Shooka Darabi for the degree of Master of Science in Industrial Engineering presented on
July 12, 2013.
Title: Modeling and Analyzing Processes with Infrequent Failures: Implications for
Process Monitoring Policies
variation and eliminating extraneous variation whenever possible to push the target
measures as close to the ideal value as possible. Statistical Processes Control (SPC)
technique is a very efficient statistical technique for monitoring and controlling processes
called g-chart has been developed for controlling process variation where failure is
However, g-chart, EWMA chart and probability based control charts, which are the main
control charts used for monitoring infrequent failure processes fail to detect changes in
the process quickly enough. As a solution it was proposed to inspect and follow-up on
model is used to compare the cost of following up on every failure policy with the cost of
By comparing the total cost per hour of using statistical control chart method with the
total cost per hour of following up on every failure, it is concluded that following up on
every failure policy results in quick detection of changes in the process and it is also a
by
Shooka Darabi
A THESIS
submitted to
in partial fulfillment of
degree of
Master of Science
APPROVED:
I understand that my thesis will become part of the permanent collection of Oregon State
University libraries. My signature below authorizes release of my thesis to any reader
upon request.
First and foremost, I would like to thank my professor and adviser Dr.David Kim for
his valuable guidance and advice. Dr. Kim was helpful, supportive and offered invaluable
assistance and guidance. I also want to express my deepest gratitude to the members of
the supervisory committee, Dr. Sarah Emerson and Dr. Javier Calvo-Amodio. Without
their support and assistance this thesis wouldnt have been successful.
Finally, an honorable mention goes to my family for their love, understanding and
support.
TABLE OF CONTENTS
1Introduction ............................................................................................................................................... 1
2Literature review ....................................................................................................................................... 4
2.1Conventional Shewhart statistical control charts and infrequent failure processes ..................... 4
2.1.1Non-Shewhart statistical methods for infrequent failure process ......................................... 6
2.1.2Statistical charts for infrequent failure processes .................................................................. 7
2.2Non-parametric charts for infrequent failure processes ............................................................... 9
2.2.1Traditional non-parametric charts for infrequent failure processes ...................................... 9
2.3Summary .................................................................................................................................... 12
3Methodology ........................................................................................................................................... 13
3.1Quality control scenario and assumptions.................................................................................. 13
3.2Review of control chart methods ............................................................................................... 15
3.2.1The Shewhart g control chart .............................................................................................. 16
3.2.2g-chart with probability distribution-based control limits................................................... 18
3.2.3Exponential Weighted Moving Average control chart ....................................................... 21
3.2.4Control chart performance measurements........................................................................... 23
3.2.5Inspection and following up every failure........................................................................... 24
3.3Control charts versus follow up of every failure what failure probability? ............................ 25
3.4Quality control cost model ......................................................................................................... 26
3.4.1Parameters ........................................................................................................................... 26
3.4.2Cost components ................................................................................................................. 27
3.4.3Fraction of time a process is in control ............................................................................... 27
3.4.4Total cost per unit time ....................................................................................................... 28
3.4.5 and using control charts ................................................................................... 28
3.4.6Approximation .................................................................................................................... 30
3.4.7 and follow-up each failure ............................................................................ 31
3.5Scenario1- hospital acquired infections ..................................................................................... 32
3.5.1Probability based control chart for hospital acquired infections ......................................... 33
3.5.2Shewhart g-chart for hospital acquired infections............................................................... 43
3.5.3EWMA chart for hospital acquired infections .................................................................... 47
3.5.4Following up on every failure policy .................................................................................. 50
3.6Scenario2- pace maker manufacturing ....................................................................................... 51
3.6.1Probability based control chart for pace maker manufacturing .......................................... 51
3.6.2Shewhart g-chart for pacemaker manufacturing process ................................................... 55
3.6.3EWMA chart for pacemaker manufacturing ....................................................................... 57
3.6.4Following up on every failure ............................................................................................. 58
4Results .................................................................................................................................................... 59
4.1Comparing probability limit based g-chart with follow up of every failure - hospital
acquired infections ...................................................................................................................................... 59
4.2Comparing Shewhart g-chart with follow up of every failure - hospital infections................... 62
4.3Comparing EWMA with follow up of every failure - hospital acquired infections................... 64
4.4Comparing probability limit based Shewhart g-chart with follow up of every failure -
pacemaker manufacturing case ............................................................................................................... 67
4.5Comparing EWMA with following up of every failure pacemaker manufacturing .................. 69
5Conclusion............................................................................................................................................... 71
6References ............................................................................................................................................... 72
LIST OF FIGURES
Figure 1: Example of monitoringthe number of inspections until failure using control charts................. 15
Figure 2 :Comparing total cost per hour of using probabilit based control charts vs. following up
on every failure when d=1 and C/F = 1 ...................................................................................................... 60
Figure 3: Comparing total cost per hour of using probability based control chart vs. following up
on every failure when d = 1 and C/F = 0.4 ................................................................................................. 62
Figure 4: Comparing total cost per hour of using g-chart vs. following up on every failure when
d = 1and C/F =1 .......................................................................................................................................... 63
Figure 5: Comparing total cost per hour of using EWMA vs. following up in every failure when
d = 1 and C/F = 1 ........................................................................................................................................ 64
Figure 6: Comparing total cost per hour of using EWMA vs. following up on every failure when
d = 1 and C/F = 0.05 ................................................................................................................................... 66
Figure 7: Comparing total cost per hour of using EWMA vs. following up on every failure when
d = 1 and C/F = 20 ...................................................................................................................................... 66
Figure 8: Comparing total cost per hour of using probability based control chart vs. following up
on every failure whe d = 1 and C/F = 1 ...................................................................................................... 67
Figure 9: Comparing total cost per hour of using g-chart vs. following up on every failure when
d = 1 and C/F = 0.1 ..................................................................................................................................... 68
Figure 10: Comapring total cost per hour of using EWMA vs. following up on every failure
when d = 1 and C/F = 1............................................................................................................................... 69
Figure 11: Comparing total cost per hour of using EWMA vs. following up on every failure
when d = 1 and C/F = 0.25.......................................................................................................................... 70
Figure 12: Comparing total cost per hour of using probability based control charts vs. following
up on every failure when d = 2 and C/F = 1 ............................................................................................. 110
Figure 13: Comparing total cost per hour of using probability based control chart vs. following
up on every failure when d = 0.5 and C/F = 1 .......................................................................................... 110
Figure 14: Comparing total cost per hour of using g-charts vs. following up on every failure
when d = 2 and C/F = 1............................................................................................................................. 111
Figure 15: Comparing total cost per hour of using g-chart vs. following up on every failure when
d = 0.5 and C/F = 1 ................................................................................................................................... 111
Figure 16: Comparing total cost per hour of using EWMA vs. following up on every failure
when d = 2 and C/F = 1............................................................................................................................. 112
Figure 17: Comparing total cost per hour of using g-chart vs. following up on every failure when
d = 0.5 and C/F = 1 ................................................................................................................................... 112
Figure 18: Comparing total cost per hour of using probability based control charts vs. following
up on every failure when d = 2 and C/F = 1 ............................................................................................. 113
Figure 19: Comparing total cost per hour of using probability based control charts vs. following
up on every failure when d = 0.5 and C/F = 1 .......................................................................................... 113
Figure 20: Comparing total cost per hour of using g-chart vs. following up on every failure when
d = 2 and C/F = 1 ...................................................................................................................................... 114
Figure 21:Comparing total cost per hour of using g-chart vs. following up on every failure when
d = 0.5 and C/F =1 .................................................................................................................................... 114
Figure 22: Comparing total cost per hour of using EWMA vs. following up on every failure
when d = 2 and C/F = 1............................................................................................................................. 115
Figure 23: Comparing total cost per hour of using EWMA vs. following up on every failure
when d = 0.5 and C/F = 1.......................................................................................................................... 115
LIST OF TABLES
Appendix1 ................................................................................................................................................. 75
Appendix2 ................................................................................................................................................. 83
Appendix3 ................................................................................................................................................. 90
Appendix4 ................................................................................................................................................. 98
Appendix5 ............................................................................................................................................... 102
Appendix6 ............................................................................................................................................... 100
Appendix 7-Hospital graphs.................................................................................................................... 110
Appendix8 -Pacemaker manufacturing graphs ....................................................................................... 113
1
1 Introduction
The research presented in this thesis addresses the question of when more
common statistical quality control charts should be used to monitor processes that are
considered to have infrequent failures. Two particular applications that will be used as
examples are the monitoring of the infection rate in a hospital, and the monitoring of
clarify the general features of the quality monitoring scenario considered that are
and either pass or fail the inspection. In the infection rate application, units are patients
and inspections are patient examinations. Failure of the inspection is testing positive for
an infection. In the pacemaker assembly example units are electronic components, and
inspections are actual inspections that are either passed or failed. The probability of a
quality control chart methods can be applied, but some methods have been developed for
Relatively recently a statistical quality control chart called a g-chart has been
developed for monitoring processes similar to the infection and pacemaker assembly
applications. They have been applied to processes with infrequent failures where the
failure rate is usually less than 1%. g-charts have been applied in healthcare, where the
failure rate is infrequent. Some examples are, using g-chart for monitoring the number
needle sticks and surgical site infections (Benneyan 1999). In the g-chart, charts monitor
the numbers of cases (patients) between two successive failures (infections, heart surgery
complications )
Medical processes are a big national concern and various techniques for making
these processes more efficient and consistent have been applied. Hospitals have been
encouraged to use statistical process control tools such as Statistical Process Control
(SPC) charts for improving and monitoring their processes. The objective of SPC is to
Motivated from the use of g-charts for health care applications, the author of this
thesis tried to develop a g-chart for monitoring the failure rate in a pacemaker
manufacturing process which, because of the type of product, also has infrequent
mostly from component failure, and pacemaker failure due to other reasons, like operator
error, machine error and others can be neglected. Since pace makers are very sensitive
products, before final assembly, all the components are tested to minimize pacemaker
failure due to component failures. Because of this, the failure probability of a component
in final assembly is very small. On average, 1 in every 2000 capacitors fails (failure rate
can be placed in the infrequent failure category and quality tools designed for infrequent
The g-chart that was developed for monitoring pacemaker manufacturing process
failed to capture changes in the final assembly process. This was concluded by testing the
3
chart with historical data. Based on this result, the author became motivated to study
why a chart that is designed specifically for monitoring infrequent failures doesnt
failure processes, what are other methods that can be used in this type of situations?
The first section of this thesis is the literature review, presenting relevant research
The second section describes the methodology. It includes description of a cost model
for the long run cost of monitoring a process and applications of this model to two
specific scenarios. The third section summarizes sensitivity analyses and presents more
general results as well as the specific analyses performed and results obtained. Finally,
2 Literature review
starts with a short summary of statistical process control methods. Later the use of the
statistical process control methods and their short coming in infrequent failure processes
will be addressed. The section ends with presenting some other alternative methods that
have been developed for monitoring infrequent failure processes for both parametric and
non-parametric data.
processes
environment. This means processes are subject to variation. Some of this variation
originates from natural causes that exist in every type of environment. These types of
variation usually have been referred to as chance causes and are present to some degree
in every process. Beside chance causes, there are other sources of variation that occur due
assignable causes and eliminating assignable causes are the objective of Statistical
The basic methodology of control charts consist of sampling from a process over
time and monitoring a process measurement over time. Controls charts have a center line
and an upper and lower control limit established from process measurements. When a
measurement falls within the upper and lower control limit, no action is required, but
5
whenever the measurement falls outside of the control limits, a possible assignable cause
is deemed possible and further investigation to determine the root cause and correct if
Count charts are SPC charts that are used for situations when the monitored
defects in an item. The p-chart and the c-chart are the two most well-known types of
count charts. The p-chart is used for monitoring the ratio of the number of
nonconformities to the sample size. The c-chart is used to monitor total number of
nonconformities per unit (Montgomery, 2007). These charts have a broad range of
these charts underperform in an environment where the failure occurrence is low. In this
case, failure counts in many fixed time intervals will be zero leading to narrow control
processes is the symmetrical shape of the control charts. The 3-sigma control limits are
derived from the original Shewhart control chart which assumes the monitored data is
normally distributed or is asymptotically normal. When using p-charts and c-charts for
monitoring frequent failure processes Binomial distribution and Poisson distributions are
approximately normal and therefore having symmetrical control charts with 3- sigma
control limits is well justified. However in an infrequent failure process, the normality
In addition, using count charts for situations where the failure rate is low results in
a negative lower control limit. Since a negative lower control limit is meaningless, it is
6
downward shifts of the process which is an indication of improvement (Chan et al. 2007).
A low defect rate also negatively impacts the upper control limit. A rule of thumb
(Chan et al. 2000) says, in a p-chart the upper control limit should be smaller than the
inverse value of the sample size . In infrequent failure process this means the UCL
would be set so low that every defect exceeds the upper control limit and triggers a
signal. In this case the false alarm probability is much greater than the desired value of
In another study, Goh 1987 also confirms that even for very high yield process
with a defect rate lower than one percent there is poor low performance of the count
charts.
The inability of common Shewhart control charts for detecting out of control
processes in environments with a low failure rate has resulted in the development of other
methods and techniques for monitoring infrequent failure processes. In this section some
other control chart techniques for low failure processes are reviewed.
the basis of 3-sigma control charts in invalid, some studies suggest using control limits
derived from the distribution of the statistic monitored. Goh (2007) explains the method
used to identify control limits derived from the Poisson probability distribution for the
probability distribution based limits for monitoring low failure processes, when the
monitored statistics has a geometric distribution. In this method control limits are
In another study, Nelson (1994) suggested counting the total number of the
defects (denoted c) in a sample with size n and develop control charts for the
transformed variable
. Although, it has been demonstrated that this method works well
for infrequent event processes, a disadvantage with this method is that the interpretation
Different methods and techniques presented above, both Shewhart and non-
Shewhart techniques, all showed limitations when it comes to monitoring processes with
infrequent failures. To overcome this problem some specific control charts have been
developed specifically for this purpose. In this section a brief explanation of methods
Since the conventional charts like p, np, u and c chart often result in subgroups
being plotted too infrequently in an infrequent failure process. An alternative chart was
designed by Benneyan (1999; 2001) mainly for the purpose of monitoring failures in an
infrequent failure environment. The charts developed are the g-chart and h-chart. These
charts monitor the number of events between occurrences of two successive failures.
These occurrences can be the number of conforming items between two consequence
failures, or days between two failures. In the g-chart, the monitored statistics follows a
8
geometric distribution. In the h-chart, the monitored statistic is the number of conforming
items until the n-th failure, which also follows a geometric distribution (Benneyan 1999;
Benneyan 2001; Benneyan 2001; Benneyan and Kaminsky 2004; Glushkovsky 1994).
For infrequent failure processes, studies show that these charts can be applied to
both the first phase of the process, to bring the process into statistical control, as well as
the second phase of the process which is to monitor the in control process (Benneyan
2001).
An alternative SPC chart for infrequent failures when the failures occur
number of units until the first defect is observed and it follows an exponential
distribution. Based on the information above the control limits have been defined as,
ln
Q 2
1
ln 2
Q
ln 1 2
Q
Although the CQC chart has been showed to perform well when the failure rate
of processes is low or moderate, there is another version of the CQC chart called
CQC(2), which has the virtue of being more sensitive to smaller shifts and makes it more
nonconformities follows a Poisson process, but the statistic that is being monitored in this
9
case is the time until the second failure occurs. This statistic follows a gamma
The alternative Shewhart-like control charts that are designed specifically for
presented there are some assumptions about the distribution of the data, and charts are not
robust to these assumption. In the following chapter some non-parametric charts that
dont make any assumptions about the distribution of data are presented for infrequent
failure processes.
In this section the traditional non-parametric charts and their specifications and
Non-parametric charts which are also known as distribution free charts can be
when the distribution of the data is unknown or doesnt follow a known distribution.
Studies show that these charts are very effective in detecting shifts for non-normal
distributions, especially those with heavy tails (Chakraborti et al. 2004). Furthermore,
these charts have the advantage of having the same in-control run length for all
continuous distributions as well as being more robust to outliers (Chakraborti et al. 2004).
When failures are infrequent and the data doesnt follow a particular distribution
non-parametric charts can be applied to the situation, using the number/time of events
10
between defective units as a statistic. The two well-known non-parametric charts are the
exponential weighted moving average EWMA and the cumulative sum chart CUSUM.
These charts are very effective in detecting minor and persistent shifts (Montgomery
2007). EWMA and CUSUM control charts can be used for monitoring either attribute or
variable type data using the entire data history. EWMA charts are developed based on
CUSUM charts a sequential technique method is used which keep tracks of the
Although many studies including Montgomery (2001, p.433) have pointed out
other studies give evidence against that. For instance a study conducted by Graham et. al.
(2012) shows that, for EWMA charts it is usually assumed that the underlying
normal (Graham et. al. 2012). In addition Human et al (2011) shows that EWMA is not
robust to non-normal distributions and would incorrectly show the process is out of
control. Borror and Montgomery (1999), also show that the statement about EWMA
being a nonparametric procedure is true when the process parameters are known or
specified. When the parameters are unknown and need to be estimated from sample of
are estimated from the sample, this will result in many more false alarms in the Average
parametric charts are not effective in detecting shifts due to the violation of the
11
assumptions made about the data distribution and sees the CUSUM chart as a better
choice. However, sensitivity of Average Run Length (ARL) show significant differences
between the charts from normally distributed data and those that are based on non-normal
data.
Graham et al. (2011 a,b) proposed a nonparametric EWMA chart which is based
on a sign and rank sum test when the median is known or specified. Although this
method is non-parametric, when the median is unknown and estimated from data, the
Further analysis indicates that the EWMA-EX chart performs as good or better
than parametric CUSUM and EWMA charts for detecting small mean shifts . Even in
comparison with other recently developed nonparametric CUSUM or EWMA chart like
better. These points as are proved by comparing the ARL charts of the different methods
developed based on the exceedance statistic called exceedance CUSUM chart. This new
non-parametric CUSUM chart like the EWMA_EX chart is more effective method in
Further analysis show that these charts are more powerful in detecting shifts in the
process than traditional CUSUM and rank-based CUSUM charts. It is showed that these
12
charts perform very well in cases when the underlying distribution is heavy-tailed and
right-skewed.
2.3 Summary
In this section a brief summary of earlier literature was presented. The main
objective was to describe methods that were developed for use in an infrequent failure
environment, both for monitoring statistics that follows a specific distribution and when
the monitored statistic does not follow any particular distribution. No research was found
that examines when some type of statistical quality control chart should be used to
monitor a process with infrequent failures. Not using statistical control charts to monitor
a process, and investigating every failure is a reasonable alternative when failures are
infrequent and the cost of running a process out of control is high. The research in this
In the next section, a general mathematical model for the long run cost of
3 Methodology
The question that is addressed in this research is: What is an infrequent failure
question can be asked as: At what failure probability will statistical quality control result
To answer this question a mathematical model for the long-run cost of a quality
control policy is developed. This model is then used to compare the policy of following
up on every failure versus using some statistical control chart method and following up
the specific quality control scenario being addressed. This is followed by a review of the
different control chart methods that will be analyzed. The details of the cost model for a
In the scenario considered the rational sub-group is equal to one; meaning a data
sample of size one is collected and monitored over time. The particular data item
represents the number of units inspected until a failure occurs. In this scenario a unit
may represent a variety of entities depending on the specific application. For example a
unit may be a product, a patient, etc. It is assumed that there are individual inspections on
each unit and a unit either passes the inspection, or fails to meet the predefined
the application. For example if the unit is a product, a failure means the product did not
14
meet specific product requirements, and if the unit is a patient then failure may mean the
patient suffered an infection. The monitored statistic is the number of units between
failures and this statistic can be converted to time between failures if the inspection time
It will be assumed that each unit has the same probability of passing an
inspection, and that this probability remains fixed between changes at discrete time
points. These changes normally represent a shift to an out-of-control state, or the return
a decrease in process quality. Under the scenario described the monitored statistic (units
If a quality control chart is being utilized to monitor the number of units between
failures the x-axis represents the failure number, and the y-axis is the number of
Figure 1: Example of monitoringthe number of inspections until failure using control charts
When using a control chart lower and upper control limits are computed and set
according to the control chart assumptions. Every point (failure) that plots above the
upper control limit or below the lower control limit generates a signal. For the situation
described, when a point plots above the upper control limit the quality of the process may
have improved, and when a point plots below the lower control limit it is an indication
that process quality may be worse. Note that in cases where the LCL is equal to zero, the
control chart loses its ability to detect increases in the failure probability.
results in potentially catastrophic (or very costly) results. By nature the products being
monitored are very sensitive and even one failure is not acceptable.
16
sticks, and other patient acquired outcomes (Benneyan 1999). Some examples of
Monitoring processes using standard charts like (p, np, u and c chart) results in
subgroups being plotted too infrequently for timely control chart feedback (Benneyan
2001). Therefore, these control chart methods are not informative and are not helpful in
In the following sections, three control charts that were developed for the scenario
described, and which have been applied detect process changes in infrequent failure
process are introduced. These charts are the Shewhart g-chart, the g-chart with
probability based control chart limits, and the Exponential Weighted Moving Average
A control chart is developed for when the number of trials between events is
monitored, and the distribution of the number of trials between events is geometrically
distributed is called a g-chart. Each trial is an independent Bernoulli trial, and the number
of trials until the first event (failure), Z, follows a geometric distribution with probability
mass function:
'() # 1, 2, 3,
(1)
17
1
-!"$
%
(2)
1%
./)!"$
%0
(3)
1
%
"
(4)
(for example the average number of failed capacitors or average days between
infections). The g-chart was developed specifically for a process similar to the scenario
being addressed.
The parameter k represents the number of standard deviations from the centerline
used to define upper and lower control limits. Usually k is three; meaning control limits
are 3 sigma from the center line. The lower control limit (LCL), and upper control limit
(UCL) are:
;
1 1%
234 56 7 8 #9 0 :
% %
(5)
<
1 1%
434 56 7 #9 0 :
(6)
% %
For these limits the actual type I error may be less than stated due to the discrete
nature of the geometric distribution. Alternative control limits where the type I error level
<
1 1%
234 56 7 8 #9 0 :
(7)
% %
;
1 1%
(8)
434 56 7 #9 0 :
% %
p Known p Estimated
;
1 1% " 8 #=" !" 1$;
Upper Control Limit
56 7 8 #9 0 :
(UCL) % %
1 "
%
Center Line (CL)
<
1 1% " #=" !" 1$<
56 7 #9 0 :
Lower Control Limit
(LCL) % %
Table 3.1 summarizes the computation of the Shewhart upper control limit
(UCL), center line (CL) and lower control limit (LCL) for a g control chart used to
An alternative that is helpful for situations where Shewhart control charts limits
(particularly the LCL) would be well below zero (and thus set to zero) is to derive the
control limits directly from the assumed distribution of the data. This method can be
applied to almost any control chart but it is especially helpful when the distribution of the
19
monitored statistics is not normal or symmetric as is the case when the monitored statistic
is from a geometric distribution with a small p value. Note that this method can be
applied to any other distribution with known Cumulative Distribution Function (CDF).
Assume >?@? and >A@? are the desired type I error rates (false alarm rate) for the
lower control limit and upper control limit respectively. In most cases it is assumed
>?@? >A@?
BCDEFG
0
.
'() # 1,2,3,
(9)
where %I is the in-control probability that a unit will fail inspection. The
probability of having a random variable Z smaller or equal to the LCL when the process
By using the geometric CDF the probability that a random variable plots at or
beyond the upper control limit when the process is in control is computed as:
By setting the above results equal to the desired type I error rate result in:
ln!>A@? $ ;
234
ln!1 %I $
(14)
For these limits the actual type I error is less than stated due to the discrete nature
of the geometric distribution. Alternative control limits where the type I error level is
ln!>A@? $ <
234
ln!1 %I $
(16)
PQ!1 >?@? $ ;
434
PQ!1 %I $
(17)
Similarly the center line can be computed by setting the probability of a random
PQ0.5
34
PQ!1 8 %I $
(19)
ln!>A@? $ ;
234
ln!1 %I $
Upper Control Limit (UCL)
21
PQ0.5
34
PQ!1 8 %I $
Control Limit (CL)
Table 3.2: Probability distribution based control limits for the known rate of
alternative control chart method often used for in-control processes. EWMA charts are
useful for detecting small mean shifts and are applicable to situations where the rational
subgroup size is one. The other advantage with EWMA charts is that unlike other
Shewhart control charts; EWMA charts are insensitive to the normality assumption and
thus perform well when the sample size is equal to one. However recent research showed
that robustness to the normality assumption is true under the condition that distribution
parameters are known (Borror, et al. 1999). The EWMA statistic that is monitored is
defined as:
(0 M X M 1). KW is the current observation. The starting value VI for sample size Y 1 is
defined as:
VI ZI
X
234 ZI 8 4[9! $ 1 !1 X$0W
2X
(21)
X
434 ZI 4[9! $ 1 !1 X$0W
(22)
2X
approaches to one. Thus, when running a process for a long time the control limits
approach a steady limit. Due to discrete nature of monitoring statistics here are two
X
234 ZI 8 4[96 7 ;
2X
(23)
X
434 ZI 4[9!
(24)
$ <
2X
For these limits the actual type I error is less than stated due to the discrete nature
of the geometric distribution. Alternative control limits where the type I error level is
X
434 ZI 4[9! $ ;
2X
(25)
stage
X X
ZI 8 4[ 9! $ 1 !1 X$0W ; ZI 8 4[9!
Upper Control
2X $ ;
Limit (UCL) 2X
X X
ZI 4[ 9! $ 1 !1 X$0W < ZI 4[9!
Lower Control
2X $ <
Limit (LCL) 2X
Control chart performance can be evaluated based on how frequent false alarms
are generated when the process is in control, and how frequent an alarm is signaled when
the process is out of control. This performance is measured by Average Run Length
(ARL). ARL0 is a measure of the average time between false alarms when in control.
ARL1 is a measure of the average time between alarms when out of control.
1
]^4
%)(_/_YP`a (b cYdQ/P
(27)
24
1
]^4I
1 !434 M " M 234 | % %I $
(28)
1
]^4
1 !434 M " M 234 | % f %I $
(29)
The Average Time to Signal (ATS) is the average time until an out of control
In more common Shewhart control charts, plotting beyond control limits generate
signals with different meanings. Signals above the UCL normally are an indication of a
decrease in the quality, and signals below the LCL normally indicate an improvement in
process quality. For g-charts this is the opposite, since the monitored statistic is the
number of units until a failure occurs. A lower statistic means the decreased quality.
For the cases when the LCL is equal to zero, the g-chart loses the ability to detect
To overcome the problem of zero LCL, and improve the power of detection of
processes changes, several design considerations have been proposed. Some of these
rules(Benneyan 2001).
conduct follow-up after every failure to determine if a process shift has occurred. This
25
can be a useful approach if failures occur infrequently and the costs of increased failures
are very high. This is the situation for the production of pacemakers.
probability?
The different control charts reviewed have been proposed for use in the situations
described in section 3.2. Regardless of what control chart that is being used, when the
LCL is equal to zero, the charts lose their ability to detect increase in the failure
probability and may not be very useful and informative. This occurs when the failure
failure. This method is not practical and not a cost-efficient approach when the failure
important question is: what failure probability defines infrequent failures, or what is the
cut-off between infrequent failures and frequent failures? When is it better to inspect and
follow-up every failure, and when should control charts be used for monitoring
processes? Which one of these methods is more beneficial from a cost perspective?
To answer these questions a cost model is developed that can be used to identify
infrequent failures for specific scenarios. This model is presented in the next section.
26
3.4.1 Parameters
The model developed generates a cost per unit of time for following a specific
quality control policy. For simplicity and clarity hours will be used as the time unit. The
overall cost per hour is comprised of two components. The first component is the cost per
hour of signal follow-up, when the process is in control (Cost IC). The second component
is the cost per hour of operating a process when the probability of failure has shifted from
%I to %I 8 k%I , plus the cost of follow-up (Cost OC). The total cost per hour is a
weighted average of the two cost components and is computed based on the percentage of
being in control (% IC), and the percentage of being out of control (% OC). The overall
The long-run average in control cost per hour is the cost of signal follow-up when
the probability of a failure is %I . This cost per hour is the signal follow-up cost divided by
the average time between signals when the process is in control. The average time
between signals when the process is in control is ]ghI and 3(c` n3 is computed as:
'
3(c` n3
]ghI
(31)
The out of control cost per hour is the cost of signal follow-up when the
probability of failure has shifted from %I to %I 8 k%I and process is out of control (OC).
This cost is the cost per hour of running an out-of-control process plus the signal follow-
up cost divided by the expected time to signal after the process is out of control.
'
3(c` l3 3 8
-o
(33)
Y is the time the process stays in an out of control state. The calculation of -o
A process can be in one of the two states: in control, or out of control. Let
switches to out of control and stays out of control for time o ; then the process is brought
back to an in-control state and remains in control for time j0; then it goes out of control
If the sequence of pjq r and sequence of poq r for Q L 1, are sequences of X and Y
and are independent and identically distributed, then the long-run fraction of time the
-j
% n3
-j 8 -o
(34)
Similarly
-o
% l3
-j 8 -o
(35)
The total cost per hour for a quality control policy is,
'
(36)
8 $
-o
The values of ]ghI and -o are determined by the quality control policy being
adopted.
1
]^4I ! 1$ -"|434 M " M 234
!cYdQ/P|% %I$
(37)
8 -"|" M 434 () " s 234
-"|434 M " M 234 is the expected value of " given all the values of " (the
Wu?@?;
!" Y$
!" Y|434 M " M 234$
!434 M " M 234$
(39)
-"|" M 434 () " s 234 is the expected value of " given the values of " (the
number of inspections until a failure) are out of control and is computed as following.
-"|" H 434
!" H 234$
!" H 434$ 8 !" L 234$
8 -"|" L 234
!" L 234$
!" H 434$ 8 !" L 234$
30
-o is the average time a process remains out of control, and equals the average
time until a signal is generated when the process is out of control. -o is computed by
multiplying the average time between inspections by the average number of inspections
1
]^4 ! 1$ -"|434
!hYdQ/P|%I %I 8 k%I $
M"
(43)
3.4.6 Approximation
as 0.98 of w . As an example compare -"|434 M " M 234s exact value and their
x
For %I = 0.01 when UCL =658 and LCL=0 , -"|0 M " M 658 = 99.108 and the
For pI = 0.02 when UCL =658 and LCL=0 , EZ|0 M " M 658 = 49.998 and the
For %I = 0.05 when UCL =129 and LCL=0 , -"|0 M " M 129 = 190.42 and the
1
]^4 ! 1$ -"|434 M " M 234 8 -"|" M 434 () " s 234
!cYdQ/P$
Although the value of -"|" M 434 () " s 234 is usually greater than
wx
, the
difference made by including this value in the computation of ]^4 is very small and the
For example for %I 0.01, when UCL =658 and LCL= 0,the exact value of the
in control cost using a probability based control charts is $0.054175 and the approximate
value is $0.055355. For the same example with an out-of-control probability of0.02 the
exact value of the out of control cost is $1000 and the approximate value is also $1000.
The exact total cost is $999.978 and the approximated cost is $999.977.
If %I 0.0005, and the UCL =13212 and LCL= 2,the exact value of the in
control cost using probability based control charts is $0.006626 and the approximate
value is $0.023. For the same example with an out-of-control probability of 0.001 the
exact value of out of control cost is $1000.04 and the approximate value is $1000.041.
The approximated total cost is $997.143 and the exact value is $999.978.
The tables presented in the appendices have used the approximations above.
In this quality control policy each failure constitutes a signal, and the average time
to signal when a process is in control = ]ghI ` ]^4I and ]^4I w . ]^4I equals
x
When the process is out of control the failure probability has shifted from %I to
%I 8 k%I . so -o
wx ;wx
.
which control charts are more costly than follow up of each failure, the model can
be used to identify this value (which will be denoted U ). By comparing the costs
of two quality control policies as a function of the failure probability, the model
In the next section the three control charts are compared to the policy of
following up every failure for two scenarios. These are the Shewhart g-chart, g-
chart with probability based control limits, and the EWMA chart assuming a
process. Three different control charts are used for monitoring the processes and
probability.
infections in the United States in 2002 was estimated at 10% ( Klevens, Monina et al
33
2007). In this scenario every patient examined is considered an inspection and every
patient that acquires a hospital related infection is considered a failure. The statistic that
infection (the number of inspections until failure). Under the assumptions presented
earlier the number of patient examinations until a failure follows a geometric distribution,
and the sub-group size, which is the number data points used to generate the monitored
Since the infection rate was estimated at 10%, failure rates from 1% to10% were
considered and total cost per hour was estimated for this range.
In this section probability based g-chart control limits are examined for
Typical 3-sigma control limits are supposed to be values such that with
probability of 0.9973, the monitored data will fall between the control limits if the system
is in control. This means the equivalent cumulative probabilities are 0.00135 for the LCL
(>?@? $ and .99865 for the UCL (>A@? $. In this case the desired type I error is 0.0027.
Using (6) and (7) the values of LCL and UCL can be computed.
For example, for %I 0.01 and the desired type I error is 0.0027, LCL = 0 and
UCL = 658. For these limits the actual type I error is less than stated due to the discrete
nature of the geometric distribution. Alternative control limits where the type I error level
The UCL in this case is chosen as 658, can be interpreted as the case when the
analysis is required. Similarly, when a point plots beyond the LCL the signal indicates a
possible reduction in the quality of the process. In this specific example since the LCL
value is 0, and a negative number of inspected patients until an infected patient is not
possible, the chart loses its ability to detect a reduction in the quality.
The probability of not getting a signal when the system is in control is the
probability that a point falls between the LCL and UCL when the probability of a failure
is %I . Using formulas (10) and (11) when %I is equal to 0.01 the probability a failure does
1 !1 0.01$ 1 !1 0.01$I
0.99864
Note that in this example control limits are chosen in a way that type I error is less
than stated.
!hYdQ/P|% %I $ 1 !( hYdQ/P|% %I $
This value can be interpreted as the probability the number of inspections from
the prior infected patient triggers a signal with a value that plots above the UCL or below
1
])/d Q_) (b b/YP)c `( cYdQ/P
!cYdQ/P|% %I $
So in the example when %I = 0.01 the average number of infected patients until a
1
])/d Q_) (b YQb
`k %/`YQ`c Q`YP / cYdQ/P 6 7 737.35585
0.00135
]^4I is the average number of examined patients until an infected patient trigger
a signal.
1
]^4I ! 1$ -"|434 M " M 234
!cYdQ/P|% %I $
between infections given that this number is within the control limits and -"|" M
434 () " s 234 is the expected value of " given the values of " (the number of
A@?<
!" Y$
!" Y|434 M " M 234$
!434 M " M 234$
Wu?@?; Y %I !1 %I $
-"|434 M " M 234=A@?< /[1 !1 pI $< !1
W<
!1 pI $
and
This value can be interpreted as the number of inspected patients until an infected
In the section above, it was assumed the system is in control, the probability of a
failure is %I and all the formulas were computed accordingly. In this part of the example
it is assumed there is a shift in the failure probability from %I to %I 8 k%I . In this case k
represents the percentage of the shift. In the example the failure probability shifts from
%I 0.01 to % 0.02. In this case the probability of having an infected patient has
When the system is in control and has reached a steady state the UCL and LCL is
set and doesnt change afterwards. So, when the system is out of control, the UCL and
LCL remains the same as when the system was in control. So even in this case LCL is 0
and UCL is 658. Since the LCL is equal to zero, the control chart doesnt have the power
1 !1 0.02$ 1 !1 0.02$I
0.99999828
So the Probability of having a signal when the system is out of control is:
occurring until a point plots beyond control limits and triggers a signal is computed as:
1
])/d Q_) (b b/YP)c `( cYdQ/P
!cYdQ/P|% %I 8 k%I $
So in the example when the shifted probability is equal to 0.02 the average
1
])/d Q_) (b YQb
`k %/`YQ`c Q`YP / cYdQ/P 6 7
0.00000172
581390.9538
The ]^4 is the average number of examined patients until an infected patient
1
]^4 ! 1$ -"|434 M "
!hYdQ/P|%I %I 8 k%I$
infections given that this number is within the control limits. This value can be computed
triggered. In this example it means when the probability of having infection is doubled it
takes on average about 28488157 inspected patients until an infected patient triggers a
signal.
3.5.1.3 Cost model for hospital acquired infections using probability distribution
based control limits
Let 3 be the cost per hour of running an out of control process. Montgomery
2007, categorizes these kinds of costs as external failure costs. External failure costs are
the costs that arise when the quality of the product doesnt meet the requirements after it
adjustment, returned product, warranty charges, liability costs and other indirect costs. In
the hospital acquired infection case, these costs range from additional health care cost to
the death of a patient due to infections. Some statistics indicate the national costs of
nosocomial infections cause 8.7 million additional hospital days and death rates range
In this example it is assumed the cost per hour of running the hospital with an
Let ' be the cost of following up on a signal. These types of costs are categorized
as internal failure costs. The important sub-categories of internal failure costs are scrap,
rework, retest, failure analysis, downtime, yield losses and downgrading (Montgomery
2007)
healthcare costs for infected patients and additional costs for searching locating and
eliminating the source of infections. In this example it is assumed the total cost per follow
up on a signal is $1000.
-jis the expected time that a process is in control after control is established. In
the hospital example it is assumed on average a process goes out of control after one
month. If it is assumed every month consist of 160 working hours, -j = 160 hours.
` is the average time between inspections. In the hospital example this can be
viewed as the arrival rate of patients. In this example it is assumed patients arrive every
15 minutes.
Parameter Value
($/hour)
($/Signal)
(hours/unit)
established
Shift Fraction 1
The total cost per hour of using a quality control policy was modeled as,
When the system is in control the in control cost per hour is:
'
3(c` n3
]ghI
'
3(c` n3
]^4I `
The ]^4I was computed as 72260.88 so the in control cost per hour is:
1000
3(c` n3 0.054175
73834.7 0.25
42
This cost is the total cost per hour of using a g-chart with probability-based
control limits for monitoring a process when the process is in control. The low cost is
driven by the large ]^4I . If a probability based quality control chart is used for
monitoring the hospital acquired infection process, when the process is in control and the
failure probability is %I =0.01, on average a signal occurs after every 72261 patients are
inspected.
When a shift occurs and the probability of having a failure shifts from %I to
%I 8 k%I the cost of using probability based control chart for monitoring the process is:
'
3(c` l3 3 8
-o
-o is the expected time until an out of control process is brought back in to
control by using process control method and the value is computed as:
-o ` ]^4
In the hospital example when using probability based control chart the cost of
1000
3(c` l3 1000 8 1000
7267273
This cost is primarily driven by C. The follow-up cost per hour used to bring the
process back into control is low due to a high value of ]^4 . When the probability of
having an infection has increased from 0.01 to 0.02, it takes on average about 7122039
signal. In this case since the LCL value is 0, the control chart doesnt have the power to
-j
% n3
-j 8 -o
-o
% l3
-j 8 -o
In this example the in control percentage using probability based control charts is
computed as:
160
% n3 0.00002016
160 8 7267273
And the out of control percentage using probability based control charts is:
7267273
% l3 0.999978
160 8 7267273
The total cost per hour of using a g-chart with probability distribution-based
control limits for monitoring hospital acquired infections (probability of infection = 0.01,
and d =1 ) is
The cost per hour with %I The cost per hour for %I ranging from 0.01 to 0.1 is
presented in appendix1.
The Shewhart g-chart is the second method analyzed for monitoring hospital
acquired infections. The g-chart is a control chart designed specifically for processes
Using equations (5) and (6) the g-chart UCL and LCL, for hospital example can
be computed as:
1 1 1
434 39 6 17< 198
0.01 0.01 0.01
In this case since the LCL is negative it is replaced with 0. The calculations for
the cost model follow the same steps as the prior section, but with different control limits.
1 !1 0.01$ 1 !1 0.01$I
0.9815
!hYdQ/P|% %I $ 1 !( hYdQ/P|% %I $
1
])/d Q_) (b b/YP)c !YQb
`k %/`YQ`c$ `( cYdQ/P
!hYdQ/P|% %I$
1
6 7 54.05399
0.0185
This value can be interpreted as the number of inspected patients until an infected
398 | % 0.02$
1 !1 0.02$ 1 !1 0.02$I
0.99967
1 1
6 7 3042.62164
!cYdQ/P|% %I 8 k%I $ 0.00033
triggered. In this example it means when the probability of having infection is doubled it
takes about 149089 examined patients until an infected patient trigger a signal.
1000
3(c` n3 0.737
5427.93 0.25
This cost is the total cost per hour of using g-chart for monitoring process when
When the process is out of control the cost per hour of using Shewhart g-control
1000
3(c` l3 1000 8 1000.26
38051.1
Using equation (34) in this example the in control percentage using g-chart is
computed as:
160
% n3 0.0042
160 8 38051.1
38051.1
% l3 0.996
160 8 38051.1
Using information above the total cost per hour using g-chart for monitoring
hospital acquired infections when the probability of having an infection is 0.01, and the
The same procedure is repeated for %I for a range of 0.01 to 0.1 and the results
EWMA is the third method used for monitoring hospital infection processes.
1 1 0.01 0.2
234 8 39 9! $ ; 199
0.01 !0.01$ 0 2 0.2
1 1 0.01 0.2
434 39 9! $ < 1
0.01 !0.01$ 0 2 0.2
In this case the UCL is 199, which means if the number of examined patients until
possible improvement in the quality of the process. However, to make a solid conclusion
further analysis is required. In this example the LCL is one, which is still very low.
1 !1 0.01$ 1 !1 0.01$
0.85329
!hYdQ/P|% %I $ 1 !( hYdQ/P|% %I $
1
])/d Q_) (b b/YP)c!YQb
`k %/`YQ`c$ `( cYdQ/P
!hYdQ/P|% %I$
1
6.81663
0.1467
]^4I 692.53
This value can be interpreted as the number of inspected patients until an infected
1 !1 0.02$ 1 !1 0.02$
0.96169
1 1
6 7 26.1007
!cYdQ/P|% %I 8 k%I $ 0.03831
]^4 1354.920267
triggered. In this example it means when the probability of having infection is doubled, it
takes about 1279 examined patients until an infected patient trigger a signal.
49
1000
3(c` n3 5.78
692.53 0.25
This cost is the total cost per hour of using EWMA chart for monitoring in control
process.
When the process is out of control the cost per hour of using EWMA chart for
1000
3(c` l3 1000 8 1002.95
338.73
Using equation (34) in this example the in control percentage using g-chart is
computed as:
160
% n3 0.321
160 8 338.7301
338.7301
% l3 0.6797
160 8 338.7301
Using information above the total cost per hour using EWMA for monitoring
hospital acquired infections when the probability of having an infection is 0.01, and the
The same procedure is repeated for %I from a range of 0.01 to 0.1 and the net
1
3.5.4.1 When process is in control:
1
]^4I
%I
3(c` n3 40
III
x 0
12.5
I.0
I.I0
' 1000
3(c` l3 3 8 1000 8 1080
-o 12.5
carefully monitored assembly process. Pacemaker failures are mostly from specific
component failures. Since pace makers are very critical products, , all the components are
tested before final assembly to minimize the chance of pacemaker failure due to
assembly is very small. In order to simplify the terminology, each component will be
referred to as a unit.
Units are inspect individually so the number of successful units until a failure
follows a geometric distribution. The statistic being monitored is the number of good
The author of this thesis worked as an intern in a pacemaker company, and the
data being used is representative of this and similar production environments. The
average rate of unit failure in production is 1 in every 2000 units, which means the
Since the failure probability was estimated at 0.0005, a range of failures from
0.0005 to 0.005 was considered and total cost was estimated for this range.
In this section a g-chart with probability based limits is developed for monitoring
1
])/d Q_) (b b/YP)c!kb
`Y QY`c$ Q`YP cYdQ/P 6 7
0.00235
425.45524
]^4I 3018322
This value can be interpreted as the number of inspected units until a defective
1
])/d Q_) (b !b/YP)c$kb
`Yc Q`YP cYdQ/P 6 7 499.79551
0.002
]^4 500311.957
triggered. In this example it means when the probability of having a defective unit is
doubled it takes about 500312 defective units until a defective unit triggers a signal.
53
3.6.1.3 Cost model for pacemaker manufacturing using probability based control
chart
Let 3 be the cost per hour of running an out of control process. In the pacemaker
manufacturing example, the cost of running a process with a shift can be ranged from
scrapping and reworking more pacemakers, which can cost about $10,000 per pacemaker,
to delivering pacemakers that are defective or a providing a pacemaker with a life time
shorter than the specified life time and potentially causing death.
In this scenario it is assumed the total cost per hour of a running pacemaker
Let, ' be the cost of following up on a signal. This cost includes, the cost of
searching and testing for the root cause of changes plus the downtime, if the production is
stopped during this period. Additionally, this cost includes the cost of eliminating the
signal is $1000.
Let, -j be the expected time that a process is in control after in control is
goes out of control every two months. If it is assumed every month consists of 160 hours,
then on average after 320 hours the process goes out of control.
54
Let ` be the average time between inspections. In this example, since every unit is
inspected, 1/t is the rate of assembling units. On average every 3 minutes, a unit is
The summary of values used in the example is presented in the table below.
established
Shift Fraction 1
Using the information in table 3.5 and computed values of ]^4I the in
3(c` n3 0.006626
Similarly the out of control cost can be computed using equation (23).
3(c` l3 1000.04
Using equations (34) and (35) in control and out of control percentages can be
computed.
55
% n3 0.012
And the out of control percentage using probability based control charts is:
% l3 0.98
Using information above the total cost per hour, using probability based control
chart for monitoring defective units in pacemaker manufacturing process, when the
computed as:
The same procedure is repeated for %I from a range of 0.0005 to 0.005 and the
net cost of using probability based control chart is estimated. The table is presented in
appendix4.
pacemaker manufacturing process. The UCL and LCL is 7998 and -3998 respectively. In
this case the LCL is a negative value and since negative LCL is meaningless in this
1
])/d Q_) (b b/YP)c !kb
`Y QY`c$ Q`YP cYdQ/P 6 7
0.01832
54.57086
]^4I 363953
The ]^4I is the number of inspected units until a defective unit trigger signal.
1
])/d Q_) (b b/YP)c!kb
`Y QY`c$ Q`YP cYdQ/P 6 7
0.00034
2983.94392
]^4 2984910.11
3(c` n3 0.055
3(c` l3 1000.007
% n3 0.0021395% l3 0.997
Using information above the total cost per hour using g-chart for monitoring
The same procedure is repeated for %I from a range of 0.0005 to 0.005 and the
net cost of using g control chart for monitoring the system is estimated. The table is
presented in appendix5.
In the pacemaker manufacturing when using EWMA for monitoring the system.
In this example the UCL and LCL values are 3999 and 1 respectively. Although the LCL
is not zero but still is very low and the chart cant capture the increased shift in the
process.
1
])/d Q_) (b b/YP)c!kb
`Y QY`c$ Q`YP cYdQ/P 7.35819
0.1359
]^4I 37175.15
1
])/d Q_) (b kb
`Y QY`c Q`YP cYdQ/P 6 7 51.77154
0.019316
]^4 52891.99
3(c` n3 0.538
3(c` l3 1000.38
% n3 0.108
% l3 0.892
The same procedure is repeated for %I from a range of 0.0005 to 0.005 and the
4 Results
In the methodology section, two infrequent failure scenarios were analyzed. Three
different control chart methods used to monitor the processes were analyzed with the cost
model developed. A step by step computation of the total cost per hour of using each of
the specific control charts was explained. In addition the total cost per hour of following
In this section the total cost per hour of using a specific type of control chart
(cost1), is compared to the total cost per hour of following up on every failure (cost2).
Figure 2 :Comparing total cost per hour of using probabilit based control charts vs. following up on every
failure when d=1 and C/F = 1
The research question that was posed earlier was at what frequency of failure does
statistical quality control result in a lower average cost per hour than follow up of every
failure?
From Figure 2 it can be concluded that for the specified range and parameter
When using a g-chart with probability based control limits and the process is in
control, the probability of a failure generating a signal is small, which results in large
values of ]^4I . When ]^4I is large it means the average number of inspected units until
a signal is large. This leads to less following up on infected patients which results in low
in-control cost.
However, in the out of control situation with % s %I the zero LCL value results
in a high out of control cost because the probability that a failure generates a signal goes
61
down. When the LCL is zero, the control chart loses its ability to detect an increased
expected time until an out of control process is brought back in to control (-o). The
high value of -o results in a low in-control percentage and increases the out of control
percentage which leads to an increased cost per hour of using control charts when the
The in control cost for following up on every failure is higher than when using
probability based control charts, because every infected patients is a signal. However, the
out of control cost is much smaller relative to the control chart since every infected
patient is a signal and the process is brought back to control much sooner. -o which is
the expected amount of time to bring an out of control process back in control is smaller.
Since the model parameters are not always known exactly, the model can be used
to conduct sensitivity analysis to examine the robustness of the prior conclusion for the
particular scenario considered, and other similar scenarios in other application contexts.
The conclusions are insensitive to the shift fraction d. When varying the shift
fraction between the 0.01 to 0.1 following up of every failure remains more efficient.
@
Sensitivity to the t ratio can also be analyzed. This ratio is a ratio of cost per unit
time over a cost per follow up so the exact units has no relevant interpretation, however it
is a measure of out-of-control cost, to the cost of bring a process back into control. If the
average time it takes to conduct failure follow-up is known, a dimensionless ratio can be
formed. For the specified %I range, when 1, follow-up of every failure is less costly,
@
62
however as gets smaller, %I also gets smaller. When = 0.4, %I = 0.1 as seen in Figure
@ @
3,
Figure 3: Comparing total cost per hour of using probability based control chart vs. following up on every
failure when d = 1 and C/F = 0.4
= 0.05, %I 0.01. These results imply that for a process with a failure
@
When
probability in the range of 0.01 to 0.1, the cost of follow up must be relatively large
Some additional figures with different shift fractions and different cost ratio
@
values of are presented in appendix 7.
infections
63
In the hospital acquired infection case, the total cost per hour was also computed
when using Shewhart g-chart for monitoring the infection process. Figures 4.3 shows a
charts that compares the total cost per hour of using a Shewhart g-chart versus following
The sensitivity analysis results are similar to those obtained for the g-chart with
probability based control limits. However, for Shewhart g-charts cost1 decreases as %I
increases, but the effect of this on the overall cost per hour is negligible.
The reason that cost1 is decreasing is because as %I is increasing the UCL limit is
decreasing and therefore more points plot above the UCL. In this case ]^4 decreases and
Figure 4: Comparing total cost per hour of using g-chart vs. following up on every failure when d = 1and
C/F =1
Additional figures of this example with different shift fraction and different cost
@
ratio values of are presented in appendix 7.
64
infections
In the hospital acquired infection case, the total cost per hour of using an EWMA
control for monitoring the system quality was also computed. The chart below shows the
total cost per hour of using EWMA versus following up on every failure for %I in the
Figure 5: Comparing total cost per hour of using EWMA vs. following up in every failure when d = 1 and
C/F = 1
With the specified parameters, when %I is between 0.001 and 0.04, it is more cost
efficient to follow up on every infected patient, but when %I is greater than 0.04 it is
more cost efficient to use an EWMA control chart for monitoring the process.
In the EWMA chart, when the process is in control, the probability of having an
out of control signal is small. This results in large values of ]^4I . When ]^4I is large it
65
means the average number of examined patients until an infected patient triggers a signal,
is large. This leads to less following up on infected patients when the process is in
The main feature of the EWMA charts that differs from the two other control
charts is the low value of the UCL. When the process is out of control, some infected
patients generate a signal because the number of inspections since the prior failure
exceeds the UCL . The probability of getting a signal, when the process is out of control
is higher than the two other methods resulting in a smaller ]^4 . This results in smaller
-Y , and the lower value of EY increases the in control percentage of the process, and
decreases the overall cost per hour of using EWMA control charts for monitoring the
process.
Note that, when the process is out of control the control chart signals are not
necessarily because of the right reason. When the failure probability increases ideally
points plotted below the LCL will trigger a signal, but because the LCL is very small this
doesnt happen. The primary reason signals are generated is because of the low UCL
value. These signals indicate possible quality improvements, which is not the case.
increases the %I get smaller and vice versa. . In addition, it can be concluded that there
is a positive association between and %I . As gets smaller, the %I also gets smaller.
@ @
When the cost ratio ( ) gets as small as 0.05, %I occurs at 0.01. The sensitivity
@
Figure 6: Comparing total cost per hour of using EWMA vs. following up on every failure when d = 1 and
C/F = 0.05
Figure 7: Comparing total cost per hour of using EWMA vs. following up on every failure when d = 1 and
C/F = 20
67
In the %I range of 0.01 to 0.1, EWMA charts are a viable method that may be
more cost effective than a policy of following up every failure. A more detailed estimate
Some Additional figures of this example with different shift fraction and different
@
cost ratio values of is presented in the appedix7.
The analysis presented was conducted for g-charts with probability distribution-
based control limits, but the conclusions are also applicable to the use of Shewhart g-
charts, which produced very similar results. Chart 8 compares, cost1, for probability
based control chart with cost2 , for %I ranging from 0.0005 to 0.005.
Figure 8: Comparing total cost per hour of using probability based control chart vs. following up on every
failure whe d = 1 and C/F = 1
68
From the figure 8 it can be concluded that following up on every defective unit is
a more cost-efficient strategy, when the probability of having a defective unit is from
0.0005 to 0.005.
When gets as small as 0.1, %I 0.0048. The cost of failure follow up must be
@
relatively high to justify the use of a g-chart. Results are shown in figure 9.
Figure 9: Comparing total cost per hour of using g-chart vs. following up on every failure when d = 1 and
C/F = 0.1
pacemaker production scenario are so high (with many difficult to quantify costs) that g-
Some Additional graphs of this example with different shift fractions and
@
different values of cost ratio is presented in the appedix8.
69
manufacturing
The figure 10 compares the total cost1 for using EWMA chart versus cost2 for %I
Figure 10: Comapring total cost per hour of using EWMA vs. following up on every failure when d = 1 and
C/F = 1
Figure 11: Comparing total cost per hour of using EWMA vs. following up on every failure when d = 1 and
C/F = 0.25
Some Additional figures with different shift fraction and different cost ratio
@
values of is presented in the appedix8.
71
5 Conclusion
The mathematical model developed in this research is a model for analyzing and
comparing the cost of utilizing a statistical control chart method for monitoring a
process. The model was employed to analyze the use of g-charts in a scenario (hospital
example) with relatively large failure probabilities (but still infrequent), and a relatively
low inspection rate of units (infection rate monitoring), and a scenario with much smaller
failure probabilities and higher inspection rate of units (pacemaker assembly). The
monitored statistic in both scenarios was geometrically distributed, for which g-charts
were developed. The scenario results as well as sensitivity analysis show that when
compared to a policy of following up every failure, g-charts should only be used when
the cost of failure follow up is very high. In most scenarios where low defect rates are
monitored carefully this will not be the case. This is a robust conclusion and the model
In contrast the use of EWMA did not require extreme scenarios, except with very
low failure probabilities, to generate a lower cost. The main difference between the g-
charts and the EWMA charts are the width of the control limits. This indicates that the
model developed may be useful to set specific control limits that minimize the long-run
6 References
Chakraborti, S., & Van der Laan, P. (1996). Precedence tests and confidence bounds
for complete data: an overview and some results. The Statistician, 351-369.
Chakraborti, S., Laan, P., & Wiel, M. A. (2004). A class of distributionfree control
charts. Journal of the Royal Statistical Society: Series C (Applied Statistics), 53(3), 443-
462.
Chan, L. Y., Xie, M., & Goh, T. N. (2000). Cumulative quantity control charts for
monitoring production processes. International Journal of Production Research, 38(2),
397-408.
Chan, L. Y., Ouyang, J., & Lau, H. Y. K. (2007). A two-stage cumulative quantity
control chart for monitoring poisson processes. Journal of quality technology, 39(3), 203-
223.
73
Human, S. W., Kritzinger, P., & Chakraborti, S. (2011). Robustness of the EWMA
control chart for individual observations. Journal of Applied Statistics, 38(10), 2071-
2087.
Goh, T. N. (1987). A control chart for very high yield processes. Quality Assurance,
13(1), 18-22.
Goh, T. N. (2007). Statistical monitoring and control of a low defect process.
Quality and reliability engineering international, 7(6), 479-483.
Klevens, R. M., Edwards, J. R., Richards, C. L., Horan, T. C., Gaynes, R. P.,
Pollock, D. A., & Cardo, D. M. (2007). Estimating health care-associated infections and
deaths in US hospitals, 2002. Public health reports, 122(2), 160.
Liu, J. Y., Xie, M., Goh, T. N., & Ranjan, P. (2004, October). Time-between-events
charts for on-line process monitoring. In Engineering Management Conference, 2004.
Proceedings. 2004 IEEE International (Vol. 3, pp. 1061-1065). IEEE.
Ozsan, G., Testik, M. C., & Wei, C. H. (2009). Properties of the exponential
EWMA chart with parameter estimation. Quality and Reliability Engineering
International, 26(6), 555-569.
Ross, S. M. (1996). Stochastic processes, 2nd.
75
Appendix1
Process In-Control
p0 LCL UCL Prob No Signal Prob Signal Events to Signal E[X|LCL<X<UCL] ARL0
0.01 0 658 0.998644 0.00136 737.36 98 72260.88
0.011 0 598 0.998644 0.00136 737.59 89.09090909 65712.4
0.012 0 548 0.998645 0.00136 737.82 81.66666667 60255.42
0.013 0 505 0.998633 0.00137 731.40 75.38461538 55136.32
0.014 0 469 0.998637 0.00136 733.84 70 51369.09
0.015 0 438 0.998646 0.00135 738.53 65.33333333 48250.53
0.016 0 410 0.998635 0.00136 732.83 61.25 44885.99
0.017 0 386 0.998641 0.00136 736.03 57.64705882 42429.98
0.018 0 364 0.998631 0.00137 730.35 54.44444444 39763.4
0.019 0 345 0.998638 0.00136 734.28 51.57894737 37873.37
0.02 0 328 0.998648 0.00135 739.73 49 36246.77
0.021 0 312 0.998640 0.00136 735.50 46.66666667 34323.35
0.022 0 298 0.998649 0.00135 740.22 44.54545455 32973.38
0.023 0 284 0.998619 0.00138 724.17 42.60869565 30855.79
0.024 0 273 0.998650 0.00135 740.71 40.83333333 30245.75
0.025 0 261 0.998616 0.00138 722.44 39.2 28319.53
0.026 0 251 0.998620 0.00138 724.87 37.69230769 27322.06
0.027 0 242 0.998635 0.00137 732.50 36.2962963 26586.91
0.028 0 233 0.998624 0.00138 726.82 35 25438.68
0.029 0 225 0.998629 0.00137 729.28 33.79310345 24644.53
0.03 0 217 0.998611 0.00139 719.96 32.66666667 23518.53
0.031 0 210 0.998614 0.00139 721.66 31.61290323 22813.66
0.032 0 204 0.998643 0.00136 736.72 30.625 22562.05
0.033 0 197 0.998608 0.00139 718.47 29.6969697 21336.5
0.034 0 192 0.998649 0.00135 740.23 28.82352941 21336.15
0.035 0 186 0.998627 0.00137 728.53 28 20398.82
0.036 0 181 0.998639 0.00136 734.74 27.22222222 20001.27
0.037 0 176 0.998637 0.00136 733.50 26.48648649 19427.82
0.038 0 171 0.998620 0.00138 724.83 25.78947368 18693.05
0.039 0 167 0.998645 0.00136 737.76 25.12820513 18538.69
0.04 0 162 0.998601 0.00140 715.04 24.5 17518.53
0.041 0 158 0.998602 0.00140 715.28 23.90243902 17097.04
0.042 0 154 0.998591 0.00141 709.70 23.33333333 16559.75
0.043 0 151 0.998630 0.00137 729.81 22.79069767 16632.86
0.044 0 147 0.998598 0.00140 713.10 22.27272727 15882.58
76
Prob IC-Cost1 IC- OC- OC-Cost2 % In control % In control Net Cost Net Cost 2
Cost2 Cost1 P/C No P/C 1
0.01 0.055355 40 1000 1080 2.2465E-05 0.927536232 999.9777 115.3623
0.011 0.0608713 44 1000 1088 2.4528E-05 0.933687003 999.9756 113.2308
0.012 0.0663841 48 1000 1096 2.65585E-05 0.938875306 999.9736 112.0587
0.013 0.0725475 52 1000 1104 2.90823E-05 0.943310658 999.9711 111.6372
0.014 0.0778678 56 1000 1112 3.08966E-05 0.947145877 999.9693 111.814
0.015 0.0829006 60 1000 1120 3.2457E-05 0.95049505 999.9677 112.4752
0.016 0.0891147 64 1000 1128 3.49229E-05 0.953445065 999.9653 113.5345
0.017 0.094273 68 1000 1136 3.65267E-05 0.956063269 999.9637 114.9244
0.018 0.100595 72 1000 1144 3.90124E-05 0.958402662 999.9612 116.5923
0.019 0.1056151 76 1000 1152 4.04524E-05 0.960505529 999.9598 118.4961
0.02 0.1103547 80 1000 1160 4.16571E-05 0.962406015 999.9586 120.6015
0.021 0.1165388 84 1000 1168 4.39397E-05 0.964131994 999.9563 122.8809
0.022 0.12131 88 1000 1176 4.51212E-05 0.965706447 999.9552 125.3114
0.023 0.1296353 92 1000 1184 4.89644E-05 0.967148489 999.9513 127.8739
0.024 0.13225 96 1000 1192 4.84645E-05 0.968474149 999.9518 130.5523
0.025 0.1412453 100 1000 1200 5.27276E-05 0.96969697 999.9476 133.3333
0.026 0.1464018 104 1000 1208 5.40777E-05 0.970828471 999.9463 136.2054
0.027 0.15045 108 1000 1216 5.45868E-05 0.971878515 999.9458 139.1586
0.028 0.1572409 112 1000 1224 5.70979E-05 0.972855592 999.9433 142.1846
0.029 0.1623078 116 1000 1232 5.83121E-05 0.973767051 999.9421 145.276
0.03 0.1700786 120 1000 1240 6.14745E-05 0.974619289 999.9389 148.4264
0.031 0.1753335 124 1000 1248 6.27641E-05 0.975417896 999.9376 151.6303
0.032 0.1772889 128 1000 1256 6.16743E-05 0.976167779 999.9387 154.8827
0.033 0.1874722 132 1000 1264 6.64441E-05 0.976873265 999.934 158.1795
0.034 0.1874753 136 1000 1272 6.3955E-05 0.977538185 999.9365 161.5166
0.035 0.1960898 140 1000 1280 6.751E-05 0.978165939 999.9329 164.8908
0.036 0.1999873 144 1000 1288 6.77474E-05 0.978759558 999.9327 168.2991
0.037 0.2058903 148 1000 1296 6.93551E-05 0.979321754 999.9311 171.7386
0.038 0.2139832 152 1000 1304 7.24388E-05 0.979854956 999.928 175.2071
0.039 0.2157649 156 1000 1312 7.11788E-05 0.980361351 999.9293 178.7023
0.04 0.2283297 160 1000 1320 7.72415E-05 0.980842912 999.9233 182.2222
0.041 0.2339586 164 1000 1328 7.85298E-05 0.981301421 999.922 185.7651
81
Appendix2
Process In-Control
p0 LCL UCL Prob No Signal Prob Signal Events to Signal E[X|LCL<X<UCL] ARL0
Cost
Prob IC-Cost1 IC- OC- OC- % In control % In control Net Cost 1 Net Cost 2
Cost2 Cost1 Cost2 P/C No P/C
0.01 0.755103 40 995.7555 115.3623 0.004274404 0.927536232 995.75554 115.36232
0.011 0.8281097 44 995.3805 113.2308 0.004652409 0.933687003 995.38052 113.23077
0.012 0.9006661 48 995.014 112.0587 0.005021903 0.938875306 995.01401 112.05868
0.013 0.9806383 52 994.5691 111.6372 0.005470497 0.943310658 994.56906 111.63719
0.014 1.0571312 56 994.1666 111.814 0.005876318 0.947145877 994.16662 111.81395
0.015 1.1326351 60 993.7789 112.4752 0.00626733 0.95049505 993.77894 112.47525
0.016 1.2154671 64 993.3146 113.5345 0.006735688 0.953445065 993.3146 113.53445
0.017 1.2771811 68 993.0851 114.9244 0.006967337 0.956063269 993.08511 114.92443
0.018 1.3509328 72 992.7273 116.5923 0.007328405 0.958402662 992.7273 116.59235
0.019 1.4346338 76 992.266 118.4961 0.007793892 0.960505529 992.266 118.49605
0.02 1.52546 80 991.7316 120.6015 0.008333215 0.962406015 991.73158 120.6015
0.021 1.5856206 84 991.532 122.8809 0.008534859 0.964131994 991.53202 122.88092
0.022 1.674602 88 991.0264 125.3114 0.009045302 0.965706447 991.02638 125.31139
88
Appendix3
Process In Control
p0 LCL UCL Prob No Signal Prob Signal Events to Signal E[X|LCL<X<UCL] ARL0
0.01 1 199 0.853299995 0.1467 6.816632353 98 668.03
0.011 1 181 0.852437176 0.14756282 6.776774639 89.09090909 603.749
0.012 1 166 0.851574405 0.1484256 6.737382452 81.66666667 550.2196
0.013 1 153 0.850161846 0.14983815 6.673867605 75.38461538 503.1069
0.014 1 142 0.849023834 0.15097617 6.623562017 70 463.6493
0.015 1 133 0.848986372 0.15101363 6.621918908 65.33333333 432.632
0.016 1 124 0.846470106 0.15352989 6.513389519 61.25 398.9451
0.017 1 117 0.846161574 0.15383843 6.500326538 57.64705882 374.7247
0.018 1 111 0.846398761 0.15360124 6.510364167 54.44444444 354.4532
0.019 1 105 0.844988144 0.15501186 6.451119448 51.57894737 332.742
0.02 1 99 0.841912166 0.15808783 6.325597447 49 309.9543
0.021 1 95 0.842989314 0.15701069 6.368993275 46.66666667 297.2197
0.022 1 90 0.839911321 0.16008868 6.246537882 44.54545455 278.2549
0.023 1 86 0.838631282 0.16136872 6.196987957 42.60869565 264.0456
0.024 1 83 0.8395778 0.1604222 6.233551219 40.83333333 254.5367
0.025 1 79 0.836209409 0.16379059 6.105356816 39.2 239.33
0.026 1 76 0.835349396 0.1646506 6.073466939 37.69230769 228.923
0.027 1 74 0.837406978 0.16259302 6.150325448 36.2962963 223.234
0.028 1 71 0.835025872 0.16497413 6.061556507 35 212.1545
0.029 1 68 0.83178424 0.16821576 5.944746186 33.79310345 200.8914
0.03 1 66 0.831909842 0.16809016 5.949188283 32.66666667 194.3402
0.031 1 64 0.831469838 0.16853016 5.933655963 31.61290323 187.5801
0.032 1 62 0.830470205 0.1695298 5.898668135 30.625 180.6467
0.033 1 60 0.828910082 0.17108992 5.844879776 29.6969697 173.5752
0.034 1 58 0.826781739 0.17321826 5.773063372 28.82352941 166.4001
0.035 1 57 0.829002931 0.17099707 5.848053468 28 163.7455
0.036 1 55 0.825910763 0.17408924 5.744180489 27.22222222 156.3694
0.037 1 54 0.827420753 0.17257925 5.794439466 26.48648649 153.4743
0.038 1 52 0.823347191 0.17665281 5.660821376 25.78947368 145.9896
0.039 1 51 0.824173763 0.17582624 5.687433324 25.12820513 142.915
0.04 1 49 0.819064894 0.18093511 5.526843425 24.5 135.4077
0.041 1 48 0.819210505 0.1807895 5.531294825 23.90243902 132.2114
0.042 1 47 0.819064439 0.18093556 5.526829535 23.33333333 128.9594
0.043 1 46 0.818631511 0.18136849 5.513636921 22.79069767 125.6596
0.044 1 45 0.81791476 0.18208524 5.491933328 22.27272727 122.3203
91
Cost
Prob IC-Cost1 IC- OC- OC- % In control % In control Net Cost 1 Net Cost 2
Cost2 Cost1 Cost2 P/C No P/C
0.01 5.9877553 40 1003.128 1080 0.333518445 0.927536232 670.56307 115.36232
0.011 6.6252696 44 1003.613 1088 0.366335972 0.933687003 638.3807 113.23077
0.012 7.2698251 48 1004.13 1096 0.397902902 0.938875306 607.47668 112.05868
0.013 7.9505959 52 1004.695 1104 0.428940803 0.943310658 577.15041 111.63719
0.014 8.6272095 56 1005.284 1112 0.458132282 0.947145877 548.68345 111.81395
0.015 9.2457323 60 1005.87 1120 0.484342119 0.95049505 523.16312 112.47525
0.016 10.026442 64 1006.571 1128 0.512518343 0.953445065 495.82363 113.53445
0.017 10.674503 68 1007.228 1136 0.536291157 0.956063269 472.7853 114.92443
0.018 11.284989 72 1007.894 1144 0.558104364 0.958402662 451.68199 116.59235
0.019 12.021328 76 1008.653 1152 0.580631011 0.960505529 429.97789 118.49605
0.02 12.905129 80 1009.519 1160 0.603659274 0.962406015 407.9039 120.6015
0.021 13.458059 84 1010.237 1168 0.620912847 0.964131994 391.32414 122.88092
0.022 14.37531 88 1011.176 1176 0.641335211 0.965706447 371.89253 125.31139
0.023 15.1489 92 1012.066 1184 0.658768659 0.967148489 355.32827 127.87385
0.024 15.714828 96 1012.874 1192 0.673183568 0.968474149 341.60279 130.55233
0.025 16.713326 100 1013.939 1200 0.690420734 0.96969697 325.43362 133.33333
0.026 17.473125 104 1014.905 1208 0.704554891 0.970828471 312.15935 136.20537
0.027 17.918415 108 1015.733 1216 0.715686328 0.971878515 301.61068 139.15861
0.028 18.854186 112 1016.846 1224 0.729394264 0.972855592 288.91659 142.18458
0.029 19.911253 116 1018.052 1232 0.742824698 0.973767051 276.60853 145.27597
0.03 20.582468 120 1019.082 1240 0.753278153 0.974619289 266.93416 148.4264
0.031 21.324225 124 1020.177 1248 0.763502096 0.975417896 257.55088 151.63029
0.032 22.142667 128 1021.34 1256 0.773472601 0.976167779 248.48835 154.88275
0.033 23.044764 132 1022.575 1264 0.78317582 0.976873265 239.76713 158.17946
0.034 24.038453 136 1023.886 1272 0.792606277 0.977538185 231.40054 161.51662
0.035 24.428153 140 1024.909 1280 0.799414738 0.978165939 225.10983 164.89083
0.036 25.580459 144 1026.356 1288 0.808318878 0.978759558 217.41028 168.29907
0.037 26.062988 148 1027.485 1296 0.814732071 0.979321754 211.59436 171.73863
0.038 27.399211 152 1029.083 1304 0.823109275 0.979854956 204.5877 175.20709
0.039 27.988666 156 1030.321 1312 0.829101509 0.980361351 199.28582 178.70228
0.04 29.540425 160 1032.089 1320 0.83698157 0.980842912 192.97436 182.22222
0.041 30.254569 164 1033.446 1328 0.842554654 0.981301421 188.20244 185.76515
0.042 31.017525 168 1034.858 1336 0.847960215 0.981738495 183.64116 189.32944
0.043 31.832021 172 1036.324 1344 0.8531972 0.982155603 179.29427 192.91363
0.044 32.701023 176 1037.847 1352 0.858266349 0.982554082 175.164 196.5164
0.045 33.627768 180 1039.427 1360 0.863169866 0.982935154 171.25142 200.13652
0.046 34.615798 184 1041.067 1368 0.867911149 0.983299933 167.55673 203.77288
96
Appendix4
Process In Control
p0 LCL UCL Prob No Signal Prob Signal Events to Signal E[X|LCL<X<UCL] ARL0
0.0005 2 13212 0.997649577 0.00235042 425.4552433 1960 833892.3
0.0006 2 11010 0.997449998 0.00255 392.1565217 1633.333333 640522.3
0.0007 1 9437 0.997949814 0.00205019 487.7606166 1400 682864.9
0.0008 1 8257 0.99784972 0.00215028 465.0557222 1225 569693.3
0.0009 1 7339 0.997749356 0.00225064 444.3172473 1088.888889 483812.1
0.001 1 6605 0.997649532 0.00235047 425.4472176 980 416938.3
0.0011 1 6004 0.997549033 0.00245097 408.002192 890.9090909 363492.9
0.0012 1 5504 0.997449885 0.00255012 392.1391422 816.6666667 320247
0.0013 1 5080 0.997349115 0.00265088 377.2325346 753.8461538 284375.3
0.0014 0 4717 0.998649156 0.00135084 740.2781531 700 518194.7
0.0015 0 4402 0.998648386 0.00135161 739.8563766 653.3333333 483372.8
0.0016 0 4127 0.998648969 0.00135103 740.1752039 612.5 453357.3
0.0017 0 3884 0.998648739 0.00135126 740.0496675 576.4705882 426616.9
0.0018 0 3668 0.99864851 0.00135149 739.924148 544.4444444 402847.6
0.0019 0 3475 0.998648957 0.00135104 740.1689893 515.7894737 381771.4
0.002 0 3301 0.998648593 0.00135141 739.9693846 490 362585
0.0021 0 3144 0.998649446 0.00135055 740.4366152 466.6666667 345537.1
0.0022 0 3001 0.998649487 0.00135051 740.4592972 445.4545455 329841
0.0023 0 2870 0.998648176 0.00135182 739.7410265 426.0869565 315194
0.0024 0 2750 0.998647134 0.00135287 739.1713861 408.3333333 301828.3
0.0025 0 2640 0.998647446 0.00135255 739.3420327 392 289822.1
0.0026 0 2539 0.998649653 0.00135035 740.550137 376.9230769 279130.4
0.0027 0 2444 0.998646445 0.00135356 738.7952331 362.962963 268155.3
0.0028 0 2357 0.998647841 0.00135216 739.5580632 350 258845.3
0.0029 0 2276 0.998649236 0.00135076 740.3217658 337.9310345 250177.7
0.003 0 2200 0.998649007 0.00135099 740.1962366 326.6666667 241797.4
0.0031 0 2129 0.998649184 0.00135082 740.2931238 316.1290323 234028.1
0.0032 0 2062 0.998647465 0.00135253 739.3525611 306.25 226426.7
0.0033 0 2000 0.998649943 0.00135006 740.7095369 296.969697 219968.3
0.0034 0 1941 0.998649444 0.00135056 740.4356147 288.2352941 213419.7
0.0035 0 1885 0.998647183 0.00135282 739.1985039 280 206975.6
0.0036 0 1833 0.998649256 0.00135074 740.3328841 272.2222222 201535.1
0.0037 0 1783 0.998647267 0.00135273 739.2439403 264.8648649 195799.7
0.0038 0 1736 0.998647173 0.00135283 739.1926115 257.8947368 190633.9
0.0039 0 1691 0.998644908 0.00135509 737.9573004 251.2820513 185435.4
99
Cost
Prob IC-Cost1 IC- OC- OC- % In control % In control Net Cost 1 Net Cost 2
Cost2 Cost1 Cost2 P/C No P/C
0.0005 0.0239839 10 1000.041 1020 0.012898035 0.864864865 987.14258 146.48649
0.0006 0.0312245 12 1000.059 1024 0.018463788 0.884792627 981.59449 128.58986
0.0007 0.0292884 14 1000.04 1028 0.012654404 0.899598394 987.38551 115.80723
0.0008 0.0351066 16 1000.052 1032 0.016461792 0.911032028 983.59023 106.39146
0.0009 0.0413384 18 1000.066 1036 0.020741192 0.920127796 979.32448 99.309904
0.001 0.0479687 20 1000.082 1040 0.025479912 0.927536232 974.60094 93.913043
0.0011 0.0550217 22 1000.099 1044 0.030664159 0.933687003 969.43335 89.771883
0.0012 0.0624518 24 1000.118 1048 0.036278961 0.938875306 963.83668 86.591687
0.0013 0.0703296 26 1000.138 1052 0.042308566 0.943310658 957.82662 84.163265
0.0014 0.0385955 28 1000 1056 3.30584E-05 0.947145877 999.96705 82.334038
0.0015 0.0413759 30 1000 1060 3.54366E-05 0.95049505 999.96468 80.990099
101
Appendix5
Process In-Control
Prob Shifted Prob Prob No signal Prob Signal Events to E[X|LCL<X<UCL] ARL1
Signal
0.0005 0.001 0.999664873 0.000335127 2983.94392 980 2924265.042
0.0006 0.0012 0.999665074 0.000334926 2985.737076 816.6666667 2438351.945
0.0007 0.0014 0.999665343 0.000334657 2988.129813 700 2091690.869
0.0008 0.0016 0.999665075 0.000334925 2985.740267 612.5 1828765.914
0.0009 0.0018 0.999665477 0.000334523 2989.330376 544.4444444 1627524.316
0.001 0.002 0.999665209 0.000334791 2986.939832 490 1463600.518
0.0011 0.0022 0.999665745 0.000334255 2991.730315 445.4545455 1332679.867
0.0012 0.0024 0.999665879 0.000334121 2992.931608 408.3333333 1222113.74
0.0013 0.0026 0.999665478 0.000334522 2989.340773 376.9230769 1126751.522
0.0014 0.0028 0.999666281 0.000333719 2996.536902 350 1048787.916
0.0015 0.003 0.999665881 0.000334119 2992.941704 326.6666667 977694.2898
0.0016 0.0032 0.999665613 0.000334387 2990.548157 306.25 915855.3729
0.0017 0.0034 0.999665748 0.000334252 2991.750261 288.2352941 862328.0165
0.0018 0.0036 0.999666684 0.000333316 3000.157958 272.2222222 816709.6663
0.0019 0.0038 0.999666752 0.000333248 3000.7641 257.8947368 773881.2679
0.002 0.004 0.999665883 0.000334117 2992.961759 245 733275.6309
0.0021 0.0042 0.999666285 0.000333715 2996.567543 233.3333333 699199.0934
0.0022 0.0044 0.999667221 0.000332779 3004.993294 222.7272727 669293.961
0.0023 0.0046 0.999667421 0.000332579 3006.807046 213.0434783 640580.6316
0.0024 0.0048 0.999666688 0.000333312 3000.189833 204.1666667 612538.7576
0.0025 0.005 0.99966622 0.00033378 2995.987859 196 587213.6203
0.0026 0.0052 0.999667223 0.000332777 3005.01847 188.4615385 566330.4039
0.0027 0.0054 0.999667291 0.000332709 3005.627866 181.4814815 545465.7979
104
Cost
Prob IC-Cost1 IC- OC- OC- % In control % In control Net Cost 1 Net Cost 2
Cost2 Cost1 Cost2 P/C No P/C
0.0005 0.1869878 10 1000.007 1020 0.002183805 0.864864865 997.82343 146.48649
0.0006 0.2243629 12 1000.008 1024 0.002617853 0.884792627 997.39092 128.58986
0.0007 0.2617043 14 1000.01 1028 0.003050392 0.899598394 996.95994 115.80723
0.0008 0.2992702 16 1000.011 1032 0.003487423 0.911032028 996.52452 106.39146
0.0009 0.3365443 18 1000.012 1036 0.00391695 0.920127796 996.09661 99.309904
0.001 0.3741626 20 1000.014 1040 0.00435374 0.927536232 995.66149 93.913043
0.0011 0.4113318 22 1000.015 1044 0.004779401 0.933687003 995.2375 89.771883
0.0012 0.4487256 24 1000.016 1048 0.005209547 0.938875306 994.80907 86.591687
0.0013 0.4865087 26 1000.018 1052 0.005647965 0.943310658 994.37243 84.163265
105
Appendix6
p0 LCL UCL Prob No Signal Prob Signal Events to Signal E[X|LCL<X<UCL] ARL0
0.0005 1 3999 0.864097021 0.13590298 7.358190429 1960 14422.05
0.0006 1 3333 0.864037647 0.13596235 7.354977153 1633.333333 12013.13
0.0007 1 2857 0.863951189 0.13604881 7.350303114 1400 10290.42
0.0008 1 2499 0.863756376 0.13624362 7.339793033 1225 8991.246
0.0009 1 2222 0.863737653 0.13626235 7.338784505 1088.888889 7991.121
0.001 1 1999 0.863529269 0.13647073 7.32757852 980 7181.027
0.0011 1 1818 0.863537659 0.13646234 7.32802908 890.9090909 6528.608
0.0012 1 1666 0.863356373 0.13664363 7.318306896 816.6666667 5976.617
0.0013 1 1538 0.863283477 0.13671652 7.314404835 753.8461538 5513.936
0.0014 1 1428 0.863156374 0.13684363 7.307611077 700 5115.328
0.0015 1 1333 0.863097032 0.13690297 7.304443551 653.3333333 4772.236
0.0016 1 1249 0.862847891 0.13715211 7.291174805 612.5 4465.845
0.0017 1 1176 0.862856378 0.13714362 7.29162602 576.4705882 4203.408
0.0018 1 1111 0.862837696 0.1371623 7.290632846 544.4444444 3969.345
0.0019 1 1052 0.862602144 0.13739786 7.278133945 515.7894737 3753.985
0.002 1 999 0.862393594 0.13760641 7.267103558 490 3560.881
0.0021 1 952 0.862456391 0.13754361 7.270421415 466.6666667 3392.863
0.0022 1 909 0.862437727 0.13756227 7.269434965 445.4545455 3238.203
0.0023 1 869 0.862188583 0.13781142 7.256292869 426.0869565 3091.812
0.0024 1 833 0.862156405 0.13784359 7.25459897 408.3333333 2962.295
0.0025 1 799 0.861825672 0.13817433 7.237234416 392 2836.996
0.0026 1 769 0.861983538 0.13801646 7.245512516 376.9230769 2731.001
0.0027 1 740 0.861693574 0.13830643 7.230322026 362.962963 2624.339
0.0028 1 714 0.861756431 0.13824357 7.233609528 350 2531.763
0.0029 1 689 0.861507152 0.13849285 7.22058945 337.9310345 2440.061
0.003 1 666 0.861393572 0.13860643 7.214672599 326.6666667 2356.793
0.0031 1 645 0.861497145 0.13850286 7.220067753 316.1290323 2282.473
0.0032 1 624 0.861030486 0.13896951 7.195822821 306.25 2203.721
0.0033 1 606 0.861337849 0.13866215 7.211773298 296.969697 2141.678
0.0034 1 588 0.861156482 0.13884352 7.202352818 288.2352941 2075.972
0.0035 1 571 0.860961483 0.13903852 7.192251605 280 2013.83
0.0036 1 555 0.86079358 0.13920642 7.183576732 272.2222222 1955.529
0.0037 1 540 0.860693583 0.13930642 7.178420212 264.8648649 1901.311
0.0038 1 526 0.860702235 0.13929777 7.178866056 257.8947368 1851.392
0.0039 1 512 0.860330446 0.13966955 7.159756537 251.2820513 1799.118
107
Cost
Prob IC-Cost1 IC- OC- OC- % In control % In control Net Cost 1 Net Cost 2
Cost2 Cost1 Cost2 P/C No P/C
0.0005 1.3867651 10 1000.394 1020 0.112013233 0.864864865 888.49214 146.48649
0.0006 1.6648451 12 1000.478 1024 0.132564875 0.884792627 868.07009 128.58986
0.0007 1.9435544 14 1000.563 1028 0.152582727 0.899598394 848.19065 115.80723
0.0008 2.2243857 16 1000.65 1032 0.172252542 0.911032028 828.6689 106.39146
0.0009 2.5027778 18 1000.738 1036 0.191043016 0.920127796 810.03213 99.309904
0.001 2.785117 20 1000.829 1040 0.209702826 0.927536232 791.53654 93.913043
0.0011 3.0634403 22 1000.92 1044 0.227370324 0.933687003 774.03674 89.771883
0.0012 3.3463745 24 1001.014 1048 0.244972768 0.938875306 756.61254 86.591687
0.0013 3.6271731 26 1001.108 1052 0.261830636 0.943310658 739.93729 84.163265
0.0014 3.9098179 28 1001.205 1056 0.278349035 0.947145877 723.6091 82.334038
0.0015 4.1909072 30 1001.303 1060 0.294230606 0.95049505 707.92196 80.990099
0.0016 4.4784362 32 1001.405 1064 0.310169579 0.953445065 692.18878 80.044693
109
Figure 12: Comparing total cost per hour of using probability based control charts vs. following up on
every failure when d = 2 and C/F = 1
Figure 13: Comparing total cost per hour of using probability based control chart vs. following up on
every failure when d = 0.5 and C/F = 1
111
Figure 14: Comparing total cost per hour of using g-charts vs. following up on every failure when d = 2
and C/F = 1
Figure 15: Comparing total cost per hour of using g-chart vs. following up on every failure when d = 0.5
and C/F = 1
112
Figure 16: Comparing total cost per hour of using EWMA vs. following up on every failure when d = 2 and
C/F = 1
Figure 17: Comparing total cost per hour of using g-chart vs. following up on every failure when d = 0.5
and C/F = 1
113
Figure 18: Comparing total cost per hour of using probability based control charts vs. following up on
every failure when d = 2 and C/F = 1
Figure 19: Comparing total cost per hour of using probability based control charts vs. following up on
every failure when d = 0.5 and C/F = 1
114
Figure 20: Comparing total cost per hour of using g-chart vs. following up on every failure when d = 2 and
C/F = 1
Figure 21:Comparing total cost per hour of using g-chart vs. following up on every failure when d = 0.5
and C/F =1
115
Figure 22: Comparing total cost per hour of using EWMA vs. following up on every failure when d = 2 and
C/F = 1
Figure 23: Comparing total cost per hour of using EWMA vs. following up on every failure when d = 0.5
and C/F = 1