International Journal for Quality Research 10(2) 407–420

ISSN 1800-6450

Yerriswamy Wooluru 1
Swamy D.R. PROCESS CAPABILITY ESTIMATION FOR
Nagesh P. NON-NORMALLY DISTRIBUTED DATA
USING ROBUST METHODS - A
COMPARATIVE STUDY
Article info:
Received 12.10.2014 Abstract: Process capability indices are very important
Accepted 13.01.2015 process quality assessment tools in automotive industries. The
UDC – 54.061
common process capability indices (PCIs) Cp, Cpk, Cpm are
DOI – 10.18421/IJQR10.02-11 widely used in practice. The use of these PCIs based on the
assumption that process is in control and its output is normally
distributed. In practice, normality is not always fulfilled.
Indices developed based on normality assumption are very
sensitive to non- normal processes. When distribution of a
product quality characteristic is non-normal, Cp and Cpk
indices calculated using conventional methods often lead to
erroneous interpretation of process capability. In the
literature, various methods have been proposed for surrogate
process capability indices under non normality but few
literature sources offer their comprehensive evaluation and
comparison of their ability to capture true capability in non-
normal situation. In this paper, five methods have been
reviewed and capability evaluation is carried out for the data
pertaining to resistivity of silicon wafer. The final results
revealed that the Burr based percentile method is better than
Clements method. Modelling of non-normal data and Box-Cox
transformation method using statistical software (Minitab 14)
provides reasonably good result as they are very promising
methods for non –normal and moderately skewed data
(Skewness ≤ 1.5).
Keywords: Process capability indices, Non - normal process,
Clements method, Box - Cox transformation, Burr
distribution, probability plots

1. Introduction1 specifications are different in different

products (Pearn et al., 1995). A frontline
Process mean µ, Process standard deviation manager of a process cannot evaluate
σ and product specifications are basic process performance using µ and σ only. For
information used to evaluate process this reason Dr. Juran combined process
capability indices however, product parameters with product specifications and
introduces the concept of process capability
1 indices (PCI).Since then ,the most common
Corresponding author: Yerriswamy Wooluru
email: ysprabhu@gmail.com indices being applied by manufacturing
industry are process capability index Cp and

407
process rattio for off - cen
ntre process Cp
pk are 2.1.. Weighted vaariance method
d
defined as
Hsin-Hung Wu Proposed a nnew process
Cp =
–
(1) capability index applying thhe weighted
’ variiance control charting methhod for non–
norm mal processses to immprove the

Cpk n
Min , (2) meaasurement of pprocess perforrmance when
the process data aare non-normallly distributed
Capability indices are widely useed to andd shows that tthe two weighhted variance
determine whether a pro ocess is capab ble of metthod are basedd on the same pphilosophy to
producing items within
w customer spliit a skewed oor asymmetricc distribution
specificatio
on limits or not. The prrocess fromm the mean. The main idea of the
capability indices Cp and Cpk heeavily weiighted variancce method is to divide a
depend on n an implicit assumption
a thaat the skew wed distribuution into ttwo normal
underlying quality characteeristic disttribution from its mean to creeate two new
measuremeents are indepeendent and norrmally disttributions whicch have the samame mean but
distributed. However, these basic diffferent standaard deviationns. For a
assumption ns are not fulfilled in actual poppulation with a mean of µ annd a standard
practice as many physicaal processes pro oduce devviation of σ, thhere are obsservations out
non- norm mal data and quality
q practitiioners of a total observattions which arre less than or
need to verify
v that thee assumptions hold equual to µ. Also,, there are observations
before dep ploying any PCI techniqu ues to out of n total obseervations whicch are greater
determine the capability of their proccesses. thann µ.the two new distributtions can be
Some auth hors have prrovided usefull and estaablished by usiing and observations
insightful information
i reg
garding the misstakes resppectively. Thhat is, the two new
in interprretation that occur with h the disttributions will have the samee mean µ, but
misapplication of indicess to non-normaal data diffferent standardd deviations and .For
(Choi and Bai, 1996; Montgomery,
M 1996; the estimation oof µ, and ,µ can be
Box and Cox, 1964). Alternatively,
A other estimated by ̅ ,i.ee.∑ / ,and and

authors haave introduceed new indices to can be estiimated by and
handle thee skewness in n the data (Boyels, resppectively. Stanndard deviationn with
1994) Tang g and Than (1 1999) reported d on a observations whicch are less thaan or equal to
comparativ ve analysis am mong seven in ndices the value of ̅ caan be computeed by using
designed foor non-normal distribution. milar formula too that used to calculate the
sim
mple standardd deviation for n total
sam
2. Surroogate PCIs for
f Non-Norrmal observations (Ahm med et al., 20008)
Distriibutions
=∑
= (3)
Here the following methods m have been 2∑
presented to t compute PC CIs for non-n
normal 4
distribution
n. 2 1
 Weighted
W varian
nce Method
Also, the sample standard deviaation with
 Cllements Metho od
observations w
which are greeater than the
 Bu urr Method
ue of can be calculated as
valu
 Bo ox-Cox Transfformation Meth hod
 Modelling
M nonn-normal data using 2∑
Sttatistical softwaare 5
2 1

408 u,, D.R., Swam

Y. Wooluru my, P. Nageshh
The two commonly ussed normally--based stan
ndard deviatioon, skewness aand kurtosis.
process caapability indicees are p, pk
p are Und der the assuumption that these four
modified using the weighted varriance paraameters determmine the type oof the Pearson
method as follows disttribution curvve, Clements utilized the
Cp (WV) = (6) tablle of the famiily of Pearsonn curves as a
funcction of skewnness and kurtossis. Clements
repllaced 6σ by U − L ) inn the below
Cpk (WV) index can be expressed
e as:
equuation (6).
C (8)
Cpk (WV) = min , 7

Wh here, U is the 99.865 percenntile and L is

2.2. Clemeents method
the 0.135 percenntile. For C , the process
For non-n normal pearssonian distrib bution meaan μ is estimaated by mediann M, and the
(which includes
i a wide classs of twoo 3 are estim mated by U – M) and (M
“populationns” withh non-n
normal − L respectivelyy, Figure 1 deepicts how a
characteristics) Clements (1989) has PCII are obtainned for a nnon-normally
proposed a novel meth hod of non-n normal disttributed qualityy attribute.
percentiles to calculate process
p capabillity C USSL M M LSL
C min , 9
and proceess capability y for off centre
c U
U M M L
process C indices bassed on the mean,
m

gure 1. probabiility distributio

Fig on curve for a non-
n normal daata with spec. L
Limits

Proceduree for calcula ating PCIs using  Look uup standarddized 0.135
Clements method
m (Boylees, 1994): percentil e,
 Ob btain specificaation limits USL and  Look uup standardiized 99.865
LSSL for a given quality
q percentil e
chharacteristic  Look up standardized M Median
 Esstimate samplee statistics for the
t  Calculatee estimateed 0.135
giiven sample daata: Sample sizee, Lp = x - s Lp
percentil e using Eqn. L
Mean,
M Standard deviation,  Calculatee estimated 99.865
Skkewness, Kurto osis percentil e using Eqn. Up = x + s
Up

409
  Caalculate estimmated median using Burrr XII distribut
ution can be ussed to obtain
Eqqn. M = x + s M , for po ositive the required perccentiles of var
ariate X .The
sk
kewness reverse sign, probbability densitty function off a Burr XII
variiate Y is
 or negative skewness leave
fo
poositive f( ,k) = if y ≥ 0; c ≥ 0 (10)

 Caalculate nonn-normal prrocess 0 f ( ,k) = 0 if y < 0
k≥ 0, (11)
caapability indicees using Equatiions.
Whhere c and k rrepresent the sskewness and
 C = , C = , kurttosis coefficiients of thee Burr XII
disttribution resppectively .Thherefore, the
an
nd C , C =Min cum
mulative distribbution functionn of the Burr
disttribution is deriived as:
C ,C ]
[C F( ,k) = 1 if y ≥ 0 (12)

F( ,k) = 0,if y < 0 (13)

2.3. Burr distribution
d
Burrr-XII distribuution can bee applied to
Although Clements’s
C meethod is widely y used estimate capabilitty indices to pprovide better
in industry
y today, Wu et al. (1999) indiicated estimate of the pprocess capabiility than the
that the Clements’s method can n not com
mmonly used C Clements methhod. Liu and
accurately measure thee nominal values, v Cheen introduced a modificationn based on the
especially when the underlying data Clements methodd, whereby insttead of using
distribution
n is skewed.. To conduct the Peaarson curve ppercentiles, thhey replaced
process cap pability analyssis when the quality
q them
m with percenntiles from ann appropriate
characteristic data iss non- norrmally Burrr distribution ((Castagliola, 1996).
distributed, Clements’s method can n be
modified byb replacing th he Pearson fam mily of 2.3..1 Procedure ffor calculatingg PCIs using
probabilityy curves wiith a Burr XII Burrr XII Distrib ution method d
distribution
n to improve the t accuracy of o the
estimates of the indicees for non-n normal Burrr method invo lves followingg steps:
process datta. Two reason ns justify the use
u of  Estimate s the sample m mean, sample
the Burr XII
X distribution. First reason is i that standard deeviation, skeewness and
the two paarameter Burr-X XII distributio
on can kurtosis of th e original sampple data.
be used to describe data that arise in th he real
 Calculatee standardizedd moments of
world and especially tho ose concerning g non-
skewness (α ) and kurtosiss (α ) for the
normal pro ocesses. The seecond reason is i that
given sample size n, as folloows:
the direct use
u of a fitted cumulative fun nction ̅
instead of a probability density
d function
n may  ∑ , where,
avoid the need for a numerical
n or formal
f
integration. It is found thhat a wide ran nge of ̅ is mean off the observatiions and s is
the skewneess and kurtosis coefficientts of the standard ddeviation.
various pro obability densiity functions can
c be ̅
∑ - ,
covered by y different com
mbinations of c and k
.Such prob bability densityy functions in nclude where n is thee number of obbservations in
most know wn functions, including no ormal, the data.
Gamma, Beta, Weibulll, Logistic, Log-  Use the values of and to
normal and d other functionns. select the apppropriate Burr parameters c
and k.Then uuse the standarddized Z = (x -
̅ ) / s = (Q
Q - µ)/σ, whhere x is the

410 u,, D.R., Swam

Y. Wooluru my, P. Nageshh
 randomm variate of th
he original dataa. Q is Most common transformatiions reduce
the seelected Burr variate,µ
v and σ its positive skew buut may exacerbbate negative
correspponding meaan and staandard skeww unless the vvariable is refleected prior to
deviatiion respectiveely. The mean n and tran
nsformation. B ox-Cox eliminnates the need
standaard deviations as
a well as skew wness for it (Box and Coox, 1964).
and kurtosis
k coefficcients, for a large
collecttion of Burrr distributionss are 2.5.. Modelling Non-Normal data using
found in the tables of Burr (Chou, 1996) Statistical softwaare
and (Castaglioa,
( 1996). From these
tables, the standardiized lower, median
m Quaality control engineers arre frequently
and uppper percentiless are obtained. askeed to evaluaate process sstability and
 Caalculate estimated perceentiles capability for keyy quality characteristics that
using Burr
B table for lower, median n, and follow non-normaal distributionss. In the past,
upper percentiles as follows: Lp = ̅ + demmonstrating process staability and
s . , M= ̅ + s . , Up = ̅ + s capability requiire the assuumption of
.
norm mally distribut
uted data. How wever, if data
 Caalculate processs capability in
ndices do not follow thhe normal disttribution, the
using equations
e preseented below. resu
ults generated under this asssumption will
be incorrect. W Whether it is decided to
C = , C C = nsform data to follow the normal
tran
C C ,C ]
= Min [C disttribution or iddentify an apprropriate non-
norm mal distribuution modell statistical
softtware’s can bee used. Identiffication of an
Cox power tran
2.4. Box-C nsformation apppropriate non-nnormal distribuution model is
a good
g approacch to find a non-normal
Box and Cox
C (1964) prrovides a family of disttribution that fits the data. Many non-
power tran nsformations that
t will optiimally norm mal distributioon can be usedd to model a
normalize a particular variable,
v elimin
nating resp
ponse, but if aan alternative tto the normal
the need to random mly try diffferent disttribution is ggoing to be viable, the
transformations to determ
mine the best option.
o expponential, loognormal, annd weibull
It transform
m non-normal data into normal
n disttributions usuaally works w well. Minitab
data on th y positive response
he necessarily statistical softwarre can be used to verify the
variable X as shown in th
he below equatiion proccess stabilityy and estim mate process
= For ≠ 0 (14) capability for non-normal quality
characteristics.
= ln For 0 (15)
This continnuous family depends
d on a single
parameter λ,
λ it can on ann infinite numb bers of 3. Methodologgy
values. This
T family of transformaations
incorporatees manny tradiitional Metthodology invoolves followingg steps:
transformations like:  Understaanding the basiic concepts of
process ccapability analyysis for non-
Square roo ot transformatiion, λ= 0.50, Cube
normal ddata
ormation, λ=0.33.
root transfo
 Data Colllection
Fourth rooot transformatio on, λ=0.25, Natural
N  Calculatee required staatistics of the
log transforrmation, λ =0.000 case studdy data
Reciprocal square root transformation n, λ=-  Validate the critical asssumptions.
0.50, Recipprocal transformmation, λ=-1.0 00,  Estimatioon of Cp, Cppu, Cpl, Cpk
No transformation needeed, when λ= 1..00, it using noon normal m methods and
produces reesults identicall to original datta. classical method

411
  Co
omparison of PCIs
P of non-n
normal 4. Constructioon of Contrrol chart,
methods
m with PCIs of claassical Normal proobability ploot and
method
m
histogram ffor validatin
ng the
3.1. Data collection
c stability an
nd normalityy
assumption n.
In order to discuss and d compare thee five
methods to o deal with noon-normality issues, 4.1.. Constructioon of Controol chart to
the data simmilar to an ex
xample presentted by asseess the stabilitty of the proceess
Douglas Montgomery
M in introductio on to
statistical Quality Contrrol, fifth editiion is In this
t study, in order to dem monstrate the
considered in this paperr. Table 1 preesents app
plicability of thhe method annd to make a
consecutivee measuremen nts on the resisstivity cleaar decision abbout the capaability of the
of Silicon wafers.
w Descrip
ptive statistics:: prodduction pro cess, -R chart are
Mean: 205 5.32; Standard deviation: 0.0405; nstructed usingg Minitab 14 software to
con
Skewness; 0.39; Kurto osis; 0.21; Range:
R veriify stability oof the processs. Figure 2
0.09785 dispplays that the process is in ccontrol as all
the mean and raange values arre within the
con
ntrol limits on tthe both charts

Table 1. Data
D of bore diaameter using bo
oring operation
n
X1 X2
X X3 X4 X5 X bar R
1 205.324 20
05.275 205..356 205.34
49 205.343 205.329 0.081
2 205.302 20
05.310 205..312 205.26
60 205.300 205.297 0.052
3 205.346 20
05.280 205..336 205.315 205.346 205.325 0.066
4 205.326 20
05.438 205..288 205.42
29 205.299 205.356 0.150
5 205.330 20
05.397 205..305 205.36
68 205.354 205.351 0.092
6 205.333 20
05.316 205..271 205.314 205.318 205.310 0.062
7 205.282 20
05.396 205..306 205.34
48 205.297 205.326 0.114
8 205.297 20
05.354 205..329 205.33
30 205.324 205.327 0.057
9 205.409 20
05.313 205..269 205.32
23 205.319 205.327 0.140
10 205.342 20
05.397 205..265 205.30
05 205.303 205.322 0.132
11 205.368 20
05.397 205..295 205.26
62 205.315 205.327 0.135
12 205.389 20
05.301 205..316 205.319 205.353 205.336 0.088
13 205.356 20
05.298 205..356 205.27
70 205.294 205.315 0.086
14 205.252 20
05.273 205..350 205.24
41 205.361 205.295 0.120
15 205.326 20
05.297 205..377 205.37
71 205.316 205.337 0.080
16 205.334 20
05.234 205..318 205.30
03 205.342 205.306 0.108
17 205.287 20
05.262 205..316 205.38
83 205.312 205.312 0.121
18 205.333 20
05.328 205..259 205.33
36 205.396 205.330 0.137
19 205.325 20
05.297 205..320 205.33
35 205.285 205.312 0.050
20 205.369 20
05.283 205..336 205.30
06 205.336 205.326 0.086
205.323 0.09785

412 u,, D.R., Swam

Y. Wooluru my, P. Nageshh
 Xbarr-R Chart of he
eight-1
205.38 U C L=2205.3796

205.35

Sample M ean
_
_
205.32 X=205 .3233

205.29

LC L=2005.2671
205.26
1 3 5 7 9 11 13
1 15 17 19
Sample

0.20 U C L=00.2061
Sample Range

0.15

_
0.10 R=0.09975

0.05

0.00 LC L=0
1 3 5 7 9 11 13
1 15 17 19
Sample

Figure
F 2. X and
d R chart

4.2. Construction off histogram and the normality of the data. Figuure 3 display
normal probability plot to check
c the histogram annd Figure 4 display the
normality of the data norm mal probabilityy plot for the data set. The
histtogram for sam
mple data appeaars to be non-
Graphical methods inclu uding the histo ogram norm mal.
and normall probability pllot are used to check

His
stogram of heig
ght-1
Normal
25 Mean 20
205.3
StDev 0.044050
N 100

20

15
Frequency

10
F

0
205.24 205.28 205
5.32 205.36 205.40 205
5.44
height-1

Figure 3. Histogram
H for case study data

413
 Probab
bility Plot of height-1
Normal - 95% CI
C
99.9
Mean 205.3
StDev 0.04050
99
N 100
AD 0.474
95 P-Valuee 0.237
90
80
70
Percent

60
50
40
30
20
10
5

0.1
205.20 205.25 205.30
0 205.35 20
05.40 205.45 205.50
h
height-1

Figure 4.
4 Normal Prob
bability Plot

The validity of non-norrmality is testeed by the mean value in the data set, = 52
using Andeerson – Darling g test (AD).Thee hole Nummber of obseervations greaater than the
diameter data
d is considerred as normall as it meaan value in the data set, = 48
pass normaality test becau use, the P-valuue is(>
0.005),greaater than criticcal value (0.05).This Thee sample standdard deviationn with
is done by y using Minitaab 14 software ,the observations whicch are lower thhan the value
result of tesst is shown in Figure
F 3and 4. of can be calcullated as:

∑
5. Comp
putation of PCI’s
P

(16)

For case study data using the folloowing = 0.0700

methods:
 Weighted
W varian
nce Method Also, the sample standard deviaation with
 Cllements Metho od observations w
which are greeater than the
 Bu urr Method ue of can be calculated as:
valu
 Bo ox-Cox Transfformation Meth
hod
∑
hted variance Method
5.1. Weigh M (17)

The statistiics for the obtaained sample daata: = 0.0288

Std. deviattion =0.0405, MeanM = 205.32 and
Median = 205.32, USL=205.60,
U LSL= Thee two commoonly used noormally-based
205.00 proccess capabiliity indices Cp,Cpk are
Total numb ber of observaation in the datta set, moddified using the weightted variance
n = 100 metthod as followss:
Number off observations less l than or eqqual to

414 u,, D.R., Swam

Y. Wooluru my, P. Nageshh
 . . .
Cp (WV) = = = min
m ,
. . . .

. = min
m [3.24, 4.81]] =3.24
= 3.92
.

mputation off PCIs usin

Com ng Clements
Cpk (WV)) index can bee expressed as, Cpk
metthod (Table 2)).
(WV) = miin ,

Table 2. Prrocess capabiliity calculation procedure usin method

ng the Clementts’s percentile m
Step No. Procedure Notations Calcuulations
1 Specification ns : USL L 205.600
Upper specification Limit Spec. M
Mean 205.300
Target resisttivity LSL 205.000
Lower speciification Limit
2 Estimate sammple statistics:
N 100
Sample size
̅ 205.32
Mean
S 0.0405
Standard dev viation
Sk 0.39
Skewness
Ku 0.21
Kurtosis
3 Look up stan ndardized 0.13 35 percentile 2.4676
Lp
4 Look up stan
ndardized 99.8
865 percentile 3.5037
Up
5 Look up stan
ndardized Median in table 2 0.0652

6 Calculate estimated 0.1135 percentille

using Lp 2205.220
L = ̅ - s
Eqn. Lp
7 Calculate estimated
e 99.865 percentille
Up 2205.461
using Eqn. Up = ̅ + s

8 Calculate esstimated mediaan using

M 2205.322
Eqn. M = ̅ + s

9 Calculate noon-normal proocess capabilitty

indices using
g Equations.
C 2.489
C = , C = , C 3.156
C 2.00
C 2.00
C , C M [C , C ]
= Min

415
Computation of PCIs using Burr’s
B
method (T
Table 3).

Table 3. Prrocess capabiliity calculation using the Burrr percentile metthod

Step Procedure Notations Calculations
No.
1 Speccifications :
Uppeer specification Limit
L USL 205.60
Targ
get resistivity Spec. Mean 205.30
Lower specification Limit LSL 205.00
2 Estim
mate sample stattistics:
Samp ple size N 100
Meaan of sample data ̅ 205.32
Standdard deviation ( overall) S 0.0405
Skewwness Sk 0.39
Kurtosis Ku 0.21
3 mate standard moments
Estim m of skewwness ( ) and Kurtosis
K ( ) 0.384
using
g Sk and Ku valu ues from step 2. 3.13
4 Baseed on and from step 3 ,seelect the parameeters c and k c 2.5377
valuees using the Burrr-XII distributio
on table k 12.5234
5 Withh reference to the parameters c and
a k obtained in step 4, use . -2.085
the table
t of standarrdized tails of the
t Burr XII diistribution to . -0.082
deterrmine standardizzed lower, mediaan and upper perrcentiles. 3.595
.
6 Calculate estimated 0.135
0 percentile using Lp 205.2355
Eqn. Lp = ̅ + s .
7 Calculate estimated 99.865
9 percentille using Up 205.4655
Eqn. Up = ̅ + s .
8 Calculate estimated median using M 205.3166
n. M = ̅ + s .
Eqn
9 Calculate non-normaal process capabiility indices usin
ng equations. C 2.6080
C =
= , C = , C 3.9038
C 1.9032
C , C = Min [C , C ] 1.9032
C

Cox Transform
5.2. Box-C mation Thee Lambda tabble as shown in figure 5
conntains an estimaate of lambda ((-0.21) which
The Box-C Cox transformation parameteer (λ) is the
t value usedd in the transsformation. It
is estimated
d by Minitab144 statistical sofftware also
o includes the upper Confiddence Interval
and correspponding processs capability inndices (0.4
46) and loweer Confidencee Interval (-
are determmined. The acccuracy of the Box- 0.955),which are marked on thhe graph by
Cox transfformation is robust to depaartures verttical lines .In tthis case studyy, an optimal
from norm mal and it avo oids the troub ble of lam
mbda value thhat correspondds to -0.21 is
having to search for a suitable
s methood for utiliized for trannsforming thhe data and
each distrib
bution encounttered in practicce. calcculation of PC CIs. The figuree 6 shows the
outpput of the Minnitab 14 statistiical software.

416 u,, D.R., Swam

Y. Wooluru my, P. Nageshh
 Box
x-Cox Plot of Resistivity
R Da
ata
Lower C L Upper CL
175 Lambda
(using 95.0% confideence)
Estimate --0.21
150
Lower C L --0.95
Upper C L 0.46

Rounded Value 0.00

125
StDev

100

75
Limit

50
-5.0 -2.5 0.0
0 2.5 5.0
Lambbda

Figu
ure 5. Box Cox
x plot to estimaate optimal valuue of

Process Cap
pability of Res
sistivity Data
a
Using Box-Cox Transformation Wiith Lambda = -0.221

U S L* LS LL*
P rocess D ata transformed datta Within
LS L 100 O v erall
Targget *
USL 500 P otential (Within) C apability
mple M ean
S am 241.35 Cp 0.988
S am
mple N 100 C P L 1.099
S tD ev
e (Within) 66.9729 C P U 0.877
S tD ev
e (O v erall) 75.5519 C pk 0.877
C C pk 0.988
A fter Transformation
O v erall C apa bility
LS L*
* 0.380189
Targget* * Pp 0.944
U S L* 0.271154 PPL 1.044
mple M ean*
S am 0.319472 PPU 0.833
S tD ev
e (Within)* 0.0184948 P pk 0.833
S tD ev
e (O v erall)* 0.0194258 C pm *

0.28 0.30
0 0.32 0.34 0.36 0.3
38
O bseerv ed P erformance E xp.
x Within P erformance E xp. O v erall P erform ance
P P M < LS L 0.00 P P M > LS L* 513.62 P P M > LS L* 887..14
P P M > U S L 10000.00 P P M < U S L* 4494.34 P P M < U S L* 6436..14
P P M Total 10000.00 P P M Total 5007.96 P P M Total 7323..29

Figure 6. Prrocess capability Analysis usiing Box- Cox ttransformationn

putation of PCIs
5.3. Comp P using Burr’s
B lognnormal are ussed to model the response
method (ressistivity of ssilicon wafer)) .Individual
disttribution identiification featurre in Minitab
In this case study, theooretical non-n
normal 14 isi used to com mpare the fit off distributions
distribution
ns like expon
nential, weibulll and as shown
s in the figgure 7.

417
 Probabil ity Plot for Resistivity
G ooodness of F it T est
Lognorm al - 95 % C I E xponential - 95% C I
99.9 99.9 Loggnorm al
99 90 A D = 0.435
50 P -V
V alue = 0.295
90
P er cent

P er cent
10 ponential
E xp
50
A D = 23.551
P -V
V alue < 0.003
10 1
1 Weeibull
A D = 2.286
0.1 0.1
100
1 200 500 0.1 1 .0 10.0 100.00 1000.0 P -V
V alue < 0.010
R e s istiv itty R e sistiv ity
G aam m a
% CI
Weibull - 95% G am m a - 95% C I A D = 0.793
P -V
V alue = 0.042
99.9 99.9
99
90
90
50
P er cent

P er cent
50
10

10
1
1

0.1 0.1
10 100 1000 1 00 200 500
R e s istiv itty R e sistiv ity

Figurre 7. Probability
y plots for the individual disttribution

5.3.1 Comparison
C of altern
native gooodness of fit wwith the data. IIn this study,
distributio
ons with P-vvalues (For 95% seveen distributionns are consideered to select
Confidencce Interval) the appropriate oone that fits thhe data. The
lognnormal distribbution providess the best fit
Individual distribution id dentification feature
fe in comparison
c wiith other distribbutions as its
in statisticaal software (Minitab 14) is used to p-vaalue (0.295) iss greater than critical value
construct probability plots for said (0.0
05).
distributionns in order to compare their

Table 5. Comparison of Alternative

A Disstributions usin
ng output from
m probability pllot
Distribution
n type AD
A value P-value
Weibull 2.286 < 0.010
Exponential 23.55 < 0.003
Log logistic 0.432 0.242
Largest extreme value 0.359 > 0.250
Lognormal 0.435 0.295
Gamma 0.793 0.042
Normal 2.045 < 0.005

Process cappability indicees for the case study 6. Results and

d Discussion
n
data using lognormal disstribution are found
through the output of Minitab
M 14 statiistical Thee following T Table 6 presennts the PCIs
software ass shown in the figure 8. calcculation resultss of different m
methods.

418 u,, D.R., Swam

Y. Wooluru my, P. Nageshh
Table 6. Numerical
N resullts for PCIs of Non-normal
N an
nd Classical M
Method
PCIs Obtained Results
R
Weeighted Cleements Burr Box Cox Lognormal Classical
varriance Meethod Distriibution Trannsformation Model mmethod
method (Normality
assumption)
Cp 3.9
92 2.48 2.60 0.98 0.97 22.37
Cpl 4.81 3.15 3.90 1.09 1.01 22.50
Cpu 3.2
24 2.00 1.90 0.87 0.93 22.19
Cpk 3.2
24 2.00 1.90 0.87 0.93 22.19

In this pap
per, the Clemen nts, Burr, Weiighted resiistivity of siliccon wafer is iinadequate as
variance, Box-Cox
B transsformation meethods all values
v are lesss than 1.33 soo ,the process
are revieweed and used to o estimate the PCIs disppersion need to be reducedd and process
for non-n normal quality characteristic data. meaan have to bbe shifted to ccloser to the
PCIs of Classical method d are compared d with targget value of 2225 from existting mean of
the PCIs of all the no on-normal meethods 241.79.
considered in the case study. In caase of Clements methodd is simple extension of the
classical method,
m Cpu is over estimateed and tradditional 6σ m method, whichh takes into
Cpl is undeer estimated, when
w compared d with acco ount the posssible non-norm mality of the
PCIs of other non n-normal metthods. basiic data.
Weighted variance
v (WV)) method givess good
Burrr-based methhod works well under
result but it requires manual
m calculaations.
disttributions thhat depart slightly or
Box- Cox x transformatiion method gives
mod derately from normality (Skkewness ≤
reasonably good resu ults compared d to
1.5)).
classical method.
m Burr peercentile metho od has
been used effectively anda it shows better Oveerall performannce of Box-Coox method is
results com
mpared to Clemments method. slig
ghtly inferior than thee lognormal
disttribution modeel, Lognormall distribution
7. Concllusions mod del exhibits thhat 6179 partss per million
exceeding the sppecification lim mits but in
In practicee, manufacturring processess that casee of Box –C ox transformaation method
yields nonn- normally distributed
d datta are 732 23 parts perr million exxceeding the
inevitable, therefore thee use of tradiitional speccification lim mits, hence it can be
process capability
c ind
dices to meeasure conncluded that modeling of data to
capability of
o such processses give misleeading logn normal distribuution approachh is accurate
results. onee.
Box-Cox method is su uccessfully used to Thee estimates maade using the B Burr method,
transform the non-normaal data to norrmally whiich are higheer than those made using
distributed data and estim
mated the PCIs.. Clements’s methhod are good indicator to
help p quality conntrol engineeers be more
The obtainned values off process capaability atteentive to and foocus on processs adjustment
indices shhows that thee capability of o the andd improvement..
production process fo or controllingg the

Referencces:

Ahmed, S., Abdollanian, M., Zeephong gsekul, P. (200

08). Process cappability estimaation for non-
normal quality
q characcteristics: A Comparison
C off Clements, Buurr and Box-C Cox method.
ANZIAMM Journal, 49(E EMAC 2007), C642-C665.
C

419
Boyles, R.A. (1994). Process capability with asym mmetric toleraances. Commuunications in
Statisticss: Simulation and
a Communication, 23(3), 615-643
Castagliolaa, P. (1996). Evaluation off non-normal process capabbility indices using Burr’s
distributiion. Quality Enngineering, 8, 587-593.
5
Chou, Y., Polansky,
P A.M
M., & Mason, R.L.
R (1998). Trransforming noon-normal dataa to normality
in statistiical process co
ontrol. Journal of Quality Tecchnology, 30, 1133-141.
Choi, I.S., & Bai, D.S. (1 1996). Process capability ind dices for skeweed populations .Proceedings
of the 200 th Internationnal conference on computer an nd Industrial E
Engineering, 12211-1214.
Clements, J.A, (1989). Process
P capabiility calculatio
ons for non-noormal distributtions. Quality
Progresss, 22, 95-100.
Montgomeery, D. (1996). Introduction to Statistical Quality
Q Controol. 5th edition.. Wiley, New
York.
Box, G.E.P P., & Cox, D.R.D (1964). An
A analysis off transformatioons. Journal of the Royal
Statistica
al Society. Seriies B (Methodo ological), 26(2)), 211-252.
Wu, H.H.,, Swain, J.J., Farrington, P.A., & Messim mer, L.S. (19999). A weighhted variance
capabilitty index for General
G non-noormal processses. Quality annd Reliability Engineering
Internatiional, 15, 397-402.
Pearn, W.L L., Chen, K.S., & Lin G.H. (1995). A gen neralization of Clements metthod for non-
normal pearsonian
p process with asym mmetric toleran nces. Internatiional journal of quality and
Reliabiliity Managemen nt, 16(5), 507-5
521.
Tang, L. & Than, S. (1999). Computting process capability indicces for non-noormal data: a
review and comparrative study.. Qual. Relliab. Engng. Int., 15(55), 339-353.
http://dx.doi.org/10.100 02/(sici)1099-1
1638(199909/1 10)15:5<339::aaid-qre259>3.00.co;2-a.

wamy Woolurru
Yerrisw Swa
amy D.R. Nagesh P.
JSS Acaademy of Techniical JSS Academy of Technical JSS Centre for M
Management
Educatio
on, Eduucation, studies,
Bangalo
ore Banngalore Mysore
India India India
ysprabhuu@gmail.com drsw
wamydr@gmail.com pnagesh1973@rrediffmail.com

420 u,, D.R., Swam

Y. Wooluru my, P. Nageshh