Chapter 8

Statistical Quality Monitoring of Geometric

Tolerances: the Industrial Practice

Bianca Maria Colosimo and Massimo Pacella

Abstract This chapter shows how traditional approaches for statistical process
control can be used for monitoring geometric tolerances. These approaches can be
useful to quickly detect changes in the manufacturing process. In particular, two
simple methods are presented. The first one uses standard variable control charts
for monitoring over time the error associated with the geometric tolerance at hand.
The second approach designs a control band around the mean shape of the feature
associated with the geometric tolerance. Both approaches are shown with refer-
ence to the problem of monitoring a roundness form tolerance. Given their ease of
use, the approaches are viable solutions for industrial practice.

8.1 Introduction

Geometric features are machined by processes that may experience changes due to
the material machined, improper setup, errors of the operator, wear or sudden
changes of the machine conditions, etc. Usually, these changes cause deteriorated
process performance, i.e., the process may produce an increased number of non-
conforming or defective items. This is why statistical quality monitoring – also

__________________________________
B.M. Colosimo
Dipartimento di Meccanica, Politecnico di Milano,
Via la Masa 1, 20156 Milan, Italy,
e-mail: biancamaria.colosimo@polimi.it
M. Pacella
Dipartimento di Ingegneria dell’Innovazione, Università del Salento,
Via per Monteroni, 73100 Lecce, Italy,
e-mail: massimo.pacella@unisalento.it

237
238 B.M. Colosimo and M. Pacella

known as statistical process control (SPC) – can be effectively considered as a

way to quickly detect unusual states of the manufacturing process by issuing an
alarm.
Traditional approaches for quality monitoring consists in designing and using
control charts when the quality characteristic of interest is well represented by one
(or more) variable. As an example, a dimensional tolerance, such as a diameter,
can be modeled as a random variable and then monitored by using a control chart.
However, when the quality of a manufactured product is related to a geometric
tolerance, instead of a dimensional one, the problem is how quality monitoring can
be implemented. This chapter presents two simple approaches that can be used for
this aim. The first approach consists in summarizing all the information contained
in the points measured on the shape of interest in just one synthetic variable. This
variable usually measures the geometric form error as the (maximum) distance
between the actual profile and the ideal geometry. Then a control chart of the
estimated geometric error is implemented for quality monitoring.
A second approach, which is called a location control chart, consists in design-
ing a control region around the ideal or mean shape observed on a set of profiles.
This control region is defined by two limits that can be computed by virtually
designing a control chart for points observed at any given location and controlling
the whole false-alarm rate of this set of control intervals by using Bonferroni’s
inequality. An alarm is issued when at least one point, in the whole set of data
observed in a profile, exceeds the control limits.
Both methods presented are based on Shewhart’s traditional control chart,
which is briefly summarized in Section 8.2. This section can be skipped by a
reader who is already familiar with SPC tools. In Section 8.3, the issue of round-
ness form error is discussed. In Section 8.4, the control chart for the geometric
form error is presented, while in Section 8.5 the location control chart is discussed.
In Sections 8.4 and 8.5, a real case study concerning roundness form error is used
as a reference. This case study is described in detail in Chapter 11.

8.2 Shewhart’s Control Chart

SPC may be considered to have begun with the pioneering work of Walter A. She-
whart, an engineer at Bell Telephone Laboratories, where he was faced with the
issue of obtaining good quality in the mass production of interchangeable equip-
ment for the rapidly expanding telephone system. Shewhart’s ideas (Shewhart
1931) are still relevant today. The most important technique he developed, i.e., the
control chart, is nowadays widely used for quality control of manufacturing proc-
esses and services. This section provides a brief overview of Shewhart’s approach,
while a more general description can be found in standard texts (Alwan 2000;
Montgomery 2004; Ryan 2000).
The baseline idea of Shewhart’s approach is that in any production process, no
matter how well designed it is, there exists a certain amount of natural variation in
8 Statistical Quality Monitoring of Geometric Tolerances: the Industrial Practice 239

the outcomes. This variability is always present as it results from a large number
of so-called common causes (i.e., natural causes) which are to some extent un-
avoidable (or can be removed but at prohibitive costs).
The process may also be affected by external sources of variation, which are
upsetting its natural functioning. Sources of variability that are not part of the
process and that can occur only accidentally are called special causes (i.e., assign-
able causes) of variation. The presence of special causes may lead to excessive
variation in process outcomes, possibly resulting in quality loss and customer
complaints. In such cases, quality improvement is possible by detection and re-
moval of special causes of variation. Since a special cause is not inherently part of
the process, it can usually be eliminated without revising the process itself. In
many cases, the removal of a special cause of variation is possible. An operator
can be instructed to recognize and remove it.
Therefore, it is essential to be able to distinguish between situations where only
common causes of variation affect the outcomes of a process and situations where
special causes are also present. A tool for supporting the operator’s decision is
needed for this purpose, since the effect of a possible special cause can be hidden
in the variation due to common causes. Shewhart developed the control chart for
this purpose and gave a rationale for using such a tool in process monitoring. The
control chart is intended to monitor a process by issuing an alarm signal when it is
suspected of going out of control.
The baseline idea of Shewhart (1931) is that if only common causes of varia-
tion are present, the manufacturing process should be statistically in control, i.e.,
the outcomes should be predictable according to a given statistical model. The
statistical predictability of a process that is in control is the basis for the control
chart. It does not mean that there is no variation, or that there is a small variation.
Simply, it means that the outcomes are predictable in a statistical sense. For exam-
ple, based on previous observations, it is possible for a given set of limits to de-
termine the probability that future observations will fall within these limits.
In order to apply control charting, data are collected in samples, usually re-
ferred to as subgroups. Statistics of interest (such as the mean or the standard
deviation) are computed for each sample in order to summarize the information
contained within each subgroup. These statistics are then plotted on a graph and
compared with limits, which represent the bandwidth of the natural variation due
to common causes. The idea is that as long as all statistics are within the control
limits, it is reasonable to assume that the underlying process is statistically in
control.
Given this monitoring strategy, a control chart involves measuring a quality
characteristic of interest at regular intervals, collecting n items each time and plot-
ting one or more sample statistics of the quality parameter against time. In prac-
tice, a new point is plotted each time a subgroup of items is measured.
Figure 8.1 shows an example of a control chart for individual measurements,
i.e., when the sample size n is equal to 1 and hence the single datum observed at
each time period is considered in the control chart.
240 B.M. Colosimo and M. Pacella

Figure 8.1 An example of a control chart for subgroups of size n = 1, i.e., individual measure-
ments. LCL lower control limit, UCL upper control limit

Points outside the control limits are called out-of-control signals and indicate
the likely presence of special causes of variation that are affecting the process. The
presence of special causes implies that there could be some source of variation that
causes the measurements to be variable, differently from what can be attributed to
the effect of natural causes only. If an out-of-control signal is observed, action is
required to track down the special cause that is responsible. Owing to the random
nature of the observations, there is also the possibility that an out-of-control signal
is encountered while the process is statistically in control and hence when there is
no special cause to be identified and removed. This is called a false out-of-control
signal or false alarm.
It is worth noting that control limits are not directly related to customer’s speci-
fications (i.e., quality requirements) for the process outcomes. In fact, control
limits just indicate the magnitude of the natural variability of the sample character-
istic used to monitor the process. They are based on the relevant sampling distri-
butions of process outcomes when only natural causes are present, while, in gen-
eral, quality requirements are not related to the actual performance of the process.
Besides, quality specifications are always related to single products, whereas con-
trol limits can be computed for any arbitrary statistic of the process outcome (such
as the sample mean, range, or standard deviation).
Even when a control chart for individual measurements is considered, i.e., a
single process outcome at each time period is plotted on the chart, if the variation
8 Statistical Quality Monitoring of Geometric Tolerances: the Industrial Practice 241

due to common causes is relatively large, all points on the control chart may be
within the control limits, but the process outcomes might fail to meet quality
specifications. Furthermore, the presence of special causes does not necessarily
mean that there is large variation, or that the specifications are not met.
A control chart must be sensitive enough to detect the effect of special causes
of variation, but also must not generate too many false alarms. In practice, a
balance between these two objectives can be achieved by a proper setup of the
control chart, i.e., by determining the appropriate width of the control limits. In
Shewhart’s approach, the control limits are usually set equal to three standard de-
viations from the center line, which represent in turn the expected value of the
plotted statistics. Considering the assumption of normally and independently
distributed points plotted on the chart, this control limit position corresponds to
an expected false-alarm signal in about 370 points on average.

8.2.1 Two Stages in Control Charting

Consider a process which has been set up in order to operate stably and properly,
i.e., in its (presumable) in-control state. A set of observations collected on this
process are first analyzed and then used to design a control chart. The aim is both
to evaluate the stability of the process and to estimate the in-control-state’s pa-
rameters. In this phase, also called phase I, control charts are used offline to de-
termine retrospectively from the set of collected data whether the process is indeed
in a state of statistical control.
In phase I, the control chart is used as an aid to the analyst to screen out out-of-
control data from the set of collected observations so that a set of presumably in-
control process data can be obtained to model the distribution of the monitoring
statistic (such as the mean, range, or standard deviation). To this aim, the collected
data are used to set up a set of initial trial control limits for the monitoring statistic.
If out-of-control signals are observed, the process is investigated to see if there
exist any special causes to explain these out-of-control signals. If indeed some
special causes are found, then the samples which produce the out-of-control sig-
nals are removed from the data set. Then, the remaining samples are used to re-
estimate control limits. This procedure should be repeated until no out-of-control
signals are generated, or when underlying assignable causes cannot be found. In
the latter case, even if these points exceed the limits simply by chance or because
of some uncovered assignable causes, to be conservative, one may choose to dis-
card them to avoid potential contamination in the data set.
When no out-of-control signals are eventually generated, at the end of phase I a
data set is available which provides information concerning the variability that can
be attributed to common causes of variation of the process in its in-control state.
Thus, reliable control limits of the control chart are established for online process
monitoring in subsequent phase II.
242 B.M. Colosimo and M. Pacella

In phase II (also called the operating phase), control charts are used for testing
whether the process remains in control when future subgroups are drawn. During
this phase, the goal is to monitor the online data and quickly detect changes in the
process from the baseline model established in phase I.
Since the target of control charting is detecting process changes as quickly as
possible, the performance of a monitoring approach is usually described by the
run-length distribution, where the run length is the number of samples taken be-
fore an out-of-control signal is given. In particular, the average run length is often
used as a performance index.

8.3 Geometric Tolerances: an Example of a Geometric Feature

Concerning Circularity

In order to exemplify the issue of geometric tolerance monitoring, the roundness

of a circular feature is assumed as a reference from now on. A circular feature in a
component such as a shaft or a hole is one of the most frequently encountered
features in manufacturing, because functionality of mechanical parts is very often
related to a proper rotation. According to the ISO/TS 12181 (2003) standard, a
circular profile is the line extracted on a cross section of a surface of revolution.
For a circular feature, the circularity error attained is represented by the so-called
out-of-roundness (OOR) value. This important geometric characteristic can be
estimated in a number of different ways.
Conventionally, the profile of a circular feature can be measured diametrically
by an operator. The difference between the maximum and the minimum diameters
measured is used as an indication of the OOR for that circular feature. Obviously,
this diametric measurement of the OOR can be deceptive in actual applications.
Alternatively, the profile of a circular feature can be measured by tracing its pe-
rimeter with a stylus or a probe. One approach is to revolve the item against a
displacement transducer. This is the basic approach of the so-called rotondimeter,
i.e., a measuring device equipped with a stylus which is able to measure a circular
profile with a sufficiently high number of points in a relatively short time. Such a
device can be exploited to measure axially symmetric workpieces, i.e., workpieces
which can be easily rotated around their axes. A different practice may often be
encountered in industry. It consists in measuring a circular profile by sampling
several points on its perimeter by either a mechanical probe or an optical probe on
a coordinate measuring machine (CMM). CMMs are the most-general-purpose
devices nowadays available in industry for measuring any type of physical geo-
metric characteristic of an object.
Irrespective of the specific device exploited, the OOR value must be computed
as the difference between the maximum and the minimum radial distances of the
manufactured feature with respect to a predetermined center, which is the center of
the so-called substitute geometry. Geometrically, this corresponds to finding two
8 Statistical Quality Monitoring of Geometric Tolerances: the Industrial Practice 243

concentric circles, one circumscribing and one inscribing the profile sampled on
the manufactured feature. The OOR is then estimated by the width of the annulus
determined by these two concentric circles. CMMs, which are controlled by a
computer, estimate algorithmically the geometric error on the basis of a finite
number of points sampled on the manufactured feature. The geometric best fit is a
numerical transformation between measurements and their substitute geometry,
i.e., the actual center of the substitute geometry.
Several methods can be implemented in practice. The differences in methods
(and results) essentially depend upon the algorithm used to determine the common
center of the circumscribed circle and of the inscribed one. Four procedures are
commonly used: the minimum zone (MZ), the least squares (LS), the maximum-
inscribed circle (MIC), and the minimum-circumscribed circle (MCC).
1. The MZ center is that for which the radial difference between two concentric
circles that just contain the measured points is a minimum. The mathematical
problem of finding the MZ center can be stated as a Chebychev problem. This
is a rather complicated nonlinear problem and exact algorithms for solving it
are not readily available.
2. The LS center is that of a circle from which the sum of the squares of the radial
ordinates between this circle and the measured points is a minimum.
3. The MIC center is that of the largest circle that can be inscribed within the set
of measured points.
4. The MCC center is that of the smallest circle that just contains all the measured
points.
Algorithms for calculating these centers exist, with varying computational
complexity (Carr and Ferreira 1995a, b; Gass et al. 1998). As an example, Fig-
ure 8.2 depicts a sampled roundness profile, the corresponding MZ OOR value
(the radius distance between two concentric circles which contain the sampled
profile), as well as the resulting substitute geometry.
Two procedures for determining the center of the substitute geometry are the
MZ and the LS methods. These procedures are assumed as a reference henceforth
since the former best conforms to the ISO standards for form tolerances, while the
latter is the one most frequently encountered in actual applications. The MZ algo-
rithm looks for a couple of geometric nominal features (e.g., a couple of concen-
tric circles for roundness) at the minimum distance that includes the whole set of
measurement points. The method always consists of the minimum deviation, given
a set of measurement data. However, it is very sensitive to asperities.
Different procedures to estimate the couple of minimum-distance nominal fea-
tures have been reported in the literature. In particular, Carr and Ferreira
(1995a, b) formulated roundness as nonlinear optimization problems, which are
then transformed into a series of linear problems. Their linear program models are
derived from the original optimization problem modeled as a constrained nonlin-
ear programming problem with a linear objective function. This model can be
effectively implemented by the library of optimization subroutines available in a
244 B.M. Colosimo and M. Pacella

Figure 8.2 Minimum zone (MZ) out-of-roundness (OOR) for a sampled roundness profile (bold
line). The thin continuous line represents the ideal geometry

MATLAB® environment (MathWorks 1991). The efficiency of these algorithms

with respect to computation time is reported as a linear function of the number of
data points. Gass et al. (1998) and Moroni and Pacella (2008) further improved the
solution proposed by Carr and Ferreira (1995a, b) for MZ form error computation.
The LS algorithm, which represents a widely used procedure for form tolerance
evaluation, minimizes the sum of squared deviations of measured points from the
fitted feature. The LS method associates one substitute feature with measurement
points (e.g., one circle for roundness) and calculates the maximum peak-to-valley
distance of the measurement points from the substitute feature. The LS method
requires that the sum of the squared errors be minimized. Although the form error
computed from the extreme points can be slightly higher than that obtained from
the MZ method, the LS-fitted feature is very stable and much less sensitive to the
effects of asperities, making it suitable for many practical applications. The im-
plementation developed by NPL (the UK’s national measurement laboratory) and
based on the Least Squares Geometric Elements (LSGE) library for MATLAB® is
used henceforth. The LSGE library consists of functions to find the LS fit of geo-
metric shapes to data, implementing a number of geometric fitting routine key
functions. It is based on a general-purpose, nonlinear LS solver that takes as the
input function-and-gradient routines and these routines are implementations of the
geometric evaluation key functions. This library was tried out on several bench-
mark data sets and was found to give correct results (Moroni and Pacella 2008).
8 Statistical Quality Monitoring of Geometric Tolerances: the Industrial Practice 245

Irrespective of the specific method used, a roundness profile is usually consid-

ered as conforming to the requirements when the OOR value is lower than the
specified tolerance. It should be noted, however, that many different circular pro-
files, which can induce different functional properties of the machined items, can
be characterized by the same OOR value, resulting in important differences in the
characteristics of the profile, e.g., in the assembly precision (Cho and Tu 2002).

8.4 Control Chart of Geometric Errors

When the quality of the manufactured product is related to surface texture (rough-
ness and waviness) or to geometric form errors (e.g., straightness, roundness,
cylindricity, planarity), Shewhart’s control chart can be used for process monitor-
ing if the information related to the texture or to the form is summarized in just
one (or few) synthetic variable. As an example, the roughness obtained on a ma-
chined surface is usually represented by the average roughness Ra parameter, or
the roundness characterizing a given item can be summarized by the OOR.
This section reviews the approach based on Shewhart’s control chart for moni-
toring a synthetic measure of the error between the actual profile and the ideal
geometry, which can be considered representative of industrial practice.

8.4.1 Control Limits of the Individuals Control Chart

If the distribution function of the measured data can be assumed to be normal, then
the individuals control chart can be used. Given a nominal false alarm probability
α ′ , the upper and lower control limits of the individuals control chart can be
computed as follows.
UCL = μ + Zα ' 2σ ,
CL = μ , (8.1)
LCL = μ − Zα ' 2σ ,

where UCL is the upper control limit, μ is the mean of the individual measure-
ments, σ represents the standard deviation, Zα ' 2 is the (1 − α ′ 2 ) percentile of
the standard normal distribution, CL is the center line, and LCL is the lower control
limit. Typically, μ and σ are unknown and hence they have to be estimated via a
phase I sample of independently and identically distributed measurements.
Assume we collect a sample of n profiles observed from the in-control manu-
facturing process and let o j denote the synthetic variable which summarizes the
form error for the j th profile, where j = 1, 2,… n . Classic estimators of μ and σ
246 B.M. Colosimo and M. Pacella

1 n
are the sample mean and the sample standard deviation, i.e., o = ∑ o j and
n j =1
1 n
∑ ( o j − o ) , respectively.
2
s=
n − 1 j =1
The sample standard deviation is asymptotically efficient for independently and
identically distributed normal random variables. The disadvantage is that it may be
sensitive to outliers. When outliers might occur, a different estimator of the stan-
dard deviation should be used which is less sensitive to these deviations. The av-
erage of the moving ranges, which are the absolute values of the difference of two
successive observations, can be used to measure the process variability. The aver-
age moving range, say, MR , scaled by 2 π can be used to obtain a more robust
1 n
estimator than the sample standard deviation. Let MR = ∑ o j − o j −1 , the
n − 1 j =2
individuals control chart is defined as follows. (Note that in a set of n observa-
tions sequentially sampled on a process, the number of moving ranges is n − 1 ).

π
UCL = o + Zα ' 2 MR,
2
CL = o , (8.2)
π
LCL = o − Zα ' 2 MR.
2
When the underlying distribution function is not normal, the approach de-
scribed can be still used after a suitable transformation of the OOR values has
been implemented to achieve normality of the transformed data. The transforma-
tion is obtained by applying a single mathematical function to the raw data values.
Depending on the distribution of sample data, there are many different functions,
such as square root, logarithm, power, reciprocal, and arcsine, which one could
apply to transform sample data. The Box–Cox power approach (Box and Cox
1964) finds an optimal power transformation that can be useful for correcting
nonnormality in process data.

8.4.2 An Example of Application to the Reference Case Study

As previously observed, the information contained in a roundness profile can be

summarized in the OOR value. With reference to the tolerance value t characteriz-
ing the circular profile, this OOR value is usually computed to decide whether the
machined item has to be scrapped/reworked (if OOR ≥ t) or can be considered to
conform to the requirements (if OOR < t).
8 Statistical Quality Monitoring of Geometric Tolerances: the Industrial Practice 247

The simplest approach for quality control of roundness profiles consists in

monitoring the OOR values by an individuals control chart. In order to exemplify
this approach, the case study described by Colosimo et al. (2008) is considered as
a reference. It consists of 100 roundness profiles obtained from lathe turning,
where each profile consists of a set of 748 points representing the deviations from
the nominal radius (equal to 13 mm) at equally distributed angle locations. This
case study is described in detail in Chapter 11. With reference to this case study of
100 roundness profiles, Table 8.1 reports the OOR values obtained for each item
estimated by both the MZ algorithm and the LS algorithm.

Table 8.1 Out-of-roundness (OOR) values (mm) based on minimum-zone (MZ) and least-
squares (LS) algorithms

MZ LS MZ LS MZ LS MZ LS
Sample OOR OOR Sample OOR OOR Sample OOR OOR Sample OOR OOR
1 0.0217 0.0231 26 0.0181 0.0184 51 0.0241 0.0258 76 0.0168 0.0178
2 0.0182 0.0208 27 0.0166 0.0178 52 0.0097 0.0101 77 0.0175 0.0193
3 0.0165 0.0178 28 0.0100 0.0109 53 0.0154 0.0184 78 0.0156 0.0160
4 0.0125 0.0141 29 0.0215 0.0251 54 0.0121 0.0128 79 0.0133 0.0141
5 0.0168 0.0178 30 0.0131 0.0145 55 0.0179 0.0189 80 0.0215 0.0237
6 0.0133 0.0135 31 0.0149 0.0169 56 0.0168 0.0178 81 0.0192 0.0207
7 0.0251 0.0259 32 0.0115 0.0122 57 0.0207 0.0212 82 0.0133 0.0148
8 0.0192 0.0213 33 0.0120 0.0128 58 0.0191 0.0196 83 0.0222 0.0230
9 0.0108 0.0118 34 0.0145 0.0155 59 0.0146 0.0157 84 0.0217 0.0227
10 0.0188 0.0209 35 0.0172 0.0193 60 0.0182 0.0191 85 0.0170 0.0176
11 0.0110 0.0120 36 0.0206 0.0221 61 0.0185 0.0193 86 0.0181 0.0191
12 0.0193 0.0203 37 0.0155 0.0180 62 0.0100 0.0106 87 0.0163 0.0171
13 0.0177 0.0183 38 0.0188 0.0218 63 0.0133 0.0141 88 0.0188 0.0207
14 0.0204 0.0224 39 0.0105 0.0119 64 0.0131 0.0136 89 0.0117 0.0117
15 0.0188 0.0214 40 0.0134 0.0145 65 0.0132 0.0139 90 0.0130 0.0139
16 0.0179 0.0189 41 0.0171 0.0181 66 0.0138 0.0151 91 0.0166 0.0181
17 0.0164 0.0171 42 0.0136 0.0141 67 0.0156 0.0167 92 0.0197 0.0207
18 0.0117 0.0119 43 0.0176 0.0183 68 0.0120 0.0124 93 0.0192 0.0199
19 0.0235 0.0256 44 0.0206 0.0217 69 0.0165 0.0177 94 0.0178 0.0187
20 0.0124 0.0129 45 0.0255 0.0280 70 0.0143 0.0146 95 0.0186 0.0204
21 0.0148 0.0155 46 0.0141 0.0149 71 0.0176 0.0191 96 0.0134 0.0149
22 0.0131 0.0146 47 0.0174 0.0185 72 0.0103 0.0111 97 0.0209 0.0214
23 0.0188 0.0197 48 0.0130 0.0146 73 0.0197 0.0206 98 0.0167 0.0181
24 0.0140 0.0150 49 0.0167 0.0182 74 0.0128 0.0133 99 0.0143 0.0158
25 0.0132 0.0146 50 0.0150 0.0159 75 0.0123 0.0130 100 0.0146 0.0149
248 B.M. Colosimo and M. Pacella

Figure 8.3 Individuals control chart of the transformed OOR values based on the MZ algorithm

From Table 8.1 it can be observed that for each profile the LS algorithm over-
estimates the roundness error calculated by the MZ algorithm. In fact, the sample
mean of the OOR values based on the LS algorithm is 0.0174 mm, while the cor-
responding value based on the MZ algorithm is 0.0162 mm (a t test of the mean
difference is significant with a p value less than 0.0005).
With reference to the 100 samples in phase I, summarized in Table 8.1, two in-
dividuals control charts of the OOR values are designed. The Anderson–Darling
test (Anderson and Darling 1952) was implemented in order to detect any depar-
tures from normality for the distribution function of the OOR values in the case of
both the MZ algorithm and the LS algorithm. From numerical computation on the
data reported in Table 8.1, the set of 100 OOR values is normally distributed (with
p = 0.149 in the case of the MZ algorithm and p = 0.150 in the case of the LS
algorithm for the test). Therefore, an individuals control chart is designed for the
sequence of sample OOR values, resulting in the control charts shown in Fig-
ure 8.3 and Figure 8.4. In particular, Figure 8.3 depicts the individuals control
chart with reference to the OOR values based on the MZ algorithm, while a simi-
lar control chart with reference to the OOR values based on the LS algorithm is
depicted in Figure 8.4. The two control charts depicted in Figures 8.3 and 8.4 are
based on different center lines (i.e., 0.0162 mm in the case of the MZ algorithm
and 0.0174 mm in the case of the LS algorithm) as well as on a different estimated
standard deviation for the data plotted (i.e., 0.0037 mm in the case of the MZ algo-
rithm and 0.0041 mm in the case of the LS algorithm, where the estimated value is
based on the average moving range of data).
It can be observed that no out-of-control signals are detected in the two control
charts in this design phase. In fact, there are no points exceeding the control limits
8 Statistical Quality Monitoring of Geometric Tolerances: the Industrial Practice 249

Figure 8.4 Individuals control charts of the transformed OOR values based on the least-squares
(LS) algorithm

in the control charts of the OOR values obtained by the MZ algorithm and by the
LS algorithm. As expected, irrespective of the specific method used, the patterns
of OOR values in the two control charts depicted in Figures 8.3 and 8.4 are simi-
lar. Indeed, a computer simulation not reported here showed that the performance
of an individuals control chart of the OOR values, measured in terms of a type I
error rate for in-control profiles, is typically not affected by the specific algorithm
used to estimate the geometric form error (i.e., MZ algorithm or LS algorithm).

8.5 Monitoring the Shape of Profiles

Given the case study of 100 roundness profiles obtained from lathe turning (Co-
losimo et al. 2008), where each profile consists of a set of 748 points representing
the deviations from the nominal radius at equally distributed angle locations (refer
to Chapter 11 for details), it can be observed that the data can be stored as 748-
length vectors.
Hence, each profile can be considered as a realization of a multivariate process.
A possible approach for profile monitoring consists in studying all the points of
the profile simultaneously by means of a specific multivariate technique, which
reduces the information contained within the vector down to a single metric. An
example of such multivariate metrics is the T 2 statistic (Montgomery 2004). How-
ever, the use of this technique is frequently ill-advised for profile monitoring.
When the number of monitored points exceeds the number of collected samples
250 B.M. Colosimo and M. Pacella

(as in the reference case), the sample covariance matrix of data is not invertible
and the usual statistical inference is not possible. This condition can often be en-
countered in actual applications, especially when machined profiles subject to
geometric specification are considered, where in order to have an accurate esti-
mate of the form error the number of observations sampled is usually on the order
of hundreds.
A different approach, aimed at combining simplicity with the need of keeping
all the information of the data observed at each location of the machined feature is
the location control chart proposed by Boeing (1998) and described in the follow-
ing subsections.

8.5.1 The Location Control Chart

The location control chart was presented in Boeing (1998, pp. 89–92) with refer-
ence to applications in which numerous measurements of the same variable (e.g., a
dimension such as thickness) are made at several locations on each manufactured
part, i.e., in the context of profile monitoring (Woodall et al. 2004). In practice,
the location control chart consists in applying a traditional Shewhart control chart
separately to each data point observed at a given location of the part, i.e., it con-
sists in designing a control interval at each different position of the point observed
on the shape of interest. The rationale behind this approach is that, if the observed
shape is in control, the data observed at that specific location should stay within
that interval with a given probability. On the other hand, when the process goes
out of control, it is likely that the control interval will be violated at one or more
locations.
In order to design the location control chart, the first step consists in identifying
the center of each interval, i.e., the systematic pattern of the in-control shape. This
reference for the in-control shape is usually estimated as the average of all the in-
control data observed at each location.
Starting from the mean shape, generation of the location control chart consists
in computing the upper and lower control limits at each location, using the stan-
dard approach that places the limits at ± K standard deviations from the sample
mean. According to this method, an alarm is issued when at least one point, in the
whole set of data observed, exceeds the control limits. The actual value of constant
K depends on the required false-alarm rate for the monitoring approach. As in
standard control charting, a greater value of K implies a larger control band and
hence a lower false-alarm rate.
Owing to its inner simplicity, this chart can be easily applied in industrial prac-
tice (and in fact its origin is in Boeing 1998). However, since the control limits
used at each location depend on the responses at that specific position only, the
main disadvantage with this method is that the multivariate structure of data is
ignored. The only form of the relationship between control intervals at each loca-
tion is a constraint on the false alarm, as discussed in the next subsection.
8 Statistical Quality Monitoring of Geometric Tolerances: the Industrial Practice 251

8.5.2 Control Limits of the Location Control Chart

Assume we collect a group of n profiles, where each profile is a vector of p meas-

urements observed at a fixed set of locations. The location control chart consists of
limits computed separately for each location by means of Shewhart’s approach,
i.e., by considering the mean and the standard deviation of the n data observed at
that location and by computing the common ± K standard deviations from the sam-
ple mean. Given a profile, an alarm should be considered when at least one point,
in the set of p observations, exceeds either the upper or the lower control limit.
Let yj (k) denote the data measured at a specific location of index k on the jth
profile, where k = 1, 2, …, p and j = 1, 2, …, n. The control limits for the location
of index k are as follows:
UCL ( k ) = y ( k ) + Zα 2 s ( k ) ,
CL ( k ) = y ( k ) , , (8.3)
LCL ( k ) = y ( k ) + Zα 2 s ( k ) ,

1 n 1 n
∑ y j ( k ) and s ( k ) = ∑
2
where y ( k ) = ⎡⎣ y j ( k ) − y ( k ) ⎤⎦ are, respec-
n j =1 n − 1 j =1
tively, the sample mean and the sample standard deviation of the data observed at
location k , while Zα 2 represents the (1 − α 2 ) percentile of the standard normal
distribution.
Given that p dependent control rules are simultaneously applied, Bonferroni’s
rule for dependent events should be used to attain an actual false-alarm rate not
greater than a predefined value. Therefore, let α ′ denote the upper bound of the
first type of probability error (false-alarm probability); the value α = α ′ p is used
to design the p control limits of Equation 8.3.
In other words, the constant K of the location control chart is computed as a
function (percentile of the standardized normal distribution) of the required false-
alarm rate (type I error) corrected by Bonferroni’s method. However, it is worth
noting that different procedures can also be used, for instance the Simes modified
Bonferroni procedure. Colosimo and Pacella (2010) showed that when compared
with the standard Bonferroni method, the Simes procedure does not produce sig-
nificant effects on the performance obtained by the location control chart. Fur-
thermore, since the Simes procedure does not allow the graphical representation of
the control region as the Bonferroni procedure does, the latter is usually consid-
ered for designing the location control chart in industrial practice.

8.5.3 An Example of Application of the Location Control Chart

The location control chart consists of a center line, an upper control limit, and a
lower control limit. With reference to the roundness case study previously consid-
252 B.M. Colosimo and M. Pacella

Figure 8.5 Control limits of the location control chart (748 locations) with reference to the 100
samples of the case study. Actual false-alarm rate not greater than 0.0027

ered (and detailed in Chapter 11), Figure 8.5 shows the average in-control shape
together with the corresponding control limits. From a visual inspection of Fig-
ure 8.5 it seems that no systematic pattern characterizes the roundness profiles of
the reference case study (i.e., the Cartesian plot of the average profile is almost flat
and equal to zero, i.e., the mean profile is a perfect circle).
This appearance hides a common problem of shape analysis, which consists in
feature registration or alignment. In fact, the average in-control shape and the
corresponding control band are correctly computed by using the limits in Equa-
tion 8.3, but on profiles which are actually misaligned. After applying a registra-
tion procedure on the set of roundness data (refer to Chapter 11 for details), Fig-
ure 8.6 shows the average profile and the control band of the roundness profiles
under study. From a visual inspection, it can be easily observed that the roundness
profiles share a common shape (pattern), i.e., the turning process leaves a specific
signature on the machined components (Colosimo et al. 2008).
Given that 748 dependent control rules are simultaneously applied, Bon-
ferroni’s rule for dependent events is used to attain an actual false-alarm rate not
greater than a predefined value. In particular, under the assumption of the standard
value α ′ = 0.0027 for the upper bound of the type I probability error (false-alarm
probability), the value α = 3.61 ⋅10 −6 is used to design the 748 control limits in
Equation 8.3. The corresponding percentile of the standard normal distribution is
about equal to Zα 2 = 4.63 .
8 Statistical Quality Monitoring of Geometric Tolerances: the Industrial Practice 253

Figure 8.6 Location control chart (748 locations) with reference to the 100 samples of the case
study (after the registration step). Actual false-alarm rate not greater than 0.0027

Figure 8.7 One of the 100 profiles of the reference case study depicted against the control
limits of the location control chart (after the registration step)

Figure 8.7 shows one out of the 100 profiles of the reference case study de-
picted against the control limits of the location control chart. The profile is plot-
254 B.M. Colosimo and M. Pacella

ted against this control region, with the advantage of allowing a simple identifi-
cation of the locations where problems arise. In this specific case, even if an
alarm is issued at a specific location (location no. 206), the profile is considered
to be in control. Indeed, from the visual inspection of the location control chart,
it appears that the behavior of this profile is close to the average common profile
and there is no apparent discrepancy in the shape of the profile when compared
with the center line of the location control chart.

8.6 Conclusions

This chapter reviewed industrial practice and the contributions of statistical meth-
ods to quality monitoring when a geometric tolerance is the quality characteristic
of interest.
In particular, two methods were described. First, a control chart for monitoring
a synthetic measure of the geometric error, (i.e., a scalar that measures the error
between the actual profile and the ideal geometry). Second, a control region for
monitoring the whole profile observed (where bound limits of this region are
computed by applying a control chart separately to each set of data points ob-
served at a given location).
Given the ease of the approaches presented in this chapter, they have been con-
sidered as industrial benchmarks (Colosimo and Pacella 2007, 2010; Colosimo
et al. 2008, 2010) when different and more complex methods, such as the ones
presented in the subsequent chapters, are considered for geometric tolerance moni-
toring.
A comparison of the performance presented by the two approaches detailed in
this chapter, along with those of the methods presented in the subsequent two
chapters for profile monitoring, is left to Chapter 11. The aim is to allow practitio-
ners to select a specific method in a given production scenario that mimics the
actual case study of roundness profiles.

Acknowledgements This research was partially funded by the Italian Ministry of University and
Research grant PRIN 2005 – “Characterization of the multiscale geometry of manufactured
surfaces for process control and tolerance analysis”.

References

Alwan LC (2000) Statistical process analysis. Irwin/McGraw-Hill, New York

Anderson TW, Darling DA (1952) Asymptotic theory of certain “goodness-of-fit” criteria based
on stochastic processes. Ann Math Stat 23:193–212
Boeing (1998) Advanced quality system tools, AQS D1-9000-1. Boeing Commercial Airplane
Group, Quality Assurance Department. http://www.boeing.com/companyoffices/doingbiz
/supplier/d1-9000-1.pdf
8 Statistical Quality Monitoring of Geometric Tolerances: the Industrial Practice 255

Box GEP, Cox DR (1964) An analysis of transformations. J R Stat Soc Ser B 26:211–252
Carr K, Ferreira P (1995a) Verification of form tolerances. Part I: basic issues, flatness, and
straightness. Precis Eng 17:131–143
Carr K, Ferreira P (1995b) Verification of form tolerances. Part II: cylindricity and straightness
of a median line. Precis Eng 17:144–156
Cho NW, Tu JF (2001) Roundness modeling of machined parts for tolerance analysis. Precis Eng
25:35–47
Cho NW, Tu JF (2002) Quantitative circularity tolerance analysis and design for 2D precision
assemblies. Int J Mach Tools Manuf 42:1391–1401
Colosimo BM, Pacella M (2007) On the use of principal component analysis to identify system-
atic patterns in roundness profiles. Qual Reliab Eng Int 23:707–725
Colosimo BM, Pacella M (2010) Control Charts for Statistical Monitoring of Functional Data,
Int J Prod Res 48(6):1575–1601
Colosimo BM, Pacella M, Semeraro Q (2008) Statistical process control for geometric specifica-
tions: on the monitoring of roundness profiles. J Qual Technol 40:1–18
Colosimo BM, Mammarella F, Petrò S (2010) Quality control of manufactured surfaces. In: Lenz
HJ, Wilrich PT (eds) Frontiers of statistical quality control, vol 9. Springer, Vienna
Gass SI, Witzgall C, Harary HH (1998) Fitting circles and spheres to coordinate measuring
machine data. Int J Flex Manuf Syst 10:5–25
ISO/TS 12181 (2003) Geometrical product specification (GPS) roundness. http://isotc213.ds.dk/
MathWorks (1991) MATLAB user’s guide. The MathWorks, Natick
Montgomery DC (2004) Introduction to statistical quality control, 5th edn. Wiley, New York
Moroni G, Pacella M (2008) An approach based on process signature modeling for roundness
evaluation of manufactured items. J Comput Inf Sci Eng 8:021003
Ryan TP (2000) Statistical methods for quality improvement. Wiley, New York
Shewhart WA (1931) Statistical method from an engineering viewpoint. J Am Stat Assoc
26:262–269
Woodall WH, Spitzner DJ, Montgomery DC, Gupta S (2004) Using control charts to monitor
process and product quality profiles. J Qual Technol 36:309–320