Transformer Failure Detection Using Dissolved Gas Analysis

(DGA)

1
Quist Mirable Jethro, 2Amevi Acakpovi
1. Accra Institute of Technology – Open University of Malaysia (AIT – OUM), Graduate Studies-Mechanical
Engineering, Email: [email protected], Accra, Ghana;
2. Accra Technical University, Electrical/Electronic Engineering Department, Email: [email protected],
Accra, Ghana.

Abstract – A transformer is a device that transfers electrical energy from one circuit to another through
inductively coupled conductors – the transformer coils.
Transformer is very important and plays a vital role when it comes to transmission and distribution of
electrical energy. Transformer failures are disruptive to electrical grid and costly. Transformer insulation
failures from various blasts like bombs or lightning strikes are major culprits. Methods and techniques used in
data collection, analyses, computer simulation and results interpretation on transformer failures are presented.
Dissolved Gas Analysis (DGA) of the insulation was used to detect the health condition of transformers.
Secondary data of burnt transformers were also considered.
A multi-layer casing for transformers was designed to protect transformers. The casing will protect
transformers from detonation by bombs, rocket and other explosives, and lightning strikes.
A stabilized power supply will mean higher productivity in the country as a result of good energy supply.

Keywords: distribution Transformer, void formation,transformer winding, transformer failure, box model,
deflectors, Dissolved Gas Analysis Test.

1. Introduction.
Transformer is very important and plays a vital role when it comes to transmission and distribution of
electrical energy in armed forces in developing countries. The increasing failure rate of distribution
transformers in power systems can be regarded as one of the dangerous phenomenon of distribution system
management; especially in developing countries in the armed forces (military), like Ghana Armed Forces,
Liberia Armed Forces, Nigerian Armed Forces, Sierra Leon Armed Forces, Guinea Armed Forces. A
distribution failure transformer really worries the stability and reliability of power supply to clients.
DGA is a well established diagnostic method, where oil samples are taken routinely and the composition of
the gases dissolved within the oil analyzed. There are mainly eight gases of interest for diagnostics. Namely,
Hydrogen(H), Methane (CH4), Ethane (C2H6), Ethylene(C2H4), Acetylene(C2H2), Carbon Monoxide(CO),
Carbon Dioxide(CO2), Oxygen (O2), Nitrogen(N2).
The amount and type of gases formed and dissolved into the oil can be determined with the help of Dissolved
Gas Analysis (DGA). Similar to a blood test or scanner examination of human body, dissolved gas analysis
can give an early diagnosis and increase the chances of finding the appropriate cure [ Ranjana and Amarjit
Singh, 2017].
2. 0.Background on Transformer failures by age based on Dissolved Gas Analysis.

`1
 There are general failures of Distribution transformer systems but the aspect of bomb blast, explosives
and mine detonations, bush burning and gun shot from hunters also exist.
This risks can be mitigated by examining the impact of Distribution Transformer failure on developing
nations, mitigate causes and effects of Distribution Transformer system failure in developing nations, to
promote efficiency and good energy supply in developing nations as a result of an inform policy on the
mitigating the economic consequences of Distributing Transformer Failures in developing nation.
Replacement estimate of distribution transformer
Table 1. below explain prices and the installation cost of some PMT transformers
VOLTAGE LEVEL TYPE KVA UNIT PRICE INST. FEE TOTAL
PMT/GMT RATING GH¢
33/0.433KV PMT 50 9,922.53 606.39 10528.92
33/0.433KV PMT 100 18,226.06 618.28 18844.36
33/0.433kv PMT 200 25,005.84 642.06 25647.90
33/0.433KV PMT 315 29,591.48 713.40 30304.88
11/0.433KV PMT 50 8,998.91 606.39 9605.30
11/0.433KV PMT 100 16,469.48 618.28 17,087.76
11/0.433KV PMT 200 23,756.70 642.06 24,398.76
11/0.433KV PMT 315 26,945.02 713.40 27658.42
TOTAL GHc 164,078.28

In 2010 about 46 various distribution transformers were damaged in the western region electricity company of
Ghana including those in military establishments, prison establishments, Customs Exercise and Preventive
Service Institutions, police stations, hospitals, universities, colleges, schools and other governmental and
Nongovernmental Organizations, which amounted to about GHc 865,328.54. It can be seen that losing so
many Ghana cedis in a region because of transformer failure within a year is very unwelcoming. Transformer
failures in general and Distribution Transformer systems failure have impact on the economy of developing
nation. Distribution Transformer system failure in developing nations adds to insufficient energy production
and power outage problems.
Insufficient energy production and power outage have effect on the efficiency and good energy supply in
developing nations. As a result of that there is the need for an inform policy on the mitigating the
consequences of Distributing Transformer Failures in developing nation.
Transformers are considered the most crucial and expensive piece of plant within a transmission system. Most
transmission systems currently have large populations of aging transformers. With the growing demand for
electricity, the loading of transformers is increasing. Current economic strategies call for reduced maintenance
as well as capital expenditure. These challenges, which face utilities world-wide, necessitate improved
management of transformers. The impact of a distribution transformer failure can be catastrophic. It would
therefore be beneficial to know the risk status of the transformer population in order to facilitate better asset
management, optimization of maintenance, refurbishment and replacement strategies which will ensure
maximum asset utilization and minimize system risk. Statistical analysis on historical transformer failure data
is therefore required, in order to obtain a model to determine the probability of failure of the transformers
currently in service. Currently, ECG Transmission has over 550 Power transformers in service. It is of interest
to see the age profile of the existing population, in terms of years since manufacture, since reliability is often
related to age. From an analysis performed on insurance claims for transformers, it was found that the average

`2
age of failure was approximately 15 years [Bartley, W.H. (2003). “Analysis of transformer failures, imia
wgp33(03). In 36th Annual Conference of the International Association of Engineer’s in Insurers,
Stokholm,(2003)”]. The expected design life of a power transformer is 40 years. This figure is reduced,
sometimes substantially, depending on the utilization of the transformer, i.e. its loading or the environment to
which it is exposed.

Figure 1: Breakdown of transformer age.

The breakdown of age within the population of Eskom Transmission transformers is shown in Figure 1. From
this graph it is evident that the majority (61%) of the population is above the age where probability of failure
would be expected to be high and 14% is beyond the expected design life. The average age of transformers in
Transmission is currently 29 years, showing an aging, high risk fleet, according to traditional thinking.
With the increased age of the transformers and the following additional risk factors [Bartley, W.H. (2003).
“Analysis of transformer failures, imia wgp33(03). In 36th Annual Conference of the International
Association of Engineer’s in Insurers, Stokholm, 2003)”]:
i.Increased utilization of equipment. ii. Deferred capital expenditure. iii. Reduced maintenance expenses. iii.
Increased power consumption/load demand
It seems inevitable that the transformer failure rate within Transmission can be expected to increase rapidly in
the near future. Therefore, it seems prudent to investigate the development of a model that will enable
determination of probability of failure of these transformers, and hence improve asset management.
Eskom has adopted an asset management tool for optimizing asset replacement strategies. The Condition,
Criticality and Risk Assessment (CCRA) model enables the phasing and structuring of project plans based on
suitable cost-to benefit ratios. A complete CCRA is required to avoid replacement strategies based on age and
condition alone, in an attempt to reduce failure rate. The purpose of this model is to incorporate the
consequence of failure as well as the risk into the investment decision making process.
Transformer Aging
Bartley (2016) did not categorize "age" as a cause of failure. According to him, aging of the insulation
system reduces both the mechanical and dielectric-withstand strength of the transformer. As the transformer
ages, it is subjected to faults that result in high radial and compressive forces. As the load increases with
system growth, the operating stresses increase. Bartley lamented that in an aging transformer, the conductor

`3
insulation is debilitated to the point where it can no longer withstand mechanical stress of a fault and these can
cause transformer to failure. Turn to turn insulation then suffers a dielectric failure, or a fault causes a loss of
winding clamping pressure, which reduces the transformer's ability to withstand future short circuit forces.

Table 2.3. Distribution of Losses by Age of Transformer

Age at fault Number of Failures Cost of Failure
0 to 5 years 9 $11,246,360
6 to 10 years 6 $22,465,881
11 to 15 years 9 $3,179,291
16 to 20 years 9 $10,518,283
21 to 25 years 10 $16,441,930
Over 25 years 16 $15,042,761
Age Unknown 35 $207,734,306

The table above displays the distribution of transformer failures by age.

The average age at failure was 18 years. The analysis of the ages of the transformers mentioned on the table
2.3, above shows that as transformers are growing they turned to fail. The age of transformers deserves
special attention.
Looking into the analysis above, the author of the paper is right because it could be seen from the table and
the explanation that transformer fails as they are catching up by age. But not withstanding, transformers under
certain ages must not also be left out because they are not too old. Looking into the table above, transformers
that failed between 0 to 5 year and 16 to 20 years are the same even though 16 to 20 years transformers are
older. Even the cost for transformers between 0 and 5 years cost more than transformers between 16 to 20
years. Even the difference between in terms of cost of transformers between 0 and 5 years and over 25 years is
not much.
This means that when transformers are installed equal care must be taken on all ranges of age. It also means
that transformers failing because they are older are not always true. All transformers, irrespective of age must
be given equal attention because they are very important device and also costly.

The ideal strategy is a life assessment or life cycle management programme, that sets loading priorities, and
provides direction for the identification of:
Transformer defects that can be corrected
Transformers that can be modified or refurbished

`4
 Transformers that should be relocated
Transformers that should be retired
Significance of the study.
There have been many researches done on the causes of transformer failures in developed countries and
some developing countries in Asia. Very little has been done in Africa, especially Ghana. Even most of the
research done in the developed countries is on power transformers. Since environmental, culture, attitude,
terrain and other situations differ from developed countries to developing countries, knowing the actual causes
of transformer failures in Ghana, will contribute immensely to the body of knowledge. It will also help to
maintain quality and efficient power supply and to improve the economy. For example in western region
alone during 2014 year about 46 distribution transformers failed and amounted to about GHC 865,328.54,
when taken Ghana as a whole for one year it will be very disastrous. It will also benefit the government or
utility companies when purchasing transformers for installation. The findings will be very useful to
engineering companies and student because it will be a source of reference for electrical engineering students
in the country. It will also help in designing alternative solutions to mitigate the problem, and will also add to
the body of knowledge.
Unplanned electrical outages can cost millions of dollars for manufacturers, industrial process plants,
commercial buildings and other organizations. A Lawrence Berkeley National Laboratory study estimated that
the annual cost of power interruptions in US is about $80 billion - 72% allocated to commercial customers
and 26% to industrial customers. And that’s just reflective of failures on the supply side. Electricity is a key
infrastructural element for economic growth.
Distribution transformer is employed to convert the electricity of higher voltage (11-33 kV) to a lower voltage
(250-433V). Little research conducted in the country has revealed that many transformers got damaged very
often which affects many customers in Ghana. The failure rate has been very high also in the developing
countries. Most of the failures have been credited to lightning strikes, overloading of the transformer and
insulation breakdown. Another vital area that researchers have not done much work is proper database which
monitors transformer failures in the country. It will also help to address better control and protection devices
for distribution and power transformers. Another reason being that failure of transformers impedes the growth
of a nation because investors would not want to invest in nations with frequent power outages. Stabilized
power supply means higher productivity in the country. Since the Government is seriously embarking on
nationwide electrification project, transformers are the quintessential equipment for power transmission and
distribution. Because of high failure rate the resultant reliability and safety implications particularly with
distribution transformers, in-depth valuation is indispensable to ascertain a healthy transformer is very
imperative. Experience has shown an increasing number of transformer explosions and fires in all types of
power and distribution transformers worldwide (internationally). Establishing the failures will help reduce the
failure rate.
All the findings from developed to developing countries vary from one place to the other, this also point to the
fact that distribution transformer failures in Ghana needs to be given a serious attention.
In summary, the following contribution will be made;
Causes of transformer failures will be establish since their failure is costly and affect the quality of service.
2.1. Methods and technques.
Thuraijah Haigh and Amaratunga (2016), describes research methodology as the overall approach to the
design process from the theoretical underpinning to the collection of data and analysis of data. These chapters
as well present the research approach and the methods used in achieving the objectives for the research. The
implications and concerns on the choice of the study approach, and framework validation approach are also
justified. Methods and techniques used in sampling, data collection, analyses, computer gas simulation and
interpretation are presented. It also discusses the concept formation through critical thinking, and is based on

`5
theories derived in literature, proposed study area. It also elaborates on the proposed study design, target
population, sample size and sampling procedure, as well as the proposed methods of lab test, simulation, and
data analysis methods. Methodology refers to the major approaches or paradigms, which guide the conduct of
a research. Methods conversely refer to the specific research techniques or tools which a researcher employs
to collect the relevant data for answering her/his research questions. The decision to use particular research
method is generally determined by their appropriateness, the purpose of the study, the research problem,
research questions, objectives and the researcher’s preferences.
The research employed different methods; Combination of research methods. Lab testing were key factor
using simulation software available to simulate some findings or results.
Secondary data of burnt transformers will be considered. Multiple methods will be employed.

2.2. Research Questions

The main aim of the research questions was to develop a framework for the categorization of transformers
in developing countries. The research was motivated by the continuous light outs in Ghana, with the lack of
appropriate means of determination of such causes and ways of solving them(Asamoah, 2015) . The central
research question is therefore: i. What is the appropriate method of determining causes of light outs and in
case of transformer failures that would meet the needs of power industries and clients in developing countries.
To appropriately handle the main research question, further sub-questions were considered. The entire
research methodology thus evolved around the research question and research objectives.
The framework for integrating risk during categorization of transformer failures in developing countries
can be investigated and the various risk indicators identified. The research also viewed the investigation to be
conducted as practical as possible rather than an abstract of ideas and place. A rigorous process of lab test was
conducted by electrical engineers engaged directly on electrification transformer installation and replacement
projects and existing procedures were also followed with regards to physical inspections, repair and
replacements.
Epistemologically, this research was geared towards the positivist position, with the belief that intriguing
process of categorizing transformer failures in developing countries could be explored further through a
systematic but cognitive approach. The research could be carried without bias and that objective conclusions
could be drawn from experimental lab tests, physical inspections and data collected through DGA, acidity,
moisture content test and simulation of gasses.
Two research strategies are appreciated in traditional research: qualitative and quantitative research
work. Though the philosophical stance underpinning the research has been discussed, Bryman (2014), holds
that there is the need to clarify the orientation if the researcher has to conduct the research. Due to the
difficulty in applying one method wholly, researchers have combined both approaches referred to as the
mixed approach.
2.3. Strategy Adopted in this Research
The research adopted the mixed approach using both qualitative and quantitative strategies in collecting
data. The approach to be used in doing this work is broken down into desktop study through literature, field
experiment and observation through lab tests, probability and statistical analysis of results, generation of
conceptual framework, development of models and the validation of the results of the model. The concepts
developed are subject to critical laboratory tests, probability and subsequent statistical analysis. The result of

`6
 the lab test constituted a basis for the model to be developed. It was further validated qualitatively through
vetting and testing in subsequent stages in this work. An experimental approach was adopted in this research
using both quantitative and qualitative methods. Quantitative research methods were chosen due to the
engineering nature of the problem, to deduce the relationship between risk factors which has a measurable
effect on the transformer failures in developing nations. The process involved lab test, DGA, simulation of
gases, model development, validation, probability and statistical analysis of results obtained through tests.
According to Frazer and Lawley (2014), a research design is a plan of information required to answer
research problems and how such information can be collected. Bryman (2004) and Yin (2013), hold that
research design enables the researcher to connect empirical data to its conclusions, in a logical sequence to the
initial research question of the study. Research design is the structure that guides the execution of the
technique for collecting and subsequently analyzing data. It is, therefore, the framework within which the
research method is employed. This sub-section outlines the various frameworks available for research data
collection and analysis. It first discusses research design frameworks and then justifies the choice adopted by
this research. Finally, it describes attitude measurement, which was adopted for the testing and validation of
the designed framework.
2.4. Research Design Options
The research design option selected depends on: the type of research question, the control the
investigator has over actual behavioral events and the focus on contemporary as opposed to historical
phenomena (Yin, 2003). Bryman (2004) described five main research design options: i. Experimental, ii.
Cross-sectional, iii. Longitudinal, iv. Case study and
v. Comparative research designs. It is important to realize the differences within these options and their
relationship with research designs. Yin (2013), classified the options into experiment, survey, archival
analysis, history and case study. Bryman (2014), pointed out that the selection of a particular design should be
done to reflect the importance with which the researcher attaches to: i. expressing causal connections between
variables. ii. generalizing to larger groups of individuals than those actually forming part of the investigation.
iii. understanding behaviors and the meaning of that behavior in its specific social context. iv. having a
temporal time appreciation of social phenomena and their interconnections.
Table:3.1.Research Strategy and Research Design.
Research Research Strategy (Typical Form)
Design
Quantitative Research Qualitative Research
Experimental Typical Form: Most Researchers No typical form
using an experimental design
employ qualitative comparison
Between Experimental and
control
groups with regards to the
dependent variable
Cross- Typical Form: Survey Research or Typical Form: Qualitative interviews or focus
Sectional structured observation on a sample groups at a single point in time. Qualitative
at a single point in time . Content content analysis of documents relating to a single
analysis on a sample of documents. period.

`7
Longitudinal Typical Form: Survey research on a Typical Form: Ethnographic research over a long
sample on more than one occasion, period, qualitative interviewing on more than one
as in panel and cohort studies. occasion or qualitative content analysis of documents
Content analysis of documents relating to different time periods. Such research
relating to different time periods. warrants being dubbed longitudinal when there is a
concern to map change.
Case Study Typical Form: Survey research in Typical Form: The intensive study by ethnography
which there is direct comparison or qualitative interviewing of single case which may
between two or more cases as in be an organization on life, family or community.
cross cultural research.
Comparative Typical Form: Survey research in Typical Form: Ethnographical or qualitative
which there is direct comparison interview research on two or more cases.
between two or more cases, as in
cross cultural research.
Source: Bryman (2014).
Yin (2013), holds that in the selection of a relevant strategy, the listed factors in table 3.1.may be
considered. Bryman (2014), summarizes the various research strategy and design options in their typical
forms under the two main research strategies discussed earlier.
2.5.Research design
Two main methodological approaches to the conduct of research that have been identified in literature
are quantitative and qualitative approaches (Cohen, et al., 2012,Bryman, 2015, 2015). Bryman (2015) argues
that even though many writers on methodological issues attempt to distinguish between these two paradigms,
the status fundamental, others regard it as no longer useful and still others as false. The author further
contends that quantitative research as one which lays emphasis on quantification in the collection and analysis
of data which entails a deductive approach to the relationship between theory and research, in which the
accent is placed on testing of theories and embodies a review of social reality as an external and objective
reality. Most of this distinction is unclear, because while some writers regard the distinction as Conversely,
Bryman conceives qualitative research as a research strategy which lays emphasis on words rather than
quantification in collection and analysis of data and which predominantly emphasizes an inductive approach
to the relationship between theory and research, in which the emphasis is placed on the generation of theories
and also embodies a view of social reality as a constantly shifting emergent property of individuals’ creation
(Bryman, 2004) .The researcher will employ mixed method (both qualitative and quantitative in data
collection) research design for the study as that is deemed appropriate for the study.
The research design adopted in this thesis is a modelling and simulation work in the field of electrical
engineering. Precisely, the design firstly consists of developing an adaptronic transformer casing resistance to
explosives, bombs and bullets through the use of stealth technology, blast bags, concrete walls, explosive
resistant glass and bomb proof materials as outlined in chapter one. Laboratory tests were performed
including monitory DGA (Dissolved Gas Analysis). DGA gas including, H 2, CH4, C2H6, C2H4, C2H2, CO, CO2,
O2, N2, were analysed and simulated. It can therefore be classified under experimental research method. The
research method is quantitative. The objective of quantitative research is to develop and employ mathematical
models, theories and/or hypotheses pertaining to phenomena. Similarly, this research study uses analytical
modelling which involves a lot of mathematical models and theory to prove a better way of analysing
transformer failures in developing countries. It is also known that quantitative data is any data that is in
numerical form such as statistics, percentages.

`8
 Again, DGA generated gases, H2, CH4, C2H6, C2H4, C2H2, CO, CO2, O2, N2, were tested using numerical data
and statistics in order to assess the aging failure rate of distribution transformer systems failure to improve
upon energy supply system . This method of modelling and simulating gasses is very relevant because it
provides a platform to assess the expected result. Simulations have also helped finding tune the design in case
the expected results were not met. The whole design process was then modified and the process was
continuously repeated until satisfaction. This saves cost and reduces the risk of getting unsatisfactorily results.
The relevance of the design method is also proven by similarities with related work studied in the literature
review. Transformer distribution systems failure has always been categorized through a process of modelling
and simulation.
Based on the fore gone discussions, an experimental design adopted for this research to enable the author
achieve his research objectives. The choice of this strategy is because the research design employ qualitative
comparison Between Experimental and control groups with regards to the dependent variable ,computer
simulation, official statistics, documentation and diaries.
Dissolved Gas Analysis

DGA is a well established diagnostic method, where oil samples are taken routinely and the composition of
the gases dissolved within the oil analyzed. There are mainly eight gases of interest for diagnostics, outlined in
1.6 Table 1.6. Diagnostic gases.
Gas Chemical formula

Hydrogen H2

Methane CH4

Ethane C2H6

Ethylene C2H4

Acetylene C2H2

Carbon Monoxide CO

Carbon Dioxide CO2

Oxygen O2

Nitrogen N2

The gas concentrations are measured and a trend of gas production is recorded. A developing fault
within the transformer will lead to larger quantities of gases being generated and dissolved within the oil.
The speed at which the fault gases are produced is an indication of the magnitude of the fault. For this
reason, the rate of production is often considered to be of greater importance than the actual gas
concentrations. The fault gases are also generated in small quantities during normal operation, due to
natural aging. For this reason, relatively high concentrations of gas can also not be alarming if those
concentrations are not increasing rapidly.

There are various methods of interpreting the gas results. These include well known methods such as: Total
Dissolved Combustible Gases (TDCG), Duval's triangle, Roger's ratios, etc. [ IEC 60599,1999 and IEEE
std, 2015]. An additional method developed by analyzing the DGA signatures of failed transformers and
relating the identified patterns to gassing transformers in service to determine the fault cause and potential
failure mode has been developed in [S Govender and A Singh, 2014].

The fault gases are produced in different quantities over a range of temperatures, due to the breakdown of
the solid and oil insulation. An indication of how the gases are produced relative to temperature is shown in
Figure 3.6. The temperature of the fault is an indication of the type of fault present. A partial discharge
fault will generate relatively low temperatures and therefore, the main gases indicating this type of fault are
Hydrogen and Methane. Arcing will generate extremely high temperatures and therefore, Acetylene is the
prominent gas for this type of fault. Bare metal faults generate heat in the range 300-700 CSo_ in this case,
the gases that are expected to present would be Ethane and Ethylene. Since the paper is made of cellulose
molecules, it is expected that any fault involving paper covered components would yield higher

`9
concentrations of Carbon Monoxide and Carbon Dioxide. CO/CO2 is a ratio that can be used to determine
whether the fault is thermal or dielectric[IEEE std. C57.104, 2014].

Since these gases are naturally occurring within the atmosphere, some caution must be exercised when
interpreting these results.

Figure 3.6: Gas concentrations at different temperatures [P Gervais and J Aubin,21995].

`10
In work done by Cigre working group A2-111 [In Cigre, A2-111, 2012], it was found that reliable statistical data
is often difficult to get since relatively few failures of distribution transformers are experienced over long periods
of time. A method of life determination based on measurements of condition data and mechanisms of breakdown
and failure, established from physics and chemistry tests, has been developed. The condition data is obtained
from forensic analysis of failed units on site and a \Degradation Model" based on DGA and oil condition. A \Life
Model" developed using the influence of temperature on the degradation of paper insulation and the model
developed in [AIEE Transactions, 67, 1948] is then used. Failure rates based on condition and service time are
then determined. Failure analysis using three tests: oil analysis, Furan derivatives analysis and HMM analysis is
described in [M.A.R Uzair, M Mohiuddin, and M.K Shujauddin, 2013]. Oil condition and diagnostic gases (used
in DGA, laboratory and online monitoring), as well as Furanic analysis (DP) are used to determine performance
while HMM are used to determine probability of failure. In this method, similar to [HICSS 2006], the state within
the HMM is hidden and the outcome, which is dependent on state is visible. Hidden variables are co-related
through a Markov process to determine the outcome, rather than independent of each other. In this way, the
model is trained in a similar way to an ANN.

2.7 Condition, Criticality and Risk Assessment Model Asset management model An asset management model
(CCRA) has been adopted by Eskom for optimizing investment decisions. The outcome of this model is
dependent on the determination of probability of failure based on a HI that is developed using the life assessment
parameters . It is there-fore necessary to evaluate the basis of this model, as well as the use of the HI parameters.

`11
2.8. Overview of methodology

Eskom has adopted an asset management tool for optimizing asset

refurbishment/replacement/retirement strategies. The Condition, Criticality and Risk Assessment (CCRA)
model enables the phasing and structuring of project plans based on suitable cost-to-benefit ratios.

The inputs to this model include plant age, condition/HI, probability of failure and consequences of failure. The
output of the model is the overall risk allowing analysis of cost/benefits of replacement/refurbishment. The
timing of replacement is optimized according to the strategy shown in Figure:3.6.

Figure 3.7: Optimization of asset replacements [WERF, 2009].

Bathtub curves, as shown in Figure 3.7, developed from Weibull distributions depict the probability of failure of
the plant as a function of age (P(age)). There are three distinct regions in a bathtub curve:

1.Infant mortality: indicated by a high failure rate in the first few years after manufacture, that decreases over
time. These failures are generally attributed to inherent design defects or manufacturing errors.
2.Random failures: indicated by a at region (constant failure rate) over the age range in the middle of the expected
design life, where random failures are expected throughout the population.
3.Wear out: indicated by an increasing failure rate with an increase in age. This region has the highest failure rate.
This area is related to wear out and is the area of most interest in this study.
A HI for each plant type is determined, based on the condition of the plant, relative to end-of-life. The HI is
a quantification of condition measurements that are taken on the transformer and an overall score is
obtained. This is then used to determine a probability of failure as a function of HI (P(HI)). Bathtub curves,
developed from Weibull distributions depict the probability of failure of the plant as a function of age
(P(age)). If P(HI) > P(age), then P(age) is modified according to P(HI), i.e. the bathtub curve becomes
steeper, indicating an increased rate of aging, and a new probability of failure is determined as a function of
both age and HI
(P(age, HI)). This final probability of failure is then used for further analysis. This then comprises the \Condition"
portion of the CCRA model

`12
 Figure 3.8: Example of equipment bathtub failure curve [WERF, 2009].

Figure : 3.9. Effect of Health Index modifier on probability of failure curve [CCRA Economic Life Manual,
2009].

The consequence (or impact) of failure of the plant is then determined, to ascertain the impact such a failure
would have on the system, should it occur. This is done by analyzing:

• the position of the plant in the network

• whether it has (N - 1) redundancy
• whether or not contingencies are available
• what consequential damage could occur or is expected should the failure of the plant be catastrophic the
impact of adjacent plant failing at the same time, or during the outage of the initial failure.
Different scenarios are drawn up for each failure event. These scenarios are developed from the Emergency
Preparedness Plans (EPP) of each substation in which the transformers are installed. EPPs outline the
`13
processes to follow in the case of an emergency or loss of plant, to recover load. The consequences include
the following:

• Replacement cost of failed plant

• Cost of emergency repairs, if applicable Cost of damage to adjacent plant
• Cost of unserved energy or customer interruption, should it occur Safety or environmental costs • Each
scenario is then analyzed, including its probability of occurrence. This then comprises the \Criticality"
portion of the CCRA model and is a monetary value.
Risk is defined as the product of the probability of an event occurring and the consequences associated with that
incident or, the frequency and severity of the losses [Routledge, 2009].

A method similar to that employed in the CCRA model is that of a time-dependent failure probability,
based on available condition data [Z Zhang, Y Jiang, and J.D McCalley Iowa State University]. The
primary objective of this method is for the optimization of required maintenance interventions. A
probability of failure based on the commonly used Weibull distribution is used to determine equipment
time-to-failure. A Bayesian approach is used with this model due to limitations in empirical data available.

By using a Bayesian approach, uncertainties due to lack of information are expressed via probability
distributions. Unknown parameters, such as shape and scale in the Weibull distribution are considered as
random variables. In this way, Baye's theorem is used to determine the posterior probabilities based on
condition data. This acts as a modification of the expected prior probability of failure that is assumed, based
on age and failure rate, as done in [Jagers J.N, Khosa J, de Klerk P.J, and Gaunt C.T, 2013].

Risk = Probability x Consequences (3.1)

Risk is defined as the product of probability and consequence or, the frequency and severity of the losses
[Routledge, 2013]. The risk of each item of plant is calculated from the probability of failure that was
calculated as a function of age and HI and the consequence cost that was calculated from the impact of the
failure. This comprises the \Risk" portion of the CCRA model and is also a monetary value.

Due to limitations in financial resources, it is necessary to perform some prioritization in justifiable

projects. This is done in order to justify spending alternatives. The benefit versus the risk of implementing
a project this year rather than delaying it by a year or more is determined. In order to do this, the benefit to
cost ratio is evaluated. The risk value is compared to the cost of refurbishment/replacement of the asset. If
this ratio is very high, the refurbishment/replacement strategy is prioritized. If it is very low, the
refurbishment/replacement strategy will more than likely be deferred to a later time. This is useful in
determining the impact on the business of a decision to delay projects.

An accurately defined probability of failure, relevant to the population of transformers being analyzed is
critical, since it forms the basis of this calculation. With a probability of failure that is not representative of
the population being analyzed, the risk analysis is inaccurate and decision-making will be awed and
unoptimized. This leads to an increased risk of wasteful expenditure and an attempt to decrease the asset
failure rate, rather than the real risk associated with the impact of the failure, which is the ultimate goal.

. Failure definition

Asset management tools are refined based on probability and subsequent consequences of failure. Failure
can range from anything between a minor defect which can be repaired on site, to a catastrophic event,
necessitating the replacement of the failed transformer, as well as adjacent plant and possibly interruption of
supply. Failure of a power transformer is therefore defined as per the definition of failure found in [Taleb N.,
2016].

“Failure: the termination of the ability of a circuit, bay or item to perform a required function” Failures

are then separated into three levels which are defined as follows:

• Severe: The transformer requires replacement or removal from site to facilitate repair within a factory. In
both instances, a new transformer will be installed to return the circuit to service.

• Intermediate: The transformer requires repair, but this can be implemented on site. This is usually
intrusive work to restore the transformer to working condition and return it to service.

• Minor: these are trip events that remove the transformer from service temporarily. No work is required
in order to return to the plant to service, since the transformer’s major components have not been affected.
`14
 Failures of components that are critical to the operation of the transformer are also considered failures of the
transformer. For example, the failure of an HV bushing, or an OLTC will be considered failures of the
transformer since the transformer cannot operate in the absence of those components.

For the purpose of this study, only failures classified as severe or intermediate are considered since the minor
failures have no direct impact on the end-of-life of the transformer, or the decision to refurbish/replace/retire it.

Modes of failure

[Land Y, 2017].As outlined in Cigre working group 12.05 report [Bossi A, 203].
, international surveys are performed on failures of large power transformers and failure statistics are
reported with 10 year intervals. These statistics are based on standard modes of failure with the following
root causes: Core

Windings

Bushing

Tapchangers

Main tank and oil system Auxiliaries

Other

The breakdown of failures, by cause, experienced within Eskom Transmission during the period 1996 -
2014 is as shown in Figure 4. From this graph, it can be seen that the vast majority (56%) of all failures
experienced during this period is related to external components such as bushings and tapchangers.

Model parameters

The existing HI model for distribution transformers consists of a number of measurable condition
parameters. These are separated into four categories, each with different weightings. The high level
composition of the HI is shown in Tables 1.7, 1.8, 1.9 and 2.

Table 1.7 provides the parameters related to the most influential component of the distribution transformer
impacting the end-of-life and expected remaining life. These parameters provide information about both
the solid and liquid insulation.

[Bossi A
Electra,
No 88,
Figure 4: Breakdown of failures in Eskom Transmission as per Cigre reporting structure
1983].

Insulation Oil test

Furanic analysis DP

Oil quality Moisture

IFT

`15
 Dielectric Strength

Acidity

Colour/appearance

Tan

Sludge Table: 1.7 Overview of distribution

transformer HI insulation parameters.

Table 1.8., provides the electrical tests used to identify potential faults in various components of the active part of
the transformer. Faults in any of these components can lead to catastrophic failure of the transformer.

Table: 1.8. Overview of distribution transformer HI electrical test parameters

Electrical test Sub – test

SFRA N/A

DC resistance N/A

tan and winding capacitance N/A

Core insulation resistance N/A

Infrared scanning N/A

Magnetizing current N/A

Table 1.9. outlines the various visual inspections, checks and tests that are performed during routine
inspections and maintenance activities. The purpose of these checks is to identify deterioration that has the
potential of developing into a transformer fault. Any defects identified during these inspections can be
corrected with maintenance activities.

Table: 1.9. Overview of distribution transformer HI visual inspection parameters

Component Visual inspection

`16
 Bushings tan and capacitance

Visual inspection

Tapchanger Oil test

Speed test

Contact thickness

Transition resistance

Number of operations

Visual inspections Oil leak

Conservator condition

Cooling system condition

Tank and overall physical condition

Table 2. shows the DGA results that are used in condition assessment evaluations.

Parameter Gas

Concentration H2 (Hydrogen)

CH4 (Methane)

C2H6 (Ethane)

C2H4 (Ethylene)

C2H2 (Acetylene)

CO (Carbon Monoxide)

CO2 (Carbon Dioxide)

Daily rate of production H2

CH4

C2H6

C2H4

C2H2

CO

CO2

The combustible gas concent rations as well as the rates of production are of interest and are included. Of the
four HI cate gories, only two are used in this study, namely: Insulation and DGA. The reason is that this study is
concerned with the determination of probability of failure relative to end-of-life of the transformer and not all of
these components are related to end-of-life determination.
`17
The primary purpose of the electrical tests is to diagnose faults once the transformer has been removed from
service. A deviation in the test results is indicative of an immediate threat to the transformer and would be
repaired as necessary and retested, prior to re-energization. After such repair, the test results would again be
satisfactory. For this reason, the test results are not considered useful in terms of condition monitoring or
indicative of long term plant health deterioration. The requirement/reason for repair is more useful for
identification of potential faults and investigation of root cause of failure.

Table: 2. Overview of distribution transformer HI DGA parameters

The visual inspections are related to maintenance activities and identified deficiencies can be addressed
relatively easily. Again, should any fault be identified within the tapchanger or bushings, these will be
repaired/replaced. Although these components have a large impact on the failure of power transformers, the
routine condition monitoring/checks are not useful from a modeling perspective since they have binary
condition values rather than the continuous values which are indicative of slow deterioration.
Also the following were carried out;
Lab testing was key factor .simulation software available were used to simulate some findings or results and
secondary data of burnt transformers were considered.

Data types

Data can be classified into four main types as proposed by Stevens [Afifi . A; Clark .V; and May S;(2016)
“Computer-Aided Multivariate Analysis, 4th Edition”. Chapman and Hall/CRC.], namely:

nominal, ordinal, interval and ratio. These are explained in Table 3.4 below.

Table 3.4: Steven's measurement system

Type Explanation Example

There are a number of distinct categories Names

Nominal
that the variable can be classified into. Religion

There are a number of distinct categories Service ranking

Ordinal that the variable can be classified into, and the Factory ratings
categories have a known order.

Interval This is an ordinal variable with an equal Temperature

(discrete distance between successive values. Calendar dates

/continuous)

Interval variables with fixed zero Height

measurement points, hence preserving Difference in time

Ratio

ratios independent of the unit of

measurement.

`18
The data used in this study are comprised of three of these data types. These include: nominal, interval and
ratio. Each data type is handled differently since different information is available from each variable
Missing data

When performing statistical analyses, the issue of missing data is always a concern. Due to the fact that this
study is based on historical data (18 years old), it is apparent that some of the data required in this study will
be missing. This may be due to a number of factors including: operating practices changing, carelessness of
data storage and inability to recover certain portions of transformer records, etc.

There are three categories of missing data [Rubin D.B. (1978). “Multiple imputations in sample surveys a
phenomenological bayesian approach to nonresponse”. In The Proceedings of the Survey Research Methods
Section of American Statistical Association, pages 20–34, ]: Missing Completely At Random (MCAR)
where the data are missing independently of the DV or IV, Missing At Random (MAR) where the data are
missing dependent on one or more IVs but independently of the DV, and non-ignorable where the missing
data is dependent on both the IVs and DV. There are a number of ways of handling the missing data. These
are outlined below.

Listwise/Casewise deletion

This is the simplest method of dealing with missing data. Any record that contains a missing variable is
deleted from the sample. This can cause substantial reduction in sample size and lead to large biases [Carter.
R.L. (2016) “Solutions for missing data in structural equation modeling”. Research and Practice in Assessment,
1( Oommen T.V., (2015).

Pairwise data deletion

With this method, data records with missing data variables are used in the analysis only when the analysis does
not involve the missing variable. Again this method can produce large biases and unequal sample sizes [Carter.
R.L. (2016) “Solutions for missing data in structural equation modeling”. Research and Practice in Assessment,
1([Carter. R.L. (2016) “Solutions for missing data in structural equation modeling”. Research and Practice in
Assessment, 1( Oommen T.V., (2015).

Mean substitution

Mean substitution involves substituting each missing variable with the mean of all corresponding
variables within the entire data set. This is problematic since it reduces the variance of the variables
substituted in this way, which can lead to underestimating the spread of the data [Burke. S(2011). “Missing
values, outliers, robust statistics and non-parametric methods”. Statistics and Data Analysis, 1:19–24].

Hot deck imputation

In this instance, the record that is most similar to that with a missing variable is found and the value of the
variable in this record that corresponds to the missing variable is substituted for the missing value [Allison.
P.D.(2012) “Missing Data”. Sage Publications]. The difficulty that arises is in defining similarity, since this
is contextual and is therefore not a simple task [Allison. P.D. (2012) “Missing Data”. Sage Publications].
This method also does not account for uncertainty in the approximation.

Regression methods

variables, i.e. the missing variable becomes the response variable and the other variables become the
predictor variables [Aiken .L and West. S.(2013) “Multiple regression: Testing and Interpreting
Interactions”. Sage Publications.]. This leads to a complete data set with a reduced standard error

Expectation maximization and Raw maximum likelihood

These methods can be used to handle data that is MAR. In these methods estimates are found of the most
likely value that the missing variable might have. A vector of means and a covariance matrix are developed
that are superior to those that are developed from the previous methods of approximating missing data that
are mentioned above [Rubin. D.B.(2013) “Multiple imputations in sample surveys a phenomenological
bayesian approach to nonresponse”. In The Proceedings of the Survey Research Methods Section of
American Statistical Association, pages 20–34], [Shafri .H; Suhaili. A; and Mansor. S (2015) “The
performance of maximum likelihood, spectral angle mapper, neural network and decision tree classifiers in
hyperspectral image analysis”. Journal of Computer Science, 3(6):419– 423.]. The disadvantage of these
methods is that large sample sets are required.

`19
Multiple imputation

Multiple imputation has a number of advantages over other methods of missing data approximation. This
method involves having more than one estimate for a given missing variable, computed using other values
within the data set. In this way, the variance between the estimates gives information about the uncertainty of
the imputation. In this way, biases in the data are also reduced. Since there are a number of estimates for
each missing variable, after multiple imputation, there are a number of complete data sets instead of just one.
Each data set is analyzed individually and the results are then compressed to form only one final solution [
Rubin. D.B.(2013) “Multiple imputations in sample surveys a phenomenological bayesian approach to
nonresponse”. In The Proceedings of the Survey Research Methods Section of American Statistical
Association, pages 20–34]. This method is also disadvantageous since large data sets are required.

Since the data set available for this study exhibits non-ignorable missing data and is of limited size, the
preferred methods of handling the missing data are not practical. For this reason, the listwise deletion method
was utilized. This leads to significant reduction in sample size and the data needs to be evaluated for biasing.

Data visualization

Visualization of the raw data set is a simple method for determining the integrity of the data. By plotting
the raw data, it is easy to visualize any discrepancies within the data, for example, if you have a dichotomous
variable, but have values lying at points between 0 and 1, then it is obvious that the data set contains bad
data. It is prudent to perform this quick check, prior to any data processing or analysis to confirm that the
data is in fact as expected.

Gas concentrations and production rates

The concentrations of the different gases was plotted against the failures for the data set to be used in this
study. Since the data was collected for only failed and healthy transformers, the failed variable can only have
a value of either \0" (indicating a healthy transformer) or \1" (indicating a failed transformer)

The data is examined to determine conformity to an expected pattern. In this case, the data is expected to
form two distinct groups, one at low concentrations of gas in healthy transformers and another at high
concentrations in failed units.

`20
 Figure .4.5: Plot of gas concentrations vs failures From the results shown in Figure:4.5, it can be seen
that, with the exception of Carbon Monoxide, the expected trend in gas concentrations is evident, with higher
gas concentrations present in the failed transformers and lower concentrations in the healthy transformers.
The opposite trend is present in the CO data, which is not expected.

The daily rates of production of the different gases was plotted against the failures for the data set to be used
in this study. From the results shown in Figure 4.6 it can be seen that, with the exceptions of Carbon
Monoxide and Ethylene, the expected production rates are evident, with higher production rates in the failed
transformers and lower production rates in the healthy ones.

`21
Similarly to the gas concentration data, the daily rate of production data is examined to determine conformity
to an expected pattern. In this case, the data is expected to form two distinct groups, one at low rates of gas
production in healthy transformers and another at high rates of gas production in failed units.

The trends of CO and C2H4 show higher rates of decrease in concentrations for failed transformers. This
decrease in production rate of CO is congruent with the concentrations of the gas in failed transformers.

DP and oil quality

The values of the DP and oil quality tests were plotted against the failures for the data set to be used in this
study.

The results shown in Figure 4.6 indicate that there are higher moisture and lower dielectric strength
measurements in the sample of failed transformers as expected.

The measurements of acidity and DP indicate values that are fairly evenly distributed across their ranges for
failed and healthy transformers alike. Lower DP values and higher acidity values in failed units would have
been expected.

Figure 4.6: Plot of gas rates of production vs failures

`22
Figure 4.7: Plot of insulation parameters vs failures

Health Index

The values of the DGA and Insulation HI scores were plotted against the failures for the data set to be used

`23
in this study.

The results shown in Figure 4.8, indicate the expected trend in scores for failed transformers is lower than for
healthy ones in both insulation and oil quality.

Figure 4.8: Plot of HI scores vs failures

Transformation of data

Data are usually normalized according to the min-max minimization algorithm into a range [Eskom and
ABB Powertech (2018). “Theory, Design, Maintenance and Life Management of Power Transformers,
Power Series Volume 5. Y-Land”.] according to equation. Equation 5.1 below

x xmin
x
norm = (5.1)

x x
max min

where:

xnorm is the normalized value in range [0 1] x

is the value to be normalized

xmax is the maximum possible value for a particular observation x x min is the minimum possible value
for a particular observation x

Various other transformation techniques can be employed to normalize/linearize the data as required by the
specific statistical analysis being performed. Normalization is however not a requirement for logistic
regression and such data processing is therefore not required for this data set. The logit function used in the
Logistic Regression linearizes the DV.

Outliers

Box plots are then used to determine the skewness of the data, and identify possible statistical outliers [
Schwertman .N; Owens M; and Adnan R (2014) “A simple more general boxplot method for identifying
outliers”. Computational Statistics and Data Analysis, 47:165–174]. All identified outliers are included in the
sample analyses, unless they have been confirmed to be erroneous values.

`24
This method uses examination of the statistical percentiles of the data set. The statistics of interest in this
analysis are: maximum, minimum, median, mean, 1 st quartile (25th percentile) and 3rd quartile (75th
percentile). These statistics are then plotted as follows:

1st and 3rd quadrant form the top and bottom of each box for each variable. 50% of the data is represented
within this box, with the length of the box being the interquartile range.

The median is represented by the horizontal line within the box. A line that is not perfectly centered is an
indication of skewness of the data which is a measure of asymmetry about the sample mean. Skewness in the
data is an indication that the data are not normally distributed. This is not of concern in this study since
normality is not a requirement of logistic regression.

The maximum and minimum of the sample are represented by lines (sometimes referred to as \whiskers")
extending from the top and the bottom of the box.

A general assumption is that an outlier is a value that falls more than 1.5 times the interquartile range away
from either the top or the bottom of the box [ Schwertman .N; Owens. M; and Adnan.R.(2014). “A simple
more general boxplot method for identifying outliers”.Computational Statistics and Data Analysis, 47:165–
174.]. Therefore, if no outliers are present, the maximum value within the sample falls on the top of the
upper \whisker" and the minimum falls on the bottom of the lower \whisker". Potential outliers are indicated
by the presence of data points either above or below the upper and lower \whiskers" respectively.

While outliers may be identified statistically, the data samples need to be evaluated critically since the
\outliers" may in fact just be extreme values that have a critical impact on the analysis and should not be
removed without just cause.

`25
Gas concentrations and production rates

H2 CH4 CO CO2 C2H4 C2H6 C2H2

Min 0.1 0.1 6 0.1 0.1 0.1 0.1

1st Quad 2 2 215.5 244 1 1 0.1

Figure 4.9: DGA concentrations(left), including potential outliers(right)

Both Carbon Monoxide and Carbon Dioxide ,have higher concentrations and daily rates of production than
thehydrocarbon gases. For this reason, the values of all gases were scaled to put values for the purpose of
graphical comparison. The boxplots in Figure 4.9, show the gas concentrations of the DGA data both with
and without identified outliers.

These plots indicate a number of potential outliers in the data by the points above the top whisker. Table 3.5,
shows the actual values of the percentile statistics for each gas. From these values, it can be seen that none of
the gases have excessively high maximum values and in fact the maximums of each gas are still considered
low. Generally these values would be discarded as outliers, however, the outliers that have been identified
result from failed transformers' data and can therefore not be excluded. In this case, valuable information
would be lost if these outliers were excluded from analysis. No obvious, real outliers are evident.

The gas production rates were also scaled to per unit values for graphical comparison. The boxplots in Figure
5, show the gas concentrations of the DGA data both with and without identified outliers.

These plots indicate the presence of outliers in both the upper and lower regions. On examination of the actual
values of the

`26
 Median 9 6 409 1041 2 2 0.1

Mean 11.78 11.79 611 1384.6 6.2 8.3 2.7

3rd Quad 14 14.5 783.5 2040.5 5.5 6.5 1

Max 119 100 3912 11200 80 188 44

Table 3.5: Percentile summary of DGA concentration percentile statistics shown .

Figure 5: DGA rate of production(left), including potential outliers(right)

in Table 3.5, none of the values are unrealistic and consequently cannot be eliminated as outliers.

Some uncertainty in interpretation of the production rate values is present due to the method of calculation
employed. Manual oil samples do not lend themselves to reliable production rate calculations.

`27
DP and oil quality.

H2 CH4 CO CO2 C2H4 C2H6 C2H2

Min -0.394 -0.583 -9.298 -137.4 -1.101 -0.432 -0.703

1st Quad -0.018 -0.009 -0.271 -0.520 -0.006 -0.006 0

Median 0 0 0.135 0.125 0 0 0

Mean 0.019 0.002 0.266 2.255 -0.014 0.001 0.004

3rd Quad 0.021 0.007 0.742 1.708 0.006 0.004 0

Max 2.212 0.892 36.838 273.9 0.199 1.174 0.788

The measurements recorded for DP and oil quality were reduced to per unit values for the purpose of
graphical comparison. The boxplots indicating distribution of the DP and oil quality data both with and
without possible outliers are shown in Figure 5.1.

Table: 3.6. Percentile summary of DGA production rates

`28
 Figure 5.1: Oil quality and DP(left), including potential outliers(right)

These plots indicate possible outliers in the positive extreme for moisture, acidity and DP and in the negative
extreme for dielectric strength.

Table 3.7: Percentile summary of Oil Quality and DP measurements

DP Acid kV H 2O

Min 121 0.01 25 1

1st Quad 440 0.02 67 7

Median 550 0.04 74 9

Mean 578 0.05 72 11

3rd Quad 676 0.06 78 14

Max 1300 0.25 95 72

On examination of the actual values of the statistical percentile analysis shown in Table 3.7, it can be seen
that the range of data values is reasonable, with the exception of the maximum value for DP.

This value appears to be unrealistically high since the expected value of DP for brand new paper is 1200 and
once it has been processed during the manufacture of a transformer it has reduced to approximately 900. This
value is therefore considered an outlier and removed from the dataset.

Health Index

The values of HI score are uniform for all components, in the range (0-5). For this reason, it was not
necessary to reduce the values to per unit equivalents. The boxplots of the DGA and Insulation HI scores,
both with and without potential outliers are shown in Figure 5.2. These plots indicate a number of potential
outliers.

`29
Figure 5.2: HI parameters(left), including potential outliers(right)

Table 3.8. Percentile summary of HI parameters.

DGA Insulation

Min 0.9 0

1st Quad 0.9 2

Median 1 3

Mean 1.1 2.6

3rd Quad 1.1 3

Max 2.5 4

Examination of the statistical percentiles in Table 3.8, shows values that are reasonable and cannot be
eliminated as outliers. This is congruent with the decision not to remove outliers in the base data used to
calculate the HI scores.

Conclusion

The data set to be used in this study was defined. The data was evaluated visually and found to display the
expected trend. The data set was found to contain records with missing data and these records were then
deleted, since the data set is too small to utilize methods of missing data approximation. The data was
examined graphically, using boxplots to identify potential outliers. In this case, only the outliers identified in
the DP measurements were removed since there was no evidence supporting the removal of the outliers
identified in the other variables. The data set processed, using the methods outlined in this chapter, will be
used in the development of the statistical model for probability of failure determination.

`30
REFERENCES

1. Ranjana and Amarjit Singh, 2017

2. Kumar at el 2017
3.Yin (2013),
3. Eduful & Mensah, 2016
4. Cohen, et al., 2012
5. Berg, Fritze, 2017
6. Frazer and Lawley (2014)
7. Bryman, 2004

8. [Bartley, W.H. (2003). “Analysis of transformer failures, imia wgp33(03). In 36th Annual Conference of the
International Association of Engineer’s in Insurers, Stokholm,(2003)”].

`31