You are on page 1of 7

Journal of Quality and Technology Management

Volume IV, Issue I1, Dec, 2008, pg. 5-11

Finding Characteristics of Process Capability Index, Cpk with


Different Distributions and Sample Sizes

S. M. Ali, M. A. Sarwar, A. Sultana


Department of Industrial and Production Engineering,
Bangladesh University of Engineering and Technology, Dhaka

Abstract
Quality is considered to be vital issue in improving productivity. A Process Capability Index
(PCI) has been considered as a valuable and popular tool to express the Capability of a
Process and assure quality of product. This paper summarizes the characteristics of the
Process Capability Index, Cpk with different sample size for both real and simulated data. It
also tries to examine the behavior of Cpk with different distributions. In this paper, two
population distributions were considered, i.e., Normal and Exponential with different values
of sample sizes to ascertain sample size effect on Process Capability Index, Cpk. The
Simulation was performed by using Engineering software MATLAB 7.0. From the result, it
was found that mean value of Cpk decreases up to sample size 15; then the value becomes
reasonably constant and hence Cpk Vs sample size graph tends to be flatter.
Keywords: Characteristics, Distribution, Process Capability Index, Sample, Sample size

Introduction used tool among the industrial and


Quality is a key criterion for survival of research personnel.
a product in a competitive environment.
The inspection of the quality of products Therefore value of this index tells how
or services provided by a company is an much capable the process is, assuming
infallible proof of the presence of the process follow the normal
competences within the organization distribution. But when the process does
not follow the normal distribution and
Industrial and quality control personnel the sample size is not fixed then what
use different types of indices to assure may be scenario is a matter of question.
quality of product. The most popular and Therefore this paper aims to find out the
widely used indices are Cp, Cpk and behavior of Cpk with different
Cpm. The earliest index Cp does not distributions (Normal and Exponential)
consider the mean of the process when and sample sizes for both real and
finding capability of process and another simulated data.
index Cpm does not directly relate to the
percentage of nonconforming product. Process Capability Index Cpk: An
To overcome the limitations of Cp and Overview
Cpm, another index (Cpk) is defined Process Capability Index (PCIs) have
which consider the mean of the process. generated extensive attention in
The index Cpk is defined in three-step statistical and quality control
procedure. At first Cpu which is equal to publications in recent times. The concept
allowable upper process spread divided of PCI was introduced by Juran. (1974)
by Actual upper process spread is found. as the ratio of tolerance width to a 6
Then Cpl which is equal to allowable multiple of a measure of variability for
lower process spread divided by actual the characteristic in question. The
lower process spread is found. And Cpk earliest form (First Generation Index) of
is the minimum of this two i.e. Cpk = min this type of PCI is
(Cpu, Cpl). This Cpk is the most widely
Cp  USL  LSL

……………………......(1)

5
Journal of Quality and Technology Management
Volume IV, Issue I1, Dec, 2008, pg. 5-11

distribution of X (this is 6 for a Normal


Where, USL and LSL are the upper and N (µ,) distribution). He defines Cpk as
lower specification limits respectively,
and σ is the standard deviation of the d ξ Μ
quality characteristics characteristic. It is C pk  1/2 ξ0.51a ξa  ..…………….... …(5)
apparent that Cp index measures
potential performance because only the Where, a is defined by Pr [Xa] =a,
process spread is related to the taking a=0.00135, so that 1-a, a are the
specification limits; the location of the upper and lower 0.135 percentiles of the
process mean is not considered here. It is
distribution of X, for, an N (µ,)
to avoid this situation, (Kane, 1986)
distribution 1-a = µ+3, a = µ-3 .
proposed the index Cpk that utilizes the
process mean and can be considered a
(Bai and Choi, 1997) have constructed
measure of process performance, where,
PCIs for use with possibly skewed
distribution of X, based on a ‘weighted
 USL μ μ  LSL ………..…(2)
Cpk  min ,  Variance’ (WV) approach. This utilizes
 3σ 3σ 
different divisors at the upper and lower
limits (U, L) of the specification interval.
In order to get an accurate estimate of With Pr [X≥µ] = P, they define Cpk as
process performance the sampling
conditions like normality, independence,  Uμ μL 
 min 
W
and statistical control are requirements. C pk
,
 3σ 2P 3σ 1  P 
This index, however, is only valid when
C pk ..……………..………...… (6)
all of the assumptions have met, but this 
rarely happens. The following is the way W
to extend the previous capability indices Where, W  1  1  2p
to the target value case.
In particular, ( Kane, 1986) proposed
Abbas Parchami. Mashallah Mashinchi,
 USL  T T  LSL  ………….(3)
Ali Reza Yavari and Hamid Reza Maleki
C pk   ,  (2005) defined Cpk with concept of Fuzzy
 3σ 3σ 
logic as follows:
Where, USL and LSL are the upper and au c1 2 m bu b12 m Cu a12 m 
CpkT , ,  ....….(7)
lower specification limits respectively, σ  6 6 6 
 
is the standard deviation of the
characteristic, T is the Target value of
the process. (Hsiang and Taguchi) noted Here,  au , bu , cu , Lal , bl , cl  FT ()
that if T=M, then three basic PCIs are be the upper and lower engineering
connected by the relationship, fuzzy specification limits provided that
au>cl and m = (bu + bl) =2 and t is the
1 target value.
C pk  C p  ………… (4)
 C 
2
 Kotz and Lovelace (1998) reported that
   1
3 
p
Leung et al. (997) examined five
C 
 pm   distributions with exactly 99.73%
conforming product (corresponding to a
(Clements, 1989), in an influential paper, centered, normal process with Cp = Cpk
suggested that “6” be replaced by the =1.0). The Cp values (calculated for non
length of the interval between the upper normal adjustment) are 0.5766, 0.7954,
and lower 0.135 percentage points of the 1.00, 1.221and 1.4030 when the
measurement arise from a uniform

6
Journal of Quality and Technology Management
Volume IV, Issue I1, Dec, 2008, pg. 5-11

triangular, normal, logistic and double plotted on the graph. Simulated data
exponential distribution. (Somerville and were generated for two different
Montgomery, 1996) investigated the distributions (Normal and Exponential)
errors (in parts per million) for various using MATLAB. And similarly the
degrees of freedom and five capability value Cpk was calculated for different
levels, when process data are t- sub- sample size and plotted on the
distributed, but are assumed normal. graph. Upper specification limit (USL)
They also investigated the one and two and lower specification limit (LSL) of
sided cases with Cpk =1.33 for this the sample product were given by the
situation. Even though the t-distribution management is given below:
is symmetrical and similar to the normal
when the degrees of freedom are high, Table 1: Product specification
the errors in estimating process yield are Name of the LSL
USL (cm)
substantial. (English and Taylor 1993) Organization (cm)
conducted extensive Monte Carlo MCWL 24.75 23.75
simulations of the distribution of Cpk for DABMCL 24.75 23.75
normal, symmetrical, triangular, uniform
and exponential distributions. Results and Discussion
There has been a debate in the Manufacturers typically use small
professional literature about the validity sample size for control chart purposes. A
and accuracy of the current indices. large value of sample size has been
However, (McCoy 1999) pointed out, considered to determine the sample size
“There should be more concern about effect. This paper has been depicted both
how the indexes are applied than about for real data of two ceramic industries
the indexes themselves-a point not to be and simulated data. Simulation
missed, especially when it comes to
considering the non normal issue. 1. Real Data: Mirpur Ceramic Works
(McCoy 1999) pointed out that, for the Limited
vast majority of processes, normal data The mean (μ) and standard deviation (σ)
seem impossible to find. But for of the process were 24.13 and 0.08
simplicity normality assumptions are respectively.
made.
Figure 1: Cpk Vs sub-sample size for real
Methodology and Data Collection data
Two Ceramic industries (Mirpur 5
4.5
Ceramic Works Limited (MCWL) and 4
Mean Cpk

3.5
Dhaka Automatic Bricks Manufacturing 3
2.5
Company Limited (DABMCL)) were 2
1.5
selected in this study. 10-Hole 1
0.5
Engineering Brick was considered as a 0
0 5 10 15 20 25 30 35 40 45 50 55 60
sample product. The ‘length’ of the Sub-sample Size

Brick was considered as quality


characteristics for that product. Length
of bricks (data) was taken for different From Fig. 1, we have observed that for
sub-sample sizes from each industry. smaller sample sizes the value of Cpk are
Hypothesis test was done to check the higher and for larger sample sizes the
distribution. The stability of the process value of Cpk are lower. The value of Cpk
was checked by using control chart. then decreases for sample size 5, 10 and
After that the value Cpk was calculated 15. For sample size15, the value of
for different sub-sample sizes and Cpk=1.68. After sample size 15, the

7
Journal of Quality and Technology Management
Volume IV, Issue I1, Dec, 2008, pg. 5-11

curve becomes fairly flat and the Cpk Figure 3: Cpk Vs sub-sample Size for
value reaches a reasonably constant simulated Exponential data
average for sample size 15 to 55. 3

2.5

Mean Cpk
2. Simulated Normal Data 2

The data were generated by using mean 1.5

1
and standard deviation of 24.21 cm and
0.5
0.08 cm respectively given by the
0
management. The simulations were 0 10 20 30 40 50 60 70 80 90 100 110 120 130 140

performed 20 times for each sample size


Sub-sample Size
3 to 125.

Figure 2: Cpk Vs sub-sample size for Again, it is seen from Fig. 3 that the
simulated Normal data value of Cpk decreases with small
3
fluctuations from 2.611 to 2.081 up to
2.5
sample size 3 to 15 and then the value of
Mean Cpk

1.5
Cpk attains a reasonably constant value
1
for rest of the sample size and
0.5 consequently the curve becomes fairly
0 flat.
0 10 20 30 40 50 60 70 80 90 100 110 120 130
Sub-sample Size
4. Real Data: Dhaka Automatic Bricks
Manufacturing Company limited
It is seen from Fig. 2 that the value of The mean (μ) and standard deviation (σ)
Cpk decreases from 2.7776 at point of the process were 24.25 and 0.08
having sample size 3 to 1.6373 at point respectively.
having sample size 15. Therefore, it is
seen that value of Cpk decreases with Figure 4: Cpk Vs sub-sample size for Real
slight random fluctuations up to sample data
size 15. From Fig. 3 it is seen that after 3.5

sample size 15, the value of Cpk becomes 3


2.5
Mean Cpk

constant regardless of sample size and 2

hence the curve is likely to be flatter at 1.5


1
higher value of sample size. 0.5
0
0 5 10 15 20 25 30 35
3. Simulated Exponential Data Sub-sample Size

The data were generated having taken


input 24.21 as expected mean of the
process. It is seen from Fig. 3 that the It is seen from Fig. 4 that maximum
value of Cpk at the start and end point is value of mean Cpk is 3.123 for sample
2.611 and 1.937 for sample size 3 and size 3 and the minimum value of mean
125 respectively Cpk is 2.401 for sample size 30.It is seen
from the Fig. 4 that the Cpk values
decreases smoothly starting from mean
Cpk value 3.123 for sample size 3 to
2.433 for sample size 15. After that, the
value of mean Cpk remains quite constant
with the increases of sample size up to
above 15.

8
Journal of Quality and Technology Management
Volume IV, Issue I1, Dec, 2008, pg. 5-11

5. Simulated Normal Data Figure 6: Combined graph of mean Cpk


The data were generated with mean Vs sub-sample size
24.25 as and standard deviation 0.15. 5

4.5
Real:
Figure 5: Cpk Vs Sub-sample size for 4
Simulated normal:
simulated Normal data 3.5
Simulated exponential:

Mean Cpk
3
3.5
2.5
3
2.5 2
Mean Cpk

2 1.5
1.5 1
1 0.5
0.5
0
0 0 10 20 30 40 50 60 70 80 90 100 110 120 130 140
0 20 40 60 80 100 120 140
Sub-Sample Size
Sub-sample Size

It is also seen from the Fig. 6 that all the


From Fig. 5 it is clear that for sample three curves for Cpk for real and
size 3 maximum value of mean Cpk with simulated situation tend to be flatter
value equal to 3.0142 was found. Then from Sample Size 15 to 20.
the value of mean Cpk sharply decreases
to an amount of 1.99 for sample size 5 as 2. Dhaka Automatic Bricks
indicated in the Fig. 6. Manufacturing Company Limited

After that, the mean value of Cpk Figure 7: Combined graph of mean Cpk
gradually decreases from 1.99 to 1.624 Vs sub-sample Size
at the point of having sample size 40. 5

With increase of sample size further 4.5

from 40 to 125 mean Cpk value attains a 4 Real:


3.5 Simulated normal:
somewhat constant value.
Mean Cpk

2.5
Combined Effect 2
The combined effect of real data (Mirpur 1.5
Ceramic Works Limited) with simulated 1

data (Normal and Exponential 0.5

distribution) is shown in Fig.6 and for 0


0 20 40 60 80 100 120 140
Dhaka Automatic Bricks Manufacturing Sub-Sam ple Size
Company limited in Fig.7 respectively.

1. Mirpur Ceramic Works Limited From Fig. 7, it is clear that slight higher
From Fig. 6, it is seen that for real data, values of Cpk have been obtained for real
initially mean Cpk value is slightly data than the Cpk values for simulated
higher than the Cpk values for simulated Normal data. Both curves are likely to be
Normal and Exponential distribution. flatter from sample size 15 to 20. Most
Also for real data, most of the Cpk values of the Cpk values for real data are
lie between the Cpk values for simulated reclining on the simulated Normal curve
Exponential and Normal distribution. due to the fact that real data have lower
standard deviation than simulated
Normal data.

9
Journal of Quality and Technology Management
Volume IV, Issue I1, Dec, 2008, pg. 5-11

Analysis of Effect of Sample Size on Figure 9: Sample standard deviation Vs


Cpk sub-sample size
From limiting value concepts, it has 0.09

Sample Standard deviation


0.08
been shown that when sample size 0.07
becomes large (n→ ∞), the mean and 0.06
0.05
standard deviation are moving towards 0.04

an average value or limiting value. So, 0.03


0.02
mean value of Cpk also attains a 0.01
0
reasonably constant value. 0 10 20 30 40 50 60 70 80 90 100 110 120 130
Sub-sample Size

From Fig. 1, Fig. 2, Fig. 3, Fig. 4 and


Fig. 5, it can be established the fact that From Fig. 9 it is seen that an upward
from sample size 15, mean value of Cpk trend of standard deviation from sample
attains a moderately constant value. The size 3 to 10. At sample size 15, the value
equation (2) of Cpk shows that the value of standard deviation is 0.08 around
of Cpk depends on mean and the standard which some standard deviation
deviation of the sample. So, Fig. 8 and randomly fluctuates and some are
Fig. 9 ‘mean Vs sub-sample size’ and approaching to this value for the rest of
‘standard deviation Vs sub-sample size’ the sample size. So, this standard
show the effect of mean and standard deviation can be considered as limiting
deviation on Cpk. The LSL and the USL value. For this constant value, Cpk
are constant, so the value of Cpk is only becomes fairly constant and hence the
depending on the mean and the standard curves tend to be flatter with increase of
deviation of the sample taken from the sample size after sample size 15.
process.
Conclusion
Figure 8: Sample mean Vs sub-sample
The aim of this paper was to find out the
size
effect of sample size on Process
24.144
24.142 Capability Index, Cpk and tried to find
24.14
out the pattern of behavior of Cpk with
Sample Mean

24.138
24.136
24.134
different distributions. From Figure 6
24.132
24.13
and Figure7, it has been seen that up to
24.128 sample size 15, the mean values of Cpk
24.126
24.124 decrease i.e. the Cpk values are reflecting
0 10 20 30 40 50 60 70 80 90 100 110 120 130
Sub-sample size
the variability of the process. In other
words, it has been outlined that mean
values of Cpk are showing the rejection
From Fig. 8 it is seen that at sample size rate and hence acting as a mirror of the
15, the mean value is 24.13. All other actual condition of the process i.e. voice
mean values for different sample size are of the process. However, after sample
randomly fluctuating around this mean size 15, the mean values of Cpk become
value keeping it as centered. Therefore, constant and the curves tend to be flatter.
for sample size 15, mean value of Cpk The constant value of Cpk does not give
attains a stable position and holds this any indication about variability of the
value constant throughout the entire process. It can be concluded that the
sample range after sample size 15. So sample size should be smaller than 15 to
after sample size 15 we do not get any make a clear sense about the voice of the
clear idea of the process as the process process i.e. to make decision on
does not show any shift. capability of the process.

10
Journal of Quality and Technology Management
Volume IV, Issue I1, Dec, 2008, pg. 5-11

From rule of thumb of process capability Hsiang, T. C. and Taguchi, G., “Tutorial
index, we know that the higher the Cpk on Quality Control and Assurance- The
value, the better it can meet engineering Taguchi Method”. Joint Meetings of the
tolerances or customer specifications. American Statistical Association, Las
From Fig. 6 and Fig. 7, it has also been Vegas, Nevada, pp. 188
observed that with increase of sample
size the mean values of Cpk decreases Jerry, B. 1989, “Principles of Quality
with increase of sample size. Therefore, Control”. Jhon Wiley and Sons. Inc.
for getting higher values of Cpk, sample New York, USA
size should be smaller. After considering
the facts it has been concluded that Juran, J. M., (1974), “Juran’s Quality
sample size should be less than 15 when Control handbook”, 3rd edition,
control chart would be constructed. McGraw-Hill, New York

Again from Fig. 6, it is seen that for Kane, V.E., (1986), “Process Capability
normal distribution, higher values of Cpk Indices”, Journal of Quality
have been obtained up to sample size 15, Technology, Vol. 18, pp. 41-52
so, values of Cpk are better suited for
Normal distribution as this satisfies the Kotz, S. and Lovelace, C. 1998,
thumb rule-the higher, the better. Also, it “Introduction to Process Capability
is seen that normality assumption is true Indices”. Arnold, London,UK
in this case.
McCoy, P.F. 1999, “Using Perform
Acknowledgement Indexes to Monitor Production
The authors would like to articulate their Process”. Quality Progress, pp. 49-55
indebtedness to many personnel of
Mirpur Ceramic Works Limited and Parchami, A., Mashinchi, M., Yavari
Dhaka Automatic Bricks Manufacturing A.R., and Maleki, H.R., 2005, “Process
Company Limited for their sincere Capability as Fuzzy numbers”.
cooperation through the period of their Australian Journal of Statistics, Vol. 34,
research work. Number 4, pp. 391-402

References Somerville S. E. and D. C. Montgomery,


Bai, D. S. and Choi, S. S., 1997, 1996, “Process Capability Indices and
“Process Capability Indices for Skewed Non-Normal Distributions”. Quality
Populations. Master Thesis, Department Engineering 9:2, pp. 305-316
of Industrial Engineering, Advanced
Institute of Science and Technology, Nomenclature
Taejon, South Korea Symbol Meaning
μ Mean
Clements, J.A. (1989), “Process σ Standard deviation
Capability Calculations for Non-Normal
USL Upper Specification limit
Distributions. Quality Progress, Vol.22,
pp. 95-100 LSL Lower Specification Limit

English, J.R. and Taylor, G.D. 1993,


“Process Capability Analysis-a
robustness study” Int. J. Prod. Res.
31:7, pp. 1621-1633

11

You might also like