Welcome to Scribd, the world's digital library. Read, publish, and share books and documents. See more
Download
Standard view
Full view
of .
Save to My Library
Look up keyword
Like this
0Activity
0 of .
Results for:
No results containing your search query
P. 1
2010 Augmented Measurement System Assessment JQT

2010 Augmented Measurement System Assessment JQT

Ratings: (0)|Views: 1 |Likes:
Published by souhasouha

More info:

Published by: souhasouha on Apr 18, 2013
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

04/18/2013

pdf

text

original

 
Augmented Measurement SystemAssessment
NATHANIEL T. STEVENS
Business and Industrial Statistics Research Group, Department of Statistics and Actuarial Sciences,University of Waterloo, Waterloo, N2L 3G1 Canada 
RYAN BROWNE
Department of Mathematics and Statistics, University of Guelph, Guelph, N1G 2W1 Canada 
STEFAN H. STEINER and R. JOCK MACKAY
Business and Industrial Statistics Research Group, Department of Statistics and Actuarial Sciences,University of Waterloo, Waterloo, N2L 3G1 Canada 
The standard plan for the assessment of the variation due to a measurement system involves a numberof operators repeatedly measuring a number of parts in a balanced design. In this article, we considerthe performance of two types of (unbalanced) assessment plans. In each type, we use a standard planaugmented with a second component. In type A augmentation, each operator measures a different setof parts once each. In type B augmentation, each operator measures the same set of parts once each.The goal of the paper is to identify good augmented plans for estimating the gauge repeatability andreproducibility (GR&R), a ratio that compares the contribution of the measurement system to the overallprocess variation. We show that, if there are three or more operators or if we include the possibility of part-by-operator interaction, then use of an appropriate augmented plan can produce substantial gainsin efficiency for estimating GR&R compared with the best standard plan with the same total number of measurements.Key Words: Gauge Repeatability and Reproducibility; Measurement Variation; Part-by-Operator Interaction;Repeatability; Reproducibility; Study Design.
Introduction
I
N MANY
manufacturing processes, parts will bemeasured to ensure that certain specifications are
Mr. Stevens is a Masters student in the Department of Statistics and Actuarial Science. His email address is nstevens@uwaterloo.ca.Dr. Browne is a Post Doctoral Fellow in the Departmentof Mathematics and Statistics. His email address is rbrowne@uoguelph.ca.Dr. Steiner is a Professor in the Department of Statisticsand Actuarial Science and Director of the Business and In-dustrial Statistics Research Group. He is a senior member of ASQ. His email address is shsteiner@uwaterloo.ca.Dr. MacKay an Associate Professor in the Department of Statistics and Actuarial Science. He is a member of ASQ. Hisemail address is rjmackay@uwaterloo.ca.
met. However, these measurements may be mislead-ing if the measurement system (the devices, people,and protocol used to measure a part) is itself not ad-equate. Accordingly, many quality systems requirethe periodic assessment of critical measurement sys-tems used to qualify parts or to make decisions aboutprocess control.Here we deal with the assessment of a nondestruc-tive measurement system that determines a singlecontinuous characteristic or dimension. In assessingsuch a system with
r
operators, it is standard prac-tice to randomly sample
k
parts from the currentprocess and have each of the
r
operators measureeach part
n
times for a total of 
=
krn
measure-ments. We call this a standard plan (SP), denotedby SP(
k,n
). Common choices suggested by the formsin AIAG (2003, pp. 216–217) are
k
10;
r
= 2
,
3;
Journal of Quality Technology 
388
Vol. 42, No. 4, October 2010 
 
AUGMENTED MEASUREMENT-SYSTEM ASSESSMENT 389
n
= 2
,
3 so that 40
90. In our experience,most measurement assessment studies in industryfollow these guidelines closely. In this article, we lookcarefully at the standard plan and suggest a differentallocation of the total number of measurements. Wealso consider augmented (unbalanced) plans in whichsome parts are only measured once by an operator.There has recently been considerable activity inthe development of different study plans and thecorresponding analyses for measurement system as-sessments in a wide variety of situations. De Mastet al. (2010) look at ordinal measurement systems.Van Wieringen and De Mast (2008) and Danila etal. (2010) consider the assessment of binary measure-ment systems. Van Der Muelen et al. (2009) examinethe assessment of a destructive measurement system.Browne et al. (2009b, 2010) look at the use of two-stage plans for assessing gauge repeatability & repro-ducibility (GR&R) where parts for the second stage(a standard plan) are selected based on an initialmeasurement.For a standard plan, we use the following modelto specify the attributes of interest in the measure-ment system and to analyze the data collected forthe assessment:
ijl
=
i
+
μ
j
+
PO
ij
+
ijl
(1)where
i
= 1
,
2
,...,k
(parts),
j
= 1
,
2
,...,r
(oper-ators),
l
= 1
,
2
,...,n
(repeats),
ijl
is a randomvariable representing the observed response for therepeated measurement
l
by operator
j
on part
i
,
i
is a random variable representing the variation of the true dimension of part
i
,
ijl
is a random vari-able that represents the measurement error when thesame operator repeatedly measures the same part(i.e., measurement repeatability), and
μ
j
representsthe effect of operator
j
. We also include the randomvariable
PO
ij
to allow the operator effect to changefrom part to part, i.e., to allow for interaction be-tween the operator and parts. We make the addi-tional distributional assumptions that
i
N(0
,σ
2
 p
),
PO
ij
N(0
,σ
2
 po
),
ijl
N(0
,σ
2
m
) and that all of these random variables are independent.Many authors treat the operator effects as ran-dom. That is, they suppose the
μ
j
’s are a samplefrom a normal distribution. This assumption makessense when there is a large number of operators andonly a sample is available for inclusion in the assess-ment. In this instance, efficient plans require a largenumber of operators who may not be available inmost manufacturing settings. Here we consider onlythe fixed-effects case when a small number of opera-tors are part of the measurement system and are allincluded in the assessment study. Because the truepart value is a random effect, we adopt the tradi-tional approach of describing the possible part-by-operator interaction using a random effect. The in-teraction is quantified parsimoniously by the singleparameter
σ
 po
. Burdick et al. (2005) discuss the issueof fixed versus random operator effects and providean analysis for both cases. They also provide a largenumber of references.We define
σ
2
o
=
r
j
=1
(
μ
j
¯
μ
)
2
r
as in Burdick et al. (2005, p. 83). Note that
σ
2
o
quantifies the measurement variation due to the rela-tive biases of each operator (i.e., measurement repro-ducibility) but is not a variance in the usual sense.Similarly, we define and interpret the total variation
σ
2
t
=
σ
2
 p
+
σ
2
o
+
σ
2
 po
+
σ
2
m
. If the system is auto-mated with no operator effects or has a single opera-tor (
r
= 1), we have
σ
2
o
= 0 and we cannot estimate
σ
2
 po
separately from
σ
2
m
, so we also set
σ
2
 po
= 0 in thiscase. We also separately consider the case when thereis no part-by-operator interaction by setting
σ
2
 po
= 0.A common metric for assessing a measurementsystem compares the variation due to the measure-ment system (repeatability and/or reproducibility)to the overall variation due to the process (i.e., dueto differences in the true dimensions and the mea-surement system). We define the gauge repeatabilityand reproducibility (GR&R) (AIAG, 2003) as
γ 
=
 
σ
2
o
+
σ
2
 po
+
σ
2
m
σ
2
t
.
(2)According to AIAG (2003, p. 77), a measurementsystem is deemed to be acceptable if 
γ 
is less than0.1, unacceptable if 
γ 
is greater than 0.3, and is inneed of improvement if 0
.
1
γ 
0
.
3. If the estimateof 
γ 
is large, we can examine the estimates of 
σ
o
,
σ
 po
,and
σ
m
separately to identify the source of the largemeasurement-system variation.To estimate
γ 
, we need a plan that provides anestimate of the underlying process variation
σ
 p
. Al-ternate metrics, such as the precision to tolerance(PT) ratio, depend only on the measurement-systemvariation
 
σ
2
o
+
σ
2
 po
+
σ
2
m
and do not require an es-
Vol. 42, No. 4, October 2010 www.asq.or
 
390 STEVENS ET AL.
timate of 
σ
 p
. The optimal design of an assessmentstudy to estimate the PT ratio and other such met-rics will be different from what we propose. In othersituations, the goal may be to estimate the individualvariance components. This change of goal will lead todifferent assessment plans. In this paper, we focus onfinding good plans for estimating
γ 
while preservingsome information about the separate variance com-ponents.Other authors have compared balanced plans inthe context of measurement system-assessment. If the measurement system is automated or has a singleoperator, Shainin and others (Shainin (1992), Traver(1995)) recommend an Isoplot
®
study, where
k
= 30parts are selected and each is measured twice, i.e.,
n
= 2. The Shainin plan provides better balance be-tween the number of degrees of freedom available forestimating the measurement and process variation,and, as we shall see, is the optimal SP for estimat-ing
γ 
in this case. Walter et al. (1998) look at theoptimal standard plan for estimating the interclasscorrelation using power calculations. They consideronly plans with random operator effects that are sub-sumed into the repeatability component
σ
m
. Varde-man and Van Valkenberg (1999) look at the designof standard plans for
fixed with a flexible crite-rion that depends on which parameters are of mostimportance. They point out, somewhat facetiously,that if the goal is to estimate only the numerator of 
γ 
2
(i.e., the total variance attributable to the mea-surement system), as is required for assessment cri-teria like the PT ratio, then it is best to use a singlepart. They also note that, if operator effects are ran-dom, then the usual standard plan with
r
= 2 or3 cannot estimate
σ
2
o
with any usable precision. Wecan add the same comment about estimating
σ
2
 p
withrelatively few parts (i.e., 10), as is typical in practice.The reference manual of the AIAG (2003) on page 99notes that their suggested number of appraisers, tri-als, and parts “represents the optimal conditions forconducting the study”. We shall see that this state-ment is far from the truth if we use a criterion basedon the precision of the estimate for
γ 
.As noted by a referee, there is often available in-formation from process records that can be incor-porated into the analysis. For example, the GR&Ranalysis in Minitab
TM
(2007) allows the substitutionof an available estimate of the total variation. Browneet al. (2009a) and Danila et al. (2010), in the binarysituation, demonstrate the considerable value of thisextra information. In the current context, with fixedoperator effects, we may have additional process in-formation by operator, i.e., estimates of 
μ
j
for eachoperator (and hence
σ
o
) and
 
σ
2
 p
+
σ
2
 po
+
σ
2
m
. Thepresence of this information will change the recom-mended plans described below. We do not pursue thisissue further here.In this paper, we compare standard plans with twotypes of augmented plans in which not all parts aremeasured the same number of times. In all cases, thenumber of operators
r
is fixed. The augmented planshave two components. One component is a standardplan using
k
parts with
n
repeated measurementsby each operator. There are two possibilities for theother componentType A: Randomly sample
k
A
parts (different fromthose selected in the SP component) where
k
A
is a multiple of 
r
. Each operator mea-sures
k
A
/r
different parts once. We call thisan A plan, denoted by
A
(
k,n,k
A
).Type B: Randomly sample
k
B
parts (different fromthose selected in the SP component). Eachoperator measures each of these partsonce. We call this a B plan, denoted by
B
(
k,n,k
B
).Plan
A
(
k,n,k
A
) has a total of 
=
krn
+
k
A
mea-surements using
k
+
k
A
parts. Plan
B
(
k,n,k
B
) has atotal of 
=
krn
+
rk
B
=
r
(
kn
+
k
B
) measurementsusing
k
+
k
B
parts. If we set
k
A
or
k
B
to zero, the cor-responding augmented plan is an SP. Note that thesecond component of plan B corresponds to an SPwith
n
= 1. The two components of an augmentedplan can be conducted simultaneously or in any or-der. Within each component, every part is measuredthe same number of times.The goal of this work is to identify augmentedand standard plans that efficiently estimate
γ 
when
, the total number of measurements available, isfixed. We measure efficiency of any augmented planat a particular set of parameter values by comparingthe asymptotic standard deviations (not variance) of the maximum likelihood estimates of 
γ 
from the aug-mented plan relative to the best SP. Here “best”means the standard plan with the smallest asymp-totic standard deviation at the given parameter val-ues. We search for augmented plans that have effi-ciency greater than 1 over a whole range of valuesfor the unknown parameters.If we set
n
= 1 in an SP or in the SP componentof either plan A or B, then no part is measured more
Journal of Quality Technology Vol. 42, No. 4, October 201

You're Reading a Free Preview

Download
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->