD diagnosis and defect

classification.

5 EXPERTINTERPRETATIONOFPHASE-RESOLVEDPARTIAL

DISCHARGEPATTERNS

This paper proposes a knowledge-based approach to the

interpretation of phase-resolved partial discharge patterns

that aims to mimic an experts visual interpretation of these

patterns (Figure 2).

By simply observing captured PRPD patterns, experts

are able to intuitively abstract features in the PRPD patterns

(i.e. PRPD descriptors) and associate these features with

key characteristics of the PD behavior and hence the

responsible defect geometry. Knowledge of the physical

attributes of the defect can subsequently inform the

classification and location of the PD defect source.

The Unified Modeling Language (UML) Activity

Diagram [17] of Figure 2 was defined through knowledge

elicitation meetings conducted with experts specializing in

the experimentation and understanding of the partial

discharge phenomenon and specifically the interpretation of

their causes and behavior from PRPD patterns derived from

UHF condition monitoring data. A UML Activity Diagram

can be used for process modeling in order to present

knowledge engineers and domain experts with a clear

understanding of the interaction between process tasks in

terms of the input/output objects handled and delivered by

the task, the task dependencies indicating the sequence of

execution and the flow of information between tasks. This

graphical representation enables fast and accurate

validation of the modeled process by domain experts.

Figure2. Process of visual interpretation of PRPD patterns for PD defect

diagnosis.

The Activity Diagram of Figure 2 in turn defined the

functional framework of the knowledge-based expert

system of Figure 3 and the structured approach taken to the

acquisition of domain knowledge suitable for system

implementation.

Inputs

Tasks


5/10


The functional flow of the experts PRPD pattern

interpretation process effectively mapped out the

agenda for the knowledge acquisition exercise. When

eliciting the diagnostic rules for the population of the

appropriate rule-bases, aligned to each stage of the

interpretation process, it was more intuitive for the

expert to start with a PRPD pattern of a known defect

class and/or location. It is then easier for the expert to

deconstruct the defect source diagnosis into its physical

characteristics first, before considering how these

characteristics influence the resultant partial discharge

activity and ultimately how this is evident in the

characteristics of the PRPD pattern.

6 KNOWLEDGE-BASEDEXPERTSYSTEM

FORPRPDPATTERNINTERPRETATION

The modular design of the system illustrated in Figure

3 reflects the experts incremental approach to the visual

interpretation of PRPD patterns. Each stage of the

interpretation process requires and utilizes knowledgerelating to how the PRPD pattern, the PD activity and the

responsible defect source can be recognized and

characterized. This knowledge can be used to build up a

practical explanation of the systems PD defect

classification. The diagnostic rules elicited from PD

experts, which enable the interpretation of the PRPD

patterns, form a number of rule-bases. Each rule-base

represents knowledge associated with a specific stage of

the interpretation process illustrated in Figure 2. The

objective of the rules contained in each rule base can be

summarized as follows:

Rule-base RB No.1 consists of rules that extractdeduced and statistical features from the PRPD

patterns, forming the basis for further interpretation.

Rule-base RB No.2 uses knowledge of how the derived

statistical fingerprint characterizing the PRPD pattern

translates into features (i.e. PRPD descriptors)

describing the observed characteristics of the PRPD

pattern itself in terminology familiar to human experts.

Rule-base RB No.3 uses knowledge of how these

PRPD descriptors translate into physical aspects of the

PD behavior.

Rule-base RB No.4 uses knowledge of how theidentified PD behavior translates into the physical

characteristics of the responsible defect source (i.e.

defect geometry).

Rule-base RB No.5 uses knowledge of how the defect

geometry can be used to identify suspected defect

classes.

Rule-base RB No.6 uses knowledge of the defect class

to propose plausible locations for the defect site within

a localized area of the plant item.

Figure 3. Functional description of knowledge-based PRPD pattern

interpretation system.

The nature of the rules in each rule-base is detailed in

Section 8.

This knowledge-based approach provides the user with a

justification for the systems defect classification. In

addition, the level of detail in the explanation can be

customized to suit the users level of expertise and

experience. For example, maintenance engineers may be

more concerned with identifying the type of PD defect

source and its location, while PD experts or plantmanufacturers may have a greater interest in the breakdown

mechanisms associated with the PD activity underpinning

this high-level diagnosis, and the potentially degrading

effect this has on the plant insulation in the long term under

different operating conditions (i.e. the PD behavior and

defect geometry).

The inference engine of an expert system is encoded in

software and is responsible for processing input data to

decide what should be done next, i.e. specifically which

rules in the system rule-base to invoke or fire. If the rule

conditions are satisfied and it is fired then the rule

concludes that some predefined action to be performed (e.g.

report a diagnosis to the screen, invoke another rule, etc.).

The inference mechanism of the system described in this

paper underpins the incremental approach to the

knowledge-based diagnosis through its application of a

forward-chaining reasoning mechanism across the rule-

bases. The selection and execution of PRPD pattern

interpretation rules from the respective rule-bases is

controlled by this inference mechanism, using existing

diagnostic facts (determined by the conclusions of

previously fired rules, asserted in the working memory) to

fire relevant rules in the same or subsequent rule-bases


6/10

S. M. Strachan et al.: Knowledge-Based Diagnosis of Partial Discharges in Power Transformers264[18]. Once fired, further rule conclusions are asserted,

effectively augmenting the systems diagnosis while

prompting subsequent rule firing in the appropriate rule-

base as the system moves towards its diagnostic

conclusion.

7 CAPTURINGDIAGNOSTICKNOWLEDGEFORRULE-BASED

IMPLEMENTATION

Knowledge may be considered to be either explicit or

tacit. Explicit knowledge is generally that which is well

defined, represented and understood, while tacit knowledge

is generally experiential knowledge often associated with

tasks performed by a specialist which have effectively

become second-nature, and as such can be difficult to

articulate. It is therefore tacit knowledge that often sets

experts apart from non-experts and subsequently

distinguishes expert systems from traditional information

systems. Within the context of this paper the term

knowledge relates to the expertise exhibited by domain

experts in diagnosing and classifying PD defect sources

from observations made on PRPD patterns. This tacit

knowledge can be captured and utilized in an expert system

through a structured process of knowledge engineering.

Knowledge engineering is a structured method of

transferring and transforming the tacit expert knowledge

relating to a specific domain, from the mind of an expert

into a computer-tractable form. The knowledge

engineering process involves the acquisition of expert

knowledge associated with a particular domain and task

using a structured interview process focused on pre-

selected case studies (i.e. PRPD patterns); representation

of expert knowledge in an intelligible format forvalidation and utilization (i.e. knowledge transcripts and

causal models); validation of expert knowledge before it

can be confidently utilized (e.g. within a knowledge

(rule)-based system).

As mentioned, the output from each rule-base in this

approach to PRPD pattern interpretation relies upon the

output asserted by its antecedent rule-base. A forward

chaining inference mechanism underpins the interaction

between these rule-bases. This section examines the

execution of the rules within the knowledge-based system

in the construction of a PD diagnosis for a particular case

study.

8 RULE-BASEIMPLEMENTATIONCASESTUDY

8.1 Rule-base No.1 Abstraction of

Statistical Fingerprint from PRPD

Pattern

Rule-base No.1 of Figure 3 deduces distributions from

the input PRPD pattern (Figure 4), and subsequently

derives statistical quantities from these distributions.

Figure 4. Input PRPD pattern used in the case study.

8.1.1 DEDUCED DISTRIBUTIONS

The PRPD pattern provides an insight into the nature of

the PD process, its intensity and the geometry of the

responsible defect [11]. Distributions deduced from the

PRPD pattern comprise both time varying and phase

varying features characterizing sustained PD activity over a

number of voltage cycles. Consequently, the following four

distributions deduced from a PRPD pattern representing a

one second snapshot of PD activity (Figure5) characterize

the changing partial discharge activity over a number ofconsecutive cycles (in this case 50 voltage cycles):

TheDischarge Pulse Count distribution Hn().

The Discharge Maximum Pulse Height distribution

Hqm().

TheDischarge Energy Magnitude distribution Hqs().

TheDischarge Mean Pulse Height distribution Hqn().

Figure 5. Abstraction of deduced discharge mean pulse height distribution

from phase-resolved partial discharge pattern.

8.1.2 STATISTICAL QUANTITIESSeparateHdistributions are required to characterize the

PD activity associated with each half of the voltage cycle

(i.e. H+ and H-). The H

+ and H- distributions can be

characterized and described in terms of their mean,

variance, skew and kurtosis statistical features. In addition,

comparative statistical features characterize the difference

in the mean discharge magnitude, inception phase angleand cross-correlation (i.e. shape) between the distributions

positive and negative half cycles. Rule-base (RB) No.1 of

Figure 3 calculates and combines these statistical

parameters providing a statistical fingerprint of each H

distribution and consequently the PRPD pattern.

8.2 Rule-base No.2 Abstraction of PRPD

Pattern Descriptors

Rule-base No.2 of Figure 3 employs rules operating on

specific combinations of the statistical parameters derived


7/10


via rules in rule-base No.1 in order to characterize the

salient features of the observed PRPD pattern shape (i.e.

PRPD pattern descriptors) (Figure 6).

For example, these quantities can be used to determine

whether: the PRPD pattern representing PD pulse activity,

is symmetrical or asymmetrical; pulses occur regularly

across consecutive voltage cycles or appear intermittently;

bursts of high-energy partial discharges appear; partial

discharge pulses occur at the voltage zeros or peaks, etc.

While still relatively abstract, the interpretation of the

PRPD pattern is progressed, providing the basis for a more

comprehensive and intuitive description of the PD

behavior.

Figure 6. Knowledge rules for the abstraction of PRPD pattern descriptors.

8.3 Rule-base No.3 Explanation of PD

Behavior

Rule-base No.3 of Figure 3 interprets the PRPD pattern

descriptors to provide an insight into the physical nature of

the partial discharge activity generated by one or more

defects (i.e. PD behavior). Experimental studies of the PD

behavior of various defects in transformer oil led to the

capture of the following extracts of knowledge from thecompiled transcripts.

Figure7. A knowledge interpretation rules explaining the PD activity.

1. The discharges would occur around the zero crossing

when they are more dependent on the rate of change of

voltage rather than absolute voltage. A classic example

to illustrate this is a discharge in a void in solid

insulation. Each discharge relieves the electrostatic

field across the void. There is a space charge effect inthat charge is trapped on the inner surfaces of the

void.

2. A truncated sinusoidal pattern implies low energy

discharges are occurring at a sufficient rate to form

the pattern envelope. If such a pattern is centered on

the zero crossings it suggests a void. A void is a

capacitive discharge governed by the combined

electric field of the system voltage and local space

charge. Void geometry can be symmetrical in shape.

Ratio of mean

amplitude in +ve

and ve half cycles

close to one

IF scale symmetry

AND shape symmetry

THEN high correlation between activity

associated with +ve and ve half

Causal

Knowledge

Models

Knowledge

Rules

Scale symmetry i.e.

same amount of PD

activity associated with

+ve and ve half cycles

Cross-correlation

factor close to one

Shape symmetry: i.e.

same distribution shape

of PD activity

associated with +ve

and ve half cyclesindicates

indicates

High correlation

between activity

associated with +ve

and ve half cycles

Scale

symmetry

indicatesShape

symmetry

AND

Pulses located

on zero

crossings

Pulse magnitude at

0 o, 180o, 360o

greater than zero

indicates

IF pulse magnitude at 0 o, 180o, 360o

greater than zero

THEN pulses located on zero crossings

IF positive skewness

THEN Asymmetrical PD activity:

skewing of larger PD pulse

amplitudes to earlier phase

position

Causal

Knowledge

Models

Knowledge

Rules

Asymmetrical PD

activity: skewing of

larger PD pulse

amplitudes toearlier phase

position

Skewness of

distribution in +ve

and ve half cycles>> 0

indicates

Pulses located

on zero

crossin s

Dependent on

rate of change

o vol ta e

Issue of space

charge

indicates

indicates

KnowledgeRules

IF pulses located on zerocrossings

THEN PD dependent on rate of

change of voltage

AND issue of space charge

Conditions are the

same for both

olarities

Low energy

discharge

Phase position

symmetry

Chopped sine

wave

indicates

indicates

IF phase position symmetry

THEN conditions for partial discharge

are same for both polarities

IF chopped sine wave

THEN low energy discharge

Causal

Knowledge

Model derived

from knowledge

extract No.2 & No.3

Knowledge

Rules


8/10

S. M. Strachan et al.: Knowledge-Based Diagnosis of Partial Discharges in Power Transformers2663. When looking at the PRPD pattern three types of

symmetry can be identified; phase, shape and

magnitude. Phase position symmetry means that field

conditions for partial discharge are similar for both

polarities. Shape and magnitude symmetry suggests

that the defect is physically symmetrical.

4. A PD exhibiting characteristics of a void discharge

could not be located in bulk gaseous or liquid

insulation. Therefore the PD site can be narroweddown to those regions where solid insulation is present

or gas bubbles might form in the oil. For example, in

the case of a transformer, a bushing would be a

possibility.

Figure 7 illustrates how the knowledge interpretation

rules translate the PRPD pattern descriptors (output from

RBs 1, 2 and 3), into a diagnosis of the PD behaviour.

8.4 Rule-base No.4 Diagnosis of Defect

Geometry and Characteristics

The physical characteristics of the defect source (i.e.defect geometry) can be interpreted from the PD behaviour

(output from RB No.3) using the knowledge retained in

rule-base No.4 of Figure 3.

Figure 8 illustrates how knowledge interpretation rules

translate characteristics of the PD behaviour into

characteristics of the possible defect geometry.

Figure8. Knowledge interpretation rules explaining the PD defect.

8.5 Rule-base No.5 Classification of Partial

Discharge Defect

Specific defects exhibit certain PD behaviors that, while

influenced by the dielectric properties of the insulation and

external conditions affecting the sensor signals, (e.g.

propagation, attenuation, interference), are generally

symptomatic of the physical nature and geometry of the

defect site. Figure 8 illustrates how an understanding of

the linkage between the behavior of PD activity

characterizing a defect source and the characteristics of

the defect geometry of the defect make its classification

possible.

Figure 9 illustrates how RB No.5 uses knowledge of

the defect geometry (derived from RB No.4) to classify

the defect type associated with the input PRPD pattern. It

is evident that while on the whole PDs arising from

different defects may behave differently, they can share

similar characteristics. In addition, more than one class of

defect may exist simultaneously, or a particular class of

defect may morph into another as a result of continued

thermal or mechanical stress, or sustained discharging.

Therefore, providing a diagnostic explanation of the pulse

activity observed from successive PRPD patterns

generated over a period of time supports the engineer in

determining the nature, extent and severity of PD activity

emanating from a defect source as it develops.

Figure9. Knowledge interpretation rules supporting the classification and

location of partial discharge defects.

8.6 Rule-base No.6 Description of

Plausible Partial Discharge Defect

Source Location

Knowledge of the defect class alone will not provide anindication of the precise location of the defect source.

However, considered in conjunction with other techniques

(e.g. time of flight) 0, which can be used to identify the

general locale of the defect, knowledge of the defect class

can be a useful means of homing in on potential defect sites

within the identified area of interest.

Figure 10 illustrates how RB No.6 uses knowledge of the

defect type classification (derived from RB No.5) to

propose possible defect sites.

Void

defect

Capacitivedischarge

mechanism

Geometrically

symmetrical

Low energy

discharge

indicates

IF capacitive discharge mechanism

AND geometrically symmetrical

AND low energy discharge

AND dependent on rate of change of

voltage

THEN void defect

Causal

Knowledge

Model derived

from knowledge

bulletin No.2

Knowledge

Rules

Dependent on

rate of change

o vol ta e

AND

Defect site

geometrically

s mmetrical

Capacitive

discharge

mechanism

Shape

symmetry

indicates

indicates

IF PD dependent on rate of change of

voltage

THEN capacitive discharge mechanism

IF shape symmetry

THEN geometrically symmetrical

IF magnitude symmetry

THEN geometrically symmetrical

Causal

Knowledge

Model derived

from knowledge

extract No.2 &

No.3

Knowledge

Rules

Magnitude

symmetry

indicates


9/10


Figure10 A knowledge interpretation rules suggesting sites of PD activity

A summary of the diagnostic explanation output from

each of the system rule-bases, for the case study discussed,

is provided below:

PRPD Pattern Descriptors (RB No.2)

- Asymmetric PD activity

- Scale symmetry

- Shape symmetry

- High correlation between activity in positive andnegative half cycles.

PD behaviour (RB No.3)

- PD dependant on rate of change of voltage

- Space charge present

- Low energy discharges

Defect geometry (RB No.4)

- Capacitive discharge mechanism

- Site geometry symmetrical

Defect classification (RB No.5)

- Void defect

Defect location (RB No.6)

- LV bushing

9 CONCLUSION

This paper proposes a knowledge-based system as a

means of replicating the visual recognition and

interpretation process of PD experts when classifying the

defect sources using PRPD patterns. This requires an

understanding of the traits commonly associated with, and

relationship between, observed partial discharge activity

and responsible defect sources. This paper has proposed a

means of capturing this currently expanding knowledge

base and subsequently retaining relevant domain

knowledge within a knowledge-based expert systemdesigned to support plant experts, manufacturers and PD

experts in the classification, and continued understanding,

of the partial discharge phenomenon. It is envisaged that

this knowledge-based approach will offer the advantage of

supporting the classification of PD defects with a practical

diagnostic explanation of the detected PD activity, currently

absent in many of the data-driven and machine learning

approaches advocated in adjacent areas of research.

A more detailed diagnosis of the effects of PD activity

experienced by transformer plant can subsequently be used

to better inform plant maintenance and asset management

strategies. This supports more focused maintenance action

(condition-based maintenance), supplying maintenance and

plant engineers with instructive remedial advice.

The knowledge-based approach to the interpretation of

the PRPD patterns supports a modular system design. This

in turn provides a generic framework for the incorporation

of PD knowledge relating to different items of power

system plant. As such, an extensible system design enablesthe knowledge-bases to be easily augmented with fresh

knowledge (e.g. different defect types, plant types,

insulation mediums) as knowledge and understanding of

the PD phenomenon improves through ongoing empirical

research and the growth of practical engineering

experience.

ACKNOWLEDGMENT

The UK authors would like to thank the EPSRC for

supporting this research through the Supergen V UK

Energy Infrastructure (AMPerES) grant.

REFERENCES

[1] M. Wang and A.J. Vandermaar, Review of condition assessment of

power transformers in service, IEEE Electr. Insul. Mag., Vol. 18,

No. 5, pp. 8-17, 2002.

[2] B.H. Ward, A survey of the new techniques in insulation monitoring

of power transformers, IEEE Electr. Insul. Mag., No. 3, pp. 16-23,

2001.

[3] E. Gulski, J.J. Smit, R. Schomper, J. Slangen and P. Schikarski,

Condition assessment model for power transformers, 14th Intern.

Symposium on High Voltage Engineering, Beijing, China, Tsinghua

University Press, (ISBN 7-302-01581-3) pp.1-6, 2005.

[4] E. Gulski, D. Allan, T.R. Blackburn, A. Contin, E. Gockenbach, E.

Lemke, L. Lundgaard, T. Mizutani, G.C. Montanari, M. Muhr, T.

Okamoto, B.T. Phung, H. Sedding, HJ Breen and F.J. Wester,.

Knowledge rules for partial discharge diagnosis in service, (CigrBrochures, 226). Paris: Cigr, 2003.

[5] J.C. Devins, "The Physics of Partial Discharges in Solid Dielectrics",

IEEE Trans. of Electrical Insulation, Vol.19, No. 5, pp 475-495,

1984.

[6] S.A. Boggs, Partial Discharge: Overview and Signal Generation,

IEEE Electr. Insul. Mag., Vol.6, No.4, pp33-39, 1990.

[7] G.P. Cleary and M.D. Judd, UHF and current pulse measurements

of partial discharge activity in mineral oil, Science, Measurement

and Technology, IEE Proc., Vol. 153, No. 2, pp. 47 54, 2006.

[8] M.D. Judd, L. Yang and I.B.B. Hunter, Partial discharge monitoring

for power transformers using UHF sensors Part 2: Field Experience,

IEEE Electr. Insul. Mag., Vol. 21, No. 3, pp. 5-13, 2005.

[9] M.D. Judd, S.D.J. McArthur, J.R. McDonald and O. Farish,

"Intelligent Condition Monitoring and Asset Management: Partial

Discharge Monitoring for Power Transformers", IEE Power Eng. J.,

pp. 297-304, 2002.[10] S. Meijer, P.D. Agoris, J.J. Smit, M.D. Judd and L. Yang,

Application of UHF diagnostics to detect PD during power

transformer acceptance tests, IEEE Int. Symp. Electr. Insul.(ISEI),

pp. 416-419, Toronto, Canada, 2006.

[11] E. Gulski, Computer Aided Recognition of Partial Discharges using

Statistical Tools, Ph.D. Thesis, ISBN 90-6275-728-6, Delft

University Press, 1991.

[12] S.D.J. McArthur, S.M. Strachan and G. Jahn, The Design of a

Multi-Agent Transformer Condition Monitoring System, IEEE

Trans. Power Systems, Vol. 19, pp. 1845-1852, 2004.

[13] R. Candela, G. Mirelli and R Schifani., PD Recognition by Means

of Statistical and Fractal Parameters and a Neural Network, IEEE

Trans. Dielectr. Electr. Insul., Vol.7, pp. 87-94, 2000.

Possible location of

defect site is LV

bushin A Phase

Void

Defect

Time of Flight

Output is top of A

phase LV winding

indicates

IF Void Defect classification

AND Time of Flight location is top of A

phase LV winding

THEN Possible location of defect site is LV

bushing A Phase

Knowledge

Rules

Causal

Knowledge

Model derived

from knowledge

extract No.4


10/10

S. M. Strachan et al.: Knowledge-Based Diagnosis of Partial Discharges in Power Transformers268[14] C. Cachin and H.J. Wiesmann, PD Recognition with Knowledge-

based Pre-processing and Neural Networks, IEEE Trans. Dielect.

Electr. Insul., Vol.2, pp. 578-589, 1995.

[15] E. Gulski, A. Krivda, Neural Networks as a Tool for Recognition of

Partial Discharges, IEEE Trans. Electr. Insul., Vol. 28, pp. 984-

1001, 1993.

[16] CIGRE Task Force 15.11/33.03.02, 2003, Knowledge Rules for

Partial Discharge Diagnosis in Service, CIGRE Technical Report,

2003.

[17] G. Booch, The Unified Modelling Language User Guide, Second Ed,

Addison Wesley, 1998.[18] G.F. Luger, Artificial Intelligence Structures and Strategies for

Complex Problem Solving, Addison Wesley, Fourth Edition, 2002.

[19] C.J. Bennoch and M.D. Judd, A UHF system for characterising

individual PD sources within a multi-source environment, Int.

Symp. High Voltage Engineering, Delft, The Netherland, 2003.

Scott Strachan received the B.Eng. (Hons.) and Ph.D.

degrees from the University of Strathclyde. He currently

holds the post of Research Fellow within the Institute of

Energy and Environment. His research interests include,

plant monitoring, asset management, data mining,

knowledge management & engineering and intelligent

systems applications in power engineering.

Susan Rudd received the B.Sc. (Hons.) from the

University of Strathclyde in 2004. She is a research

student in the Institute for Energy and Environment. Her

research interests include knowledge engineering,

intelligent system applications in power engineering and

condition monitoring.

Stephen McArthur (M93-SM07) received the B.Eng.

(Hons.) and Ph.D. degrees from the University of

Strathclyde. He is a Reader in the Institute for Energy

and Environment. His research interests include

intelligent system applications, condition monitoring and

multi-agent systems.

Martin D. Judd (M02-SM04) graduated from the

University of Hull in 1985 with a B.Sc. (Hons.) degree in

electronic engineering. His employment experience

includes work at Marconi Electronic Devices Ltd and

EEV Ltd, both in Lincoln, England. Martin received a

Ph.D. from the University of Strathclyde in 1996 for his

research into the excitation of UHF signals by partial

discharges in gas-insulated switchgear. From 1999 to 2004 he was an

EPSRC Advanced Research Fellow. His fields of interest include high

frequency electromagnetics, fast transients and their measurement and

partial discharges. Dr Judd is a Senior Lecturer within the Electrical Plantand Diagnostics group at the University of Strathclyde.

Sander Meijer received the M.Sc. degree in electrical

engineering in the field of electrical power systems in

1995. In 2001 he obtained the Ph.D. degree from the

Technical University of Delft. At present, he is assistant

professor in the department of high-voltage technology

and management. His main research topic is in the field

of high frequency measuring techniques for insulation

diagnosis of gas insulated systems, power transformers

and high-voltage cables and electromagnetic compatibility. He is active in

international working groups and is the secretary of Cigr WG D1.03 (gas-

insulated systems).

Edward Gulski received the M.Sc. degree in

information technology in 1982 from Dresden

University of Technology in Germany. In 1991 he

received the Ph.D. degree from Delft University of

Technology in The Netherlands. In 2004 he received

the Doctor Habilitatus degree from Warsaw University

of Technology in Poland. At present, he is an associate

professor involved in education and research in the

field of insulation diagnosis of HV components and Asset Management.

He is member of the Executive Board of KSANDR organisation and

responsible for research in education. He is chairmen of Cigre Working

Group D1.17, ''HV asset condition assessment tools, data quality, and

expert systems'', member of Cigre Study Committee D1 Materials and

Emerging Technologies, convenor of Cigre Task Force D1.33.03 Partial

Discharges Measurements, member of IEEE Working Group 400.3 and

member of the IEEE Insulating Conductors Committee.

Documents

Knowledge-Based Diagnosis of Partial Discharges in Power Transformers