Maintenance Strategic and Capacity Planning

Chapter 2
Maintenance Strategic and Capacity

Planning
2.1 Introduction
Planning is one of the major and important functions for effective management. It
helps in achieving goals and objectives in the most efcient and effective manner.
Planning is usually divided into three levels, depending on the objective and the
planning horizon. The levels are as follows:
1. long-range planning (covers a period of ve or more years);
2. medium-range planning (covers 1 month to 3 years); and
3. short-range planning (daily, weekly, and monthly plans).
Maintenance strategic planning is a long-range planning and is concerned with the
determination of maintenance mission, strategic goals, and objectives. The mission
and the strategic goals are derived and aligned with organizations mission and
goals. Then, maintenance strategies and programs are developed to achieve a set
mission and goals. Maintenance capacity planning is medium- to long-range
planning for maintenance involves the determination of the maintenance resources
that are needed to perform the maintenance load in order to achieve organizational
objectives such as availability, reliability, quality rates and delivery dates.
Maintenance forecasting is an essential activity for planning. Maintenance
forecasting comprises the estimation and prediction of the maintenance load. The
maintenance load drives the whole maintenance system, and it consists of two
major categories. The rst category is the scheduled and planned maintenance
which is formed of (1) routine and preventive maintenance, (2) scheduled overhauls
which involve closure or plant shutdown, (3) corrective maintenance that involves
determining the causes of repeated breakdown and substandard performance as a
result of design malfunction, and (4) scheduled overhaul, repair, or building of
equipment which is not covered under item 2. The second category is emergency or
breakdown maintenance load. This category depends primarily on the failure pat-
tern, and it is a major source of uncertainty in the planning process. The sum of the
Springer International Publishing Switzerland 2015 19

S.O. Duffuaa and A. Raouf, Planning and Control of Maintenance Systems,
DOI 10.1007/978-3-319-19803-3_2
20 2 Maintenance Strategic and Capacity Planning
maintenance load for the two categories is a random variable, and it is the major
factor in determining maintenance capacity.
In this chapter, techniques of maintenance load forecasting, strategic planning,
and capacity planning are presented. Section 2.2 covers a brief introduction to
forecasting. Section 2.3 describes qualitative forecasting techniques. Section 2.4
presents a host of quantitative forecasting models including moving averages,
regression analysis, exponential smoothing, and seasonal forecasting. Section 2.5
covers error analysis and measures for testing forecasting models. Section 2.6
outlines the procedure for maintenance load forecasting. Maintenance strategic
planning is outlined in Sect. 2.7, and Sect. 2.8 introduces the problem of capacity
planning in maintenance. Deterministic techniques for capacity planning are pre-
sented in Sect. 2.9. Section 2.10 outlines stochastic techniques for capacity plan-
ning. The chapter is summarized in Sect. 2.11.
2.2 Forecasting Preliminaries
Forecasting techniques can be classied into two approaches: qualitative and

quantitative. Qualitative forecasting is based on the expert or engineering experi-
ence and judgment. Such techniques include historical analogy, surveys, and the
Delphi method. Quantitative techniques are based on mathematical models that are
derived from the historical data estimates for future trends. These models are either
time series-based data such as moving averages and exponential smoothing or
structural such as regression models Montgomery and Johnson [7].
A forecasting model is judged by the following criteria: (1) accuracy, (2) sim-
plicity of calculation, (3) data needed for the model and storage requirements, and
(4) flexibility. Accuracy is measured by how accurate the model predicts future
values and is judged by the difference between the model forecasts and the actual
observed values. In general, high accuracy requirements demand a complex rela-
tionship and therefore increase the complexity of computation. Flexibility is the
ability to adjust to changes in the conditions. In other words, it is a measure of the
robustness of the forecasting model. Important considerations in the selection of the
forecasting approach are as follows: (1) the purpose of the forecast, (2) the time
horizon for the forecast, and (3) the availability of the data for the particular
approach. The following are the steps for developing a quantitative forecasting
model.
1. Identify the characteristic/item to be forecasted and understand its nature. Dene
the purpose of the forecast and its time horizon.
2. Screen and validate available data for errors and outliers. Identify the additional
data needed and the methodology for collecting it.
3. Use the available data and graphical techniques to hypothesize appropriate
models. The model represents a relationship that describes the historical pattern
of the data or a relationship between the dependent and independent variables.
2.2 Forecasting Preliminaries 21
4. Use the major part of the data to estimate the parameters of the models. Keep
part of it for testing and validating the model. The estimation can be accom-
plished by an appropriate statistical method such as the least squares or the
maximum likelihood.
5. Test and validate the models and select the most appropriate one. Simulation
and error analysis are useful tools for testing, validating, and selecting the most
appropriate one.
6. Monitor the selected forecasting process and model to detect out-of-control
conditions and nd opportunities for improving forecasting performance.
Improvement can be made by rening parameter estimation or changing the
forecasting model. The cycle of the forecasting process is shown in Fig. 2.1
Identify characteristic/item to
be forecasted. Define
forecasting horizon.
Is
historical
Use qualitative No Yes Use quantitative
data forecasting.
forecasting.
available
Use historical data and

graphical techniques to
hypothesize statistical models
that appear to
fit the data.
Test and validate Estimate model

forecasting models and
Parameters.
select the most appropriate
one .
Generate Forecasts.
Monitor and
control the
forecasting
process
Fig. 2.1 The cycle of the forecasting process

2.3 Qualitative Forecasting Techniques
In the absence of data, the analyst must rely on estimates of experts and their
judgment. The role of the analyst in qualitative forecasting is to systematically
extract information from the mind of the expert by using structured questionnaires
or interviews. He should help the expert or management to quantify their knowl-
edge. Techniques such as cause-and-effect diagrams and the Delphi method can be
helpful in identifying relationship among the variables. The analyst should identify
which variables influence the forecast and the impact of each one.
After identifying the variables and their impact, the next step is to get an
agreement on the magnitude of the variables. Best-case, expected-case, and
worst-case scenarios are usually used to estimate the magnitude of the variables. An
interactive approach can be used to present arguments to the expert, such as why his
estimate differs from the average estimate, and he is asked to revise his/her estimate
until a reasonable consensus is reached. When no further reduction in variation
about the consensus is possible, the result is used as a forecast.
2.4 Quantitative Forecasting Techniques
In this section, quantitative forecasting techniques are presented. The models pre-
sented depend on the availability of the historical data and are usually referred to as
time series or structural models. These models either assume future values follow
historical trends or that a predictor (independent) variable exists that can provide a
model or a functional relationship that predicts the characteristic under study. For
example, the age of the equipment can predict the number of maintenance hours
required on the equipment. The models presented here include moving averages,
regression analysis, exponential smoothing, and seasonal forecasting.
2.4.1 Simple Moving Average
Suppose the characteristic under study is generated by a constant process, plus a

random error. An example of this could be the load xt exerted on an electronic
component. Mathematically, this can be represented as
xt b t
where b is a constant and t is a random variable with mean 0 and variance r2 .
To forecast the future value, we need to estimate the parameter b. Suppose we
have the time series observations, x1 ; x2 ; . . . xn : If all the observations are assumed
to be equally important, i.e., of equal weight, and if we use the least squares method
2.4 Quantitative Forecasting Techniques 23
(see Appendix A, Section A8), then we select a value of b that minimizes the sum
of squared error denoted by SSE.
X
n
SSE xt b2 2:1
t1
Equation (2.1) is differentiated with respect to b and equated to zero, and Eq. (2.2)
is obtained.
dSSE Xn
2 xt b 0 2:2
db t1
The estimator of b, ^b, is given as

Pn
^b i1 xt
2:3
n
which is just the average of the observations on hand. This method generates the
next periods forecast by averaging the actual observations for the last n periods.
Example 1 If the maintenance load in man-hours for the last 6 months is given as
Month 1 2 3 4 5 6
Maintenance load 200 300 200 400 500 600
Find the load forecast for periods 7 and 8 using a 3-month moving average.
The forecasted load for month 7 using n = 3 is
400 500 600

^x7 500
3
We did not observe x7, so if the load in month 7 is estimated as 500 as calculated
above. The forecast for the 8th month is obtained as follows:
500 600 500

^x8 533:33
3
2.4.2 Weighted Moving Average
It is logical to assume that the most recent observations should have more contri-
butions to future forecasts than to more distant observations, especially when the
data are not stable (e.g., there are signicant changes). The idea of the weighted
moving average is to give each observation a different weight. The forecasting
relationship is
X
n
xn1 wt xt 2:4
i1
where wi = weight for the ith actual observation
X
n
wi 1 2:5
i1
The formula given above can be obtained mathematically by minimizing the sum of
the weighted square error using the same procedure as in Sect. 5.4.1. The values of
wi can be determined empirically or be estimated based on experience. If the values
of wi are determined based on experience, this method combines qualitative and
quantitative forecasting approaches.
Example 2 Assume, for the data used in example 1, that the most recent period
should weigh twice as much as the previous months. Find the forecasts for x7 and x8
using three-period moving average:
w4 w5 w6 1
w6 2w5 2w4 2w
4w 1
w 1=4; w5 w4 0:25 w6 0:5
Therefore,
^x7 0:25400 0:25500 0:5600 525
Assuming the load for the seventh month x7 525, the forecast for month 8 is
given as
^x8 0:25500 0:25600 0:5525 537:5
2.4.3 Regression Analysis
If an independent variable exists that can predict a characteristic (dependent vari-

able) and a reasonable correlation exists between the two variables, then a
regression model can be used. For example, if the cost of maintenance for this
period, y(t), is a linear function of the number of operational hours in the previous
period, x(t 1), then the model would be
yt a bxt 1 t 2:6
where r2 :
y(t) desired forecast value for the cost in period t
x(t 1) operational hours in period t 1
a, b parameters to be determined
t a random variable with mean 0 and variance
After estimation, the resulting model that can be used for predicting is of the form:
^yt â ^bxt 1
It could be possible that the dependent variable is a function of more than one
independent variable. In the case of the maintenance cost, it could be a linear
function of operational hours in the previous period, x(t 1), and the age of the
plant, t. Mathematically, this can be expressed as
yt a bxt 1 ct t 2:7
where a, b, and c are parameters to be determined.

In case of one variable, the parameters a and b can be determined or estimated
by nding a good trend line that ts the data points by visual estimation. The
parameters a and b are the intercept and the slope of the line, respectively. If more
precision is desired, a regression analysis is used. Regression analysis refers to the
process of estimating the model parameters using the least squares method. This
method ts a line to the observations such that the sum of the square vertical
distances from the line is minimized.
The basic equation of a straight line showing a linear trend between an inde-
pendent variable and a dependent variable x(t) that represents demand for main-
tenance work is
xt a bt t 2:8
where a is the intercept and b is the slope that needs to be estimated. The parameter
estimation is to determine a and b. The least squares method estimates b and a in
terms of xt and t; t 1; 2; . . .; n, as follows:
Pn Pn Pn
^b n txt i1 t t1 xt
t1
P Pn 2 2:9
n nt1 t2 t1 t
Table 2.1 Data for parameter t x(t) tx(t) t2

estimation for the linear
regression model 1 15 15 1
2 25 50 4
3 30 90 9
4 45 180 16
5 50 250 25
6 70 420 36
7 85 595 49
Summed values 28 320 1600 140
â x ^bt 2:10
where
X
n X
n
x xt=n and t t=n 2:11
t1 t1
The resulting equation that can be used for prediction is
xt â ^bt
If the model is taken to be constant plus random variation, i.e., xt a t , then

the estimate for a is the average as obtained earlier in Sect. 2.4.1.
Regression analysis can easily be generalized to the case of multiple variables
and a polynomial relationship between dependent and independent variables.
Example 3 The monthly maintenance load in man-hours is given in the table
below. Develop a straight line that best ts the data and can be used to predict future
maintenance load.
Month t 1 2 3 4 5 6 7
Load x(t) 15 25 30 45 50 70 85
The data set and intermediate computation are given in Table 2.1.
The equation of the straight line is
xt a bt
The slope of the line is estimated as follows:

P7 P P
7 7
n t1 txt t1 t t1 xt 71600 28320
^
b 11:43
P7 2 P7 2 7140 282
n t1 t t1 t
320 11:4328
^
a xt bt 0:005
7 7
Therefore, xt 0:005 11:43t.

The above equation can be used for forecasting future load. For example, the
load in the ninth month is obtained by substituting 9 in place of t in the equation
obtained and is equal to 102.82 man-hours.
2.4.4 Exponential Smoothing
Exponential smoothing is a widely used forecasting method that is simple, efcient,

and easy to apply. It assigns weight to observations of previous periods in an
inverse proportion to their age. It does this in a very ingenious manner in which
only three pieces of data are required to generate the next periods forecast. These
are (1) last periods forecast, (2) last periods actual observation, and (3) a
smoothing factor, which determines the relative weight given to the recent obser-
vation. The basic equation that is the heart of the exponential smoothing is the
following:
^xt a xt 1 1 a^xt 1 2:12
where
^xt the forecast for period t and all future periods in the case of a constant
model
x(t 1) actual demand at period t 1
^xt 1 the forecasted value for t 1
smoothing constant, 0\a\1:
The exponential smoothing approach can be used to estimate the parameters for a
constant model, linear model, and any polynomial functional form. The parameter
estimation for the constant and the linear cases is given below:
In the constant case, or zero growth, the process model is given as
xt b t 2:13
where
b expected demand in any period
t random component having mean 0 and variance r2 :
At the end of the period t 1, we have the observations x(1),, x(t 1), from
which we need to estimate b and r2 . This model is referred to as simple exponential
smoothing or
xt axt 1 1 a^xt 1 2:14
where is determined using experimentation or judgment. A large value of

(closer to 1) indicates a belief that the current observation carries a high weight. In
other words, the system has shifted and the most recent observations resemble its
behavior. However, small values of indicate a belief that the past still resembles to
a great extent the system.
If the plotted historical data suggest a linear growth over time, the model is
considered a linear model. In that case, the process mean changes linearly with time
according to the following equation:
xt a bt t 2:15
where the expected demand at time t is a linear function of time.
Extjt a bt 2:16
It is known that the lag (the amount by which the forecast deviates from the most
recent data value) at the most recent data value in the case of the linear model is

b b
lag slope b 2:17
a a
If we apply the exponential smoothing again (double smoothing) denoted by
^xt a^xt 1 1 a^^xt 1 2:18

h i b
lag xt 1 ^xt 1 ^xt 1 ^^xt 1 b 2:19
a
at each period t 1, the values of a and b are updated as follows:
at 1 xt 1 ^xt 1 lag 2bx t 1 b

^ bx t 1 2:20
h i
^bt 1 a ^xt 1 ^
^xt 1 2:21
b
Initial conditions to start the process are

^b1 xt 1 x1
N2
â1 x1
b
^x1 â1 ^b1
a
^^x1 â1 2^b1 b
a
Example 4 Consider the data given in example 3 for a linear t by regression. Use
exponential smoothing with a linear growth model and a 0:2. Find ^x8 and
^x10:
a 0:2
b 1 a 0:8
Computing Initial Conditions
â1 x1 15
^b1 xN x1 85 15 70 11:67
N1 6 6
0:8
^x1 15 11:67 31:68
0:2

^^x1 15 211:67 0:8 78:36
0:2
The rst and double exponential smoothing are given in Table 2.2.
Estimates for a(7) and b(7) can be obtained using Eqs. 2.20, 2.21, respectively.
^
a7 218 10:67 36 10:67 46:67
^ 0:2 1
b7 18 10:67 18 10:67 7:17
0:8 4
Table 2.2 The rst and double exponential smoothing

t x(t) x(t) b^xt 1 ^xt a^xt b^
^xt - 1) ^
^xt
1 15 (31.68) (78.36)
2 25 5.0 25.34 20.34 4.07 62.69 66.76
3 30 6 16.27 10.27 2.05 53.41 55.46
4 45 9 8.22 0.78 0.16 44.37 44.21
5 50 10 0.62 10.62 2.12 35.37 33.25
6 70 14 8.50 22.5 4.5 26.0 22.10
7 85 17 19 35 7 17.68 10.67
The prediction for any period after L units from 7 is given as
^x7 L â7 ^b7l

^x8 46:67 7:17 53:84
^x10 a7 b73 68:18
2.4.5 Seasonal Forecasting
In many cases, the process under study might exhibit seasonality characteristics.
For example, demand for electricity is high during summer months in the Middle
East, or the rate of absenteeism might be higher at the beginning or at the end of the
week. Also, the maintenance load might be higher in certain seasons due to weather
and operational conditions. The appropriate period to look for seasonality should be
driven by the nature of the operation under study. A quick way to check seasonality
and growth is by plotting the historical data.
A set of logical steps to be followed when forecasting a characteristic with
combined seasonality and growth is as follows:
1. Plot the data and visually determine clear time series characteristics.
2. Determine the growth model and remove the growth component from the data.
One way to remove the growth component from the data is to determine an
average period for each cycle and divide each data value by the average.
3. Determine whether a signicant seasonality is present in the data as it appears
with the growth component removed (degrowthed). The seasonality index can
be computed by averaging the degrowthed data over the seasons (periods
exhibiting the similar behavior).
4. Deseasonalize the original data and analyze the growth factor. A plot of the
deseasonalized data will reveal the form of the growth component. The des-
easonalizing is accomplished by dividing each data by the appropriate seasonal
index.
5. Fit the data by some appropriate method, least squares regression, exponential
smoothing, etc.
6. A forecast for the future consists of a combination of seasonal and growth
trends.
The above steps are demonstrated on the following data. The following data show
the number of hours lost due to absenteeism and late arrivals at the maintenance
department for a period of 4 weeks.
The data in Table 2.3 exhibit seasonality. It is clear that the data values for
Mondays and Fridays are high. The data for Tuesdays, Wednesdays, and Thursdays
are relatively low. The daily averages for each week exhibit a growth model (7, 9,
11, 13). To remove the growth (step 2 above), the data for each day in the week are
Table 2.3 Seasonal data for number of hours lost with ve seasons (days) across the period
Week Days
Monday Tuesday Wednesday Thursday Friday Average
1 9 3 2 5 16 7
2 11 5 5 9 15 9
3 13 6 7 10 19 11
4 16 8 7 12 22 13
Table 2.4 Seasonal data for number of hours lost with ve seasons (days) across the period
Week Days
Monday Tuesday Wednesday Thursday Friday Average
1 1.29 0.43 0.29 0.71 2.29 5.01
2 1.22 0.55 0.56 1.0 1.67 5.01
3 1.18 0.55 0.64 0.91 1.73 5.01
4 1.23 0.62 0.54 0.92 1.69 5.0
Total 4.92 2.16 2.03 3.54 7.38
Average (Id) 1.23 0.54 0.51 0.89 1.85 5.02
divided by the daily average (e.g., the Monday value of week 1 is found by dividing
7 1:20. This has been performed in Table 2.4. The seasonality index for each day
9
is obtained by averaging the degrowthed data over the 4 weeks. Also, Table 2.4
shows the seasonality index for each day in the last row.
The next step is to deseasonalize the data and determine a model for the growth
component. One approach for this is to divide each data point by the seasonal index,
and then a growth model should be developed for the daily averages. This is shown
in Table 2.5.
The daily averages seem to grow linearly. Using linear regression, the parameter
for the growth component is given as
^xt 4:15 2:36t 2:22
Table 2.5 Deseasonalized data

Week Days
Monday Tuesday Wednesday Thursday Friday Daily Average
total
1 7.32 5.56 3.92 5.62 8.65 31.07 6.21
2 8.94 9.26 9.80 10.11 8.11 46.22 9.24
3 10.57 11.11 13.73 11.24 10.27 56.92 11.38
4 13.01 14.81 13.73 13.48 11.89 66.92 13.38
Table 2.6 Weeks 5 and 6 daily forecast with xt Id 4:15 2:36t

Week Days
Monday Sunday Wednesday Thursday Friday Daily average
5 19.62 8.61 8.13 14.20 29.51 15.95
6 22.52 9.89 9.34 16.29 33.87 18.31
Table 2.7 Weeks 5 and 6 daily forecast with xt Id 5:0 2t

Week Days
Monday Tuesday Wednesday Thursday Friday Daily average
5 18.45 8.1 7.65 13.35 27.75 15
6 20.91 9.18 8.67 15.13 31.45 17
To obtain the daily forecasts for the 5th and 6th week, use the equation above to
forecast the weekly average. Then, multiply the weekly average by the daily sea-
sonal index. For example, the forecast for the 5th week average is
x5 4:15 2:365 15:95
To obtain the forecast for Monday in the 5th week, this value is multiplied by 1.23,
the seasonal index for Monday Im. It is 1:2315:95 19:62. Table 2.6 presents
the daily forecast for the 5th and 6th week.
An alternative approach to estimate the growth component is to total or average
the data in each cycle. This approach would lead to the daily average for each week
and is shown in Table 2.3. These data show a perfect linear trend. Regression
analysis for the four daily averages would yield the model
^xt 5:0 2t 2:23
Equation (2.23) predicts the daily averages for the 5th and 6th week as 15 and 17.
The daily forecast is obtained by multiplying the daily average for the week by the
daily seasonality index Id. The daily forecasts are shown in Table 2.7.
The forecasts given in Tables 2.6 and 2.7 are both logical and sound. The logical
questions are which one is better and which one should the decision maker adopt for
planning purposes. Error analysis provides an approach for making the selection.
2.5 Error Analysis
Forecast error analysis provides a valid approach for checking the effectiveness of a
forecasting model. It also provides a sound methodology for evaluating and
selecting from several forecasting models that are available for a particular
situation.
2.5 Error Analysis 33
The forecast error at period t is the difference between the actual data value x(t)
and the forecasted value for it
et xt ^xt 2:24
The sum of the errors
X
N X
N
et xt ^xt 2:25
t1 t1
is not a valid measure of effectiveness of a forecasting model, but it is a measure of

bias. The sum of the errors should approach zero if the model is tted using the least
squares method. The sum of the errors has the problem that large positive errors, e
(t), can offset large negative errors. To eliminate this problem, we take either
absolute errors or squared errors. The following error measures are commonly used
for error analysis and evaluation of forecasting models.
1. Mean absolute deviation (MAD)
PN
jxt ^xtj
MAD t1
2:26
N
2. Mean-squared error (MSE)
P
N
xt ^xt2
MSE t1 2:27
N
3. Mean absolute percent error (MAPE):
N
100 X xt ^xt

MAPE 2:28
N t1 xt
4. Mean-squared percent error (MSPE)
" 2 #
100 X
N
xt ^xt
MSPE 2:29
N t1 xt
One of the above measures can be calculated for all the available models, and the
one with the minimum value is selected. This approach can be applied easily for
more on some of the models presented in Sect. 2.4 see [3, 7].
2.6 Forecasting Maintenance Work
Prior to performing capacity planning or designing a new maintenance organiza-

tion, it is essential to have some forecast of the expected maintenance load. The
load comprises the following:
1. Emergency maintenance workload. This can be forecasted using actual histor-
ical workloads and the appropriate techniques of forecasting and/or management
experience. This component of the load is random and can be minimized by
having a well-designed planned maintenance.
2. Preventive maintenance workload. This can be forecasted using actual historical
records coupled with newly developed preventive maintenance programs. This
should include routine inspection and lubrications.
3. Deferred corrective maintenance. This can be forecasted based on historical
records and future plans.
4. A forecast for overhaulremoved items and fabrication. This can be estimated
from historical records coupled with future plans for improvements.
5. Shutdown, turnarounds, and design modications. This can be forecasted from
actual historical records and the future maintenance schedule.
The forecasting of the maintenance load for a new plant is more difcult and must
rely on similar plants experience, benchmarking, management experience, and
manufacturers information (Table 2.8).
Once a plant is in operation, errors in forecasting and job standards may lead to a
backlog. An alternative approach is to do forecasting by examining the maintenance
backlog. The usefulness of calculating a backlog for planned work is seen when
load forecasting is done for the coming weeks or months work. Table 2.8 illus-
trates this.
The table shows a year-to-date average emergency work level in mechanical
trades of 400 h and in electrical, 80 h. This week PM requirements are 1200 h for
mechanical and 600 h for electrical. There are an average number of hours required
for the minor unplanned work that is done with non-emergency, yet high priority,
work orders. Some work was begun last week but was not nished at weeks end; it
will be completed this week. The total planned priority 3 work in backlog awaiting
schedulingthat is work that has been planned and the materials required are
Table 2.8 Weekly backlog Mechanical Electrical

report in hours
YTD average emergency 400 80
This weeks PM level 1200 600
YTD average minor unplanned 600 400
This weeks carryover work 800 40
Total planned priority 3 work 12,000 900
backlog
Total shutdown backlog 10,000 3000
2.6 Forecasting Maintenance Work 35
availableis 12,000 h for the mechanical and 900 h for electrical. Finally, there is a
total of 13,000 h of mechanical and electrical work awaiting the next plant
shutdown.
This represents the total backlog that is relevant for scheduling the maintenance
work for the upcoming week.
2.7 Maintenance Strategic Planning
This section addresses maintenance planning at the strategic level which falls under
the long-range planning. In the past, maintenance is not considered as a strategic
unit within the organization, but lately this has changed and researchers and
maintenance decision makers have brought forward the strategic role of mainte-
nance. Maintenance strategic planning is the process of ensuring the alliance of
maintenance mission, goals, objectives, and programs with the organizations
mission and objectives. Recent work of Tsang [9], Murthy et al. [8], and Al-Turki
[1] has addressed and identied several important issues that are essential in
deciding maintenance strategic plan. These issues include maintenance mode of
delivery, organization, methodology, and support. The following ten steps are
suggested to put to work the frameworks suggested in the recent literature:
Top management develops organization mission, strategic goals, and objectives.
Maintenance management (MM) reviews the organization mission and objec-
tives and identies the role of maintenance in achieving them.
MM analyzes the current internal and external situations related to the main-
tenance function using the methodology for determining strengths, weaknesses,
opportunities, and threats known as SWOT analysis. SWOT leads to the iden-
tication of the strategic issues that need to be addressed. The strategic issues
are expected to include mode of delivery, maintenance organization, method-
ology, manning, and training.
MM formulates the identied role as a mission statement for maintenance and
develops strategic goals and objectives to address the strategic issues.
Align and prioritize maintenance strategic goals, objectives, and programs with
maintenance mission.
MM checks the alignment of maintenance strategic goals and objectives with
organization set mission and strategic goals and objectives.
MM develops maintenance programs to achieve set strategic goals and
objectives.
Develop a set of quantitative performance measures for the maintenance stra-
tegic goals and objectives.
Evaluate periodically the progress toward achieving the goals and the objectives
using the performance measures and identify the gap between the actual and the
desired situation.
Identify the root causes of the gap and implement effective corrective actions.
2.7.1 Maintenance Strategic Issues
In this section, several strategic issues in maintenance that are identied in the
literature are discussed. The rst issue is the delivery mode. This issue deals with
whether maintenance is outsourced, conducted in-house or a combination of both.
Each option has its pros and cons. The selection of the right option should be made
in light of the set maintenance strategic goals and objectives at the same time to
minimize risks and threats to maintenance systems efciency and effectiveness.
The second maintenance strategic issue is maintenance organization as identied by
Tsang [10]. Several options are available in this issue. The options include func-
tional, process, or network. Each of the previous organization options can be
crossed with centralized, decentralized, or a cascade option. The third issue is the
work structure and control. Many options are available for the work structure
depending on the reporting and supervision structure. The fourth strategic main-
tenance issue is the selection of the support system that includes information sys-
tem, training, and performance management and reward system. Each element has
to be carefully selected to support the overall objective of the organization.
2.8 Maintenance Capacity Planning
Maintenance capacity planning determines the optimal level of resources (crafts,

skills, spares, inventory equipment, and tools) to meet the forecasted maintenance
load that consists of future load forecast plus the maintenance backlog. An essential
element of capacity planning is the determination of the skills of craftsmen, the exact
number of various types of craftsmen, the healthy level of backlog, overtime capacity,
and contract maintenance. The optimal allocation of the maintenance resources to
meet a random and varying workload is a complex and challenging problem.
Capacity planning techniques play very important roles in handling this complex
problem. The steps involved in capacity planning can be summarized as follows:
1. Determine the total maintenance load.
2. Estimate the required spares and material to meet the load.
3. Determine equipment and tools that are necessary for all types of maintenance
work.
4. Determine the skills and the number of crafts from each skill. A special attention
should be given to multi-skill crafts.
5. Sometimes in highly computerized equipment, there may be a need to provide a
certain specialty that may require special plans.
In maintenance capacity planning, a major issue is to determine the optimal mix of
skills of crafts from the available sources to the organization. The usual available
sources are regular and overtime in-house crews and contract maintenance. The best
mix using these sources is determined using cost and availability measures.
2.8 Maintenance Capacity Planning 37
Multi-skilling provides capacity planners with options and alternatives that

enrich and generate a wide range of options for determining the optimal capacity.
Crafts with multi-skills can be utilized in more than one type of maintenance work.
This usually improves manpower utilization.
Techniques for capacity planning can be divided into two major categories:
deterministic and stochastic. The deterministic approach assumes that the forecasted
maintenance load, standard times, and other random variables are xed constants.
A plan that minimizes cost or maximizes availability depending on the organization
objectives that may include reliability, availability, and cost is then determined. In
this chapter, two deterministic techniques will be presented, namely:
1. Heuristic tableau method, and
2. linear programming.
The stochastic approach models the maintenance load, standard times, job arrival
times, and other variables as random variables with certain probability distributions
and uses standard statistical techniques to identify these distributions. Then, a
stochastic model is utilized to determine the optimal capacity that meets the
maintenance load. Two stochastic techniques will be presented in this chapter,
namely:
1. Queuing models and
2. stochastic discrete event simulation
Next, a brief description for these techniques is provided.
2.9 Deterministic Approaches for Capacity Planning
In Sects. 2.9.1 and 2.9.2, two deterministic techniques for capacity planning are
presented. These are the heuristic tableau method and linear programming.
2.9.1 Heuristic Tableau Method
In maintenance capacity planning, the heuristic tableau method derives intuitively

appealing plans to determine a feasible craft mix, based on sound principles and
guidelines. The tableau is used to evaluate the cost of each alternative, and the plan
with minimum cost is selected. Sound principles and guidelines include, providing
sufcient in-house crafts for high priority work, a reasonable ratio of overtime work
to regular time work and a xed level of a healthy backlog. The following dis-
cussion presents the reasoning that may be adopted for determining the required
crafts and skills to meet the forecasted mechanical maintenance load using this
approach. The same approach can be repeated for other types of maintenance load
such as electrical or instrument. As an example, the mechanical workload is rst
classied into two grades. The skill the work requires and its impact on the facility
determines the grade. In this example, it is classied as grade one and grade two
mechanical workload, although other classications based on skill and priority are
possible. Prior to presenting the plan in Table 2.9, the following notation is dened
as follows:
FMt Total forecasted mechanical load

Bt1 Mechanical workload backlog from period t 1
TMt Total mechanical workload, for period t. TMt = FMt + Bt 1
TMi,t Total mechanical workload of grade i in period t, i = 1,2
RMi,t Regular in-house capacity for mechanical workload of grade i in period t
OMi,t Overtime capacity for mechanical workload of grade i in period t
CMi,t Contract capacity for mechanical workload of grade i in period t
The plan is derived based on the following commonsense principles and guidelines:
1. All priority work is met by regular in-house crafts as much as possible.
2. If it is not possible to satisfy priority one work by regular in-house crafts use
overtime.
3. No backlog is allowed for grade 1 work.
4. The manning level must be determined based on the average maintenance load
plus a healthy backlog from grade 2 work.
5. The priority two work is met with overtime or contract maintenance.
6. The overtime capacity is at most 25 % of the regular in-house capacity.
7. In the example, the maximum for the backlog is 100 man-hours. If the backlog
exceeds this limit, subcontracting is utilized, and it is assumed that subcon-
tracting can provide as much capacity as needed.
If the periods in Table 2.9 are taken to be weeks and the guidelines (16) are taken
into consideration, a flexible plan for meeting the maintenance load from the
available sources is shown in Table 2.9. Column two contains the forecasted
mechanical maintenance load, and column three has the expected backlog from the
previous period. Column four contains the total load. Columns ve and six present
the two grades of maintenance work. The sum of these columns gives column four.
Columns seven and eight present the amount of work that will be met by regular
in-house maintenance crafts. Columns nine and ten contain the work that is
expected to be performed by overtime in-house maintenance. Columns eleven and
twelve present the amount of contract maintenance for both grades of mechanical
work. The hours in all the columns are standard hours.
It should be noted that this is a target plan; the realized workload might be
different from the forecasted load, and an adjustment to the plan at the execution
stage might be needed. Different plans can be generated based on the same
guidelines, and the one with minimum cost is selected. The required number of
employees can be based on the number of standard hours in the regular in-house
column, after making an allowance for the expected productivity of the trades.
Table 2.9 Sample data for heuristic tableau capacity planning
Col. # 2 3 4 5 6 7 8 9 10 11 12
Period t FMt Bt1 TMt TM1,t TM2,t RM1,t RM2,t OM1,t OM2,t CM1,t CM2,t
1 150 100 250 150 100 150 10 40
2 200 50 250 150 100 150 10 40
3 250 50 300 150 150 150 10 40
4 180 100 280 180 100 160 20 20
5 200 80 280 200 80 160 40
6 150 80 230 120 110 120 40 40
2.9 Deterministic Approaches for Capacity Planning
7 200 30 230 150 80 150 10 40

8 250 30 280 150 130 150 10 40
9 200 80 280 150 130 150 10 40
10 300 80 380 200 180 160 40 10 110
11 250 70 320 210 110 160 40 10
12 200 60 260 200 60 160 40
39
The above heuristic tableau approach can be formalized by using a tableau

similar to the one used in production planning with some modications. In this
tableau, the sources of manpower supply are shown on the left side of the tableau
and the capacity of each source on the right-hand side on the same row. The
maintenance load for each period is shown in the columns of the tableau. The cost
of meeting the load from a particular source is shown in the corner of each cell in
the table.
In production planning products, from the current period can be kept in
inventory to satisfy future periods demand, however, this is not the case in main-
tenance planning. In maintenance, unnished work is backlogged and performed in
future periods at an added cost to the system. It is also possible to divide the
maintenance load in each period by skill or priority, and the same method can be
applied. Table 2.10 shows a tableau for a three-period maintenance plan for one
kind of maintenance work which is the mechanical workload. The same table can
be replicated for other types of work. The notations on the tableau are as follows:
Cr hourly cost of mechanical trade on regular time
Co hourly cost of mechanical trade on overtime
Cs hourly cost of subcontracting
Bt backlog in man-hours at the beginning of period t
CRt capacity of in-house regular time in period t
COt capacity of in-house overtime in period t
CSt capacity of subcontracting in period t
FMt forecasted maintenance load in period t
Table 2.10 Data for capacity allocation problem

2.9 Deterministic Approaches for Capacity Planning 41
If a work is backlogged for r period and performed at period r + 1 with regular in-house
manpower, it will have a cost of Cr + r per hour. Table 2.10 shows the data needed for
a three-period planning horizon. The table shows the costs, capacities, and mainte-
nance load. The symbol in the cost cell means that a work cannot be done in this
period. For example, work that came in period 2 cannot be done in period 1.
A simple least cost heuristic method can be used to compute the allocation of the
maintenance load to different sources of manpower supply. The method starts with
the least cost cell and satises the load as much as possible and then moves to the
next least cost until all the load is met. It is highly likely to attain near-optimal
solutions with the least cost method. Next, an example is given to demonstrate the
tableau approach.
Assume that we have three time periods with maintenance loads 400, 300, and
500, respectively. The in-house regular capacity is 200, 350, and 300 for periods 1,
2, and 3 respectively. The overtime is at most 25 % of the in-house capacity.
Subcontracting is abundant, and technically, there is no limit on this source. The
cost of performing one in-house man-hour is taken to be 1 unit, overtime man-hour
costs 50 % more than this regular time, i.e., 1.5 units, and the subcontracting costs 2
units. Backlogging of one man-hour costs = 0.3. The capacity for subcontracting
can be taken a large number in this example, and for purpose of demonstration, it is
taken to be 500 man-hours. The data and the solution for this situation are shown in
Table 2.11.
Table 2.11 Data and solution of the example

Applying the least cost heuristic, the plan in Table 2.11 is obtained. The plan
calls for meeting the maintenance load by 200 h of regular in-house, 40 h of
overtime, and 30 h of subcontracting in the rst period. In addition to 50 h of
regular time, and 80 h of overtime in the second period. The number of hours
backlogged from the rst period to the second period is 130. The second-period
load is met by 300 h of regular in-house maintenance in the same period. The
third-period load is met by 300 h of regular in-house, 75 h of overtime, and 125 h of
subcontracting. The total cost (TC) of the plan is sum of the costs in each cell,
which is obtained by multiplying the hours in the cell by the cost of the hour.
TC 1 2000 1:5 40 2 30 1:3 50

1:8 80 1 300 1 300 1:5 75 2 125 1491:5 units:
2.9.2 Linear and Integer Programming for Maintenance

Capacity Planning
Linear programming is a mathematical model that optimizes a linear function

subject to linear inequalities. The linear programming model determines the optimal
values of decision variables that optimize a given objective such as minimizing cost
or maximizing prot. Decision variables are elements under the control of the
decision maker, and their values determine the solution of the model. The objective
function in linear programming is the criteria with which feasible solutions are
evaluated. A feasible solution is a solution that satises all the constraints in the
system. A constraint is a condition on the system that must be satised.
In the case of maintenance capacity planning (MCP), the decision variables
could be the number of hours from different skills and trades made available for
maintenance capacity through regular in-house, overtime, or contract maintenance.
The objective could be to maximize resource utilization or minimize total cost. An
example of a constraint is the ratio of overtime hours to regular in-house hours
should not exceed certain percentage. In this section, linear programming is
introduced via a simple hypothetical example.
Suppose it is forecasted that the mechanical maintenance load is 100 man-hours
divided into two grades: 60 h of grade 1 and 40 h of grade 2. Grade classication is
based on the skill the work requires and the impact the work has on the facility. The
capacity in man-hours for this work can be provided from two skills of workers,
i.e., skill 1 and skill 2. Both skills of workers can perform the two types of
mechanical work but with different productivity. The productivity of skill one
worker is 0.75 and 0.8 for grade 1 and grade 2 mechanical works, respectively. The
productivity of skill two worker is 0.5 for grade 1 and 0.7 for grade 2 works. The
cost of one man-hour of skill 1 worker is 30 dollars, and the cost of one man-hour
of skill 2 worker is 20 dollars. The objective is to determine the number of
man-hours from each skill level workers assigned to perform the maintenance load
(work) at a minimum cost.
Prior to stating the model constraints and objective, let xij be the number of
man-hours of skill i assigned to perform the mechanical work of grade j,
i 1; 2; j 1; 2. The rst constraint is to meet the load for grade 1 mechanical
work. This is stated mathematically as
0:75x11 0:5x21 60 2:30
The second constraint is to meet the grade 2 mechanical work, mathematically

stated as
0:8x12 0:7x22 40 2:31
The non-negativity restrictions state that the man-hours obtained from each skill of
workers should be nonnegative, mathematically stated as
x11 ; x12 ; x21 ; x22 0
The objective function is to minimize the sum of the cost of meeting the required
load from the two skill levels. The linear programming model for determining the
capacity for the mechanical work in the example above is
Min 30x11 x12 20x21 x22 2:32
Subject to
0:75x11 0:5x21 60
0:8x12 0:7x22 40
x11 ; x12 ; x21 ; x22 0
Solving the above model will determine the allocation of man-hours from different
skills to perform the two grades of mechanical work at minimum cost. The solution
of the linear program could be fractional values. If we want to obtain the solution in
whole hours, an additional restriction must be imposed. That restriction requires the
variables to be integers. Adding this restriction transforms the linear program to an
integer program. In many real situations rounding the linear programming, solution
to the nearest hour would be a good solution.
The above formulation can be employed for maintenance capacity planning
(MCP) with minor generalization. The formulation will be given for planning the
capacity for the mechanical work. The objective is to determine the number of
hours from different skill levels from the mechanical trade made available from
different sources to perform different grades of mechanical work. The available
sources include regular in-house trades, overtime, and contract maintenance. The
same model can be used for determining the number of hours needed from other
trades such as electrical, instrument, and civil. The following notations are neces-
sary for stating the linear programming model: of a mechanical trade of skill i, from
source j when performing mechanical work of grade k
xijkt Number of man-hours from the mechanical trade of skill i (i = 1, 2, , I)
from source j (j = 1, 2, , J) made available to perform the mechanical work
of grade k (k = 1, 2, . K) in period t, t = 1, 2, , T
Cij Hourly cost of a mechanical trade of skill i from source j
Pijk Productivity
Fkt Forecasted mechanical load of grade k in period t
Bkt Backlog of work of grade k work in period t
UBk Upper limit for a healthy backlog for grade k work
LBk Lower limit for a healthy backlog for grade k work
Uijt Upper limit on the availability of skill i mechanical trade from source j in
period t
rkt Cost of backlogging one man-hour of grade k work in period t
The linear programming model for determining the required number of

mechanical man-hours of different skills, and sources made available to perform
all grades of the mechanical work, consists of an objective function and a set of
constraints. The objective function is to minimize manpower and backlog cost.
The constraints include the work balance constraints, a reasonable ratio of
in-house man-hours and overtime to limits on manpower availability. The model
is stated as follows:
!
XX XX XX
Min Cij xijkt rkt Bkt 2:33
i j k t k t
Subject to:
Work balance constraints,
XX
Pijk xijkt Bkt Fkt Bk;t1 2:34
i j
Limit on overtime man-hours in terms of in-house man-hours (taken as 25 %)
xi2kt 0:25xi1kt 0 2:35
Limits on man-hours availability,
xijkt Uijt 2:36

Lower and upper limits on backlog on different work grades,
LBk Bk UBk
i = 1, 2, , I (number of skills), j = 1, 2, , J (number of sources), k = 1, 2, ,

k (number of grades), t = 1 T (time period in the planning horizon).
The above model can be solved by any linear programming algorithm such as
the simplex-based code in the package Linear Interactive and Discrete Optimizer
(LINDO) or the one in the package International Mathematical Software Library
(IMSL) or the Optimization Software Library (OSL).
The output of the linear programming model will determine the optimal number
of man-hours of different skill levels from a specic trade (mechanical) made
available to perform the maintenance work. Based on the number of man-hours
from source 1 (in-house), the manning level will be determined.
The linear programming formulation assumed that the parameters of MCP
problem are xed constants, which in reality is not true. In order to examine the
sensitivity of the solution obtained, the standard linear programming sensitivity
analysis must be conducted. Sensitivity analysis addresses issues, such as what will
happen to the solution if the demand changed or a new cheaper resource became
available. Also, the complete linear programming solution that includes the dual
variables provides information about which resource to increase its level if the
demand for maintenance work increased. The interested reader is referred to Taha
[9] for more detail on the subject of dual variables and sensitivity analysis.
If we require the crafts and skill levels to be specied in terms of employees to
be hired in a maintenance department, we either round up the linear programming
solution to the nearest number of employees or reformulate the problem as an
integer programming model. This is done as follows:
Let employee of skill i, from source j in period t
nijkt Number of employees from the mechanical trade of skill i, from source j
made available to perform the mechanical load of grade k in period t
NR Sj Number of hours worked per period for an employee from source j (Sj is
a constant, and it is usually taken to be 1 if an employee is full time
in-house, and between 0 and 1 if an employee is part time or overtime.
NH is the number of hours worked by a regular in-house employee per
period)
Uijt Upper bound on the availability
Cijt Salary of an employee with skill i, from source j in period t
Assuming that the periods are in months and NH = 160, the integer programming
model for determining the number of employees from the mechanical trade of
different skills to perform the mechanical load is stated as follows:
XXXX XX
Min Cijt nijkt rkt Bkt 2:37
i j k t k t
Subject to:
XX
160Sj Pijk nijkt BRt F k;t Bk;t1 2:38
i j
ni2k 0:25ni1k 0 2:39
nijkt Uijt 2:40
LBR Bk;t UBk 2:41
nijk ; integer
The packages mentioned in this section have the capabilities of solving the integer
programming model.
2.10 Stochastic Techniques for Capacity Planning
Queuing models and stochastic simulation are two important techniques for
capacity planning. Queuing models address the situation where customers arrive at
a service facility, perhaps wait in a queue, and then are served by a server and thus
leave the facility. In maintenance capacity planning, the customers may take the
form of a maintenance job arriving at the maintenance planning and scheduling
unit, that job is then planned and routed to a workshop for repair (which represents
the service facility). The results from queuing theory allow us to evaluate the
performance of such systems under different congurations. Queuing theory has
been used to determine maintenance stafng. However, when queuing models are
used for capacity planning, the measures of performance are obtained for the system
under steady-state conditions which do not represent the transient system behavior
which is more close to day-to-day operations.
Stochastic simulation offers a viable alternative when the decision maker is
interested in a transient situation or when the system under consideration is complex
(which is usually the case in maintenance systems). In stochastic simulation, the
maintenance system is represented on the computer and well-designed experiments
(scenarios) are used for system performance evaluation. The experiment described
for capacity planning of a maintenance system is an allocation of maintenance crews
and skills under certain maintenance policy and procedures. The performance mea-
sures can be cost, utilization of resources, and availability of critical and major
equipment.
2.10 Stochastic Techniques for Capacity Planning 47
In this section, these two stochastic approaches for capacity planning are
presented.
2.10.1 Queuing Models
A queuing model can be described as follows: If the customers arrive at a facility,

they join a working queue (line). A server chooses a customer according to a certain
discipline from the queue and serves him. Upon service completion, the customer
leaves the service facility. A queuing model can represent many situations in real
life. A relevant example is the arrival of maintenance jobs to a workshop. Other
examples include the arrival of customers to a bank and the arrival of telephone
calls to a telephone number. The major components of a queuing system are the
customers and the servers. The interaction of the customer with the server is gauged
by the time the server spends serving the customer. The customers drive the system,
and we are interested in their arrival pattern or in their interarrival time, which is
usually modeled as a random variable. The service time is also modeled as a
random variable with a certain probability distribution. Other important elements in
a queuing model are: (1) the service discipline, which refers to the manner in which
customers are selected for service, e.g., it could be rst-come rst-serve (FCFS) or
some other certain priority rule; (2) the design of the facility, which refers to the
number of servers and design of the queues (parallel, series, tandem, network);
(3) the queue size (nite, innite); (4) the size of the source of arrivals (nite,
innite); and (5) human behavior (jockey, balk). In summary, the elements of a
queuing model are the following:
(1) arrival distribution,
(2) service time distribution,
(3) design of service facility,
(4) service discipline,
(5) customer population, and
(6) human behavior.
Specifying the above factors will result in a specic queuing model. Some queuing
models have steady-state results that give the expected number of customers in the
system, Ls, the expected queue length, Lq, the expected waiting time in the system,
Ws, and the expected waiting time in the queue, Wq. The above results can be used
to evaluate a queuing system.
A notation that summarizes the main characteristic of a parallel queuing model is
(a/b/c), (d/e/f) where
a Arrival distribution
b Service time distribution
c Number of parallel servers
d Service discipline
e Maximum number allowed in the system
f Size of customer population
One of the simplest queuing models is the one denoted by (M/M/C) (GD//), in
which there are C servers and interarrival and service times are exponential. If all
the servers have exponential service time distribution with parameter and if we let
= /, then we have the following steady-state results:
( )1
X q
c1 n
qc
P0 2:42
n0
n! c!1 q=c
where P0 is the probability of zero customers in the queue.
qc1
Lq P0 2:43
c 1!c q2
Ls Lq q 2:44
Lq
Wq 2:55
k
1
Ws Wq 2:46
l
If the above model is used as an approximation for a maintenance department, in

which we want to determine the optimal number of repairmen (servers). The
steady-state results can be used to evaluate machines availability and utilization.
Another queuing model known as the machine servicing model is one in which
we have R repairmen servicing a total of K machines, and, because, a broken
machine cannot generate a breakdown while in service, the arrival population is
nite. The number of repairmen R is less than K. The objective is to nd a value
R that minimizes the total expected costs that consist of the cost of failure and cost
of service. This type of model is denoted by (M/M/R):(GD/K/K), R < K.
The model is a special case of the general queuing model. If is the rate of
breakdown per machine, then if there are n broken machines, the arrival rate from
the system is given as:
kn K nk 0nK
2:47
0 nK
and the service rate is as follows:

8
< nl 0 n R
ln Rl R n K 2:48
:
0 n[K
Pn is the probability of n machines in the system. The steady-state results for the
system are derived as follows:
8
>
> K
qn P0 0 n P0
<
n
Pn 2:49
>
> K n!qn
: RnK
n R!RnR
XR Xk
K K n!qn
P0 q
n
2:50
n0
n nR1
n R!RnR
X
K
Lq n RPn
nR1 2:51
keff
Ls Lq R R Lq
l
PR
R expected number of idle repairmen = n0 fR ngPn
keff l R R kK Ls 2:52
If R = 1, the model yields results for a system with single server. If a maintenance
system can be represented by such a queuing model, its effectiveness can be
evaluated using steady-state measures.
2.10.2 Stochastic Simulation
Maintenance systems have several characteristics that make capacity planning a

rather complex problem. These characteristics are as follows:
Maintenance as a function interacts with other technical and engineering
functions in a complex fashion.
The maintenance factors are highly dependent on each other.
Maintenance as a function has many uncertain elements. These elements include
demand for maintenance, time of arrival of job requests, content, time to
complete a job, tools, equipment, and spare parts availability.
The complexity of the maintenance capacity planning suggests that simulation is

one of the most desirable approaches for modeling it.
Stochastic simulation is the process of representing a system on the computer and
then employing well-designed experiments (scenarios), to evaluate the system
performance. Using this process, systems can be analyzed, planned, and designed.
Law and Kelton [6] provide ten major steps for conducting a typical simulation
study. In this section, these ten steps are summarized in eight steps and the rela-
tionships among them are outlined:
1. Purpose of simulation: The rst step toward a successful simulation study is to
state precisely the purpose of the study. Simulation has been used in mainte-
nance systems for the following purposes: to determine the optimal crew size
and stafng, to evaluate the effect of maintenance policies on production sys-
tems, to design and plan maintenance operations, and to determine the shutdown
time periods.
2. Simulation models: The conceptual model used in building the computer sim-
ulation study will affect the simulation accuracy and efciency. The simulation
model should contain only the necessary information that captures the essence
of the system under study.
3. Model assumptions: The assumptions of a simulation model will affect the
realism of the simulation results. They also may affect the way results are
interpreted. Therefore, each assumption should be reviewed carefully before
putting it into effect. Availability of manpower, equipment, job standards, and
spare parts are some of the assumptions used in maintenance systems.
4. Data Accuracy: Accurate data and their distributions are very essential for a
reliable simulation model. To simulate a maintenance system, the distribution of
equipment failures and repair times must be identied using sound statistical
methodology.
5. Simulation languages and computers: One of the major jobs in building a sim-
ulation model is to convert the conceptual model into an actual computer simu-
lation program. There are over 100 simulation software programs currently
available for a variety of computers. Computer languages such as GPSS,
SLAM II, SIMAN, and SIMSCRIPT II.5 are generally used in simulation. Several
other criteria also have been used in practice to classify the simulation software.
6. Program verication and model validation: Verication is testing and checking
the computer code to show that it performs as intended. Validation is to ensure
that the models assumptions are realistic and correct, and the simulation model
fairly represents the behavior of the modeled system. Even though this step is
fairly tedious and time-consuming, it is the most important step in simulation
studies.
7. Design of experiment: Another signicant element in any simulation study is the
design of the experiment. This comprises the following:
(a) selecting experimental factors,
(b) selecting measures of performance,
(c) determining the initial conditions,

(d) determining the steady-state conditions,
(e) determining the length of a simulation run,
(f) determining the number of replications, and
(g) applying variance reduction techniques
8. Output analysis: In any simulation study, it pays very well to spend time on
output analysis. To check for the true estimate, test, validate, and decide on the
output results from your simulation, statistical techniques that ensure reliable
estimates for system performance must be used. These include deciding on the
length of the simulation run, the number of runs, and condence intervals for
estimated measures of performance.
A machine servicing model with general input distributions and general servicing
distributions can be used to model maintenance operations and capacity planning.
Depending on the maintenance activities and the maintenance load forecasted, the
system can be divided into types of jobs requiring different crafts. Craft types
include mechanical (ME), electrical (EE), instrumental (INS), civil (CI), and jani-
torial (JAN). The historical data can be used to develop a probability distribution
that represents the arrival of maintenance jobs and the time to complete repairs for a
job. The machine servicing model can be modied so that instead of considering
physical machines, each type of work generated by the system may be considered
as being generated by a hypothetical machine. Each type of work (machine)
requires a specic craft or skills. The number of hypothetical machines is deter-
mined by the types of work generated by the system. Such a representation can be
used within a simulation model to determine the size of crews needed. Figure 2.2 is
a schematic representation of a model that can be simulated to determine the stafng
levels in a maintenance department.
Craft Size Machines
Electrical
EE
Work
Mechanical
ME
Work Maintenance
Load is
Instruments Work Forecasted. It is

INS
Work Orders a random
Variable
Civil
CI
Work
Janitorial
JAN
Work
Fig. 2.2 Representation of maintenance capacity planning in a machine servicing framework

In order to implement a simulation model for capacity planning, the following

practical steps are needed Duffuaa and Raouf [5].
1. A detailed study of the organization/plant maintenance requirements to deter-
mine the types of maintenance crafts and crews required, the types and criticality
of equipment repaired, the failure mechanism for each equipment, and the effect
of a failure on production or service provided by the organization.
2. Forecast the maintenance workload and divide it according to priority.
3. Outline the existing work order system and dene the logic of work assignments
(system discipline).
4. Set up the relevant machine servicing model after determining the failure rate of
each machine, the service rate, and the cost of each machine being out of
service.
5. Develop the simulation software and verify and validate the model.
6. After the model is validated, perform production runs and, on the basis of all
measures of performance, nd the optimal stafng levels.
2.11 Contract Maintenance
Contracting out maintenance activities is common throughout all industrial and

public sectors. The premise is that certain aspects of maintenance can be done as
effectively and at a lower cost than in-house resources can accomplish it. A key
question is whether contracting maintenance builds or diminishes the competitive
advantage of the organization. It is difcult to see how, for example, contracting out
custodial services would harm the competitive advantage of a business, unless
somehow in-house personnel could do it more effectively and signicantly cheaper
and with less management attention than a rm specializing in facility cleaning. On
the other hand, if haul truck maintenance at a remote mine site is contracted out to a
local garage and body shop, it is possible that the mining operation could suffer,
regardless of the competitive rates.
In the petrochemical and oil rening sector in North America, it is not
uncommon to see the majority of the execution of maintenance contracted to a rm
specializing in that business, whereas the reliability engineering and maintenance
management is often kept in-house. Major shutdown and overhaul maintenance
requires the contracting out of a large segment of the shutdown work backlog,
because there is usually a short, nite time period to accomplish all the work and
not enough capacity within the organization to accomplish it.
Contracting maintenance during peak periods is more effective in most working
environments when nite projects or tasks can be estimated and outsourced as a
work package, as one would normally do for a capital expansion or modication
project. This allows for a separate contracted crew to work together outside the
2.11 Contract Maintenance 53
day-to-day organizational structure of the plant or facility. The work package would
specify the contractors trade skills and numbers and contain any required
instructions. Contracting individual trades and integrating them with the in-house
staff can lead to inefciencies and conflicts, as these contracted staff may not know
the routines, equipment, or working procedures and rules, thereby reducing the
productivity of the entire crew. Further, particularly in a unionized environment,
contracting out routine work without extensive bilateral communication on the
issue of job security can often lead to labor relations difculties and some
resentments that may lead to strikes or slowdowns.
2.12 Summary
This chapter presented the essential techniques for maintenance forecasting and
strategic and capacity planning. Forecasting is a prerequisite for capacity planning,
and capacity planning is a major element in maintenance organization and deter-
mines its ability to perform its mission in an effective way. The techniques covered
in the forecasting sections of this chapter are moving averages, linear regression,
exponential smoothing, and seasonal forecasting. Steps for developing a strategic
plan for maintenance are outlined. The methods presented for capacity planning are
simple heuristic tableau form, linear programming, queuing theory, and stochastic
simulation. The application of the stochastic techniques requires some maturity on
the part of the analyst and is introduced only briefly here (for more detail on
stochastic techniques see [14]).
Exercises
1. The forecasting of the maintenance load that is generated by a newly developed

sophisticated equipment does not seem to be amenable to time series fore-
casting. Why? Suggest a procedure to predict the maintenance load for the next
two years for this equipment.
2. In the absence of data, you may have to resort to qualitative forecasting.
Explain how would you perform qualitative forecasting and how to validate it?
3. When is it possible to use a control chart to control the qualitative forecasting
process? What type of control charts will you use? Explain how.
4. Prove that using the least squares method to estimate the slope, b, and the
intercept, a, of the single variable linear t, x(t) = a + bt results in estimates for
b and a as follows:
Pn Pn Pn
^b n t1 txt t1 t t1 xt
P Pn 2
n nt1 t2 t1 t
â xt bt
5. Use the least squares method to estimate the parameters of the quadratic model
xt a bt bt2
6. Derive the simultaneous equations a and b for the model xt abt using the
least squares method.
7. Given the following data:
t 1 2 3 4 5 6
x(t) 5 4 7 8 7 10
Predict x(t) for periods 7 and 8 using ve-period moving average and linear
regression model. Which resulting prediction will you recommend for use in
this case and why?
8. Show quantitatively that the estimate you obtain from the basic exponential
smoothing equation is a function of all previous observations; however, the
most recent ones are heavily weighted than the far-distant ones.
9. Use a linear model and exponential smoothing to predict the values of x(t) for
periods 7, 8, and 10 for the data in Exercise 7.
with slope b, the rst
10. Prove that if a set of data have a noise-free ramp
b
exponential smoothing will lag the true value by a b where b 1 a
11. How would you determine optimally in the exponential smoothing model?
12. Consider the following quarterly data for the maintenance load in man-hours
for the last 5 years.
Year Quarter
I II III IV
1 215 120 150 100
2 250 175 75 150
3 300 250 75 165
4 350 275 100 200
5 400 300 125 225
(a) Determine an appropriate seasonal index for each quarter.

(b) Deseasonalize the data and t it to an appropriate growth model.
(c) Predict the quarter values for the 6th and 7th years.
Exercises 55
13. Discuss the advantages and disadvantages of the approaches given in this
chapter for capacity planning.
14. Suppose you were asked to develop a capacity plan for a maintenance
department. Which approach will you select from the approaches given in this
chapter? and why?
15. For the data in Table 2.9, evaluate the cost of the plan given in the table if an
in-house regular hour is 10$, overtime hour is 15$, and a contract man-hour
costs 25. In addition, every backlog man-hour delayed from period t to t + 1
costs 5. Develop an alternative competitive plan in which you can have instead
of 5 skilled employees have a mixture of skill and unskilled. The hour of
unskilled worker is 8$.
16. Develop a linear programming model for the data given in Table 2.9. Assume
that the trades in problem 15 are available.
17. Explain how would you use the machine servicing model for determining
optimal mix of crafts and skills to meet the maintenance load.
18. What are the disadvantages of using linear programming for capacity planning.
19. Locate a factory near your area and study its operations and maintenance.
(a) Forecast their maintenance load.
(b) Use the structured tableau method to determine their maintenance capacity
in terms of staff only.
(c) Use linear and integer programming to plan their maintenance capacity.
20. Apply stochastic simulation to plan the maintenance capacity for the factory in
problem 19. Is there a difference between the approaches in problems 19 and
20. Why do you expect such a difference? Which is more appropriate for this
case and why?
References
1. Al-Turki OM (2011) Strategic maintenance management. J Qualit Maint Eng 18(2):8794

2. Andijani AA, Duffuaa SO (1996) A critical evaluation of simulation study in maintenance. In:
Proceedings of the 1st international conference on industrial engineering, theory and
applications, Houston, Texas, Dec 1996
3. Bedworth DD, Bailey JE (1987) Integrated production systems and contro, 2nd edn. Wiley,
New York
4. Duffuaa SO, Andijani AA (1996) Stochastic simulation: an effective tool for modeling
maintenance systems. A working paper, Department of Systems Engineering, KFUPM,
Dhahran 31261, Saudi Arabia
5. Duffuaa SO, Raouf A (1995) A simulation model for determining maintenance stafng in an
industrial environment. Simulation 59:293299
6. Law AM, Kelton WD (1993) Simulation modeling and analysis. McGraw Hill, New York
7. Montgomery DC, Johnson LA (1976) Forecasting and time series analysis. McGraw Hill,
New York
8. Murthy DN, Atrens P, Eccleston JA (2002) Strategic maintenance management. J Qualit

Maint Eng 8(4):287305
9. Taha H (1992) Operations research: an introduction, 5th edn. Macmillan Publishing Company,
New York
10. Tsang AHC (1998) A strategic approach to managing maintenance performance. J Qualit
Maint Eng 4(2):8794
http://www.springer.com/978-3-319-19802-6

Maintenance Strategic and Capacity Planning

Uploaded by

Copyright:

Available Formats

Maintenance Strategic and Capacity Planning

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Maintenance Strategic and Capacity Planning

Uploaded by

Copyright:

Available Formats

Chapter 2

Maintenance Strategic and Capacity

Springer International Publishing Switzerland 2015 19

2.2 Forecasting Preliminaries

Forecasting techniques can be classied into two approaches: qualitative and

Use historical data and

Test and validate Estimate model

Fig. 2.1 The cycle of the forecasting process

2.3 Qualitative Forecasting Techniques

2.4 Quantitative Forecasting Techniques

2.4.1 Simple Moving Average

Suppose the characteristic under study is generated by a constant process, plus a

The estimator of b, ^b, is given as

The forecasted load for month 7 using n = 3 is

400 500 600

500 600 500

2.4.2 Weighted Moving Average

where wi = weight for the ith actual observation

^x7 0:25400 0:25500 0:5600 525

^x8 0:25500 0:25600 0:5525 537:5

2.4.3 Regression Analysis

If an independent variable exists that can predict a characteristic (dependent vari-

where a, b, and c are parameters to be determined.

Table 2.1 Data for parameter t x(t) tx(t) t2

The resulting equation that can be used for prediction is

If the model is taken to be constant plus random variation, i.e., xt a t , then

The slope of the line is estimated as follows:

Therefore, xt 0:005 11:43t.

2.4.4 Exponential Smoothing

Exponential smoothing is a widely used forecasting method that is simple, efcient,

^xt a xt 1 1 a^xt 1 2:12

xt axt 1 1 a^xt 1 2:14

where is determined using experimentation or judgment. A large value of

where the expected demand at time t is a linear function of time.

If we apply the exponential smoothing again (double smoothing) denoted by

^xt a^xt 1 1 a^^xt 1 2:18

at each period t 1, the values of a and b are updated as follows:

at 1 xt 1 ^xt 1 lag 2bx t 1 b

Initial conditions to start the process are

Table 2.2 The rst and double exponential smoothing

The prediction for any period after L units from 7 is given as

^x7 L ^a7 ^b7l

2.4.5 Seasonal Forecasting

^xt 4:15 2:36t 2:22

Table 2.5 Deseasonalized data

Table 2.6 Weeks 5 and 6 daily forecast with xt Id 4:15 2:36t

Table 2.7 Weeks 5 and 6 daily forecast with xt Id 5:0 2t

x5 4:15 2:365 15:95

^xt 5:0 2t 2:23

2.5 Error Analysis

The sum of the errors

is not a valid measure of effectiveness of a forecasting model, but it is a measure of

2. Mean-squared error (MSE)

3. Mean absolute percent error (MAPE):

100 X xt ^xt

4. Mean-squared percent error (MSPE)

2.6 Forecasting Maintenance Work

Prior to performing capacity planning or designing a new maintenance organiza-

Table 2.8 Weekly backlog Mechanical Electrical

100 X xt ^xt

0:75x11 0:5x21 60 2:30

0:8x12 0:7x22 40 2:31

x11 ; x12 ; x21 ; x22 0

xi2kt 0:25xi1kt 0 2:35

xijkt Uijt 2:36

ni2k 0:25ni1k 0 2:39

nijkt Uijt 2:40

LBR Bk;t UBk 2:41