Management Syst ems

i n
Pr oduc t i on Engi neer i ng

2012,No3(7),pp2630

Abstract:
Studydealswithananalysisofdatatotheeectthatitimprovesthequalityofstascaltoolsinprocessesofassembly
of automobile seats. Normal distribuon of variables is one of inevitable condions for the analysis, examinaon, and
improvement of the manufacturing processes (f. e.: manufacturing process capability) although, there are constantly
moreapproachestononnormaldatahandling.Anappropriateprobabilitydistribuonofmeasureddataisrstlytested
by the goodness of t of empirical distribuon with theorecal normal distribuon on the basis of hypothesis tesng
using programme StatGraphics Centurion XV.II. Data are collected from the assembly process of 1
st
row automobile
seatsforeachcharacteriscofquality(SafetyRegulaonS/R)individually.Studycloselyprocessesthemeasureddataof
an airbags assembly and it aims to accomplish the normal distributed data and apply it the stascal process control.
Resultsofthecontribuonconcludeinastatementofrejeconofthenullhypothesis(measuredvariablesdonotfollow
thenormaldistribuon)thereforeitisnecessarytobegintoworkondatatransformaonsupportedbyMinitab15.Even
thisapproachdoesnotreachanormaldistributeddataandsoshouldbeproposedaprocedurethatleadstothequality
outputofwholestascalcontrolofmanufacturingprocesses.
IMPROVINGQUALITYOFSTATISTICALPROCESSCONTROLBYDEALINGWITHNONNORMALDATAIN
AUTOMOTIVEINDUSTRY
INTRODUCTION
Manufacturing process quality of automobile seat in
organizaonismonitoredandevaluatedbyvariousstas
cal methods; one of them is the evaluaon of manufactu
ring process capability on the basis of control charts and
indexes of capability C
p
and C
pk
. Important condion for
evaluangtheprocesscapabilityisafactthatstudiedvaria
bles follow the normal distribuon. A mediumdicult
stascal tools of quality for hypothesis tesng about a
goodness of t (f. e.: ShapiroWilk, KolmogorovSmirnov,
etc.)areusedforanempiricaldistribuontwiththeore
caloneaerthebasicstascaltoolsfornormalityevalua
on such as histogram, density funcon and probability
plot. Contribuon focuses on the applicaon of tools for
stascalprocesscontrolinautomoveindustrywherewe
canoenndthenonnormaldatadistribuons.Measured
values are analysed so it is possible to reject or accept the
null hypothesis at the predetermined level of signicance
. Technique for the manufacturing process observaon
according to the quality of data entering into analysis is
determinedaertherevealedresults.
In case of acceptance of the null hypothesis, process
capability evaluaon of automove seats would be carried
on. Otherwise, it is necessary to approach to the transfor
maon of data, to alteraon of the indexes of manufactu
ring process capability C
p
and C
pk
for variables that do not
follow the normal distribuon or, as menoned by Haridi
(2011), to approach to a method which was developed by
Ford Motor Company (esmated cumulave frequencies,
mirroringtechnique).
OBJECTANDMETHODS
Datacollection
Measured data are collected from the process of as
semblyofautomobileseatsforAUDIQ7andregisteredinto
a database of screwing torques of S/R quality characteri
scs.Casestudycloselyfocusesontheanalysisofvariables
N=250(screwingtorques,Nm)fromtheprocessofairbag
assembly(gure1)whicharedenedfortheprocesscapa
bilityobservaonandevaluaon(table1).Nominalvalueof
thescrewingtorqueforairbagis8.10.81Nmwithatole
rance.
ZuzanaANDRSSYOV,TomPAULIEK,PavolPICHA,MarnKOTUS
SlovakUniversityofAgricultureinNitra
MariadelCarmenBASCERD
UniversidadPolitcnicadeValencia
Keywords:quality,stascalcontrol,hypothesistesng,probabilitydistribuon,normality,datatransformaon

ManagementSystemsinProduconEngineering3(7)/201227
Z.ANDRSSYOVetal.Improvingqualityofstascalprocesscontrolbydealingwithnonnormaldatainautomoveindustry
Fig.1.Componentsofautomobileseang
Table1
Measuredvariablesofscrewingtorqueforairbag,Nm

1 2 3 4 5 6 7 8 9 10 11 12 13
1 8.141 8.141 8.108 8.132 8.213 8.108 8.141 8.213 8.165 8.213 8.132 8.165 8.124
2 8.181 8.189 8.124 8.124 8.100 8.108 8.149 8.157 8.197 8.108 8.157 8.173 8.222
3 8.100 8.213 8.181 8.173 8.108 8.132 8.124 8.132 8.173 8.124 8.100 8.116 8.124
4 8.173 8.149 8.230 8.222 8.124 8.205 8.124 8.108 8.116 8.157 8.108 8.132 8.181
5 8.197 8.157 8.197 8.116 8.124 8.100 8.100 8.181 8.116 8.100 8.165 8.149 8.197
6 8.181 8.222 8.181 8.116 8.100 8.149 8.141 8.246 8.132 8.141 8.124 8.165 8.157
7 8.116 8.181 8.189 8.132 8.108 8.132 8.181 8.124 8.141 8.197 8.173 8.197 8.100
8 8.124 8.157 8.165 8.181 8.141 8.124 8.141 8.141 8.173 8.100 8.173 8.181 8.157
9 8.141 8.116 8.213 8.157 8.132 8.149 8.157 8.124 8.132 8.141 8.149 8.116 8.157
10 8.141 8.108 8.157 8.173 8.173 8.189 8.108 8.246 8.124 8.100 8.165 8.108 8.149
14
8.149
8.173
8,181
8.124
8.141
8.157
8.116
8.108
8.108
8.108
15
8.181
8.197
8.132
8.132
8.124
8.173
8.141
8.108
8.181
8.189
16
8.173
8.181
8.108
8.124
8.197
8.165
8.213
8.213
8.116
8.173
17
8.100
8.100
8.116
8.116
8.124
8.149
8.157
8.173
8.116
8.108
18
8.149
8.108
8.116
8.116
8.116
8.116
8.141
8.149
8.124
8.165
19
8.141
8.181
8.100
8.173
8.197
8.181
8.116
8.124
8.141
8.141
20
8.141
8.189
8.213
8.149
8.157
8.222
8.181
8.157
8.116
8.108
21
8.108
8.124
8.181
8.230
8.197
8.181
8.189
8.165
8.213
8.157
22
8.132
8.124
8.173
8.222
8.116
8.116
8.132
8.181
8.157
8.173
23
8.213
8.100
8.108
8.124
8.124
8.100
8.108
8.141
8.132
8.173
24
8.108
8.108
8.132
8.205
8.100
8.149
8.132
8.124
8.149
8.189
25
8.141
8.149
8.124
8.124
8.100
8.141
8.181
8.141
8.157
8.108
Histogram,probabilitydensityfunctionandprobability
plot
Firstly,allmeasuredvariablesareanalysedtodenethe
normal distribuon by histogram and probability density
funcon. Variables are ed into idencal intervals with a
number of intervals k and an interval width d. Apart from
calculaon of absolute and relave frequencies, values of
theprobabilitydensityfunconhavetobecalculated.Pro
bability density funcon is ploed on the histogram of
frequencies,itlinksmeansofindividualintervalsandcoun
ted values f(x). Probability plot and other resultant images
showifnormalornonnormaldistribuon.
Goodnessofitofempiricalwiththeoreticaldistribution
Pearson, ShapiroWilk, KolmogorovSmirnov tests of
goodness of t are used for the stascal hypothesis H
0

tesng about the goodness of t of empirical distribuon

withtheorecalnormaldistribuon.
Pearsoncoecient

where:
m
Ti
expectedfrequencyofenlargediinterval,
m
i
observedfrequencyofenlargediinterval,

ShapiroWilk

where:
x
(i)
orderedsamplevalues(x
(1)
isthesmallest),
a
i
constantgeneratedfromthemean,varianceandcova
rianceoftheorderstascsofasampleofsizenfrom
anormaldistribuon,

KolmogorovSmirnov

where:
sup
x
supremumofthesetofdistances,
F(x)cumulavedistribuonfuncon,
F
n
(x)empiricaldistribuonfuncon.

,
k
i
Ti
Ti i
m
) m m (
1
2
2

(1)

w =
( o
i
x
(i)
n
i=1
)
2
(x
i
-x )
2 n
i=1

(2)

n = sup
x
|F
n
(x) - F(x)] (3)

28ManagementSystemsinProduconEngineering3(7)/2012
Z.ANDRSSYOVetal.Improvingqualityofstascalprocesscontrolbydealingwithnonnormaldatainautomoveindustry
Levelofsignicanceisnormallydenedas=0.05(5%)
for hypothesis tesng which divides the set of values into
the rejected and accepted ones. A dierence between sta
scal expected results for normal distribuon and ob
served results would dene a nal stascal signicance
(probability)ofgoodnessoft.
Finally, formulated null hypothesis H
0
is rejected or
accepted. So the measured variables are evaluated as fol
lowing the normal distribuon or not, next moves are dis
cussed.
Transformationofnonnormaldata
The most common dealing with nonnormal data is ap
plyingatransformaontechnique.Transformaonchanges
themeasuredvariablexintotransformedvariableyaccord
ingtoaselectedmethod(logarithmic,squareroot,inverse,
arcsine, etc.). Family of transformaon Box Cox is consid
ered,byOsborne(2010),tobetheonewhichdevelopsand
improveconvenonaltransformaonmethods(logx,x,x

1
, arcsine x, etc.). So it is also used for an experiment de
pending on the esmated parameter (the most common
from5upto5).
BoxCox

There are several methods but they are used on

the basis of the distribuon of analysed data. Phase of de
terminaon of measured data distribuon and suitable
transformaon method is one of the most dicult. The
programme StatGraphics is used for graphical image of
funcon of measured data (lognormal, uniform, Weibull,
etc.) which could help with a following decision. Then a
selectedtransformaonisappliedondatainordertoreach
their normality for evaluaon of manufacturing process
capability of assembly of automobile seats and so improve
the quality of applicaon of stascal tools in automove
industry.
RESULTSANDDISCUSSION
Calculated frequencies were used for plong on
histogram and for a visual presentaon of distribuon of
the airbags measured data by programme StatGraphics.
The average value were counted x = 8.14932 and the
standarddeviaon=0.0353557,alsothenumberofinter
valsk=16andtherangeofintervalsd=0.0101.Probability
densityfunconf(x)fornormaldistribuonwasploedon
the histogram. Histogram, probability density funcon and
probability plot for normal distribuon shows that meas
uredvariablesofairbagsscrewingtorques(gure2)donot
followingthenormaldistribuon.Inthiscase,itwasneces
sarytovalidatethisresultbythetestofgoodnessoftfor
normaldistribuon.
Hypothesistestingaboutthegoodnessofit
The null hypothesis H
0
was formulated as a no
signicant dierence between the expected and observed
frequencies.Datafollowthenormaldistribuon.
H
0
:F(x
e
)=F(x
t
)
Alternave hypothesis H
1
was formulated as a signi
cantdierencebetweentheexpectedandobservedfrequ
encies.Datadonotfollowthenormaldistribuon.
H
1
:F(x
e
)F(x
t
)
Levelofsignicanceweredenedas=0.05.
Because the calculated values of resultant probability
are p < , the null hypothesis H
0
was rejected at the level of
signicance 5% (table 2). Detected dierence is very large
and so it cannot be random (chance), it is proved as the
stascal signicant dierence. Therefore, the alternave
hypothesis about the nonnormal distributed data was ac
ceptedatthelevelofsignicance95%.
Thefact,thatmeasuredvariablesdonotfollowthenor
mal distribuon and so they do not meet the condion of
normalityforthecapabilityevaluaonofairbagsassembly
by control charts and indexes of manufacturing process
capability C
p
and C
pk
, urges to propose next steps for study
ingandevaluangtheprocess.

(4)

(5)
Fig.2.Histogram,probabilitydensityfunconandprobabilityplotformeasuredvariables(StatGraphics)
Typeofthetest Stascs Probabilityp
ShapiroWilk W=0.933168 1.89e15
KolmogorovSmirnov Dn=0.111087 0.00418192
Pearson 2=51.7028 6.99822e07
Table2
Resultsofthetestofnormalityformeasureddataofscrewingtorque(StatGraphics)

ManagementSystemsinProduconEngineering3(7)/201229
Z.ANDRSSYOVetal.Improvingqualityofstascalprocesscontrolbydealingwithnonnormaldatainautomoveindustry
Datatransformation
Accordingtoafactthatitwasdiculttodeterminethe
distribuon of measured data of screwing torque from the
graphical image unequivocally, we applied the transforma
on methods for nonnormal data which were the most
edonesinpracce:inverse,squareroot,logarithmicand
family of Box Cox ( = 5, esmated by soware) with a
support of soware Minitab15. Further, the transformed
data were analysed in view of their distribuon (gure 3).
Noneoftransformedfunconfollowedthenormaldistribu
on.Resultsshowthatnoneofthetransformaonmethods
of this study can be used for measured data of screwing
torqueofairbagwithaimtoreachtheirnormality.Transfor
maon is not always the best choice for handling the non
normaldataalthoughitisthemostcommonmethod.The
refore it is needed to consider an applicaon of another
methodology for nonnormal data handling that areneces
sary for further stascal quality control of manufacturing
processes.
Johnson (2007) introduces a usage of Pearsons resp.
Johnson technique for goodness of t and a determinaon
of process capability with suitable percentage points of
distribuon in his study. He also states that whether a
transformaon method for nonnormal data has to be se
lectedornonnormaldistribuonmodelisgoingtobeiden
editisusefultochooseaStascalSoware(f.e.:Mini
tab can be used to accurately verify process stability and
calculate process capability for nonnormal quality charac
teriscs)forthatinvesgaon.
Thesamemanufacturingprocessesusuallydemonstrate
very similar data behaviour. If data distribuon is determi
nedcorrectlyandtransformaonappliedwithsuccess,itis
possibletousethistransformaonmethodforothersimilar
manufacturingprocesseswithoutthecloseranalysisofnon
normal data. Sharman (2012) states two types of non
normal data: data that follows the other distribuon and
datathatcontentsmixtureormulpledistribuons.Osbor
ne(2002)depictseectsofapplicaonofthedatatransfor
maonaccordingtoselectedtransformaonmethod.
CONCLUSION
Determinaonofthemeasureddatadistribuonbythe
basicstascaltoolsofquality(histogram)showsthatvalu
esdonotfollowthenormaldistribuon.Thisfactisvalida
tedbyhypothesistesngaboutthegoodnessoftofempi
ricalandtheorecalnormaldistribuon,thenullhypothesis
H
0
isrejected.Therefore,furtherstudyoftransformaonof
nonnormal data into data following the normality is nee
ded for Shewharts control charts and indexes of manufac
turingprocesscapabilityC
p
andC
pk
forthestascalquality
control. Procedure for normality sasfacon resp. calcula
on of indexes of process capability for nonnormal data
distribuon is going to be dened aer a closer analysis of
measured variables. Because there were not accomplished
theaimtoreachthecondionforstascalprocesscontrol
used in the organizaon, another technique for normality
or modicaon of indexes of process capability for non
normaldistributeddata.
Fig.3.Resulngprobabilityplotsoftransformeddataforeachtransformaonmethod(Minitab15)

30ManagementSystemsinProduconEngineering3(7)/2012
Z.ANDRSSYOVetal.Improvingqualityofstascalprocesscontrolbydealingwithnonnormaldatainautomoveindustry
REFERENCES
[1]HaridyS.,WuaZ.,CastagliolaP.:Univariateandmulva
riate approaches for evaluang the capability of dyna
micbehaviour processes. Stascal Methodology. Vol.
8,Issue2,2011,pp185203.
[2]Sherman J. P.: Tips for Recognizing and Transforming
NonnormalData.iSixSigma.hp://www.isixsigma.com/
toolstemplates/normality/psrecognizingand
transformingnonnormaldata/(23.03.2012).
[3]Osborne J.: Notes on the use of data transformaons.
Praccal Assessment, Research & Evaluaon. Vol. 8,
Issue 6, 2002. hp://PAREonline.net/getvn.asp?
v=8&n=6(25.03.2012).
[4]STN ISO 8258: 1995 (010271) Shewhart control charts.
SR:SlovakStandardsInstute(SUTN).1995.
[5]Johnson L.: Modelling nonnormal data using stascal
soware. Process control and process capability can
now be modelled using nonnormal distribuons. R&D
Magazine. hp://rdmag.com/FeaturedArcles/2007/
08/ModelingNonNormalDataUsingStascal
Soware/(26.03.2012).
[6]Osborne, J.: Improving your data transformaons: Ap
plying the BoxCox transformaon. Praccal Assess
ment, Research & Evaluaon. Vol. 15, Issue 12, 2010.
hp://pareonline.net/pdf/v15n12.pdf(18.05.2012).
Ing.ZuzanaAndrssyov
Ing.TomPauliek
Ing.PavolPicha
Ing.MarnKotus,PhD.
SlovakUniversityofAgriculture,FacultyofEngineering
DepartmentofQualityandEngineeringTechnologies
Tr.A.Hlinku2,94976Nitra,SLOVAKIA
email:xandrassyova@is.uniag.sk

MariadelCarmenBasCerd
UniversidadPolitcnicadeValencia
CentreforQualityandChangeManagement
CaminodeVera,s/n46022Valencia,SPAIN
email:maibacer@doctor.upv.es