Journal of Quality Technology

A Quarterly Journal of Methods, Applications and Related Topics

ISSN: 0022-4065 (Print) 2575-6230 (Online) Journal homepage: https://www.tandfonline.com/loi/ujqt20

Spatial Control Charts for the Mean

Scott D. Grimshaw, Natalie J. Blades & Michael P. Miles

To cite this article: Scott D. Grimshaw, Natalie J. Blades & Michael P. Miles (2013)
Spatial Control Charts for the Mean, Journal of Quality Technology, 45:2, 130-148, DOI:
10.1080/00224065.2013.11917922

To link to this article: https://doi.org/10.1080/00224065.2013.11917922

Published online: 21 Nov 2017.

Submit your article to this journal

Article views: 58

View related articles

Citing articles: 3 View citing articles

Full Terms & Conditions of access and use can be found at

https://www.tandfonline.com/action/journalInformation?journalCode=ujqt20
 Spatial Control Charts for the Mean
SCOTT D. GRIMSHAW, NATALIE J. BLADES, and MICHAEL P. MILES
Brigham Young University, Provo, UT 84602

Developments in metrology provide the opportunity to improve process monitoring by obtaining many
measurements on each sampled unit. Increasing the number of measurements may increase the sensitivity
of control charts to detection of flaws in local regions; however, the correlation between spatially proximal
measurements may introduce redundancy and inefficiency in the test. This paper extends multivariate sta-
tistical process control to spatial-data monitoring by recognizing the spatial correlation between multiple
measurements on the same item and replacing the sample covariance matrix with a parameterized covari-
ance based on the semivariogram. The properties of this control chart for the mean of a spatial process are
explored with simulated data and the method is illustrated with an example using ultrasonic technology to
obtain nondestructive measurements of bottle thickness.

Key Words: Hotelling T 2 Control Chart; Multivariate EWMA; Optimal Allocation of Sample Resources;
Semivariogram; Spatial Covariance Models.

1. Introduction gies: Megahed et al. (2011) review the use of a sin-

gle control chart from a machine-vision image of a
A DVANCES in metrology are changing the landscape
of industrial-process monitoring: machine-vision
systems, 3-d laser scanners, improved coordinate-
manufactured part. These spatial control charts have
been used to detect defects in woven fabrics, liquid-
crystal display (LCD) panels, and thin-film transistor
measuring machines, and ultrasound technology are
panels; however, while these charts scan across one
all facilitating faster, higher dimensional monitoring.
spatial dimension of an image, plotting a statistic for
While increasing the number of measurements taken
each slice or pixel, they do not account for the spatial
on a single item may increase the sensitivity of con-
correlation in the image.
trol charts to process aberrations, it also induces
a correlation between spatially proximal measure- When measurements are taken during process op-
ments. Multivariate statistical process control (SPC) eration, the opportunity to monitor a profile instead
typically uses Shewhart charts and multivariate ex- of a final quality measurement can provide insight
ponentially weighted moving average (EWMA) to into the nature of the out-of-control process opera-
monitor multiple quality characteristics simultane- tion. Woodall et al. (2004) present an introduction to
ously; however, these techniques, as shown in this profile monitoring, example applications, and litera-
paper, can be extended to account for spatial auto- ture review where profiles are modeled by a linear ap-
correlation when a single characteristic is measured proximation, nonlinear parametric model, wavelets,
at many locations on each sampled item. or splines; the text by Noorossana et al. (2011) pro-
Recent papers address the process-monitoring vides a more in-depth introductory treatment. These
challenges presented by many of these new technolo- methods often fit parametric models to the profiles
and monitor the coefficients of the model to deter-
mine outlying profiles. Thus, the data that form the
Dr. Grimshaw is Professor in the Department of Statistics.
profiles are reduced to a smaller set of values that
He is an ASQ member. His email address is grimshaw@byu
simplify the monitoring scheme. Zhang and Albin
.edu.
(2009) apply profile monitoring directly to the vector
Dr. Blades is Assistant Professor in the Department of assembled from measurements of a single character-
Statistics. Her email address is [email protected]. istic at many locations on a sampled item. While
Dr. Miles is Associate Professor in the Department of this avoids the difficult parametric summarization of
Manufacturing Engineering Technology. His email address is the multiple measurements and maintains the intu-
[email protected]. itive appeal of interpreting the actual measurements,

Journal of Quality Technology 130 Vol. 45, No. 2, April 2013

 SPATIAL CONTROL CHARTS FOR THE MEAN 131

the possible high dimension of p produces problems Monitoring the manufacture of plastic bottles
for the sample covariance matrix used in T 2 . Zhang demonstrates some of the issues introduced by metro-
and Albin (2009) ignore spatial or temporal correla- logic advances. Historically, control charts for bot-
tion by assuming independence and require estima- tle thickness required a destructive test measure-
tion of the diagonal elements of a potentially high- ment: a sample of bottles was dissected and cal-
dimensional covariance matrix. Ding et al. (2006) ad- lipered to measure wall thickness. Recent improve-
dress the high dimensionality by performing princi- ments in metrology allow a nondestructive thickness
pal components analysis (PCA) on the Phase I data measurement to be made at any selected location on
with the intention of finding a projection to a smaller the bottle using an ultrasonic micrometer with access
dimension that is sensitive to out-of-control profiles. to only one side of the material being measured. The
Chicken et al. (2009) use wavelets to reduce the high tradeo↵ between nondestructive evaluation and de-
dimension to a set of parameters that can be moni- structive testing typically involves a choice between
tored. the greater precision of destructive testing and the
lower cost of nondestructive evaluation (Reese et al.
Profile-monitoring papers have acknowledged cor- (2008)); however, the ultrasonic micrometer used to
relation in time or space within a profile may occur measure bottle thickness provides a nondestructive
because of the frequency of data collection. Staud- measurement with equivalent precision to the de-
hammer et al. (2007), Jensen et al. (2008), and structive measurement.
Jensen and Birch (2009) include correlation in the
This paper proposes a control chart for the mean
profile model and investigate the impact of correla-
when spatial data are collected for a control chart,
tion in simulation studies. Colosimo et al. recognize
such as the nondestructive thickness measurements
not only profile parameters for evaluating roundness
taken at many bottle locations. These control charts
(Colosimo et al. (2008)) or cylindricity (Colosimo et
are based on multivariate control charts with very
al. (2010)) of a lathe-turned object but that the pa-
large p. Recent advances in spatial data analysis al-
rameters of a spatial-correlation model must be in-
low for the design of control charts that take into
cluded in the T 2 statistic for profile monitoring. Qiu
account the spatial correlation structure where mea-
et al. (2010) combine flexible nonparametric estima-
surements that are close together are correlated. In
tion of profiles with a flexible correlation structure.
this paper, the spatial data are assembled into a
In their simulation study, they found that ignoring
vector and the process is monitored using a T 2 or
strong correlation in Phase II results in much shorter
EWMA statistic where the in-control covariance ma-
in-control average run length (ARL) values than the
trix is based on the semivariogram that represents
nominal value. Wei et al. (2012) suggest robust esti-
the spatial correlation. The proposed control charts
mation of Phase II profiles and derive the properties
are demonstrated on an example from bottle manu-
of their test statistics under a dependence condition
facturing. A simulation study indicates that account-
that implies the intraclass correlation decays expo-
ing for spatial correlation improves the ARL.
nentially.
2. Characteristics of Spatial Data
Additionally, SPC has been leveraged in biomed-
ical and epidemiologic applications with a spatial While the recognition of spatial e↵ects in statistics
component: Lindquist et al. (2007) use EWMA to de- dates to the earliest randomized plot designs for agri-
tect changepoints in voxel-wise functional magnetic cultural experiments, widespread application of mod-
resonance imaging (fMRI) data; they do not directly els to account for spatial correlation were computa-
account for spatial correlation but rather use a false tionally intractable for most of the 20th Century. As
discovery rate control to account for spatial correla- computational tools for modeling spatial data have
tion. Unkel et al. (2012) review the use of cumulative become more accessible, the use of spatial data has
sum (CUSUM) charts and scan statistics in public- expanded from geography to health, environmental
health surveillance systems for prospective identifi- science, ecology, and economics; however, the prac-
cation of disease clusters and monitoring of hospital tical use of spatial statistics has penetrated but not
performance. Diggle et al. (2005) detect changepoints permeated industrial applications. A brief introduc-
in inhomogeneous Poisson point processes where the tion to spatial data analysis is included below; for a
intensity of cases or events reflects spatial and tem- more complete treatment, Schabenberger and Got-
poral variation around the incidence rate. way (2005) and Cressie (1993) have written excellent

Vol. 45, No. 2, April 2013 www.asq.org

132 SCOTT D. GRIMSHAW, NATALIE J. BLADES, AND MICHAEL P. MILES

introductions and Kaluzny et al. (1998) and Bivand

et al. (2008) provide resources for computation.
To understand the potential impact of spatial
statistics in quality improvement, let Zj (si ) denote
the measurement of an important quality character-
istic on the jth sample at location si . In the bottle
measurement example in Section 4, si is a point in
three-space, but other applications such as measure-
ments on a semiconductor wafer would have locations
represented in two-space. The spatial process is typ-
ically denoted {Z(s) : s 2 D}, where the locations D
are a continuous fixed set, often called geostatistical,
but the control charts described in this manuscript
are also applicable if the data are collected on a lat-
tice of fixed discrete locations or random locations.
One of the characteristics of spatial measurements
is that the covariance between measurements at si
and si0 will depend on the distance between the lo-
cations. That is, observations close together will be FIGURE 1. Typical Shape of the Semivariogram. The sill
highly correlated while observations far apart will be denotes the variance of points sufficiently far apart to be
uncorrelated. uncorrelated.

In spatial data analysis, the covariance structure

for a stationary function Z(·) is specified by the semi- The diagonal elements of ⌃ are given by the sill,
variogram defined as ii = Var[Z(si )] = limd(si ,si0 )!1 (d(si , si0 )). The
o↵-diagonal elements are ii0 = Cov[Z(si ), Z(si0 )] =
1 (d(si , si0 )), where observations the same dis-
(d(si , si0 )) = Var[Z(si ) Z(si0 )] ii
2 tance apart have the same covariance. Estimation
= Var[Z(si )] Cov[Z(si ), Z(si0 )], of the semivariogram is difficult because it must be
positive definite. A number of simple models for the
where d(si , si0 ) measures the distance between mea-
semivariogram depend only on the distance between
surements. Samples that are close are expected to
si and si0 and guarantee positive definiteness.
yield similar values of Z(·); consequently, (d) will
be small when si and si0 are close, implying the The most common semivariogram models are the
covariance o↵sets the variance. As the distance be- exponential and the Matern, and both of these fam-
tween si and si0 increases, the spatial correlation ilies can be parameterized without or with a nugget
between them will decrease until the semivariogram e↵ect. The exponential model, the default semivari-
approaches (d(si , si0 )) = Var[Z(si )], called the sill. ogram model in geoR (Ribeiro and Diggle (2001)), is
The distance at which the semivariogram reaches the attractive because it has a simple form that defines
sill is called the range; when the distance between si a valid semivariogram at all possible distances; how-
and si0 is greater than the range, Z(si ) and Z(si0 ) ever, the semivariograms fit using the Matern fam-
are uncorrelated and the semivariogram is flat. If ily provide a useful and flexible quantification of the
the semivariogram reaches the sill only asymptoti- spatial dependence (Diggle and Ribeiro (2007)).
cally, the practical range is defined as the distance
at which the semivariogram reaches 95% of the sill. The exponential model is given by
The expected shape of the semivariogram is shown h i
in Figure 1; the semivariogram for observations that (dii0 ) = ⌧ 2 + 2 1 e dii0 /↵ ,
are far apart approaches the asymptote at Var[Z(si )]
and approaches zero for observations that are close where dii0 = d(si , si0 ) measures the distance between
together. two measurement locations, 2 = Var[Z(si )] or the
sill, and ↵ is a function of the practical range that
The semivariogram defines the covariance ma- scales the distance to represent the correlation. Fig-
trix ⌃ when the measurements at p locations are ure 1 was constructed from an exponential model
assembled as the vector Z = [Z(s1 ), . . . , Z(sp )]0 . with 2 = 0.03, ↵ = 3, ⌧ 2 = 0.

Journal of Quality Technology Vol. 45, No. 2, April 2013

 SPATIAL CONTROL CHARTS FOR THE MEAN 133

The semivariogram, (d(si , si0 )), in Figure 1 is forms. A common approach is to fit the exponential
zero as d(si , si0 ) approaches zero, but Cressie (1993) or Matern model either with or without a nugget ef-
points out that, in some spatial applications, there fect. The parameters of the semivariogram model are
is a discontinuity at the origin, called the nugget, estimated on the empirical semivariogram cloud.
reflecting that some sources of variation operate on
Third, with improvements in metrology making
distances smaller than the shortest distance sampled.
it easier and faster to obtain observations at many
For the exponential model, the no nugget e↵ect is
locations, the dimension p in the multivariate chart
specified by setting ⌧ 2 = 0 and the exponential model
will be large—probably larger than n, the number of
with a nugget e↵ect allows ⌧ 2 > 0, where ⌧ 2 is an es-
sampled products used to compute Z(si ) to monitor
timated parameter.
the process. If the measurements Zj (si ) are not nor-
The more flexible Matern semivariogram model is mally distributed, there will be no central limit e↵ect
given by on Z because n ⌧ p. Therefore, it is important that
 p the stationary Gaussian process assumption be in-
(2 dii0 /↵) p
(dii0 ) = ⌧ 2 + 2 1 K (2 dii0 /↵) , vestigated, with particular exploratory data analysis
2 1 ()
of the assumptions of normality, constant variance at
where (·) is the gamma function and K (·) is the all s, and isotropy, the dependence of the semivari-
modified Bessel function of order . The parameter ogram only on the distance between points not the
 allows for flexibility in the shape of the curve, with location of the points.
 = 1/2 reducing to the exponential model. As de-
scribed above, when ⌧ 2 = 0, there is no nugget e↵ect Creation of a control chart for the mean of a spa-
and, when ⌧ 2 is estimated, there is a nugget e↵ect. tial process is a straightforward application of a mul-
tivariate control chart: The mean of the j = 1, . . . , n
3. Control Charts for the Mean sampled products at each of the p locations is as-
of a Spatial Process sembled in a vector denoted by Z. The multivariate
control chart produces an out-of-control signal when
While the spatial process is typically denoted
{Z(s) : s 2 D}, the control chart for µ(s) = E(Z(s)) T 2 = n(Z µ0 )0 ⌃(✓0 ) 1
(Z µ0 )
is based on recognition that the measurements on
exceeds the upper control limit 2 (1/ARL0 ; p), where
the jth product can be assembled as a p ⇥ 1 vec-
ARL0 denotes the in-control ARL and ✓0 holds
tor as in Zhang and Albin’s (2009) profile-monitoring
the parameters of the semivariogram; for example,
approach. That is, Zj = [Zj (s1 ), . . . , Zj (sp )]0 . We as-
✓0 = ( 02 , ↵0 ) for the exponential without a nugget
sume Zj is a stationary Gaussian process with mean
e↵ect semivariogram. Because the in-control covari-
E(Z(s)) = µ(s). When the process is in control, the
ance matrix is a function of the parameters ✓0 and
spatial process has constant mean, µ(s) = µ, for all
the distances dii0 , the spatial control chart T 2 can
s, but when the process is out of control, the mean
be computed even if the spatial measurement loca-
µ(s) 6= µ for at least one location s. While based
tions and/or the number of spatial measurements p
on a multivariate control chart, the control chart for
change. For example, if, after obtaining the Phase I
the mean of this spatial process will di↵er in three
samples, a di↵erent spatial grid density is chosen for
important ways from a typical multivariate control
Phase II, there is no need to repeat Phase I sampling
chart.
for the spatial T 2 , as would be required for the reg-
First, the in-control mean vector µ0 = [(E(Zj (si ))] ular T 2 . Here the control-chart statistic is denoted
describes a single quality characteristic measured by T 2 even though the covariance matrix is based on
at di↵erent locations. This interpretation simplifies the parameters of the semivariogram instead of S,
much of the complexity that arises from the typical the sample covariance matrix. While the in-control
application where di↵erent measurement characteris- mean may be known from the design specifications,
tics are combined. When diagnosing an out-of-control in most applications, the semivariogram will be esti-
signal, the problem simplifies to identifying the col- mated from a Phase I sample of in-control products
lection of locations that depart from specifications. resulting in ⌃(✓ˆ0 ). If µ(s) is unknown, the estimated
spatial mean µ̂(s) from a Phase I sample could be
Second, the covariance matrix ⌃ is determined
used to construct the vector µ0 .
by the semivariogram that models the spatial cor-
relation. While the semivariogram need not be pa- Alternatively, the ideas in the multivariate
rameterized, there are several common parametric EWMA (Lowry et al. (1992)) can be modified to de-

Vol. 45, No. 2, April 2013 www.asq.org

134 SCOTT D. GRIMSHAW, NATALIE J. BLADES, AND MICHAEL P. MILES

fine a spatial multivariate EWMA control chart that thickness when manufacturing Boston round bot-
should detect small and medium shifts that persist tles, basic round-shaped bottles commonly used for
over time. If Zt denotes the sample mean vector at soft drinks and other beverages, water bottles, phar-
time t, a weighted average is formed by maceuticals, cosmetics, and hair-care products when
paired with regular or flip-top caps, sprayers, and
Wt = (Zt µ0 ) + (1 )Wt 1,
lotion-pump lids. Nondrink bottles are often made
where W0 = 0 and the choice of determines the using a continuous extrusion method, where the pari-
amount of smoothing (0 < < 1). Small values of son is made by extruding a cylindrical tube that is in-
pool the data over a wide time interval and pro- troduced directly from the extrusion part of the blow
duce a control chart that e↵ectively identifies small, molding machine into the mold. The mold closes and,
persistent changes from µ0 . Large values of heav- as it does, clamps the bottom and top of the cylin-
ily weight the current observation, yielding a control der (except for a small inlet to allow air to be intro-
chart instantaneously sensitive to large shifts from duced). Pressurized air inflates the preheated parison
µ0 . The multivariate EWMA chart signals a process inside the mold and then the cool walls of the mold
is out of control at time t when cause the bottle to solidify. The mold is opened and
the bottle is ejected.
Qt = Wt0 ⌃W1 Wt
While blow molding is a robust, high-volume man-
exceeds an upper control limit. The variance-
ufacturing process, there are several factors that
covariance matrix ⌃W depends on the semivari-
can a↵ect bottle-wall thickness during production,
ogram, , and t and is given by
✓ ◆ nonuniformity of parisons and nonuniformity over
[1 (1 )2t ] 1 time of the air pressure, for example. The customer
⌃W = ⌃(✓0 ),
2 n will recognize cosmetic defects in bottles that are not
nearly uniform in bottle thickness because shrinkage
where ⌃(✓0 ) is constructed from the semivariogram
stresses after hot filling can exploit a thin spot, caus-
as in the T 2 statistic. There is the potential that
ing the surface of the bottle to curve inward.
the multivariate EWMA is more robust to nonnor-
mality than the T 2 , as found in Testik et al. (2003), Bottle-production quality is monitored or con-
but the issue merits further study because spatial trolled by taking a sample of n bottles at regular
control charts could have much larger p than they time intervals, and, on each product, making p mea-
investigated in their simulation study. surements at di↵erent spatial locations. This exam-
ple is based on quality control of 32-ounce Boston
When either the T 2 or Qt produces an out-of-
round bottles, which, if the thread top and bottom
control signal, at least one of the physical loca-
are ignored, are basically 6-inch-tall cylinders with a
tions departs from the specifications. However, the
circumference of 12 inches. Thickness, using a nonde-
out-of-control diagnostic does not stop there, as it
structive ultrasound micrometer, is measured at the
is possible to construct an interpolation to loca-
same locations on each sampled bottle. As these bot-
tions where measurements were not taken because
tles are blow molded by expanding a preform against
of the spatial correlation. Ordinary kriging is the
a mold, the distance metric used for the spatial cor-
minimum mean-square error prediction given the ob-
relation is arc length (the natural distance metric for
served Z(s1 ), . . . , Z(sp ). That is, for any location s,
points on the surface a cylinder). Each bottle has
p
X two ‘parting lines,’ which are the observable seams
Ẑ(s) = wi (s)Z(si ), where the two halves of the mold join together. The
i=1 two parting lines are used to consistently orient the
where the wi (s) are explicit functions of the semi- bottles for measurement.
variagram (d). Therefore, an out-of-control signal
Let Zj (si ) denote the bottle thickness of the jth
can be diagnosed with a graphic representing mea-
bottle at location si . The three-dimensional measure-
surements at all locations.
ment locations are represented more simply in two-
space by imagining an ‘unwrapped’ cylinder. While
4. A Spatial Control Chart
actually doing so would be wasteful, visualize cut-
for Monitoring Bottle Thickness
ting the bottle on one of the parting lines as shown
To demonstrate the application of a spatial control in Figure 2, unwrapping the bottle at the cut, and
chart, consider the problem of monitoring bottle-wall laying it flat. Then the measurement locations, rep-

Journal of Quality Technology Vol. 45, No. 2, April 2013

 SPATIAL CONTROL CHARTS FOR THE MEAN 135

FIGURE 2. Two-Space Representation of the Bottle Surface. While a Boston round bottle is a three-dimensional object, the
use of arc length as a distance metric renders the problem simpler: ‘unwrapping’ the bottle yields measurements in two-space
instead of three-space.

resented by dots on the grid in Figure 2, are locations ment grid (four locations around the cylinder at two
in two-space on a 6-inch by 12-inch rectangle. One of heights, p = 8). In a departure from traditional Phase
the parting lines is in the center of the unwrapped I sampling, intended to provide insight into choos-
cylinder (parting line B) but the other, where we ing the semivariogram model and to improve semi-
imagined cutting the bottle, is at both the far left variogram estimation by having observations at dis-
and far right (parting line A). While it appears the tances near zero, data were collected on two other
right and left boundaries of the unwrapped cylinder days where one bottle each day was measured on two
are borders, the cylinder is a closed surface and these di↵erent denser spatial measurement grids (a 5 ⇥ 5
boundaries are actually neighbors. This must be ad- grid with locations 0.25 inches apart one day and a
dressed when computing the arc-length distances be- 5 ⇥ 5 grid with locations 0.125 inches apart the other
tween spatial measurement locations and explains day) but at only two primary locations (one on the
why no two points would be separated by a distance parting line and one halfway between parting lines);
as large as the 12-inch rectangle width. this yields data on m · n + 1 + 1 = 26 bottles at three
di↵erent spatial measurement grids.
The first step in creating a spatial control chart
is to estimate the in-control semivariogram from a The semivariogram cloud is constructed in Figure
Phase I sample. Historical data that displayed no 3. The semivariogram cloud is a plot of the squared
signs of flaws or defects were selected by operators pairwise di↵erences (1/2)[Z(si ) Z(si0 )]2 against the
from di↵erent production days. While these sampled arc distance dii0 = d(si , si0 ) for each pair of points
bottles were not a random sample, they represent on bottle (Diggle and Ribeiro (2007)). The Z(si ) are
in-control process operation. The 32-ounce Boston the log-transformed thickness measurements. While
round bottles in this example are designed to have parameter estimates are computed on the large num-
uniform thickness of 0.055 inch. These measurements ber of pairs of points, the plot creates small bins
are skewed, but the log transformation yields data and plots the mean in each bin (to demonstrate the
that are approximately normally distributed and ap- variation within each bin, the error bars denote the
pear to satisfy the assumptions of constant variance first and third quartile). The parameters of the ex-
at the same measurement location between bottles ponential semivariogram model without a nugget ef-
and constant variance at di↵erent locations. fect, (dii0 ) = 2 [1 e dii0 /↵ ], are estimated using
n-weighted least squares and a numerical minimiza-
To estimate the semivariogram, pairs of measure- tion algorithm. Cressie (1993) describes how a sim-
ment locations must be chosen to span distances from ple nonlinear least-squares estimator would ignore
near zero to the range. As shown in a later simu- the fact that the variance is a function of distance
lation, the semivariogram parameters are well esti- (notice in Figure 3 that small d have smaller error
mated when the Phase I sample is as large as re- bars than larger d), and a weighted least-squares ap-
sources permit. There were m = 4 days where n = 6 proach could improve model fit by allowing di↵erent
bottles were measured on a coarse spatial measure- variances. The weights are 1/[ (dii0 )]2 and the algo-

Vol. 45, No. 2, April 2013 www.asq.org

136 SCOTT D. GRIMSHAW, NATALIE J. BLADES, AND MICHAEL P. MILES

FIGURE 3. Semivariogram Cloud Where All Squared

Pairwise Di↵erences Are Plotted Against the Arc-Length FIGURE 4. Hotelling T 2 for Spatial Control Chart of
Distance. The solid line is the estimated semivariogram (d) Four Days of In-Control and Four Days of Out-of-Control
based on the exponential model without a nugget e↵ect. The Process Operation. Comparisons include the spatial control
plot suggests there is spatial correlation of bottle thickness chart using the exponential without a nugget e↵ect (black
at di↵erent measurement locations on the same bottle. The solid line), the Matern without a nugget e↵ect (black dotted
dotted line denotes the Matern model without a nugget ef- line), the regular T 2 based on S (gray solid line) with Phase I
fect to estimate (d). and Phase II control limits, and the naive T 2 ignoring spatial
correlation (Zhang and Albin (2009)) (black dash-dot line).

rithm must iterate to find the weighted least-squares

uniform thickness. In terms of the in-control mean
estimates. Cressie (1993) points out that the obvious
vector µ0 = 2.9 · 18 , where 18 denotes an 8 ⇥ 1 col-
correlation between the squared pair-wise di↵erences
umn vector of ones. In other applications, the bottle
is ignored in the weighted least-squares algorithm for
design may have specifications where the thickness
computational simplicity, but specifying a full V ma-
depends on s for decorative reasons or to enhance
trix could be performed if desired. There is no evi-
structure and reduce cosmetic defects like collapses
dence of a nugget e↵ect because the empirical semi-
or bulges after filling and, in those cases, the µ0 could
variogram cloud approaches zero for small distances.
come from the specifications or be estimated from
Based on these Phase I sample bottles, the spatial
Phase I sampling.
control chart will use the covariance based on the
exponential model without a nugget-e↵ect parame- The in-control covariance matrix ⌃0 is an 8 ⇥ 8
ter: ˆ 2 = 0.034 and ↵ˆ = 3.17 (solid line overlaid in matrix computed from the semivariogram. The ⌃0 =
Figure 3). ⌃(0.034, 3.17) reflects the e↵ect of spatial correlation
While future control charts may investigate the where the measurement locations that have small dis-
bottle thickness at more measurement locations, a tances are more correlated than observations with
primary concern of the manufacturer was behavior larger distances.
on the parting lines and so the coarse spatial mea-
Figure 4 plots the spatial control chart T 2 for
surement grid with p = 8 nondestructive bottle-
the m = 4 days of Phase I sampling, where p = 8,
thickness measurements are taken at the same loca-
and four days when later testing indicated a flaw in
tion as the m = 4 days (on the parting lines and mid-
the molds that created unevenness along the parting
way between the parting lines at two heights, near
lines. The upper control limit for the spatial control
the top and near the bottom).
chart T 2 is 2 (0.01, 8) = 15.51. Notice that all the
The in-control mean is µ(s) = µ = ln(0.055) ⇡ days of out-of-control process operation are correctly
2.9 for all s because the bottle specifications are a detected by the spatial control chart.

Journal of Quality Technology Vol. 45, No. 2, April 2013

 SPATIAL CONTROL CHARTS FOR THE MEAN 137

This holds true even when the Matern model,

which is more flexible than the exponential, is es-
timated for the semivariogram. That is, while there
is a small change in the Matern fitted semivariogram
(dotted line in Figure 3), which better estimates the
semivariogram for small d, the 4 in-control days have
nearly identical control chart statistics and the out-
of-control days have slightly lower T 2 values (black
dotted line in Figure 4).
Because the spatial measurements are arranged
into a vector, one might consider applying the regu-
lar Hotelling T 2 , where the covariance is estimated
by S, the mean of the sample covariance matrices. In
contrast to the spatial control chart that estimates
a few parameters of a semivariogram, S estimates
p(p 1)/2 parameters. When m, the number of Phase
I samples is small, the control limits must be ad-
justed to recognize the impact of estimating so many
parameters. While the constants in the Phase I and
Phase II upper control limits di↵er, both are based FIGURE 5. Spatial Interpolation from the Eight Measure-
on the F distribution. To obtain a positive denom- ment Locations from the Bottles Sampled on Day 7, Which
inator degree of freedom, m must be greater than Was Identified by the Spatial Control Chart as Out of Con-
(p 1)/(n 1). In this application, with a coarse spa- trol. Grayscale plot of Ẑ(s) for all measurement locations,
tial measurement grid, the m = 4 Phase I samples not just the eight observed. The dark and light clusters rep-
are sufficient but, in a denser grid case with measure- resent departures from uniform bottle thickness.
ments taken 1.5 inch apart (p = 40), the regular T 2
could not be applied with m = 4 because a minimum
of m = 8 is required. The regular T 2 is denoted by a upper control limit, consider the spatial interpola-
gray solid line on Figure 4. Notice that the upper con- tion of bottle thickness using ordinary kriging from
trol limits di↵er for Phase I and Phase II. With the the bottles sampled on day 7. One of the advantages
adjustment to the usual control limit, all the Phase of accounting for the spatial correlation is that bot-
I bottles are correctly identified as in control. How- tle thickness can be interpolated using simple kriging
ever, two of the four out-of-control bottles fall under to measurement locations other than those that were
the Phase II control limit. part of the data collection. Figure 5 plots the bottle
The naive multivariate control chart that ignores thickness according to measurement location on an
the spatial correlation proposed by Albin and Zhang ‘unfolded’ bottle, where the vertical lines on the plot
(2009) would use ⌃0 with o↵-diagonal elements zero represent the locations of the parting lines. The dark
(assuming uncorrelated) and the variances computed cluster at the top left and right of the plot indicates
from the historical data by pooling the variances the bottle thickness exceeds the specifications at the
from the repeated observations at each measurement top of one of the parting lines and the light cluster
location ( 2 = 0.009). The result is a naive T 2 mul- at the bottom center of the plot indicates the bot-
tivariate control-chart statistic that is artificially in- tle thickness is lower than the specifications at the
flated and leads to a false alarm in the in-control bottom of the other parting line. While greater wall
sample, as shown by the gray dotted line in Fig- thickness may not be a problem, wall thickness be-
ure 4. low a certain level is a flaw that will likely lead to a
collapsed bottle after hot filling.
Not only can the spatial control chart do a better
job than the regular T 2 and the naive T 2 at identi- 5. Simulation Study to Investigate
fying in-control and out-of-control bottles, but spa- Practical Guidelines
tial statistics can also help to diagnose the nature
of the flaw. Returning attention to the days of out- To illustrate some of the practical issues in ap-
of-control process operation where T 2 exceeded the plying control charts to spatial data, this section

Vol. 45, No. 2, April 2013 www.asq.org

138 SCOTT D. GRIMSHAW, NATALIE J. BLADES, AND MICHAEL P. MILES

presents a simulation study based on the features of failure. Two sizes of circular flaws are investigated:
the bottle-thickness application in the previous sec- a local or small radius flaw (1.5 inch), where only a
tion. Data are generated from a two-dimensional sur- few locations will be out of control, and a regional or
face based on the ‘unwrapped cylinder’ with height large radius flaw (3 inch), where a large number of
6 inches, circumference 12 inches. The arc length is locations will be out of control.
the distance metric, and the distances are adjusted
Two spatial measurement grid densities are ex-
because the ‘unwrapped cylinder’ is a closed surface.
plored: a coarse grid, with 3 inches between points,
The data follow a stationary Gaussian spatial model.
and a fine grid, with 1.5 inch between points. The
When the process is in control, the mean µ0 = µ(s) =
spatial measurement grid is oriented to include the
2.9 for all s matches the uniform bottle thickness
parting lines. While circular flaws could occur any-
specification from the previous section. ARL curves
where on the bottle, in the simulation, the flaw lo-
are constructed on a grid of = |µ µ0 |/ , where
cation is centered at one of the spatial measurement
is constant for all spatial measurements in the flawed
grid locations.
region. The spatial correlation is modeled by the ex-
ponential semivariogram without a nugget e↵ect with Figure 6 displays the ARL curves for the scenar-
parameter values similar to those from the example ios defined above. In each of the six plots, the ARL
in the previous section ( 2 = 0.003 and ↵ = 3 unless performance improves as the degree of spatial corre-
the simulation explores the e↵ect of spatial correla- lation, ↵, increases. To understand how spatial corre-
tion by varying ↵ or by specifying other semivari- lation can improve a control chart, consider the two
ogram models). Spatial measurements will be taken spatial measurement locations (A and B) presented
on an evenly spaced grid or lattice. All results were in Figure 7. Notice that, in the left panel, the control
obtained from 100,000 simulations unless indicated limit for ↵ = 1 is nearly circular, as expected when
otherwise. the two spatial locations are nearly independent but,
in the right panel for ↵ = 4, the control limit is a
5.1. E↵ect of Spatial Correlation tilted ellipse, with the tilt reflecting the spatial cor-
The degree of spatial correlation is controlled by relation. As is well known from multivariate analysis,
varying ↵ in the semivariogram. Small values of ↵ the point plotted with ‘*’ will be an out-of-control
result in a covariance matrix close to the scaled iden- signal only for the case where the spatial measure-
tity matrix, 2 I, while large values of ↵ reflect high ments are correlated.
correlation even for measurement locations on op-
When the out-of-control distribution has a shift in
posite ends of the bottle. This simulation examines
the mean at location A but the mean at location B
values of ↵ = 1, 2, 3, 4, corresponding to vanish-
remains in control, it appears the high spatial corre-
ing correlation between points separated by at least
lation (right panel) has more probability outside the
3, 6, 9, or 12 inches, respectively. On the 32-ounce
control limit than the nearly independent case (left
Boston round bottle with measurements taken ap-
panel). This behavior corresponds to the ARL curves
proximately 3 inches apart, measurements from these
for local circular flaws (center column) in Figure 6 be-
four cases demonstrate no spatial correlation, cor-
cause only a few spatial measurement locations will
relation between observations on the same side at
be out of control.
approximately the same height, correlation between
observations on opposite sides but near the same The most dramatic improvements in the ARL
height, and correlation between observations on op- curves occur for large circular flaws (right column)
posite sides and di↵erent heights. and the denser spatial grid (bottom row). This is be-
cause there are many out-of-control locations, and
Two types of flaws common to blow molding are
one can imagine that, if both locations in Figure 7
investigated: parting line defects and circular defects.
had mean shifts, the density is mostly beyond the
Parting line defects occur when the seams from join-
control limit.
ing the two-part mold are thicker or thinner than
specifications. The parting line region runs the length These simulations also demonstrate the benefit
of the bottle but it is narrow because it is restricted of using a finer spatial measurement grid: in each
to the width of the seam. Circular defect regions oc- situation, the out-of-control ARL is demonstrably
cur when a flaw in the blow-molding process pro- shorter when the fine spatial measurement grid is
duces a deviation in bottle thickness that may lead used. While this improvement is expected because
to bottle collapse after filling due to stress material more measurement locations will fall within the de-

Journal of Quality Technology Vol. 45, No. 2, April 2013

 SPATIAL CONTROL CHARTS FOR THE MEAN 139

FIGURE 6. ARL Curves Demonstrating the E↵ect of Correlation Between Spatial Measurements. Small ↵ corresponds
to no correlation between observations close together and large ↵ corresponds to high correlation between all measurement
locations on a cylinder. Higher spatial correlation results in improved out-of-control ARL curves, with more dramatic impact
on denser spatial measurement grids and smaller, local flaws.

FIGURE 7. Control Limits for Two Measurement Locations. A mean shift at location A yielding log thickness measurements
of 3.0 and 2.8 at the point labeled ‘*’ will be detected when the two measurement locations are spatially correlated (right
panel) but not when the two measurement locations are independent (left panel).

Vol. 45, No. 2, April 2013 www.asq.org

140 SCOTT D. GRIMSHAW, NATALIE J. BLADES, AND MICHAEL P. MILES

FIGURE 8. ARL Curves for Choices of n and p Corresponding to 30 Minutes of Data Collection.

fective region, it is interesting that parting-line de- grid as there will be several spatial measurement lo-
fects are more frequently detected than circular flaw cations that have out-of-control measurements, and
regions—particularly as the narrow defect region for so the improvement in ARL will come from taking a
the parting line will contain fewer measurement lo- larger sample size.
cations. This improvement comes at a cost: the dis-
tance between points is halved when moving from Small local flaws are the most difficult flaws to
the coarse to the fine spatial measurement grid but detect because even a dense measurement grid may
the dimension of p increases more than three-fold, have only a few measurement locations within the
from 12 to 40. With n = 6 for all simulations in this flaw region: the three ARL curves in the center panel
section, the change in grid density results in a signif- of Figure 8 demonstrate considerably worse perfor-
icant di↵erence in the amount of work required for mance than for parting-line and large circular flaws.
data collection, which is further explored in the next One might expect that small local flaws would be
section. best detected by the allocation scheme with a dense
grid but a small sample size, but this is actually
5.2. Allocation of Sampling Resources the poorest choice: apparently, even a large number
When evaluating the work of data collection for a of measurement locations within the out-of-control
spatial control chart, the efficient allocation of sam- region will not o↵set the small sample size. While
pling resources demands a tradeo↵ between sampling the coarse grid with large n is a reasonable choice
many bottles (n) or many locations (p). The total for local flaws, a compromise between many coarsely
sampling time is n(⇠o + ⇠1 p) when it takes ⇠o seconds sampled bottles and few densely sampled bottles im-
to select a new bottle and prepare it for measure- proved the detection of small flaws: the best ARL
ment and ⇠1 seconds to move to a new location and performance occurred with n = 7 samples with mea-
measure the bottle thickness. surements every 1.5 inches (yielding five measure-
ment locations within the flawed region).
Figure 8 presents the simulated ARL curves for
possible data-collection strategies that could be exe- While Reynolds and Stoumbos (2004) recognized
cuted in approximately 30 minutes when ⇠0 = 60 and that efficient sample allocation requires a tradeo↵ be-
⇠1 = 5. At one extreme, a large sample size (n = 15) tween frequent sampling and sample size, this trade-
is paired with a coarse measurement grid (3 inches, o↵ is transformed when many spatial measurements
p = 12), and, at the other extreme, a dense mea- are collected on each item. In practice, the optimal
surement grid (1 inch, p = 84) is paired with a small allocation scheme depends on the type of defects
sample size (n = 4). It appears that, for the parting- that are of concern: the best performance is obtained
line flaw (left panel) and the large circular flaw (right when the measurement grid is chosen to cover the
panel), the best data-collection strategy takes many locations and types of anticipated defects and then
bottles on a coarse grid. This is perhaps not surpris- the sample size maximized subject to resource con-
ing because a large flaw will not be missed by a coarse straints.

Journal of Quality Technology Vol. 45, No. 2, April 2013

 SPATIAL CONTROL CHARTS FOR THE MEAN 141

have similar variance to the without-nugget counter-

parts.
Figure 10 presents the ARL curves for the correct
semivariogram model when µ is known. The results
are similar when µ must be estimated, with a slight
bias toward lower ARL values, as might be expected
with the estimation of an additional parameter.
While m = 1 may be adequate, there is a clear
bias toward false alarms (dotted lines in Figure 10).
This is somewhat surprising, as replacing unknown
values with parameter estimates was shown to in-
crease in-control ARL for the X-chart (Quesenberry
(1993)) and for the Hotelling T 2 chart (Champ et al.
(2005)). Larger m results in a movement toward the
specified in-control ARL, but it appears to depend on
the semivariogram model and grid density whether
FIGURE 9. Semivariograms Used in Simulation Study m = 4 is sufficient. For the Matern without a nugget,
Where Parameters Are Estimated from Phase I Sampling. both grid densities appear adequate with m = 4 but
only the 3-inch grid density is adequate at m = 4 for
the exponential with nugget and the Matern with
5.3. Estimation of Semivariogram in Phase I nugget. The di↵erent results for each semivariogram
Sampling model indicate further research is needed to identify
Finally, consider the impact on Phase II sampling a ‘small sample’ Phase II control limit that would
of estimating the parameters for the spatial con- perform better than the 2 control limit. The practi-
trol chart through Phase I sampling. Jensen et al. cal advice in choosing the Phase I sample data would
(2006) review the e↵ect of replacing unknown pa- be to choose m as large as reasonable.
rameters with estimates from Phase I sampling for
It will be surprising to some that, even though in-
a wide collection of control charts. For the spatial
creasing the density of the spatial measurement grid
control chart, the mean parameter may be given by
results in more data, the Phase II ARL curves are
the product specifications but the parameters of the
similar for the coarse and dense cases. While Zimmer-
semivariogram (d) that capture the spatial corre-
man and Zimmerman (1991) observed an improve-
lation would commonly be estimated from Phase I
ment in semivariogram parameter estimates with a
sampling.
more dense spatial grid in a simulation study, later
Simulated Phase II ARL curves are constructed to research showed that in-fill asymptotics (where more
detect a small radius flaw (1.5 inch) in two situations: observations are taken more densely in a fixed do-
when µ is known from specifications but (d) is esti- main) are quite complicated. Under in-fill asymp-
mated from Phase I sampling and when µ and (d) totics, the parameter estimates for the exponential
are both estimated from Phase I sampling. The ARL semivariogram are not consistent (Ying (1991)) nor
curves are based on 1,000 simulations at each . are the parameter estimates for the Matern semi-
The semivariogram parameters are estimated from variogram (Zhang (2004)); however, the parameter
m Phase I in-control samples ( = 0) of n = 6. estimates of the exponential semivariogram with a
Two Phase I sample sizes, m = 1 and m = 4, and nugget e↵ect are consistent and asymptotically nor-
two spatial measurement grids, a coarse 3-inch grid mal (Chen et al. (2000)) though the rate of conver-
(p = 12) and a dense 1.5-inch grid (p = 40), are gence is N 1/4 , which is much slower than the usual
explored. Four common semivariogram models are N 1/2 . Zhang (2004) saw no significant decrease in
investigated. Figure 9 plots the four models: the ex- the variance of the parameter estimates when a dense
ponential in black and the Matern in gray—each with grid was used. Therefore, the practical guideline is to
(dotted) and without (solid) the nugget e↵ect. The choose a sufficiently dense spatial measurement grid
cases without a nugget e↵ect are similar to the fits to detect local defects, but to otherwise prefer larger
to the semivariogram cloud for the example from the m in Phase I sampling during data collection over
previous section and the cases with a nugget e↵ect increased spatial measurement grid density.

Vol. 45, No. 2, April 2013 www.asq.org

142 SCOTT D. GRIMSHAW, NATALIE J. BLADES, AND MICHAEL P. MILES

FIGURE 10. Phase II Sampling ARL Curves Demonstrating the E↵ect of Parameter Estimation for Di↵erent m (Phase I
Sample Size) and Spatial Measurement Grids for Each Semivariogram Model.

Finally, consider the impact of misspecifying the the exponential is a special case of the Matern. Con-
semivariogram model. In Figures 11–15, the black sider the semivariogram cloud of a typical realization
solid line denotes the true model, the gray solid line of the simulation for the dense grid with m = 4 in
denotes correct family specification but misspecifi- Figure 12 to investigate the estimated Matern model.
cation of the nugget e↵ect (estimating the nugget Instead of the Matern estimating ̂ close to 1/2 (the
when there is no nugget e↵ect or modeling without special case of the exponential), a larger ̂ value
a nugget when there is a nugget e↵ect), the black results in a lower estimate of Var[Z(si )], which bi-
dotted line denotes misspecification of the family (es- ases the control chart statistic to have frequent false
timating Matern when true model is exponential or alarms. The poor Matern fit is because the spatial
estimating exponential when true model is Matern) measurement grid does not provide sufficient data
but correct specification of the nugget e↵ect, and the for small distances. If a nugget e↵ect is erroneously
gray dotted line denotes misspecification of both the estimated, the Phase II ARL is reasonable, but the
family and the nugget e↵ect. reasons di↵er for the exponential and Matern cases.
For the exponential, the estimated nugget e↵ect is
When the true semivariogram is exponential with- near zero, and there is a small cost in ARL due to
out a nugget (Figure 11), the results are fairly robust the estimation of an additional parameter. As shown
to model misspecification when the coarse measure- in Figure 12, it is interesting that fitting the Matern
ment grid was used. However, extremely poor perfor- with a nugget results in poor estimates of the correct
mance occurs if the Matern without a nugget is esti- parameters (non-zero nugget and  6= 1/2) but rea-
mated using a dense grid. This is surprising because sonable estimates of the spatial covariances. Because

Journal of Quality Technology Vol. 45, No. 2, April 2013

 SPATIAL CONTROL CHARTS FOR THE MEAN 143

FIGURE 11. Exponential Semivariogram Phase II Sampling ARL Curves Demonstrating the E↵ect of Parameter Estimation
for Di↵erent m (Phase I Sample Size) and Spatial Measurement Grids. The regular T 2 using S, which can only be applied
in the m = 4 and 3-inch grid case (upper right panel) because the dimensionality exceeds m in other cases, and the
naive T 2 ignoring spatial correlation (Zhang and Albin (2009)) are included for comparison. The e↵ect of misspecifying the
semivariogram depends on both m and spatial measurement grid.

the control chart relies on spatial covariances and false alarms. The bias is worst when the spatial mea-
not parameters, the control chart performs well even surement grid is dense and the Matern without a
if the semivariogram parameters don’t correspond to nugget model is estimated.
the true model.
If the Matern without a nugget e↵ect is the true
When the true semivariogram is exponential with model (Figure 14) but a nugget e↵ect is estimated,
a nugget (Figure 13), the Phase II ARL curves are the gray solid line in the lower left panel shows the
similar when the nugget e↵ect is correctly specified. Phase II ARL is positively biased; the bias is more
If the nugget e↵ect is ignored, there is a bias toward severe for small but acceptable for large . Mis-

Vol. 45, No. 2, April 2013 www.asq.org

144 SCOTT D. GRIMSHAW, NATALIE J. BLADES, AND MICHAEL P. MILES

a semivariogram with more complicated shape and

nugget e↵ect.
The performance of two alternatives to the pa-
rameterized covariance matrix are included in this
simulation study. The regular T 2 uses the sample
covariance matrix S and can be thought of as esti-
mating each spatial covariance without a parametric
model or borrowing information from data separated
by the same spatial distance. One of the frustrations
in applying the regular T 2 is because, in the Phase I
sample, size m must be sufficiently large to o↵set the
dimensionality p in order for the denominator degrees
of freedom in the F used for the Phase I and II crit-
ical values to be positive. For the cases investigated
in Figures 11–15, the regular T 2 can only be applied
in the m = 4 and 3-inch grid case (upper right panel)
because the other cases yield mn m p + 1 < 0
for denominator degrees of freedom in the F used
to compute the critical value. Even if it could be
FIGURE 12. Exponential Semivariogram Cloud for a Typ- applied, the regular T 2 has a much larger than nom-
ical Simulation for a Phase I Sample with m = 4 and 3-inch inal in-control ARL value and poor ARL for out-
Density. Instead of the estimated Matern fitting the expo- of-control cases, regardless of the spatial correlation
nential special case, without data at distances near zero the model.
Matern underestimates Var[Z(si )], which biases the control The other alternative is the control chart proposed
chart statistic to have frequent false alarms. Surprisingly, by Zhang and Albin (2009). This naive T 2 assumes
the Matern with nugget yields reasonable spatial covariance independence between spatial locations because the
estimates with poor parameter estimates. o↵-diagonal elements of ⌃0 are zero. While the Phase
II ARL curves indicate the naive T 2 could be applied
specification of the family by fitting the exponential in some applications with a coarse spatial measure-
(with or without a nugget e↵ect) leads to an unac- ment grid, this control chart should be avoided for
ceptably conservative control chart for dense grids. denser grids when the spatial correlation has a more
This was expected because, while the exponential is obvious presence.
a special case of the Matern, when the Matern has In conclusion, there is great benefit to taking as
a nonexponential shape, the Var[Z(si )] is overesti- large a Phase I sample of m as reasonable. While in-
mated and the spatial covariance estimates from the creasing the spatial grid density provides more data,
exponential are poor. the in-fill asymptotics indicate it is better to have
large m. When the spatial measurement grid that
When the true model is Matern with a nugget
can detect the desired local flaws is rather coarse,
(Figure 15), on a coarse grid the curves suggest ro-
there appears to be little sensitivity to semivari-
bustness to semivariogram model choice. However,
ogram model selection. However, when the spatial
on a dense grid, the consequences of a misspecified
measurement grid is dense, there are biases in the
model are severe. If the estimated model is Matern
Phase II ARL curves when the semivariogram model
without a nugget, the bias in Phase II ARL results
is misspecified and there isn’t a robust semivariogram
in more frequent false alarms. The case of fitting the
model that performs competitively in all cases.
exponential (with or without a nugget) also creates
biased Phase II ARL, but the bias is positive, mean-
ing an overly conservative control chart. It appears
6. Summary
that, while the exponential is a special case of the
Matern, even if the nugget e↵ect is correctly spec- Spatial data are increasingly prevalent across all
ified, the exponential can be a poor approximation areas of statistical practice. Within manufacturing,
to the spatial correlation structure corresponding to metrologic improvements are generating increasingly

Journal of Quality Technology Vol. 45, No. 2, April 2013

 SPATIAL CONTROL CHARTS FOR THE MEAN 145

FIGURE 13. Exponential with Nugget Semivariogram Phase II Sampling ARL Curves Demonstrating the E↵ect of Parameter
Estimation for Di↵erent m (Phase I Sample Size) and Spatial Measurement Grids. The regular T 2 using S, which can only
be applied in the m = 4 and 3-inch grid case (upper right panel) because the dimensionality exceeds m in other cases, and
the naive T 2 ignoring spatial correlation (Zhang and Albin (2009)) are included for comparison. The e↵ect of misspecifying
the semivariogram depends on both m and spatial measurement grid.

massive spatial datasets. This paper was motivated duces a correlation structure. This paper uses well-
by an opportunity with a bottle manufacturer to use known models from spatial statistics to represent the
a nondestructive instrument to measure bottle thick- correlation and uses a parameterized covariance ma-
ness at any bottle location. While the manufacturer trix in a multivariate control chart.
was excited about the opportunity to detect smaller
flaws by measuring more locations, the statistical in- While recent work in profile monitoring has ac-
tuition is that measuring locations close together in- counted for spatial correlation, this correlation has

Vol. 45, No. 2, April 2013 www.asq.org

146 SCOTT D. GRIMSHAW, NATALIE J. BLADES, AND MICHAEL P. MILES

FIGURE 14. Matern Semivariogram Phase II Sampling ARL Curves Demonstrating the E↵ect of Parameter Estimation for
Di↵erent m (Phase I Sample Size) and Spatial Measurement Grids. The regular T 2 using S, which can only be applied in the m
= 4 and 3-inch grid case (upper right panel) because the dimensionality exceeds m in other cases, and the naive T 2 ignoring
spatial correlation (Zhang and Albin (2009)) are included for comparison. The e↵ect of misspecifying the semivariogram
depends on both m and spatial measurement grid.

been treated as a nuisance parameter and added to improves the ARL curves—as the degree of correla-
the list of parameters monitored in the profile. This tion increases, the spatial chart that accounts for this
paper also uses a parametric model for spatial cor- correlation will signal more quickly.
relation, but instead of including those parameters
in the multivariate characteristics vector, they are The proposed control charts are applied to the
used to parameterize the covariance matrix. The sim- manufacture of 32-ounce Boston round bottles,
ulation study demonstrates that correlation is not a where the in-control semivariogram is estimated from
nuisance but rather that accounting for correlation measurements on historical data from in-control bot-

Journal of Quality Technology Vol. 45, No. 2, April 2013

 SPATIAL CONTROL CHARTS FOR THE MEAN 147

FIGURE 15. Matern with Nugget Semivariogram Phase II Sampling ARL Curves Demonstrating the E↵ect of Parameter
Estimation for Di↵erent m (Phase I Sample Size) and Spatial Measurement Grids. The regular T 2 using S, which can only
be applied in the m = 4 and 3-inch grid case (upper right panel) because the dimensionality exceeds m in other cases, and
the naive T 2 ignoring spatial correlation (Zhang and Albin (2009)) are included for comparison. The e↵ect of misspecifying
the semivariogram depends on both m and spatial measurement grid.

tles. The proposed methods successfully identify out- This spatial interpolation provides a novel diagnostic
of-control production. While the example and simu- tool for SPC: when a process has signaled as out-of-
lations used the T 2 chart, which will be sensitive to control, the kriged estimates provide a prediction of
large jumps, this paper also presented a spatial mul- the quality characteristic across the entire surface—
tivariate EWMA chart for detecting small shifts. not just at the observed locations but also at every
location that was not measured. In the worked ex-
Additionally, the spatial interpolation constructed ample, the kriging revealed a defective region along
from bottles generating an out-of-control signal pro- the parting line—not just a defective point on the
vides insight into the nature of the flawed bottles. parting line.

Vol. 45, No. 2, April 2013 www.asq.org

148 SCOTT D. GRIMSHAW, NATALIE J. BLADES, AND MICHAEL P. MILES

Acknowledgments Megahed, F. M.; Woodall, W. H.; and Camelio, J. A.

(2011). “A Review and Perspective on Control Charting with
The authors thank Amy Royer for creating Figure Image Data.” Journal of Quality Technology 43, pp. 83–98.
2. The authors greatly appreciate the helpful com- Noorossana, R.; Saghaei, A.; and Amiri, A. (2011). Statis-
ments of the Editor and anonymous referees that tical Analysis of Profile Monitoring. Hoboken, NJ: Wiley.
greatly improved the paper. Qiu, P.; Zhou, C.; and Wang, Z. (2010). “Nonparametric
Profile Monitoring by Mixed E↵ects Modeling (with Discus-
sion).” Technometrics 52, pp. 265–293.
References Quesenberry, C. P. (1993). “The E↵ect of Sample Size on
Bivand, R. S.; Pebesma, E. J.; and Gómez-Rubio, V. (2008). Estimated Limits for X and X Control Charts.” Journal of
Applied Spatial Data Analysis with R. New York, NY: Quality Technology 25, pp. 237–247.
Springer. Reese, C. S.; Deininger, P.; Hamada, M. S.; and Krabill,
Champ, C. W.; Jones–Farmer, L. A.; and Rigdon, S. E. R. (2008). “Exploring the Statistical Advantages of Nonde-
(2005). “Properties of the T 2 Control Chart when Parame- structive Evaluation Over Destructive Testing.” Journal of
ters are Estimated.” Technometrics 47, pp. 437–445. Quality Technology 40, pp. 259–267.
Chen, H.-S.; Simpson, D. G.; and Ying, Z. (2000). “Infill Reynolds, M. R. and Stoumbos, Z. G. (2004). “Control
Asymptotics for a Stochastic Process Model with Measure- Charts and the Efficient Allocation of Sampling Resources.”
ment Error.” Statistica Sinica 10, pp. 141–156. Technometrics 46, pp. 200–214.
Chicken, E.; Pignatiello, Jr., J. J.; and Simpson, J. R. Ribeiro, Jr., P. J. and Diggle, P. J. (2001). “geoR: A Pack-
(2009). “Statistical Process Monitoring of Nonlinear Profiles age for Geostatistical Analysis.” R-NEWS, 1, pp. 15–18.
Using Wavelets.” Journal of Quality Technology 41, pp. 198– Schabenberger, O. and Gotway, C. A. (2005). Statistical
212. Methods for Spatial Data Analysis. Boca Raton, FL: Chap-
Colosimo, B. M.; Semeraro, Q.; and Pacella, M. (2008). man and Hall.
“Statistical Process Control for Geometric Specifications: On Staudhammer, C.; Maness, T. C.; and Kozak, R. A. (2007)
the Monitoring of Roundness Profiles.” Journal of Quality “Profile Charts for Monitoring Lumber Manufacturing Using
Technology 40, pp. 1–18. Laser Range Sensor Data.” Journal of Quality Technology
Colosimo, B. M.; Mammarella, F.; and Petro, S. (2010). 39, pp. 224–240.
“Quality Control of Manufactured Surfaces.” In Frontiers
Testik, M. C.; Runger, G. C.; and Borror, C. M. (2003).
in Statistical Quality Control 9, Lenz, H. J.; Wilrich, P.-T.;
“Robustness Properties of Multivariate EWMA Control
and Schmid, W., eds., pp. 55–70. Berlin, Germany: Physica-
Charts.” Quality and Reliability Engineering International
Verlag.
19, pp. 31–38.
Cressie, N. A. C. (1993). Statistics for Spatial Data, revised
Unkel, S.; Farrington, C. P.; Garthwaite, P. H.; Robert-
edition. New York, NY: Wiley.
son, C.; and Andrews, N. (2012) “Statistical Methods for
Diggle, P. J.,; Rowlingson, B.; and Su, T.-L. (2005). “Point
the Prospective Detection of Infectious Disease Outbreaks:
Process Methodology for On-Line Spatio-Temporal Disease
A Review.” Journal of the Royal Statistical Society, Series
Surveillance.” Environmetrics 16, pp. 423–434.
A 174, pp. 49–82.
Diggle, P. J. and Ribeiro Jr., P. J. (2007). Model-based
Wei, Y.; Zhao, Z.; and Lin, D. K. J. (2012) “Profile Control
Geostatistics. New York, NY: Springer.
Charts Based on Nonparametric L-1 Regression Methods.”
Ding, Y.; Zeng, L.; and Zhou, S. (2006). “Phase I Analy-
Annals of Applied Statistics 6, pp. 409-427.
sis for Monitoring Nonlinear Profiles in Manufacturing Pro-
cesses.” Journal of Quality Technology 38, pp. 199–216. Woodall, W. H.; Spitzner, D. J.; Montgomery, D. C.;
Kaluzny, S. P.; Vega, S. C.; Cardoso, T. P.; and Shelly, and Gupta, S. (2004) “Using Control Charts to Monitor
A. A. (1998). S+ Spatial Stats. New York, NY: Springer. Process and Product Quality Profiles.” Journal of Quality
Jensen, W. A. and Birch, J. B. (2009). “Profile Monitoring Technology 36, pp. 309–320.
via Nonlinear Mixed Models.” Journal of Quality Technology Ying, Z. (1991). “Asymptotic Properties of a Maximum Like-
41, pp. 18–34. lihood Estimator with Data from a Gaussian Process.” Jour-
Jensen, W. A.; Birch, J. B.; and Woodall, W. H. (2008). nal of Multivariate Analysis 36, pp. 280–296.
“Monitoring Correlation Within Linear Profiles Using Mixed Zhang, H. and Albin, S. (2009). “Detecting Outliers in Com-
Models.” Journal of Quality Technology 40, pp. 167–183. plex Profiles Using a 2 Control Chart Method.” IIE Trans-
Jensen, W. A.; Jones-Farmer, L. A.; Champ, C. W.; and actions 41, pp. 335–345.
Woodall, W. H. (2006). “E↵ects of Parameter Estimation Zhang, Z. (2004). “Inconsistent Estimation and Asymptot-
on Control Chart Properties: A Literature Review.” Journal ically Equal Interpolations in Model-Based Geostatistics.”
of Quality Technology 38, pp. 349–364. Journal of the American Statistical Association 99, 250–261.
Lindquist, M. A.; Waugh, C.; and Wager, T. D. (2007). Zimmerman, D. L. and Zimmerman, M. B. (1991). “A Com-
“Modeling State-Related fMRI Activity Using Change-Point parison of Spatial Semivariogram Estimators and Corre-
Theory.” NeuroImage 35, pp. 1125–1141. sponding Ordinary Kriging Predictors.” Technometrics 33,
Lowry, C. A.; Woodall, W. H.; Champ, C. W.; and Rig- pp. 77–91.
don, S. E. (1992). “A Multivariate Exponentially Weighted Zou, C.; Tsung, F.; and Wang, Z. (2008). “Monitoring Pro-
Moving Average Control Chart.” Journal of Quality Tech- files Based on Nonparametric Regression Methods.” Techno-
nology 34, pp. 46–53. metrics 50, pp. 512–526.

Journal of Quality Technology Vol. 45, No. 2, April 2013