Electric Power Systems Research 119 (2015) 471–477

Contents lists available at ScienceDirect

Electric Power Systems Research

journal homepage: www.elsevier.com/locate/epsr

Transformer risk modelling by stochastic augmentation of

reliability-centred maintenance
Burnet O’Brien Mkandawire a,∗ , Nelson Ijumba b , Akshay Saha a
a
University of KwaZulu-Natal, School of Electrical Engineering, Howard College, Durban 4041, South Africa
b
University of Rwanda, Box 4285, Kigali, Rwanda

a r t i c l e i n f o a b s t r a c t

Article history: This paper reviews the literature and then applies power transformer case studies to develop a risk
Received 8 August 2014 assessment model that augments the probabilistic capabilities of Reliability-centred maintenance (RCM).
Received in revised form 24 October 2014 It is usually very difficult to assess the effectiveness of the RCM at its inception, when data is inadequate or
Accepted 1 November 2014
when it is implemented on a large population of assets, especially when most of them are small in size as
Available online 22 November 2014
in the distribution systems. Thus, the major contribution of the study rests on applying a key performance
indicator (KPI) for assessing the effectiveness of the RCM programmes, obtained by trending the profile
Keywords:
of the mean-time-to-first-failure (MTTFF) and the average annual repair costs. The MTTFF, determined
Markov processes
Power asset management
using Markov analysis, is inversely proportional to the costs. Besides, the method of moments is applied to
Reliability-centred maintenance statistical, historical data to generate a failure probability distribution comparative model, which lacks
Risk analysis in the current practices for conducting a failure mode effect and criticality analysis (FMECA). Finally,
Transformer the Markov derives complement of uptime-steady-state probability as input for FMECA, using limited
Reliability modelling data; which is an improvement on the current approaches. The overall approach developed offers a cost
effective risk-priority-screening model for transformers, which can be applied prior to rigorous testing
and inspection procedures on individual items during the RCM application.
© 2014 Elsevier B.V. All rights reserved.

1. Introduction characterisation, data requirements and reliability modelling.

Then, Section 3 outlines the methodology. Thereafter, Section 4
Many electric power utilities install Reliability-centred mainte- presents and discusses the results. Finally, Section 5 provides the
nance (RCM) programs for establishing maintenance requirements conclusions.
[1,2]. Much data and long (vast) experience is needed to success-
fully conduct the RCM [3]. Despite being around for a long time, its 2. Critical review of the RCM
application in the power sector is still at an infancy stage [4].
Significant changes in business environment make electric 2.1. Merits, demerits and opportunities
power utility decision making and risk management, especially the
need for improved risk characterisation, increasingly challenging The RCM, initially applied to the aviation industry, establishes
[5]. Every stage of the physical asset management process should maintenance and refurbishment needs in complex, critical assets
incorporate and build on the best attributes of the RCM for the [2,7] based on constant failure and repair rates [8]. Critical assets are
process to succeed [6,7]. those for which the financial and service level impacts of failure jus-
Section 2 provides a critical examination of the RCM and tify proactive assessment and restoration [1]. The main value of the
challenges faced during its implementation with respect to risk RCM is in answering the seven questions about the functions, func-
tional failures and their causes as well as remedial actions needed
to determine the most appropriate maintenance strategies [2,9].
Abbreviations: FMECA, failure mode effect and criticality analysis; FMEA, failure The RCM has been described as a structured methodology and
mode effect analysis; ICT, information and computer technology; MOM, method maintenance organisation or process for establishing the most cost
of moments; MTTR, mean time to repair; MTTFF, mean time to first failure; RCM,
effective level of reliability [9–12]. The RCM is claimed to reduce
reliability-centred maintenance; RPN, risk priority number.
∗ Corresponding author. Tel.: +27 603297130. 11 kV transformer maintenance costs by 30–40% and routine pre-
E-mail addresses: [email protected] (B.O. Mkandawire), ventive maintenance costs by up to 50% [13]. The US navy reported
[email protected] (N. Ijumba), [email protected] (A. Saha). the following tangible benefits of the RCM: life cycle cost reductions

http://dx.doi.org/10.1016/j.epsr.2014.11.005
0378-7796/© 2014 Elsevier B.V. All rights reserved.
472 B.O. Mkandawire et al. / Electric Power Systems Research 119 (2015) 471–477

of 15%, representing $1.7 billion; increased availability by 17%; and protocols to incorporate CM transducers [6], and also developed a
an 8-year ship life extension [6]. standard for the FMECA and fault root cause analysis [22].
The major flaws of the RCM are as follows: it lacks prioritisation
needed for general industrial application [6], it is costly to imple- 2.4. Reliability modelling
ment and requires components of Total Productive Maintenance
(TPM) to sustain its full capability [14]. Furthermore, it lacks the Models that quantify effects of maintenance on reliability in the
flexibility and full merits of probabilistic models [15]. Finally, it is power systems are few [26]. The RCM has been applied to critical
unable to quantify the benefits of maintenance on system reliability systems and equipment [6], however, it is heuristic and requires
and costs [16,17]. judgement and experience which can take long before enough
Risk Based Inspection (RBI) may be incorporated in the RCM to data is collected for the decision making purposes [3]. The Markov
fortify it [1,6]; thereby helping to select appropriate condition mon- analysis can use a few data sets to cost effectively derive mean-
itoring methods [12]. However, the RBI is neither able to quantify time-to-first-failure (MTTFF), model reliability and measure the
costs of inspection/condition monitoring nor indicate the alterna- effectiveness of strategies, based on component failure and repair
tive risk treatment options [12]. The RCM’s fault root cause analysis rates [27,29]. These rates are assumed constant [26,27]; a notion
usually comprises a Failure mode effect analysis (FMEA) which is viewed as flawed [30], but it simplifies the analysis [31]. Section
used to analyse potential failure modes and their impacts; and the 3 presents an analytical model that exploits the opportunities and
Failure Mode Effect and Criticality Analysis (FMECA) which extends addresses the challenges that have been outlined in Section 2.
the FMEA to include measures to rectify the faults [18]. Major con-
cerns in the FMECA are the tendency to eliminate cascade failures 3. Risk modelling methodology
[19], and the use of simple uptime/downtime data to compute risk
indexes which can affect the validity of the results [20]. For power This section proposes a multi-method approach involving sta-
transformers, a risk index is given as the product of consequence tistical data analysis and simulation of the MTTFF using stochastic
factor and failure probability [21,22]. Sections 2.2, 2.3 and 2.4 crit- Markov processes, as a way of determining probability of failures
ically examine challenges pertaining to the risk characterisation, with reduced level of physical equipment inspection and testing;
data requirements and reliability modelling, respectively. and even where limited data is available. The probabilities deter-
mined can be used to compute risk indexes within the FMECA stage
2.2. Risk characterization of the RCM.

Physical asset management centres on optimisation of risks, cost 3.1. Comparison of methods and overall approach
and reliability [23]. The risk management process consists of the fol-
lowing seven stages [1]: risk contextualisation, risk identification, There are four main methods for reliability and or statistical data
risk exploration, risk assessment, risk treatment, risk monitoring analysis, namely: least squares method (LSM), maximum likelihood
and review, and risk reporting. Risk characterisation refers to a syn- estimation (MLE), method of moments (MOM) [30,32], and Markov
thesis of the seven stages, to provide a conclusion about the risk, the process [27,29]. The LSM gives accurate results for large sample
nature of the inherent and residual risk; and a rethink in strategy sizes and for non-censored data, whereas the MLE and MOM yield
or policy due to changes in the risk profile over time. accurate results for all sample sizes and even when the data is cen-
The success of the risk characterisation process requires a sored [32]. The results from the MLE and MOM are usually almost
comprehensive database of faults, failures, operations, and main- the same although the MLE is preferred to the MOM because its ana-
tenance as in a surveyed breakdown strategy; whereby a fault and lytical approach is simple. In electric power systems, data is likely
damage database is combined with Supervisory Control and Data to be censored; hence either the MLE or the MOM can be applied.
Acquisition (SCADA), Enterprise Resource Planning (ERP) software However, it has been shown that the application of the MOM yields
and Geographical Information System (GIS) [24]. The cost impli- more satisfactory standard errors than the MLE, thus the MOM can
cations and integration challenges tend to limit the application of be used to validate the results of the MLE [33]. Unlike the LSM, MLE
such strategies in the power sector. and MOM, the Markov technique is not suitable for drawing statis-
tical inference, but its advantage over the other methods is that it
2.3. Data requirements can resolve state probabilities and simulate the MTTFF, using data
from a very short time span (horizon) [27]. Based on the above
The RCM is one of the proactive equipment management prac- background, this study applies the MOM to process statistical data
tices, with probabilistic inferences, that have characterised the to draw the relevant statistical inferences; and the Markov analy-
current asset management paradigms [7,24]. Data requirements sis to deal with the computation of transitional probabilities and
for probabilistic concepts are huge, and it is usually difficult to MTTFF.
get the data [3,24]. The validity of the risk evaluation processes
by line managers, who normally conduct the risk analysis [25], can 3.2. Overall analytical approach
be adversely affected by the data unavailability.
Parameters for probabilistic models include mean times to Fig. 1 outlines the overall methodology, whereby the risk ana-
failure, inspection rates and probabilities of state transitions lyst must decide whether the equipment has a long history, with
[26,27]. Information and computer technology (ICT) models are a reasonable quantity of data, or not. If yes, the MOM should be
useful for capturing these parameters, but power utilities have applied, otherwise the Markov simulation models should be used.
applied the models inconsistently, in a fragmented way and Fig. 1 (dotted) also shows that existing models on spare part
face challenges in integrating them in the data mining process contingency plans should be applied and the optimum number of
[28]. Open System Architecture Condition-Based Maintenance spares determined to reduce the mean time to repair (MTTR) [22].
(OSA-CBM) and Machinery Information Management Open Sys- A new way of doing this was presented by [34]; it uses the Poisson
tems Alliance (MIMOSA) initiated the integration of standard ICT distribution to satisfy the minimum requirements on mean time
protocols in condition monitoring (CM) and maintenance, but between failure, reliability and statistical economics. The statistical
most organisations have not embraced their use [28]. The Institute economics model is the best as it reduces the total system costs
of Electrical and Electronics Engineers (IEEE) standardised the as well as the cost of spares carried in the system [34]. Besides,
 B.O. Mkandawire et al. / Electric Power Systems Research 119 (2015) 471–477 473

Start

No Is it long term risk

planning?
Develop state Determine and report risk profile
space diagram Yes
Reduce MTTR Input failure
statistics
Determine transitional
probability matrices Empirical deductions Evaluate RPN &
on cost and risk benchmark
Compute life
Develop algorithm for modelling
Compute transitional Use probabilities in parameters with
transitional probabilities FMECA
probabilities /MTTFF MOM
and MTTFF
Fit another
Input transitional Stack data in a Compute reliability parametric
rates special vector functions/probabilities model

Yes Are parameters from the

purported Weibull? No
Apply existing models: Optimise Model aging
Contingency plans for Spare parts failures
timely spare installation

Fig. 1. Overall risk modelling methodology. MTTFF: mean time to first failure; MOM: method of moments; FMECA: failure mode effect and criticality analysis; RPN: risk
priority number; MTTR: mean time to repair.

Table 1 terms of the coefficient of variation (cv) and the estimate of  (i.e.
Times-to-failure [×105 h] for 12 MVA transformers.
),
ˆ respectively, as [33]:
1.892 1.971 1.971 2.182 2.31 2.418
2.365 3.715 3.925 4.03 4.188 4.366 1/2
[ (1 + (2/ˇ)) −  2 (1 + (1/ˇ))]
cv = (4)
 2 (1 + ((1/ˇ)))

aging failure probabilistic test models should be incorporated to  ˇˆ

x̄
accurately evaluate the system reliability [35]. ˆ=
 (5)
 (1/ˇˆ + 1)

3.2.1. Method of moments (MOM) where x̄ is the mean, ˇˆ is the estimate of ˇ, derived from Eq. (4)
The MOM is applied to time-to-failure data for transformers using the tabulation method or software packages. Eq. (4) and (5)
as outlined in Table 1. In order to apply the MOM, data must be are applied to the data in Table 1.
applied to a known distribution [36]. Hence, a statistical hypoth- Section 3.2.2 shows how the Markov approach is used to com-
esis test may be required to prove that the chosen distribution is pute the transitional probabilities and the MTTFF.
acceptable. This has been conducted using Kolmogorov–Smirnov
(K–S) test implemented by the R-Statistical software package. In
this study, the Weibull distribution is used, because it offers the 3.2.2. Markov and mean-time-to-first-failure
following advantages: flexibility, as it can be used to model many If a repairable system or component exists in either a failed or
types of distributions; and the relationship of the shape and scale a non-failed state, probabilities associated with these states can
parameters to the bath-tub curve [32]. The Weibull probability den- be defined using a Markov process based on the repair and failure
sity function (PDF), cumulative density function (CDF) and hazard rates [27]. These rates can be used to produce stochastic transitional
rate h(x) are, respectively, expressed as [30,32]: probability matrices [27,29]. These are needed to determine the
optimum maintenance policy or strategy for physical assets [37].
 x ˇ−1   ˇ 
ˇ x Fig. 2 shows a case of two independent, non-identical compo-
PDF = f (x) = exp − (1) nents (T1 and T2) with four states, which is chosen for modelling
  
and simulating the transient probabilities and the MTTFF, assuming
perfect upstream reliability and a fuse blowing philosophy applies
x   ˇ 
x to the system. In the figure, 1 and 2 are failure rates 1 and 2,
CDF = f (x)dx = 1 − exp − (2) respectively; whereas 1 and 2 are repair rates 1 and 2, respec-

0 tively. Numbers 1–4 in Fig. 2(b) represent States 1 to 4 (i.e. P1 –P4 ),
respectively.
f (x) ˇ
 x ˇ−1 The general Markov process assumes conditional probabilities
h (x) = = (3) of events. If t = time and x = incremental time, transition from i to j
R(x)  
can be generalised by [27,29]:
where ˇ = shape parameter,  = scale parameter (i.e. when 63% will 
have failed), x = random variable, and R(x) = reliability function. The pi,j (t + x) = pik (t)pkj (x) (6)
h(x) represents the probability that an item that has survived up to
time, x, will fail after that time. 
pij (x) ≤ 1, for all i and j. (7)
The moment generating function can be applied to Eq. (1) to
derive expressions for ˇ and  [36]. These expressions are given in j
474 B.O. Mkandawire et al. / Electric Power Systems Research 119 (2015) 471–477

Fig. 2. Two-component-four-state model. T1 and T2: transformers; 1 and 2 : failure rates; 1 and 2 : repair rates; LP1 and LP2: load points.

The stochastic transitional probability (state-space) matrix for the

four states, that is, Fig. 2(b), is determined from the rates of arrival
and departure by the following matrix:
⎡ ⎤
1 − 1  t − 2  t 1  t 2  t 0
⎢ ⎥
⎢ 1  t 1 − 2  t − 1  t 0 2  t ⎥
P=⎢
⎢
⎥
⎥ (8)
⎣ 2  t 0 1 − 1  t − 2  t 1  t ⎦
0 2  t 1  t 1 − 1  t − 2  t

The probability of being in a given state after n time steps is

expressed as:
Expected failure cost per year can be presented in terms of the
lim P (n) = ˛
P (n) = P(0)P n and (9) MTTFF by [26] as:
n→∞
FC
˛P = ˛ (10) CF = (16)
TR + MTTFF
where P(0) is the initial state vector and ˛ is the limiting probability
where CF is the expected failure cost per year, FC is the failure cost,
vector = [ P1 P2 P3 P4 ].
and TR is the repair time.
By applying Eq. (8) according to (10), with t implicit, time and
frequency domains (treating the problem as an initial value prob-
lem) can be computed and four limiting state probabilities (P1 –P4 ) 4. Results and discussion
for each state arise as follows:
1 2 1 2 4.1. Parameter estimates and reliability modelling
P1 = , P2 = ,
(1 + 1 )(2 + 2 ) (1 + 1 )(2 + 2 )
Table 2 presents the parameters that were estimated by the
1 2 1 2 MOM and accepted by the Kolmogorov–Smirnov (K–S) test that
P3 = , and P4 =
(1 + 1 )(2 + 2 ) (1 + 1 )(2 + 2 ) was implemented by the R-Statistical software (see, e.g., Section
(11) 3.2.1). MATLAB algorithms generated the CDF, the hazard rate and
the PDF from Eqs. (1)–(3) as shown in Fig. 3.
If Sk is a transient set of states with matrix G obtained by truncating
The CDF is a failure distribution function; hence it can be used
P, the mean number of times the process is in Sj before absorption,
to represent the failure probability [30]. Thus, Fig. 3(a) can be used
if it started in Si , gives the MTTFF as [29]:
 as a model from which the failure probabilities to be used for risk
S(Nij ) = nij ≤ S ∈ Sk (12) modelling can be extracted; and for benchmarking against similar
assets currently in operation. Point A on Fig. 3 signifies the end of
N = (I − G)−1 (13) life, with probability (CDF) of 1 at the age of 65.84 years. Point D
on the PDF curve corresponds to the steepest slope on the CDF plot
where nij are the elements of N and I = identity matrix.
(point C) and it occurs at 34 years of age. In terms of risk mitiga-
Alternatively, using a mean first passage time (MFPT) matrix, if
tion, that is the right time to refurbish the transformers, i.e., before
MF is a fundamental matrix and T̄ be the MFPT matrix, then:
failure rate exceeds survival rate.
MF = [I − (P − A)]−1 (14)

T̄ = [I − MF + UMD ]D (15) Table 2

Estimates of parameters ˇ and  by the MOM.
where I = identity matrix, U = unit matrix, P = transition matrix,
A = matrix each row of which is the limiting probability vector, Weibull parameters ˇ  [×105 ]
˛i = (˛o , ˛1 , . . ., ˛n ); MD = a diagonal matrix of MF , T̄ is such that Estimated parameter values 3.4988 3.2786
t̄ij represents the MTTFF or mean number of steps from state i to j, Standard error (se) 0.6697 0.3025
and D = diagonal matrix such that dii = 1/˛i . Confidence interval (2.5%), (97.5%) (2.9048), (4.5105) (2.7049), (3.828)
 B.O. Mkandawire et al. / Electric Power Systems Research 119 (2015) 471–477 475

(a) CDF and hazard rate h(t) at β=3.4988,η=327860 hrs. (b) PDF for β=3.4988,η=327860 hrs.
1 0.04
CDF
A D
0.9 hazard rate
0.035

0.8
0.03
h(x) [Magnitude] and CDF [Probability]

0.7

0.025

Probability density
0.6

C
0.5 0.02

B
0.4
0.015

0.3
0.01
0.2

0.005
0.1

0 0
0 10 20 30 40 50 60 70 0 10 20 30 40 50 60 70
Technical life, x [yr.] Technical life, x[yr.]

Fig. 3. Comparative model: CDF, hazard rate and PDF. ˇ and : Weibull shape and scale parameter, respectively.

The probabilities in Fig. 3(a) can be used as inputs into the 4.2.2. Application of transient probability inferences
FMECA (Table 3b). The analysis shows that the value of  for the Fig. 5 presents plots of the transient probabilities. From the fig-
transformers is 327860 h or 37 years. ure, State 1 has the highest survival probability, represents the least
Table 3a shows the first stage of the FMECA, with current meas- level of risk and can be predicted with certainty; hence it is the
ures (business as usual); whereas Table 3b presents the second most desirable and the one which asset managers would want to
stage, that is, after proactive measures are taken to reduce the risk. sustain. From Fig. 5, the risk of system failure is the complement of
The computed failure probabilities are incorporated in Table 3b. For the steady-state value of the transient probability of the uptime-
example, from Fig. 3(a), at 30 years of age the probability is 0.387. state (i.e. 1 − P1 ), which is 29.6% (0.296) for case # 1 and 36.7%
This is entered in column F of Table 3b, and the risk priority number (0.367) for case # 2 (see, e.g., Table 4 for the cases). Similar results
(RPN) is given as the product of columns F, G and H (see column I). are obtained from the complement of P1 in Eq. (11) as well as from
By comparing the RPN under the current actions, for each failure the value of the complement of transitional probability where the
mode, with that after proactive actions are taken, the asset manager 1 − P1 in Fig. 5 approximates a constant. Therefore, these probabil-
can determine whether strategies implemented were effective in ities are the ones to be used as inputs into the FMECA [see, e.g.,
mitigating the risk or not. The heading of column E in Table 3b will Table 3b column F].
change to reflect the next type of strategy applied to improve the
RPN. Typical failure probabilities for different power system assets
at the age of 30 years are presented in [8]; those for transformers MTTFF- Cost profile
compare very well with the results in Fig. 3(a). Detection [Table 3b 550
B
H] shows the relative difficulty (in terms of the skill needed) in Case# 2
which a failure mode is detected on a scale of 1 to 10 [1]. Condi- 500
tion monitoring can improve the detection capabilities. Section 4.2
demonstrates how the Markov process analyses transitional prob- 450
abilities and the MTTFF. In the process, failure and repair rate data
Annual cost rate (US$)

is applied to the state-space model that was presented in Fig. 2(b). 400

350 A
Case# 1
4.2. Simulation of MTTFF and transient probabilities
300
4.2.1. Application of failure and repair data in MTTFF analysis
Eq. (8) is implemented according to (9) and (10), assuming 250
P(0) = [1000], using the values from Table 4; whereas the MTTFF
is computed from Eq. (15). From Eq. (16), the annual cost rates (at 200
FC = US$123,000 and TR = 60 h) are computed and plotted against
the simulated MTTFF results as shown in Fig. 4. 150
The profile in Fig. 4 shows an inverse proportional relation- 100 200 300 400 500 600 700
MTTFF
ship between the two parameters and can be used as a metric
(key performance indicator, KPI) for the effectiveness of the RCM Fig. 4. Profile of MTTFF vs. Costs as a metric (KPI) of RCM programs. MTTFF: mean-
implementation. time-to-first-failure.
476 B.O. Mkandawire et al. / Electric Power Systems Research 119 (2015) 471–477

Table 3a
FMECA Stage 1 [from system to failure mode].

A B C D
System Function Functional failure Failure mode

Transmission transformer To transmit power Failure to transmit power Decrease of mechanical, thermal and electrical strength

Table 3b
FMECA Stage 2 [from current measures to risk priority number (RPN)].

E F G H I
Current measures to rectify failure Probability Severity (1–10) Detection (1–10) RPN (F × G × H)

Planned maintenance 0.387 3.35 3 3.889

Table 4
Average repair and failure rates for 12 MVA transformers.

Case #1: 25 years old transformers Case #2: 30 years old transformers

Repair rate 1 599.999 repairs/year 1 = 0.0685 674.52 repairs/year 1 = 0.077

Repair rate 2 729.997 repairs/year 2 = 0.0833 762.12 repairs/year 2 = 0.087
Failure rate 1 149.998 failures/year 1 = 0.0171 192.72 failures/year 1 = 0.022
Failure rate 2 100.004 failures/year 2 = 0.0114 175.2 failures/year 2 = 0.02

Fig. 5. Transient probabilities for cases 1 and 2. 1 and 2 : failure rates; 1 and 2 : repair rates.

4.3. Comparison of effectiveness of the adopted approaches and 0.387, respectively. As expected, the probabilities from the two
approaches cannot be exactly the same because they are derived
The application of the Markov approach enables the use of from data from different time horizons, and for different operating
data from very short time periods, for example, one year. In con- conditions. So what is remarkable about the chosen approaches is
trast, the MOM uses data collected from longer periods of time the capability to generate failure probabilities that can be used in
as is the case with most statistical models. However, in compar- the FMECA and fault root cause analysis. This can strengthen or
ison with other models like the LSM (see, e.g., Section 3.1), the augment the current approaches which have based the computa-
MOM is more desirable as it can utilise very few data sets. In tion of these probabilities only on condition monitoring as alluded
addition, the MOM generates a failure probability distribution com- to by [22].
parative model which current approaches to FMECA/root cause
analysis don’t do. Neither is this possible with the Markov. How- 5. Conclusions
ever, the strength of the Markov lies in the generation of transient
probabilities and the MTTFF that can be used for measurement of In the study, the literature review highlighted major challenges
performance of the RCM strategies. The Markov can also model in the current approaches for conducting the RCM and risk assess-
derated states, which the MOM cannot do. Despite the differences ment. Next, case study data was applied to illustrate the application
between the two techniques, the failure probabilities computed of the risk assessment model developed. The originality of the study
compare very well. For example, the Markov yields probabilities is three-fold. First, the inverse proportional relationship between
of 0.296 and 0.367 for the 25-year old and 30-year old transform- the MTTFF and costs was trended (profiled) and applied as a KPI
ers, respectively; whereas the MOM generates probabilities of 0.23 of the effectiveness of the RCM strategies. Since average annual
 B.O. Mkandawire et al. / Electric Power Systems Research 119 (2015) 471–477 477

costs, and repair and failure rates are used to produce the KPI, [10] J. Endrenyi, Impact of Maintenance Strategy on Reliability of the Reliability,
the approach is a convenient and cost effective way of providing IEEE/PES Task Force Final Rep., Risk and Probability Sub-Committee, 1999,
July.
a broad view of the effectiveness of the strategies on a group of [11] J.K. August, RCM Guide Book: Building a Reliable Plant Maintenance Pro-
network assets. Second, the MOM was applied to statistical data to gramme, Penn Well, Tulsa, Oklahoma, 2004.
generate a failure probability distribution benchmarking model; [12] J. Woodhouse, Combining Best Bits of RCM, RBI, TPM, TQM, Six-Sigma and Other
Solutions, Woodhouse Partnership Ltd., Newbury, UK, 2001.
in addition to the current practices for conducting the FMECA, [13] A.K.S. Jardine, A.H.C. Tsang, Maintenance, Replacement and Reliability: Theory
where failure probabilities are based only on condition monitoring, and Applications, Taylor and Francis, Boca Raton, FL, 2006.
a notion that [22] describes as inadequate. Third, Markov tech- [14] R. Moore, D. Roth, Combining TPM and RCM, in: Proceedings of the Maintenance
and Reliability Conference, Knoxville, Tennessee, May, 1998.
niques generated failure probabilities for use in the FMECA, with
[15] M. Schwan, K.H. Weck, M. Roth, Assessing the impact of maintenance strategies
data from as short as a 1 year time frame, using the complement on supply reliability in asset management methods, in: Proceedings of the 40th
of the uptime-steady-state transitional probability values. The cur- CIGRE Session, Rep. C1-108, August, 2004.
[16] L. Betling, Tutorial on asset management and replacement strategies, in:
rent approaches have not employed the use of data from such short
Proceedings of the IEEE/PES General Meeting, Stockholm, Sweden, 2004.
time horizons, although it may be necessary to do so in some cases. [17] P. Hilber, Maintenance optimization for power distribution systems (Doctoral
This is especially useful in giving a strategic direction when data is thesis), Royal Institute of Technology, Stockholm, Sweden, 2008.
inadequate or when a large population of transformers or equip- [18] MIL-STD 1619A, Procedures for performing a FMECA, Military Standards, USA.
[19] D. White, Application of systems thinking to risk management: a review of the
ment is involved. The proposed method forms a priority-screening literature Manage. Decis. 33 (1995) 33–45.
phase so that any rigorous testing procedures, especially for an aged [20] B. Keeter, Using reliability engineering methods as a tool for continuous process
system (which has a significant impact on the system reliability) improvement, in: 18th International Maintenance Conference – Clearwater,
Florida, 2003.
[35], can still be carried out, but only on the critical assets; thereby [21] W.H. Bartley, Life cycle management of utility transformer assets, in: Break-
reducing operating costs or effort/time. through Asset Management for the Restructured Power Industry Conference,
10–11 October, Salt lake City, Utah, 2002.
[22] IEEE Guide for the Evaluation and Reconditioning of Liquid Immersed Power
Acknowledgements Transformers, IEEE Standard C57.140, 2006.
[23] Publicly Available Specification for Optimised Management of Physical Assets,
The authors are grateful to the University of KwaZulu-Natal’s British Standard PAS 55-1, 2008.
[24] J. Schneider, J. Gaul, C. Neumann, J. Hogräfer, W. Wellßow, M. Schwan, A.
High Voltage Direct Current Centre (HVDC) and the Centre for Post- Schnetter, Asset management techniques, Int. J. Electr. Power Energy Syst. 28
graduate Engineering Studies (CEPS) of South Africa (Grant No. (2006) 643–654.
HVDC/CEPS/2014/PHD/BOM/ND19), for supporting this study. [25] G.J. Anders, J. Endrenyi, C. Yung, Risk-based planner for asset management, IEEE
Comput. Appl. Power 14 (4) (2001) 20–26.
[26] P. Jirutitijaroen, C. Singh, The effect of transformer maintenance parameters
References on reliability and cost: a probabilistic model, Electr. Power Syst. Res. 72 (2004)
213–224.
[1] B.O. Mkandawire, N. Ijumba, H. Whitehead, Application of maintenance tools [27] R. Billinton, R.N. Allan, Reliability Evaluation of Engineering Systems: Concepts
and strategies in integrated risk management of critical physical assets, Int. J. and Techniques, Plenum Press, New York, 1983.
Agile Syst. Manage. 4 (2011) 261–279. [28] J. Campos, Development in the application of ICT in condition monitoring and
[2] J. Moubray, Reliability Centred Maintenance, 2nd ed., Butterworth-Heinemann, maintenance, J. Comput. Ind. 60 (2009) 1–20.
Oxford, 1997. [29] C. Singh, R. Billinton, System Reliability Modeling and Evaluation, Hutchinson
[3] J. Endrenyi, G.J. Anders, A.M. Leite da Silva, Probabilistic evaluation of the effect Educational, London, 1977.
of maintenance on reliability – an application, IEEE Trans. Power Syst. 13 (1998) [30] D.J. Smith, Maintainability and Risk: Practical Methods for Engineers, 8th ed.,
576–583. Butterworth-Heinemann, London, 2011.
[4] E.B. Abu-Elanien, M.M.A. Salama, Asset management techniques for transform- [31] S. Asgarpoor, M.J. Mathine, Reliability evaluation of distribution systems with
ers, Electr. Power Syst. Res. 80 (2010) 456–464. non-exponential down times, IEEE Trans. Power Syst. 12 (1997) 579–584.
[5] Electric Power Research Institute (EPRI), Third Power Delivery Asset Manage- [32] P.D.T. O’Connor, A. Kleyner, Practical Reliability Engineering, 5th ed., John Wil-
ment Conf. Proc.: Decision Support and Information Technology, EPRI, Palo Alto, ley, Chichester, UK, 2012.
CA, 2007, Rep. 1012497, March. [33] Y. Lei, Evaluation of three methods for estimating the Weibull distribu-
[6] J.S. Mitchell, in: T.H. Bond, N. Nodianos (Eds.), Physical Asset Management tion parameters of Chinese pine (Pinus tabulaeformis), J. For. Sci. 54 (2008)
Handbook, 3rd ed., Clarion Technical Publishers, Houston, Texas, 2002. 566–571.
[7] B.O. Mkandawire, N.M. Ijumba, H. Whitehead, Asset management optimisa- [34] A.A. Chowdhury, D.O. Koval, Development of probabilistic models for comput-
tion through integrated systems thinking and N-1 contingency capability for ing optimal distribution spare transformers, IEEE Trans. Ind. Appl. 41 (2005)
refurbishment, IEEE Syst. J. 5 (2011) 321–331. 1493–1498.
[8] X. Zhang, E. Gockenbach, Age-dependent maintenance strategies of medium- [35] W. Li, Incorporating aging failure in power system reliability evaluation, IEEE
voltage circuit breakers and transformers, Electr. Power Syst. Res. 81 (2011) Trans. Power Syst. 17 (2002) 918–923.
1709–1714. [36] G. James, D. Burley, D. Clements, P. Dyke, J. Seal, N. Steele, J. Wright, Advanced
[9] J.K. August, Reliability Centred Maintenance effectiveness in the power indus- Modern Engineering Mathematics, 3rd ed., Pearson, Harlow, UK, 2004.
try: how much is real and how much hype? in: Proceedings of the Joint Power [37] G.K. Chan, S. Asgarpoor, Optimum maintenance policy with Markov processes,
Generation Conference, Denver, Colorado, 1997, pp. 347–359. Electr. Power Syst. Res. 76 (2006) 452–456.