GD&T
GD&T
GD&T
Geometric Tolerances
Impact on Product Design, Quality Inspection
and Statistical Process Monitoring
123
Prof. Bianca M. Colosimo Prof. Nicola Senin
Politecnico di Milano University of Perugia
Via La Masa 1 Via Duranti 67
20156 Milano 06125 Perugia
Italy Italy
biancamaria.colosimo@polimi.it nsenin@unipg.it
MATLAB is a registered trademark of The MathWorks, Inc., 3 Apple Hill Drive, Natick, MA, 01760-
2098, USA, http://www.mathworks.com
Apart from any fair dealing for the purposes of research or private study, or criticism or review, as
permitted under the Copyright, Designs and Patents Act 1988, this publication may only be
reproduced, stored or transmitted, in any form or by any means, with the prior permission in writing of
the publishers, or in the case of reprographic reproduction in accordance with the terms of licences
issued by the Copyright Licensing Agency. Enquiries concerning reproduction outside those terms
should be sent to the publishers.
The use of registered names, trademarks, etc. in this publication does not imply, even in the absence of
a specific statement, that such names are exempt from the relevant laws and regulations and therefore
free for general use.
The publisher makes no representation, express or implied, with regard to the accuracy of the
information contained in this book and cannot accept any legal responsibility or liability for any errors
or omissions that may be made.
In Short
Geometric tolerances are increasingly being adopted in the design and manufac-
ture of products, and awareness of their importance is now widespread.
This is not a book on geometric tolerances per se. A significant amount of lit-
erature work, including dedicated textbooks, is available that illustrates the main
principles, definitions, and international standards related to geometric tolerances.
Instead, this book explores the impact that geometric tolerances are having in
specific areas of product development and manufacture, namely, product design,
product quality inspection, and statistical process monitoring.
The book is structured as a collection of contributions from different authors,
each highlighting one or more specific aspects related to the impact of geometric
tolerances. New issues and also new opportunities are investigated.
v
vi Preface
While this opens up more opportunities for the accurate capture of a designer’s
intent, at the same time it raises new issues. With more options to define the
boundaries of what could be considered as acceptable geometry, more choices
must be made on what, when, and how to use such new resources. In order to
make these choices wisely and effectively, deeper investigations are needed,
aimed at gaining a better understanding of the potentially numerous effects, either
direct or indirect, that any constraint on the shape/position/orientation of a geo-
metric element, such as a feature, a part, or an entire assembly, may have on func-
tional performance, manufacturability, assemblability, and ultimately cost and
time to market.
Similarly, novel problems are raised at quality inspection. As quality is now re-
lated to a more comprehensive and complex representation of geometric variation
– especially when compared with what could be previously achieved by looking
only at linear dimensions, i.e., with dimensional tolerances – it is clear that quality
inspection needs to react accordingly, and traditional approaches may not be capa-
ble of successfully supporting novel requirements. It goes without saying that
verifying whether a linear dimension is in tolerance or not is definitely simpler
than, for example, verifying that the deviation of a free-form surface with respect
to its nominal counterpart is within the specifications. The latter problem implies
the existence of a solid background of solutions for solving not-so-straightforward
issues, such as how to measure the actual geometry and how to assess from meas-
urement data if the feature variation is confined within acceptable boundaries.
Countless other examples could be made to testify how this conceptual transition
from dimensional requirements to geometric requirements is raising significant
issues, while at the same time providing the opportunity to explore new promising
ground.
A similar scenario is faced by researchers operating in statistical process moni-
toring. Once more, how can monitoring solutions evolve so that they can handle
the geometric variability associated with an entire surface, starting from traditional
approaches where monitoring relies on the variations of a few synthetic indicators,
such as those obtainable from a set of linear dimensions? Again, new issues ap-
pear alongside new opportunities to explore promising paths toward the develop-
ment of innovative solutions.
delve deep into the process, affecting design, manufacturing process planning,
quality inspection, and process monitoring activities in lots of direct and indirect
ways.
This book chooses to deal with such complexity by collecting separate testimo-
nies from different authors, each contribution concerning one or more specific
issues that are deemed relevant within one of the subject areas listed earlier as
product design, product quality inspection, and statistical process monitoring.
Manufacturing plays a central role in all the contributions: since geometric
variability is induced by the process, manufacturing-related aspects cannot be
ignored when dealing with any problem related to geometric variation. The manu-
facturing process, with its central importance, permeates the entirety of this book.
The contributions include original research, current literature reviews, and/or
industrial practice experiences related to one of the subject areas listed above. This
approach is intended to provide a significant breadth of views, while by no means
attempting to achieve completeness. The hope is that a collection of significantly
disparate topics covering such a wide range of subjects within product develop-
ment may prove inspirational for researchers pursuing original research in one of
the areas identified.
Reading Guide
The book is divided into three sections, each collecting contributions related to
one of the subject areas:
• Part I: impact on product design;
• Part II: impact on product quality inspection;
• Part III: impact on statistical process monitoring.
In the following, a brief overview of the structure and contents of the book is
presented. This is meant to provide an overall conceptual framework that allows
the chapters to be placed in their correct context, and it also works as a reference
guide, or gateway, for taking the reader toward the section he or she may be more
interested in, for further reading.
Fundamental challenges are related to how to acquire and analyze geometric in-
formation more efficiently and effectively in order to cope with the increased
requirements; new opportunities surface as well.
This part of the book addresses two main subjects in the field of inspection: ge-
ometry measurement solutions (i.e., measurement instruments and measurement
processes) and measured data analysis. For both subjects, the selection of au-
thored contributions was driven by the desire to highlight specific theoretical and
applied subject areas that have recently attracted considerable interest, such as the
inspection of microtopographic features, inspection through the adoption of sen-
sor-fusion techniques, the development of innovative measurement process plan-
ning solutions, and the development of new formal representations for encoding a
shape and its variations. Shape (form) and shape error (form error), in particular,
emerge as the central subject shared by most of the contributions presented in this
part of the book.
In detail, Chapter 3 is about measurement instruments such as profilometers
and 3D microscopes and their application to the assessment of form error on mi-
crotopographic surface features. The problem is attracting considerable interest
owing to the increased production of items such as microelectromechanical sys-
tems, semiconductors, other types of micromanufactured goods, and also more
standard-sized parts characterized by microtopographic surface features. As form
error assessment becomes relevant for such products, several issues must be faced
to make profilometers and 3D microscopes adequate – with their peculiar modes
of measurement and performance features – for such quality inspection tasks.
In Chapter 4, the problem of measurement process planning is analyzed for co-
ordinate measuring machines (CMMs) involved in form error assessment tasks. As
constraints on variability of surface shape evolve toward more comprehensive
solutions, it becomes increasingly necessary to ensure that inspection captures all
the relevant aspects of a surface geometry: for point-based measurement, this
implies denser point clouds and/or a more thought-out placement of measurement
points themselves. Significant measurement process planning problems arise, as
the requirement of a more detailed inspection clashes with measurement time and
cost.
Chapter 5 explores some issues related to the analysis of measured data. The
focus is again shifted toward microtopography, and problems related to the as-
sessment of form error on a microtopographic surface feature are analyzed, start-
ing from a cloud of points as can be obtained by means of a profilometer or 3D
microscopes. Peculiar aspects include the fact that the point cloud may be poten-
tially suboptimal for a given geometry, owing to it being generally acquired by
means of raster scanning, and that the exact localization of the surface feature
within the acquired surface region may not be known, thus implying the need for
feature identification and alignment with a nominal reference, before the actual
form error can be assessed.
In Chapter 6, recent trends toward the development of multisensor measure-
ment solutions are explored, and their role and applicability to inspection scenar-
ios involving form error and geometric tolerances are discussed. The analysis is
Preface xi
model representing surface data (namely, linear regression with spatially corre-
lated noise and principal component analysis). Chapter 10 shows a “model-free”
approach where no model of the machined surface is assumed for the development
of the monitoring strategy. In this case, the monitoring tool consists of a neural
network operating in unsupervised mode to cluster the machined surfaces. Given a
set of in-control surface data collected over a long period of time, the network is
trained to detect any different patterns as representative of a different (and possi-
bly) out-of-control state.
The performances of all three classes of approaches presented in the previous
chapters of this part are compared in Chapter 11, where a real case study concern-
ing roundness form errors of lathe-turned items is considered as a starting refer-
ence and different production scenarios are derived by slightly perturbing the case-
study features. In this chapter, the best approaches in each production scenario are
outlined in order to let the reader gain some insight into the advantages and disad-
vantages of the alternative solutions.
xiii
xiv Contents
__________________________________
A. Armillotta
Dipartimento di Meccanica, Politecnico di Milano, Via La Masa 1, 20156 Milan, Italy,
e-mail: antonio.armillotta@polimi.it
Q. Semeraro
Dipartimento di Meccanica, Politecnico di Milano, Via La Masa 1, 20156 Milan, Italy
3
4 A. Armillotta and Q. Semeraro
1.1 Introduction
For a long time, tolerancing was almost exclusively a design task. The engineering
designer had the responsibility of assigning linear tolerances on some dimensions
of parts according to all the precision requirements he could recognize on the
product. These specifications were applied by manufacturing people (engineer,
machinist, inspector), who contributed to design improvement by giving feedback
on manufacturability and costs.
Things have changed considerably during the last few decades. Increasing mar-
ket competition has forced companies to reduce defects and production costs.
Technical standards have encouraged the adoption of geometric tolerancing, thus
giving the designer more tools to ensure the assemblability and the correct func-
tion of the product. The statistical approach to production control and quality
management has made it necessary to include processing constraints in product
specifications. As a result, tolerancing is now a complex activity which involves
both design and manufacturing personnel.
However, it is at the design stage that the diffusion of geometric tolerances has
had the main impact. While geometric specifications can still be converted in the
former representation language for manufacturing purposes, the designer cannot
avoid dealing with the many types of geometric controls and datums now avail-
able. The complexity and the continuous evolution of standards make this adapta-
tion process difficult and slow.
In response to such problems, much research has been done with the aim of
supporting the assignment of geometric tolerances on manufactured parts (Hong
and Chang 2002). The methods developed can be classified into the following
three categories, related to the main design tasks of geometric tolerancing (Fig-
ure 1.1):
As is well known, linear tolerances have strong limitations with respect to in-
terpretation, inspection, and effectiveness. The specification problem should defi-
nitely consider the use of geometric tolerances, in order to allow full control of
geometric characteristics (form, orientation, location, profile, runout). The funda-
mental condition to be satisfied by specified tolerances is compliance with rele-
1 Geometric Tolerance Specification 7
vant ASME or ISO standards (ASME Y14.5.1 1994; ISO 1101 2004). Considering
the difficulty of ensuring such a condition even in common design practice, it is
not surprising that few of the proposed tolerancing methods generate fully consis-
tent specifications. For example, it has been noted that some of them do not ac-
count for some key concepts of standards, such as the priority of datums in a fea-
ture control frame (Kandikjan et al. 2001).
It is convenient to decompose the general objective of compliance to standards
into simpler acceptance conditions for tolerances resulting from a specification
procedure. A set of geometric tolerances is usually considered acceptable if it
guarantees an unambiguous positioning of each feature within controlled devia-
tions (Shah et al. 1998). The degree of positioning depends on part function. For
example, Figure 1.3 shows three different ways to tolerance an end plane of a
pump housing with respect to the opposite end plane. Option a does not constrain
the orientation of the feature to the datum, a condition that must always be satis-
fied for correct tolerancing. Option b constrains the orientation of the feature, but
not its distance to the datum. Option c is preferred if additional position control is
needed by some technical function (as is the case, owing to some tolerance chains
involving the part).
This acceptance criterion can be split into three conditions: validity, suffi-
ciency, and consistency (Willhelm and Lu 1992). A tolerance can be invalid, i.e.,
impossible to satisfy in practice, owing to a poor choice of its value (not accounted
for in the tolerance specification). The condition of insufficiency has already been
discussed for the example in Figure 1.3. An inconsistent tolerance loses its mean-
ing owing to poor control of a referenced datum; in Figure 1.3, the lack of a flat-
ness control on datum A affects the unambiguous positioning of the toleranced
feature.
8 A. Armillotta and Q. Semeraro
Figure 1.4 gives an overview of the data that a designer should collect in order to
specify geometric tolerances. Except for some additional information that may be
required when reasoning about geometric controls, the input data are quite similar
to those needed in the context of linear tolerancing. In general, two types of data
are absolutely needed for both empirical and computer-aided specification:
• a geometric description of parts and assembly relations; and
• a list of design requirements involved arranged by either product function or
the need for a correct assembly.
They will be discussed in the following subsections. The remaining types of
data (nongeometric and process-related) may not be strictly needed for a qualita-
tive tolerancing of parts, although they could be of help in downstream design
tasks (tolerance allocation and analysis). For instance, materials and surface fin-
ishes selected on parts lead to the choice of manufacturing processes, which may
influence applicable tolerances. Similarly, any information that may be available
on the assembly process (order of operations, level of automation, fixtures) could
impose special precision requirements on apparently nonfunctional features (Roy
and Bharadwaj 1996). Obtaining such data in advance can support early choices
among alternative specifications.
The complexity of the product and the type of software support desired for tol-
erance specification suggest to what extent input data should be structured. To
achieve a reasonable level of integration, they should be organized into formal
assembly models to be extracted from product data with the interactive support of
a 3D solid modeler (Kim et al. 2004).
A product has a nominal geometry, upon which allowable variations will be de-
fined through tolerances. Geometric data describe the individual parts and their
mutual relations in the assembly.
Each part carries some general properties, such as the material (if deemed use-
ful for tolerance specification) and the possible existence of identical parts in the
assembly. The detailed description of the part is a list of its functional features,
i.e., those surfaces or features of size1 that have geometric relations with features
of mating parts. The list may be extended to nonfunctional features if they are to
be controlled, as is advocated by practitioners in some industrial sectors comply-
ing to ASME standards.
1
A feature of size is a cylindrical or spherical surface, or a set of two opposed elements or op-
posed parallel surfaces, associated with a size dimension (ASME Y14.5.1 1994). Pins, holes,
tabs, slots, and any other feature that can be associated with a size dimension are examples of
features of size.
10 A. Armillotta and Q. Semeraro
The functional features identified for an example part (crankshaft) are shown in
Figure 1.5, along with a possible description of one of its features. Essential in-
formation for any tolerancing procedure includes the shape of the feature and the
spatial orientation of its representative geometric entity (derived feature), as well
as the number of equal features within a possible pattern. Further information can
help to select features to be put in the datum reference frame (DRF) of the part.
Depending on the specification method, this choice could privilege features with
such favorable properties as large size, symmetry, and ease of access to reference
surfaces of fixtures and gages.
Assembly-level data describe all pairwise relations between features of differ-
ent parts. Figure 1.6 shows the assembly relations identified for an example prod-
uct (rotary compressor) which includes the part shown in Figure 1.5. As will be
clarified later, each relation is usually associated with one or more design re-
quirements and can thus influence the choice of tolerance types on the features
involved.
Assembly relations can be regarded as simple contacts between parts, without
further detail. For a generative specification procedure, however, it is better to
distinguish assembly relations with respect to the actual contact conditions they
establish between parts. A possible classification of the most common cases is
exemplified in Figure 1.6. It includes fits between features of size (a), seating
relations with at least three points of actual contact (b), simple relations with at
least one point of contact (c), and nominal relations with no actual contact re-
quired but a coincidence of ideal surfaces (d). Similar classifications have been
proposed with an emphasis on the degrees of freedom subtracted to the relative
motion of mating parts (Wang et al. 2003).
For empirical tolerancing, a designer should recognize geometric product data
on engineering drawings and annotate them properly for later reference. Differ-
ently, the development of a generative tolerancing tool may require the extraction
of product data from CAD models of assemblies and parts. Actually, only a few
specification methods proposed in the literature provide such an integration as an
alternative to a tedious input of data through formatted text files or graphical user
interfaces.
1 Geometric Tolerance Specification 11
What is the purpose of asking for precision for a product? A first answer is related
to product function, which may impose an accurate manufacture of parts in order
to preserve the kinematics of mechanisms or the positioning of machine compo-
nents with respect to external constraints. Some conditions of this type, which will
be referred to as functional requirements, are shown in Figure 1.7, part a for an
example product (belt drive assembly). The grooves of the pulley must be accu-
12 A. Armillotta and Q. Semeraro
Regardless of the way they are collected (interactive, data-driven), design re-
quirements should be carefully treated as they are in a close relationship with the
tolerances to be specified. Any specification method should include a flowdown
procedure from requirements to tolerances. Although no complete flowdown pro-
cedure is available in the literature for geometric tolerances, some reasoning rules
have been proposed. In one case (Wang et al. 2003) some types of functional
requirements deriving from mechanical functions (sealing, rotation, balance, gear-
ing, fastening, sliding, press fit, and clearance fit) are classified and related to
typical choices for datums and tolerances. In another approach, not reported in
detail (Roy and Bharadwaj 1996), such high-level requirements are converted into
an intermediate form (equivalent functional specifications) which is claimed to be
linked to tolerances through geometric rules.
control frame, which is shown in Figure 1.8 for the requirements discussed above
for Figure 1.7. The same notation has been used to define functional tolerances in
a generative specification method (Mejbri et al. 2005).
In another proposed definition, assembly requirements are treated as generali-
zations of fits between features of size (Voelcker 1998). This concept, referred to
as maximum material part in analogy to the maximum material principle for fits,
does not seem to have been developed further.
Definitions and drawing rules are not sufficient for a generative specification
procedure, which also needs a representation model for design requirements. Such
a model should codify requirements in a format allowing their easy association
with predefined tolerance specifications.
The model of virtual boundary requirements has been proposed for require-
ments involving one or two parts, such as a fit or the protection of a minimum
amount of material (Jayaraman and Srinivasan 1989; Srinivasan and Jayaraman
1989). A requirement is a set of conditions, each involving a pair of features of
two different parts. For each pair of features, a virtual boundary is defined as the
geometric entity which separates them at their maximum material conditions (a
concept deriving from virtual conditions of clearance fits). The half-spaces con-
structed on the virtual boundary define the conditions of the requirement (e.g., a
hole surrounds a peg, and simultaneously two planes are in contact). Should the
concept be generalized to a sufficiently large set of functional and assembly re-
quirements, it could be the basis for a tolerance specification procedure (Willhelm
and Lu 1992).
A more recent approach treats precision requirements in close connection with
geometric relations and tolerances (Dufaure et al. 2005). The product is decom-
posed from both a structural and a functional viewpoint by defining three entities:
parts, interfaces (surfaces, points, and lines), and functions. A function can repre-
sent a functional requirement, a contact condition between surfaces on different
parts, or a geometric specification between surfaces on the same part. A graph
provides a synthetic view of the model by showing possible relationships among
entities: for example, functions are relationships among two or more interfaces,
and parts are in turn technical solutions of functions. A software implementation
of the model has been provided along with a demonstrative example (Teissandier
and Dufaure 2007).
Another representation model connects geometric relations, precision require-
ments, and tolerances in a single hierarchy of hypergraphs (Giordano et al. 2005).
The model describes all the above-mentioned items at different levels of detail,
from the simple kinematic scheme of a mechanism to a full geometric description
of parts. The authors foresee an automatic generation of the model, which would
be the basis of an approach to tolerance specification.
As has already been noted, some design requirements can be identified from geo-
metric data without having to be explicitly set by the designer. Although it is not
clear how an automated recognition can be done in generic cases, some indications
could derive from approaches proposed for linear tolerances.
In some cases, requirements are identified by search algorithms. In Mullins and
Anderson (1998), search in a graph model of the assembly and geometric reason-
ing on contact surfaces among parts allowed the identification of some types of
assembly requirements, such as the ending gaps of tolerance chains (mating condi-
16 A. Armillotta and Q. Semeraro
tions). A similar method was proposed in Zou and Morse (2004) for different
types of gaps that can be created among parts in an assembly.
In Islam (2004a, b), a systematic procedure for identifying functional require-
ments was proposed within a method for the allocation of tolerance values on 1D
chains. Again, the objective is to develop high-level requirements involving tech-
nical and safety considerations to a level of detail corresponding to part features
(gaps, fits, flatness, location, etc.). The procedure is based on the function analysis
and system technique (FAST), commonly used in the redesign of products by the
value analysis method. The determination of limiting values for requirements is
left to the experience of the designer. The development of a CAD-based software
tool for the same purpose was described in Mliki and Mennier (1995), but details
of identification procedures were not given.
Besides identifying design requirements, some methods can even generate their
algebraic relations with dimensions of the parts involved. These are referred to as
functional equations and are essential for tolerance analysis and allocation. The
generation of functional equations has been studied in the context of linear toler-
ances. In Ramani et al. (1998), equations generated by a CAD-based tool which
identifies all possible tolerance chains from functional surfaces and critical dimen-
sions selected by the user were reported. In Söderberg and Johannesson (1999),
tolerance chains were identified as loops in a network of relations among part
features in an assembly model. For each part or subassembly, the model includes
functions and reference frames deriving from relations with mating parts. Wang
et al. (2006) generated 3D tolerance chains by a variational approach based on a
solid modeler for geometric computations. Once a functional requirement has been
interactively defined, the procedure identifies the surfaces involved in the chain.
For this purpose, each surface is displaced from its original position by a small
perturbation, and possible changes in the value of the functional requirement are
detected after regenerating the geometric model. The chain results from the adja-
cencies of parts along the three reference directions.
Two additional problems can be cited to conclude the discussion of design re-
quirements. They are less studied, but potentially important in order to integrate
tolerance specification into the general process of product development. As a first
question, how are design requirements (and indirectly tolerance specifications)
influenced by choices related to the assembly process? Preliminary discussions
and conceptual examples are included in some of the already-cited papers (Whit-
ney 1996; Whitney et al. 1999; Sudarsan et al. 1998; Narahari et al. 1999). Further
developments in this direction could help to plan assembly processes with reduced
need for part tolerancing.
Equally interesting is the chance to reduce the number of requirements through
product design choices. Empirical knowledge for this purpose exists and could be
incorporated in computer-aided procedures. Some general rules are collected in
McAdams (2003) and were classified according to their relevance in the different
stages of product development and to the type of redesign involved (parameter
calibration, detail changes, revision of product architecture).
1 Geometric Tolerance Specification 17
Datum selection has been treated in tolerancing handbooks based on ASME stan-
dards (Meadows 1995; Drake 1999; Cogorno 2006) and case studies of tolerance
specification (Wang et al. 2003). Figure 1.9 shows valid DRFs for some parts of
an example product (wheel assembly, Figure 1.9a) in order to clarify the use of
available rules. For all parts, datum precedences follow alphabetical order (A,
B, C).
Figure 1.9 Examples of datum reference frames: a wheel assembly, b support, c axle, d wheel,
and e plates
18 A. Armillotta and Q. Semeraro
Since datums are often chosen among the functional features of a part, a basic
rule suggests selecting the features that establish the most important assembly
relations for product function. Such features may be rotation axes (support in Fig-
ure 1.9b, axle in Figure 1.9c, wheel in Figure 1.9d) or mating planes through which
major forces are transmitted to the mechanism (support in Figure 1.9b, upper plate
in Figure 1.9e). Actually, the importance of a functional relation cannot be easily
evaluated in a data-driven procedure without decomposing it into geometric rea-
soning rules. A criterion proposed for this purpose is a large contact surface with
mating parts, which allows repeatable fixturing and measurement (primary datums
in Figure 1.9b–d). Another one is the number of degrees of freedom that mating
parts subtract from part motion through the feature: rotational (seating) constraints
suggest the choice of primary datums, whereas translational (locating) constraints
determine secondary or tertiary datums (all parts in Figure 1.9).
Any different configuration of the DRF requires a specific design of manufac-
turing fixtures. Therefore, it is also desirable that datum features are easily ac-
cessed and simulated in machining and inspection, a condition that usually sug-
gests a preference for planar and cylindrical surfaces.
The above rules are only a partial answer to the need for a guided choice of da-
tums. To restrict the number of possible solutions, handbooks describe the most
common types of DRFs. These include rotation axes defined by single or multiple
features (Figure 1.9c, d), 3-2-1 schemes with three perpendicular planes (Fig-
ure 1.9e) or two planes and an axis (Figure 1.9b), and plane–cylinder sets with or
without clocking. Each frame usually comes with suggestions for datum priorities
and tolerance types to be selected for datums and targets. In the choice among
alternative solutions, another suggested criterion is the minimum number of datum
features, cited in Farmer and Gladman (1986) and Weill (1988) and demonstrated
with examples of part redesign. As to current knowledge, the judgement of the
designer is ultimately essential to select datum priorities for parts which do not
match the most common cases (Mejbri et al. 2005).
Figure 1.10 Tolerancing guidelines as per ASME Y14.5.1 (1994). DRF datum reference frame
1 Geometric Tolerance Specification 19
Once a DRF has been selected, tolerances have to be chosen on all target fea-
tures of the same part. Apart from the geometric classifications described below,
some empirical guidelines for this purpose are provided by standards (ASME
Y14.5.1 1994). For instance, the tolerance on a feature of size is chosen according
to feature location (coaxiality to a datum axis) and to explicit design requirements
(interchangeability, alignment, protection of minimum volumes, etc.). A partial
outline of the guidelines is shown as a schematic flowchart in Figure 1.10.
(rotation about z) applies to the DRF. According to the above rule, plane C has to
be controlled on a translation and two rotations perpendicular to it, which requires
a profile tolerance. Controls of remaining features are similarly determined. It
should be noted that kinematic control may not be the only criterion for the selec-
tion of the tolerance type: for example, the function of plane G (mating with a
screw head) can be fulfilled by a simple orientation control rather than by a loca-
tion control as suggested by its invariances.
A first list of tolerancing cases can be built by the tolerance representation
model proposed in Clément et al. (1994) and also described in Clément and
Rivière (1993), Weill (1997), and Chiabert and Orlando (2004). The model is
based on a classification of part surfaces in seven classes according to their invari-
ance with respect to translations and rotations. The defined types are spherical,
planar, cylindrical, helical, rotational, prismatic, and generic (i.e., noninvariant)
surfaces (shown in reverse order in Figure 1.13, part a). Each invariance class is
characterized by a minimum geometric datum element (MGDE), i.e., the set of
geometric entities (point, line, and plane) which is sufficient to determine the
spatial position of the surface through dimensions and tolerances. Figure 1.13,
part b shows the MGDE and the number of translational, rotational, and roto-
translational invariances for each class.
The possible associations between two surfaces belonging to the same part are
called technologically and topologically related surfaces (TTRSs). The composi-
tion of the MGDEs of the two associated surfaces allows one to determine the
MGDE of the resulting TTRS, which is reclassified within one of the original
invariance classes. In this way, the TTRS can be associated in turn with other
surfaces or TTRSs, thus iterating the tolerancing process. Considering all cases
deriving from possible geometries of associated surfaces, 44 composition rules are
identified. These correspond to an equal number of tolerancing cases of a feature
with respect to another.
1 Geometric Tolerance Specification 21
For example, Figure 1.13, part c shows the three cases of TTRSs that occur
when associating a cylinder (target) with a plane (datum). If the cylinder is per-
pendicular to the plane, the resulting TTRS belongs to the invariance class of
rotational surfaces; according to the same considerations made for the example in
Figure 1.12, such a case corresponds to a perpendicularity tolerance on the target.
If the cylinder is parallel to the plane, the TTRS belongs to the invariance class of
prismatic surfaces and corresponds to a position tolerance with a basic dimension
(distance of the cylinder axis to the plane). If the cylinder is oblique to the plane,
the TTRS belongs to the invariance class of generic surfaces and corresponds to a
position tolerance with two basic dimensions (angle of the cylinder axis to the
plane, distance of an arbitrary point on the cylinder axis to the plane).
An extension to the TTRS model has been proposed in order to classify asso-
ciations between surfaces of different parts (Clément et al. 1995). Such associa-
tions, called pseudo-TTRSs, are an additional way to represent precision require-
ments. Based on them, it has been demonstrated how functional requirements
defined at a high level (e.g., minimize the noise of a gear train) can be hierarchi-
cally translated down to a geometric level (Toulorge et al. 2003).
An interesting consequence of the above concepts is that a limited set of toler-
ancing cases exist and can be explored in a generative specification procedure.
Regarding the number of cases, the classification based on the TTRS model is not
the only one, as it assumes a definition of invariance classes which may not cover
the diversity of features designed on manufactured parts.
The ASME “math standard” (ASME Y14.5.1.M 1994) classifies the cases of
DRFs that can be constructed from elementary geometric entities (points, lines,
and planes), considering the possible priority orders among datums. Each combi-
nation is described by its invariants, defined as dimensions (distances or angles)
22 A. Armillotta and Q. Semeraro
Figure 1.13 Classification of tolerancing cases by the technologically and topologically related
surfaces (TTRS) model. MGDE minimum geometric datum element, T translational, R rotational
RT roto-translational
which do not change after free transformations allowed by a datum system. Possi-
ble invariants include the three principal translations and rotations, the distance to
an axis or point, and the angle to an axis. A set of 52 cases of DRFs is classified, a
larger number than the number of associations between TTRSs since datum priori-
ties are considered. Classified datum systems are then grouped according to in-
variants, thus giving six generic cases (one less than the number of TTRS invari-
ance classes as the helical surface is not explicitly considered in the standard).
Another tolerance representation model (Shah et al. 1998; Kandikjan et al.
2001) is based on elementary entities coinciding with (or derived from) features
on a part. Each entity has its translational and rotational degrees of freedom, which
can be regarded as the complementary set of invariances. Any combination be-
tween two entities corresponds to the relation between a target and a datum, and
maintains a number of degrees of freedom depending on the two original entities.
Considering all possible combinations between two or three entities, 31 cases of
DRFs and related tolerance cases are classified. Despite the analogy between the
two approaches, the number of cases is smaller than in the TTRS model as the
classification does not explicitly account for degenerate cases (such as coincident
points, lines, or planes) and does not treat helical surfaces as separate cases.
A further model is based on similar considerations based on degrees of freedom
(Hu et al. 2004). It consists in classifying all combinations between elementary
1 Geometric Tolerance Specification 23
Several methods have been proposed for the selection of DRFs and tolerance types
on part features. In some cases, the results are directly transferred to tolerance
allocation and analysis, thus using the specification method as a part of a more
comprehensive tolerancing procedure. Some solutions have been implemented in
software tools with possible CAD integration.
The methods developed at sufficient completeness and generality are based on
six different approaches to tolerance generation:
Table 1.1 compares them with respect to the overall strategy adopted for the
treatment of product data. Basic criteria include the existence of an a priori classi-
fication of tolerancing cases, the use of a global description model of assembly
relations, and the explicit treatment of design requirements.
As said before, a classification of tolerancing cases allows one to treat single
steps of the problem with reference to a set of predefined cases. To develop a
specification method from a classification, features to be toleranced on each part
have to be properly ordered. The approaches based on this principle differ in the
way they analyze product geometry to identify priorities among features. A classi-
fication helps to improve the correctness and completeness of the tolerances gen-
erated, and usually has a mathematical foundation which can allow one to generate
functional equations for downstream tolerancing tasks.
24 A. Armillotta and Q. Semeraro
The first and most cited specification method is based on the TTRS model for
tolerance representation and classification (Clément et al. 1994). The general
approach, which has been described before, models each surface by an invariance
class and a set of reference geometric entities (MGDEs). When two surfaces are
considered in association with each other, these data allow one to determine the
tolerances needed to control their relative position. The association itself (TTRS)
is modeled with an invariance class and a MGDE. This concept is exploited in a
recursive strategy, which builds a hierarchy of TTRSs on the functional features
(here called “surfaces”) of each part.
The sequence of surface associations on a part derives from a graph which
represents surface relations among parts in an assembly. Any of the loops of the
graph involve two or more parts: for each part, the loop includes two surfaces,
which can be associated in a TTRS. By exploration of all loops in a user-selected
order, the TTRSs defined on the parts are composed to form a hierarchy. This
allows one to identify applicable tolerance types according to the classified toler-
ancing cases from the properties of the TTRSs (invariance class and MGDE).
Figure 1.15 shows the recursive procedure that builds the TTRS hierarchy for
the example. The graph is initialized with the relations between pairs of surfaces
corresponding to functional features of parts (Figure 1.15a). The loops of the
graph are ordered in a sequence (L1, L2, L3, L4) which is then followed when
associating surfaces into TTRSs. For example, loop L1 involves two surfaces on
part 1 (S12 and S13) which are associated into T1, and two surfaces on part 2 (S21
and S22) which are associated into T2. The next loops allow one to build the other
TTRSs, namely, T3 and T4 (Figure 1.15b), T5 and T6 (Figure 1.15c) and eventu-
ally T7, T8, and T9 (Figure 1.15d).
For complex assemblies, the loops have to be properly ordered in order to con-
struct a TTRS hierarchy consistent with design practice. For this purpose, prede-
fined tolerancing templates are provided for common functional requirements on
the basis of technical information published by vendors of components (bearings,
slideways, gears, seals, etc.).
The TTRS-based specification method has been developed into an interactive
software tool (Salomons 1995; Salomons et al. 1996). In addition to the original
method, some criteria have been introduced to order the loops of the graph and to
construct the MGDE of a TTRS from the MGDEs of associated features. These
26 A. Armillotta and Q. Semeraro
tasks remain mostly left to user interaction, along with the initial preparation of
the assembly model in the lack of a direct CAD interface.
Loop and MGDE selection are two critical steps for an implementation of the
method at a higher automation level. It has been suggested (Desrochers and Ma-
ranzana 1995) that priority should be given to the loops that (1) involve the mini-
mum number of parts, (2) are related to tolerance chains detected from sequences
of contacts along a same direction, and (3) are of large size or involve specific
types of features such as planes. For the choice of the MGDE, the following crite-
ria help to treat the cases in which a TTRS has a different invariance class from
associated features: (1) specification of datums and basic dimensions with a func-
tional interest and (2) predefinition of clearances or interferences on fits.
The construction of TTRSs on geometric models of assemblies and parts has
been implemented in commercial CAD packages. In one case (Toulorge et al.
2003) a semiautomatic software tool is integrated in the product life cycle man-
agement system of an automotive manufacturer, with the already-described exten-
sion of the model (pseudo-TTRSs) to represent functional requirements (here
referred to as use aptitude conditions). A more complete implementation in the
Figure 1.15 TTRS hierarchy: a loops,surfaces, and parts, b building of T3 and T4, c building
of T5 and T6, and d building of T7, T8, and T9
1 Geometric Tolerance Specification 27
same context was reported in Buysse et al. (2007): to streamline data input,
TTRSs are selected by the user through a graphical procedure which displays the
related MGDE as a skeleton geometric representation of parts.
Figure 1.16 Metric relations in the degree-of-freedom method. SP shape, r radius, PP perpen-
dicularity, CI coincidence, C coaxiality
28 A. Armillotta and Q. Semeraro
1.5.3 Mirrors
previously toleranced part, the selection of datums is influenced by the relation (in
practice, the datum system of the previous part is “reversed” on the new part).
Form, orientation and location, tolerances are then assigned to datums and targets
by a rule-based procedure. In the second step (final tolerancing), other rules are
applied to harmonize tolerances on parts based on the whole set of assembly rela-
tions with other parts.
Figure 1.17 shows the use of mirrors on the example. Part 1 is considered as the
most important since it has two mirrors, both strong owing to the occurrence of a
minimum of three points of actual contact. One of the two mirrors (feature 1.2) is
preferred as the primary datum of the part since it is more readily accessible in
manufacturing and inspection. Each mirror is taken as a starting point to cluster
adjacent features, which are temporarily toleranced by the already-cited ASME
classification of tolerancing cases. For both clusters of the example, the temporary
tolerancing scheme consists of a flatness control on the plane and a perpendicular-
ity control on the cylindrical feature.
The mirror method has been demonstrated on a realistic example with results in
good accordance with design practice. No details are available on software im-
plementation and further validation of the approach.
(first and third steps) and in series (second step). By keeping note of graph simpli-
fications, one obtains a decomposition of the functional requirement as a hierarchy
of features (Figure 1.18b). Within the hierarchy, features which do not influence
the requirement are discarded from further consideration: specifically, all cylindri-
cal contacts do not contribute to the translational constraint of functional require-
ment and thus do not need to be toleranced (at least with respect to the functional
requirement).
The procedure described has been incorporated in interactive specification
tools, aimed at supporting designers during the whole product development cycle,
starting from conceptual design (Dantan et al. 2003a). To better link input data to
product design information, functions defined at a high level are converted into a
set of key characteristics related to part features. In addition, an attempt is made to
consider the effect of the assembly plan on tolerances. In some cases, the fulfill-
ment of a functional requirement (e.g., a clearance fit) depends on the configura-
tion established in previous assembly operations. The dependence is modeled by
means of rules for the selection of maximum-material and least-material modifiers
(Dantan et al. 2003b, 2005), defined from a tolerance representation based on the
virtual boundary model (Jayaraman and Srinivasan 1989).
Each design requirement usually involves only a limited number of parts. For this
reason, the generation of the tolerances needed to satisfy a requirement can be
done without considering the whole graph of relations among part features in an
assembly. This basic consideration introduces the specification method proposed
in Anselmetti and Mawussi (2003) and Mejbri et al. (2003, 2005).
A functional requirement is defined as a geometric tolerance assigned to an
ending entity on a part with respect to external datums on a base component. Such
a “generalized” tolerance must be converted into a set of regular tolerances (i.e.,
with only internal datums) on the individual parts involved in the requirement. For
this purpose, a positioning table is defined for each part to represent contact rela-
tions with mating parts. The table lists the contacts in an ordered set, correspond-
ing to a priority order of datums for all requirements involving the part. Datum
order is defined by the designer according to the types of actual contacts required
between features (seating contacts on three actual points, simple contacts, etc.) and
to the need to define complex cases of datum frames (composite datums, symme-
tries, feature patterns, etc.).
With use of the positioning tables of parts, a chain of relations (mechanical
joints) up to the base component is constructed for each functional requirement.
The chain is then explored recursively from the first part by following the chain of
relations. At each step, the external datums of the geometric tolerance related to
the requirement are converted to internal datums.
The DRF used for each tolerance generated is validated, i.e., reduced to the
simplest subsystem that is able to unambiguously control the spatial position of
the tolerance zone. The validation criterion is based on the invariant degrees of
freedom, i.e., the translations and rotations with no effect on the tolerance zone. If
a datum deriving from the contact with another part is not needed to constrain the
degrees of freedom of the tolerance zone, that part is considered inactive. The
validation is based on a classification of tolerance zones, considered in combina-
tion with possible feature types. A set of validation rules allow one to verify if a
certain system of datums contributes to constraining the degrees of freedom, i.e.,
eliminates degrees of freedom which are noninvariant for the tolerance zone.
Figure 1.19, part a shows the positioning table for part 3 of the example. The
first column of the table means that a contact exists between planar features 3.1
and 1.4. The remaining columns are similarly interpreted with other types of fea-
tures and relations. In Figure 1.19, part b, each relation generates a tolerance on
the feature involved according to a rule drawn from a classification. The column
order is set by the designer and determines the priority of datums (A and B for this
part).
The approach was later completed with an automated generation of functional
equations for tolerance allocation (Anselmetti 2006). In addition to functional
requirements defined by the user, some assembly requirements are automatically
identified from geometric data. These include interferences and alignments be-
32 A. Armillotta and Q. Semeraro
Figure 1.20 Variational loop circuit for a functional requirement. DRF datum reference frame,
NF nominal feature, AF associated feature
Figure 1.20 shows the corresponding graphs for the functional requirement
considered for the example. Since the requirement involves only planar surfaces of
just two parts (1 and 3), the circuit is very simple and does not have derived fea-
tures (e.g., axes). The NLC is first built with only nominal features. Each nominal
feature is linked to the features (a single plane in this case) of the DRF of the same
part (DRF1 or DRF3). Each link represents the geometric relation to the datum,
which consists of one of distances D1, D2, and D3. To build the VLC, the associ-
ated features are added. Each of them is linked to the corresponding nominal fea-
ture through geometric parameters (not shown in the figure) that allow one to
select the proper type of tolerance and the possible need for basic dimensions from
a classification of tolerancing cases.
1.6 Conclusions
What can be learned from this review on tolerance specification? It seems clear
that the introduction of geometric tolerances in design practice has created a prob-
lem that did not exist with linear tolerancing. The designer is bound to declare the
degree of geometric control of every part feature, in order to guarantee the fit and
function of the product. This is not easy, since many types of geometric controls
exist and their standard representation language is complex. While this difficulty
can be overcome through training and experience, it remains hard for a designer to
reason on the combinatorial explosion of assembly relations and design require-
ments of a complex product.
Research is attempting to remove these obstacles. Classifications of tolerancing
cases and guided procedures are increasingly available for training and design
practice. Generative methods are being developed and integrated in CAD software
as a further help when dealing with complex assemblies. They are not intended as
push-button tools and usually involve a moderate amount of user interaction.
34 A. Armillotta and Q. Semeraro
Some indications can be drawn for future work on tolerance specification. The
main developments are expected in three areas: the treatment of product data, the
general strategy for problem solution, and the compatibility with downstream
tolerancing tasks.
To allow a complete description of the product, some improvements need to be
made to the treatment of design requirements. Owing to their large number, it is
impractical for the identification of requirements to be left completely to the de-
signer. Procedures will have to be developed for the automated identification of
assembly requirements, and functional requirements will have to be better modeled
and classified. The effect of the assembly process on tolerances will have to be bet-
ter clarified. Improved CAD interfaces will help to streamline the construction of
product description models without tedious and error-prone interactive procedures.
To improve tolerance generation, a balance will have to be found between an
explicit processing of precision requirements, ideal to treat complex chains of
geometric tolerances, and a global strategy guided by feature relations at the as-
sembly level. The formalization of technical knowledge will have to be extended
by rules for the selection of tolerance types and datums in typical design situa-
tions. The need to cover the whole set of tolerances provided by technical stan-
dards will probably require an extension of existing reasoning rules and classifica-
tions of tolerancing cases.
To extend the scope of specification methods, an integration with computer-
aided tools for tolerance analysis will have to be ensured. The interface will have
to involve an automated generation of geometric tolerance chains and related func-
tional equations. This will allow a seamless integration of tolerancing activities
and a significant efficiency gain in the design of mechanical assemblies.
References
Abdullah TA, Popplewell K, Page CJ (2003) A review of the support tools for the process of
assembly method selection and assembly planning. Int J Prod Res 41(11):2391–2410
Anselmetti B (2006) Generation of functional tolerancing based on positioning features. Comput
Aided Des 38:902–919
Anselmetti B, Mawussi K (2003) Computer aided tolerancing using positioning features. ASME
J Comput Inf Sci Eng 3:15–21
ASME Y14.5.1 (1994) Dimensioning and tolerancing. American Society of Mechanical Engi-
neers, New York
ASME Y14.5.1.M (1994) Mathematical definition of dimensioning and tolerancing principles.
American Society of Mechanical Engineers, New York
Ballu A, Mathieu L (1999) Choice of functional specifications using graphs within the frame-
work of education. In: Proceedings of the CIRP seminar on computer-aided tolerancing,
Twente, pp 197–206
Bjørke O (1989) Computer-aided tolerancing. ASME Press, New York
Buysse D, Socoliuc M, Rivière A (2007) A new specification model for realistic assemblies
simulation. In: Proceedings of the CIRP international conference on computer-aided toleranc-
ing, Erlangen
1 Geometric Tolerance Specification 35
Carr CD (1993) A comprehensive method for specifying tolerance requirements for assembly.
ADCATS report 93-1
Chiabert P, Orlando M (2004) About a CAT model consistent with the ISO/TC 213 last issues. J
Mater Proc Technol 157/158:61–66
Clément A, Rivière A (1993) Tolerancing versus nominal modeling in next generation
CAD/CAM system. In: Proceedings of the CIRP seminar on computer-aided tolerancing, Ca-
chan, pp 97–113
Clément A, Rivière A, Temmerman M (1994) Cotation tridimensionelle des systèmes mé-
caniques. PYC, Yvry-sur-Siene
Clément A, Rivière A, Serré P (1995) A declarative information model for functional require-
ments. In: Proceedings of the CIRP seminar on computer-aided tolerancing, Tokyo, pp 3–16
Cogorno GR (2006) Geometric dimensioning and tolerancing for mechanical design. McGraw-
Hill, New York
Dantan JY, Anwer N, Mathieu L (2003a) Integrated tolerancing process for conceptual design.
CIRP Ann 52(1):135–138
Dantan JY, Ballu A, Martin P (2003b) Mathematical formulation of tolerance synthesis taking
into account the assembly process. In: Proceedings of the IEEE international symposium on
assembly and task planning, Besançon, pp 241–246
Dantan JY, Mathieu L, Ballu A, Martin P (2005) Tolerance synthesis: quantifier notion and
virtual boundary. Comput Aided Des 37:231–240
Desrochers A, Maranzana R (1995) Constrained dimensioning and tolerancing assistance for
mechanisms. In: Proceedings of the CIRP seminar on computer-aided tolerancing, Tokyo,
pp 17–30
Drake PJ (1999) Dimensioning and tolerancing handbook. McGraw-Hill, New York
Dufaure J, Teissandier D, Debarbouille G (2005) Influence of the standard components integra-
tion on the tolerancing activity. In: Proceedings of the CIRP international seminar on com-
puter-aided tolerancing, Tempe, pp 235–244
Farmer LE, Gladman CA (1986) Tolerance technology: computer-based analysis. CIRP Ann
35(1):7–10
Giordano M, Pairel E, Hernandez P (2005) Complex mechanical structure tolerancing by means
of hyper-graphs. In: Proceedings of the CIRP international seminar on computer-aided toler-
ancing Tempe, pp 105–114
Hong YS, Chang TC (2002) A comprehensive review of tolerancing research. Int J Prod Res
40(11):2425–2459
Hu J, Xiong G (2005a) Dimensional and geometric tolerance design based on constraints. Int J
Adv Manuf Technol 26:1099–1108
Hu J, Xiong G (2005b) Concurrent design of a geometric parameter and tolerance for assembly
and cost. Int J Prod Res 43(2):267–293
Hu J, Xiong G, Wu Z (2004) A variational geometric constraints network for a tolerance types
specification. Int J Adv Manuf Technol 24:214–222
Islam MN (2004a) Functional dimensioning and tolerancing software for concurrent engineering
applications. Comput Ind 54:169–190
Islam MN (2004b) A methodology for extracting dimensional requirements for a product from
customer needs. Int J Adv Manuf Technol 23:489–494
ISO 1101 (2004) Geometrical product specifications (GPS) – geometrical tolerancing – toler-
ances of form, orientation, location and run-out. International Organization for Standardiza-
tion, Geneva
Jayaraman R, Srinivasan B (1989) Geometric tolerancing: I. Virtual boundary requirements. IBM
J Res Dev 33(2):90–104
Kandikjan T, Shah JJ, Davidson JK (2001) A mechanism for validating dimensioning and toler-
ancing schemes in CAD systems. Comput Aided Des 33:721–737
Kim KY, Wang Y, Mougboh OS, Nnaji BO (2004) Design formalism for collaborative assembly
design. Comput Aided Des 36:849–871
36 A. Armillotta and Q. Semeraro
McAdams DA (2003) Identification and codification of principles for functional tolerance de-
sign. J Eng Des 14(3):355–375
Meadows JD (1995) Geometric dimensioning and tolerancing: applications and techniques for
use in design, manufacturing and inspection. Dekker, New York
Mejbri H, Anselmetti B, Mawussi B (2003) A recursive tolerancing method with sub-assembly
generation. In: Proceedings of the IEEE international symposium on assembly and task plan-
ning, Besançon, pp 235–240
Mejbri H, Anselmetti B, Mawussi K (2005) Functional tolerancing of complex mechanisms:
identification and specification of key parts. Comput Ind Eng 49:241–265
Mliki MN, Mennier D (1995) Dimensioning and functional tolerancing aided by computer in
CAD/CAM systems: application to Autocad system. In: Proceedings of the INRIA/IEEE
symposium on emerging technologies and factory automation, Paris, pp 421–428
Mullins SH, Anderson DC (1998) Automatic identification of geometric constraints in mechani-
cal assemblies. Comput Aided Des 30(9):715–726
Narahari Y, Sudarsan R, Lyons KW et al (1999) Design for tolerance of electro-mechanical
assemblies: an integrated approach. IEEE Trans Robot Autom 15(6):1062–1079
Nassef AO, ElMaraghy HA (1997) Allocation of geometric tolerances: new criterion and meth-
odology. CIRP Ann 46(1):101–106
Pérez R, De Ciurana J, Riba C (2006) The characterization and specification of functional re-
quirements and geometric tolerances in design. J Eng Des 17(4):311–324
Ramani B, Cheraghi SH, Twomey JM (1998) CAD-based integrated tolerancing system. Int J
Prod Res 36(10):2891–2910
Roy U, Bharadwaj B (1996) Tolerance synthesis in a product design system. In: Proceedings of
NAMRC, Ann Arbor
Salomons OW (1995) Computer support in the design of mechanical products – constraints
specification and satisfaction in feature based design and manufacturing. PhD thesis, Univer-
sity of Twente
Salomons OW, Jonge Poerink HJ, Haalboom FJ et al (1996) A computer aided tolerancing tool I:
tolerance specification. Comput Ind 31:161–174
Shah JJ, Yan Y, Zhang BC (1998) Dimension and tolerance modeling and transformations in
feature based design and manufacturing. J Intell Manuf 9:475–488
Söderberg R, Johannesson H (1999) Tolerance chain detection by geometrical constraint based
coupling analysis. J Eng Des 10(1):5–24
Srinivasan B, Jayaraman R (1989) Geometric tolerancing: II. Conditional tolerances. IBM J Res
Dev 33(2):105–125
Sudarsan R, Narahari Y, Lyons KW et al (1998) Design for tolerance of electro-mechanical
assemblies. In: Proceedings of the IEEE international conference on robotics and automation,
Leuven, pp 1490–1497
Teissandier D, Dufaure J (2007) Specifications of a pre and post-processing tool for a tolerancing
analysis solver. In: Proceedings of the CIRP international conference on computer-aided tol-
erancing, Erlangen
Toulorge H, Rivière A, Bellacicco A et al (2003) Towards a digital functional assistance process
for tolerancing. ASME J Comput Inf Sci Eng 3:39–44
Voelcker HB (1998) The current state of affairs in dimensional tolerancing: 1997. Integr Manuf
Syst 9(4):205–217
Wang H, Roy U, Sudarsan R et al (2003) Functional tolerancing of a gearbox. In: Proceedings of
NAMRC, Hamilton
Wang H, Ning R, Yan Y (2006) Simulated tolerances CAD geometrical model and automatic
generation of 3D tolerance chains. Int J Adv Manuf Technol 29:1019–1025
Weill R (1988) Tolerancing for function. CIRP Ann 37(2):603–610
Weill RD (1997) Dimensioning and tolerancing for function. In: Zhang HC (ed) Advanced
tolerancing techniques. Wiley, New York
1 Geometric Tolerance Specification 37
Whitney DE (1996) The potential for assembly modeling in product development and manufac-
turing. Technical report. The Center for Technology, Policy and Industrial Development,
MIT, Cambridge
Whitney DE, Mantripragada R, Adams JD et al (1999) Designing assemblies. Res Eng Des
11:229–253
Willhelm RG, Lu SCY (1992) Tolerance synthesis to support concurrent engineering. CIRP Ann
41(1):197–200
Wu Y, Shah JJ, Davidson JK (2003) Computer modeling of geometric variations in mechanical
parts and assemblies. ASME J Comput Inf Sci Eng 3:54–63
Zou Z, Morse EP (2004) A gap-based approach to capture fitting conditions for mechanical
assembly. Comput Aided Des 36:691–700
Chapter 2
Geometric Tolerance Analysis
Wilma Polini
Abstract This chapter focuses on five main literature models of geometric toler-
ance analysis – vector loop, variational, matrix, Jacobian, and torsor – and makes
a comparison between them in order to highlight the advantages and the weak-
nesses of each, with the goal of providing a criterion for selecting the most suit-
able one, depending on the application. The comparison is done at two levels: the
first is qualitative and is based on the analysis of the models according to a set of
descriptors derived from what is available in the literature; the second is quantita-
tive and is based on a case study which is solved by means of the five models.
Finally, in addition to providing comparative insight into the five tolerance analy-
sis models, some guidelines are provided as well, related to the development of a
novel approach which is aimed at overcoming some of the limitations of those
models.
2.1 Introduction
__________________________________
W. Polini
Dipartimento di Ingegneria Industriale, Università degli Studi di Cassino,
Via Gaetano di Biasio 43, 03043 Cassino, Italy.
e-mail: polini@unicas.it
39
40 W. Polini
approach where product design and manufacturing process planning activities are
carried out in parallel in highly communicative and collaborative environments,
with the aim of reducing reworking times and discard rates. In recent years, the
importance of assessing the effects of part tolerances on assembly products has
increasingly been acknowledged as one of the most strategic activities to be pur-
sued with the goal of ensuring higher production qualities at lower costs. In fact,
while the need for assigning some type of dimensional and geometric tolerances to
assembly components is widely recognized as a necessary step for ensuring a stan-
dardized production process and for guaranteeing the correct working of the as-
sembly to the required levels of satisfaction, the relationships between the values
assigned to such tolerances and final product functionality are more subtle and
need to be investigated in greater detail. As defined by designer intent, assembled
products are built to satisfy one or more functional prerequisites. The degree to
which each of such functional prerequisites is satisfied by the product is usually
strongly related to a few key dimensions of the final assembly itself. In related
literature, such key dimensions are often called key functional parameters or de-
sign dimensions, or functional requirements, which is the term that will be adopted
in this chapter. Functional requirements are typically the result of the stack-up of
geometries and dimensions of several parts; and their final variability, also funda-
mental in determining overall functional performance, arises from the combined
effects of the variabilities associated with the parts involved in the stack-up. Toler-
ance values assigned to parts and subassemblies become critical in determining the
overall variability of functional requirements and, consequently, the functional
performance of the final product. Moreover, when analyzing the chain of con-
nected parts, it becomes relevant to identify those tolerances that – more than oth-
ers – have an influence on the final outcome, since it has been shown that often the
70–30 rule applies, meaning that 30% of the tolerances assigned to the components
are responsible for 70% of the assembly geometric variation. The variabilities
associated with dimensions and geometries of the assembly components combine,
according to the assembly cycle, and generate the variability associated with the
functional requirements. Tolerance stack-up functions are mathematical models
aimed at capturing such combinations. They have that name because they are de-
signed as functions whose output is the variation range of a functional requirement,
and whose inputs are the tolerances assigned to assembly components. In other
words, a tolerance analysis problem implies modeling and solving a tolerance
stack-up function, in order to determine the nominal value and the tolerance range
of a functional requirement starting from the nominal values and the tolerance
ranges assigned to the relevant dimensions of assembly components.
In a typical tolerance analysis scenario, linear or nonlinear tolerance stack-up
functions may need to be solved, depending on the problem being studied, and on
the way it was modeled in the analysis. Alternative assembly cycles may be con-
sidered within the analysis in order to identify the one allowing assembly func-
tionality with the maximum value of the tolerance range assigned to the compo-
nents. The importance of an effectual tolerance analysis is widely recognized:
significant problems may arise during the actual assembly process if the tolerance
2 Geometric Tolerance Analysis 41
analysis on a part or subassembly was not carried out or was done to an unsatisfac-
tory level (Whitney 2004). It may even happen that the product is subjected to
significant redesign because of unforeseen tolerance problems, which were not
detected prior to actual assembly taking place. In this case business costs may be
significantly high, especially when considering that 40–60% of the production
costs are estimated as due to the assembly process (Delchambre 1996).
Many well-known approaches, or models, exist in the literature for tolerance
analysis (Hong and Chang 2002; Shen et al. 2004). The vector loop model adopts
a graph-like schematization where any relevant linear dimension in the assembly
is represented by a vector, and an associated tolerance is represented as a small
variation of such a vector (Chase et al. 1997; Chase 1999). Vectors are connected
to form chains and/or loops, reflecting how assembly parts stack up together in
determining the final functional requirements of the assembly. Stack-up functions
are built by combining the variations associated with vectors involved in each
chain into mathematical expressions, which can then be solved with different
approaches.
The variational model has its roots in parametric geometric modeling, where
geometry can be modeled by mathematical equations that allow shape and size
attributes to be changed and controlled through a reduced set of parameters (Gupta
and Turner 1993; Whitney et al. 1999; Li and Roy 2001). Parametric modeling
can be used as the starting point for reproducing small variations in an assembly
part, within the ranges defined by a given tolerance. In the matrix model, the ap-
proach aims at deriving the explicit mathematical representation of the geometry
of each tolerance region (the portion of space where a feature is allowed to be,
given a set of tolerances); this is done through displacement matrices, which de-
scribe the small displacements a feature is allowed to have without violating the
tolerances (Desrochers and Rivière 1997; Clément et al. 1998).
In the Jacobian model, tolerance chains are modeled as sequences of connected
pairs of relevant surfaces; displacement between such surfaces (whether nominal or
due to variations allowed by tolerances) is modeled through homogeneous coordi-
nate transformation matrices (Laperrière and Lafond 1999; Laperrière and Kabore
2001). The way the matrices are formulated draws inspiration from a common
approach adopted in robotics which involves the use of Jacobian matrices.
Finally, in the torsor model, screw parameters are introduced to model 3D tol-
erance zones (Ballot and Bourdet 1995, 1997). The name derives from the data
structure adopted to collect the screw parameters (i.e., the torsor).
The five modeling approaches introduced above propose different solutions to
specific aspects of the tolerance analysis problem; all have strong points and
weaknesses that may make them inadequate for specific applications. Most aspects
differentiating the approaches are related to how geometric variability of parts and
assemblies is modeled, how joints and clearance between parts are represented,
how stack-up functions are solved, and so forth. Moreover, it is difficult to find
literature work where the different approaches are compared systematically with
the help of one or more case studies aimed at highlighting the advantages and
disadvantages of each.
42 W. Polini
Section 2.2 describes a practical case study, which will be used as a reference
to illustrate the tolerance analysis models presented in Sections 2.3–2.5. Sec-
tion 2.6 provides some guidelines for the development of a new model aimed at
overcoming the weaknesses of the literature models.
To compare the tolerance analysis models, the case study shown in Figure 2.1 is
introduced. The 2D geometry of the example assembly is made of a rectangular
box containing two disk-shaped parts. The width g of the gap between the top disk
and the upper surface of the box is assumed as the functional requirement to be
investigated by the analysis. The goal of the tolerance analysis problem is to iden-
tify the tolerance stack-up function that defines the variability of g, and describes
it as a function of the geometries and tolerances of the components involved in the
assembly.
Tolerance analysis is based on the dimensional and geometric tolerances illus-
trated in Figure 2.1. The example is adapted from a real-life industrial application
and properly simplified to make it easier to present and discuss in this context. The
tolerancing scheme applied, which may not appear as entirely rigorous under the
Figure 2.1 Dimensional and geometric tolerances applied to the case study
2 Geometric Tolerance Analysis 43
The variability of the gap is the difference between the maximum or the minimum
values and the nominal one:
( )
Δg dim1 = + g maxdim − g N = ( 2.1064 − 1.2702 ) ≅ +0.84 mm,
(2.3)
Δg dim2 = −(g N − g mindim ) = − (1.2702 − 0.4909 ) ≅ −0.78 mm.
Albeit operating on a simplified problem (geometric tolerances are neglected)
the manual computation of the gap boundary values provides a useful support for
the quantitative comparison of the five methods, at least when they are applied by
considering dimensional tolerances only. The manually obtained, extreme gap
values will be used as reference values later on, then the results of the five meth-
ods will be discussed.
Furthermore, the manual computation procedure highlights one of the funda-
mental issues analyzed in this chapter, i.e., how hard it actually is to include geo-
metric tolerances in any model attempting to represent geometric variability. The
investigation of how this challenge is handled in the analysis of tolerance chains is
of fundamental importance when analyzing the performance and limitations of the
five approaches: under this assumption, it will be shown how each approach pro-
vides a different degree of support to the inclusion of geometric tolerances, and
how each requires different modeling efforts, simplifications, and workarounds, in
order to let geometric tolerances be included in the tolerance chain analysis problem.
44 W. Polini
The vector loop model uses vectors to represent relevant dimensions in an assem-
bly (Chase et al. 1995, 1996). Each vector represents either a component dimen-
sion or an assembly dimension. Vectors are arranged in chains or loops to repro-
duce the effects of those dimensions that stack together to determine the resultant
assembly dimensions. Three types of variations are modeled in the vector loop
model: dimensional variations, kinematic variations, and geometric variations.
In a vector loop model, the magnitude of a geometric dimension is mapped to
the length (Li) of the corresponding vector. Dimensional variations defined by
dimensional tolerances are incorporated as ± variations in the length of the vector.
Kinematic variations describe the relative motions among mating parts, i.e., small
adjustments that occur at assembly time in response to the dimensional and geo-
metric variations of the components. In the vector loop model, kinematic varia-
tions are modeled by means of kinematic joints, i.e., schematizations such as the
slider. In vector loop models, there are six common joint types available for 2D
assemblies and 12 common joints for 3D assemblies. At each kinematic joint,
assembly adjustments are turned into ranges for the motions allowed by the joint
(i.e., degrees of freedom). A local datum reference frame (DRF) must be defined
for each kinematic joint.
Geometric variations capture those variations that are imputable to geometric
tolerances. These are modeled by adding additional degrees of freedom to the
kinematic joints illustrated above. This introduces a simplification: although geo-
metric tolerances may affect an entire surface, in vector loop models they are
considered only in terms of the variations they induce at mating points, and only in
the directions allowed by the type of kinematic joint. Depending on what type of
geometric variation is represented by the tolerance and what motions are allowed
at the kinematic joint, a geometric tolerance is typically modeled as an additional
set of translational and rotational transformations (e.g., displacement vectors,
rotation matrices) to be added at the joint.
To better understand the vector loop model, the basic steps for applying it to a
tolerance analysis problem are provided below (Gao et al. 1998; Faerber 1999;
Nigam and Turner 1995):
1. Create the assembly graph. The first step is to create an assembly graph. The
assembly graph is a simplified diagram of the assembly representing the parts,
their dimensions, the mating conditions, and functional requirements, i.e., the
final assembly dimensions that must be measured in order to verify that the
product is capable of providing the required functionality. An assembly graph
assists in identifying the number of vector chains and loops involved in the as-
sembly.
2. Define the DRF for each part. The next step is to define the DRF for each part.
DRFs are used to locate relevant features on each part. If there is a circular con-
tact surface, its center is considered as a DRF too.
2 Geometric Tolerance Analysis 45
3. Define kinematic joints and create datum paths. Each mating relation among
parts is translated into a kinematic joint. Kinematic joints are typically located
at contact points between parts. Datum paths are geometric layouts specifying
the direction and orientation of vectors forming the vector loops; they are cre-
ated by chaining together the dimensions that locate the point of contact of a
part with another, with respect to the DRF of the part itself.
4. Create vector loops. With use of the assembly graph and the datum paths, vec-
tor loops are created. Each vector loop is created by connecting datums. Vector
loops may be open or closed; an open loop terminates with a functional re-
quirement, which can be measured in the final assembly (it could be either the
size of a relevant gap in the final assembly, or any other functionally relevant
assembly dimension); a closed loop indicates the presence of one or more ad-
justable elements in the assembly.
5. Derive the stack-up equations. The assembly constraints defined within vector-
loop-based models may be mathematically represented as a concatenation of
homogeneous rigid body transformation matrices:
R1 ⋅ T1 ⋅ ... ⋅ R i ⋅ Ti ⋅ ... ⋅ R n ⋅ Tn ⋅ R f = H , (2.4)
where R1 is the rotational transformation matrix between the x-axis and the first
vector; T1 is the translational matrix for the first vector, Ri,n and Ti,n are the cor-
responding matrices for the vector at node i or node n, and Rf is the final clo-
sure rotation, again with respect to the x-axis. H is the resultant matrix. For ex-
ample, in the 2D case the rotational and the translational matrices are as
follows:
⎡cos φi sin φi 0⎤ ⎡1 0 Li ⎤
R i = ⎢⎢ sin φi cos φi 0 ⎥ and Ti = ⎢⎢0 1 0 ⎥⎥ ,
⎥
⎢⎣ 0 0 1 ⎥⎦ ⎢⎣0 0 1 ⎥⎦
where φi is the angle between the vectors at node i, and Li is the length of vector
i. If the assembly is described by a closed loop of constraints, H is equal to the
identity matrix, otherwise H is equal to the g vector representing the resultant
transformation that will lead to the identification of a functional requirement.
6. Tolerance analysis – assuming an assembly as made of p parts. Each part is
represented by an x vector of its relevant dimensions and by an α vector con-
taining additional dimensions, added to take into account geometric tolerances.
When parts are assembled together, the resulting product is characterized by
a u vector of the assembly variables and by a g vector of measurable functional
requirements. It is possible to write L = J – P + 1 closed loops, where J is the
number of the mates among the parts and P is the number of parts. For each
closed loop
H ( x, u , α ) = 0 , (2.5)
46 W. Polini
while there is an open loop for each functional requirement that looks like
g = K ( x, u , α ) . (2.6)
Equation 2.5 allows one to calculate g after having solved the system of equa-
tions in Equation 2.4. The equations in Equation 2.4 are usually not linear; they
can be solved in different ways, for example, by means of the direct lineariza-
tion method:
dH ≅ A ⋅ dx + B ⋅ du + F ⋅ dα = 0 , (2.7)
dH ≅ −B −1 ⋅ A ⋅ dx − B −1 ⋅ F ⋅ dα , (2.8)
dg ≅ C ⋅ dx + D ⋅ du + G ⋅ dα = 0 , (2.9)
dg ≅ [C − D ⋅ B −1 ⋅ A] ⋅ dx + [G − D ⋅ B −1 ⋅ F] ⋅ dα
(2.10)
= S x ⋅ dx + Sα ⋅ dα,
while in the statistical scenario the solution can be obtained as a root sum of
squares, as follows:
where k and l are the number of x dimensions and α geometric tolerances that
influence the variable gi, Sxik is the matrix of the coefficients of the k x variables
inside the i-stack-up function of Equation 2.10, Sαil is the matrix of the coeffi-
cients of the l α variables inside the i stack-up function of Equation 2.10, and
txk and tαl are the vectors of the dimensional or the geometric tolerances of the
xk and αl variables, respectively.
The direct linearization method is a very simple and rapid method, but it is ap-
proximated too. When an approximated solution is not acceptable, it is possible
to use alternative approaches, such as numerical simulation by means of a
Monte Carlo technique (Gao et al. 1998; Boyer and Stewart 1991).
2 Geometric Tolerance Analysis 47
With reference to Figure 2.2, let x1 and x2 be the dimensions of the box, and x3
and x4 the diameters of the two disks; u1, u2, u3, and u4 are the assembly (depend-
ent) dimensions and g is the width of the gap between the top side of the box and
the second disk. The dimension g is the functional requirement. Therefore, the
assembly graph in Figure 2.3 has been built. It shows two joints of “cylinder
slider” kind between the box and disk 1 at point A and point B, respectively; one
joint of “parallel cylinder” kind between disk 1 and disk 2 at point C; one joint of
“cylinder slider” kind between disk 2 and the box at point D; and the measurement
to be performed (g).
DRFs have been assigned to each part; they are centered at point Ω for the box
and at the centers O1 and O2 of the two disks. All the DRFs have a horizontal
x-axis. Datum A has been assumed to be nominal. The DRF of the box is also
assumed as the global DRF of the assembly. Figure 2.4 shows the created datum
paths that chain together the points of contact of a part with another with respect to
the DRF of the part itself. Vector loops are created using the datum paths as a
guide. There are L = J – P + 1 = 4 – 3 + 1 = 2 closed loops and one open loop. The
first closed loop joins the box and disk 1 passing through contact points A and B.
F G
H
x4
O2
D
x2 = 80 ± 0.50
x4
C
x4
x3 = 20 ± 0.05
u4 x3
B O1
x4 = 20 ± 0.05
x3
x3
u2
E
Ω u1 A
u3
x1 = 50 ± 0.20
Figure 2.2 Assembly variables and tolerances of the vector loop model with dimensional toler-
ances
48 W. Polini
The second closed loop joins the subassembly box disk 1 and disk 2 through con-
tact points D and C. The open loop defines the gap width g. All the loops are de-
fined counterclockwise. The R and T matrices are 2D; their elements are shown in
Table 2.1.
Circle 1
L1
L2
Box Circle 2
L3
g
F G
g
H
x4
O2
D
x4
C
x4
u4 x3
B O1
x3
x3
u2
E
Ω u1 A
u3
Table 2.1 Elements of R and T matrices when the case study considers dimensional tolerances
which gives
g = x2 − u4 − x4 . (2.18)
where
⎡ 0 0 1 0 ⎤
⎢ 0 0 1 0 ⎥⎥
⎢
u
⎢ −0.2582 0 2.2910 1.2910 ⎥
S =⎢ ⎥. (2.23)
⎢ 0 0 0 0 ⎥
⎢ −0.0258 0 0.0323 0.0323 ⎥
⎢ ⎥
⎣⎢ 0.0258 0 −0.0323 −0.0323⎦⎥
The gap g depends on the following x variables through the sensitivity coeffi-
cients:
dg = dx2 − dx4 − du4 = 0.2582 ⋅ dx1 + dx2 − 2.2910 ⋅ dx3
(2.24)
−2.2910 ⋅ dx4 .
The solution obtained is lower than the value obtained by means of the manual
resolution method of about 4% [= (1.56–1.62)/1.62].
It is possible to calculate the solution in the statistical scenario (root sum of
square) as
2 1/ 2
⎢⎣ (
Δg Stat = ± ⎡ ∑ S xik ⋅ t xk ) ⎤⎥⎦ = ±0.5158 ≅ ±0.52 mm. (2.26)
With reference to Figure 2.5, let x1 and x2 be the dimensions of the box, and x3
and x4 the diameters of the two disks; u1, u2, u3, and u4 are the assembly (depend-
ent) dimensions and g is the width of the gap between the top side of the box and
the second disk. The DRFs and the datum paths are the same as in the previous
case (see Section 2.3.1).
The vector loops are the same as in the previous case, but they have to take into
consideration the geometric tolerances. To include geometric tolerances, the fol-
lowing variables must be added to the x vector (note that mating points A and B
are named after datums A and B the points lie upon):
2 Geometric Tolerance Analysis 51
− The flatness tolerance applied to the bottom surface of the box (datum A in
the drawing) can be represented as a translation of the A point in the direc-
tion perpendicular to datum A, i.e., perpendicular to the x-axis; this transla-
tion is described by the variable α1 = TA1 = 0 ± 0.10/2 = 0 ± 0.05 mm.
− The perpendicularity applied to the vertical left surface of the box (datum B)
can be represented as a translation of point B in the direction perpendicular
to datum B (the y-axis); this translation is described by the variable α2 = TB2
= 0 ± 0.10/2 = 0 ± 0.05 mm.
− The parallelism applied to the right side of the box (with respect to datum B)
can be represented as a translation of point D, again in the direction perpen-
dicular to datum B. It can be described by the variable α3 = TD3 = 0 ± 0.20/2
= 0 ± 0.10 mm.
− The circularity applied to disk 1 can be seen as points A, B, and C translat-
ing along the radius, and can be described by the variables α4 = TA4 = 0 ±
0.05/2 = 0 ± 0.025 mm, α5 = TB4 = 0 ± 0.05/2 = 0 ± 0.025 mm, and α6 = TC4
= 0 ± 0.05/2 = 0 ± 0.025 mm.
− The circularity applied to disk 2 can be represented as points C, D, and H
translating along the radius, and can be described by the variables α7 = TC5 =
0 ± 0.05/2 = 0 ± 0.025 mm, α8 = TD5 = 0 ± 0.05/2 = 0 ± 0.025 mm, and α9 =
TH5 = 0 ± 0.05/2 = 0 ± 0.025 mm.
− The parallelism applied to the top side of the box can be represented as a
vertical translation of point G, and is described by the variable α10 = TG6 = 0
± 0.10/2 = 0 ± 0.05 mm.
The R and T matrices are 2D; their elements are shown in Table 2.2.
// 0.10 A
F G
g // 0.20 B
B
0.05
H
⊥ 0.10 A
x4
0.05 O2
D
x2 = 80 ± 0.50
x4
C
x4
x3 = 20 ± 0.05
u4 x3
B O1
x4 = 20 ± 0.05
x3
u2 x3
E
Ω u1 A
A 0.10
u3
x1 = 50 ± 0.20
Figure 2.5 Assembly variables and tolerances of the vector loop model with geometric tolerances
52 W. Polini
Table 2.2 Elements of R and T matrices when the case study considers geometric tolerances
Once the vector loops have been generated, the relative equations can be de-
fined and solved. For the first loop, Equation 2.4 becomes
R1 ⋅ T1 ⋅ R 2 ⋅ T2 ⋅ R 3 ⋅ T3 ⋅ R 4 ⋅ T4 ⋅ R 5 ⋅ T5 ⋅ R 6 ⋅ T6 ⋅ R 7 ⋅ T7 ⋅ R8 ⋅ T8 ⋅
(2.27)
R f = I,
x1 − ( x4 + α 3 + α 8 ) + ( x3 + x4 + α 6 + α 7 ) ⋅ cos(180 + ϕ24 ) +
( x3 + α 2 + α 5 ) ⋅ cos(180 + ϕ24 + ϕ26 ) = 0,
u4 + ( x3 + x4 + α 6 + α 7 ) ⋅ sin(180 + ϕ24 ) + ( x3 + α 2 + α 5 ) ⋅ (2.30)
sin(180 + ϕ24 + ϕ26 ) − u2 = 0,
ϕ24 + ϕ26 = 0.
2 Geometric Tolerance Analysis 53
which gives
g = x2 + α10 − u4 − x4 − α 9 . (2.32)
where
T
⎧0.05, 0.05, 0.10, 0.025, 0.025, 0.025, ⎫
dα = {dα1 ,..., dα10 }
T
=⎨ ⎬ , (2.36)
⎩0.025, 0.025, 0.025, 0.05 ⎭
⎡ 0 0 1 ⎤ 0
⎢ 0 0 1 ⎥ 0
⎢ ⎥
⎢ −0.2582 0 2.2910 ⎥ 1.2910
Sud =⎢ ⎥, (2.37)
⎢ 0 0 0 ⎥ 0
⎢ −0.0258 0 0.0323 0.0323 ⎥
⎢ ⎥
⎢⎣ 0.0258 0 −0.0323 −0.0323⎥⎦
T
⎡0 1 1 0 0 0 ⎤
⎢1 0 0.2582 0 0.0258 −0.0258 ⎥⎥
⎢
⎢0 0 0.2582 0 0.0258 −0.0258 ⎥
⎢ ⎥
⎢0 1 1 0 0 0 ⎥
⎢1 0 0.2582 0 0.0258 −0.0258 ⎥
Suα =⎢ ⎥ . (2.38)
⎢0 0 1.0328 0 0.0064 −0.0064 ⎥
⎢0 0 1.0328 0 0.0064 −0.0064 ⎥
⎢ ⎥
⎢0 0 0.2582 0 0.0258 −0.0258 ⎥
⎢ ⎥
⎢0 0 0 0 0 0 ⎥
⎢⎣ 0 0 0 0 0 0 ⎥⎦
54 W. Polini
The solution shows that the variability of the gap width g can be represented by
the following function of the x vector:
dg = dx2 + dα10 − dx4 − du4 − d α 9 = 0.2582 ⋅ dx1 + dx2
−2.2910 ⋅ dx3 − 2.2910 ⋅ dx4 − dα1 − 0.2582 ⋅ d α 2 − 0.2582 ⋅ d α 3
(2.39)
− dα 4 − 0.2582 ⋅ dα 5 − 1.0328 ⋅ dα 6 − 1.0328 ⋅ dα 7 − 0.2582 ⋅ dα 8
− dα 9 + dα10
It is also possible to compute the solution with the statistical approach (root
sum of squares):
2 1/ 2
Δg Stat = ± ⎡ ∑ S xik ⋅ t xk
( ) ( ) ⎤⎦⎥
2
+ ∑ Sαij ⋅ Δα j = ±0.5361 ≅ ±0.54 mm. (2.41)
⎣⎢
A mathematical foundation of this model was proposed first by Boyer and Stewart
(1991), and then by Gupta and Turner (1993). Later, several additional variants
were proposed as well, and nowadays commercial computer aided tolerancing
(CAT) software packages are based on this approach, such as eM-TolMate from
UGS®, 3-DCS from Dimensional Control Systems®, and VisVSA from UGS®.
The basic idea of the variational model is to represent the variability of an as-
sembly, due to tolerances and assembly conditions, through a parametric mathe-
matical model.
To create an assembly, the designer must define the nominal shape and the di-
mensions of each assembly component (this information is usually retrieved from
CAD files). Then, the designer identifies the relevant features of each component
and assigns dimensional and geometric tolerances to them. Each feature has its local
DRF, while each component and the whole assembly have their own global DRF. In
nominal conditions, a homogeneous transformation matrix (called TN) is defined
that identifies the position of the feature DRF with respect to the part DRF. In real
conditions (i.e., manufactured part), the feature will be characterized by a roto-
translational displacement with respect to its nominal position. This displacement is
modeled to summarize the complete effects of the dimensional and geometric varia-
tions affecting the part by means of another matrix: the differential homogeneous
transformation matrix (called DT). The variational model may take into account the
precedence among the datums by setting the parameters of the DT matrix.
2 Geometric Tolerance Analysis 55
The variational model is not able to deal with the form tolerances such as the
vector loop model does; this means that the actual feature shape is assumed un-
changed, i.e., feature shape variations are neglected. The position of the displaced
feature in the part DRF can be simply obtained by matrix multiplication as a
change of DRF.
The model is parametric because different types and amounts of variations can
be modeled by simply altering the contents (parameters) of the DT matrix. In
some cases, the localization of a feature affected by a variation may be defined by
a transformation with respect to another feature in the same part which is affected
by variations as well. Therefore, the material modifier condition is modeled by
setting the parameters of the DT matrix.
Once the variabilities of the parts have been modeled, they must be assembled
together. Another set of differential homogeneous transformation matrices is in-
troduced to handle the roto-translational deviations introduced by each assembly
mating relation. Such matrices are named DA, with the letter A (for “assembly”) to
distinguish them from the matrices that have been used for parts. Those matrices
are hard to evaluate, since they depend on both the tolerances imposed on the parts
in contact and the assembly conditions. This model is not able to represent mating
conditions with clearance. The problem of evaluating the differential matrix is
analyzed in several literature works. A possible strategy consists in modeling the
joint between the coupled parts by reconstructing the coupling sequence between
the features (Berman 2005). Another possibility is to impose some analytical con-
straints on the assembly parameters (Whitney 2004).
When all the transformation matrices have been obtained, it is possible to ex-
press all the features in the same global DRF of the assembly. Finally, the func-
tional requirements can be modeled in the form of functions, as follows:
FR = f ( p1 , p2 ,..., pn ) , (2.42)
where FR is the assembly functional requirement, p1,…, pn are the model parame-
ters, and f(p) is the stack-up function (usually not linear) obtained from the matrix
multiplications described above. This model may be applied to assemblies involv-
ing joints which make a linear structure among the parts (linear stack-up function,
see Figure 2.6a) and joints which make a complex structure among the parts (net-
works of stack-up functions, see Figure 2.6b), such as a vector loop does.
Once the stack-up functions have been modeled, there are two approaches to
solve them: the worst-case approach and the statistical approach. The worst-case
analysis consists in identifying the extreme configurations of the assembly under a
given set of tolerances. In the variational approach, the problem is generally han-
dled as an optimization (maximization and/or minimization) problem, under con-
straints defined by the tolerances themselves. The statistical approach is generally
handled by assigning predefined probability density functions, e.g., Gaussian, to
the parameters identifying the main elements that contribute to the variation of
each feature (often assumed independent, by simplification), and then solving the
stack-up functions accordingly (Salomons et al. 1996).
56 W. Polini
To better illustrate the variational method, its basic steps are illustrated in the
following:
1. Create the assembly graph. The first step is to create an assembly graph. The
assembly graph is a simplified diagram of the assembly representing the parts,
the features, the mating conditions, and the functional requirements.
2. Define the DRF of each feature, of each part, and of the assembly. The next
step is to identify the local DRF of each feature and the global DRF of each
part and of the assembly (usually the DRF of the assembly coincides with the
DRF of the first part). DRFs are positioned depending on the surface type; from
the DRFs, local parameters and the differential homogeneous transformation
matrices DT are defined.
3. Transform the features. Once the transformation matrices are known, each
feature of a part is transformed into the global DRF of the part.
4. Create the assembly. With use of the assembly graph and the transformed fea-
tures, the assembly conditions are extracted, i.e., the assembly parameters in-
cluded in the matrix DA are calculated.
5. Derive the equations of the functional requirements. Once the assembly pa-
rameters are known, all the features can be expressed in the same global DRF
of the assembly. At this point, the functional requirements are defined in terms
of functions that can be solved by means of the previously described worst-case
and/or statistical approaches.
ous transformation matrices are again considered as the foundation of the mathe-
matical representation. A displacement matrix DT is used to describe any roto-
translational variation a feature may be subjected to; the matrix is defined with
respect to a local DRF. Since the goal is to represent the boundaries of the region
of possible variations (i.e., extreme values), the approach is intrinsically a worst-
case approach. No statistical approach may be implemented, such as vector loop
and variational models do. To represent boundaries, constraints must be added to
the displacements modeled within the DT matrices. Displacement boundaries
resulting from complex series of tolerances are solved by modeling the effects of
each tolerance separately and by combining the resulting regions. Analogously,
gaps/clearances are represented as if they were tolerance regions. Finally, by clas-
sifying the surfaces into several classes, each characterized by some type of in-
variance with respect to specific displacement types (e.g., a cylinder is invariant to
any rotation about its axis), one can simplify displacements and the resulting dis-
placement matrix (Clément et al. 1994).
A similar approach is followed to model the dimensions acting as functional
requirements of the assembly; since in this case the resulting region (of possible
values) is essentially contained in a segment, segment boundaries must be com-
puted by means of a worst-case approach (minimum–maximum distances between
the two points). The two points defining the boundaries of the segment must be
defined as the result of stack-up functions (Desrochers and Rivière 1997).
The matrix model is based on the positional tolerancing and the technologically
and topologically related surfaces (TTRS) criteria (Clément et al. 1998). Geomet-
ric features are assumed as ideal, i.e., the form tolerances are neglected, such as in
the variational model. To better understand the matrix method for tolerance analy-
sis, its basic steps are provided below:
1. Transform the tolerances applied to the drawing. The first step is to transform
the tolerances applied to the drawing to make them compliant with the posi-
tional tolerancing and the TTRS criteria.
2. Create the assembly graph. The second step is to create an assembly graph.
The assembly graph allows for identification of the global DRF and the link-
ages among the features to which the tolerances are assigned. The assembly
parts should be in contact; the joints with clearance may not be considered.
3. Define the local DRF of each part feature. A DRF must be assigned to each
part feature.
4. Identify the measurable points for each functional requirement. Points that
locate the boundaries of each functional requirement must be identified and the
path that connects them to the global DRF must be defined, taking into account
all the tolerances stacking up along the way.
5. Define the contributions of each single displacement and the related con-
straints. It is necessary to define the contribution of each displacement to the
total displacement region, and the constraints necessary to identify its bounda-
ries. Each surface can be classified into one of the seven classes of invariant
surfaces; this allows one to discard some displacements and to obtain a simpli-
58 W. Polini
In the terminology adopted by the Jacobian model approach, any relevant surface
involved in the tolerance stack-up is referred to as a functional element. In the
tolerance chain, functional elements are considered in pairs: the two paired sur-
faces may belong to the same part (internal pair), or to two different parts, and are
paired since they interact as mating elements (kinematic pair, also referred to as an
external pair). The parts should be in contact to be modeled by this model.
Transformation matrices can be used to locate a functional element of a pair
with respect to the other: such matrices can be used to model the nominal dis-
placement between the two functional elements, but also additional small dis-
placements due to the variabilities modeled by the tolerances. The form tolerances
are neglected. The main peculiar aspect of the Jacobian approach is how such
matrices are formulated, i.e., by means of an approach derived from the descrip-
tion of kinematic chains in robotics. The transformation that links two functional
elements belonging to a pair, and that includes both nominal displacement and
small deviations due to tolerances, can be modeled by a set of six virtual joints,
each associated with a DRF. Each virtual joint is oriented so that a functional
element may have either a translation or a rotation along its z-axis. The aggrega-
tion of the six virtual joints gives rise to the transformation matrix linking one
functional element to the other functional element of the pair (Laperrière and La-
fond 1999; Laperrière and Kabore 2001). The position of a point lying on the
second functional element of a pair, which may be assumed as depicting the func-
tional requirement under scrutiny, with respect to the DRF of the first functional
2 Geometric Tolerance Analysis 59
element (assumed as the global DRF) may be expressed by considering the three
small translations and the three small rotations of the point in the global DRF
through the product of a Jacobian matrix associated with the functional element
with tolerances of all the functional element pairs involved (internal or kinematic)
and a vector of small deviations associated with the functional element with toler-
ances of all the functional element pairs involved, expressed in the local DRF. The
main element of the expression is the Jacobian matrix, which is relatively easy to
compute, starting from the nominal position of the geometric elements involved.
The tricky part, however, is to turn the assembly tolerances into displacements to
assign them to the virtual joints defined for each functional element pair in the
chain.
The main steps of the approach are described below:
1. Identify the functional element pairs. The first step is the identification of the
functional element pairs (i.e., pairs of relevant surfaces). The functional ele-
ments are arranged in consecutive pairs to form a stack-up function aimed at
computing each functional requirement.
2. Define the DRF for each functional element and the virtual joints. The next
step is to define a DRF for each functional element, and to create the chain of
virtual joints representing the transformation that links the pair of functional
elements. Once such information is available, the transformation matrix for
each functional element can be obtained.
3. Create the chain and obtain the overall Jacobian matrix. The transformation
matrices can be chained to obtain the stack-up function needed to evaluate each
functional requirement. This model has been developed for assemblies involv-
ing joints which make a linear structure among the parts (linear stack-up func-
tion), while it is not able to deal with joints which make a complex structure
among the parts (network of stack-up functions), such as the matrix model
does.
4. Once the required stack-up function has been obtained, it may be solved by the
usual methods in the literature (Salomons et al. 1996) for the worst-case or sta-
tistical approaches.
5. Finally, it is necessary to observe that this model is based on the TTRS crite-
rion (Clément et al. 1998) and on the positional tolerancing criterion (Legoff
et al. 1999). Therefore, the tolerances of a generic drawing need to be con-
verted in accordance with the previously defined criteria, before carrying out
the tolerance analysis.
The torsor model uses screw parameters to model 3D tolerance zones (Chase et al.
1996). Screw parameters are a common approach adopted in kinematics to de-
scribe motion, and since a tolerance zone can be seen as the region where a sur-
60 W. Polini
face is allowed to move, screw parameters can be used to describe it. Each real
surface of a part is modeled by a substitution surface. A substitution surface is a
nominal surface characterized by a set of screw parameters that model the devia-
tions from the nominal geometry due to the applied tolerances. Seven types of
tolerance zones are defined. Each one is identified by a subset of nonzero screw
parameters, while the remaining ones are set to zero as they leave the surface in-
variant. The screw parameters are arranged in a particular mathematical operator
called a torsor, hence the name of the approach. Considering a generic surface, if
uA, vA, and wA are the translation components of its point A, and α, β, and γ are the
rotation angles (considered small) with respect to the nominal geometry, the corre-
sponding torsor is
⎧α uA ⎫
⎪ ⎪
TA = ⎨ β vA ⎬ , (2.43)
⎪γ wA ⎭⎪R
⎩
where R is the DRF that is used to evaluate the screw components.
To model the interactions between the parts of an assembly, three types of tor-
sors (or small displacement torsor, SDT) are defined (Ballot and Bourdet 1997):
a part SDT for each part of the assembly to model the displacement of the part;
a deviation SDT for each surface of each part to model the geometric deviations
from the nominal geometry; a gap SDT between two surfaces linking two parts to
model the mating relation. The form tolerances are neglected and they are not
included in the deviation SDT.
A union of SDTs is used to obtain the global behavior of the assembly. The ag-
gregation can be done by considering that the worst-case approach computes the
cumulative effect of a linear stack-up function of n elements by adding the single
components of the torsors. This is not true for a network of stack-up functions,
which has not been developed by the torsor model yet. The torsor method does not
allow one to apply a statistical approach, since the torsor’s components are inter-
vals of the small displacements; they are not parameters to which it is possible to
assign easily a probability density function.
The torsor model operates under the assumption that the TTRS and the posi-
tional tolerancing criteria are adopted, which means that the tolerances in the
drawing may need to be updated before carrying out the tolerance analysis. The
solution of stack-up functions arranged in a network has not been completely
developed. Finally, it is worth pointing out that, in the relevant literature, the use
of SDTs for modeling tolerance analysis problems tends to follow two main ap-
proaches: on one hand, SDTs are used to develop functions for computing the
position of geometric elements (belonging to the assembly) as they are subjected
to displacement allowed by tolerances (e.g., see Chase et al. 1996); on the other
hand, SDTs are used to model entire spatial volumes that encapsulate all the pos-
sible points in space that may be occupied by geometric elements during their
variations (e.g., see Laperrière et al. 2002). In the analysis of the case study, only
the second approach has been considered, since it looks more promising.
2 Geometric Tolerance Analysis 61
The basic steps of the torsor model are described in the following (Villeneuve
et al. 2001; Teissandier et al. 1999):
1. Identify the relevant surfaces of each part and the relations among them. The
first step is to identify the relevant surfaces belonging to each part and the rela-
tionships among them; this information is usually collected in a surfaces graph.
In this step the chains to relate the functional requirements to the relevant sur-
faces are identified.
2. Derive the SDTs. A deviation SDT needs to be associated with each relevant
surface of each part. This leads to the evaluation of a global SDT for each part.
Finally, the shape of the gap SDT is associated with each joint according to the
functional conditions of the assembly.
3. Obtain the functional requirement stack-up functions. Compute the cumulative
effects of the displacements and obtain the final linear stack-up function of
each functional requirement.
A first comparison of the previously described five models can be done by devising
a set of indicators describing features, capabilities, and issues related to the applica-
tion of such models to given tolerance stack-up problems. The indicators and their
results for the five models are summarized in Table 2.3. The indicators were de-
signed by drawing inspiration from what is available from the literature (Salomons
et al. 1996), with the necessary adaptations. Each descriptor may assume one of the
following three states: “X” if the model has a property, or it is capable of handling
a specific aspect or issue of the problem, “–” if the same property is missing, or the
model is not able to handle the aspect of the problem; “?” if the answer is uncertain,
because it may depend on the specific tolerance analysis problem, or because there
is not enough information to verify the capability. The first descriptor is the “analy-
sis type”, which refers to the type of approach that can be adopted to solve the
stack-up functions, i.e., worst-case or statistical. The descriptor “tolerance type”
indicates the kind of tolerance that the model may take into account: dimensional,
form, or other geometric (no form) tolerances. The “envelope and independence”
descriptors refers to the possibility of the model representing a dimensional toler-
ance when the envelope principle or the independence principle is specified. The
“parameters from tolerances” descriptor indicates whether the model allows for
translation of the applied tolerance ranges into the model parameter ranges. The
“tolerance stack-up type” descriptor refers to the possibility of a model building
and solving linear stack-up functions or networks of stack-up functions. The “joint
type” descriptor refers to the joint types that the model may take into account, ei-
ther with contact between the surfaces or with clearance. The “functional require-
ment schematization” descriptor refers to how a functional requirement can be
represented by a feature or by a set of points belonging to a feature. The “tolerance
zones interaction” descriptor indicates the capability of representing the interaction
62 W. Polini
among more than one tolerance applied to the same surface. The “datum prece-
dence” descriptor indicates whether a model can represent a sequence of datums.
Finally, the “material modifiers condition” descriptor indicates the capability of a
model to take into account material modifiers.
According to the results of the first comparison, the vector loop model and the
variational model appear more developed than the others; they are the only ones
that provide support for solving tolerance stack-up functions involving networks.
Moreover, they provide a method for assigning probability density functions to
model parameters, given the applied tolerances. However, the vector loop model
and the variational model are not completely consistent with the actual ISO and
ASME standards and they do not provide support handling interactions among
tolerance zones.
The vector loop model is the only model providing actual support for modeling
form tolerances; all the other models adopt the simplification consisting in consid-
ering the real features as coincident with their substitute ones.
The variational model supports the inclusion of precedence constraints among
datums, and also the presence of material modifiers conditions.
The matrix model and the torsor model support only the worst-case approach
for solving the tolerance analysis problem. This is a limitation, but their formaliza-
tion allows them to handle joints with clearance, and interaction among tolerance
zones.
The Jacobian model has the advantage that the Jacobian matrix can be easily
calculated from nominal conditions, while displacements of the functional re-
quirements can be directly related to displacements of the virtual joints; however,
it is difficult to derive such virtual joint displacements from the tolerances applied
to the assembly components. On the other hand, the torsor model may allow for
an easy evaluation of the ranges of the small displacements directly from the toler-
ances applied to the assembly components, but then it is very difficult to relate
these ranges to the ranges of the functional requirements of the assembly.
These two considerations have suggested the idea of a unified Jacobian–torsor
model to evaluate the displacements of the virtual joints from the tolerances ap-
plied to the assembly components through the torsors and, then, to relate the dis-
placements of the functional requirements to the virtual joint displacements
through the Jacobian matrix (Laperriére et al. 2002; Desrochers et al. 2003). Al-
though this is theoretically possible, since the deviations are usually small and,
therefore, the equations can be linearized, the actual feasibility of this approach is
still the subject of research.
Finally, the models considered have some common limitations. The first deals
with the envelope rule: the models do not allow one to apply the envelope rule and
the independence rule to different tolerances of the same part. The second is that
there do not exist any criteria to assign a probability density function to the model
parameters joined to the applied tolerances and that considers the interaction
among the tolerance zones. The last point deals with the assembly cycle: the mod-
els are not able to represent all the types of coupling with clearance between two
parts.
The solution of the case study by the five models considered is described in de-
tail in Marziale and Polini (2009b). Table 2.4 summarizes the results for the func-
tional requirement Δg as obtained by the application of the five models and com-
pared with the solution obtained with manual computation when only dimensional
Table 2.4 Results of the comparison among the models applied to the case study with dimen-
sional tolerances
tolerances are considered. Table 2.5 shows the results when both dimensional and
geometric tolerances are applied.
The results obtained by considering only the dimensional tolerances show that
all the models give slightly underestimated results with the worst-case approach,
when compared with the results obtained manually with the approach described
earlier. The matrix model has the highest error (–14%), while all the other models
provide the same result (–4%). This is probably due to the way the dimensional
tolerances are schematized (i.e., the first datum is nominal, the variability due to
the dimensional tolerance is considered to be applied only on one of the two fea-
tures delimiting the dimension). Moreover, the statistical approach gives similar
results for all the models considered.
The results obtained by considering both dimensional and geometric tolerances
show that all the models, except the vector loop model, give similar results with
the worst-case approach. This is probably due to the fact that the vector loop
model considers the effect of a set of tolerances applied to a surface as the sum of
the effects due to each single tolerance applied to the same surface. The effects of
the different tolerances are considered to be independent. Therefore, increasing the
number of tolerances applied to the same surface increases the variability of the
functional requirement. This means that the interaction among tolerances defined
on the same surface are not properly handled.
All five models produce very similar results when the statistical approach is
applied.
Moreover, the results in Tables 2.4 and 2.5 obtained from the Jacobian model
and from the torsor model are basically identical. This is due to the fact that a
simplification was adopted when modeling the problem, i.e., the angles of the box
were considered fixed at 90°. This assumption is due to the need to avoid the
networks of stack-up functions that the two models are not able to deal with. It
means that all the tolerances applied may involve only translations of the sides of
the box.
Table 2.5 Results of the comparison among the models applied to the case study with dimen-
sional and geometric tolerances
None of the models proposed in the literature provide a complete and clear
mechanism for handling all the requirements included in the tolerancing standards
(Shen et al. 2004). This limitation is reflected also in the available commercial
CAT software applications, which are based on the same models (Prisco and Gior-
leo 2002). As already discussed in detail in previous work (Marziale and Polini
2009b), the main limitations of the actual models are the following: they do not
properly support the application of the envelope rule and of the independence rule
to different dimensional tolerances on the same part as prescribed by the ISO and
ASME standards; they do not handle form tolerances (except for the vector loop
model); they do not provide mechanisms for assigning probability density func-
tions to model parameters starting from tolerances and considering tolerance zone
interactions; finally, they are not capable of representing all the possible types of
part couplings that may include clearance.
Some guidelines are now presented, aimed at the development of a new model
that addresses at least some of the limitations highlighted above.
A dimensional tolerance assigned to the distance between two features of a part
or of an assembly (Figure 2.7) may be required with the application of the enve-
lope rule or of the independence principle. In the second case, to correctly define
the relationship between the features, it is necessary to add a geometric tolerance
in order to limit the form and the orientation deviations. The envelope principle
states that when a feature is produced at its maximum material condition (MMC),
the feature must have a perfect form. The MMC of a feature is the size at which
the most material is in the part. The MMC size establishes a perfect form bound-
ary and no part of the feature must extend outside this boundary. As the size of a
feature departs from the MMC, its form is permitted to vary. Any amount of varia-
tion is permitted, as long as the perfect form boundary is not violated. However,
the size limits must not be violated either. Therefore, if the envelope principle is
applied to a dimensional tolerance, an additional constraint has to be considered to
build and to solve the stack-up functions of the assembly. Both the ASME
Y14.5M (1994) and the ISO 8015 (1985) standards foresee the possibility that,
also on the same part, dimensional tolerances may be assigned with or without the
application of the envelope principle. Consequently the mathematical model that is
used to schematize a dimensional tolerance in order to build and to solve the
stack-up functions should necessarily take into consideration these two possibili-
ties. To overcome this limitation of the models in the literature that are not able to
consider these two cases (Desrochers et al. 2003), a possible solution is to con-
sider a greater set of parameters to model the degrees of freedom of the planes
delimiting the dimension considered due to the applied tolerances (dimensional,
orientation, form). The envelope rule and the independence rule constrain those
parameters in different ways.
66 W. Polini
2.7 Conclusions
In this work, five different models available from the literature on tolerance analysis
were compared through their application to a case study. None of the models pro-
posed in the literature provide a complete and clear mechanism for handling all the
requirements included in the tolerancing standards, and this limitation is reflected
also in the available commercial CAT software. The main limitations include the
following: no proper support for the application of the envelope rule and of the inde-
pendence rule; cannot handle form tolerances (except for the vector loop model); no
mechanisms for assigning probability density functions to model parameters starting
from tolerances and considering tolerance zone interactions; no proper representa-
tion of all the possible types of part couplings that include clearance.
Guidelines were presented for the development of a new model aimed at ad-
dressing such highlighted limitations. Some suggestions were given to consider
dimensional tolerance with the application of both the envelope principle and the
independence principle, to take into account the real features and the interaction of
the tolerance zones, to consider joints with clearance among the assembly compo-
nents, and to adopt both the worst-case and the statistical approaches to solve the
stack-up functions. The implementation of those suggestions in a new model and
its application to case studies is the subject of ongoing research.
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
ASME Y14.5M (1994) Dimensioning and tolerancing. American Society of Mechanical Engi-
neering, New York
Ballot E, Bourdet P (1995) Geometrical behaviour laws for computer aided tolerancing. In:
Proceedings of the 4th CIRP seminar on computer aided tolerancing, University of Tokyo,
April 1995
Ballot E, Bourdet P (1997) A computational method for the consequences of geometric errors in
mechanisms. In: Proceedings of the 5th CIRP seminar on computer aided tolerancing, To-
ronto, 27–29 April 1997
Berman YO (2005) Shape and position uncertainty in mechanical assembly. PhD thesis, The
Hebrew University, Jerusalem
Boyer M, Stewart NF (1991) Modeling spaces for toleranced objects. Int J Robot Res 10:470–
582
Chase KW (1999) Multi-dimensional tolerance analysis (automated method). In: Drake PJR (ed)
Dimensioning and tolerancing handbook. McGraw-Hill, New York
Chase KW, Gao J, Magleby SP (1995) General 2-D tolerance analysis of mechanical assemblies
with small kinematic adjustments. J Des Manuf 5:263–274
Chase KW, Gao J, Magleby SP et al (1996) Including geometric feature variations in tolerance
analysis of mechanical assemblies. IIE Trans 28:795–807
Chase KW, Gao J, Magleby SP (1997) Tolerance analysis of 2- and 3D mechanical assemblies
with small kinematic adjustments. In: Zhang HC (ed) Advanced tolerancing techniques.
Wiley, New York
68 W. Polini
__________________________________
N. Senin
Dipartimento di Ingegneria Industriale, Università degli Studi di Perugia,
Via G. Duranti 67, 06125 Perugia, Italy,
e-mail: nsenin@unipg.it
G. Campatelli
Dipartimento di Meccanica e Tecnologie Industriali, Università degli Studi di Firenze,
Via S. Marta 3, 50139 Florence, Italy,
e-mail: gianni.campatelli@unifi.it
71
72 N. Senin and G. Campatelli
3.1 Introduction
these values were obtained by combining the information publicly available at the
Web sites of several manufacturers, and should be representative of their current
offerings, at least at the time of writing. While specific numbers may change, a
comparison in terms of orders of magnitude should suffice in delineating the main
performance differences among instrument families.
In this section, the most common classes of profilometers and 3D microscopes are
introduced and discussed. The focus will be on those instruments that are currently
used for microtopography analysis and visual inspection, but that could be adapted
to operate for the quantitative assessment of geometric error on microfabricated
parts or surface features. The proposed list is not meant to be complete: the selec-
tion criterion was to highlight those instruments which are more successful and
widespread, and therefore more commonly found in industrial practice.
Before proceeding with the illustration of the main instruments, it is convenient
to provide a brief introduction concerning the main terms that are commonly used
to refer to such instrument classes. In general, the term “profile” refers to some
digital data aimed at depicting the outline/contour formed by the reliefs of a sur-
face; in this sense, a profilometer, or profiler, is an instrument devoted to acquir-
ing profiles. Profilometers acquire surface topography information by having a
probe tracing the surface along a specified traversal trajectory. As the probe
moves, surface height data are collected. The profilometers analyzed in this chap-
ter are all 3D profilometers, i.e., they are devoted to acquiring 3D surface topog-
raphy information; this is generally done through the sequential acquisition of a
series of profiles, conveniently placed on the original surface, which are then
combined – via software – into a single coherent representation of surface topog-
raphy.
The term “microscope” is commonly used to indicate an instrument that pro-
duces an image, where a small (hence the term “micro”) surface region under
scrutiny is magnified to aid visual observation (hence the term “scope”). When
used in quality inspection of mechanical parts, traditionally microscopes have had
the main role of supporting visual inspection for investigating surface finish and
material microstructure (also in cross sections).
In this work, the term “3D microscope” is used to refer to a variation of the
digital microscope, designed to acquire topography information as a height map,
i.e., the equivalent of a digital image, but whose pixels contain surface height data.
Some 3D microscopes operate very similarly to 3D profilometers, and scan the
surface as if they were acquiring a series of profiles; others acquire surface data
points simultaneously, thus resembling more a conventional optical microscope,
where the light needed to form the image reaches the observer simultaneously.
3 Quality Inspection of Microtopographic Surface Features 75
Some instruments that go under the name of microscopes, namely, optical instru-
ments, may provide additional information in the form of color associated with the
points of the height map, thus providing something closer to the results of visual
observation.
In some industrial environments, the term “3D microscope” is also commonly
used to describe instruments which are capable of providing volume data, i.e.,
digital information concerning the entire 3D structure of the specimen (including
the internal parts). An example of an instrument of this type is the confocal laser
scanning microscope, which is included in the list of selected instruments and will
be illustrated later. Instruments capable of providing volume data are generally
also capable of providing surface topography data, as a subset of the measured
data: this applies to the confocal laser scanning microscope as well, which is why
it was included in the selection.
z displ.
tip
Voltage proportional to z
surface displacement
Figure 3.1 Architecture of a stylus probe. LVDT linear variable displacement transducer
76 N. Senin and G. Campatelli
Profile points are acquired sequentially, during traversal. Measurement can take
place continuously, the acquisition being triggered at constant time intervals while
the stylus travels at constant velocity (which gives rise to uniformly spaced points
in the traversal direction, at least nominally) or it can take place in a step-by-step
fashion, with the probe coming to a complete halt before z measurement takes
place (allowing for points to be located anywhere and at any reciprocal distance
along the probe traveling path). The actual implementation and the traveling strat-
egy depend on the instrument type and the intended usage.
Δx Δx
Δy
z
z Δx= Δy
y
y
x
x
a b
Figure 3.2 Profile scanning solutions for stylus-probe systems: a single-profile scanning with
uniform point spacing, and b parallel-profile scanning with uniform point spacing (raster scan-
ning)
With a single scan along a straight path a 2D profile can be acquired (see Fig-
ure 3.2a). The term “2D profile” derives from the fact that such a profile can be
completely defined in a 2D space. Traditional profilometers usually acquire a 2D
profile by scanning the surface along a straight path. In this work they are referred
to as 2D profilometers to distinguish them from more complex solutions illustrated
later. The 2D profilometer equipped with a stylus probe is the single most estab-
lished and widespread solution for surface finish analysis; its main architectural
elements are defined by international standards, see, in particular, ISO 3274
(1998). Measured data consist of a series of x, z coordinate pairs, or of a simpler
series of z values with common x spacing (Δ x) value, in the case of uniformly
spaced acquisition along a straight path. The stylus-based 2D profilometer is used
to assess surface finish properties related to form error, waviness, and roughness
in the form of quantitative parameters, also specified in detail by international
standards, in particular, ISO 4287 (1997), ISO 4288 (1996), ISO 12085 (1996),
ISO 8785 (1998), ISO 13565-1 (1996), ISO 13565-2 (1996), and ISO 13565-3
(1998).
While acquiring surface topography information through one or multiple scans
along straight paths may be sufficient for several quality inspection applications,
it was highlighted that the reconstruction of a complete 3D representation of sur-
face topography is often capable of providing superior results (Lonardo et al.
1996). Assuming 3D topography data are available, a significant research effort
3 Quality Inspection of Microtopographic Surface Features 77
has been long under way to define form error, waviness, and roughness areal
parameters to replace the 2D parameters in capturing the topographical properties
of a 3D surface topography. For this purpose, an international standard has been
published (ISO 25178-6:2010) and several others are currently being prepared:
ISO/CD 25178-1, ISO/DIS 25178-2, ISO/DIS 25178-3, and ISO/DIS 25178-7.
Three-dimensional surface topography can be reconstructed from a series of 2D
profiles (Figure 3.2b) with known relative displacement.
The most common measurement approach implemented by stylus-probe sys-
tems consists in doing parallel passes, which is compatible with the preferred
mode of operation of the stylus probe. The simplest way to recombine the profiles
into a single surface representation consists in having profiles equally spaced in
the y direction (Δy spacing) and using the same spacing between points in each
profile (Δx spacing), the overall result being a grid of equally spaced z coordi-
nates; this is usually referred to as raster scanning. Profile data obtained by means
of raster scanning are often referred to as a height map, or image, as it resembles a
uniform grid of grayscale pixels, each z coordinate being equivalent to a gray
level.
Stylus-based 3D profilometers can be synthetically depicted as 2D profilome-
ters with an additional y-axis drive. They are usually implemented as depicted in
Figure 3.3 (the illustration does not include the control unit, usually a PC-based
system).
column
stylus pick up body
x-axis drive
specimen
specimen
holder
base
y-axis
drive
Figure 3.3 Architecture of a Cartesian, stylus-based 3D profilometer (does not include the
controller unit)
The architecture was originally developed for 2D profilometers, and hence cen-
tered about the x-axis drive, whose predominant role is enforced by the very na-
ture of operation of the stylus; 3D profilometers have an additional y-axis drive. A
variant of this 3D architecture has the x-axis mounted under the specimen holder,
together with the y-axis drive. Both architectures are very delicate in their imple-
mentations, as different effects related to the kinematics and dynamics of the axes
have to be taken into account when designing and manufacturing the instruments,
with significant influence on measurement error.
A problem shared by both 2D and 3D profilometers is the establishment of a
reliable datum for the z measurement. Typical solutions for a 2D profilometer may
78 N. Senin and G. Campatelli
include the use of an external reference (e.g., optical flat) to retrieve the z error
introduced by the probe while it translates along the scanning direction, and/or the
use of an analogue high-pass filtering device, with behavior defined by interna-
tional standards such as ISO 3274 (1998) and ISO 11562 (1996). to remove this
type of error. These solutions are generally not viable for 3D profilometer archi-
tectures, because it is important to maintain the same datum for all measured pro-
files. The preferred strategy for 3D profilometers consists in not adopting any type
of mechanical/analogue filter devoted to such a task (i.e., acquisition of true pro-
files) and to proceed at the identification of a measurement datum at a later stage,
after measurement, operating directly via software on the acquired digital data.
A particular type of mechanical filtering solution is worth analyzing in more
detail: the stylus with a skid. Typical architectures of styluses with a skid are
shown in Figure 3.4: the probe consists of a main body terminating with a skid
that is in contact with the surface; the stylus is mounted within the main body.
Both the main body and the stylus freely follow the surface reliefs during scanning
since the probe body is connected to the instrument with a joint that allows it to
follow surface reliefs; only stylus displacement (relative to the main body) is re-
corded, since the vertical position of the main body acts as the reference datum.
main body
main body
stylus
stylus
skid skid
Figure 3.4 Two typical architectures for skidded styluses. Stylus displacement is recorded rela-
tive to the main body axis, whose orientation is determined by the skid sliding over the surface
The result is a mechanical high-pass filtering effect; therefore, this type of sty-
lus is incapable of recording a complete profile shape, but can just record its high-
est-frequency components. For this reason, profilometers equipped with such a
stylus are usually referred to as roughness measurement instruments, as the only
information they are capable of acquiring pertains to the roughness (high-
frequency components) of a surface. A skidded stylus has a more limited meas-
urement range than a stylus without a skid, but has higher vertical resolution and
accuracy; thus, it is preferentially adopted to measure surfaces of high nominal
finish quality. The use of such a solution is quite common in 2D profilometers,
because a surface with high surface finish requires a high-resolution probe, whose
limited range may cause out-of-range measurements if the specimen surface is not
perfectly leveled; the stylus with a skid solves this problem.
Nevertheless, the use of skidded styluses in 3D solutions should be avoided
when possible, or at least it should be considered with care, as different geometric
3 Quality Inspection of Microtopographic Surface Features 79
effects of the skid in the x and y directions may give rise to different filtering re-
sults, making it hard to identify a common datum for all the traces.
The last notable instrument belonging to the stylus-based class is the portable
roughness measurement system. It is usually a simplified version of a Cartesian
2D profilometer where the x-axis drive, pickup body, and stylus are enclosed,
together with a controller unit and minimal display, in a compact, portable unit.
The probe is equipped with a stylus with a skid, so the reference datum for the
single trace is provided directly, independently of the placement of the instrument.
The system is usually configured to compute roughness and waviness parameters
only.
scanned
stylus tip profile
narrow hill
dale
step
Figure 3.5 Interaction of the stylus tip with the measurand surface during scanning
As depicted in the figure, the interaction of the tip with the surface results in a
mechanical filtering effect. Consequently, steep slopes, narrow cavities, and spikes
may be smoothened, or remain undetected. The overall performance and issues
related to sylus-based profile measurement, discussed in detail in the following
paragraphs, are significantly influenced by the interaction of the stylus tip geome-
try (conicity and tip radius) and the profile aspect ratio (ratio of the length and
height of the profile). The steeper the slopes (higher aspect ratios), the more diffi-
cult it is to acquire them correctly, for a given tip geometry.
80 N. Senin and G. Campatelli
be provided as well. It is now generally accepted that information about the meas-
urement error should be provided in some rigorous statistical form, e.g., as meas-
urement uncertainty (i.e., a characterization of the range of values within which
the true value is asserted to lie with some level of confidence). Measurement un-
certainty is considered a valid quantitative indicator of measurement accuracy,
i.e., the combination of measurement trueness (closeness of agreement between
the average of an infinite number of replicate measured quantity values and a
reference quantity value – mainly related to systematic error), and measurement
precision (closeness of agreement between measured quantity values obtained by
replicate measurements – a measure of dispersion, mainly related to random er-
ror). Further details about the terminology and procedures for computing meas-
urement uncertainty can be found in ISO/IEC Guide 99:2007(E/F) (2007),
ISO/IEC Guide 98-3 (2008), ISO 5725-1,2,3,4 and 6 (1994), and ISO 5725-5
(1998).
The problem of obtaining reliable quantitative information concerning meas-
urement uncertainty for a given instrument is complex, and is undermined by
different interpretations concerning terminology and statistical procedures that
have been adopted by the manufacturers and by the industrial community in gen-
eral over the years. For example, the term “accuracy”, now generally assumed as
encompassing both trueness and precision, has been previously associated with
the meaning of trueness alone; the term “precision” itself has sometimes been
mistaken for accuracy, at least in colloquial terms, and used consequently. As a
consequence of this, the procedures for calibrating the instruments, and thus ensur-
ing metrological traceability, have sometimes suffered as well. Currently, the
more traditional and well-established stylus-based 2D profilometers are the solu-
tion with the best metrological traceability among all the instrument classes that
are discussed in this chapter. This is due to their widespread, long-term adoption,
and to the existence of measurement standards and procedures dedicated to such
instruments and their calibration; see, for example, ISO 5436-1 (2000), ISO 5436-2
(2001), and ISO 12179 (2000). Stylus-based 3D profilometers are a somewhat
newer breed and two ISO standards have recently become available for them:
ISO 25178-601:2010 and ISO 25178-701:2010.
Except for portable roughness measurement instruments, which can almost always
be brought directly onto the surface to be measured (unless specific impediments
exist on the part), all other stylus-based profilometers are generally implemented
as fixed installations, and require the specimen to be placed on its proper fixture
within the instrument. Compared with noncontact solutions, stylus-based systems
are generally more robust to environmental conditions since measurement origi-
nates from physical contact between the probe and the surface. Nevertheless, like
in any other type of measurement solution, the higher the accuracy of the instru-
82 N. Senin and G. Campatelli
ment, the more the instrument is sensitive to environmental factors such as tem-
perature, vibrations, atmospheric conditions, and electromagnetic fields, stylus-
based instruments being no exception. Typical 2D and 3D profilometers, including
also portable roughness testers, usually operate without the need for enclosures:
this means that often even large parts can be measured as long as a suitable orien-
tation is found that allows the part to be placed without interfering with the in-
strument, while at the same time making the (usually small) region to be measured
properly aligned with the stylus probe.
Very high resolution instruments may require a specifically sized specimen to
be prepared. In this case the specimen may be physically extracted from the part
(destructive testing), or may be manufactured to mimic the surface on the actual
part. In general, as measurement resolutions increase, specifically prepared speci-
mens need to be smaller and more accurately manufactured. Particular attention is
usually dedicated to ensure that the specimen is manufactured so that the surface
to be measured can be accurately aligned to the x, y plane defined by the instru-
ment. Leveling errors can be compensated via software by acting on the coordi-
nates of measured points, but only as long as out-of-range conditions are not en-
countered by the stylus probe.
As stated earlier, since the stylus is in physical contact with the surface during
scanning, stylus-based profilometers are classified as contact-based measurement
instruments. Contact techniques raise several concerns: specimen and/or stylus
damage due to grooving, chipping, or wear, stick–slip interactions, stylus stuck
between rough surface features, and bouncing behavior during traversal are all
aspects that may introduce significant measurement error when not also damaging
the inspected part, or the instrument as well. Key aspects to consider when evaluat-
ing the suitability of a material for being acquired by means of a stylus-based pro-
filometer are the stylus contact force, typically in the range 9.81 × 10–8 –9.81 × 10–5 N
for stylus-based instruments, traversal speed, friction, and the presence of topog-
raphy features that may negatively affect the scanning process (e.g., high-aspect-
ratio features). The suitability ultimately depends also on the instrument and stylus
type, the required performance, and the application domain. When the prerequi-
sites are satisfied, stylus-based profilometers are the most robust solution currently
available.
Two main types of architectures are typically found for instruments equipped
with optical probes. The single-point architecture is shown in Figure 3.6a: the
probe acts as a distance measurement device for the point (actually a small region)
targeted by the light spot. The distance to the probe can be turned into a point
height with respect to a reference datum. Motorized x, y axes can be used to ac-
quire multiple points at different x, y coordinates. Single and multiple profile scan-
ning strategies can be implemented, the most common being raster scanning; as
opposed to stylus-based profilometers, the main scanning direction is not con-
strained to be on the x-axis.
specimen
specimen
specimen
holder column
x-axis
base
y-axis
a b
Figure 3.6 Example architectures for instruments equipped with optical probes: a single-point
measurement system (raster scanning achieved by a motorized x, y table), and b wide-field meas-
urement system (simultaneous acquisition of a surface region; a motorized x, y table can be used
to acquire multiple regions sequentially, and collate them later)
detector
focusing lens
in-focus spot
focal plane
As stated earlier, many types of focus-detection probes are available since there
are many ways to analyze the reflected light pattern formed at the detector, and
many ways to relate it to vertical displacement from the focal plane. Although one
may measure the degree of defocusing at the detector and turn it into a vertical
distance between the measured region and the focal plane, it is usually preferred to
implement a controller-driven lens that is translated vertically until the optimal
3 Quality Inspection of Microtopographic Surface Features 85
focus conditions are achieved at the detector. At that point, height information is
computed directly from the vertical travel of the lens. Vertical translation can be
assigned to the probe or to the specimen; traditionally actuated solutions (electric
motors) are slower but allow for greater translations, while PZTs allow for faster
movements but more limited ranges (better for measuring flatter surfaces).
Profilometers based on focus-detection probes such as the one described above
are implemented as single-point measurement instruments (see Figure 3.6a); the
raster scanning velocity is somewhat limited as the probe must be allowed to have
enough time to achieve the optimal focus conditions at each x, y coordinate.
While solutions such as the one implemented by the simplified probe depicted in
Figure 3.7 have been classified as active autofocusing solutions, since they involve
the emission of a ray or light and the analysis of the light reflected back from the
specimen surface in order to assess focus conditions, and thus distance, passive
autofocusing solutions also exist, where the focus conditions are detected by ana-
lyzing information pertaining to the digital image formed at the detector: in con-
trast measurement, for example, a region of an image is in focus when there is
there maximum intensity difference between two adjacent pixels; in phase detec-
tion two images are formed by dividing the incoming light – images are aligned
and compared in terms of phase differences in order to identify in-focus regions.
Wide-field focus-detection 3D microscopes can be implemented that make
use of passive autofocusing solutions: a sequence of images of the specimen are
taken while the optics (or the specimen) are lowered/raised by means of a motor-
ized z-axis, thus changing the focal plane at each image of the sequence. For each
image, in-focus regions (pixels) are detected and associated with a reference
height. Data extracted from the series of images are then combined into a single
height map.
detector
pinhole
pinhole
focusing lens
in-focus spot
focal plane
specimen profile
a b
Figure 3.9 Chromatic aberration probe for point-based distance measurement: a example fre-
quency spectrum of the signal received at the detector, and b position of the focal planes
are called the reference path and the measurement path). Such a difference in the
path lengths can be converted into a height measure for the corresponding speci-
men surface point, as long as the position of the corresponding point on the refer-
ence mirror (i.e., the length of the reference path) is known.
A common variation of the architecture depicted in Figure 3.10 specifically de-
signed to operate at the micro and submicro scales is illustrated in Figure 3.11a
and is known as the Mirau interferometer, or Mirau objective (Bhushan et al.
1985).
Detector
Light
emitter
Specimen
surface
detector
Objective
lens Beam
splitter
Light
Reference
emitter
mirror Objective
beam
splitter lens
beam
Reference
splitter
mirror
Specimen
Region being surface
acquired Specimen
surface
a b
Figure 3.11 Mirau interferometric microscope architecture: a Mirau objective (Mirau interfer-
ometer), and b microscope assembly
3 Quality Inspection of Microtopographic Surface Features 89
interferogram. As the probe (or the surface) translates vertically by a step (hence
vertical scanning), local maxima (constructive interference) in the interferogram
allow for identification of those surface points whose height matches the above-
mentioned conditions; their actual z height can be retrieved by combining the
information obtained by the PZT with the known reference path length. The proc-
ess continues until all the x, y pairs populating the detector grid have been as-
signed a z coordinate. Different algorithms exist for detecting with effectiveness
and reliability maximum constructive interference points in an interferogram ob-
tained at a given z position of the probe, for example coherence correlation inter-
ferometry (Taylor Hobson 2009). In terms of the light source, recent mainstream
VSI solutions have privileged the use of white light; VSI using white light is often
referred to as white light scanning interferometry or scanning white light interfer-
ometry. White light encompasses a wide array of wavelengths (theoretically, all of
them; however, as a side note, some instruments are currently narrowing the
bandwidth adopting colored light), and more importantly is characterized by short
coherence lengths. A short coherence length means that a large positive interfer-
ence effect will be visible only when the measurement and reference path lengths
match with great accuracy, while it will decay significantly when the measurement
length is equal to any multiple of the reference length. White light is therefore
ideal for conditions where only the perfect match must be reliably identified,
which is compatible with the mode of operation of VSI solutions.
Instruments adopting PSI or VSI usually acquire the points located within the
measured region simultaneously, each point being associated with a specific x, y
grid cell at the detector (i.e., they can be classified as wide-field instruments);
larger surface regions may be acquired by implementing stitching solutions paired
with motorized x, y tables.
Other types of profilometers use similar interferometric principles to implement
single-point measurement approaches (i.e., they are equipped with interferometric
probes for single-point distance measurement); in this case a PZT or motor is used
for vertical translation of the probe (in the search for maximum constructive inter-
ference), while a motorized x, y table is used to implement single-profile or raster
scanning.
interference
pattern formed
detector at the detector
conoscopic
holography polarizer
optical
assembly
birefringent
crystal
polarizer
lens
laser
beam beam
splitter
laser
source reflected
light
specimen
surface
In detail, the first lens of the assembly is responsible for reproducing a conicity
effect in the parallel beam, so that each ray is sent out at different angles. The
polarizer splits each ray into two components of different polarization and sends
them to the crystal. The crystal receives the two components and, since it is bire-
fringent (i.e., double refractive), slows down one of them, inducing a phase shift,
which is proportional to the incidence angle of each ray hitting the uniaxial crys-
tal. The phase shift creates the holographic interference effect between the two
92 N. Senin and G. Campatelli
light components, one acting as the reference beam for the other, when the two
phase-shifted components are recombined by the second polarizer. Finally, the
interference ray exiting from the second polarizer is sent to the detector, where it
forms a high-contrast interference pattern together with all the other rays. The
interference pattern is analyzed by a computer to obtain a distance measurement
for the surface point being targeted by the spot.
As for stylus-based instruments, also for optical instruments a wide array of ranges
and resolutions are available; the performance depends essentially on the specific
optical technique adopted by the instrument.
For focus-detection and confocal instruments, a survey of current offerings by
several manufacturers (Carl Zeiss 2009; Fries Research & Technology 2009; Hi-
rox 2009; Leica Mikrosysteme Vertrieb 2009; Olympus 2009; Sensofar Tech
2009; Solarius Development 2009; Veeco Instruments 2009) shows that the best
performing instruments may achieve vertical resolutions smaller than 1 nm over a
vertical range below 5 µm, while coarser instruments are found with vertical reso-
lutions of 1–2 µm, operating over a vertical range of approximately 30 mm. The
lateral resolution depends on several factors, and on whether the architecture is
wide field or single point. For single-point architecture, significant constraints on
the lateral resolution come from the probe spot size (the maximum resolution may
be limited by measurement averaging effects within the spot region) and x, y ta-
ble/probe positioning mechanisms; for wide-field instruments, probe components
such as the magnification lens and the pixel spacing at the detector play a relevant
role. For high-performance focus-detection and confocal instruments, a maximum
lateral resolution below 1 µm can be achieved. Analogously to stylus-based sys-
tems, the maximum lateral range is constrained by the vertical range of the probe,
since the likelihood of out-of-range measurement errors caused by specimen level-
ing errors increases with the width of the region measured.
Chromatic aberration sensors (Fries Research & Technology 2009) are charac-
terized by vertical resolutions varying from approximately 3 nm (within a range
below 300 µm) to 250 nm over 25 mm for coarser probes. The maximum horizon-
tal resolution is within the 1–14-µm range (usually, the higher the vertical resolu-
tion, the higher the lateral resolution).
Interferometric instruments (Ambios Technology 2009; FOGALE nanotech
2009; Fries Research & Technology 2009; Novacam Technologies 2009; Sensofar
Tech 2009; Solarius Development 2009; Taylor Hobson 2009; Veeco Instruments
2009; Zygo 2009) are characterized by the highest vertical resolutions among all
3 Quality Inspection of Microtopographic Surface Features 93
thus result in measurement errors. Small features (with respect to lateral resolution
power/focus spot size), for example, small holes and pins, may also result in er-
rors, or be averaged out during measurement. One of the key performance aspects
that must be considered when assessing the applicability of an optical measure-
ment technique is referred to as the maximum detectable slope: information taken
from a recent review paper (Hansen et al. 2006) indicates that focus-detection
instruments can detect a maximum slope of approximately 15°, while confocal
microscopes achieve up to 85°; interferometric instruments (white light scanning
interferometry in particular) lie somewhere in between, with detectable slopes of
approximately 30°. Conoscopic holography probes may achieve detectable slopes
up to 85° (Sirat and Paz 1998).
Heterogeneity of the measured material may cause problems as well, again
leading to local variations of reflectivity and optical constants. Impurities (dust,
liquids, and particles) may also cause local measurement errors, like for contact-
based measurement, and often require the surface to be cleaned before meas-
urement. For the same reasons, optical instruments are in general more sensitive
to environmental conditions, including atmosphere composition, and ambient
illumination.
The scanning electron microscope (SEM) is type of electron microscope that ac-
quires an image of the surface by scanning it with a high-energy beam of elec-
trons, in a raster scan pattern (Thornton 1968). It is classified as a nonoptical mi-
croscope since light is replaced by a beam of electrons. The SEM is not capable of
acquiring a 3D profile from a specimen surface topography per se; however, if
properly equipped and configured, a SEM can produce a 3D topography represen-
tation by means of a tilting specimen holder (tilting table) and the application of
stereophotogrammetry, giving rise to a stereoscopic SEM (stereo SEM).
The basic architecture of a SEM fit for 3D stereo scanning is illustrated in Fig-
ure 3.13. A high-energy beam of electrons, known as the primary beam, is gener-
ated by an electron gun acting as the cathode and is accelerated toward the speci-
men surface, which acts as the anode. The beam is guided by magnetic fields and
is focused onto a single small point of the specimen surface (the spot being usually
a few nanometers in diameter). Magnetic scanning coils control the beam deflec-
tion and drive the beam spot over the surface according to an x, y raster scanning
pattern.
As the primary beam hits the surface, different types of emissions occur owing
to the interaction between the high-energy electrons of the beam with the atoms of
the specimen. Three main types of emissions can usually be captured by equipping
the microscope with suitable detectors: primary electrons (electrons belonging to
the primary beam), which are reflected by the specimen surface (this is known as
3 Quality Inspection of Microtopographic Surface Features 95
elastic scattering and these electrons are also called backscattered electrons);
secondary electrons, which are produced by the surface itself thanks to the energy
received from the primary beam (this is known as inelastic scattering); and X-rays
generated as a consequence of the interaction. Each main emission type carries
different types of information. Primary electrons and X-rays provide indications
on the chemical composition of the specimen; secondary electrons can be used to
obtain indications of the local slope, as the intensity of their emission is related to
the incidence angle between the primary beam and the specimen topography at the
spot. Given this latter point, the preferred approach for obtaining surface topogra-
phy information is to record the emissions of secondary electrons during a raster
scanning process, and use the information to produce a grayscale image that re-
sembles the topography, as if it were observed by optical means.
The maximum lateral resolution that can be achieved with a raster scanning
process depends mainly on the energy of the electron beam: compromises must be
found between the desired resolution and the risk of damaging the specimen.
Scanning electron microscopy requires the specimen to be conductive (since it
acts as the anode), and measurement must take place in a vacuum, to avoid de-
grading the electron beam. Both aspects pose constraints on the specimen: non-
conductive surfaces can be coated, but the coating layer introduces changes in the
topography which may be undesirable, especially when measuring at such small
scales; specimens with risk of releasing vapors need to be dried or frozen. A varia-
tion of a traditional SEM, the environmental SEM (ESEM), allows for operation at
lower pressures and humid atmosphere (Danilatos and Postle 1982) by adopting
more sophisticated detectors.
electron gun
selectable
magnetic aperture
lenses
scanning
coils
electron
detectors
tilting sample
holder
vacuum
The only type of nonoptical microscope discussed in this work is the SEM, which
by itself is only a 2D imaging device and thus can only be described in terms of
lateral resolution. As stated earlier, the main factor in determining lateral resolu-
tion is electron beam energy. An overview of the offerings of several manufactur-
ers indicates maximum lateral resolutions in some cases smaller than 1 nm, or
more often of a few nanometers depending on the beam energy (Aspex 2009;
CamScan Electron Optics 2009; FEI 2009; Hitachi High Technologies America
2009; JEOL 2009; Nanonics Imaging 2009; VisiTec Microtechnik 2009). The
actual performance may be limited in some real-life application scenarios for ma-
terials not withstanding the energies required to achieve the highest resolutions.
The lateral range for typical SEMs is a few hundred micrometers. ESEM perform-
ance varies compared with that of conventional SEM: the energy of the electron
beam is dissipated more in ESEM instruments owing to the atmosphere, thus po-
tentially resulting in lower contrast. Lateral resolution in ESEMs can be lower
than, approximately equal to, or higher than in a conventional SEM, depending on
beam energy.
3 Quality Inspection of Microtopographic Surface Features 97
Vertical resolutions for SEM devices producing 3D output by means of the ste-
reo-pair technique have been reported varying from a few nanometers to a few
tens of nanometers. However, the performance may be significantly affected by
several factors: pixel pair-matching may not work as desired, and even assuming
that an ideal pixel pair-matching performance could be achieved, the smallest
measurable height difference would be related to the smallest measurable slope,
and thus to a combination of pixel color resolution, spacing, and the topographic
properties of the surface measured. Additional issues that may degrade the resolu-
tion are related to different focus conditions associated with the image pair, which
in turn may lead to slightly different magnifications. ESEM instruments usually
produce noisier images than conventional SEM instruments, thus affecting also the
quality of 3D reconstruction.
Some information is available concerning the accuracy and precision of con-
ventional SEMs, and nominal data are available from most manufacturers. Cali-
bration artifacts are commercially available for SEMs operating as 2D imaging
devices, which reproduce various types of 2D patterns. However, scarce informa-
tion exists for stereo SEM solutions, and – as stated earlier – the performance
remains strongly related to the surface type (material and texture).
The two main drawbacks of conventional scanning electron microscopy are that
the specimen must be conductive (since it acts as the anode), and that measure-
ment takes place in a high vacuum, to avoid degrading the electron beam. Both
aspects pose constraints on the specimen: nonconductive surfaces can be coated,
but the coating layer introduces changes in the topography which may not be de-
sirable, especially when measuring at such small scales; specimens with risk of
releasing vapors need to be dried or cryogenically frozen. ESEM could be used to
reduce the need for vacuum conditions; the price to pay is a more significant dis-
sipation of electron beam energy. High-energy beams are necessary to achieve the
highest resolutions. Again, not all materials may withstand such energies without
experiencing degradation effects. Three-dimensional surface topographies ob-
tained by stereo pairs through photogrammetry are seldom reliable enough for
extended quantitative inspection tasks, owing to the number of issues related to the
identification of correct pixel pairs and the significant reconstruction errors asso-
ciated with wrong matches. For this reason, 3D surface topographies reconstructed
through stereophotogrammetry on SEM images are mainly confined to laboratory
usage and research work. However, qualitative inspection of surface topography
by SEM 2D imaging remains one of the strongest and most useful approaches and
often provides invaluable support to other 3D surface topography measurement
techniques.
98 N. Senin and G. Campatelli
z
tip tip
x
tunneling
current
surface +
surface
a b
Figure 3.14 Scanning tunneling microscope: a overall schema, and b close-up view of the tip–
surface interaction
The most widespread SPM alternative to the STM is the AFM, which is essen-
tially a derivation of the STM itself. In the AFM (see Figure 3.15), the probe con-
sists of a microscopic cantilever terminating with a tip, with a radius of a few
nanometers. The tip is brought very close to the surface, so it is affected by the
interatomic forces existing between the atoms of the surface and the tip itself. As
the tip approaches the surface, the atomic forces are initially attractive, and then
become repulsive, which is basically “contact” at such small scales. Atomic forces
can be indirectly measured from the deflection of the cantilever, since its stiffness
is known. The deflection, in turn, can be measured in several ways, the most
common being by means of a laser beam reflected by the cantilever and hitting an
array of photodiodes (see Figure 3.15a).
laser
detector emitter interatomic
repulsive
tip force
regime
cantilever interatomic
force
surface 0
d d
surface
attractive
regime
a b c
Figure 3.15 Atomic force microscope: a overall schema, b close-up view of the tip–surface
interaction, and c interatomic force curve
100 N. Senin and G. Campatelli
Several approaches exist for acquiring a surface topography. Recall the tradi-
tional stylus-based profilometer illustrated in Section 3.2.1. In the stylus-based
profilometer, deflection is measured in an open-loop manner while the tip is trans-
lated at constant reference z height over the surface; this is typically not applicable
in AFMs, given the increased risk of collision and damage of the delicate tip at
such small scales. Instead, a wide variety of closed-loop solutions are available,
the most important being listed in the following.
In contact mode (or static mode), the tip is so close to the surface that the
atomic force is repulsive (see Figure 3.15c) and deflects the cantilever upward. A
PZT raises/lowers the cantilever with respect to the measured surface point so that
the deflection force is kept constant. The vertical displacements generated by the
transducer during traversal are recorded as height points, and are used to recon-
struct the surface topography.
In noncontact mode, the tip is subjected to forced oscillation and is kept at a
distance from the surface so that the atomic forces are mainly attractive; changes
in the resonant frequency or amplitude in the cantilever oscillation can be related
to changes of material and topographic properties of the underlying surface point.
In tapping mode, the tip is subjected to forced oscillation and is placed at a dis-
tance from the surface so that, while oscillating, it is partially subjected to attrac-
tive forces, and partially to repulsive forces (this is known as intermittent contact);
again the oscillation can be studied to obtain surface material and topographic
properties. With respect to the contact mode, the tapping mode eliminates lateral
forces, such as drag, and reduces the risk of damaging the surface.
Additional approaches exist that apply variations of the illustrated AFM tech-
niques for obtaining surface information that go beyond simple topography (such
as friction properties and elasticity); the most widely known include lateral force
microscopy, force modulation microscopy, magnetic force microscopy, and elec-
trostatic force microscopy. Since most such techniques can be obtained from the
same basic SPM structure by simply changing the probe, several commercial mul-
timode SPM instruments are currently available.
STMs have the highest resolutions amongst all the instruments illustrated in this
chapter. A survey of the offerings of several STM and AFM manufacturers
(Agilent Technologies 2009; AIST-NT 2009; Ambios Technology 2009; Asylum
Research 2009; Bruker AXS 2009; Fries Research & Technology 2009; JEOL
2009; Micro Photonics 2009; Nanonics Imaging 2009; Nanosurf 2009; Novascan
Technologies 2009; NT-MDT 2009; Park Systems 2009; Veeco Instruments 2009)
3 Quality Inspection of Microtopographic Surface Features 101
shows that the maximum vertical resolution of a STM is approximately 1 pm, over
a vertical range of 200 nm, and the maximum lateral resolutions are less than
10 pm. AFMs have slightly poorer performance, but still can achieve vertical
resolutions of 0.1 nm, over a vertical range of 10–20 µm, with lateral resolutions
of a few nanometers.
Information on accuracy and precision is available from several manufacturers;
some high-resolution, traceable calibration artifacts are also commercially avail-
able that reproduce specific patterns that can be used to determine SPM perform-
ance experimentally. Once again, however, performance tends to vary signifi-
cantly depending on materials and surface properties, as discussed also in the next
section.
As for the SEM, the main drawback of the STM is that the surface to be measured
must be of conductive or semiconductive material. This greatly limits the applica-
bility of the instrument. Nonconductive surfaces can be coated, but the coating
layer introduces changes in the topography which may not be desirable, especially
when measuring at such small scales. The AFM mode of operation makes the
AFM appear very similar to a stylus-based instrument; thus, it may appear more
robust and generally applicable, although at slightly inferior resolutions. However,
the nature of the contact at such small scales makes the application of the AFM for
quantitative measurement of surface topography slightly more complex, as the
signal detected by the probe may be influenced by factors other than topography,
such as material properties.
SPMs such as the AFM and STM are usually built so that measurement takes
place in a confined enclosure, to achieve a great degree of separation from the
surroundings; as a result, specimens are usually very small, and must be specifi-
cally prepared to be compatible with the instrument.
The classes of profilometers and microscopes that have been illustrated so far have
been conceived to fulfill specific functional roles in conventional quality control.
Profilometers come from the domain of surface finish analysis at micro and sub-
micro scales, where topography is seen as a combination of wavelike height varia-
tions describing roughness, waviness, and form error, and the main objective is to
provide a quantitative measurement of such height variations. Microscopes are
conceived to support visual inspection from a mainly qualitative standpoint, while
102 N. Senin and G. Campatelli
A selection of topics has been identified in this work, each constituting a relevant
subject to be investigated when assessing the possibility to use profilometers and
3D microscopes for evaluating geometric error on microfabricated parts and sur-
face features. These topics can be synthetically ascribed to the following terms:
• different measurement performance in x, y and z;
• unidirectional probing;
• raster scanning; and
• image-inspired data processing.
Each one raises specific issues, as discussed in the following sections.
by PZT or electric motors. More accurate solutions are based on direct measure-
ment of displacement and adopt additional distance measurement probes (e.g.,
interferometric). Since different solutions are adopted for acquiring the x, y and z
coordinates, the accuracy, precision, resolution, range, and any other property
related to measurement performance for each coordinate are potentially different.
Since traditional surface finish analysis is mainly concerned with the study of
deviations of surface heights, it is more likely to have information about the meas-
urement performance about the z-axis when using a profilometer, while informa-
tion about the x- and y-axes may be lacking. Conversely, when a SEM or a similar
device rooted in the 2D imaging realm is used, it is often more likely to provide
metrological information about lateral accuracy and precision, rather than about
height measurement performance.
Since profilometers and 3D microscopes are designed with the main goal of ac-
quiring vertical distance (height) measurements, they are designed so that the
probe approaches the surface from the same direction. This has several important
consequences.
The first is a limitation on the types of geometric features that can be success-
fully measured, as geometric undercuts cannot be acquired. At any x, y position,
the probe can usually measure only the distance to the closest surface point
aligned on the vertical (z) axis. A notable exception is 3D microscopes capable of
acquiring volume data (e.g., confocal laser scanning microscopes), but only if the
material properties are compatible with volume data acquisition.
The second consequence of unidirectional probing – actually, strictly related to
the first – is that measurement performance is strongly related to the orientation of
the surface being measured. Data on vertical or high-slope surfaces are usually
difficult to acquire and measurement performance is degraded with inclination to
the point that factors such as the maximum detectable slope are key for preferring
a measurement technique over another. It is often problematic, if not impossible,
to acquire data on complex multifaceted surfaces in a single measurement session
(i.e., without repositioning the probe or the specimen), or even data on single sur-
faces with high curvatures.
The third and last consequence of unidirectional probing is related to the nature
of the geometric data available as a result of the measurement, which is often
colloquially referred to as 2.5D geometry, i.e., not fully 3D geometry. Geometric
models result from height maps and, again, the only way to obtain a full 3D geo-
metric model of a specimen (e.g., a microfabricated part) is to reconstruct it from
multiple measurements (i.e., multiple probe/workpiece orientations) and through
stitching techniques in the attempt to compensate for the accuracy losses due to
repositioning.
All the limitations of unidirectional probing have been long accepted in the
domain of surface finish analysis, but less so in the new domain of quality inspec-
104 N. Senin and G. Campatelli
tion of microfabricated parts and surface features, where the relevance of such
issues depends on the type of feature investigated and on the manufacturing proc-
ess that was used to fabricate it. For manufacturing processes that are essentially
2.5D themselves (e.g., LIGA, etching, lithography), unidirectional probing may be
tolerated as long as the functional requirements and types of tolerances to be in-
spected can be expressed in a 2.5D space as well. The same considerations apply
to the maximum measurable angle, as with some manufacturing processes high-
slope surfaces cannot be obtained anyway. However, for some micromanufactur-
ing processes that make it easier to obtain full 3D geometries (e.g., micromilling,
laser micromachining), the limitations of unidirectional probing may be unaccept-
able and more complex measurement strategies may need to be developed.
In the conceptual paradigm underlying surface finish analysis, the nominal geome-
try is assumed to be smooth and regular (e.g., a horizontal plane) and error is as-
sumed as a composition of height deviation components represented as harmonic
oscillations of different amplitudes and wavelengths. Assuming such a paradigm
exists, it is perfectly reasonable to implement a raster scanning process where the
surface is sampled with uniform spacing according to a regular grid. Spacing al-
lows for determining the range of wavelengths captured by the measurement proc-
ess. When form error on an average-sized manufactured part is inspected, in con-
trast, the nominal geometry is usually provided in the form of a multifaceted
geometric model, and edges, regions with high curvatures, and other significant
features can be used to plan a measurement process optimized for the geometry at
hand.
If profilometers and 3D microscopes are to be used to inspect microfabricated
parts and surface features, this is like being in the condition where possibly com-
plex geometries are to be measured with raster scanning, i.e., rather suboptimal,
strategies. An example of this type of problem is illustrated in Chapter 5. Trade-
off solutions must be pursued to identify optimal sampling resolutions that take
into account instrument capability, cost of the measurement process, representa-
tiveness of the result, measurement uncertainty.
to surface topography analysis, on the other hand, it implicitly drives the analysis
to consider geometric data as images, and to treat any transform as a transform
whose output is an image. To give a simple example, while alignment between
geometries in a 3D space implies rigid roto-translation transforms, on images any
translation and/or rotation takes place in the x, y plane, and usually implies pixel
remapping, and potentially, resampling. In the domain of surface finish analysis,
this image-inspired approach has usually been considered acceptable; when deal-
ing with the assessment of geometric error on microfabricated parts, this may no
longer be the case.
Some of the problems that must be faced and solved when applying profilometers
and microscopes to the measurement of micromanufactured surface features and
parts are actually also common to more traditional metrology applications, such as
the measurement of an average-sized mechanical part by means of a CMM. In
some cases, the solution is scale- and instrument-independent, in other cases, spe-
cific aspects must be taken into account.
A shared problem when assessing geometric error on any part and with any in-
strument is being able to spatially relate (register) the measurement to a reference
(nominal) geometry. The reference geometry could be provided as an explicit
model (e.g., CAD model for quality inspection of an average-sized mechanical
part) or as a set of mathematical equations (a plane, a cylinder, a parametric sur-
face, etc.). In conventional surface topography analysis at the micro and submicro
scales, which is the domain from which profilometers and microscopes come, the
reference geometry is usually assumed to be a plane, or similarly a simple surface
(e.g., second-order or third-order polynomial). Once properly placed with respect
to measurement points, height differences are used to compute roughness, wavi-
ness, and form error parameters.
Regardless of the scale, application, and model type, the registration of the ref-
erence geometry and the measured points is almost invariably determined through
some sort of point fitting. The way fitting is accomplished may change depending
on the application. For example, in roundness evaluation least-squares fitting and
minimum-zone fitting are the two most popular approaches; in surface finish
analysis, since most of time the reference geometry is a plane with limitless exten-
sions, least-squares fitting is used to properly place the reference surface in its
106 N. Senin and G. Campatelli
As mentioned several times earlier, data stitching is a process that is used to regis-
ter data from different measurements so that wider portions of a given geometry
can be obtained even from instruments with more limited ranges. Stitching is
based on the alignment of those portions of each data set which are known to be
derived from the same region of the measured geometry, thus implying that the
measurement itself must be planned so that it provides some degree of overlap-
ping. Stitching techniques have always been popular in geometric reconstruction,
as they allow a fully 3D representation of a manufactured part to be obtained by
collating measurements taken from different directions. In the domain of surface
microtopography analysis by means of profilometers and microscopes, the appli-
cation of stitching techniques is more recent, especially in noncontact instruments,
and is mainly aimed at achieving high resolutions together with high measurement
ranges.
When dealing with the inspection of microfabricated parts and surface features
by means of profilometers and microscopes, stitching techniques may play a fun-
damental role, especially when considering the aforementioned issues related to
undercuts and maximum detectable slope, and when considering also the overall
issues related to resolution and range for high-aspect-ratio surface features. Simi-
lar to the issues of registering the nominal and measured geometry, the problem of
collating measurements together in a successful way is shared with conventional
inspection of mechanical parts, and most of the algorithmic solutions developed
are scale-independent.
3.4 Conclusions
References
ISO 3274 (1998) Geometric product specifications (GPS) – surface texture: profile method –
nominal characteristics of contact (stylus) instruments. International Organization for Stan-
dardization, Geneva
ISO 4287 (1997) Geometrical product specifications (GPS) – surface texture: profile method –
terms, definitions and surface texture parameters. International Organization for Standardiza-
tion, Geneva
ISO 4288 (1996) Geometrical product specifications (GPS) – surface texture: profile method –
rules and procedures for the assessment of surface texture. International Organization for
Standardization, Geneva
ISO 5436-1 (2000) Geometrical product specifications (GPS) – surface texture: profile method;
measurement standards – part 1: material measures. International Organization for Standardi-
zation, Geneva
ISO 5436-2 (2001) Geometrical product specifications (GPS) – surface texture: profile method;
measurement standards – part 2: software measurement standards. International Organization
for Standardization, Geneva
ISO 5725-1,2,3,4 and 6 (1994) Accuracy (trueness and precision) of measurement methods and
results. Parts: 1,2,3,4 and 6. International Organization for Standardization, Geneva
ISO 5725-5 (1998) Accuracy (trueness and precision) of measurement methods and results – part
5: alternative methods for the determination of the precision of a standard measurement
method. International Organization for Standardization, Geneva
ISO 8785 (1998) Geometrical Product Specifications (GPS) – surface imperfections – terms
definitions and parameters. International Organization for Standardization, Geneva
ISO 11562 (1996) Geometrical product specifications (GPS) – surface texture: profile method –
metrological characteristics of phase correct filters. International Organization for Standardi-
zation, Geneva
ISO 12085 (1996) Geometrical product specifications (GPS) – surface texture: profile method –
motif parameters. International Organization for Standardization, Geneva
ISO 12179 (2000) Geometrical product specifications (GPS) – surface texture: profile method –
calibration of contact (stylus) instruments. International Organization for Standardization,
Geneva
ISO 13565-1 (1996) Geometrical product specifications (GPS) – surface texture: profile method
– surfaces having stratified functional properties – part 1: filtering and general measurement
conditions. International Organization for Standardization, Geneva
108 N. Senin and G. Campatelli
ISO 13565-2 (1996) Geometrical product specifications (GPS) – surface texture: profile method
– surfaces having stratified functional properties – part 2: height characterization using the
linear material ratio curve. International Organization for Standardization, Geneva
ISO 13565-3 (1998) Geometrical product specifications (GPS) – surface texture: profile method
– surfaces having stratified functional properties – part 3: height characterization using the
material probability curve. International Organization for Standardization, Geneva
ISO 25178-6:2010 Geometrical product specifications (GPS) – Surface texture: Areal – Part 6:
Classification of methods for measuring surface texture, International Organization for Stan-
dardization, Geneva
ISO 25178-601:2010 Geometrical product specifications (GPS) – Surface texture: Areal – Part
601: Nominal characteristics of contact (stylus) instruments, International Organization for
Standardization, Geneva
ISO 25178-602:2010 Geometrical product specifications (GPS) – Surface texture: Areal – Part
602: Nominal characteristics of non-contact (confocal chromatic probe) instruments, Interna-
tional Organization for Standardization, Geneva
ISO 25178-701:2010 Geometrical product specifications (GPS) – Surface texture: Areal – Part
701: Calibration and measurement standards for contact (stylus) instruments, International
Organization for Standardization, Geneva
ISO/IEC Guide 98-3 (2008) Uncertainty of measurement – part 3: guide to the expression of
uncertainty in measurement (GUM:1995), 1st edn. International Organization for Standardi-
zation, Geneva
ISO/IEC Guide 99:2007(E/F) (2007) International vocabulary of basic and general terms in
Metrology (VIM). International Organization for Standardization, Geneva
Abbott EJ, Firestone FA (1933a) A new profilograph measures roughness of finely finished and
ground surfaces. Autom Ind 204
Abbott EJ, Firestone FA (1933b) Specifying surface quality: a method based on accurate meas-
urement and comparison. Mech Eng 55: 569–572
Agilent Technologies (2009) Agilent Technologies – home page. http://www.home.agilent.com.
Accessed 30 Sep 2009
AIST-NT (2009) AIST-NT – home page. http://www.aist-nt.com. Accessed 30 Sep 2009
Ambios Technology (2009) Ambios Technology – home page.
http://www.ambiostech.com/index.html. Accessed 30 Sep 2009
Aspex (2009) Aspex Corporation – home page. http://www.aspexcorp.com. Accessed 30 Sep
2009
Asylum Research (2009) Asylum Research – home page. http://www.asylumresearch.com.
Bhushan B, Wyant JC, Koliopoulos CL (1985) Measurement of surface topography of magnetic
tapes by Mirau interferometry. Appl Opt 24(10):1489–1497
Binnig G, Rohrer H (1983) Scanning tunnelling microscopy. Surf Sci 126:236–244
Binnig G, Quate CF, Gerber C (1986) Atomic force microscope. Phys Rev Lett 56(9):930–933
Bruker AXS (2009) Bruker AXS – home page. http://www.bruker-axs.com. Accessed 30 Sep
2009
CamScan Electron Optics (2009) CamScan Electron Optics – home page.
http://www.camscan.com. Accessed 30 Sep 2009
Carl Zeiss (2009) Carl Zeiss International – home page. http://www.zeiss.com/explore. Accessed
29 Sep 2009
Danilatos GD, Postle R (1982) The environmental scanning electron microscope and its applica-
tions. Scanning Electron Microsc 1:1–16
Dupuy MO (1967) High-precision optical profilometer for the study of micro-geometrical sur-
face defects. In: Proceedings of the Institute of Mechanical Engineers, vol 182, part 3k,
pp 255–259
FEI (2009) FEI – home page. http://www.fei.com. Accessed 30 Sep 2009
FOGALE nanotech (2009) FOGALE nanotech – home page. http://www.fogale.fr. Accessed
30 Sep 2009
3 Quality Inspection of Microtopographic Surface Features 109
Sirat G, Paz F (1998) Conoscopic probes are set to transform industrial metrology. Sens Rev
18(2):108–110
Solarius Development (2009) Solarius – home page. http://www.solarius-inc.com/index.html.
Accessed 29 Sep 2009
Taylor Hobson (2009) Taylor Hobson – home. http://www.taylor-hobson.com. Accessed 30 Sep
2009
Thornton PR (1968) Scanning electron microscopy. Chapman and Hall, London
Veeco Instruments (2009) Veeco – solutions for a nanoscale world. http://www.veeco.com
VisiTec Microtechnik (2009) VisiTec – home page. http://http://www.visitec-em.de/. Accessed
30 Sep 2009
Weckenmann A, Estler T, Peggs G et al (2004) Probing systems in dimensional metrology. CIRP
Ann Manuf Technol 53(2):657–684
Werth Messtechnik (2009) Werth Messtechnik – home page. http://http://www.werth.de.
Accessed 17 Dec 2009
Zygo (2009) Zygo Corporation – home page. http://www.zygo.com. Accessed 28 Sep 2009
ISO/CD 25178-1 Geometrical product specifications (GPS) – surface texture: areal – part 1:
indication of surface texture. International Organization for Standardization, Geneva
ISO/CD 25178-604 Geometrical product specifications (GPS) – surface texture: areal – part 604:
nominal characteristics of non-contact (coherence scanning interferometry) instruments. In-
ternational Organization for Standardization, Geneva
ISO/CD 25178-605 Geometrical product specifications (GPS) – surface texture: areal – part 605:
nominal characteristics of non-contact (point autofocusing) instruments. International Or-
ganization for Standardization, Geneva
ISO/DIS 25178-2 Geometrical product specifications (GPS) – surface texture: areal – part 2:
terms, definitions and surface texture parameters. International Organization for Standardiza-
tion, Geneva
ISO/DIS 25178-3 Geometrical product specifications (GPS) – surface texture: areal – part 3:
specification operators. International Organization for Standardization, Geneva
ISO/DIS 25178-7 Geometrical product specifications (GPS) – surface texture: areal – part 7:
software measurement standards. International Organization for Standardization, Geneva
ISO/DIS 25178-603 Geometrical product specifications (GPS) – surface texture: areal – part 603:
nominal characteristics of non-contact (phase-shifting interferometric microscopy) instru-
ments. International Organization for Standardization, Geneva
Chapter 4
Coordinate Measuring Machine
Measurement Planning
Abstract Once a measuring instrument has been chosen for the control of the
quality of a part of known design and specifications, the measurement process must
be planned. For coordinate measuring instruments implementing point-based
measurement, planning implies appropriately choosing the number and placement
(pattern) of the points to be measured. In fact, coordinate measuring instruments
sample points on features to be measured, but how should points be located on the
feature itself? This problem is particularly relevant with measuring instruments
which require a long time to sample dense clouds of points, e.g., most coordinate
measuring machines (CMMs). In this chapter the problem of planning the inspec-
tion strategy, i.e., defining the number and pattern of sampling points, provided the
measuring systems allow the operator to define the inspection strategy, will be
addressed, with particular reference to CMMs. Sample size planning will be ap-
proached as an economic problem, because as the sample size increases, uncer-
tainty is reduced and measurement cost rises, and a trade-off has to be searched for.
Then, a few different criteria for defining the sampling pattern are proposed; these
differ in terms of the accuracy and the information required for their application.
These criteria can be categorized as blind, adaptive, and process-based sampling
strategies. A few examples are proposed, outlining the effectiveness of different
approaches to sampling strategy planning. In order to better understand the prob-
lem of strategy planning, a brief description of the main CMM features is provided.
__________________________________
G. Moroni
Department of Mechanical Engineering, Politecnico di Milano,
Via La Masa 1, 20156, Milan, Italy,
e-mail: giovanni.moroni@mecc.polimi.it
S. Petrò
Department of Mechanical Engineering, Politecnico di Milano,
Via La Masa 1, 20156, Milan, Italy,
e-mail: stefano.petro@polimi.it
111
112 G. Moroni and S. Petrò
4.1 Introduction
The CMM concept may be associated with a very large family of measuring sys-
tems, so there will be CMMs with contacting or noncontacting probes, CMMs
with bridge structures, CMMs with a horizontal arm, manual or motorized CMMs,
and small or large CMMs. Anyway, regardless of the implementation of the spe-
cific CMM concept, CMMs are flexible. Their ability to sample points in a 3D
space allows them to check any geometric or dimensional tolerance. Of course,
different CMMs are more or less suitable for different measurement tasks, but
broadly speaking any CMM can measure any part subject to geometric specifica-
tions, provided the CMM measuring volume is adequate (the part is not too large).
A CMM is essentially constituted by the following parts (Figure 4.1) (Bosch
1995):
• mechanical setup (machine axes and transducers);
• sensor;
• control unit; and
• computer with software for data processing.
The CMM configuration describes the overall structure of the machine. Even
though the commercial solutions are very different, they can be reduced to five
basic configurations: bridge, cantilever, horizontal arm, gantry, and non-Cartesian.
ISO 10360-1 (2000) further details these structures, distinguishing machines with
a fixed or moving table, column-type CMM, etc., which are variations of the pre-
viously mentioned configurations.
Differing measurement tasks require different characteristics from CMMs.
Therefore, different CMM configurations have been proposed, each one with its
advantages and disadvantages.
4 Coordinate Measuring Machine Measurement Planning 113
A cantilever CMM allows for good accessibility of the measuring volume, but
the cantilever arm reduces the structure stiffness, thus reducing the overall accu-
racy. This may be avoided by adopting a bridge structure, which improves stiff-
ness, even though accessibility is reduced. Gantry and horizontal arm CMMs are
very effective for large volume measurement, owing to their inherent structure
characteristics. However, because in these kinds of CMMs the fixture supporting
the part and the structure moving the measuring sensor are often not directly con-
nected, accuracy is reduced with respect to other solutions. Finally non-Cartesian
CMMs are usually less accurate than any other configuration. However, because
of their ability to move “around” the part, this kind of CMM may be really flexible
from the accessibility point of view and really suitable to sample large clouds of
points in a short time, as is required for reverse modeling purposes, when coupled
to a sensor like a laser stripe.
A CMM sensor, or probe, is that part of the machine which “touches” (physically,
or optically, or with any other principle) the part to be inspected. Most of the
CMM accuracy relies on sensor accuracy, which strongly influences the overall
114 G. Moroni and S. Petrò
The control unit of a CMM coordinates the various parts of the CMM itself. Dif-
ferent kinds of controls are possible. The simplest control is manually driven, both
with or without axis motorization. However, most of the present CMMs are con-
trolled by a computerized numerical control (CNC) system. The CNC system may
control the machine in discrete-point probing, or scanning (ISO 10360-1 2000). In
discrete-point probing, after probing a point, the CMM probe leaves the surface
and moves toward the next point, and so on. In scanning, the probe is in continu-
ous contact with the surface, so a line is sampled. Scanning may be performed
both on a predefined path or on an unknown path. In this last situation, the control
system has to able to adapt the probing path in order to move from a point on the
surface to be inspected to another point without losing contact with the surface
itself.
4 Coordinate Measuring Machine Measurement Planning 115
The control unit is also responsible for software compensation (Sartori and
Zhang 1995; Schwenke et al. 2008). In recent years, CMM manufacturers have
pointed out that further increasing CMM mechanical accuracy is not economic.
The solution has been to “map” the measurement errors in the CMM measuring
volume, based on repeated measurement of specifically designed, calibrated or
uncalibrated, artifacts. Then, because the errors due to mechanical inaccuracy are
known, they may be compensated in successive measurement tasks, thus reducing
measurement uncertainty.
The output of the CMM itself is just a cloud of points. In order to evaluate dimen-
sions and geometric errors, the cloud of points has to be further analyzed. This
analysis is performed by specific software. The principal function of the software
is then to fit the clouds of points in order to extrapolate measurements.
MZ tolerance zone
LS tolerance zone
LS Circumference
Figure 4.2 Least-squares (LS) and minimum-zone (MZ) fitting for a circle
Two fitting principles (Anthony et al. 1996) are mainly used for analyzing
clouds of points: least-squares (also known as Gaussian) and minimum-zone (also
known as Chebyshev) fitting (Figure 4.2).
Least-squares fitting consists in solving the problem
n
min ∑ d i2 ( p ) , (4.1)
p
i =1
116 G. Moroni and S. Petrò
where p is a vector of parameters defining the fitting geometry, di(p) is the signed
distance of the ith point from the fitting geometry, and is, of course, a function of
p, and n is the number of sampling points.
Minimum-zone fitting is defined by the solution of the following minimum–
maximum problem:
gauge blocks and step gauges to be adopted as length standards, but the standard is
under revision, in order to allow the use of instruments such as laser interferome-
ters, which are required for large-volume CMMs. Three repetitions of a set of five
length standards have to be performed, in seven different positions within the
measuring volume of the CMM.
Probing performance is defined by MPEP, the “extreme value of the probing er-
ror P permitted by specifications, regulations, etc.”, where P is the “error of indi-
cation within which the range of radii of a spherical material standard of size can
be determined by a CMM” (ISO 10360-1 2000). Its test is performed by measur-
ing a reference (calibrated) sphere according to a pattern of 25 points specified by
the ISO 10360-2 (2001) standard. Similar performance indicators have been pro-
posed for scanning CMMs (ISO 10360-4 2000) and multistylus or articulating
probe CMMs (ISO 10360-5 2000).
Currently, these procedures are suitable only for CMMs equipped with contact-
ing probes. The ISO 10360-7 standard is currently under development for CMMs
with imaging probes, and at present it exists only as a draft (ISO/DIS 10360-7).
This lack of a reference standard makes noncontact CMMs not completely trace-
able.
• hardware uncertainty sources, that is, anything related to the structure (sensor,
mechanical structure, etc.) of the CMM;
• workpiece uncertainty sources, related to properties of the workpiece and
measurement interaction with the workpiece;
• sampling strategy, including inadequate sampling or datums, and interaction
between the sampling strategy and the actual geometric error;
• fitting and evaluation algorithms; and
• extrinsic factors, such as temperature, operator, and dirt.
This variety of uncertainty sources has led to a few uncertainty evaluation pro-
cedures, summarized in the ISO 15530 series of standards. Currently, only parts
three (ISO/TS 15530-3 2004) and four (ISO/TS 15530-4 2008) have been pub-
lished.
ISO/TS 15530-3 (2004) proposes a way of evaluating uncertainty which is
based on repeated measurement of one or more calibrated artifacts in the same
conditions in which the following measurements will be performed. Even though
the procedure is quite simple, for each uncertainty evaluation a calibrated artifact,
which has to be as similar as possible to the real parts, is required. Therefore, this
118 G. Moroni and S. Petrò
After a CMM has been selected to perform a specific measurement task, the in-
spection has to be planned. Inspection planning consists in defining the single
aspects of the measurement:
• fixturing definition;
• sensor configuration;
• sampling strategy definition; and
• path planning.
Inspection
GeometricModeler Feature Selection
+
ToleranceModeler
Sampling Strategy
Sensor Configuration
Fixture Configuration
Path Planning
Simulation
Figure 4.3 An automatic inspection planning system for a coordinate measuring machine
4 Coordinate Measuring Machine Measurement Planning 119
• Select the features to be inspected (Fiorentini et al. 1992); in fact, the number
of features to be inspected in a single part is often very large. This could lead to
a long measurement time. Therefore, methods are required to choose which
features have to be inspected.
• Configure the sensor and the fixture (Moroni et al. 1998), that is, define how
the fixture and the sensor should be designed in order to minimize part defor-
mation, ensure surface accessibility, and minimize risk of collision between the
probe and the part.
• Plan the probe path (Yau and Menq 1995), avoiding collision between the
part/fixture and the probe.
• Plan the sampling, or measurement, strategy. Strategy planning consists in
defining how many sampling points should be taken, and in which locations.
Because this last factor is responsible for most of the measurement uncertainty,
the rest of this chapter will illustrate methods of measurement strategy planning.
Under the assumption that a coordinate measuring system has been set up, i.e., a
suitable fixture for the part has been defined, allowing for easy access to every
feature to be inspected, the probe is correctly configured and qualified, and so on,
the next problem to solve is the choice of where to sample the part feature in order
to obtain a description of the surface complete enough to estimate substitute fea-
tures with sufficient accuracy for the measurement task, but at the same time
without exceeding the sample size, because the measurement time and cost tend to
increase as the sample size increases; this is in part true even for standstill scan-
ners. In fact, if the lateral resolution of the scanner is not adequate in the current
setup, the operator will have to take multiple scans of the part and then register
them – of course, more scans means a longer measuring time. This problem is
known as “defining a sampling strategy”, where the sampling strategy is the com-
plete description of locations of the sampling points.
When dealing with geometric error estimates, the sampling strategy is one of
the most relevant contributors to measurement uncertainty. Oversimplifying, it
120 G. Moroni and S. Petrò
could be stated that, according to definitions in ISO 1101 (2004), geometric error
is defined by the point that deviates the most from the nominal geometry; how-
ever, in a feature sampling process the actual measurement is not a complete de-
scription of the feature itself, it is just an approximation. Therefore, a sampling
strategy is effective if it is able to catch the most deviating point of the feature
most of the times. In fact, according to the actual definitions of tolerance zones,
only these points influence compliance or noncompliance of parts, that is, if these
points fall inside the tolerance zone, then the part is compliant. Of course, if the
most deviating point is not caught by the sampling, some measurement error will
be present. Figure 4.4 exemplifies how this error originates. First of all, it may be
interesting to point out that this measurement error will be present even if the
sampling process does not generate any measurement error, that is, e.g., for a
CMM equipped with a touching probe, when the measured coordinates of the
contacting point coincide with the coordinates of the real point for any sampling
point. In this case, the measurement error will consist of an underestimate of the
geometric error (see Figure 4.4a, which shows what happens if a nominally circu-
lar, truly elliptical profile is sampled with an insufficient density of points), and
the underestimate is inversely related to the sample size. However, because any
real sampling process generates some sampling error, it is possible to overestimate
the geometric error. Consider, for example, the sampling of a perfect geometry,
i.e., a manufactured geometry which is identical to its nominal counterpart. Of
course, the geometric error should be null, but because of probing error, the sam-
pling points will probably not belong to the perfect geometry, so the geometric
error estimate will be greater than zero, and will tend to increase as the sample
size increases, because measurement error dispersion will tend to form a sort of
envelope around the real geometry. Similarly, if a nonperfect feature is inspected,
the sampling error tends to inflate the geometric error estimate (Figure 4.4d). Un-
fortunately, in most practical application scenarios involving coordinate measuring
systems the underestimation due to undersampling is more relevant than the over-
estimation due to probing error, so geometric error estimates usually include some
residual systematic measurement error.
Therefore, in order to obtain geometric error estimates as accurate as possible, a
correct sampling strategy has to be defined, able to identify those areas of the
surface which have the maximum local geometric deviation, that is, the maximum
deviation of the real feature from the nominal feature (ISO 14660-1 1999). This
can be obtained by sampling a large number of points uniformly distributed across
the feature to be inspected. If the measuring system adopted is fast (that is, it is
able to sample a large number of points in a short time), this is probably the best
way to sample, because there is no risk of missing some relevant areas of the fea-
ture itself. However, measuring instruments characterized by the minimum prob-
ing error are often slow, so the problem of choosing the correct sampling strategy
may become relevant, because a complete description of the feature would require
an unaffordable measurement time.
4 Coordinate Measuring Machine Measurement Planning 121
Figure 4.4 Effects of measurement error and sample size on the estimated geometric error (the
real geometric error is 2): a no probing error, five points, estimated geometric error 1.523, b with
probing error, five points, estimated geometric error 1.6396, c no probing errors, 40 points,
estimated geometric error 2, and d with probing error, 40 points, estimated geometric error
2.6303
Moreover, it should be pointed out that often with typical coordinate measuring
systems the definition of the sampling strategy is not completely free. Consider,
for example, measurement by means of a structured light scanner. The number and
the pattern of the sampling points will depend on the relative position of the struc-
tured light projector, camera(s), and the object to be measured, and the sampling
points will be measured more or less simultaneously. If more sampling points
were to be measured, one would have to modify this setup, but this may imply –
among other things – an increase of probing error or difficulties in registering
clouds of points. Measurements could be repeated within the same setup to aver-
age the results. Averaging will lead to a reduction of probing error, and improve
accuracy; but if the overall coverage of the sampling strategy does not change, that
is, points are located in the same positions in every measurement repetition, it is
still possible to miss critical areas. Therefore, the sampling strategy may tend to
122 G. Moroni and S. Petrò
the situation depicted in Figure 4.4a: some very accurate points, missing the areas
of maximum geometric deviation, thus introducing a measurement bias (a priori
impossible to evaluate). However, when adopting these measuring systems, the
sample size obtained in a single acquisition is often enough to describe the whole
surface with a sufficient point density. Differently, if a profiler or a scanning
CMM (Weckenmann et al. 2004) is adopted as the measuring instrument, the
sampling points will be grouped in profiles, even when the inspected feature is a
surface. The measurement of a single point will be quite fast, but covering the
whole surface with the same density of points along profiles and in any other di-
rection may be very time-consuming. Finally, if a CMM is adopted which is capa-
ble of point-to-point measuring, significant freedom is granted to define the sam-
pling strategy; unfortunately, these measuring systems are usually the slowest.
Therefore, the problem of choosing the correct sampling strategy is most critical in
this situation.
Summarizing, with fast surface scanners there should be no significant prob-
lems in planning the sampling strategy, and measurement uncertainty is mainly
influenced by probing error; with profile scanners or CMMs equipped with scan-
ning probing systems, even though it is possible to sample a large number of
points, the number and the placement of profiles to be acquired must be carefully
selected, in order to ensure a good coverage of the surface; finally, with point-to-
point measuring systems, densely measuring the whole feature will often require
an unaffordable measuring time. Because the last problem is related to CMMs and
is the most critical situation for sampling strategy planning, it will be considered
as the reference situation in the chapter.
The problem of choosing the correct sampling strategy may be split into two
subproblems: sample size choice and sampling pattern choice.
The first problem is economic: measurement accuracy and measurement cost
are, most of the time, directly related to sample size, so a trade-off between accu-
racy and cost has to be identified. In the last part of this chapter the problem of
economically evaluating accuracy will be addressed, and an economic model for
inspecting cost will be proposed that, when optimized, defines the optimal sample
size.
The choice of the sampling pattern is less straightforward. The effectiveness of
a sampling strategy has been defined as its ability in identifying the most deviating
point. Provided one is completely free to choose the sampling strategy, in most
situations the only information available about the measurement to be made per-
tains to nominal geometry, tolerances, and fixtures. A sampling strategy based on
this information alone is defined as a “blind sampling strategy”, because it does
not consider any information on the actual geometric error of the feature. How-
ever, it is possible to develop a sampling strategy based on the actual geometric
deviation of the part. If the sampling strategy is based on information on the geo-
metric deviation acquired during measurement, that is, a set of points is sampled,
then depending on such points, one or more additional points are sampled accord-
ing to some criterion, and so on, until some terminating condition is reached, the
strategy is an “adaptive sampling strategy”, which tries to “adapt” to the actual
4 Coordinate Measuring Machine Measurement Planning 123
geometric error of the inspected part. Finally, if the information available on the
geometric deviation relies on a preliminary study of the manufacturing process,
and the measurement strategy is planned on the basis of this information, then this
is a “process-based” sampling strategy. Consistency of process-based strategies is
guaranteed by the fact that geometric deviations generated by a particular manu-
facturing process tend to be similar in each manufactured part.
Because the choice of the sample size is strongly affected by the criterion cho-
sen for planning the pattern, blind, adaptive, and process-based strategies will be
introduced before addressing the problem of choosing the sample size.
Several kinds of sampling pattern definitions have been proposed in the literature.
They may be categorized according to the following classification:
1. blind sampling strategies;
2. adaptive sampling strategies; and
3. process-based sampling strategies.
Moreover, sampling patterns for blind sampling strategies are easy to define.
Often, blind sampling strategies are completely defined by the sample size only,
that is, given the sample size, the sampling point pattern is completely defined, or
at most depends only on random parameters (e.g., if the distribution of sampling
points is random, then having chosen the sample size to be n, there will be n ran-
domly spread sampling points throughout the surface). Blind sampling strategies
are therefore easy to implement in CMM control systems, and often strategies
based on such a pattern, like a uniform strategy, may be automatically generated
by CMM software.
The most common blind sampling strategies, including those proposed in the
international standards, are now briefly illustrated.
Figure 4.5 Uniform sampling strategies for a a plane, b a cylinder, c a circumference, and
d a straight profile
the same sample (a phenomenon known as “aliasing”, see Figure 4.6). Real fea-
tures are seldom entirely defined by a finite harmonic content; however, in most
situations, the first few harmonics of higher amplitude will be sufficient to recon-
struct the geometry to a sufficient degree of approximation. If a threshold has been
defined beyond which the amplitude of the harmonics is negligible, this threshold
can be considered for application of Nyquist’s criterion when choosing the density
of the sampling points.
Figure 4.6 Effects of undersampling (real frequency 16, sampling frequency 10)
However, when dealing with surfaces, applying a uniform strategy which re-
spects Nyquist’s criterion (even after having chosen a proper threshold) tends to
lead to huge sample sizes; therefore, a uniform sampling strategy may be afford-
able only for fast measuring systems. On the other hand, the use of a uniform
strategy that does not respect Nyquist’s criterion may again lead to aliasing-related
problems. For example, if the surface is characterized by a dominant wave and the
sampling point frequency is not adequate, sampling points may end up being lo-
cated on wave crests, without any point in valleys, thus resulting in an incorrect
assessment of the underlying geometry.
If a uniform strategy respecting Nyquist’s criterion is unaffordable owing to the
excessive sample size required, a few strategies have been proposed that are capa-
ble of avoiding problems related to the adoption of a uniform grid of points. One
is the random strategy (Dowling et al. 1997). In a random strategy a given number
of sampling points are randomly scattered throughout the feature to be inspected.
This avoids the generation of any pattern, so no bad interaction should be created
between the geometry of the real feature (ISO 14660-1 1999) and the sampling
126 G. Moroni and S. Petrò
Some strategies which avoid defects of regular patterns without placing points
in random locations come from the Monte Carlo method and are based on the so-
called quasi-random sequences (Hammersley and Handscomb 1964). Probably,
the most famous strategy of this kind (for surfaces) is the Hammersley strategy.
The Hammersley sequence sampling strategy is mathematically defined as fol-
lows. If n points have to be sampled, then define
k −1
i
ui = vi = ∑ bij 2− j −1 , i ∈ {0,1, …, n − 1} , (4.3)
n j =0
where ui and vi are normalized coordinates of the ith point (i.e., they are defined in
the [0,1] interval and must be rescaled to retrieve the actual surface coordinates),
bi is the binary representation of i, and bij is the jth bit of bi (so bij∈{0,1}), and
k = ⎡log2n⎤, where ⎡x⎤ is the smallest integer greater than or equal to x. As high-
lighted in Equation 4.3, once the sample size n is defined, the Hammersley strat-
egy is univocally defined. However, as Figure 4.8 shows, points are spread
throughout the whole surface without any apparent pattern, and this should avoid
any interaction with any harmonic content in the surface, or any other systematic
behavior. Lee et al. (1997) have shown that, given the sample size, the Hammer-
sley sequence based sampling strategies outperforms both the uniform and the
random sampling strategies for a plane, cylinder, cone, and dome (sphere). The
4 Coordinate Measuring Machine Measurement Planning 127
authors claim that the Hammersley sequence has a nearly quadratic reduction in
the number of points needed by a uniform sampling for the same level of accu-
racy.
Similar to the Hammersley sampling strategy is the Halton–Zeremba strategy
(Figure 4.8), which is defined only if the sample size is a power of 2. The mathe-
matical definition is as follows:
k −1 k −1
ui = ∑ bij 2 − k − j , vi = ∑ bij′ 2− j −1 , i ∈ {0,1, …, n − 1} , (4.4)
j =0 j =0
where b′ij is equal to 1-bij if j is odd, otherwise b′ij is equal to bij if j is even. Kim
and Raman (2000) claim the Halton–Zeremba strategy not only outperforms the
uniform and random strategies, but also outperforms the Hammersley strategy for
flatness.
nologies. Profile strategies may be adopted even with systems which measure
single points, simply by grouping single sampling points along profiles. Several
patterns for profiles are proposed in standards for the plane and the cylinder, and
are illustrated in Figures 4.9 and 4.10.
In order to choose how many points have to be sampled in each profile, the
same criterion proposed for roundness and straightness is adopted (i.e., a uniform
Figure 4.9 Profile extraction strategies proposed in ISO technical specifications for flatness:
a rectangular grid, b triangular grid, c polar grid, d Union Jack, and e parallel profiles
Figure 4.10 Profile extraction strategies in ISO technical specifications for cylindricity: a bird
cage, b roundness profiles, and c generatrix
4 Coordinate Measuring Machine Measurement Planning 129
the inspection, and the part to be measured. An adaptive sampling strategy starts
from a few sampling points scattered throughout the surface to be inspected; then,
depending on the information acquired through such points, more sampling points
are added. Point addition may go on until some criterion is met (e.g., the computed
geometric deviation does not vary significantly anymore, or some evaluation of
the lack of information in unsampled zones is sufficiently small, or a given maxi-
mum sample size has been reached).
Adaptive sampling is therefore a technique which aims to find critical areas of
the surface to be inspected with a reduced sample size (if compared with blind
sampling strategies). Reduction is made possible by not taking all points at one
time, but by varying the sampling pattern according to the information that may
be extracted from sampling points as they are sampled. This kind of strategy
may be effective for slow measuring systems, which require every effort to re-
duce the sample size to control the measurement cost. Another advantage of an
adaptive sampling strategy is that no planning of the pattern is required, and in
some implementations not even the sample size is required – the operator
chooses only the target accuracy for the measurement task and then the measur-
ing system will automatically choose the measuring strategy. Adaptive sampling
may then be considered as a fully flexible automatic way of planning the meas-
urement strategy.
However, flexibility is the Achilles’ heel of adaptive sampling; in fact, even
though some methods for the automatic planning of the probe path have been
proposed (Moroni et al. 2002), current CMM software applications are not able to
automatically define a path for the probe which is guaranteed to be collision-free,
owing to the presence of fixtures and other features of the part itself. Therefore,
for commercial instruments, the adoption of fixed patterns for sampling points is
justified by the need to avoid damage risks for the measuring system, risks that
may be solved once and for all for static sampling patterns by correctly defining
just one sequence of CMM movements. Adaptive sampling strategies may there-
fore be considered as the most promising strategies for the future; however, the
integration of automatic probe path planning methods in current CMM control
systems is required to allow adaptive sampling. If collision detection problems are
eventually solved, adaptive strategies could be able to combine the advantages of
both blind (suitability for inspection of a few parts and ease of definition) and
manufacturing-signature-based (reduced uncertainty, given the sample size)
strategies; however, more research is required to make adaptive sampling feasible.
An adaptive sampling strategy is defined essentially by three elements: a start-
ing (blind) sample set, a criterion aimed at choosing the next sampling point(s),
and a stopping rule.
The starting sample set is usually generated by means of a simple uniform
sampling strategy, and it is characterized by a few points (at most ten or 20). Only
in Badar et al. (2005) is it suggested that the starting set pattern should be based
on the manufacturing process, the pattern itself being chosen empirically, by op-
erator choice.
4 Coordinate Measuring Machine Measurement Planning 131
Since when verifying compliance to tolerances, only points showing the maxi-
mum local deviation are relevant, the criterion for choosing the next sampling
point usually looks for such points. In particular, three ways have been proposed:
1. Sampling point deviations from the nominal geometry are fitted by means of a
spline (Edgeworth and Wilhelm 1999). Then, this spline is adopted to predict
deviations in those areas of the feature which have not been sampled yet. The
area (or areas) that shows the largest deviation is chosen for sampling in the
next step of the algorithm. However, it should be pointed out that a spline fit-
ting a set of unevenly spaced sampling points may tend to introduce unwanted
undulation, which could lead to a wrong choice of the location of the next
sampling point.
2. Focusing on roundness, Rossi (2001) proposed that lobes on the roundness
profile are identified from the starting set. Then, if lobing is significant, sam-
pling points are concentrated in crests and valleys of the profile; otherwise, a
uniform strategy is adopted. Because lobing is often found in roundness pro-
files, the criterion is efficient for finding maximum and minimum geometric
deviations. Some limitations of this technique include its scarce adaptability
to tolerances different from roundness; furthermore, if the initial sample size
is small, some undulations may not be detected;
3. Direct search techniques are adopted (Badar et al. 2003, 2005). Most adaptive
sampling techniques involve fitting the surface in order to choose the next
sampling point(s) as the most deviating one. They may therefore be assimi-
lated to numerical optimization techniques, aimed at identifying maxima and
minima of a function, which is nothing other than the actual deviation of the
geometric function as a function of the sampling coordinates. Differently
from the previously indicated techniques, direct search optimization tech-
niques do not require any surface fitting in order to find maxima (i.e., choose
the next point or points). Owing to the nature of most inspected surfaces,
which often have sharp variations or noise present, this may be an interesting
feature.
4. Barbato et al. (2008), adopting kriging interpolation (Cressie 1993) to fit
sampled points, proposed a different criterion: because kriging interpolation
allows for an evaluation of the sample variance of fits, the location character-
ized by the maximum variance is chosen. This approach is intended to ensure
that the amount of information pertaining to the behavior of the surface is as
uniform as possible in every area of the surface itself.
A few stopping criteria have been proposed. The easiest one is to stop when, af-
ter performing some iterations of the algorithm, a fixed number of sampling points
is reached, leading to strategies characterized by a constant sample size. A more
interesting criterion was proposed by Edgeworth and Wilhelm (1999), which evalu-
ates the uncertainty for each iteration (i.e., points addition), and stops when the
uncertainty, which tends to decrease as the sample size grows, is sufficiently low.
132 G. Moroni and S. Petrò
Regardless of the method adopted, the research works mentioned above show
that adaptive sampling is capable of reducing the number of sampling points re-
quired to achieve a given uncertainty by at least of an order of magnitude with
respect to blind strategies. This result is really encouraging; however, the problem
of online path planning has to be solved before adaptive sampling can be widely
adopted.
Figure 4.11 Signature examples: a shaft turned between centers, and b face-turned plane
4 Coordinate Measuring Machine Measurement Planning 133
adopted for this preliminary data acquisition will consist of a blind strategy char-
acterized by a point density as high as possible, to ensure that the amount of in-
formation collected is adequate. Most algorithms for defining strategies based on
process raw data require the initial sampling strategy to be the same throughout
the preliminary set. Because of these requirements, a dense uniform sampling
strategy is adopted most of the time. In most cases, the same measuring instrument
is selected to perform the preliminary measurements and for successive inspection.
This choice ensures that the accuracy of the measurement system is implicitly
considered in the definition of the sampling strategy; if different measuring sys-
tems are considered, different performances should be taken into account in the
definition of the strategy.
Once all the required information is available, the method for defining the sam-
pling strategy is directly applied to it, without any complex preliminary elabora-
tion. The methods for setting up a measurement strategy based on raw data usually
do not require significant human intervention, apart from data collection, thus
making this kind of sampling strategy easy to plan, even for inexperienced opera-
tors. These strategies provide the typical accuracy improvements which are to be
expected from a process-based strategy, while only adding the cost of acquiring
raw data.
As previously pointed out, the performance of process-based sampling strate-
gies is sensitive to manufacturing process instability; therefore, the success of such
techniques is strongly dependent on successful monitoring of the process itself.
However, the most efficient process control techniques available from the litera-
ture are all based on the availability of a model of the manufacturing process sig-
nature (Colosimo et al. 2008b, 2010), and therefore cannot be applied to this case,
where a signature model is not available. If such a model were available, sampling
strategies based on the manufacturing signature model should be adopted instead.
A few strategies based on manufacturing process raw data have been proposed
in the literature for the evaluation of form error For several geometric features
(e.g., plane and straight line) the points that deviate the most from the nominal
geometry always belong to the convex hull of the point cloud; given this, some
strategies have been proposed that choose sampling points among those belonging
to the convex hull of raw data. In Raghunandan and Rao (2007), it is proposed that
sampling points are randomly chosen from these convex hull points; in order to
ensure efficiency, it is also suggested that the strategy is tested on a few more
parts before it is adopted. Buonadonna et al. (2007) proposed a hybrid approach
where point selection is based on using the information provided by a signature
model (hence the hybrid nature of the approach), with the application of the D-
method (Montgomery 2004) (i.e., minimizing the variance of the model parame-
ters), and combining such information with the extraction of some more sampling
points from the convex hull. This further addition is driven by point ranking,
which in turn is based on the distance from the fitted model. The presence of a
signature model in the method could suggest it should be listed as a manufactur-
ing-signature-based approach; however, the authors suggest one “avoid a complete
signature description” to enhance performance; therefore, not requiring a full sig-
136 G. Moroni and S. Petrò
If a model of the manufacturing signature is available for the actual process, then a
sampling strategy based on this model may be proposed. In other words, these are
sampling strategies that take as input the nominal geometry and tolerances, metro-
logical characteristics of the measuring system, and – in addition – a model of the
manufacturing signature. On the basis of this information, strategies are proposed
whose pattern is defined once and for all, optimized for the manufacturing signa-
ture. The model required by these methods may be either experimentally deter-
mined or generated through the application of some predicting technique that links
the manufacturing signature to manufacturing process parameters. If the model is
derived experimentally, the same considerations that were made for raw-data-
4 Coordinate Measuring Machine Measurement Planning 137
based strategies apply here as well for the selection of sampling pattern and the
size of the preliminary set. Since the manufacturing signature model usually does
not describe the measurement error of the measuring system adopted, the measur-
ing instrument adopted to collect data from which to derive the signature model
does not necessarily have to be the same as that adopted for inspection, and there-
fore a high-accuracy measuring system should be adopted in order to generate
models that are as accurate as possible. However, regardless of the origin of the
signature model, the metrological characteristics of the measuring system adopted
for successive inspection should be taken into account when dealing with the defi-
nition of the sampling pattern.
Most of the methods for planning signature-based strategies require particular
kinds of signature models. As an example, the extended zone criterion, which will
be introduced later, requires an ordinary least-squares regression model to describe
the signature. In general, the choice of the applicable planning method is restricted
to those methods suitable for the signature model available, or, if the method has
been chosen, the correct kind of model has to be selected. Anyway, the signature
models adopted will be mostly of statistical nature because they have to be able to
capture the variability of the manufacturing signature. Moreover, models such as
pure time series models are not suitable for modeling the signature because the
description of a feature, even when considering correlation, does not consider
systematic behavior. Regression models (Draper and Smith 1998) are the most
common choice.
The identification of a manufacturing signature model requires greater effort
than does the simple acquisition of raw data. However, one should remember that
knowledge of the manufacturing signature leads to several advantages (Dowling
et al. 1997; Colosimo et al. 2008b), in addition to allowing for optimization of the
inspection strategy.
Although the problem of modeling the signature will not be directly addressed
in this work, it is important to recall the three parts that make a signature model:
1. a structure, which defines the kind of behavior geometric deviations show (e.g.,
roundness profiles are usually characterized by the presence of lobes, flat sur-
faces often show a polynomial surface, etc.);
2. an average amplitude, which defines, on average, the geometric deviation
induced by the structure; and
3. a random noise quantification, which characterizes residuals not explained by
the rest of the model.
has the drawback of considering only the structure of the signature, and not its
average amplitude or random noise.
To overcome the limitations of the extended zone method, one of the authors of
this work (Petrò 2008) proposed developing a sampling strategy based on toler-
ance intervals for a regression model describing the signature. Regression inter-
vals, given the coordinates, define the upper and the lower bounds within which a
given fraction of future observations at those coordinates will fall. The amplitude
of the tolerance intervals depends on the concentration of the sampling points
around the coordinates, on residual dispersion of random noise, and on the model
structure, while the values of the upper and lower bounds of the interval are influ-
enced also by the average amplitude at the coordinates considered. Therefore, it is
proposed to choose the pattern of sampling points by minimizing the difference
between the maximum value of the upper bound, and the minimum value of the
lower bound. This approach has proven capable of concentrating the sampling
points in those areas of the feature that typically deviate the most from the nomi-
nal behavior, taking into account every part of the signature.
Similarly to raw-data-based approaches, if a technique for evaluating the uncer-
tainty associated with a manufacturing signature model is available, then it is pos-
sible to directly optimize the sampling strategy by minimizing such uncertainty.
This method will be further discussed later.
actual geometry, which solves the problem of the model choice. For kriging re-
construction, Yang and Jackman suggested a random sampling strategy, and for
Shannon reconstruction, a uniform sampling strategy was chosen.
Some signature-based strategies have been proposed that make use of recon-
struction techniques. Summerhayes et al. (2002) adopted the extended zone crite-
rion, reconstructing the signature by means of ordinary least-squares regression.
Moroni and Pacella (2008) proposed a reconstruction approach that is very similar
to Yang and Jackman’s (2002), as both of them adopt Shannon reconstruction in
order to fit the model. The two approaches differ in the greater attention Moroni
and Pacella gave to the harmonic content in the feature: Yang and Jackman
adopted Shannon reconstruction “blindly”, without considering the real behavior
of the feature; Moroni and Pacella analyzed the influence of the real behavior of
the feature on the effectiveness of reconstruction.
The first case study considers face-milled planes, as defined in the ISO 10791-7
(1998) standard. A series of nine 160 mm × 160 mm planar surfaces have been
machined on an MCM Synthesis machining center, as illustrated by Figure 4.12.
The material chosen was a 6082-T6 aluminum anticordal alloy. The cutting pa-
rameters were as follows:
• spindle speed 3,000 rpm;
• feed rate per tooth 0.12 mm;
• depth of cut 0.5 mm;
• diameter of the mill 100 mm; and
• number of teeth 7.
140 G. Moroni and S. Petrò
Figure 4.12 Milling path superimposition in face milling. Tool path segments are supposed to
be horizontal
Machined surfaces were then measured with a Zeiss Prismo Vast HTG CMM,
adopting a uniform strategy. A square 157×157 grid of points (a point per millime-
ter, leaving 2 mm of clearance from the edges) was sampled on each workpiece.
Figure 4.13 shows the mean surface obtained by averaging the z coordinates of
the nine acquired grids, aligned in the xy plane. The mean surface gives an idea of
the manufacturing signature: the surface is characterized by an overall smooth
saddle geometry, a region with steeper variations and an abrupt discontinuity
which is imputable to the superimposition of the two milling path segments (as
illustrated in Figure 4.12).
To obtain the raw data needed to investigate this family of strategy planning
approaches, a larger number of points is needed: to generate such additional data,
a simulation-based solution is introduced. Of course, for real applications of the
method, such points would be better obtained experimentally.
The simulation solution is based on generating additional points affected by
random error. To assess the random error, the geometric deviation evaluated on
24,649 sampling points belonging to the experimental measurement of a single
surface specimen has been taken as the reference value, with 1.5 μm standard
uncertainty, as defined in ISO/IEC Guide 99:2007(E/F) (2007). Therefore, to gen-
erate a cloud of points constituting raw data, a random error, uniformly distributed
in the ±2-μm interval, was added to the measured cloud of points. Ten clouds of
points for each surface were simulated this way. The result of this simulation con-
stitutes the raw data feeding the strategy planning method.
4 Coordinate Measuring Machine Measurement Planning 141
where k is the expansion factor. The ISO 15530-3 specification allows one to
compensate for the b term, thus reducing its influence; anyway, in the following
discussion it will be supposed not to be compensated.
Mathematical expressions found in the technical specification allow for uncer-
tainty evaluation if a single calibrated artifact is adopted in the procedure. How-
ever, a single artifact is not sufficient to describe the manufacturing signature
142 G. Moroni and S. Petrò
∑∑ ( y
j =1 i =1
i, j − xcal, j )
b= . (4.6)
mrm
In Equation 4.6 m is the number of calibrated artifacts adopted, rm is the measure-
ment repetitions number for each artifact, yi,j is the measurement result (estimated
geometric error) of the ith measurement repetition of the jth artifact, and xcal,j is the
reference value for the jth artifact (calibrated geometric error). It is supposed that
each calibrated workpiece is measured the same number of times; to be as similar
to the ISO 15330-3 standard as possible, it rm ≥ 10 is suggested.
Then, to estimate up, a pooled standard deviation could be used:
m rm
∑∑ ( y )
2
i, j − yj rm
yi , j
yj = ∑ (4.7)
j =1 i =1
up = , .
m ( rm − 1) i =1 rm
Substituting Equations 4.6 and 4.7 into Equation 4.5, the required evaluation of
U results.
A final note on uW: ISO 15530-3 introduces it to take into account the “variabil-
ity of the production”, that is, part-to-part differences in local form deviations;
however, if more than one calibrated workpiece is measured, then up should con-
tain this uncertainty contribution, and then uW = 0, as the international standard
itself suggests.
Having identified a method to evaluate uncertainty based on raw data, the next
step is to choose a sampling strategy that minimizes U (given the sample size n).
Suppose that the raw data consist of rm dense measurements of m calibrated parts,
and that the sampling strategy is the same for every measurement. Then, sampling
points corresponding to any sampling strategy may be extracted from these clouds
of points. Extracted subsets of points can be introduced in the measurement uncer-
tainty estimation procedure. The uncertainty evaluation obtained is, of course, in-
fluenced by the interaction between the sampling pattern considered and any typi-
cal geometric error left by the manufacturing process. If the sampling pattern is
effective, i.e., it is able to detect regions of the feature that deviate the most, then
the uncertainty will be low. The identification of an optimal pattern can be seen as
an optimization problem where several different alternative patterns are compared.
For the search of the optimal strategy, an optimization algorithm may be suggested.
Given the discrete nature of the problem, genetic algorithms (Holland 1992) and
simulated annealing algorithms (Kirkpatrick et al 1983) are suitable for the task.
4 Coordinate Measuring Machine Measurement Planning 143
Figure 4.14 Strategies considered in the comparison: a uniform sampling, n = 16, b Hammers-
ley sampling, n = 16, c optimized sampling, n = 16, d uniform sampling, n = 100, e Hammersley
sampling, n = 100, and f optimized sampling, n = 100
144 G. Moroni and S. Petrò
they may be relevant if some anomaly happens in the manufacturing process, thus
providing better robustness for the strategy.
Now, consider Figure 4.15, which shows the expanded uncertainty U for the
uniform, the Hammersley, and the proposed strategy (with a coverage factor
k = 2). First of all, it can be noted that, in contrast with what Lee et al. (1997)
stated, the Hammersley strategy does not outperform the uniform one in terms of
expanded uncertainty. Moreover, the uncertainty does not monotonically decrease
with increase of the sample size. This is an example of negative interaction be-
tween the part form error and the sampling strategy: since extreme points are lo-
cated in the same regions of the surface throughout the whole set of workpieces, if
the (blind) sampling strategy adopted does not place any points in these regions,
the uncertainty will be high, regardless of the sample size. On the other hand,
being optimized for the particular manufacturing process being considered, the
raw-data-based sampling strategy shows a significantly lower uncertainty. More-
over, for a sample size larger than 100 points, the uncertainty is nearly constant
and mainly depends on the workpiece calibration uncertainty.
The different influences of bias and measurement procedure repeatability on the
overall uncertainty are depicted in Figures 4.16 and 4.17. As is evident, the raw-data-
based sampling strategy outperforms the uniform and Hammersley strategies in that
it is both less biased and less uncertain. As highlighted in Figure 4.16, the bias term
for the Hammersley and uniform strategies is about 5 times larger than the meas-
urement procedure repeatability. This proves that blind sampling strategies, if the
sample size is not large enough, tend to fail to detect extreme points, thus underesti-
mating (biasing) the geometric deviation,. Differently, an optimized sampling strat-
egy, if the sample size is large enough, can make the bias negligible. Finally, it can be
pointed out that Figures 4.16 and 4.17 show really similar plots, as long as the sign is
25
Uniform
Hammersley
20 Minimum U
15
U [μm]
10
0
0 100 200 300
Number of Sampling Points
Figure 4.15 Expanded uncertainty for the case study considered and uniform, Hammersley, and
minimum U strategies
4 Coordinate Measuring Machine Measurement Planning 145
−5
Bias [μm]
−10
−15 Uniform
Hammersley
Minimum U
−20
0 100 200 300
Number of Sampling Points
Figure 4.16 Bias for the case study considered and uniform, Hammersley, and minimum U
strategies
1.2
1
Repeatability [μm]
0.8
0.6
0.4
Uniform
0.2 Hammersley
Minimum U
0
0 100 200 300
Number of Sampling Points
Figure 4.17 Measurement procedure uncertainty for the case study considered and uniform,
Hammersley, and minimum U strategies
146 G. Moroni and S. Petrò
Usually, the standard strategy for profile measurement consists in uniform sam-
pling. Consistently, roundness profiles are usually sampled at evenly spaced
points. However, if a manufacturing process signature model is available, different
strategies may be considered.
In the following, a signature model for turned profiles will be considered, and a
signature based strategy will be proposed.
Along the lines of the raw data based sampling strategy proposed for a flat surface,
the main idea here is to minimize the measurement uncertainty, once the sample
size has been chosen. Uncertainty evaluation can be performed with the support of
a signature model.
A similar method was proposed by the authors of the present work (Moroni and
Petrò 2009), and is based on the “virtual CMM” concept. In particular, this strat-
egy planning method is based on the idea of a virtual CMM which takes into ac-
count the presence of the manufacturing signature and its interaction with the
actual sampling strategy when evaluating the measurement uncertainty.
A classic virtual CMM (Wilhelm et al. 2001; Balsamo et al. 1999) is based on
the simulation of ideal, but not necessarily geometric-error-free, geometric features
for which the geometric error x is known; a sampling error is simulated according
to a model of the real behavior of the CMM for which the uncertainty is being
evaluated, and is added to the ideal feature (which is referred to as “feature pertur-
bation”). Measurement uncertainty is evaluated by comparing geometric errors
evaluated on the perturbed features and the known real geometric errors of ideal
features. The overall method may be regarded as a Monte Carlo simulation whose
aim is to obtain values for the real geometric error x and the measured geometric
error y, which are then subtracted to yield the measurement error x-y. A Monte
Carlo simulation based virtual CMM able to take into account signature–sampling
strategy interaction is not difficult to obtain: it is sufficient to simulate ideal fea-
tures according to some signature model. If the simulated ideal features are gener-
ated according to some real signature model instead of “perfect features”, then the
uncertainty evaluation will implicitly consider the presence of the signature.
Various error sources should be considered in the simulation, including meas-
urement strategy, environmental conditions, and CMM volumetric errors; a com-
plete list of these sources may be found in the ISO/TS 15530-4 (2008) technical
specification. This recently published standard deals with the problem of validat-
ing virtual CMM models, proposing four validation methods.
Several methods have been proposed to extrapolate uncertainty from simulation
results. Here, the approach proposed by Schwenke et al. (2000) was adopted. This
approach does not allow one to explicitly calculate a standard uncertainty u , but
only an expanded uncertainty U characterized by some coverage probability p
4 Coordinate Measuring Machine Measurement Planning 147
G (U ) − G ( −U ) = p . (4.8)
Note that the resulting evaluation of U is coherent with the definition of the
coverage probability, and that any uncorrected bias is considered (Schwenke et al.
2000).
To simulate geometric errors, an approach based on results obtained by van
Dorp et al. (2001) was considered. The model developed is based on the frequency
content of the error signal, and may be applied to a Zeiss Prismo CMM.
ε ~ N 748 ( 0, σ 2 I ) ,
which, for a roundness profile, describes the deviation r(θi) from the average ra-
dius as a function of the angle θi; in the model, βj are coefficients, W(i) is the
neighborhood matrix of the ith order, ρi is the ith spatial correlation coefficient,
and σ 2 is the residuals variance.
The results obtained by applying a virtual CMM to estimate the measurement
uncertainty and adopting a simulated annealing algorithm to minimize this esti-
mated uncertainty by considering different sampling point patterns are similar to
those proposed for the planar surface sampling strategy. The comparison is out-
lined in Figure 4.18, which shows the decrease of the expanded uncertainty as the
sample size increases. Even though the uncertainty decreased for both the uniform
and the optimized sampling strategy, it is evident that the uncertainty is consis-
tently smaller for the optimized strategy.
Finally, one could think that because the same sets of 1,000 simulated profiles
and 1,000 perturbed profiles was adopted to optimize the sampling strategy, the
sampling strategy itself may be effective only for these specific simulated profiles.
In order to prove the general effectiveness of the strategy, a new, independent
148 G. Moroni and S. Petrò
14
Uniform
12 Minimum U
10
U [μm]
2
0 20 40 60 80 100
Number of Sampling Points
Figure 4.18 Expanded uncertainty for the case study considered and uniform and minimum U
strategies
virtual CMM uncertainty evaluation was performed for every strategy proposed.
The results are not distinguishable from those illustrated here, thus proving the
general validity of the result.
The case studies show that process-based sampling strategies may outperform
traditional, blind sampling strategies. Of course, if there is no need to reduce the
sample size, and if the measurement uncertainty that can be obtained with a blind
sampling strategy is adequate, there is no need to adopt a process-based sampling
strategy. However, when the number of parts to be inspected is very large, or the
tolerance is critical for the functional characteristics of the part, care should be
taken to choose the correct sampling pattern.
The criteria for planning a sampling strategy described so far allow one to define a
sampling pattern having defined the sample size. However, the problem of choos-
ing how many points to sample has not been addressed yet. This problem is known
as sample size definition.
As may be gathered from Figures 4.15 and 4.18, as the sample size increases,
the measurement uncertainty U tends to decrease. This progressive reduction of
the measurement uncertainty may be explained by considering the more complete
description of the measured feature that can be achieved with a larger sample size.
Sometimes (e.g., see Figure 4.15), when the sample size is small, the interaction
between the manufacturing form error and the blind sampling strategy may make
the decrease nonmonotonous; however, for large sample sizes, the decreasing
behavior is restored. Moreover, even though the uncertainty decreases as the sam-
4 Coordinate Measuring Machine Measurement Planning 149
ple size increases regardless of the sampling strategy chosen (blind, signature-
based etc.), the connection between the sample size and the uncertainty depends
on the criterion adopted for pattern generation. For example, given the sample
size, process-based sampling strategies may show a lower uncertainty with respect
to blind sampling strategies.
Measurement uncertainty definition, and then sample size definition, is a prob-
lem of agreement between the customer and the manufacturer. This, in turn, is
related to economic aspects of measurement, and, more generally, of inspection.
Low uncertainty measurement procedures are usually more expensive. This is
evident for coordinate metrology. It has already been pointed out that uncertainty
is inversely proportional to sample size: for coordinate measuring systems which
do not sample points simultaneously, such as 3D scanners, but sequentially (point-
to-point such as a touch-trigger CMM or in a set of profiles such as profilers), a
larger sample size often implies a longer measurement time, and a consequently
higher measurement cost. However, uncertainty quantifies the dispersion of the
values which may be attributed to a measurand (ISO/IEC Guide 99:2007(E/F)
2007), so larger uncertainty usually implies more inspection errors, and excessive
inspection errors may lead to unexpected costs. When dealing with inspection of
any product, two kinds of inspection errors are possible (Burdick et al. 2003):
rejecting a conforming product (“false failure”) and accepting a nonconforming
error (“missed faults”). Both errors may generate an “inspection error cost” for the
manufacturer and the customer. False failure generates a cost because a part that
could be sold is discarded, or has to be reworked. Missed faults cost is usually
indirect: the customer may reject a batch because of an excessive fraction of non-
conforming parts in the batch itself or, if the part has to be assembled, it can make
the final product defective. Of course, the probabilities of both false failure and
missed faults increase with measurement uncertainty, so the inspection error cost
is directly proportional to uncertainty. Measurement uncertainty definition is
therefore the search for the optimal trade-off between the measurement cost and
the inspection error cost, that is, minimization of the inspection cost.
The ISO/TS 14253-2 (1999) and ISO/TS 14253-3 (2002) technical specifica-
tions deal with the problem of finding an agreement between the customer and the
manufacturer. After an agreement has been reached, the “Procedure for Uncer-
tainty Management” (PUMA) method proposed in ISO/TS 14253-2 may be
adopted to evaluate the uncertainty and define the overall measurement procedure,
including the choice of the measuring system, part fixturing, identification of un-
certainty sources, and so on (for further details on the PUMA method see the
ISO/TS 14253-2 technical specification). Since the sample size definition is part
of the measurement procedure definition, it is implicitly defined during the appli-
cation of the PUMA method.
Anyway, when choosing the sample size, complete freedom is normally guar-
anteed only by point-to-point measuring systems, for example, a CMM equipped
with a digital probe. Unfortunately, this kind of measuring system is the slowest,
and thus the one with the higher measurement cost (even though it is usually accu-
rate, thus reducing the inspection error cost). Another motivation for freedom
150 G. Moroni and S. Petrò
reduction when defining the sample size is the presence of a time limit. If every
manufactured part has to be inspected, the measurement time will have to be ade-
quate with respect to the cycle time. Because an increase in the sample size causes
an increase in the measurement time, a time limit creates an upper bound to the
sample size. In this situation it may be important to adopt a nonblind strategy in
order to reduce measurement uncertainty.
In the literature the problem of sample size definition has been addressed
mainly for blind sampling strategies. A technique to estimate the expected evalua-
tion error (essentially systematic measurement error) for the straightness error
evaluation has been proposed: Namboothiri and Shunmugam (1999) proposed a
method that is capable of choosing the right sample size once an upper bound for
the measurement error has been defined. Some objections may be suggested: the
criterion considers only systematic errors, but random error in measurements are
at least as important as systematic ones, such as usual uncertainty evaluations
show; the sampling strategies analyzed were limited to random sampling; and an
arbitrary choice of the error level may be suitable when calibrating a single arti-
fact, but the choice of the sample size for quality check of mass production should
be based on agreement between the manufacturer and the customer. Anyway, an
interesting consideration found in this article is:
“Furthermore, this study clearly underlines the importance of the sampling pattern to be
followed on the surface during form error measurement. The guiding principle should be
to catch that point that has got maximum error. If we get this point at the initial stages of
measurement then further prolonging the measurement process is not necessary. Hence
the measurement time can be considerably reduced.”
In this section, a cost function is proposed which is aimed at supporting the deci-
sion procedure for finding an agreement between the manufacturer and the cus-
tomer on the measurement uncertainty, which, in turn, depends on the sample size.
The cost function provides a measure of the overall “inspection cost”, i.e., the sum
of the measurement cost and the inspection error cost.
The general form of the inspection cost CI is
C I = CM + C E , (4.10)
where CM is the measurement cost and CE is the inspection error cost. The cost
function will therefore separately define these contributions.
Assuming that the cost associated with the measurement process depends es-
sentially on the measurement time, and assuming the measurement time is related
to the sample size (which, in particular, applies to point-to-point measuring sys-
tems), we can express CM as
CM = cM t ≈ cM t p n = c p n , (4.11)
where cM is the hourly cost of the measuring system, t is the time required to per-
form the measurement task, tp is the time for sampling a single point, cp = tpcM is
the cost of sampling a single point, and n is the number of sampling points.
In Equation 4.11 the cost related to each sampling point is supposed to be con-
stant, meaning that each point requires the same time to be sampled. This is not
always true, because point-to-point distances may differ, and the travel time varies
depending on the location of the point. Therefore, this evaluation of the measure-
ment cost is only approximate. Moreover, the time required to set up the machine,
align the part, etc., is not considered because it does not depend on the sampling
strategy.
Evaluation of CE, the inspection error cost, is quite subjective, and depends on
the approach chosen by the manufacturer to deal with inspection errors. In fact,
declaration of a part to be defective while it is actually conforming may lead to it
being discarded, or reworked (if possible), or to other expensive and unnecessary
152 G. Moroni and S. Petrò
actions. It is even harder to define the cost associated with declaring as conform-
ing an actually defective part, being related to the possibility of making some
finished product not working, or with decreasing customer satisfaction. A simple
approach is proposed in the following.
Suppose that conformance to a tolerance has to be verified, and the ISO 14253-1
(1998) standard is followed. According to the standard, a part is assumed as con-
forming if and only if the measurement result y is smaller than the specification
limit SL reduced by the expanded uncertainty U. Moreover, suppose that x, the
real geometric error of the part, behaves according to some statistical distribution
(e.g., a Gaussian distribution): if the uncertainty increases, a higher number of
parts will be rejected, even if they should be accepted. Under these assumptions,
CE can be evaluated as
CE = cw P ( SL − U < x < SL ) , (4.12)
will not be treated here; any uncertainty evaluation technique may be fit for this
method anyway.
Having defined a model for the inspection cost, and assuming that a sampling
pattern (uniform, adaptive, process-based, etc.) has been chosen, so that the link
between uncertainty and sample size is known, finding a trade-off between meas-
urement cost and inspection error cost is straightforward. It is sufficient to identify
the sample size that minimizes the inspection cost.
The case studies previously introduced for comparing different sampling patterns
in terms of uncertainty may also be useful to compare different strategies in terms
of inspection cost. The data required to perform this calculation are already avail-
able from preceding sections, with the exclusion of the statistical distribution of x,
which may be derived from the parts measured to obtain the raw data, or from the
manufacturing signature model.
In order to find the required trade-off between the measurement cost and the er-
ror cost, the optimization algorithm adopted in Sections 4.3.4.3 and “Example
Strategies for Roundness” may be modified to choose the pattern, leaving the
sample size free to change. This is computationally quite expensive, but leads to
the correct trade-off.
Of course, if no process-based sampling strategy is of interest, a blind sampling
strategy (e.g., a uniform strategy) could be simply selected, and then the sample
size may be left free to change: if the pattern is fixed, the link between sample size
and uncertainty is univocally determined, and then the identification of the optimal
Flatness Roundness
25 3
Uniform
Hammersley 2.5
20
Minimum U
Inspection Cost [ ]
2
15
1.5
10
1
5
0.5 Uniform
Minimum U
0 0
0 100 200 300 0 20 40 60 80 100
Sample Size
Figure 4.20 Inspection cost for flatness inspection, and roundness inspection
154 G. Moroni and S. Petrò
sample size is straightforward. Figure 4.20 clearly identifies the trade-off: if the
sample size is small, then the inspection error cost is high, that is, several parts are
wrongly rejected; if the sample size is large, inspection errors seldom occur, but
the measurement cost is high because the sample size is large. Optimal sampling
strategies may be identified between these conditions.
Different methods for defining the sampling pattern yield different inspection
costs: in particular, process-based strategies show a significantly lower inspection
cost, and a lower cost of the trade-off solution. However, it should be noted that as
evident in the case of roundness, if the sample size is large enough there is no cost
difference between blind and process-based strategies. This is because if the sam-
ple size is large enough, the blind sampling strategy sufficiently covers the feature
to be inspected and provides an uncertainty that is not significantly different from
the one granted by process-based strategy. Therefore, if there is no need to reduce
the sample size because the inspection costs are considered not relevant anyway, a
blind sampling strategy as dense as possible may be the correct solution, avoiding
any effort to acquire data on the actual manufacturing process. However, if reduc-
ing the sample size is unavoidable (e.g., because a very expensive measuring in-
strument is adopted or the measurement time available is short), the cost of acquir-
ing raw data or a modeling signature, which are not accounted for in the inspection
cost model, may be justified.
4.5 Conclusions
Finally, the problem of defining the correct sample size was depicted in its eco-
nomic nature. The choice of the sample size, significantly influencing the meas-
urement uncertainty, should be considered as an economic problem. A correct
choice of the sample size should balance the inspection error cost, inversely pro-
portional to the sample size, and the measurement cost, usually proportional to the
sample size. A cost model was proposed as a support for finding this trade-off.
References
ISO 1101 (2004) Geometrical product specifications (GPS) – tolerances of form, orientation,
location and run out, 2nd edn. International Organization for Standardization, Geneva
ISO 10360-1 (2000) Geometrical product specifications (GPS) – acceptance and reverification
tests for coordinate measuring machines (CMM) – part 1: vocabulary, 1st edn. International
Organization for Standardization, Geneva
ISO 10360-2 (2001) Geometrical product specifications (GPS) – acceptance and reverification
tests for coordinate measuring machines (CMM) – part 2: CMMs used for measuring size,
2nd edn. International Organization for Standardization, Geneva
ISO 10360-3 (2000) Geometrical product specifications (GPS) – acceptance and reverification
tests for coordinate measuring machines (CMM) – part 3: CMMs with the axis of a rotary ta-
ble as the fourth axis, 1st edn. International Organization for Standardization, Geneva
ISO 10360-4 (2000) Geometrical product specifications (GPS) – acceptance and reverification
tests for coordinate measuring machines (CMM) – part 4: CMMs used in scanning measuring
mode, 1st edn. International Organization for Standardization, Geneva
ISO 10360-5 (2000) Geometrical product specifications (GPS) – acceptance and reverification
tests for coordinate measuring machines (CMM) – part 5: CMMs using multiple-stylus prob-
ing systems, 1st edn. International Organization for Standardization, Geneva
ISO 10791-7 (1998) Test conditions for machining centres – part 7: accuracy of a finished test
piece, 1st edn. International Organization for Standardization, Geneva
ISO 14253-1 (1998), Geometrical product specifications (GPS) – inspection by measurement of
workpieces and measuring equipment – part 1: decision rules for proving conformance or
nonconformance with specifications, 1st edn. International Organization for Standardization,
Geneva
ISO 14660-1 (1999) Geometrical product specifications (GPS) – geometrical features – part 1:
general terms and definitions, 1st edn. International Organization for Standardization, Geneva
ISO/DIS 10360-7 (2008) Geometrical product specifications (GPS) – acceptance and reverifica-
tion tests for coordinate measuring machines (CMM) – part 7: CMMs equipped with imaging
probing systems. International Organization for Standardization, Geneva
ISO/IEC Guide 98-3 (2008) Uncertainty of measurement – part 3: guide to the expression of
uncertainty in measurement (GUM:1995), 1st edn. International Organization for Standardi-
zation, Geneva
ISO/IEC Guide 99:2007(E/F) (2007) International vocabulary of basic and general terms in
metrology (VIM), 1st edn. International Organization for Standardization, Geneva
ISO/TS 12180-2 (2003) Geometrical product specifications (GPS) – cylindricity – part 2: speci-
fication operators, 1st edn. International Organization for Standardization, Geneva
ISO/TS 12181-2 (2003) Geometrical product specifications (GPS) – roundness – part 2: specifi-
cation operators, 1st edn. International Organization for Standardization, Geneva
ISO/TS 12780-2 (2003) Geometrical product specifications (GPS) – straightness – part 2: speci-
fication operators, 1st edn. International Organization for Standardization, Geneva
ISO/TS 12781-2 (2003) Geometrical product specifications (GPS) – flatness – part 2: specifica-
tion operators, 1st edn. International Organization for Standardization, Geneva
156 G. Moroni and S. Petrò
Hammersley JM, Handscomb D (1964) Monte Carlo methods. Wiley, New York
Holland JH (1992) Adaptation in natural and artificial systems, 2nd edn. MIT Press, Cambridge
Hwang I, Lee H, Ha S (2002) Hybrid neuro-fuzzy approach to the generation of measuring
points for knowledge-based inspection planning. Int J Prod Res 40(11):2507–2520.
doi:10.1080/00207540210134506
Kim WS, Raman S (2000) On the selection of flatness measurement points in coordinate measur-
ing machine inspection. Int J Mach Tools Manuf 40(3):427–443
Kirkpatrick S, Gelatt CD, Vecchi MP (1983) Optimization by simulated annealing. Science
220(4598):671–680
Lee G, Mou J, Shen Y (1997) Sampling strategy design for dimensional measurement of geomet-
ric features using coordinate measuring machine. Int J Mach Tools Manuf 37(7):917–934
Lin ZC, Lin WS (2001a) Measurement point prediction of flatness geometric tolerance by using
grey theory. Precis Eng 25(3):171–184
Lin ZC, Lin WS (2001b) The application of grey theory to the prediction of measurement points
for circularity geometric tolerance. Int J Adv Manuf Technol 17(5):348–360
Montgomery DC (2004) Design and analysis of experiments, 6th 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(2):021003.
doi:10.1115/1.2904923, http://link.aip.org/link/?CIS/8/021003/1
Moroni G, Petrò S (2008) CMM measurement uncertainty reduction via sampling strategy opti-
mization. In: Proceedings of the 9th biennial ASME conference on engineering systems de-
sign and analysis, Haifa, Israel, CD-ROM
Moroni G, Petrò S (2009) Virtual CMM based sampling strategy optimization. In: Proceedings
of the 11th CIRP international conference on computer aided tolerancing, Annecy, France,
CD-ROM
Moroni G, Polini W, Semeraro Q (1998) Knowledge based method for touch probe configuration
in an automated inspection system. J Mater Process Technol 76(1):153–160
Moroni G, Polini W, Rasella M (2002) CMM trajectory generation. CIRP J Manuf Syst
31(6):469–475
Namboothiri VNN, Shunmugam MS (1999) On determination of sample size in form error
evaluation using coordinate metrology. Int J Prod Res 37(4):793–804
Petrò S (2008) Geometric tolerances verification: strategy optimization for CMM measurement.
PhD thesis, Politecnico di Milano, Milan
Raghunandan R, Rao PV (2007) Selection of an optimum sample size for flatness error estima-
tion while using coordinate measuring machine. Int J Mach Tools Manuf 47(3–4):477–482
Rossi A (2001) A form of deviation-based method for coordinate measuring machine sampling
optimization in an assessment of roundness. Proc IME B J Eng Manuf 215(11):1505–1518
Sartori S, Zhang GX (1995) Geometric error measurement and compensation of machines. CIRP
Ann Manuf Technol 44(2):599–609
Savio E, De Chiffre L, Schmitt R (2007) Metrology of freeform shaped parts. CIRP Ann Manuf
Technol 56(2):810–835. doi:10.1016/j.cirp.2007.10.008
Schwenke H, Siebert BRL, Wäldele F, Kunzmann H (2000) Assessment of uncertainties in
dimensional metrology by Monte Carlo simulation: proposal of a modular and visual soft-
ware. CIRP Ann 49(1):395–398
Schwenke H, Wäldele F, Weiskirch C, Kunzmann H (2001) Opto-tactile sensor for 2d and 3d
measurement of small structures on coordinate measuring machines. CIRP Ann Manuf Tech-
nol 50(1):361–364
Schwenke H, Neuschaefer-Rube U, Pfeifer T, Kunzmann H (2002) Optical methods for dimen-
sional metrology in production engineering. CIRP Ann Manuf Technol 51(2):685–699.
doi:10.1016/S0007-8506(07)61707-7
Schwenke H, Knapp W, Haitjema H, Weckenmann A, Schmitt R, Delbressine F (2008) Geomet-
ric error measurement and compensation of machines – an update. CIRP Ann Manuf Technol
57(2):660–675. doi:10.1016/j.cirp.2008.09.008
158 G. Moroni and S. Petrò
Summerhays KD, Henke RP, Baldwin JM, Cassou RM, Brown CW (2002) Optimizing discrete
point sample patterns and measurement data analysis on internal cylindrical surfaces with
systematic form deviations. Precis Eng 26:105–121
van Dorp BW, Haitjema H, Delbressine F, Bergmans RH, Schellekens PHJ (2001) Virtual CMM
using Monte Carlo methods based on frequency content of the error signal. Proc SPIE
4401:158–167. doi:10.1117/12.445616
Weckenmann A, Estler T, Peggs G, McMurtry D (2004) Probing systems in dimensional metrol-
ogy. CIRP Ann Manuf Technol 53(2):657–684
Wilhelm RG, Hocken R, Schwenke H (2001) Task specific uncertainty in coordinate measure-
ment. CIRP Ann 50(2):553–563
Yang TH, Jackman J (2000) Form error estimation using spatial statistics. J Manuf Sci Eng
122:262–272
Yang TH, Jackman J (2002) A Shannon sampling approach to form error estimation. Proc IME B
J Eng Manuf 216(2):225–233
Yau HT, Menq CH (1995) Automated CMM path planning for dimensional inspection of dies
and molds having complex surfaces. Int J Mach Tools Manuf 35(6):861–876
Zayed AI (1993) Advances in Shannon’s sampling theory. CRC, Boca Raton
159
160 N. Senin et al.
5.1 Introduction
5.1.1 Scenario
This work is concerned with assessing form error of surface topography features
manufactured at the micro and submicro scales, and whose geometry has been
acquired by means of 3D profilometers and/or 3D scanning microscopes. This is a
common scenario in quality inspection of microtopographic features manufactured
on the surface of small parts such as semiconductor products, microelectrome-
chanical systems (MEMS), and a wide array of other types of microcomponents.
Additional valid examples include standard-sized parts on whose surfaces micro-
topographic patterns have been manufactured.
The peculiar aspects of the problem being considered, with respect to other
form error assessment tasks, reside in the following considerations:
• Generally, the surface feature of interest is extremely small; under such prem-
ises, 3D profilometers and 3D scanning microscopes are usually selected as
measurement instruments for quality inspection, since they are the only types of
instruments which are capable of operating at such small dimensional scales
with the required precision.
• Microtopography acquisition with 3D profilometers and 3D scanning micro-
scopes is usually done through a raster scanning process, which consists in
sampling z coordinates on points lying on a uniform x, y grid. Geometric data
are available as a height map, essentially a discrete 2.5D geometry, which may
also be considered as formally equivalent to a digital grayscale image (z coor-
dinates being equivalent to pixel gray levels). Sampling is therefore often
suboptimal with respect to the geometry and orientation of the feature being in-
spected.
• Given the demanding requirements in terms of precision, measurement instru-
ments are often limited in terms of range. This leads to the difficulty of acquir-
ing –within the same measurement – the entire topography of the feature being
inspected together with the topography of additional datum surfaces which may
be needed for localization of the feature itself, which may be located further
apart; when this happens, feature localization errors are difficult to assess.
• When accurate localization is hard or impossible, common practice consists in
acquiring a rectangular portion of the surface topography, covering a region
larger than the feature, but within which the feature is known to be found; in
order to assess form error, it is necessary to spatially register the measured ge-
ometry of the feature with its nominal counterpart.
Under these premises, current industrial approaches adopt ad hoc procedures
depending on each specific application scenario, and a generalized solution is not
available.
To favor repeatability and reproducibility, a generalized approach should in-
clude the adoption of algorithmic solutions for the issues being considered: com-
5 Identification of Microtopographic Surface Features and Form Error Assessment 161
puter-assisted data processing solutions are needed to separate the points belong-
ing to the actual feature of interest from the remaining acquired scene points, for
registering identified feature points to the nominal geometry, and for computing
form error from the aligned geometries.
In this work, the nominal geometry of the feature of interest is assumed to be
available as a CAD model, while the manufactured feature is supposed to be
available as a part of a larger surface topography region, acquired by means of
raster scanning with a 3D profilometer or a 3D microscope. Only 2.5D geometries
are considered, as these are what most raster scanning devices are capable of ac-
quiring. A novel algorithmic solution for identifying the subset of points actually
belonging to the feature of interest, for aligning such points with the nominal ge-
ometry, and for performing a comparative assessment of the aligned geometries to
evaluate the overall form error is presented, and is validated through the applica-
tion to several case studies.
The nominal geometry of the feature of interest is referred to as the template. The
template is supposed to be available as a triangulated surface, defined by a stan-
dard tessellated language (STL) model (STL is a common format for storing tri-
angulated geometry), even though the proposed approach is generally applicable
to templates defined by any mathematical surface representation.
The measured surface topography is referred to as the scene. Consistent with
3D profilometers for microtopography and 3D scanning microscopes, the scene is
available as a simple set of height points organized over a regular grid (a height
map). Raster scanning ensures that the topological arrangement of points is known
(i.e., they form a structured data set); thus, intermediate surface coordinate values
can be obtained through interpolation on request.
The proposed approach comprises two main steps, briefly summarized in the
following:
• Feature identification and extraction: A moving window, the size of the tem-
plate, is used to scan the scene to search for positive matches with the template.
Best-matching regions are selected as candidate regions (i.e., scene regions
containing the feature). It is assumed that in general more than one feature in-
stance may be present in the scene; hence, more than one candidate may be
found. Best-matching candidate regions are extracted and turned into stand-
alone surfaces. At this point it is assumed that each stand-alone surface con-
tains an instance of the feature.
• Feature form error assessment: Each extracted feature is aligned with the tem-
plate through a sequence of coarse and fine alignment steps. The final align-
ment geometries can be used as the starting point for evaluating the form error.
162 N. Senin et al.
From an overall point of view, the problem discussed here is fundamentally novel:
• The need to assess form error on micromanufactured surface features is recent
and comes from the increased production of semiconductors, MEMS, and other
microcomponents; current solutions are based on ad hoc approaches.
• The widespread usage of measurement solutions based on 3D profilometers and
microscopes, combined with raster scanning, raises specific issues, unusual in
other form error assessment problems, and related both to uniform sampling
and to the 2.5D nature of measured geometry.
As briefly stated in the previous sections, the proposed approach identifies two
main steps: feature identification and extraction, and form error assessment, the
latter including geometry alignment. Each of these steps is related to specific lit-
erature domains, as illustrated in the following.
The analysis of the topography of manufactured surfaces at the micro and submi-
cro scales, in particular as acquired by means of 3D profilometers, is currently
based for the most part on approaches that rely on the computation of descriptors
of mean surface texture properties (e.g., roughness parameters) (Dong et al. 1994;
Lonardo et al. 1996). Feature identification is still a relatively uncommon issue in
the realm of micro- and nanotopography, and most of the previous work focused
on the identification of basins and watersheds (Scott 1998; Barré and Lopez 2000)
and on the application of filtering techniques such as wavelet filtering (Jiang et al.
2004; Zeng et al. 2005) for surface feature recognition.
Under the assumption of uniform sampling and the measured geometry being
intrinsically 2.5D, the analogy with grayscale images holds, and surface topogra-
phy analysis can obtain a great benefit from the adaptation of techniques originally
developed within the realm of digital image analysis and processing. Feature rec-
ognition is no exception, as novel solutions can draw inspiration from the litera-
ture on pattern recognition for digital images. Overviews of pattern-matching-
related issues for digital images can be found in Brown (1992) and Zitovà and
Flusser (2003). The specific approach proposed in this work is based on preproc-
essing topography data before searching for matches. Preprocessing is done with a
combination of segmentation and edge detection techniques: references on such
techniques as applied to digital images can be found in Pal and Pal (1993) and
Lucchese and Mitra (2001), while a solution specifically addressing surface topog-
raphy data is proposed in previous work by the authors (Senin et al. 2007). The
proposed approach for template matching is a novel solution resulting from the
adaptation of a recent technique developed for images and based on the ring pro-
5 Identification of Microtopographic Surface Features and Form Error Assessment 163
jection transform (Lin and Chen 2008). Transforms such as this one are shape
coding solutions, i.e., they turn shape information into a simpler numerical form,
more suitable to be handled in quantitative shape comparison tasks, and generally
designed to be invariant to shape localization, orientation, and sometimes scale as
well (also called shape descriptors). Other notable approaches to shape coding
include moment-based encoding (Flusser 2006), ridgelets, fast Fourier transforms,
and wavelet transforms (see Chen et al. 2005 for a combined use of the aforemen-
tioned techniques).
Form error assessment implies aligning the geometry to be evaluated with its
nominal counterpart. Since the goal is to evaluate form error from a complete 3D
perspective, alignment must take place in three dimensions; therefore, alignment
techniques developed for digital images – that operate on the more constrained
image plane – cannot be used in this case. Instead, the literature related to align-
ment in 3D space must be considered.
One of the most established and well-known approaches for aligning geome-
tries is the iterative closest point (ICP) technique (Besl and McKay 1992; Zhang
1994), with its variants (Audette et al. 2000; Rusinkiewicz and Levoy 2001). The
ICP technique allows for alignment of geometries with great precision, but since
the result is affected by the initial placement of the geometries, it is mainly used as
a fine alignment technique. Thus, a coarse alignment solution must be obtained as
a starting point for the ICP technique: among the most popular approaches avail-
able in the literature for coarse-alignment, many of which are summarized in
Audette et al. (2000), it is worth mentioning those based on matching relevant
axes, such as symmetry axes (Kazhdan et al. 2004), and principal component
analysis axes, and those based on matching significant geometric features belong-
ing to both geometries. In the latter case, features can be recognized by looking at
local curvatures, e.g., in the curvature scale space method (Mokhtarian et al.
2001), or by looking at other local shape descriptors defined to ease the recogni-
tion process, such as spin images (Johnson and Hebert 1999) and shape contexts
(Kortgen et al. 2003).
Form error assessment is concerned with obtaining a quantitative measure of
the difference between two geometries. The approach that has been followed in
this work consists in aligning geometries first, and then computing the form error
by combining local surface-to-surface Euclidean distances into a measure of the
volume. Most 3D shape comparisons in the literature rely on shape descriptors
instead, which, as mentioned earlier, allow for encoding shape information into
simpler numerical forms. With descriptors, shape comparison becomes a matter of
comparing descriptor results. Many of the techniques mentioned previously for
164 N. Senin et al.
In this section, the proposed approach is illustrated with the help of an artificially
generated case study, designed to highlight the specific aspects of the problem and
the steps of the procedure.
To better describe the steps of the proposed approach, a simulated case study is
introduced. The STL geometry shown in Figure 5.1a represents the template of an
example surface feature to be recognized. In Figure 5.1b an artificially generated
(i.e., simulated) scene is shown, containing several instances of the template with
different placement and orientation.
a b
Figure 5.1 Geometries for the artificially generated case study: a template standard tessellated
language (STL) geometry, and b simulated scene. The template is not visualized at the same
scale as the scene
5 Identification of Microtopographic Surface Features and Form Error Assessment 165
a b
c d
Figure 5.2 Creation of the simulated scene geometry: a original STL model, b new triangula-
tion and application of algorithmic deformation effects, c virtual raster scanning, and d close-up
view of two feature instances on the raster-scanned surface
166 N. Senin et al.
Figure 5.3 The overall process of feature recognition, extraction, and form error assessment for
surface features against a template geometry
5 Identification of Microtopographic Surface Features and Form Error Assessment 167
Through comparison with the template, best matches for the nominal feature are
identified on the scene surface (feature identification) and are extracted as separate
geometries (feature extraction). Each geometry extracted from the scene is then
aligned to the template geometry to assess nominal-to-measured form error.
The single steps contained in the schema depicted in Figure 5.3 are discussed in
detail in the next sections.
The detailed breakdown of the feature identification step is shown in Figure 5.4.
Identification is performed through a main scanning loop, where a moving win-
dow that scans the original scene is used to select a portion of the surface to be
subjected to pattern matching with the template. Both the template and the scene
Figure 5.4 The feature identification step. MRPT modified ring projection transform
168 N. Senin et al.
geometry enclosed in the window are preprocessed first, to enhance salient attrib-
utes and to make the approach more robust to noise and other variations. The re-
sult of pattern matching is a matching score for each position of the moving win-
dow. At the end of the scanning process, window positions corresponding to the
highest matching scores are selected as the best candidate results of feature identi-
fication, and the corresponding surface regions are extracted from the original
topography as representative of positive identifications.
The moving window is translated over the original scene surface by selecting its new
center point at each step, according to a raster scanning strategy that covers all sur-
face points. The window itself extends about the center point, and zero padding is
used to make up for the missing points when the window is close to the scene
boundaries. The region enclosed within the moving window is called the candidate
region. For coarse investigation of a large scene, raster scanning can be modified to
consider only a subset of scene points, by skipping a predefined number of positions
while moving over the scene surface. The size of the moving window/candidate
region is important, as it must match the spatial extensions of the template geometry.
Before the actual comparison between the template and each candidate region can
take place, the template, originally available as an STL model, must be turned into a
form suitable for being compared with the candidate. For this purpose, the template
is subjected to the same virtual raster scanning process the simulated scene was
subjected to; if the scene is the result of an actual measurement process, virtual raster
scanning of the template ensures that the measurement conditions applied when
acquiring the surface topography of the scene are replicated for the template. This
step must be done only once, and ensures that both the template and the candidate
region topographies contain the same spatial frequencies, and thus are comparable.
Both the candidate region and the template undergo a preprocessing stage before
being compared. Preprocessing is aimed at highlighting the most relevant attrib-
utes of the surface feature to be identified, and it makes identification less sensi-
tive to nonrelevant forms of variation (disturbances). A correct choice of preproc-
essing operations may have a fundamental impact on the performance of the
feature identification step. Since it acts as a filter for highlighting what is relevant,
5 Identification of Microtopographic Surface Features and Form Error Assessment 169
it must be designed depending on the specific feature one is after; on the other
hand it should be considered that while the template geometry is preprocessed
only once, the candidate region changes at each position of the moving window,
and thus it must be preprocessed many times. Computational load may become an
issue to consider before opting for overly complex preprocessing solutions.
Preprocessing may include leveling (by means of subtracting the z coordinates of
the least-squares mean plane from the coordinates of the actual surface) if the rele-
vant attributes of the feature are not deemed to be related to the overall orientation of
the surface normal for the candidate region with respect to the overall surface. Some
of the most common preprocessing approaches are illustrated in the following.
a b c
d e f
Figure 5.5 Preprocessing with height-based binarization: a original raster-scanned template,
b binary map of the template, c binarized template topography, d candidate region from the
scene, e binary map of the candidate, and f binarized candidate topography
170 N. Senin et al.
With reference to Figure 5.5 one can see that in this case the effect of preproc-
essing is to turn the identification problem into a 2D binary pattern matching prob-
lem. Of course this is applicable when a 2D silhouette is deemed as sufficient for
discrimination; other, more complex cases may require more complex preprocess-
ing strategies.
a b c
Figure 5.6 Preprocessing with slope-based segmentation to detect feature edges: a example
surface topography, b segment map (each color identifies a region with uniform slope), and
c edge map (edges are the borders of the uniform-slope regions)
5 Identification of Microtopographic Surface Features and Form Error Assessment 171
peculiar aspects of the feature recognition problem. In any case, the result of the
preprocessing step applied to a rectangular portion of surface topography (whether
the template or a candidate region) consists of a map of scalar values, which in
terms of data structures can be assimilated to a grayscale image and handled as
such, thus paving the way to use pattern matching algorithms developed in the
realm of digital image processing.
Once the template and the candidate region geometries have been preprocessed
and the corresponding maps of scalar values have been generated, a quantitative
comparison of the template and the candidate must be done. From a general stand-
point, a quantitative comparison between shapes relies on the availability of quan-
titative measures of shape, i.e., shape descriptors, mapping shape properties to
numerical values. Once such values are known, shape comparison can be mapped
to a simple difference of values.
The choice of an appropriate shape descriptor depends on the nature of the
data structures it must operate upon, and on the specific requirements of the fea-
ture identification process. In this case, the requirements are, on one hand, to
cooperate with the preprocessing substep to ensure there is the required robust-
ness (to noise and partial occlusion) and, on the other hand, that the descriptor
must capture the salient traits of the shape without being influenced by shape
localization and orientation. Robustness and invariance requirements are often
counterbalanced by discrimination power, i.e., with the capability of telling apart
two different shapes.
The proposed approach draws inspiration from the ring projection transform,
recently applied in the literature to pattern matching problems for digital images
(Lin and Chen 2008). This work adopts a modification of the transform, which
will be referred to as the modified ring projection transform (MRPT), illustrated
by Equation 5.1 and with the support of Figure 5.7.
nθ
C ( rj ) = ∑ rj dθ ⋅ z ( xc + rj cos θ i , yc + rj sin θ i ) . (5.1)
i =1
y C
rj
θi
xc, yc
C (rj)
x rj r
a b
Figure 5.7 Conceptual schema of the MRPT: a geometric construction of the circumference
each integral is computed upon, and b resulting MRPT curve
To compute the similarity (or difference) of two shapes, one can build the two
MRPT curves, and then compute their difference in several ways. In this work
three main approaches were investigated:
1. the cross-correlation coefficient computed on the two curves;
2. the maximum value of the cross-correlation function computed on the two
curves; and
3. the integral of the point distances of the two curves.
The second solution captures curve-shape differences but is less sensitive to the
relative positioning of the two curves. The first and third solutions are sensitive to
relative positions of the curves. The third approach was selected as the most prom-
ising, and was implemented according to the expression illustrated by Equa-
tion 5.2:
nr
1
Errtot = ∑ Ctpl ( rj ) − Cscn ( rj ) . (5.2)
rmax j =1
The cumulative error Errtot is computed as the summation of the errors at each
rj value, Ctpl being the MRPT curve of the template, and Cscn the MRPT curve of
the candidate region (portion of the scene). To make the cumulative error inde-
pendent of the feature size, the cumulative error is divided by the maximum radius
of the MRPT. Finally, remember also that this approach returns an error, while
cross-correlation-based approaches return a similarity measure. To treat this as a
similarity problem, the error function must be inverted.
Finally, notice that all the approaches to MRPT curve comparison, Equation 5.2
being no exception, capture cumulative differences between the two curves, but
not local ones, which is like saying that more sophisticated ways to compare the
MRPT curves may lead to an increase in the discrimination power of the similarity
metric selected.
5 Identification of Microtopographic Surface Features and Form Error Assessment 173
Given that the MRPT integral is computed over a circumference, the angular ori-
entation of the shape does not change its value. However, the position of the cen-
ter each circumference is referred to does make a significant difference; hence, the
approach is independent of orientation but not of localization of the shape. Sec-
ondly, it should be noted that any shape containing the same amount of “mass”
(i.e., values) at the same radial distance from the center will be encoded with the
same C(r) value; this is a limitation in terms of discriminative power of the ap-
proach. In other words, the “averaging” effect of integration, on one hand, in-
creases the robustness to small shape variation, but, on the other hand, decreases
the discriminative power as different shapes may be encoded similarly owing to
what was illustrated above. Remember, however, that the descriptor is sensitive to
localization of the given feature. This is actually not a problem, as the center of the
transform is actually the center of the moving window translating on the original
surface topography (see Section 5.4.4). The best match between MRPT curves
will be obtained corresponding to the match of the template center and the feature
center (see the examples in Figure 5.8).
a b c
Figure 5.8 MRPT construction on binarized topography, and corresponding MRPT curves:
a template, b candidate feature, center-aligned with template, and c off-centered candidate fea-
ture
a b
Figure 5.9 Match response surface for the ongoing example; a z-const contour plot has been
overlaid to better visualize the location of the local maxima: a perspective view, and b top view
(local maxima are circled)
viewable as a surface, whose z coordinates are the actual scores themselves. This
surface is referred to as the match response surface; the match response surface
obtained for the ongoing example is shown in Figure 5.9. In a match response
surface, peaks (local maxima) are the points where the best matches between the
candidate region and the template were found; i.e., they correspond to the best
guesses in terms of identifying the feature of interest on the scene surface. In the
result obtained for the ongoing example, shown in Figure 5.9, the first eight high-
est maxima correspond exactly to the centers of the features in the scene surface,
meaning that all eight instances of the feature were positively identified, regard-
less of their orientation. In the example, the highest matching scores are for the
scene features that have undergone the least deformation with respect to the tem-
plate, while the most deformed ones achieve lower scores.
It is expected that as the differences between the scene and template features
increase, some positive identifications may be missed; at the current state of re-
search it is difficult to determine when this may happen, as it also depends on the
intrinsic properties of the feature being analyzed, on measurement process pa-
rameters, and on the preprocessing step discussed previously.
The local maxima of the match response surface correspond to the centers of the
scene regions to be extracted as the best candidates for the feature identification
problem. The process of extracting such regions from the original surface topog-
raphy is called feature extraction: since it basically involves just an extraction of a
finite set of points belonging to the original scene set, the process is straightfor-
ward and does not need to be discussed further. Consistently with the raster scan-
ning process, if the center point is located close to one of the scene borders, the
extraction will use zero padding to determine the values of the missing points.
5 Identification of Microtopographic Surface Features and Form Error Assessment 175
Figure 5.10 Regions to be extracted as the best-matching candidates in the original scene
topography (their centers correspond to the first eight local maxima in the match response sur-
face shown in Figure 5.9)
The last part of the proposed approach is related to performing an accurate com-
parison of the template and each measured feature (the regions extracted from the
scene as the best-matching candidates) in order to assess the form error. The over-
all process is described in Figure 5.11.
Form error assessment implies aligning the geometries of the template and the
recognized feature first, and then computing the differences between the two
aligned geometries. However, the entire process is completely dependent on the
type of form error that must be assessed. Several types of form error may require
datum surfaces to be identified, if this is the case, datum surfaces should be de-
fined on the template, their counterparts located on the candidate region, and both
used for alignment. In this work it is assumed that no datum is available; therefore,
the only alignment option consists in globally aligning the entire geometries,
minimizing the sum of squared distances computed between each template point
and its closest counterpart on the candidate region.
The global alignment problem discussed here is special since at the end of the
feature identification and extraction steps, the template (nominal feature) and the
candidate region (measured feature) are already partially aligned. In fact, feature
identification itself, by locating the points corresponding to the best matches, pro-
vides a hint on the relative placement of feature centers; however, no information
is given concerning rotational alignment, the MRPT being invariant to rotation.
176 N. Senin et al.
Therefore, the alignment problem becomes for the most part a rotational align-
ment problem (about the z-axis), even though some rotational and translational
alignment about the other axes may still be necessary to compensate for discretiza-
tion and preprocessing-induced errors in the matching stages.
Figure 5.11 Template versus measured feature comparison step. ICP iterative closest point
5 Identification of Microtopographic Surface Features and Form Error Assessment 177
The most widespread technique available in the literature for aligning two geome-
tries is known as iterative closest point (ICP) (Besl and McKay 1992; Zhang
1994). ICP is suitable for application to the problem discussed here. However, it is
well known that ICP tends to converge to local minima; therefore, geometries
need to start from an initial position which is quite close to the optimal result.
While in this case this is true for translational alignment, it definitely does not hold
for rotational alignment, since the matched feature and the template may be in
completely different angular orientations with respect to the z-axis. In order to deal
with this problem, in this work ICP was used as a fine alignment solution, while a
preliminary coarse alignment step was accomplished by means of another tech-
nique based on rotational alignment about the z-axis only.
Both the template and the candidate region need to be preprocessed for better
alignment. Recall that the candidate region originates as a rectangular portion of the
original scene, defined by a regular grid of equally spaced points. The candidate
region is turned into a triangulated geometry so that a continuous (C0) surface re-
presentation is available in case interpolated points are needed (and in fact they will
be needed by both the coarse and the fine alignment procedures). The template
geometry is retriangulated as well, even though it was already provided as a trian-
gulated geometry from the beginning (recall that the template was originally de-
fined from STL geometry). The goal of triangulation is to produce a mesh finer than
the original to facilitate optimal alignment with algorithms that operate on points.
To accomplish the coarse alignment step (rotational and about the z-axis), the follow-
ing procedure is proposed. First, the template and candidate geometries are trimmed
into circular shapes (as seen from a z-aligned viewpoint), the diameters being se-
lected so that the feature is completely enclosed within the circumference. This is
done to eliminate the peripheral points lying in the rectangular corners of the topog-
raphies, which would influence too much the overall outcome of angular alignment.
The template and candidate geometric centers are registered within a common
Cartesian reference frame, then both geometries are sliced with diametral planes
passing through the z-axis and through the geometric centers, each creating a dia-
178 N. Senin et al.
a b
c d
Figure 5.12 Coarse alignment procedure applied to the ongoing example: a template with
radial profiles, b unaligned candidate and radial profiles, c match score profile (minimum corre-
sponding to optimal alignment angle), and d rotationally aligned candidate
5 Identification of Microtopographic Surface Features and Form Error Assessment 179
a b
c
Figure 5.13 ICP fine alignment applied to the ongoing example: a model and scene in their
original position, b model and scene in their final aligned position, and c displacement error plot
180 N. Senin et al.
belongs to the template, while the light-gray mesh belongs to the candidate geome-
try; the final alignment, reached after about 30 iterations, is shown in Figure 5.13b.
Figure 5.13c shows the cumulative displacement error as a function of the ICP
iteration; in this implementation convergence to a final alignment is assessed by
setting a minimum threshold value on the variation of the displacement error.
Given the underlying assumption of the absence of reference datums (hence the
global alignment approach), form error can be treated as a quantitative measure
related to overall differences between the template and candidate geometries.
However, the candidate geometry cannot be directly compared with the template
geometry yet. The template in fact represents a geometry which, even when ide-
ally replicated on the manufactured surface, would still appear different at meas-
urement, owing to the very same nature of the acquisition process: raster scanning
introduces a discretization which – at these scales – usually cannot be neglected.
Form error assessment becomes the problem of comparing the candidate geometry
with an ideal geometry, as it would appear after measurement.
candidate surface
measured point
on candidate
z
unit volume error
y
x
error vector (length: |ΔZ|)
template surface
x point spacing
y point spacing
Figure 5.14 Geometric construction for the error vector and the unit volume error
a b
Figure 5.15 Error vectors in the ongoing example: a error vector map, and b unit volume
errors
virtual raster scanning). The error vector can be used to compute a unit volume
error by multiplying it by the unit measurement area, which in raster scanning is x
point spacing times y point spacing (see Figure 5.14).
The sum of all unit volume errors computed for all candidate points that have a
corresponding counterpart on the template is the total volume error. The approach
is also summarized by Equation 5.3:
nc
Etot = A0 ⋅ ∑ Δzi , (5.3)
i =1
To validate the proposed approach, the application to a real case study is now
illustrated. The subject of the analysis is the microtopography of the surface of a
steel embossing roller for texturing a thin polymeric film. The embossing pattern,
whose functional role requires a certain degree of geometric uniformity, has been
182 N. Senin et al.
a b
Figure 5.16 Surface topography of embossing roller: a template geometry (STL), and b
measured surface (approximate projected area 1,473 µm × 1,531 µm). The template is not
visualized at the same scale as the scene
engraved on the roller. The measured surface topography (a portion of the roller
surface) is shown in Figure 5.16b; it was obtained through raster scanning with a
3D profilometer equipped with a noncontact laser sensor (conoscopic holography;
Optimet Conoscan 1000), consists of 280 × 290 scanned points with 5.28-µm
spacing in x, y and z resolution less than 0.1 μm, and covers an approximate area
of 1,473 µm × 1,531 µm. To apply the proposed approach, the nominal geometry
shown in Figure 5.16a was defined to represent the characteristic shape of a pat-
tern element, and was modeled as an STL triangulated geometry by means of a
CAD system.
Before the MRPT was applied, the template was slightly extended at the sides
with flat surfaces so that the circular patterns of the MRPT would cover the entire
feature definition area, as shown in Figure 5.17a.
The combination of slope-based segmentation and region edge detection
illustrated earlier was then selected as the preprocessing step for the virtual raster
scanned template and the candidate regions. Figure 5.17b shows the preprocessed
template; slight asymmetries in the final edge map are due to discretization errors
in raster scanning, segmentation, and edge detection. Figure 5.17c shows the
MRPT curve computed on the preprocessed template. Finally, Figure 5.17d shows
the preprocessed scene; notice how the irregularities on the measured surface
result in a slightly more irregular edge map.
At the end of the main scanning loop, a match score surface is obtained, as shown
in Figure 5.18a. Local maxima represent the locations of the highest-scoring matches
of the feature-identification process. The match score surface is complex, as ex-
pected given the many self-similar regions generated by the preprocessing transform.
5 Identification of Microtopographic Surface Features and Form Error Assessment 183
a b
c d
Figure 5.17 Preprocessing steps: a template surface extended for the MRPT and subjected to
virtual raster scanning, with overlaid circular paths, b template binary map after preprocessing
(slope-based segmentation and edge detection on the segment map), c MRPT curve for the pre-
processed template, and d scene binary map after preprocessing (slope-based segmentation and
edge detection on the segment map)
a b
Figure 5.18 Feature identification results: a match score surface, and b corresponding locations
of the regions defined by the 16 local maxima identified
184 N. Senin et al.
While some maxima are strong where we expected them to be, and thus prop-
erly locate some of the features, others are weaker and are as high as other
peaks referring to features that are partially located outside the acquired area.
To recognize the 16 features properly, rules were added to the extraction of
local maxima so that they could not be too close to each other and/or too close
to the border of the acquired region. This is basically imposing a constraint on
feature localization of the feature recognition process. Given such constraints,
the 16 highest matching scores correspond to the 16 regions highlighted in
Figure 5.18b.
a b c
Figure 5.19 Fine alignment results with one of the extracted candidates: a template and candi-
date before fine alignment, b fine-aligned geometries, and c unit volume errors
5 Identification of Microtopographic Surface Features and Form Error Assessment 185
5.7 Conclusions
In this work a novel algorithmic approach was proposed for form error assessment
for microtopography features such as those that are commonly found on semicon-
ductor products, MEMS and microcomponents in general, as well as for micro-
topographic patterns that can be micromanufactured on the surfaces of standard-
sized parts. The surface topography is assumed to be available as a height map,
i.e., 2.5D geometry, as is typically acquired through raster scanning by means of a
3D profilometer or a 3D scanning microscope. The approach comprises several
steps. First, the feature of interest is localized and identified within the measured
region by means of a modified version of the ring projection transform, adapted to
operate on topography data. Then it is extracted and aligned with a CAD-modeled
reference (representing the nominal geometry): alignment comprises two steps
(coarse alignment, through an exhaustive search over discrete angular positions,
and fine alignment, with ICP). Finally, the form error is computed as the volume
enclosed between the two aligned geometries: computation also takes into account
discretization induced by raster scanning. The approach was illustrated and vali-
dated first through its application to an artificially generated case study, then to a
real-life case of industrial relevance. The results show that the approach is indeed
effective in identifying the features of interests, aligning them to their nominal
references (templates), and obtaining a measure of the form error. However, some
open issues still remain.
The performance of the proposed approach for coarse alignment depends on the
selection of the angular discretization parameter, which must be carefully selected
depending on the geometric properties of the feature to be aligned. This depend-
ency is not that important for features with strong asymmetry traits, but becomes
relevant for quasi-symmetric features, where a too-coarse discretization may lead
to unsatisfactory alignment. Algorithmic determination of suitable angular discre-
tization should be pursued.
Types of form error assessment requiring reference datums to be registered first
require changes to the alignment procedure, as a global alignment solution would
not be adequate. In such situations, alignment should take identify the datums on
the candidate (the datum should be explicitly defined on the template) and then
overlay them; then, the remaining alignment transforms should be constrained to
move the features without breaking datum alignment.
Final mention should be reserved for the problem of measurement uncertainty,
which was not dealt with in this chapter, but is, however, of fundamental importance
for any quality inspection procedure. Owing to their rapid evolution, most pro-
filometers and microscopes dedicated to the analysis of 3D surface microtopography
lack proper measurement traceability, and their accuracy (trueness and precision)
cannot be easily determined. International standards are beginning to be available
for specific classes of instruments (ISO 25178-601:2010, ISO 25178-602:2010 and
ISO 25178-701:2010) and others are being developed (ISO/DIS 25178-603,
ISO/DIS 25178-604, ISO/DIS 25178-605); however, some more time is necessary
before such efforts reach maturity and gain widespread acceptance.
References
ISO 25178-601:2010 Geometrical product specifications (GPS) – Surface texture: Areal – Part 601:
Nominal characteristics of contact (stylus) instruments, International Organization for Stan-
dardization, Geneva
ISO 25178-602:2010 Geometrical product specifications (GPS) – Surface texture: Areal – Part 602:
Nominal characteristics of non-contact (confocal chromatic probe) instruments, International
Organization for Standardization, Geneva
ISO 25178-701:2010 Geometrical product specifications (GPS) – Surface texture: Areal – Part 701:
Calibration and measurement standards for contact (stylus) instruments, International Organiza-
tion for Standardization, Geneva
Audette MA, Ferrie FP, Peters TM (2000) An algorithmic overview of surface registration tech-
niques for medical imaging. Med Image Anal 4(3):201–217
Barré F, Lopez J (2000) Watershed lines and catchment basins: a new 3D-motif method. Int J
Mach Tools Manuf 40: 1171–1184
Besl PJ, McKay ND (1992) A method for registration of 3-D shapes IEEE Trans Pattern Anal
Mach Intell 14(2):239–256
Brown LG (1992) A survey of image registration techniques. Comput Surv 24(4):325–376
5 Identification of Microtopographic Surface Features and Form Error Assessment 187
Chen GY, Bui TD, Krzyzak A (2005) Rotation invariant pattern recognition using ridgelets,
wavelet cycle-spinning and Fourier features. Pattern Recognit 38(12):2314–2322
Dong WP, Sullivan PJ, Stout KJ (1994) Comprehensive study of parameters for characterising
3D surface topography III: parameters for characterising amplitude and some functional
properties, & IV: parameters for characterising spatial and hybrid properties. Wear 178:29–60
Flusser J (2006) Moment invariants in image analysis. Proc World Acad Sci Eng Technol
11:196–201
Funkhouser T, Kazhdan M, Min P et al (2005) Shape-based retrieval and analysis of 3D models.
Commun ACM 48(6):58–64
Horn BKP (1987) Closed-form solution of absolute orientation using unit quaternions. J Opt Soc
Am A 4:629–642
Jiang XQ, Blunt L, Stout KJ (2004) Third generation wavelet for the extraction of morphological
features from micro and nano scalar surfaces. Wear 257:1235–1240
Johnson AE, Hebert M (1999) Using spin images for efficient object recognition in cluttered 3D
scenes. IEEE Trans Pattern Anal Mach Intell 21(5):433–449
Kazhdan M, Chazelle B, Dobkin D et al (2004) A reflective symmetry descriptor for 3D models.
Algorithmica 38(1):201–225
Kortgen M, Park G-J, Novotni M et al (2003) 3D shape matching with 3D shape contexts. In:
Proceedings of the 7th central European seminar on computer graphics, Budmerice, Slovakia
Lin YH, Chen CH (2008) Template matching using the parametric template vector with transla-
tion, rotation and scale invariance. Pattern Recognit 41:2413–2421
Lonardo PM, Trumpold H, De Chiffre L (1996) Progress in 3D surface microtopography charac-
terization. Ann CIRP 42(2):589–598
Loncaric S (1998) A survey of shape analysis techniques. Pattern Recognit 31(8):983–1001
Lucchese L, Mitra SK (2001) Color image segmentation: a state-of-the-art survey. Image proc-
essing, vision, and pattern recognition, Indian National Science Academy, New Delhi
Mokhtarian F, Khalili N, Yuen P (2001) Curvature computation on free-form 3-D meshes at
multiple scales. Comput Vis Image Understand 83(2):118–139
Pal NR, Pal SK (1993) A review on image segmentation techniques. Pattern Recognit
26(9):1277–1294
Rusinkiewicz S, Levoy M (2001) Efficient variants of the ICP algorithm. In: Third international
conference on 3-D digital imaging and modeling (3DIM ’01), p 145
Scott PJ (1998) Foundation of topological characterisation of surface texture. Int J Mach Tools
Manuf 38(5–6): 556–559
Senin N, Ziliotti M, Groppetti R (2007) Three-dimensional surface topography segmentation
through clustering. Wear 262(3–4):395–410
Tangelder JWH, Veltkamp RC (2004) A survey of content based 3D shape retrieval methods. In:
Proceedings shape modeling international, pp 145–156
Zeng W, Jiang X, Scott P (2005) Complex ridgelets for the extraction of morphological features
on engineering surfaces. J Phys Conf Ser 13:246–249
Zhang Z (1994) Iterative point matching for registration of free-form curves and surfaces. Int J
Comput Vis 13:119–152
Zitovà B, Flusser J (2003) Image registration methods: a survey. Image Vis Comput
21:97721000
ISO/CD 25178-604 Geometrical product specifications (GPS) – surface texture: areal – part 604:
Nominal characteristics of non-contact (coherence scanning interferometry) instruments. In-
ternational Organization for Standardization, Geneva
188 N. Senin et al.
ISO/CD 25178-605 Geometrical product specifications (GPS) – surface texture: areal – part 605:
nominal characteristics of non-contact (point autofocusing) instruments. International Or-
ganization for Standardization, Geneva
ISO/DIS 25178-603 Geometrical product specifications (GPS) – surface texture: areal – part 603:
nominal characteristics of non-contact (phase-shifting interferometric microscopy) instru-
ments. International Organization for Standardization
Chapter 6
Geometric Tolerance Evaluation Using
Combined Vision – Contact Techniques
and Other Data Fusion Approaches
Gianni Campatelli
Hybrid coordinate measuring machines (CMMs) use more than one sensor in
order to obtain information faster regarding the test piece geometry. The main
difference with respect to traditional CMMs is the use of optical sensors to acquire
geometric data. The optical acquisition source can be mounted on the arm of the
__________________________________
G. Campatelli
Dipartimento di Meccanica e Tecnologie Industriali, Università di Firenze,
Via di S. Marta 3, 50139 Florence, Italy,
e-mail: gianni.campatelli@unifi.it
189
190 G. Campatelli
CMM (i.e., camera) or can be external to the CMM structure depending on the
technology and acquisition strategy chosen. The use of optical sensors does not
exclude the successive use of a contact probe in order to collect a high-resolution
dataset in a time-consuming manner.
The classic use of a hybrid CMM is for medium-sized to small products; so the
traditional implementation of such systems is a medium-sized CMM with a work-
space volume ranging from about 0.5 to 1 m3. Products with larger dimensions are
usually measured by using a pure optical system or, at least, with a system where the
trigger probe position is measured using optical sensors (i.e., laser triangulation or
optical orientation of the touch probe, a solution often used for naval and car body
measurements); however the aim of this chapter is the presentation of a hybrid
CMM where optical and touch probe approaches are both used in order to obtain a
measurement, so these large-scale applications will not be treated in this chapter.
The optical systems with which a hybrid CMM is equipped are usually based
on techniques that allow a high rate of point acquisition, such as laser triangula-
tion, structured light projection, and interferometric systems. Other optical sensor
techniques characterized by higher precisions but lower acquisition rates, such as
laser interferometry and conoscopy, are not used, at least in an industrial field of
application, in a hybrid CMM.
The hybrid CMM idea is spreading (Chen and Lin 1997) because the actual op-
tical probes can provide good precision in measuring well-defined geometries,
theoretically as good as the precision obtained with contact probes in certain spe-
cific and controlled conditions (Chen and Lin 1991). Within the category of hybrid
CMMs there are two main groups:
1. The imaging probe CMM group, which encompasses, as defined by the
ISO/DIN 10360-7/2008 standard, all the CMMs equipped with a vision probe
system. A subcategory of these is the optical CMM (OCMM), where the sensor
is a camera.
2. The group of CMMs with standstill optical systems, which include all the ap-
proaches that have the optical sensors not rigidly connected to the CMM (i.e.,
mounted on the arm) but used externally to provide information for the succes-
sive use of a touch probe. An example of an application is the use of an inde-
pendent structured light system supporting the CMM.
It is, however, necessary to point out that the use of optical probes introduces
some specific error sources, such as the type of illumination, the optical lens image
distortion, the distance from the part, and the optical characteristics of the measured
part. For these reasons the real resolution of a CMM equipped with an optical sensor
is typically much lower than that of a CMM with a touch probe. By “resolution”, we
refer to the smallest spatial distance that a measuring device can distinguish.
A CMM could have a resolution as fine as 0.5 µm, while an optical system
mounted on a CMM typically has a resolution of the order of 10 µm (Shen et al.
2000). Laser scanners have a very high data acquisition rate (up to 2,500 data
points per second), and good resolution, of the order of 10 µm (Chan et al. 1997).
Vision systems have lower resolutions (about 20–30 µm); however, they can
6 Geometric Tolerance Evaluation Using Combined Vision – Contact Techniques 191
acquire thousands of data points simultaneously over a large spatial range (El-
Hakim and Pizzi 1993) without moving the optical head.
Optical methods, such as interferometry, stereo vision, structured light projec-
tion, and shape from focus/defocus, have long received extensive attention. Unlike
classic touch probe CMMs, measurement systems developed from optical princi-
ples are noncontact in nature. This characteristic is responsible for the absence of
wear and deformation during the measuring process that sometimes can happen
when contact systems are used with low hardness and low stiffness measurands.
A typical application of an OCMM is in the field of reverse modeling, where a
large number of data points must be acquired and the touch probe approach would
be too slow. However, if high precision is required for reverse modeling, the use
of a touch probe is mandatory as it is if there is a reduced number of points. For
this reason, an OCMM system usually integrates both contact and noncontact
acquisition systems.
Nashman et al. (Nashman 1993; Nashman et al. 1996) were among the first to
develop integrated vision–touch probe systems, which emphasize the integration
of a CMM, a 3D analog touch probe, a video camera, and a laser triangulation
probe, all managed by a unique system controller. Their work shows that it is
possible to create multisensor cooperative interaction of a vision and a touch probe
system which provides sensory feedback to a CMM for a measuring task.
Examples of the application of OCMMs can be found in many companies. In
Figure 6.1 a CMM with an optical sensor mounted and a detailed picture of an
optical sensor (laser line scanner) are illustrated.
Figure 6.1 a Touch probe for a coordinate measuring machine (CMM), and b a laser line scanner
192 G. Campatelli
Currently, the standard systems for surface digitization are the digital light proc-
essing (DLP) projector and laser line scanners.
The DLP projector is one of the most used solutions for surface digitization
thanks to the flexibility and high precision that can be obtained with this system,
especially if compared with the older technology of LCD projectors. The DLP
projector is based on the technology of a digital micromirror device (DMD) (Horn-
beck 1998). The DMD is an array of micromirrors fabricated onto a memory chip
and directly controlled by it. Each micromirror is able to turn on or off a single
pixel of light controlled by the input digital signal. The low inertia of the system
allows the mirrors to be switched at very high speed. The switching creates an
image that is reflected from the surface of the DMD. This allows the DLP projector
to create easily, quickly, and without interpolation error a large variety of digital
patterns with high resolution and precision. The most used structured light patterns
range from a single light beam to a grid light, a single light stripe, and multiple
light stripes. The DLP projector is associated with a CCD camera that grabs images
of the object. Through the sampling of the 2D image coordinates of the surface
points illuminated by the light patterns, the 3D coordinates of these surface points
can be calculated using a triangulation algorithm. The principle of projection light
reconstruction is that projecting a narrow band of light onto a three-dimensionally
shaped surface produces a line of illumination that appears distorted from perspec-
tives other than the perspective of the projector, and can be used for an exact geo-
metric reconstruction of the surface shape. For this reason the camera is rotated by
a certain angle with respect to the axis of the projector. The modern systems also
use multiple cameras to increase the precision of acquisition by averaging the co-
ordinates of the point acquired. Usually patterns with many stripes at once, or with
arbitrary fringes, are used in order to have a fast acquisition of the whole surface.
This initial phase, not considering calibration where the reference system of the
projector and the CMM are to be roughly aligned, is very fast. A simple scheme of
the process is shown in Figure 6.2.
Another approach used with DLP projectors is to use a light stripe scanning
method, which sequentially projects a single light stripe, or more frequently a
pattern of stripes in order to save time, onto the object surface with the aim of
exploring the whole product. The advantage of this approach is that there are no
moving parts in the acquisition system and all the scanning is controlled by soft-
ware. Moving mechanisms often introduce errors and usually complicate the
whole system (Yang et al. 1995).
The other most used type of sensor is the laser scanning head. This system is
based on the projection of a laser line (sometimes also only a dot) onto a surface
and the acquisition of the resulting image using a CCD camera. The camera has to
be positioned in a known position with respect to the laser source in order to cal-
culate the position of the surface points. Depending on how far away the laser
light collided with the surface, the laser line (or dot) appears at different places in
the camera’s field of view. This technique is called triangulation because the laser
line, the camera, and the laser source form a triangle. The triangulation evaluates
the position of each point of the laser line because the distance and the angle be-
tween the camera and the laser source are known. Given the position of the pixels
on the CCD camera that see the projected line, it is possible to reconstruct the
position of the whole curve where the line is projected. As for the DLP approach,
this method is very fast but it is surface-dependent. For reflective surfaces the
projected light may not be seen, and for translucent surfaces the signal could be
very noisy; only for Lambertian surfaces the method provides very good results
(Forest and Salvi 2002; Zussman et al. 1994). Often the products to be measured
with this approach are painted with a matt white paint. An example of the behavior
of laser light on different surfaces is shown in Figure 6.3.
Figure 6.3 Effect of the surface type on the laser light reflection: a specular surface, b Lamber-
tian surface, and c translucent surface
194 G. Campatelli
Both of these systems use CCD cameras for the image acquisition. When this
solution is adopted, it is necessary to evaluate and consider lens distortion. The
lens distortion is a characteristic feature of all real optical systems. The lens distor-
tion can be evaluated using different models (Weng et al. 1992); the basic distor-
tion components are defined as radial distortion, decentering distortion, and thin
prism distortion, but more complex models are also available. For the lenses that
are suitable for machine vision, the radial distortion is usually much more signifi-
cant than the other two types of distortion. Therefore, usually only this term is
included in the lens distortion model. This model can be evaluated and the vision
of the CCD corrected by acquiring and processing a known image, usually a
chessboard. Similar corrective actions are developed for the lens and DLP system
of the projector (Chen et al. 2009; Zexiao et al. 2007).
The starting problem for the integration of the data from multiple sources is given
by the need for data registration in order to obtain a spatial “closeness” among the
same points acquired by two, or more, measuring systems. The techniques used to
fuse together close points acquired also at different resolutions will be treated in
the following section, but each method applied needs a first preprocessing step of
registration in order to provide reliable results.
Each measuring instrument provides a cloud of points, sometimes obtained us-
ing a regular grid but often with an irregular pattern, that have to be related to, and
later integrated with, the output of the other acquisition instruments. The charac-
teristics of the optical spatial data, which are usually unordered, with fixed density
and independent of the surface curvature, make this integration nontrivial.
The main problem is that the different acquiring instruments nearly always use
different spatial references (origin and orientation of the axes). Even if the part is
measured on a single measuring platform carrying both a touch and a laser probe,
independent calibration is usually needed for each probe, resulting in a misalign-
ment.
The misalignment among the datasets is usually small because, during the cali-
bration of the tools, it is normal practice to refer the acquiring devices to the same
spatial coordinate system. However, also using this precaution, a small misalign-
ment is always found among the many datasets. The misalignment is very impor-
tant because the numerical algorithms to achieve an alignment are different if one
starts from two or more clouds of points that are heavily misaligned or if one starts
from clouds that are nearly aligned. The algorithms to be used in the first case are
based on pattern recognition or on the evaluation of the moments of inertia of the
cloud; these approaches are very fast but provide only a rough alignment of the
clouds. In the second case, the one that is usually applied in the case of multisen-
sor measuring machines such as the OCMM, the approaches are mainly based on
the popular iterative closest point (ICP) algorithm.
6 Geometric Tolerance Evaluation Using Combined Vision – Contact Techniques 195
The ICP algorithm proposed by Besl and McKay (1992) is a standard solution
to the alignment problem. The ICP algorithm has three basic steps, as reported
below considering P and M as two set of points to be aligned:
1. Pair each point of P to the closest point in M.
2. Compute the motion that minimizes the mean square error (MSE) between the
paired points.
3. Apply the optimal motion to P and update the MSE. Some iterations are needed
to obtain a satisfactory solution.
The ICP algorithm is a very popular algorithm, so many variants have been
proposed over the years. These variants can be classified according to how the
algorithms:
• select subsets of P and M;
• pair points;
• weight the pairs;
• reject some pairs;
• assign the error metric; and
• minimize the error metric.
Many modifications of the ICP algorithm have been proposed in order to im-
prove the robustness of the algorithm. In the case of noisy clouds of points, the
ICP algorithm sometimes provides faulty results. The source of the problem is that
the original algorithm assumes outlier-free data and P being a subset of M; this
implies that each point of the first dataset has a corresponding point in the second
dataset. Therefore, this approach requires M has a larger number of points than P,
as usually happens using optical and touch measuring machines.
Many attempts have been conducted in order to make the ICP algorithm more
robust by rejecting wrong pairs. An example of this strategy can be found in the
works by Pulli (1997) and Rousseeuw and Van Zomerman (1990) that introduced
two algorithms called least median of square (LMedS) and least trimmed square
(LTS). The general idea is to make the linear regression used to evaluate the least
mean square insensitive to outliers.
Consider the expression
n m n
yi = β 0 + ∑∑ β j xij + ∑ ei = β 0 + β1 x11 + … + β m xn m + e1 + … + en (6.1)
i =1 j =1 i =1
for i = 1,…, n. Here yi are the response variables, xii to xim the explanatory vari-
ables, and βj for j = 0, …, m the coefficients (the model parameters). In the classi-
cal regression theory the errors ei are assumed to have a Gaussian distribution with
zero mean. The standard least (mean) squares method computes the parameters βj
such that the sum of the squares of the residuals ri = f (β0,…,βm) is minimal.
A simple but useful measure of robustness for the proposed variant to the ICP
algorithm is the breakdown point, that is, the fraction of the dataset that can be
contaminated with outliers before the algorithm crashes (wrong results). Both the
196 G. Campatelli
LTS and the LMedS have a breakdown value of 50%. On the basis of the study of
the breakdown values of the algorithms, research into a new solution has started to
consider with more care the statistical properties of the method adopted.
Among the many solutions developed over the years, it is worth mentioning the
algorithm proposed by Chetverikov et al. (2005), which in the application to real
registration problems has proved its robustness and usefulness. This is a good
model to explain the “trimmed” approach of the ICP algorithm used by many
other authors. The algorithm proposed is called the trimmed ICP (TrICP) algo-
rithm. The advantages of the proposed TrICP algorithm are its robustness to noise
and to poor initial estimation.
The variant with respect to the original ICP algorithm is mainly due to the in-
troduction of an overlap function (ξ), the fraction of the points of the smaller set
(P) that have to be paired, which can be computed automatically by the algorithm
with some iteration. The steps of the TrICP algorithm are as follows:
1. For each point of P; find the closest point in M and compute the individual
distances of each point to the paired one.
2. Sort the individual distances in increasing order, then select the ξ lowest values
and calculate their sum (STS).
3. If the value of the MSE or number of iterations has reached the stopping value,
stop the algorithm, otherwise update the value of STS and continue.
4. Compute for the selected pairs the optimal Euclidean motion that minimizes
STS.
5. Transform P according to the motion and return to step 1.
On the other hand, data fusion has the aim of creating a complementary acqui-
sition of surface data by multiple sensors. The difference from the serial data in-
tegration is that the resultant dataset is a fusion of the data acquired both by the
low-resolution and by the high-resolution sources. Actually there are only a few
applications of data fusion in the metrology field, while this approach is exten-
sively used in other fields of research such as the nondestructive testing, geospatial
applications, weather forecasting, and airplane control.
Data fusion is generally defined as the use of techniques that combine data
from multiple sources and gather that information in order to make inferences,
which will be more efficient and potentially more accurate than if they were made
by means of a single source. Data fusion is a low-level fusion process. Fusion
processes are often categorized as low, intermediate, or high, depending on the
processing stage at which fusion takes place. Low-level fusion combines several
sources of raw data to produce new raw data. The expectation is that fused data
are more informative and synthetic than the original inputs.
From this definition it is possible to evaluate the differences among serial data in-
tegration and data fusion. Serial data integration has the main advantages of simplic-
ity of application, the deterministic choice of the precision of the result (given by the
second measuring tool to be used), and the possibility of creating general-purpose
systems to analyze a large variety of products. On the other hand, the use of serial
data integration has a lower efficiency than the use of data fusion because some of
the acquired data are discarded during the process and the geometric information has
to be reacquired with a slower high-precision system. Data fusion, however, has
some major problems: the main one is the inability to determine a priori the preci-
sion of the information for a specific measurement. The limit of the precision that
can be reached, for classical application of data fusion that does not acquire replicas
of the measurement, is always a fraction of the precision of the better performing
device used. Moreover, the data fusion cannot be used for general-purpose meas-
urement but, at the current state of the art, the strategy and the following characteris-
tic parameters have to be evaluated for each specific case. The most used techniques
for data fusion are the neural network, the fuzzy logic, and, specifically for the meas-
urement field, the Bayesian approaches. These limitations, and the greater simplicity
of the serial data acquisition approaches in terms of practical applications, have
undermined the development of “pure” data fusion approaches in the field of indus-
trial measurement. Nowadays, many OCMMs work with a serial data integration
approach, while data fusion is the preferred strategy in other fields, such as geo-
graphical and environmental measurements. However, hybrid models have been
proposed in order to implement some of the advantages of both approaches. These
hybrid models provide a resultant dataset that takes the information mainly from the
touch probes but partly also from the optical sensors. These approaches are based on
the implementation of advanced feature recognition strategies based on the output of
the optical sensors. Currently they are also providing promising results, albeit lim-
ited only to applications with a small number of specific features. In the following
sections more detailed descriptions of some exemplary implementations of the serial
and hybrid data integration and data fusion approaches are reported.
198 G. Campatelli
Serial data integration is used both for reverse modeling of free-form surfaces and
for inspection of geometric tolerances. In the first case the optical sensors have the
aim of acquiring the general dimensions and boundaries of the sample in order to
automatically generate and optimize the path of the touch probe. In the second
case the aim of the optical sensors is to locate a sample of known geometry inside
the measuring range and execute a preset trigger probe tolerance inspection. In
both cases the main advantage of a serial approach is the reduction of the time
needed for the geometry/tolerance acquisition of the sample. In the traditional
touch-probe-only approach there is a relevant manual activity that can be auto-
mated using a serial approach. The results in terms of performances are very inter-
esting; some researchers (Chen and Lin 1997) reported a reduction in the time
needed for the reverse modeling, maintaining the same precision of traditional
approaches, of about 85–90%.
The general scheme of a serial data integration approach is always based on
four simple steps:
arm and a trigger touch probe. The limitations of this approach are that the recon-
structed surface must satisfy two conditions:
The first condition is essential for the edge detection algorithm to work prop-
erly, the second is necessary for the following digitization by a touch probe. The
approach is constituted by two main steps:
Final Model
In the first step, a vision system is used to detect 3D surface boundaries by us-
ing a 3D stereo detection method based on multiple image analysis (stereo vision).
Obviously, before image processing the images are filtered in order to reduce
noise to an acceptable level. Then a classical Laplacian edge detection algorithm is
used to extract the surface boundaries effectively. Later the boundaries are re-
duced to a number of characteristic points using a data reduction strategy in order
to create a first dataset that can be processed with a constrained Delaunay triangu-
lation algorithm.
The geometric features extracted from the edge detection algorithm applied to
the acquired image normally contain many redundant points which are not re-
quired for the initial surface triangulation. To achieve a reduction in terms of the
number of points, an algorithm is used that computes the relative angle between
two consecutive points that have been obtained. Only when the angle between the
approximating curve of the sequence of points until the selected point and the
segment that links this point to the following point is greater than a specific
threshold, usually 20° is used, the point is considered in the dataset. This algo-
rithm was developed by Fujimoto and Kadya (1993). For each point chosen the
vertical coordinate is evaluated by using stereo-vision techniques. This approach
provides only approximate values (tolerance of 0.1–0.3 mm), but it is enough for
the rough model of the product. The next step is the development of a Delaunay
triangulation based on the boundary points chosen. Other points internal to the
boundary are chosen automatically in order to reduce the deformation of the trian-
gles using the Delwall algorithm. This is a very fast step because this first model is
created on the basis of only a few points. For all the vertices of the internal points
of this first triangulation the vertical coordinate is acquired, either using again
Figure 6.5 Digitizing of a ceramic plate: a boundary points, b boundary schematization, c first
triangulation, and d refined triangular model
6 Geometric Tolerance Evaluation Using Combined Vision – Contact Techniques 201
The second approach is intrinsically different from the VAREA because in this
case the data acquired with both vision and a touch probe are organized in order to
create a CAD model based on Bézier curves. This is a more complex representa-
tion with respect to the triangulation approach and it allows one to reconstruct also
surfaces that are not always visible from a single direction like in the first case.
This approach is also a serial one, where the points obtained with the optical probe
are finally discarded and the object model is based only on the higher-accuracy
touch probe data; however, the information from the optical device is considered
to create a stopping condition and to define if the model has to be refined during
the acquisition steps.
Briefly, the method can be summarized in the following steps. The first step is
represented by the optical digitization of the object using many clouds of points.
The clouds are merged together using the “bandwidth approach”, which allows for
an easy reconstruction of the surface based on Bézier curves. Basically the cloud
of points rough model is somehow sliced by many bands and each band consti-
tutes the basis to create a curve. This first rough model is used to define the tool
path for the touch probe depending on the surface curvature (Yau and Menq
1996). Then the process is iterated until the error between the actual curve and the
one computed in the previous step is under a user-defined threshold. Usually one
or two iterations are enough for most surfaces, while a higher number of iterations
could be needed for complex shapes. The scheme of the approach, as presented by
Carbone et al. (2001), is shown in Figure 6.6.
202 G. Campatelli
Physical object
Vision system
Point acquisition digitization
Rough Model
CMM
Evaluation No digitization and
Error <tol inspection
Yes
Individual surface entities
Final Model
The data obtained from the structured light processing projector are then fil-
tered thanks to the choice of arbitrary bands in which to group the points. The
bands are chosen starting from the definition of the reference plane, the ordering
direction of the cloud of points, and the bandwidth between the arrays of points.
The enveloping direction of the bands is perpendicular to the defined ordering
direction and relates to the reference plane. The data are divided into bands and
the points in the same band are ordered along the ordering direction. After that, the
direction of adjacent curves is alternately changed.
An associated style curve is computed for each band using the Bézier represen-
tation. The style curve is obtained by filtering the huge number of points of the
band with a threshold filtering angle similar to the one already used by the
VAREA. From the filtered points the Bézier spline is automatically calculated,
taking care to segment the curve where a discontinuity arises. The rough model is
constituted by nearly parallel Bézier curves (the bands are parallel, while the 3D
curves do not have this constraint, but they only have to be inside the associated
band) that can be processed in order to create a rough surface schematization of
the object. This aggregation into Bézier surfaces is usually carried out automati-
cally by many commercial CAD software packages. The final accuracy of this
model is usually of the order of 0.5 mm, enough to generate a collision-free touch
probe tool path.
The touch probe acquisition process, which includes the definition of the meas-
urement sequence, the number of measurement points, the number of probes, and
their configuration, is planned using an algorithm that considers the curvature of
the surface and defines the density of points to be acquired accordingly. Surfaces
are then digitized using the touch probe, and the measurements are carried out until
the stopping condition is found. This condition is given by the achievement of a
specific user-defined value of the deviation between nominal points (those on the
reconstructed Bézier surface) and actual points (those acquired by the CMM). If
this value is greater than the threshold, a new set of points is acquired and the sur-
faces are iteratively reconstructed. In this case the first step of the iteration uses the
surfaces generated by the structured light acquisition as references, while the fol-
lowing steps use the surface generated by the points acquired by the touch probe.
The geometric approaches are hybrid approaches that consider the information
from the optical and the touch probes for the final digitization of the surfaces but
do not apply a real fusion process. The hybrid approaches are used mainly for
reverse modeling and scarcely for tolerance verification. This is because for the
tolerance verification the information regarding the orientation of the sample,
which a simpler serial approach can also provide, is deemed sufficient. In the
hybrid approaches the information from the two sources is merged together in
order to create a final representation of the product that is constituted mainly by
204 G. Campatelli
the points acquired by the touch probe but where the geometric features are evalu-
ated thanks to the vision system. In this case higher-level information is extracted
from the vision system with respect to the other approaches that also includes also
geometric information. These approaches provide an advantage in terms of the
time needed for the analysis of the sample with respect to the serial approaches
when the surface can be simplified as an assembly of basic features, such as in the
case of mechanical products; there is no advantage when the surface is a sculp-
tured free form. In Figure 6.7 two examples of a sculptured surface and a me-
chanical product are shown, where the geometric approach has higher efficiency.
Many multisensor approaches have been developed on the basis of the geometric
feature evaluation; the one that it is explained in detail in this chapter was devel-
oped at Columbus University by Shen et al. (2000). It was chosen because of the
high degree of integration of the solution proposed and it could be considered, in
its general definition, as a representative scheme for all similar approaches.
The multiple-sensor integrated system developed consists of three main subsys-
tems:
1. a CMM with an active 3D vision system and a touch probe;
2. an information aggregation module for data fusion based on feature recogni-
tion; and
3. an inspection planning algorithm for the final touch probe surface digitization.
The vision system chosen is constituted by a DLP projector and a CCD camera.
The DLP projector is one of the most used solutions for surface digitization thanks
to the flexibility and high precision that this system can provide, especially when
compared with the older technology of LCD projectors. The information aggrega-
tion module responsible for the data aggregation analyzes the acquired coordinate
data from the 3D vision system to extract the surface geometric information of the
object. This module performs three main tasks:
6 Geometric Tolerance Evaluation Using Combined Vision – Contact Techniques 205
Points Cloud
Neighbor searching
Data grouping
Patch Grouping
The extracted features are then grouped on the basis of the content of the in-
formation:
1. complete feature information;
2. incomplete feature information; and
3. unavailable feature information.
For complete feature information the CMM can be instructed to acquire the
points needed to digitize the surface features with a touch probe that permits
higher accuracy with respect to the optical system. If the feature information is
incomplete or unavailable for the surface reconstruction, other optical acquisitions
are needed in order to create a low-level geometric model useful for planning the
touch probe acquisition. An example of complete and incomplete data is shown
for a mechanical product in Figure 6.9.
The touch probe path is defined on the basis of the geometric feature: planar,
circle, free form, etc. For each type of feature an optimal strategy can be planned
depending on the precision that the measurement or the reverse modeling has the
aim of achieving.
Similar to the geometric reasoning for the sequence of phases, but very different in
terms of the algorithms used, is the self-organizing map (SOM) feature recognition
approach proposed by Chan et al. (2001). The relevant aspect of the proposed ap-
6 Geometric Tolerance Evaluation Using Combined Vision – Contact Techniques 207
proach is the use of an advanced neural network, a Kohonen SOM, for the surface
segmentation and feature extraction. This approach is useful because its further
implementation could be the basis for a real multisensor data fusion (Section 6.6) at
least for simple features such as holes. The use of a neural network for feature recog-
nition is widely spread in medical science, especially for the study of tumors, lesions,
and other abnormalities in medical images. The general idea is to use only a CCD
camera to acquire an image of the object and to extract the feature using the SOM.
The patches are created by the study of the image color and intensity gradient.
A competitive learning network, the SOM, consists of n layers of 2D arrays of
neurons; usually a 5 × 5 strategy is considered a good compromise between data
processing time and precision. Each neuron is connected to its immediate neighbor
on its own layer and to (n – 1) neurons on the layers immediately below and above
its location. The input neurons are on the first layer and receive the input from the
CCD image. For each iteration of the SOM, the strength of the neurons on each
layers is measured and for each input dataset it is possible to evaluate which layer
has the greatest strength; this is the winning layer and the pixel associated with the
5 × 5 dataset is associated with this layer. Each layer represents a possible patch,
so the SOM has to be designed with a number of layers compatible with the char-
acteristic of the geometry that has to be acquired; the number of layers has to be
greater than the number of possible patches that the SOM has to recognize.
Given the patch of the object, it is possible to search for simple feature such as
holes. The routine proposed is based on the evaluation of the external and internal
boundaries of each patch. The internal boundaries are then processed in order to
evaluate if the boundary is associated with a hole or another type of feature. On
the basis of this analysis, a specific CMM acquisition process is defined automati-
cally. The scheme for the hole recognition, as presented by Chan et al. (2001), is
shown in Figure 6.10.
SOM defined
patch
Is Std. Dev.
Radius Yes
Label as hole
<15%
radius?
No
Label as other
feature
The integration approaches presented in the previous sections use the data from the
lower-resolution system, usually the optical system, in order to optimize and auto-
mate the touch probe acquisition. The higher-resolution data of the touch probe sim-
ply overwrite the data already acquired with the other systems. Data fusion has the
aim of integrating together the measurements from the single acquisition systems in
order to obtain a measurement that has higher precision than each single dataset.
The data fusion approach applied to a hybrid CMM is still a new field of research.
While multisensor data fusion is widely used in many fields such as weather forecast-
ing, medical science, acoustics, and nondestructive testing, very few attempts have
been made to integrate together the data from different sources in the mechanical
measurement field. Huang et al. (2007) stated that “Multi sensor data fusion (MDF)
is an emerging technology to fuse data from multiple sensors in order to make a more
accurate estimation of the environment through measurement and detection” and
“Although the concept is not new, MDF technology is still in its infancy”.
The difficulty in applying the data fusion approaches already used in other
fields is related to the differences among the variables involved. The use of data
fusion in weather forecasting, for example, takes into consideration factors such as
air humidity, soil and air temperature, and wind velocity in order to predict the
probability of rain. The fusion of data from many sensors adds a variable to a
problem that is usually underdefined, and with an output that is different from the
input. In this case the data fusion has the aim of creating a model to link together
complementary information on the same process. The methods used for this data
fusion have to create a model to link multiple parameters to a different output. The
data fusion has to find a model to approximate an unknown model. For this reason
the approaches used are usually without a predetermined model such as fuzzy
logic (Chen and Huang 2000) or neural network (Huang et al. 2007).
The case of mechanical measurement is completely different. In this case the
inputs are all measurements of the product and so is the result. Moreover, in
Soil temperature
Air temperature
Weather
Probability of rain
Wind speed forecating
Humidity
xyz measure
CMM measure xyz measure
xyz measure
Figure 6.11 The data fusion model for weather forecasting and a mechanical measurement
6 Geometric Tolerance Evaluation Using Combined Vision – Contact Techniques 209
order to improve the precision of the data, both inputs have to be referred to
the same spatial position. The strategy used to solve the data fusion process in
the mechanical measurement field is to create a linkage model between the
high-resolution data and the low-resolution data. Such a model would allow the
prediction of the high-resolution data starting from the low-resolution data.
From such a model it would be possible to obtain measurements with the same
spatial coordinate from the low-resolution and high-resolution data that could
be fused together. A proposal for such model was made by Xia et al. (2008a),
based on a Bayesian hierarchical model, and it is explained briefly in the fol-
lowing paragraphs.
The difference between input and output for weather forecasting and measure-
ment using a hybrid CMM is reported in Figure 6.11.
Recent studies on the problem of synthesizing spatial data collected at differ-
ent scales and resolutions by Ferreira et al. (2005) and on the problem of cali-
brating computer simulation models of different accuracies by Kennedy and
O’Hagan (2001) and Qian and Wu (2006) have proven that multiresolution data
can be fused together, once a fit linkage model has been chosen. These studies
however do not consider the misalignment issue, which is a fundamental aspect
for multiresolution data in the field of mechanical measurements.
The approach proposed by Xia et al. (2008a) was called a Bayesian hierarchical
model by the authors. This starts from the modeling of the low-resolution data as a
Gaussian process – this was already proven to be a valid approach for manufac-
tured parts by Xia et al. (2008b) – and then the low-resolution and high-resolution
data are aligned with a variant of the TrICP algorithm (Section 6.2) in order to
create a linkage model represented as a K kernel function. The linkage model is
then used to fuse together the low-resolution and high-resolution data thanks to a
Bayesian prediction. The general scheme of the approach, as presented by Xia
et al. (2008b), is shown in Figure 6.12.
Modeling the low-resolution data as a Gaussian process means that the value of
each measurement acquired for coordinate xi is composed of a mean value and an
error:
yl ( xi ) = ηl ( xi ) + ε l ( xi ) , (6.2)
where yl (xi) is the acquired low-resolution data, ηl (xi) is the actual value, and
εl (xi) is the error. As in the Gaussian process, the mean value is made up of two
components, the mean component and the covariance function:
where k is the variance and R the correlation function, which can be expressed as
d
R ( xi , x j ) = ∏ exp ⎡⎢ −ν k ( xki − xkj ) ⎤⎥ ,
2
(6.4)
k =1
⎣ ⎦
where d is the dimension of the input variable and νk are the scale parameters that
control the “damping” of the correlation with respect to the distance from xi. The
modeling of the low-resolution data with a Gaussian process is useful for the defi-
nition of the kernel function, which is necessary for the creation of a linkage
model between the high-resolution and the low-resolution data.
After the modeling, the two datasets are aligned in order to create a linkage
model. For the precise alignment the algorithm proposed is a small variation of the
TrICP algorithm presented in Section 6.2. The general idea is to perform the
alignment using only a subset of the data. The points considered are the whole
high-resolution dataset and an iteratively chosen group of low-resolution data. The
iterations of the algorithm that choose the new dataset at each iteration are carried
out until there is no improvement in the objective function, defined as in the TrICP
algorithm.
The neighborhood linkage model has a double objective: to create a linking
model between low-resolution and high-resolution data and to enable a fine
alignment with greater precision with respect to the previous one. The first align-
ment carried out with the TrICP algorithm is necessary to give a first approximate
alignment for the creation of a reliable linking model. The linking model is consti-
tuted by an adaptive kernel function:
l
yh ( x j ) = α1 ∑ K ( x j , xi )ηl ( xi ) + α 0 + e ( x j ), (6.5)
i =1
where yh(xj) is the high-resolution data, α0 and α1 are the scale and location coef-
ficients of the alignment, e is the residual, and K(xj, xi) is the kernel function de-
fined as
⎡ d ⎛ xki − xkj ⎞ ⎤
K ( xi , x j ) = D ⎢ ∑ ⎜ ⎟⎥ , (6.6)
⎣⎢ k =1 ⎝ λk ⎠ ⎥⎦
6 Geometric Tolerance Evaluation Using Combined Vision – Contact Techniques 211
with D(t) = (1 – t 3/2)3 if t < 1 or 0 if t > 1 and λ the size of the neighborhood auto-
matically assigned and controlled by the kernel.
In order to fuse together the data, a prior distribution p(θ) is defined as the
product of the prior distributions of the individual parameters. From the prior it is
possible to define a Bayesian predictor that integrates the multiresolution data in
order to provide the value of yh(x0) at the input location x0 considering the
neighbor data from the low-resolution and high-resolution datasets. The predictor
can be defined as
The introduction of OCMMs in the field of product inspection and reverse model-
ing has brought a great advance in measurement technologies. The advantages are
mainly related to the possibility to automate the entire process and to obtain a
result in ever-decreasing processing time. Most of the approaches developed in the
last decade are based on a sequential acquisition of data: a fast acquisition of low-
resolution data is done first, followed by a more time-consuming acquisition of
higher accuracy data. This second acquisition can be optimized thanks to the
analysis of the low-resolution dataset; many approaches that have been developed
are also able to create the CMM operation code for the high-resolution data acqui-
sition with almost entirely automated procedures. This field of research is ex-
tremely active. Apart from continuous work to improve the alignment algorithms,
the new frontier of the research is to shift from a sequential approach to a real data
fusion in order to save even more time and to improve the precision of the final
dataset. In the case of data fusion, a general model capable of handling various
applications is very difficult to achieve. For a specific area of research and appli-
cation, a specific data fusion model must be built. Recent activities in this field
212 G. Campatelli
have shown encouraging results that, we hope, will motivate more and more re-
searchers to develop new approaches and seek the automation of the data fusion
process in the field of CMMs.
References
Besl P, McKay N (1992) A method for registration of 3-D shapes. IEEE Trans Pattern Anal
Mach Intell 14:239–256
Carbone V, Carocci M, Savio E, Sansoni G, De Chiffre L (2001) Combination of a vision system
and a coordinate measuring machine for the reverse engineering of freeform surfaces. Int J of
Adv Manuf Technol 17:263–271
Cernuschi-Frias B (1984) Orientation and location parameter estimation of quadric surfaces in
3-D space from a sequence of images. PhD dissertation, Brown University of Providence
Chan V, Bradley C, Vickers W (1997) Automating laser scanning of 3-D surfaces for reverse
engineering. SPIE 3204:156–164
Chan VH, Bradley C, Vickers GW (2001) A multi-sensor approach to automating co-ordinate
measuring machine-based reverse engineering. Comput Ind 44:105–115
Chen LC, Lin GCI (1991) A vision-aided reverse engineering approach to reconstructing free-
form surfaces. Robot Comput Integr Manuf 13(4):323–336
Chen LC, Lin GCI (1997) Reverse engineering of physical models employing a sensor integra-
tion between 3D stereo detection and contact digitization. Proc SPIE 3204:146–155
Chen YM, Huang HC (2000) Fuzzy logic approach to multisensor data association. Math Com-
put Simul 52:399–412
Chen X, Xia J, Jin Y, Sun J (2009) Accurate calibration for a camera–projector measurement
system based on structured light projection. Opt Lasers Eng 47(3–4):310–319
Chetverikov D, Stepanov D, Krsek P (2005) Robust Euclidean alignment of 3D point sets: the
trimmed iterative closest point algorithm. Image Vision Comp 23:299–309
Daniels J, Ha HK, Ochotta T, Silva CT (2007) Robust smooth feature extraction from point
clouds. In: Proceedings of IEEE international conference on shape modeling and applications
El-Hakim SF, Pizzi N (1993) Multicamera vision-based approach to flexible feature measure-
ment for inspection and reverse engineering. Opt Eng 32:2201–2215
Ferreira M, Higdon D, Lee H (2005) Multi-scale random field models. ISDS discussion paper
05-02. http://ftp.stat.duke.edu/WorkingPapers/05-02.html
Forest J, Salvi J (2002) A review of laser scanning three dimensional digitizers. Intell Robot Syst
73–78
Fujimoto M, Kadya K (1993) An improved method for digitised data reduction using an angle
parameter. Measurement 12:113–122
Hornbeck LJ (1998) From cathode rays to digital micromirrors: a history of electronic projection
display technology. Texas Instruments Technol J 15(3):7–40
Huang J, Menq CH (1999) Feature extraction from incomplete surface data. Coordinate Metrol-
ogy Measurement Laboratory technical report CMML-99-03
Huang YB, Lan YB, Hoffmann WC, Lacey RE (2007) Multisensor data fusion for high quality
data analysis and processing in measurement and instrumentation. J Bionic Eng 4:53–62
Kennedy M, O’Hagan A (2001) Bayesian calibration of computer models. J R Stat Soc Ser B
63:425–464
Lee N L, Menq CH (1995) Automated recognition of geometric forms from B-rep models. In:
Proceedings of ASME international computers in engineering conference, pp 805–816
Pulli K (1997) Surface reconstruction and display from range and color data. PhD thesis, Univer-
sity of Washington
6 Geometric Tolerance Evaluation Using Combined Vision – Contact Techniques 213
Qian Z, Wu CFJ (2006) Bayesian hierarchical modeling for integrating low-accuracy and high-
accuracy experiments. Technometrics 50(2):192–204
Rousseeuw P, Van Zomeren B (1990) Unmasking multivariate outliers and leverage points.
J Am Stat Assoc 85:633–651
Nashman M (1993) The use of vision and touch sensors for dimensional inspection tasks. Manuf
Rev 6(2):155–162
Nashman M, Hong TH, Rippey WG, Herman M (1996) An integrated vision touch-probe system
for dimensional inspection tasks. In: Proceedings of SME Applied Machine Vision ’96 con-
ference, Cincinnati, pp 243–255
Shen TS, Huang J, Menq CS (2000) Multiple-sensor integration for rapid and high-precision
coordinate metrology. IEEE/ASME Trans Mechatronics 5(2):110–122
Spitz SN, Spyridi AJ, Requicha AAG (1999) Accessibility analysis for planning of dimensional
inspection with coordinate measuring machines. IEEE Trans Robot Automat 15(4):714–727
Weng J, Cohen P, Herniou M (1992) Camera calibration with distortion models and accuracy
evaluation. IEEE Trans Pattern Anal Machine Intell 14:965–980
Xia H, Ding Y, Mallicky BK (2008a) Bayesian hierarchical model for integrating multi-
resolution metrology data. In: Proceedings of European Network for Business and Industrial
Statistics conference
Xia H, Ding Y, Wang J (2008b) Gaussian process method for form error assessment using coor-
dinate measurements. IIE Trans 40(10):931–946
Yang J, Lee NL, Menq CH (1995) Application of computer vision in reverse engineering for 3D
coordinate acquisition. In: Proceedings of the symposium on concurrent product and process
engineering, ASME international mechanical engineering congress and exposition, pp 143–
156
Yau HT, Menq CH (1996) A unified least-squares approach to the evaluation of geometric errors
using discrete measurement data. Int J Mach Tools Manuf 36(11):1269–1290
Zexiao X, Jianguo W, Ming J (2007) Study on a full field of view laser scanning system. Int J
Mach Tools Manuf 47:33–43
Zussman E, Schuler H, Seliger G (1994) Analysis of the geometrical feature detectability con-
straints for laser-scanner sensor planning. Int J Adv Manuf Technol 9:56–64
Chapter 7
Statistical Shape Analysis
of Manufacturing Data
7.1 Introduction
In statistical shape analysis (SSA) the shape of an object is defined as all the in-
formation of the object that is invariant with respect to similarity transformations
on the Euclidean space (rotations, translations, and dilations or changes of scale).
The goal of SSA is to analyze the shapes of objects in the presence of random error.
Analysis of shapes in manufacturing is critical because geometric tolerances
(specifications) of roundness, flatness, cylindricity, etc., need to be inspected,
controlled, or optimized based on a cloud of two- or three-dimensional measure-
ments taken on the machined surfaces of the part. These tasks are even more com-
plex if the part geometry has a “free form”, i.e., there is no standard geometric
construction that can represent the shape, a situation common in advanced manu-
facturing applications such as in the aerospace sector.
__________________________________
E. del Castillo
Department of Industrial and Manufacturing Engineering, The Pennsylvania State University,
University Park, PA 16802, USA,
e-mail: exd13@psu.edu
215
216 E. del Castillo
Over the last 20 years, SSA techniques have been developed and applied in
many areas of the natural sciences where there is interest in characterizing differ-
ences between and variability within shapes, e.g., biology, paleontology, and ge-
ology. A considerable intersection of ideas also exists with image and pattern
recognition in computer science. In particular, SSA is known as geometric mor-
phometrics in biology, and the type of techniques developed in the past two dec-
ades has been called the “morphometrics revolution” by some authors (Adams
et al. 2004) given the success SSA had over previous techniques used to analyze
shapes. For more on the history and foundations of SSA with applications in bio-
logical sciences we refer readers to the books by Dryden and Mardia (1998) and
Zelditch et al. (2004).
Our interest in shape analysis stems in part from recent interest on “profile
analysis” in the field of statistical process control (SPC) (Kang and Albin 2000;
Colosimo et al. 2008; Woodall et al. 2004) (although we do not discuss SPC based
on shape analysis in this chapter, this is certainly another potential area of research
where SSA ideas can be used). In profile-based SPC, a parametric model is sought
that describes the form that the response follows with respect to some variable of
interest (in essence, one performs functional data analysis). The parameters of this
model are fitted on the basis of process data and then multivariate SPC methods
are applied to the estimated parameters.
By working with the shape directly, SSA techniques avoid the parametric
model definition step, allow complicated shapes to be studied, and simplify the
presentation of results. In SSA, one works with the whole shape of the object, so
the geometry is not “thrown away” (Dryden and Mardia 1998).
The remainder of this chapter is organized as follows. Section 7.2 describes
methods to solve the landmark matching problem, which occurs when two objects
have point labels that do not correspond to each other. In Section 7.3, we review
the main ideas of SSA based on the so-called Procrustes method. In this section,
the notions of shape space, the generalized Procrustes algorithm (GPA), and tan-
gent space coordinates are discussed. The chapter concludes with a discussion of
other shape analysis techniques, including areas for further work.
In most of SSA, the main goal is statistical inference with shapes, in particular, to
test if two or more objects have an equal shape or not, or to determine the direc-
tions in which most of the variability of a shape occurs. Some other authors’ main
interest (e.g., in biology) is to describe how shapes of objects (e.g., species of
animals) change with time. In our case, the main goal is to study the shape of
manufactured parts.
The techniques considered herein are based on shape data obtained by measuring
the parts at specific landmarks, points of special interest or unique characteristics.
In order to be amenable to data analysis, landmarks should refer to homologous
7 Statistical Shape Analysis of Manufacturing Data 217
called the “context” of point i by these authors. The idea then is to match those
points between two different shapes that have the most similar “contexts”. For this
purpose, define the cost of matching point i in part 1 and point j in part 2 to be
2
L S ⎡ h ( l, s ) − h ( l, s )⎤
Cij = ∑∑ ⎣ ⎦ , i = 1, 2,… , k , j = 1, 2,… , k ,
i j
(7.1)
l =1 s =1 h i ( l , s ) j( )
+ h l , s
which is the classic χ 2 statistic (with L ⋅ S − 1 degrees of freedom) used to test for
the difference between two distributions. Note that Cij ≠ C ji . Let B = (U , V , E ) be
a graph with two disjoint sets of points (U and V ), i.e., a bipartite graph, and a
set of edges ( E , to be decided) joining a point in U with a point in V (the
“matching” set). Define the decision variables X ij = 1 if the edge joining points
vi ∈ V and u j ∈ U is included in the matching, and X ij = 0 otherwise (Papadimi-
trou and Steiglitz 1982). Belongie et al. (2002) proposed solving the landmark
matching or labeling problem by solving the following weighted matching prob-
lem (in our notation):
min ∑ i =1 ∑ j =1 Cij X ij
k k
subject to
∑
k
j =1
X ij = 1, i = 1, 2,… , k , (7.2)
∑
k
i =1
X ij = 1, j = 1, 2,… , k ,
X ij ≥ 0, i = 1,… , k ; j = 1,… , k .
The problem is then one of linear programming, for which, as is well known,
there exist efficient algorithms. Note that the formulation does not include the con-
straints X ij ∈ {0,1} , which turn out to be redundant (the linear programming solu-
tion is always binary) so the problem is not an integer programming problem, which
would imply a considerably harder optimization problem. The matrix formulation
of the problem is based on defining the k 2 × 1 vector of decision variables
x ' = ( X 11 , X 12 ,… , X 1k , X 21 , X 22 ,… , X 2 k ,… , X k1 , X k 2 ,… , X kk )
where all empty spaces are zeroes. If matrix C = {Cij } is put into vector form as
follows
min c′x
subject to
Ax = b,
x ≥ 0,
Table 7.1 Input matrices for the “digit 3’s” problem (first four columns). The last two columns
are the output, sorted matrix
14 41 21 25 9 39
21 42 22 19 15 39
29 42 25 22 21 40
35 37 9 39 25 36
32 33 21 27 23 31
26 30 21 40 21 27
16 26 15 39 19 25
25 26 8 17 21 25
29 24 25 36 23 24
33 20 15 17 25 22
30 16 19 25 22 19
23 11 23 31 15 17
16 12 23 24 8 17
220 E. del Castillo
Table 7.2 The cost matrix C for the two digit 3’s problem. This is a k × k = 13 × 13 nonsym-
metric matrix. Bold numbers correspond to costs for the optimal matching solution
1 2 3 4 5 6 7 8 9 10 11 12 13
1 7.0 8.3 7.8 4.7 9.3 6.8 4.4 11.5 10.2 9.5 10.1 10.2 7.3
2 7.7 7.0 9.5 6.8 9.2 5.5 2.0 10.4 8.0 6.7 10.1 9.2 8.0
3 6.2 3.5 6.3 5.7 8.2 1.6 5.9 6.0 6.7 2.7 10.7 7.0 6.8
4 9.6 7.7 8.8 9.0 6.3 7.5 8.2 7.0 2.6 5.3 9.0 3.8 8.5
5 6.3 6.1 8.2 10.7 7.0 7.2 7.7 8.8 5.6 5.8 8.8 3.5 8.7
6 8.5 9.3 12.5 10.7 6.5 7.8 4.8 10.1 7.2 6.9 11.0 9.2 9.7
7 7.8 8.0 10.0 9.0 7.8 6.5 3.2 6.3 7.3 5.3 9.0 7.5 8.5
8 5.0 8.5 9.2 11.7 5.7 6.6 7.7 11.7 7.5 6.8 10.2 7.2 9.4
9 3.0 7.1 6.7 9.8 6.2 4.7 7.7 11.0 6.8 6.7 9.5 7.0 5.2
10 5.2 6.3 4.5 7.5 8.3 4.8 5.8 10.6 7.6 7.7 9.7 8.2 4.3
11 5.8 1.3 4.5 5.8 6.3 3.2 4.2 8.9 7.0 5.4 11.0 5.5 5.7
12 5.3 5.3 8.1 6.1 8.6 2.8 4.7 6.7 4.6 2.0 12.5 9.5 6.1
13 6.3 6.9 7.7 7.7 10.2 3.1 8.9 4.5 5.1 3.3 12.0 9.4 6.2
tances) and S = 9 bins were used for θ . Specifically, the bin edges where set at
[0, exp(−4), exp(−3.5), exp(−3), exp(−2.5), exp(−2), exp(−1.5), exp(−1), exp(−0.5),1]
and [−π , −3* π / 4, −π / 2, −π / 4, 0, π / 4, π / 2,3* π / 4, π ] . (This is a higher-reso-
lution histogram than that used by Belongie et al. 2002; we found the results vary
considerably with the resolution of the histogram, given by the number of bins.
Figure 7.1 Two handwritten digit “3”s, each with k = 13 landmarks. The labels of the land-
marks of the second digit were shuffled and do not correspond to those of the first digit 3
7 Statistical Shape Analysis of Manufacturing Data 221
Figure 7.2 The A matrix in the weighted matching linear programming formulation applied to
the two digit 3’s matching problem (where k = 13). Dots indicate 1’s, empty spaces indicate 0’s.
This is a 2k = k2 = 26 × 169 matrix
Figure 7.3 Optimal solution to the linear programming matching problem, digit 3 problem.
Compare with Figure 7.1
222 E. del Castillo
We point out how the two configurations in the example did not have the
same scale. We would like a matching algorithm to work even if the configura-
tions are not equally oriented (i.e., to be rotation-invariant; the method is already
location- and scale-invariant). For configurations with different orientation, a
simple solution, which we used in the previous example, is to compute the an-
gles θ in the histogram with respect to the line defined by the two closest points
to the point in question (this is a procedure similar to that suggested by Belongie
et al. 2002, who suggested using the “tangent” line to each point as the axis of
reference).
An apparently open problem in the literature is how to solve similar matching
problems when there are n shapes, not only two. A first attempt to solve such a
problem may involve matching shapes (1, 2), giving 2’ (the relabeled object 2),
then matching (2’, 3), (3’, 4), ... , (n–1’, n) and then repeating matching (n’, 1),
(1’, 2’), etc., until there is convergence. It is unknown how effective such an ap-
proach is, and whether or not convergence is guaranteed.
There is a very large body of literature on SSA techniques. Only the main precepts
and techniques are presented here. For a more thorough presentation of SSA, we
refer the reader to Dryden and Mardia (1998) and Goodall (1991) for develop-
ments up to 1998 and for more recent developments we refer the reader to Adams
et al. (2004), Kent and Mardia (2001), Klingenberg and Monteiro (2005), and
Green and Mardia (2006).
The mainstream of SSA that followed the “revolution” in morphometrics is
based on two main steps. First, the objects under consideration are registered or
superimposed with respect to each other in order to filter out rotation, translation,
and isometric scaling (dilation) effects. This is done because the objects may have
different orientations on the Euclidean space or have different locations or sizes,
and therefore their shapes cannot be initially compared. The main technique for
this task is the GPA.
An underlying assumption of the GPA is that landmarks refer to homologous or
corresponding points in each object. Since this is not always the case in CMM
data, the landmark matching problem discussed in the previous section must be
solved first before attempting the registration. Matching before registering seems
to be a simpler and better strategy than trying to jointly match the landmarks and
register the objects, as attempted in Gold et al. (1998) and Chui and Rangarajan
(2000).
Once objects have been registered, multivariate statistical methods of inference
can be performed on the projections of the shapes on the space tangent to the mean
shape. These two steps are explained below. We first give some geometric notions
necessary to understand the algorithms.
7 Statistical Shape Analysis of Manufacturing Data 223
∑∑
2
denotes the Frobenius norm of a matrix (i.e., A = i j
aij ). If we define
h j = −[ j ( j + 1)]−1/ 2 , then H is a matrix whose jth row is
(h j , h j ,… , h j , − jh j , 0,… , 0 ) for j = 1,… , k − 1 . Note that HH ′ = I k −1 and that the
j times k − j −1 times
rows of H are contrasts. Alternatively, one could start with the centered pre-
shapes, defined by Z c = H ′Z (these are k × m matrices).
The transformation in Equation 7.3 removes location effects via the numerator,
and rescales the configurations to unit length via the denominator. Since we have
not removed rotations from Z , it is not yet the shape of X , hence the name pre-
shape. The centered preshapes are equivalent to centering each coordinate of each
configuration by its centroid and dividing each by its norm.
The shape of configuration X , denoted [ X ] , is defined as the geometric in-
formation that is invariant to similarity transformations. Once location and scale
effects have been filtered as above, the shape is then defined as
[ X] = {ZΓ : Γ ∈ SO ( m )} , (7.4)
where Z is the preshape of X and Γ is a rotation matrix [i.e., a matrix such that
Γ′Γ = ΓΓ′ = I m with det ( Γ ) = +1 ] and SO (m) is the space of all m × m rota-
tion matrices, the special orthogonal group. Multiplication by a suitable matrix Γ
reorients (rotates) the object. Note that a shape is therefore defined as a set.
The following geometric interpretation of these transformations is due to Kend-
all (1984, 1989). Given that preshapes are scaled and centered objects, they can be
represented by vectors on a sphere of dimension (k − 1)m , because the numerator
in Equation 7.3 removes m degrees of freedom for location parameters and the
denominator removes one additional degree of freedom for the change of scale.
The preshapes, having unit length, are therefore on the surface of this (hyper-
spherical) space, which has (k − 1)m − 1 dimensions by virtue of being on the
224 E. del Castillo
surface of a unit sphere. As one rotates a preshape Z via Equation 7.4, the vectors
ZΓ describe an orbit, in effect, a geodesic, on the preshape space. All these vec-
tors correspond to the same shape, since by definition the shape of an object is
invariant to rotations. Thus, the orbits (also called fibers) of the preshape space are
mapped one to one into single points in the shape space, the space where shapes
reside. Two objects have the same shape if and only if their preshapes lie on the
same fiber. Fibers do not overlap. The shape space, the space of all possible
shapes, has dimension M = (k − 1)m − 1 − m(m − 1) / 2 since in addition to losing
location and dilation degrees of freedom we also lose m(m − 1) / 2 degrees of
freedom in the specification of the (symmetric) m × m rotation matrix Γ .
Example. Preshape space and shape space. In order to explain these ideas,
consider one of the simplest possible cases, where we have two lines in R2 . Thus,
we have that m = 2 and k = 2 , where the obvious landmarks are the endpoints of
the lines. After centering and scaling the two lines using Equation 7.3, one obtains
the preshapes with matrices Z1 and Z 2 . Since the original objects evidently have
the same shape (that of a line in Euclidean space), these two preshapes lie on the
same fiber or orbit, generated as the preshapes are rotated using Equation 7.4. The
shape space is of dimension (k − 1)m − 1 = 1 , namely, the circumference of a circle.
As the preshapes rotate (they can rotate clockwise or counterclockwise) they will
eventually coincide, which corresponds to the centered and scaled lines coincid-
ing. Finally, since there is a single shape, the shape space is the simplest possible,
namely, a single point [the dimension is M = (k − 1)m − 1 − m(m − 1) / 2 = 0 ].
In general, the shape space will also be a spherical, nonlinear space, of reduced
dimension compared with the preshape space.
and the “full Procrustes distance”, where the minimization is also done over a
scale parameter:
d F ( X1 , X 2 ) = min Z 2 − β Z1Γ . (7.6)
Γ∈SO ( m ) , β ∈
7 Statistical Shape Analysis of Manufacturing Data 225
Figure 7.4 Distances between two shapes in preshape space. ρ is the Procrustes distance (along
a geodesic), dF is the full Procrustes distance (along a tangent), and dp is the partial Procrustes
distance (along the secant). The preshapes have ⎜⎜Zi ⎜⎜= 1
1 n n
G ( X1 , X 2 ,… , X n ) = min
βi , Γi , γ i
∑ ∑ | | βi Xi Γi + 1k γ i′ − (β j X j Γ j + 1k γ ′j ) ||2 , (7.7)
n i =1 j = i +1
The mean shape of the n objects is simply the average of the n configurations,
∑
n
namely, μ = 1
n i =1
Xip .
The minimization (Equation 7.7) needs to be subjected to a constraint that limits
the scaling done, otherwise the optimal value of G will be zero. One such restric-
226 E. del Castillo
tion is to use a constraint on the size of the mean shape, S (μ ) = 1 , where the size of
∑ ∑
k m
any configuration X is defined as S ( X) = i =1 j =1
( X ij − X j ) 2 =|| CX || , where
∑
k
C = I k − k −11k1′k , X j = 1
n i =1
X ij and X ij is the jth coordinate of the ith point in
the configuration. Another common constraint, used in what follows, is to make the
average of the squared sizes of the registered configurations Xip given by Equation
7.8 equal to the average of the squared sizes of the original objects:
1 n 2 p 1 n
∑
n i =1
S ( Xi ) = ∑ S 2 ( X i ) .
n i =1
(7.9)
The GPA, as developed by Gower (1975) and Ten Berge (1977), proceeds as
follows to solve Equation 7.7 subject to Equation 7.9:
1. Center (but do not scale) the configurations X1 ,… , X n by initially defining
[alternatively, we can define H ′HXi = CXi = Xip and the resulting matrices
will be k × m ; note that Xip as defined above is instead (k − 1) × m ]
2. Let X(i ) = 1
n −1 ∑ j ≠i
X pj , i = 1,… , n . These are the “jackknifed” average shapes
excluding object i .
3. Do a Procrustes fit (rotation only) of the current Xip ’s onto X(i ) . This yields
rotation matrices Γi , from which we let
∑
n 2
X pj
βˆi = j =1
2
φi , i = 1,… , n
Xip
and let X ip ← βˆi Xip . The algorithm repeats steps 2–5 until there is conver-
gence.
7 Statistical Shape Analysis of Manufacturing Data 227
The algorithm is guaranteed to converge (in the sense that the fitted Xip cease
to vary as i increases), usually in just a few iterations (Ten Berge 1977). The
exact solution to the Procrustes registration problem between two objects X1 and
X 2 required in step 3, implies finding Γ ∈ SO (m) that minimizes d p ( X1 , X 2 )
(see Equation 7.5) for X1 = Xip and X 2 = X (i ) , i = 1,… , n . The exact solution to
this problem is well known in both the statistics (Jackson 2003) and the computer
vision (Horn et al. 1988) fields and is given by Γ = UV ′ , where U and V are
obtained from the singular value decomposition Z′2 Z1 = VΛU . An important
implementation detail of singular value decomposition for shape analysis is that to
ensure we have det (Γ) = +1 and hence a rotation matrix (as opposed to –1 and a
reflection matrix), we can make instead Γ = URV ′ , where R is the identity ma-
trix except for the last diagonal element, for which we use det (UV ′) .
The GPA as described assumes the statistical model
where μ is the mean shape of the objects and the k × m matrix of errors Ei is
such that vec(Ei ) ~ (0, σ 2 I km× km ) , where 0 is a vector of km zeroes and I km× km is
the km × km identity.
The model then assumes isotropic variance, i.e., the variance is the same at
each landmark and at each of the m coordinates. Modification of the GPA for the
case of a general covariance matrix of the errors Σ requires a straightforward
modification of the definition of the d F distances minimized in Equation 7.7 that
accounts for Σ . However, given that in general Σ is unknown and needs to be
estimated, there is no known registration algorithm which guarantees convergence
in the nonisotropic case. Common practice is to initially set Σ = I , run the GPA,
then estimate Σ with
1 n
Σ= ∑ vec(Xip − μ)vec(Xip − μ)′ ,
n i =1
run the GPA again with the squared full Procrustes distances in Chui and Rangara-
−1
jan (2000) replaced by the Mahalanobis squared distance vec( Xi )′Σ vec( X j ) ,
and iterate this process (but convergence is not guaranteed).
Equation 7.10 implies that each object results from the rotation, scaling, and
translation of the mean shape in the presence of random noise, i.e., similarity
transformations of the mean shape observed with noise generate the observed
profiles of the objects.
228 E. del Castillo
Figure 7.5 Example of generalized Procrustes algorithm (GPA) registration applied to the
contours of ten simulated “circular notched” parts, each with k = 200 (labeled) landmarks:
a original, unregistered shapes, and b shapes registered using the GPA
Once n configurations or profiles have been registered using the GPA, the main-
stream of the SSA literature (see, e.g., Dryden and Mardia 1998; Goodall 1991;
Adams et al. 2004) recommends that further statistical analysis of shape variabil-
ity and any desired inferences be made on the basis of the resulting registered
shapes Xip using the full Procrustes distances from the mean shape (or pole),
called the tangent space coordinates. This is suggested in contraposition to work-
ing with the Procrustes distances which are not linear. A principal component
analysis (PCA) on the tangent space coordinates is then recommended to better
understand the directions in which the shapes are varying the most.
7 Statistical Shape Analysis of Manufacturing Data 229
For a preshape Xip and mean shape μ , the tangent coordinates vi are the dis-
tances along the tangent at the mean shape corresponding to the projection of Xip on
μ and are given by
⎡ ⎤
⎛ μ ⎞ ⎛ μ ⎞′ ⎥ ⎛ Xp ⎞
v i = ⎢ I ( k −1) m − vec ⎜ ⎟ vec ⎜ ⎟ vec ⎜ i
⎟ , i = 1,… , n . (7.11)
⎢ ⎜ || μ || ⎟ ⎜ || μ || ⎟ ⎥ ⎜
⎢⎣ ⎝ ⎠ ⎝ ⎠ ⎦⎥ ⎝ || Xi
p
|| ⎟⎠
The tangent coordinates are (k − 1)m vectors. If the centered configurations are
used, then v i is a km dimensional vector.
An alternative approach which is close to the tangent coordinates if the pre-
shapes are not too different from the mean shape (see Figure 7.6) is to use the
Procrustes residuals ri defined by
that is, we work with the secants instead of the tangents. Again, for small differ-
ences about the mean, the conclusions of the analysis would be very similar. Re-
gardless of how one computes the tangent coordinates, either using Equation 7.11
or using the approximate tangent coordinates v i ≈ ri , the mainstream approaches
to SSA recommend using a PCA on the v i ’s (Goodall 1991; Adams et al. 2004).
The theoretical justification for this recommendation comes from work by Kent
230 E. del Castillo
and Mardia (2001), who have shown that an isotropic distribution of the landmarks
results in an isotropic distribution in the tangent space (given that small changes in
a configuration matrix X induce an approximately linear change in the tangent
coordinates v ), and, hence, PCA in tangent space is valid for shape analysis.
Let V = [ v1 , v 2 ,… , v n ] be all the tangent coordinates of all n objects under
study. An estimate of the covariance matrix cov(V ′) , giving the between-shape
variances and covariances at the landmarks, is given by
1 n
Sv = ∑ ( vi − v)( vi − v)′ ,
n i =1
where v is the average of the v i ’s. This is a (k − 1)m × (k − 1)m matrix if the pre-
shapes are only scaled and a km × km matrix if the centered preshapes are used
instead. In the first case, the rank of this matrix is p = M = (k − 1)m − 1 − m(m − 1) / 2
and in the latter case the rank is p = M + 1 , since the mean is not lost. Let {λ j } pj =1
and {e j } pj =1 be the p eigenvalues and eigenvectors of S v . Dryden and Mardia
(1998) suggested computing
v(c, j ) = v + c λ j e j , j = 1,… , p
⎡ H′ 0⎤
vec( X I ) = ⎢ [ v (c, j ) + vec(μ)] (7.12)
⎣0 H ′ ⎥⎦
for all principal components j and for all multiplies c . These are the coordinates
of the original shapes where the (registered) objects exist, and indicate the direc-
tions in which the principal components indicate movement – variability – around
the mean shape (if the v i ’s are km -dimensional vectors, there is no need to
premultiply the block matrix of Helmert matrices). Just as in regular PCA, the
percentage of variation explained by the jth principal component is given by
∑
p
100λ j / j =1
λ j . Once the tangent coordinates have been computed, multivari-
ate analysis techniques can be applied in the usual way until the point where visu-
alization using Equation 7.12 is necessary.
Example. PCA of the circular notch shape data. Consider the ten shapes shown
in Figure 7.5. These shapes were simulated by superimposing sinusoidal variabil-
ity along the circle at a second harmonic (inducing a bilobed shape) and inducing
7 Statistical Shape Analysis of Manufacturing Data 231
Figure 7.7 Example of principal component analysis applied to the data of the ten circular
notched parts: a the first principal component detects a sinusoidal variation, and accounts for
56.7% of the variability, and b the second principal component, which corresponds to variation
in the depth of the notch, accounts for 8.3% of the variability. In these plots, values of c ∈ – (3,3)
were used. The dark line is the mean shape
variability in the depth of the notch. Additional random normal bivariate variabil-
ity was added at each landmark, which masks the first two sources of variability in
such a way that they are not obvious to the eye. These sources of variability where
added to demonstrate the power of PCA in the tangent space. Figure 7.7 shows the
first two principal components in this example, which together account for more
than 65% of the variability. Note how the first component is precisely the simu-
lated bilobed shape and the second component refers to the depth of the notch. The
remaining principal components do not show any obvious pattern.
Performing a PCA on the tangent coordinates is of value when one is interested
in analyzing how the variability of the shapes behaves around the mean. To ana-
lyze the effect of factors (varied during an experiment) on the mean shape (as
required when conducting manufacturing experiments that may improve the
shapes of parts produced by a process) one needs to perform an analysis of vari-
ance. This was first discussed by Goodall (1991) (see also Dryden and Mardia
1998) for the one-way case and was studied in the two-way layout case, with an
application in manufacturing, by Del Castillo and Colosimo (2009).
The methods presented in this chapter assume all parts contain the same number of
corresponding landmarks, or locations of interest. If there is a different number of
landmarks between two objects, the iterative closest point (ICP)) algorithm (Besl
and McKay 1992; Zhang 1998) has been proposed to obtain the same number of
corresponding landmarks between the parts. If the landmarks do not correspond, a
matching algorithm such as the context-labeling algorithm in Section 7.2 could be
applied after the ICP algorithm. An alternative to the use of the ICP algorithm
232 E. del Castillo
(Belongie et al. 2002) is to simply add dummy landmarks to the smallest landmark
matrix to get max(k1 , k 2 ) landmarks, and assign a large cost between these points
and all others (this is also a mechanism to handle outlier landmarks, since they
would be matched to the dummy points).
A matching algorithm notably different from the one presented in Section 7.2
was proposed recently by Green and Mardia (2006). It also applies to the case of
two objects. Another possibility is to use the two-dimensional context histograms,
but to use a statistic other than the χ 2 used here, to measure distances (costs)
between two multivariate distributions, e.g., a two-dimensional Kolmogorov–
Smirnov test or other recent alternatives (e.g., that in Rosenbaum 2005). Such an
approach would still use the weighted matching linear programming formulation
presented here, but with a different way of getting the cost matrix C . Even in the
Belongie et al. (2002) approach, it is not clear how to best scale the X matrices,
how many bins to use in each dimension, or what is the best way to measure an-
gles for differently oriented objects in order to achieve effective rotation invari-
ance. An interesting embellishment of the landmark matching algorithm (Belongie
et al. 2002) is to iterate the matching algorithm with an algorithm for the estima-
tion of the registration transformation between the objects. This may result in
better matching (and hence registration) because the initial matching may be sensi-
tive to the different orientations of the parts due to the ambiguities mentioned
earlier about defining the histogram resolution. These authors suggested using thin
plate spline transformations, popular also in the area of morphometrics, as op-
posed to the GPA considered here. A similar iterative procedure could be at-
tempted with the context labeling algorithm and the GPA applied iteratively. A
recent description of the matching problem from a computer-vision perspective is
given in the book by Davies et al. (2008).
As mentioned earlier, a generalization of two-object matching methods to the
case of n objects is desirable, since once labeled (corresponding) landmarks are
available (assuming there is the same number of landmarks in each object), the
SSA methods presented herein can be implemented. The advantages and disadvan-
tages of the “context labeling” method compared with the Green and Mardia
(2006) approach need to be investigated.
In this chapter we did not discuss tests for comparing the mean shapes between
two or more populations, which can be done using analysis of variance methods
applied to the shapes. For more information on this topic, see Del Castillo and
Colosimo (2009).
Most of the work on SSA has focused on two-dimensional shapes. Extensions
to the three-dimensional case are evidently practical (but the landmark matching
problem becomes more difficult). The context labeling approach presented here
was extended recently by Frome et al. (2004) to the three-dimensional case using
three-dimensional histograms. The implementation details mentioned above re-
main and need to be studied. For three-dimensional objects, the GPA and the
PCA can be used without any change, but visualization of the PCAs is challeng-
ing if k (number of of landmarks) is large.
7 Statistical Shape Analysis of Manufacturing Data 233
Finally, some authors (e.g., Lele and Richtsmeier 2001) have proposed using
the interlandmark Euclidean distance matrix [dij ] to make inferences on the
shapes of objects, with application to testing for the difference between shapes.
Lele and Richtsmeier (2001) suggested using the GPA to estimate the mean shape,
but there is debate about how to estimate the covariance matrix of the landmarks
in the nonisotropic case. This series of methods do not have an easy way to visual-
ize the results, and require more information ( ( k2 ) distances instead of km), al-
though this information is implicit in the k × m matrix X . In addition, there
seems to be no counterpart to the PCA of variability in distance-based methods.
There is considerable debate about which method is more powerful to detect dif-
ferences in shapes, and it is of interest to compare distance-based methods with
those studied in Del Castillo and Colosimo (2009) for a variety of shapes of rele-
vance in manufacturing, since the power of these methods appears to depend on
the shape in question. Dryden and Mardia (1998) presented a good overview of
distance-based methods.
MATLAB programs that perform the computations required for the context label-
ing algorithm in Section 7.2 and for the GPA, including visualization of PCAs,
were written for this research and can be downloaded from
http://www2.ie.psu.edu/Castillo/research/EngineeringStatistics/software.htm.
The programs posted contain several programs for SSA. Two of the programs are
related to what is discussed in the present paper: ContextLabeling.m, which
implements the context labeling algorithm presented in Section 7.2 for two two-
dimensional objects, and GPA23.m, which implements the GPA (assuming iso-
tropic variance), and performs, if desired, the PCA in tangent space, including the
corresponding visualization.
Acknowledgements This research was partially funded by the National Science Foundation
grant CMI-0825786 and a grant from the Ministry of Science of Italy.
References
Adams DC, Rohlf FJ, Slice DE (2004) Geometric morphometrics: ten years of progress follow-
ing the “revolution”. Ital J Zool 71:5–16
ASME Y14.5M (1994) Dimensioning and tolerancing. American Society of Mechanical Engi-
neers, New York
234 E. del Castillo
Belongie S, Malik J, Puzicha J (2002) Shape matching and object recognition using shape con-
texts. IEEE Trans Pattern Anal Mach Intell 24:509–522
Besl PJ, McKay ND (1992) A method for registration of 3-D shapes. IEEE Trans Pattern Anal
Mach Intell 14:239–256
Chui H, Rangarajan A (2000) A new algorithm for non-rigid point matching Proc IEEE Conf
Comput Vis Pattern Recognit 44–51
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
Davies R, Twining C, Taylor C (2008), Statistical models of shape, optimisation and evaluation.
Springer, London
Del Castillo E, Colosimo BM (2010) Statistical shape analysis of experiments for manufacturing
processes. Technometrics (in press)
Dryden IL, Mardia KV (1998) Statistical shape analysis. Wiley, Chichester
Frome A, Huber D, Kolluri R, Bulow T, Malik J (2004) Recognizing objects in range data using
regional point descriptors. In: Proceedings of the 8th European conference on computer vi-
sion, vol 3, pp 224–237
Gold S, Rangarajan A, Lu C-P, Pappu S, Mjolsness E (1998) New algorithms for 2D and 3D
point matching: pose estimation and correspondence. Pattern Recognit 31:1019–1031
Goodall C (1991) Procrustes methods in the statistical analysis of shape. J R Stat Soc B,53:285–
339
Gower JC (1975) Generalized Procrustes analysis. Psychometrika 40:33–51
Green PJ, Mardia KV (2006) Bayesian alignment using hierarchical models, with applications in
protein bioinformatics. Biometrika 93:235–254
Horn BKP, Hilden HM, Negahdaripour S (1988) Closed-form solution of absolute orientation
using orthonormal matrices. J Opt Soc Am A 5:1127–1135
Jackson JE (2003) A user’s guide to principal components. Wiley, New York
Kang L, Albin SL (2000) On-line monitoring when the process yields a linear profile. J Qual
Technol 32:418–426
Kendall DG (1984) Shape manifolds, Procrustean metrics, and complex projective spaces. Bull
Lond Math Soc 16:81–121
Kendall DG (1989) A survey of the statistical theory of shape Stat Sci 4:87–89
Kent JT, Mardia KV (2001) Shape, Procrustes tangent projections and bilateral symmetry. Bio-
metrika 88:469–485
Klingenberg CP, McIntyre GS (1998) Geometric morphometrics of developmental instability:
analyzing patterns of fluctuating asymmetry with Procrustes methods. Evolution 52:1363–
1375
Lele SR, Richtsmeier JT (2001) An invariant approach to statistical analysis for shapes. Chap-
man & Hall/CRC, Boca Raton
Papadimitrou CH, Steiglitz K (1982) Combinatorial optimization, algorithms and complexity.
Prentice Hall, Englewood Cliffs
Rosenbaum PR (2005) An exact distribution-free test comparing two multivariate distributions
based on adjacency. J R Stat Soc B 67:515–530
Ten Berge JMF (1977) Orthogonal Procrustes rotation for two or more matrices. Psychometrika
42(2):267–276
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
Zelditch ML, Swiderski DL, Sheets HD, Fink WL (2004) Geometric morphometrics for biolo-
gists, a primer. Elsevier, San Diego
Zhang Z (1998) Iterative point matching for registration of free-form curves. Reports de recher-
che no. 1658. IRIA, Sophia Antipolis
Part III
Impact on Statistical Process Monitoring
Chapter 8
Statistical Quality Monitoring of Geometric
Tolerances: the Industrial Practice
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
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.
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.
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
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.
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.
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).
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.
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
1 n 1 n
∑ ∑
2
where y ( k ) = y j ( k ) and s ( 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.
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
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
Chapter 9
Model-based Approaches for Quality
Monitoring of Geometric Tolerances
Abstract A relatively new area of research in the field of statistical process con-
trol has been named profile monitoring. It includes a collection of methods and
techniques used to check the stability of a functional relationship over time. This
chapter shows how this approach can be usefully considered as a viable solution to
form error monitoring when geometric tolerances are of interest. In this case, qual-
ity monitoring consists in detecting deviations of the shape from its nominal or in-
control state. This task is accomplished by firstly modeling the functional relation-
ship representing the manufactured shape and then checking whether or not the
estimated model is stable over time. The goal of this chapter is to introduce profile
monitoring, show how it works, and then illustrate how this approach can be effec-
tively used for quality control of geometric form errors.
9.1 Introduction
__________________________________
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
257
258 B.M. Colosimo and M. Pacella
y
y = f (x)+ε
y = f (x)
x
Figure 9.1 A simple example of functional data
tional data refer to information summarized in the form of profiles where the re-
sponse variable y can be modeled as a function of one (or more) independent vari-
able x plus random noise ε, i.e., y = f (x) + ε (Figure 9.1). Usually, the independent
variable is used to define a spatial or temporal location. Very often, functional data
are related to quality characteristics of the product or process and hence monitoring
the stability of the functional data allows one to detect out-of-control states that are
usually associated with deteriorated process performances (quality loss).
Since functional data are usually observed only at a finite set of locations, a
vector can be used to store the observed set of responses. Therefore, the most
direct way to deal with these data vectors consists in using multivariate control
charting (Montgomery 2000), thus treating each profile as a realization of a mul-
tivariate process. The use of standard multivariate charts for monitoring profile
data is not recommended. When the number of points of the profile exceeds the
number of profiles used during the design stage, the common multivariate statis-
tics cannot be estimated. In actual applications, the number of monitored points on
each profile is usually greater than the number of samples collected during the
design stage. This is especially true in the case of machined profiles subject to
geometric specification (e.g., roundness, straightness, free-form tolerance) where,
in order to have an accurate estimate of the form error, the number of observations
sampled can be on the order of hundreds. Hence, when the quality of a process or
product is characterized by functional data, new approaches are required.
Recently, there has been much research activity in a new area of statistical
process control, named profile monitoring. Woodall et al. (2004) and Woodall
(2007) discussed general issues to be addressed when monitoring quality profiles,
and presented a complete review of the literature on the topic of profile monitor-
ing. The approaches for profile monitoring proposed in the literature share a com-
mon structure which consists of:
9 Model-based Approaches for Quality Monitoring of Geometric Tolerances 259
used in Section 9.2 to show how profile monitoring works. The following sections
describe approaches aimed at monitoring geometric tolerances by using spatial
regression models and PCA-based ones. Numerical examples are used to illustrate
all the approaches presented.
15
90 15
120 60 14 y(k)
10 13
150 30
12
5
k)
11
y(
180
x(k) 0 10
8
210 330
7
240 300 6
270
5 x(k)
0 1 2 3 4 5 6 7
a b
Figure 9.2 An example of a roundness profile: a the polar graph shows the radius y(k) modeled
as a function of the angle x(k), and b the Cartesian diagram describes the roundness profile as
functional data
y(k)
z(k)
x(k)
a b
Figure 9.3 The equally spaced grid of locations where the radius has to be measured to model
a a cylindrical surface, and b the coordinates of the kth point [radius y(k), angle x(k), and vertical
position z(k)]
9 Model-based Approaches for Quality Monitoring of Geometric Tolerances 261
Linear profile monitoring is a broad topic applied for a wide variety of applica-
tions. Most of the work on profile monitoring proposed in the literature is related
to the case in which the profile can be adequately represented by a simple straight
line (Woodall et al. 2004; Woodall 2007).
Calibration processes are often characterized by such linear functions. For in-
stance, Stover and Brill (1998) studied multilevel ion chromatography linear cali-
brations to determine the instrument response stability and the proper calibration
frequency. Furthermore, Kang and Albin (2000) presented two examples of proc-
ess profiles. One of them was a semiconductor manufacturing application in which
the process is represented by a linear calibration function. Several other case stud-
ies and approaches for linear profile monitoring can be found in the literature
(Kim et al. 2003; Mahmoud and Woodall 2004; Chang and Gan 2006; Gupta et al.
2006; Zou et al. 2006; Mahmoud et al. 2007). These approaches share a common
basic idea for implementing linear profile monitoring, i.e., to use control charts
based on the estimated regression parameters.
Assume we collect a group of n profiles, where each profile consists of p
measurements observed at a fixed set of locations. Let y j ( k ) denote the depend-
ent variable measured at a specific location of index k on the j th profile, and let
x ( k ) represent the value of the independent variable at the same location
( k = 1, 2,… p and j = 1, 2,… n ). We also assume that, as in many profile monitor-
ing applications, the x -values are known constants and have the same values in
all samples.
The observed data collected over time are n random samples, with each sam-
ple consisting of a sequence of p pairs of observations ⎡⎣ x ( k ) , y j ( k ) ⎤⎦ . For each
sample of index j = 1, 2,… n , it is assumed that the model which relates the inde-
pendent variable x to the response y is the following:
y j ( k ) = b0 j + b1 j x ( k ) + ε j ( k ) , (9.1)
y j = ∑ k =1 y j ( k ) p , x = ∑ k =1 x ( k ) p , S xx = ∑ k =1 ⎡⎣ x ( k ) − x ⎤⎦ , and
p p p 2
where
S xy ( j ) = ∑ k =1 ⎡⎣ x ( k ) − x ⎤⎦ ⎡⎣ y j ( k ) − y j ⎤⎦ . Furthermore, σ 2j can be estimated by the
p
is the residual error at the location of index k = 1, 2,… p on the profile of index
j = 1, 2,… n , in the formula e j ( k ) = y j ( k ) − bˆ0 j − bˆ1 j x ( k ) .
Assume the linear profile process is in control, with fixed effects and independ-
ent, identically distributed residual errors N ( 0, σ 2 ) , then the least-squares estima-
tors of the intercept bˆ0 j and slope bˆ1 j are distributed as a bivariate normal distri-
bution with the mean vector and the variance–covariance matrix, respectively:
⎡ σ 2 σ 01 ⎤
β′ = [ β 0 β1 ] , Σ=⎢ 0 2 ⎥
, (9.3)
⎣σ 01 σ 1 ⎦
y j ( k ) = a0 j + a1 j x′ ( k ) + ε j ( k ) , (9.4)
ing the residual variation about the regression line involves plotting the quanti-
ties Fj , j = 1, 2,… , n , on a chart with the following upper and lower limits:
Figure 9.4 A linear profile data set consisting of 12 samples each with ten points. Each point
represents a functional relationship between an independent variable (abscissa) and a dependent
one (ordinate). The independent variable assumes the values are equally spaced on the abscissa.
The continuous line in each graph represents the least-squares fitted line where the intercept and
slope are based on Equation 9.2
266 B.M. Colosimo and M. Pacella
Figure 9.5 Residual errors between observed points depicted in Figure 9.4 and the least-squares
fitted line. A standard assumption in the monitoring of simple linear regression profiles is that the
errors are independent and identically distributed, usually with an assumed normal distribution
Table 9.1 Least-squares estimation of the intercept and slope for 12 samples of ten points each
j bˆ0 j bˆ1 j
1 1.0013 4.2896
2 1.2310 2.3770
3 1.0253 2.4904
4 1.3588 2.3424
5 0.5803 2.7451
6 1.1055 2.5346
7 1.2253 3.7798
8 0.9457 3.5215
9 1.1022 2.2775
10 0.9039 2.8027
11 0.5335 3.4190
12 0.6243 3.4545
9 Model-based Approaches for Quality Monitoring of Geometric Tolerances 267
The first step in the Kim et al. (2003) approach is to code the independent vari-
able x ( k ) so that the average coded value is zero. In the numerical example at
hand, as the independent variable x assumes discrete values equally spaced in the
interval [ −0.5, 0.5] , we have x = 0 . Therefore, a0 j = b0 j , a1 j = b1 j , and
x′ ( k ) = x ( k ) , i.e., Equation 9.4 is equal to Equation 9.1. For an in-control process
with fixed effects (i.e., b0 j = β 0 , b1 j = β1 ), the slope bˆ0 j and the intercept bˆ1 j are
statistically uncorrelated, normally distributed random variables [as
ε j ( k ) ∼ N ( 0, σ 2 ) ], with means β 0 and β1 and variances σ 2 p and σ 2 S xx ,
respectively.
In this numerical example, the in-control parameters of the simulated linear
profile process are β 0 = 1 , β1 = 3 , and σ 2 = 1 . Therefore, the estimated least-
square estimate slopes bˆ have mean equal to 1 and variance equal to 0.1. On the
0j
other hand, the estimated least-squares intercepts bˆ1 j have mean equal to 3 and
variance approximately equal to 0.9818. These true values of the process moni-
tored can be used to implement two control charts in order to monitor the intercept
and the slope separately (Kim et al. 2003).
When the in-control process parameters β 0 , β1 , and σ 2 are not known, the
control limits of Equations 9.5 and 9.6 can be used for phase I control charting. In
particular, the limits computed for the numerical example at hand are reported in
Table 9.2. The control chart for the intercept is depicted in Figure 9.6a, while the
chart of the slopes is shown in Figure 9.6b. Similarly, the control limits in Equa-
tion 9.7 are used to monitor the statistics Fj . The numerical values of such limits
are also reported in Table 9.2, while the corresponding control chart is depicted in
Figure 9.6c.
A nominal false-alarm probability (type I error rate) equal to α ′ = 5% was used
to compute the control limits of the three charts. Given three control charts are
contemporarily used, Bonferroni’s rule for independent events was used to design
the control limits of the three charts. Therefore, given a nominal false-alarm prob-
ability α ′ = 5% , the control limits of each of the three control charts in Equations
9.5–9.7 were set by assuming a false-alarm probability α = 1 − 3 1 − α ′ . From
Figure 9.6, it can be noted that there are no profiles, in the set of 12 samples,
which produce an alarm in any of the three control charts.
Table 9.2 Phase I control limits for each parameter of the linear profile model for the numerical
example. Type I error rate 5%
a b c
Figure 9.6 a Control chart of the least squares estimates of the intercept (phase I), b control chart of the least-squares estimates of the slope
(phase I), and c control chart for monitoring the residual variation about the regression line (phase I)
B.M. Colosimo and M. Pacella
9 Model-based Approaches for Quality Monitoring of Geometric Tolerances 269
Although most approaches for profile monitoring focus on linear profiles, when
geometric tolerances are of interest, nonlinear profiles are usually necessary. Dif-
ferent authors have focused on nonlinear profile monitoring (Williams et al. 2007;
Walker and Wright 2002; Jin and Shi 1999, 2001; Young et al. 1999; Ding et al.
2006; Zou et al. 2008; Zhang and Albin 2009; Jensen and Birch 2009).
Walker and Wright (2002, p. 124) mentioned that “autocorrelation is frequently
present in data that are observed within small intervals of time or space”. Despite
the specific mention of the autocorrelation problem, the approach proposed by
Walker and Wright (2002) was based on independent data, whereas most of the
other approaches focused on profile monitoring. The only exceptions to this gen-
eral rule are the papers by Jensen et al. (2008) and Colosimo et al. (2008, 2010).
In particular, while the first paper dealt with autocorrelated profile data, the sec-
ond ones focused on spatially autocorrelated data and are more suitable for model-
ing geometric form features (profiles and surfaces). When data refer to surface
texture or profile tolerances, measurements are often spatially correlated because
they are obtained in similar conditions of the machining process and are related to
similar (local) properties of the machined material. Spatial autocorrelation is dif-
ferent from temporal autocorrelation, which is usually represented through time-
series models. Spatial autocorrelation models allow one to represent contiguity in
space rather than in time. With reference to profiles, contiguity in space implies
that the dependency among data on a profile is bidirectional (i.e., a given point is,
in principle, correlated to points located on its left and on its right, regardless of
the specific direction), while time-series models are suitable for representing just a
one-direction dependency (i.e., past data influence future ones) (Whittle 1954).
Second, spatial models allow one to represent a specific type of relationship
among points observed in closed profiles (e.g., roundness profile). In fact, when
data on a closed or circuit profile are sequentially numbered (by defining an arbi-
trary starting point), observations at the beginning and at the end of the profile are
spatially correlated (Colosimo et al. 2008).
With reference to the vector of points measured on a machined profile which is
subject to geometric specification, the predictable behavior can be referred to as
the process signature and is defined as the systematic pattern that characterizes all
the features machined with a given process. Knowledge of this signature can be
used to design proper tools for profile monitoring. Colosimo et al. (2008, 2010)
presented an approach for modeling the manufacturing signature based on fitting a
SARX model (Cressie 1993). The regression model presented in these papers is
summarized below.
270 B.M. Colosimo and M. Pacella
y j = Xb j + υ j ,
(I − R ) υ
j j = εj, (9.8)
q
R j = ∑ asj W ( s ) .
s =1
The first expression in Equation 9.8 describes the p × 1 vector of the response
for the j th profile y j as formed by a large-scale and a small-scale component
(Cressie 1993). The large-scale component is given by Xb j where X is a p × r
matrix of r regressor variables that are assumed to be known and constant and
b′j = ⎡⎣b1 j … blj … brj ⎤⎦ is the r × 1 vector of regression parameters which are nor-
mally distributed with mean β′ = [ β1 … βl … β r ] and covariance matrix B
⎡⎣b j ∼ N ( β, B ) ⎤⎦ . With reference to the large-scale component, fixed-effect models
are usually assumed in traditional approaches for profile monitoring. This situation
can be seen as a special case of the general model in Equation 9.8, where B = 0 is
the variance matrix of the large-scale model coefficients b j . Alternatively, when
B ≠ 0 is considered, random effects are included in the large-scale component of
the model.
The small-scale component is the p × 1 vector of error terms υ j in Equa-
tion 9.8. Error terms are assumed to be spatially correlated and are represented as
a generic spatial autoregressive process (SAR) of order q . The SAR ( q ) model
expression is given in the last two expressions in Equation 9.8, where I is the
p × p identity matrix, ε j is a p × 1 vector of independently and normally dis-
tributed errors ⎡⎣ε j ∼ MN ( 0, σ 2 I ) ⎤⎦ and a′j = ⎡⎣ a1 j … asj … aqj ⎤⎦ is the vector of the
coefficients of the SAR ( q ) model for the j th profile, which is assumed to be
9 Model-based Approaches for Quality Monitoring of Geometric Tolerances 271
element w(1) ( k , t ) is set equal to 1 if the t th point is the neighbor of the k th point
and is set to 0 otherwise. Analogously, the element of the second-order adjacency
matrix, w( 2) ( k , t ) , is set equal to 1 if the t th point is a neighbor of the original
first-generation neighbors of the k th point, and so on. By definition, all the adja-
cency matrices are binary and symmetric matrices whose diagonal elements are
zero (Cressie 1993).
For each profile, two vectors of coefficients b j and a j are considered. In order
to let the model have the most general form, we further assume that these two
vectors could also be correlated, i.e., cov ( b j , a j ) = D (Colosimo and Pacella
2010). In other words, with reference to the parametric model structure given in
Equation 9.8, we merge the two vectors characterizing the observed pattern into a
single coefficient vector related to the j th profile:
The signature model for the j th profile shown in Equation 9.8 is completely de-
fined by a SARX model which requires one to estimate the d = r + q parameters
which are components of the vector c j , besides the residual variance σ 2 . Let
cˆ ′j = ⎡⎣bˆ ′j aˆ ′j ⎤⎦ represent the vector of the estimates of the d parameters for the
j th profile. A T 2 control chart can be designed with reference to the statistics,
T j2 = ( cˆ j − μ )′ Σ −1 ( cˆ j − μ ) , (9.10)
where j = 1, 2,… , μ and Σ are, respectively, the mean vector and the covariance
matrix of the d -dimensional vectors of the coefficients. Here, the control chart
parameters are assumed to be known or estimated from a large data set of in-
control profiles.
272 B.M. Colosimo and M. Pacella
( )
number of parameters estimated d + ⎡⎣ d ( d + 1) 2 ⎤⎦ , the upper control limit based
2
on the asymptotic distribution of T can be used:
UCL = χα2 , d , (9.11)
(
e j = ( I − R j ) y j − Xbˆ j , ) (9.12)
by
e′j e j
s 2j = . (9.13)
p −1
where χα2 2, p −1 and χ12− (α 2), p −1 are the upper and lower α 2 percentage points of
the χ 2 distribution with p − 1 degrees of freedom associated with the residuals
(Montgomery 2000).
The case study used as a reference was described in detail by Colosimo et al.
(2008). It consists of items obtained by turning (cutting speed, 163 m/min; feed rate,
0.2 mm/rev; two steps of cutting depth 1 mm) C20 carbon steel cylinders, where
9 Model-based Approaches for Quality Monitoring of Geometric Tolerances 273
each item was characterized by a roundness profile of 748 evenly distributed meas-
urements of its radius. The original measurements sampled using a coordinate
measuring machine were scaled by subtracting the least-squares estimate of the
radius and centered at the least-squares estimate of the center. A further step was
eventually applied to register all the sampled profiles by minimizing the phase delay
caused by the random contact angle (Colosimo and Pacella 2007) (see Chapter 11).
Starting from actual measurements, the general SARX model of Equation 9.8
was fitted to the data. In this case, y j represents the column vector of the radial
deviation from the nominal radius measured at the angular position
θ k = k ( 2π p ) , where k = 1, 2,… p ( p = 748 ) is the index of equally spaced
observations on each profile.
Consider the l th column vector of matrix X in Equation 9.8, and denote such a
p × 1 vector by xl , where l = 1, 2,… r . In general, each element of vector xl is
described as a function of the index location k . In the specific case of roundness
profiles, xl ( k ) can be expressed either as xl ( k ) = cos ⎣⎡( k − 1) f h ⎦⎤ or as
xl ( k ) = sin ⎡⎣( k − 1) f h ⎤⎦ , i.e., as a sinusoidal function of frequency equal to
f h = h ( 2π p ) rad per sample. h is the frequency ( h ∈ {1, 2,… p 2} ) measured in
undulations per revolution, which is fixed for all the elements of xl .
Two harmonics were selected to model the radial deviations in the actual test
case, namely, the second and the third harmonics. Indeed, the process signature
was mainly affected by ovality and triangularity (Moroni and Pacella 2008). The
oval contour was possibly due to a bilobe error motion affecting the spindle’s lathe
or to eccentricity caused by an improper setup, while the three-lobe pattern was
due to a similar error motion of the spindle.
Therefore, the regressor matrix X in Equation 9.8 has four columns ( r = 4 )
since two sinusoidal functions are needed to model the amplitude and
phase of each specific harmonic. The k th row of matrix X is equal
to cos ⎡⎣( k − 1) f 2 ⎤⎦ sin ⎡⎣( k − 1) f 2 ⎤⎦ cos ⎡⎣( k − 1) f 3 ⎤⎦ sin ⎡⎣( k − 1) f 3 ⎤⎦ , where
f h = h ( 2π p ) represents the frequency in radians per sample of the h th har-
monic ( h = 2,3) .
The vector of random error υ j in Equation 9.8 was fitted as a SAR model of
order 2 ( q = 2 ) (Colosimo et al. 2008), using the algorithm implemented in the
Spatial Econometrics toolbox (LeSage 1999).
With reference to the model of Equation 9.8, each specific profile of index j is
associated with the vector of d = p + q = 4 + 2 parameters c j ∼ N ( μ, Σ ) , where
⎡ B D⎤
μ = [α ′ β′] and Σ = ⎢ ⎥ . Values of the vector and the matrices for the
⎣ D′ A ⎦
actual roundness data are shown in Table 9.3 (Colosimo et al. 2008).
274 B.M. Colosimo and M. Pacella
⎡ B D⎤
Table 9.3 Parameters characterizing the mean μ = [α ′ β′] and the variance Σ = ⎢ ⎥
⎣ D′ A ⎦
of the distribution of coefficients c j ∼ N ( μ, Σ ) for the actual roundness data (Colosimo et al.
2008)
Figure 9.7 Twelve samples of a roundness profile, each of 748 points. Each point represents a
functional relationship between an independent variable (abscissa) and a dependent one (ordinate).
The dependent variable is the radial deviation from the nominal radius measured at the angular
position. The independent variable is the index of equally spaced observations on each profile
Table 9.4 Spatial autoregressive regression (SARX) model parameters estimated for a group
of 12 roundness profiles shown in Figure 9.7
Figure 9.8 Residual errors between observed points depicted in Figure 9.7 and the SARX
models for each roundness profile
Table 9.5 T 2 statistics for the 12 roundness profiles shown in Figure 9.7
Figure 9.9 Multivariate control chart for the regression-based approach applied to the 12 ran-
domly generated roundness profile data sets shown in Figure 9.7
z j = U ′ ( y j − y ) = ⎡⎣ z j1 … z jk … z jp ⎤⎦′ , (9.16)
where z jk are the so-called scores. Each profile can then be expressed as a linear
combination of loadings u k , where the weights of the linear combination are the
scores z jk :
y j = y + z j1u1 + z j 2 u 2 + + z jp u p . (9.17)
Since the PCs are statistically uncorrelated and each PC has variance equal to
the corresponding eigenvalue ( lk ) , we can rank the PCs according to the associ-
ated eigenvalue and decide to retain just the most important ones (i.e., the ones
which are associated with larger variances). Different approaches can be used to
select the proper set of PCs (Jackson 2003; Jolliffe 2002). For instance, cross-
validation can be effectively used to choose the number m of significant PCs (Co-
losimo and Pacella 2007). When a subset m of the whole number of p PCs is re-
tained ( m < p ) , the original profile can be estimated as
yˆ j ( m ) = y + z j1u1 + z j 2 u 2 + + z jm u m . (9.18)
Given the estimate in Equation 9.18, the Q statistic can be computed as the sum of
the squared errors obtained by reconstructing each observation by the first m PCs:
(
Q j = y j − yˆ j ( m ) )′ ( y j )
− yˆ j ( m ) . (9.20)
The upper control limit of the T 2 statistics in Equation 9.19 can be computed
as (Williams et al. 2006)
UCL = χα2 , m , (9.21)
With reference to the aforementioned case study of roundness profiles (see Sec-
tion 9.3.1.2), Colosimo and Pacella (2007) used a cross-validation approach in
order to determine the number of significant PCs to be retained. With reference to
the roundness profiles of the reference case study, the number of significant PCs is
equal to m = 3 .
Once the PCA has been performed, the retained PCs should be interpreted to
gain more insight into the systematic pattern characterizing the machined profiles.
To this aim, each eigenfunction uk (i.e., the coefficients of each eigenvector, also
known as loadings) can be graphically represented as a function of the location.
Following this practice, Figure 9.10 reports the diagrams of the first three eigen-
functions (u1, u2, u3), which are related to the three retained PCs (Colosimo and
Pacella 2007). The first PC, which describes the most important component of
variability (13.87%), has a bilobe form. This qualitative observation indicates that
the main variability around the mean profile is due to a periodic function charac-
terized by a frequency of two undulations per revolution. The second PC (which
accounts for the 10.41% of the total variability) is a mixture of a bilobe and
a trilobe contour. This mixture is obtained by combining two periodic functions,
namely, harmonics with two and three undulations per revolution. Eventually, the
third PC (which accounts for 7.10% of the total variability) has a trilobe contour.
Assume we consider a subgroup of n = 12 roundness profiles, where each pro-
file consists of p = 748 measurements observed at a fixed set of locations. In par-
ticular, the 12 samples depicted in Figure 9.7 are considered. A PCA-based control
280 B.M. Colosimo and M. Pacella
chart can be used to monitor the vector of the first m retained PCs. In this case, the
plotted statistics are based on the values of the eigenvalues associated with the
three retained PCs. With reference to the case study, these values are
l1 = 0.59 ×10−3 , l2 = 0.44 × 10−3 , and l3 = 0.30 × 10−3 .
The scores z j1 , z j 2 , and z j 3 for each profile of index j = 1, 2,…12 are reported
in Table 9.6. In particular, the T j2 statistics computed for each of the 12 samples is
based on Equation 9.19. Since T j2 follows a central χ 2 distribution with a number
of degrees of freedom equal to the number of retained PCs ( m = 3) , the upper
control limit of this chart that is used is based on Equation 9.21 and consists of a
100 (1 − α ) percentile of the χ 2 distribution with three degrees of freedom (in the
reference case study, χ 3,2 α = 7.81 ). The numerical values of the T j2 statistics are
summarized in Table 9.7, while the T 2 control chart is depicted in Figure 9.11.
Table 9.6 Scores associated with the first three retained principal components (PCs) for each
profile
Profile index z j1 z j2 z j3
1 –0.0437 0.0032 –0.0290
2 –0.0264 0.0317 0.0318
3 0.0083 0.0172 0.0087
4 0.0182 –0.0097 0.0063
5 –0.0109 –0.0181 –0.0206
6 0.0031 0.0022 0.0073
7 –0.0412 –0.0375 –0.0062
8 0.0392 –0.0247 –0.0252
9 0.0353 –0.0043 0.0101
10 –0.0237 –0.0058 –0.0121
11 0.0085 –0.0174 0.0234
12 –0.0140 0.0134 0.0072
Table 9.7 T 2 statistics for roundness profiles: case of known in-control parameters (phase II)
Figure 9.10 The first three eigenfunctions are related to the three retained PCs: a The first PC
(bilobe contour) describes 13.87% of the variability, b the second PC (a mixture of a bilobe and a
trilobe contour) accounts for the 10.41% of the total variability, and c the third PC (trilobe con-
tour) accounts for 7.10% of the total variability
Figure 9.11 PC-analysis-based multivariate control chart assuming known in-control parame-
ters (phase II)
9.4 Conclusions
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
Chang TC, Gan FF (2006) Monitoring linearity of measurement gauges. J Stat Comput Simul
76:889–911
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
Cressie NAC (1993) Statistics for spatial data. Wiley, New York
Ding Y, Zeng L, Zhou S (2006) Phase I analysis for monitoring nonlinear profiles in manufactur-
ing processes. J Qual Technol 38:199–216
Gupta S, Montgomery DC, Woodall WH (2006) Performance evaluation of two methods for
online monitoring of linear calibration profiles. Int J Prod Res 44:1927–1942
Jackson JE (2003) A user’s guide to principal components. Wiley, New York
Jensen WA, Birch JB, Woodall WH (2008) Monitoring correlation within linear profiles using
mixed models. J Qual Technol 40:167–183
Jin J, Shi J (1999) Feature-preserving data compression of stamping tonnage information using
wavelets. Technometrics 41:327–339
Jin J, Shi J (2001) Automatic feature extraction of waveform signals for in-process diagnostic
performance improvement. J Intell Manuf 12:257–268
Jolliffe IT (2002) Principal component analysis. 2nd edn. Springer, New York
Kang L, Albin SL (2000) On-line monitoring when the process yields a linear profile. J Qual
Technol 32:418–426
Kim K, Mahmoud MA, Woodall WH (2003) On the monitoring of linear profiles. J Qual Tech-
nol 35:317–328
LeSage JP (1999). The theory and practice of spatial econometrics. Available via
http://www.spatialeconometrics.com/
MacGregor JF, Kourti T (1995) Statistical process control of multivariate processes. Control Eng
Pract 3:403–414
9 Model-based Approaches for Quality Monitoring of Geometric Tolerances 283
Mahmoud MA, Woodall WH (2004) Phase I analysis of linear profiles with calibration applica-
tions. Technometrics 46:377–391
Mahmoud MA, Parker PA, Woodall WH, Hawkins DM (2007) A change point method for linear
profile data. Qual Reliab Eng Int 23:247–268
Montgomery DC (2000) Introduction to statistical quality control, 4th 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
Nomikos P, MacGregor JF (1995) Multivariate SPC charts for monitoring batch processes.
Technometrics 37:41–59
Ramsay JO, Silverman BW (2005) Functional data analysis. 2nd edn. Springer, New York
Stover FS, Brill RV (1998) Statistical quality control applied to ion chromatography calibrations.
J Chromatogr A 804:37–43
Walker E, Wright SP (2002) Comparing curves using additive models. J Qual Technol 34:118–
129
Whittle P (1954) On stationary processes in the plane. Biometrika 41:434–449
Williams JD, Woodall WH, Birch JB, Sullivan JH (2006) Distribution of Hotelling’s T2 statistic
based on the successive differences estimator. J Qual Technol 38:217–229
Williams JD, Woodall WH, Birch JB (2007) Statistical monitoring of nonlinear product and
process quality profiles. Qual Reliab Eng Int 23:925–941
Woodall WH (2007) Current research on profile monitoring. Producao 17:420–425
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
Young TM, Winistorfer PM, Wang S (1999) Multivariate control charts of MDF and OSB verti-
cal density profile attributes. For Prod J 49:79–86
Zhang H, Albin S (2009) Detecting outliers in complex profiles using a 2 control chart method
IIE Trans 41:335–345
Zou CL, Zhang YJ, Wang ZJ (2006) A control chart based on a change-point model for monitor-
ing linear profiles. IIE Trans 38:1093–1103
Zou CL, Tsung FG, Wang ZJ (2008) Monitoring profiles based on nonparametric regression
methods. Technometrics 50:512–526
Chapter 10
A Model-free Approach for Quality Monitoring
of Geometric Tolerances
__________________________________
M. Pacella
Dipartimento di Ingegneria dell’Innovazione, Università del Salento,
Via per Monteroni, 73100 Lecce, Italy,
e-mail: massimo.pacella@unisalento.it
Q. Semeraro
Dipartimento di Meccanica, Politecnico di Milano,
Via la Masa, 20156 Milan, Italy,
e-mail: quirico.semeraro@polimi.it
A. Anglani
Dipartimento di Ingegneria dell’Innovazione, Università del Salento,
Via per Monteroni, 73100 Lecce, Italy,
e-mail: alfredo.anglani@unisalento.it
285
286 M. Pacella et al.
10.1 Introduction
quality practitioners. This is the aim in the present chapter, in which a different
approach is discussed for profile monitoring. This approach does not require an
analytical model for the statistical description of the profiles considered, and it
does not involve control charting. In particular, an algorithm which allows com-
puters to automatically learn from data the relationship to represent the profiles in
space, is described. This algorithm is commonly referred to as a neural network
and the data set from which the relationship is automatically learned consists just
of profiles representative of the process in its natural, or in-control, state. Further-
more, the neural network can produce a signal when an input profile does not fit,
according to a specific criterion, to the learned relationship to represent the pro-
files in space.
The approach presented in this chapter can be exploited for profile monitor-
ing when a geometric tolerance is the quality characteristic of interest. The neu-
ral network is defined as a model-free approach because it can be exploited
when an analytical model for the statistical description of the profiles considered
is not available. Furthermore, learning from in-control data and profile monitor-
ing can be automatically implemented with a computer. The practitioner just
needs to set one input parameter of the computer algorithm, depending on the
performance in terms of the false-alarm rate required for the actual application.
The specific method proposed in the present chapter is based on the adaptive
resonance theory (ART) neural network trained by an unsupervised approach.
The proposed approach will be mainly applicable to quality characteristics
which are related to a two-dimensional profile (e.g., roughness, waviness,
roundness, straightness).
An approach based on a ART neural network was first presented by Pacella and
Semeraro (2007b) as an adaptation of their previous work, related to univariate
quality characteristics (Pacella et al. 2004a, b), to the case of geometric tolerances
on manufactured items, with a focus on monitoring roundness. Indeed, profile
monitoring based on the use of neural networks is a promising area of research
since a neural network, when compared with other profile monitoring approaches,
has the main advantage of being easily implemented in practice. Furthermore,
recent achievements in the field of neural network theory related to the develop-
ment of new models able to handle functional data (Rossi and Conan-Guez 2005;
Rossi et al. 2005) open interesting scenarios for future research on innovative
approaches for automatic profile monitoring.
The remainder of this chapter is organized as follows. Section 10.2 gives a
brief overview of machine learning and neural network theory. The expert
reader can skip this section and go on to the next one, which provides a review
of the literature on the general topic of neural networks for quality monitoring.
A step-by-step description of the model-free approach based on the ART neural
network is provided in Section 10.4. A verification study on the efficacy of the
neural network approach for profile monitoring is discussed in Section 10.5.
The last section provides the conclusions and some final remarks. An appendix
is also included, which provides some specific details on the neural network
algorithm.
288 M. Pacella et al.
ble to use a quantitative measure, by a qualitative criterion (e.g., how the learning
procedure can help in obtaining understanding about the system).
There are essentially two types of learning procedure, referred to as supervised
and unsupervised tasks, which are briefly discussed in the following subsection.
For a more complete overview of learning theory, there are many standard refer-
ences, such as the books by Vapnik (2000) and Bishop (2006). Also, the first part
of the book by Bishop (1995) presents the subject of pattern recognition.
The basic idea of supervised learning is that data samples are composed of an
input and a target part. The target is usually a scalar value, while the input is often
described by a vector. The goal of the supervised learning is to derive from the
data set the dependency of targets on inputs, so that the model is capable of return-
ing target predictions for new inputs.
When one has some knowledge of the structures that should be present in the
data, a supervised learning procedure can be adopted. However, in practical appli-
cations there is no such prior knowledge of the rules in the system, either because
it is too complex to characterize these rules or because the system is significantly
stochastic. This is typically the case of manufacturing processes, where observed
data are noisy, dependencies are highly stochastic, and there is no simple physical
rule to represent them.
When one has little knowledge about the patterns present in the data, looking
for clusters in the observed data is a good starting point. The second type of learn-
ing task is the unsupervised one, a.k.a. clustering. The goal of clustering is to
identify inherent separations in the data. One could view clustering as a classifica-
tion problem without label information. Data belonging to a cluster should be
close to one another, while data from different clusters should be far apart. While
in the case of supervised learning adjusting a model requires the definition and the
minimization of a loss function, for unsupervised learning the loss function is a
measure of the similarity of samples within each cluster. Different natural cluster-
ing would be found just by changing the similarity measure. Setting the number of
clusters is also an important issue as it might depend on the similarity measure.
A task related to clustering is quantization (Graf and Luschgy 2000). The aim of
quantization is to encode a large set of input vectors by finding a smaller set of
representative prototypes (or templates). Each prototype provides a good approxi-
mation to the original data in a cluster by summarizing the specific characteristics
of that cluster. The principle is to use a discrete set of prototypes to approach con-
tinuous stochastic vectors based on the minimization of an error cost function. That
means the prototypes can represent the original distribution with the least error.
The basic motivation of vector quantization is dimensionality reduction or data
compression. Actually, the term “quantization” originates in the theory of signal
processing. In this context, “quantization” means a process of discretizing signals.
290 M. Pacella et al.
Σ f
a b
Figure 10.1 a The structure of a three-layer neural network, and b a model of a neuron with
four input connections
10 A Model-free Approach for Quality Monitoring of Geometric Tolerances 291
In the following subsections, two neural networks are discussed, namely, the
multilayer perceptron (MLP) and the ART. The MLP represents the common
model used for supervised learning tasks, while the ART is related to unsupervised
tasks.
MLP is a flexible neural network model configured by setting the number of lay-
ers, the number of neurons in each layer, and the types of link functions.
The training stage of MLP corresponds to the optimization of the different lay-
ers of weights in order to accomplish a specific supervised learning task.
The success of MLP is basically due to the approximation property of this
model. Indeed, theoretical results (Hornik et al. 1989) demonstrate that a MLP
with one hidden layer of nonlinear processing functions (followed by a linear one)
is capable of approximating “any measurable function to any desired degree of
accuracy”. According to the results of Hornik et al. (1989), “… any lack of suc-
cess in applications must arise from inadequate learning, insufficient numbers of
hidden units or the lack of a deterministic relationship between input and target”.
The main drawback with MLP is that this model can easily become “too flexi-
ble” and hence it can be difficult to avoid overfitting of observed data (the net-
work fits the observed data very well but generalizes very poorly). The overfitting
problem (a.k.a. overtraining) is more likely to occur in MLPs than in other models
owing to the large parameter set to be estimated. Indeed, the most difficult prob-
lem is how to develop a network of appropriate size for capturing the underlying
patterns in the data. Although MLP theory suggests that more hidden nodes typi-
cally lead to improved accuracy in approximating a functional relationship, they
also cause the problem of overfitting. Therefore, determination of how many hid-
den layers and hidden neurons to use is often chosen through experimentation or
by trial and error.
Different weight elimination and node pruning methods have been proposed in
the literature for building the optimal architecture of MLP (Reed 1993; Roy et al.
1993; Wang et al. 1994; Murata et al. 1994; Cottrell et al. 1995; Schittenkopf et al.
1997), but none of these methods can guarantee the best solution for every possible
situation. The basic idea with these methods is to find a parsimonious model that fits
the data well. Generally, a parsimonious model not only gives an adequate represen-
tation of the data, but also has the more important generalization capability.
Another way to tackle the overfitting problem is to divide the set of observed
data into three subsets, namely, training, validation, and testing data. The training
and validation parts are used for model building, and testing is used for evaluation
of the model. In particular, the training set is used for computing and updating
iteratively the network weights. During training, the error on the validation set is
monitored, and this set of data is not used for updating the network weights. The
validation error normally decreases during the initial phase of training, as does the
292 M. Pacella et al.
training set error. However, when the network begins to overfit the training data,
the error on the validation set begins to rise, while the training set error continues
to decrease. When the validation error increases for a specified number of itera-
tions, the training is stopped (early stopping). The network with the best perform-
ance on the validation set is then evaluated on the testing data set.
As previously mentioned, the overfitting problem is more likely to occur in a
MLP model since it presents a large parameter set to be estimated. If the number
of parameters in a network is much smaller than the total number of samples in the
training set, then there is little or no chance of overfitting. Furthermore, overfitting
does not apply to some neural network paradigms, such as the unsupervised ones
(e.g., the ART model) because they are not trained using an iterative process.
Recently, MLP models have been adapted to functional data. In particular, a
few studies in which the extension of MLP to functional inputs was proposed from
a theoretical point of view are reported in the literature. One of these approaches
was first proposed by Stinchcombe (1999). Subsequently, Rossi and Conan-Guez
(2005) proposed a functional MLP, although this model suffers from the need for
a specialized implementation and from long training times. Rossi et al. (2005)
proposed another functional MLP based on projection operators. This method has
some advantages over a specialized implementation of MLP, especially because
the projections can be implemented as a preprocessing step that transforms func-
tions into adapted vector representations. The vectors obtained like this are then
processed by a standard MLP model. In general, the functional MLP models pro-
posed in the literature are interesting from a theoretical point of view. From all of
these studies it appears that functional MLPs are a valuable tool for data analysis
when a functional representation of input variables is possible.
ART is an algorithm able to cluster input vectors which resemble each other ac-
cording to the stored prototypes. ART can adaptively create a new cluster corre-
sponding to an input if this specific pattern is not similar to any existing prototype.
In a physical system when a small vibration of a proper frequency produces a
large amplitude vibration, it is defined as resonance. Indeed, ART gets its name
from the fact that information, i.e., the output of neurons, reverberates back and
forth between two layers, namely, F1 (the comparison layer) and F2 (the recogni-
tion layer), which are fully connected by weights. On the one hand, the compari-
son layer (F1) acts as a feature detector that receives external input; on the other
hand, the recognition layer (F2) acts as a category classifier that receives internal
patterns (a schematic representation of the ART model is depicted in Figure 10.2).
The application of a single input vector leads to a set of activity that the neural
network develops in the so-called resonant state and which produces different top-
down templates (prototypes) from layer F2 to layer F1. Each template is associ-
ated with one of the cluster nodes in layer F2.
10 A Model-free Approach for Quality Monitoring of Geometric Tolerances 293
category
index
F2 Reset Signal
Recognition Layer
bottom-up top-down
pattern template
F1
ρ
Comparison Layer
Vigilance
coded Test
input
pattern
Orienting
Subsystem
Figure 10.2 Adaptive resonance theory architecture
The orienting subsystem of the ART model is responsible for generating a reset
signal to the recognition layer when the bottom-up input pattern and the top-down
template mismatch according to a vigilance criterion. This signal, if sent, will
cause either a different cluster to be selected or, if no other cluster is available, the
end of the resonance state. During training, the formerly coded template associated
with the cluster node that represents the best match to the current bottom-up input
will be modified to include the input features. If there are no clusters that match to
the current bottom-up input, a new one is initialized with the incoming pattern.
This is called the vigilance test, and is incorporated in the orienting subsystem of
the neural network. The ART model allows control of the vigilance test, i.e., the
degree of similarity of patterns placed in the same cluster. The similarity depends
on the vigilance parameter ρ, where ρ ∈ [0,1]. If ρ is small, the result is inclined
to a coarse categorization. On the other hand, if ρ is chosen to be close to 1, many
finely divided clusters are formed.
The ART model operates in a plastic mode (i.e., a continuous and cumulative
training mode) as long as new patterns are presented to it. This type of neural
network was firstly introduced to solve the stability/plasticity problem, i.e., to
provide a method by which a neural network can incrementally learn new patterns
without forgetting old knowledge. During training, the ART neural network cate-
gorizes input patterns of data into clusters with similar features, and when it is
confronted by a new input, it produces a response that indicates which cluster the
294 M. Pacella et al.
pattern belongs to (if any). Detecting whether an input vector resembles the natu-
ral categories formed during training is the function of the matching algorithm.
Among different ART models, the fuzzy ART is considered henceforth. It in-
herits the design features of other ART models and incorporates computations
from fuzzy set theory by which it can cluster analog patterns. While a detailed
description of fuzzy ART can be found in the papers of Huang et al. (1995), Geor-
giopoulos et al. (1996, 1999), and Anagnostopoulos and Georgiopouolos (2002),
some analytical details of the matching algorithm along with the step-by-step
implementation of training are given in the Appendix.
hanced and the total training time was increased. They also observed that there is
no established theoretical method to determine the optimal configuration of a MLP
model, thus most of the design parameters must be determined empirically.
Guh and Hsieh (1999) presented a control system composed of several inter-
connected MLPs both to recognize the unnatural patterns and to estimate their
parameters. Guh (2005) proposed a hybrid-learning-based model integrating
MLPs and decision trees as an effective identification system of patterns in the
series of process quality measurements. This approach showed high performance
in detecting and recognizing some unnatural patterns on the control charts.
Generally, the size of a MLP neural network, i.e., the number of model parame-
ters, increases as the number of input vector elements increases. In the literature,
data-window sizes from 16 to 64 observations have been used. Even with 16 data-
window observations, the need to train hundreds of weights to classify unnatural
patterns is a normal requirement.
In most of the approaches proposed in the literature, coding schemes were applied
on the quality measurements after standardization. In the coding process, the meas-
ured variable range was divided into N zones (where the width of each zone was
prespecified), each returning an integer code. The objective of the coding process
was to reduce the effect of the noise in the input data before the data were presented
to the neural network while retaining the main features in the data. The choice of the
coding zone width was critical for MLP classification performance. Gradations that
are too small might not be able to detect the important features in the data owing to
the effect of random noises. On the other hand, if the gradations are too large, the true
process variations might be lost (Cheng 1997; Guh and Tannock 1999).
In the other category, unnatural process behavior detection, Pugh (1991) reported
the first application of a MLP model for mean shift detection in a manufacturing
process. The author also compared the performance of the implemented neural
network with that of Shewhart’s control chart. Smith (1994) implemented a four-
layer MLP to signal shifts in means or variance in X-bar and R control charts.
Chang and Ho (1999) used a neural network to discover shifts in variances in two
steps. The first part is a neural network which decides whether the pattern is in or
out of control. The second part provides a coded value for the shift magnitude.
Similarly, Ho and Chang (1999) proposed a MLP model for monitoring process
mean and variance shifts simultaneously and classifying the types of shifts. The
performance of the MLP model in detecting changes in the process mean was
found to be superior to that of a combined Shewhart–cumulative sum (CUSUM)
control scheme (Cheng 1995). Cheng and Cheng (2001) proposed a MLP model to
monitor exponential mean shifts. Pacella and Semeraro (2007a) proposed a modi-
fied MLP model which employs feedback connections between layers to discover
mean shifts in the case of serially correlated data.
296 M. Pacella et al.
An approach based on a fuzzy ART neural network for profile monitoring was
presented by Pacella and Semeraro (2007b) as an adaptation of their previous
work related to univariate quality characteristics (Pacella et al. 2004a, b) to the
case of geometric tolerances (roundness profiles). The emphasis in their work was
on signaling process changes from the underlying in-control model. In other
words, the fuzzy ART network was not intended to provide a classification of the
pattern detected during the operating phase, but to signal any unnatural behaviors
in profile data.
The main advantage of this approach is that the in-control model is autono-
mously derived by the neural network, without requiring any further intervention
from the analyst.
The issues of using a fuzzy ART neural network for profile monitoring of geo-
metric tolerances are discussed in later subsections. The case study described in
Chapter 11 is considered as a reference. It consists of roundness profiles of
p = 748 entries. Each profile is scaled by subtracting the least-squares estimate of
the radius and centered on the least-squares estimate of the center. Figure 10.3
shows both the polar diagram (Figure 10.3a) and the Cartesian diagram (Fig-
ure 10.3b) of the data set used for training the neural network algorithm.
Figure 10.3 Experimental data after the scaling steps and the alignment phase – 100 roundness
profiles of 748 data each: a polar diagram, and b the corresponding Cartesian diagram
298 M. Pacella et al.
10.4.2 Training
Training of the neural network works as follows. Given a list of profiles ob-
served while the process is in its in-control state, we want the fuzzy ART to
cluster these patterns into a suitable number of categories. The vigilance parame-
ter ρ controls the required degree of similarity among input profiles. This pa-
rameter has to be chosen at the beginning of the training phase on the basis of
the number of clusters, as well as on the fineness of each single cluster, we want
to obtain.
In principle, ρ should be chosen in order to maintain the false-alarm rate
about equal to a predefined value (Pacella et al. 2004a, b; Pacella and Semeraro
2005). This serves to provide an unbiased comparison of the approach with other
techniques when the process drifts to unnatural states. To this aim the set of in-
control profiles is split into two subsets. We refer to these two subsets as the
training and testing lists, respectively (note that as overfitting does not apply to
fuzzy ART, validation data are not required). First, we train the neural network
on the training list by using a given vigilance. Then, we check the performance
of the trained network, in terms of false alarms produced on profiles of the test-
ing list, using the same vigilance value. Training and testing are then repeated in
order to test the effect of different settings of the vigilance parameter.
Generally, the vigilance factor may have a dual effect on the false-alarm rate.
On the one hand, a greater value of vigilance may cause the false-alarm rate to
10 A Model-free Approach for Quality Monitoring of Geometric Tolerances 299
1. Consider the fuzzy ART network with p nodes in the input layer (and hence
with 2 p nodes in field F1). Initialize the iteration index i = 1 and the vigilance
step s = 0.0001 . Consider a training list of m natural profiles (in our study
m = 100 ).
2. Train the neural network on the given training list using a vigilance parameter
of ρi = 1 − i ⋅ s .
3. Repeat step 2, by setting i = i + 1 , until a single category (committed node) in
layer F2 is obtained. Set ρu equal to the maximum value inducing one cate-
gory. From this step on, only values in [ 0, ρu ] are considered for the vigilance.
It should be noted that when one category is formed, there exists no competi-
tion among alternative committed nodes in layer F2 and the choice parameter
of the network has no influence (Georgiopoulos et al. 1996, 1999).
4. Reinitialize the iteration index i = 0 and use vigilance step s = 0.00001 . Con-
sider a testing list of M natural profiles (in our study M = 100 ).
5. Train the neural network on the training list of m natural profiles by using a
vigilance parameter of ρi = ρu − i ⋅ s .
6. Disengage learning and check the performance on the M profiles in the testing
list. The goal is to evaluate the number of vectors in the testing list that mis-
match to the category formed during training in step 5. Set a equal to this num-
ber of false alarms.
7. The estimated false-alarm rate is αˆ = a M (with an approximate standard error
of αˆ (1 − αˆ ) M ). Repeat steps 5 and 6 by setting i = i + 1 until α̂ is equal to
an acceptable rate. Set ρ equal to the vigilance parameter eventually found.
300 M. Pacella et al.
In this section, a verification study on the efficacy of the proposed neural network
approach is described. To this aim, computer simulation is used.
As observed in Chapter 9, roundness profiles can be described by a harmonic
model combined with random residual errors. In this chapter, a simplified model is
used for simulation, in which a common autoregressive model of order 2 – AR(2)
– is used for the residuals. Although AR(2) is not an accurate model for describing
the spatial correlation among adjacent measurements on physical profiles, it can
be adequate for a simplified modeling (Moroni and Pacella 2008).
Let y j ( k ) represent the radial deviation from the nominal radius measured at
the angular position θ k = ( k − 1) 2pπ , where j = 1, 2,… is the index of the profile.
The simulation model is
2
y j (k ) = ⎡c1 j cos ( 2θ k ) + c2 j sin ( 2θ k ) + c3 j cos ( 3θ k ) + c4 j sin ( 3θ k ) ⎤⎦ + υ j ( k ) ,
p⎣
1
υ j (k ) = ε j (k ),
1 − c5 j B - c6 j B 2
(10.1)
Table 10.1 Values of μ and Σ for the simulation model in Equation 10.1
In order to verify the efficacy of the approach, the ability to detect unnatural
patterns during phase II is estimated. To this aim, occurrences of assignable causes
are simulated by two spindle-motion errors as follows (Cho and Tu 2001).
Bilobe out of control, which is simulated by increasing the amplitude of the
second harmonic in the baseline model:
2
y j (k ) + δ 2 ⎡ c1 j cos ( 2θ k ) + c2 j sin ( 2θ k ) ⎤⎦ , (10.2)
p ⎣
where c1 j and c2 j are the coefficients in the baseline model related to the second
harmonic, used to simulate the in-control profile of index j . δ 2 is the increasing
factor for the amplitude of the second harmonic. Values of δ 2 = 0.25, 0.5, 0.75,1
are considered to simulate different severities of this out-of-control condition.
Trilobe out of control, which can be simulated by increasing the amplitude of
the third harmonic with respect to the baseline model:
2
yj (k ) + δ 3 ⎡c3 j cos ( 3θ k ) + c4 j sin ( 3θ k ) ⎤⎦ , (10.3)
p ⎣
where c3 j and c4 j are the coefficients of the third harmonic in the baseline model,
for the in-control profile of index j . δ 3 is the increasing factor for the amplitude
of the third harmonic. Values of δ 3 = 0.25, 0.5, 0.75,1 are considered.
Table 10.2 Experimental results of the fuzzy adaptive resonance theory (ART) neural network:
a the number of categories formed during training (steps 2 and 3), and b the false-alarm rate
during tuning (steps 5–7). The value of the vigilance parameter eventually used is highlighted in
bold
a b
i ρi Categories i ρi Alarms
1 1.0000 100 0 0.98590 22
2 0.9999 100 1 0.98589 16
… … … 2 0.98588 15
21 0.9980 80 3 0.98587 12
22 0.9979 78 4 0.98586 9
… … … 5 0.98585 7
41 0.9960 37 6 0.98584 7
… … … 7 0.98583 6
61 0.9940 23 8 0.98582 5
… … … 9 0.98581 4
81 0.9920 13 10 0.98580 4
… … … 11 0.98579 4
101 0.9900 7 12 0.98578 4
… … … 13 0.98577 4
111 0.9890 4 14 0.98576 4
… … … 15 0.98575 3
121 0.9880 3 16 0.98574 3
… … … 17 0.98573 3
140 0.9861 2 18 0.98572 2
141 0.9860 2 19 0.98571 2
142 0.9859 1 20 0.98570 1
data), 1,496 nodes in field F1 (equal to double the number of input nodes) and a
single node in field F2.
The training list consisted of the m = 100 roundness profiles of the real case
study. By applying steps 2 and 3 of the procedure described in Section 10.4.3, we
found the range of vigilance values which allows clustering of training data into
one category. From our experimental results, we observed that fuzzy ART classi-
fied the training set into one category when the value of the vigilance factor is
ρ ∈ [ 0, 0.9859] .
A testing list of M in-control profiles was then considered. In this work, the
testing list was simulated using the in-control model in Equation 10.1. In spite of
this, in actual applications the testing list could be obtained directly from the
process while it is in its in-control state. Indeed, no additional parametric model-
ing of profile data is required for implementation of the proposed neural network
approach.
10 A Model-free Approach for Quality Monitoring of Geometric Tolerances 303
We decided to consider M = 100 in-control profiles also for the testing list.
From our experimental results we observed that such a dimension is adequate to
calibrate the vigilance parameter when the reference false-alarm rate is approxi-
mately 5% (or greater than 5%). However, the lower the false alarm rate is, the
higher the dimension M should be in order to reduce the uncertainty associated
with the estimate of the false-alarm rate produced by the neural network.
By applying steps 5–7 of the procedure described in Section 10.4.3, the vigi-
lance value of ρ = 0.98582 was eventually identified in order to have a false-
alarm rate of approximately 5%.
Table 10.2 details some of the experimental results obtained during the imple-
mentation of the fuzzy ART neural network. In particular, Table 10.2a reports the
effect of the vigilance parameter on the number of categories formed during train-
ing (steps 2 and 3). It can be observed that as the vigilance parameter decreases,
the number of categories formed during training decreases too. Table 10.2b reports
the effect of the vigilance parameter on the number of false alarms observed on
the testing list (steps 5–7). In this case, as the vigilance parameter decreases, the
number of false alarms decreases too.
In order to verify the efficacy of the neural network approach for profile monitor-
ing, we compare its performance with the performance of the individuals control
charts of the out-of-roundness (OOR) values computed by the least-squares and
minimum-zone algorithms (see Chapter 8). A separate control chart was imple-
mented for each of these two methods.
The performance index is the average run length (ARL) required by the com-
peting methods to signal out-of-control states in phase II. The run length is defined
as the number of samples taken until an out-of-control signal is issued.
For each new profile generated, the least-squares and minimum-zone OOR val-
ues were computed. Each of the individuals control charts released an alarm when
the corresponding statistic exceeded the control limits. Each new profile generated
was also presented as an input to the fuzzy ART neural network, formerly trained
on the set of 100 roundness profiles of the reference case study (vigilance parame-
ter 0.98582). The neural network released an alarm when the current input mis-
matched to the natural category formed during training according to the algorithm
described in the Appendix.
Profiles were simulated until each of the three competing approaches (i.e., least
squares, minimum zone, and ART) issued an alarm. For each approach, we stored
the run length (number of samples until the first alarm). This procedure was per-
formed 1,000 times. Hence, a sample of 1,000 run lengths for each case was used
to compute both the ARL value and the corresponding 95% confidence interval
(based on the t-based statistic with 999 degrees of freedom).
304 M. Pacella et al.
Figure 10.4 Bilobe. Confidence interval (95%, 1,000 replications) on the average run length
(ARL) (ordinate) versus the magnitude of the out-of-control effect (abscissa). LS least squares,
MZ minimum zone
Table 10.3 reports the confidence interval for the ARLs obtained by simulating
in-control roundness data. It can be observed that the competing methods have the
same performance when no out-of-control conditions affect profiles data (with a
5% false-alarm rate the expected in-control ARL is 1 0.05 = 20 ). Hence, the per-
formance in phase II is related to the ability to detect out-of-control profiles, given
that all the approaches are designed to achieve the same false-alarm probability of
about equal 5%.
Figures 10.4 and 10.5 graphically depict the simulation results for the two out-
of-control models. In particular, Figure 10.4 refers to bilobe out-of-control errors,
while Figure 10.5 refers to trilobe out-of-control errors. (In both figures, the data
have been jittered on the x-axis to improve the readability). In order to select the
best approach in each case study (where each case study is characterized by a spe-
Table 10.3 ARL and confidence interval (95%, 1,000 replications) of competing methods with
reference to the baseline model
Figure 10.5 Trilobe. Confidence interval (95%, 1,000 replications) on the ARL (ordinate)
versus the magnitude of the out-of-control effect (abscissa)
10.6 Conclusions
We have shown that fuzzy ART can be effectively used to this aim. The ap-
proach is quite simple to automate using a computer through the seven-step
procedure presented in Section 10.4.3 and the fuzzy ART training algorithm (in
the Appendix). In this chapter, we used 100 roundness profiles to train the
neural network, and another 100 roundness profiles to tune the vigilance pa-
rameter, which allows a 5% false-alarm rate. The proposed approach is quite
general and can be easily extended to deal with different two-dimensional geo-
metric specifications.
With reference to the ability to detect out-of-control states in phase II, a verifi-
cation study showed that the proposed neural network approach is more effective
than the usual control charting methods based on the OOR values in detecting two
types of out-of-control conditions (modeled as a systematic form deviation of
profile data due to errors of the spindle motion).
When compared with the methods of multivariate charting of fitted parameter
vectors (e.g., those discussed in Chapter 9), the neural network approach has a
basic advantage, namely, it can be exploited in those applications where an ana-
lytical model for the statistical description of profiles considered is not available.
Furthermore, a simulation study (not reported here) showed that the neural net-
work has good robustness (i.e., performance is invariant) for moderate contamina-
tion of the training data set by out-of-control profiles (about 5%).
As for the ability in signaling out-of-control states in phase II, the neural net-
work may not produce an outperforming result when compared with the model-
based approaches. As expected, the model-based techniques may be more effec-
tive than the model-free approach described in this chapter in signaling changes in
the functional data (for specific production scenarios). Furthermore, an additional
disadvantage of the algorithm presented in this chapter is that such an approach is
limited to phase II of profile monitoring only, while, in general, the control chart
approach can be used for both phase I and phase II.
Despite these drawbacks, profile monitoring based on the use of neural net-
works is a very promising area of research in the computer-integrated manufactur-
ing field. Indeed, the automation of quality monitoring is becoming more and
more important in practical applications, owing to the availability of affordable
point-based measurement systems capable of acquiring large amounts of data as
point coordinates in reasonable times and with high accuracy. The possibility of
managing large data sets, automatically acquired from the process, using auto-
matic online procedures implemented by computers paves the way to the devel-
opment of innovative approaches in the field of quality monitoring.
In order to select a specific method in a given production scenario, a compari-
son of the performance of the neural network approach with the performances of
the methods presented in the previous chapters is given in Chapter 11.
10 A Model-free Approach for Quality Monitoring of Geometric Tolerances 307
Appendix
Fuzzy ART operates over all of the committed nodes along with a single uncommit-
ted node. Each committed node (of index j ) has a vector w j = ( w j1 w j 2 ... w j 2 p ) of
adaptive weights (which represents the coded template). The number of committed
nodes n ( j = 1, 2, … , n ) is arbitrary, while the dimension of vector w j is 2 p .
Let x be a p-dimensional input vector x = ( x1 ... xk ... x p ) , where each compo-
nent xt ranges in [ 0,1] . The matching algorithm of fuzzy ART is as follows.
Step 1: Initialization
During training, initialize the number of committed nodes to n = 0 . Then, set a
choice parameter α ∈ [ 0, ∞ ] (a small value is used in this work: α = 10−6 ), and a
vigilance parameter ρ ∈ [ 0,1] .
xc ∧ w j
Tj ( x ) = , (10.5)
α + wj
where operator ∧ gives the vector u ∧ v = ( min {u1 , v1} min {ut , vt } ) , while
operator gives the scalar u = ∑ t ut .
Choose the committed node in layer F2 that receives the maximum bottom-up
input. Assume this node has index j * : T j* = max {T j : T j ≥ 0, j = 1,… , n} . If more
than one T j* is maximal, choose the category with the smallest j index.
308 M. Pacella et al.
x c ∧ w j*
≥ρ. (10.6)
xc
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
Al-Ghanim A (1997) An unsupervised learning neural algorithm for identifying process behavior
on control charts and a comparison with supervised learning approaches. Comput Ind Eng
32:627–639
Anagnostopoulos GC, Georgiopouolos M (2002) Category regions as a new geometrical con-
cepts in fuzzy-ART and fuzzy-ARTMAP. Neural Netw 15:1205–1221
Bishop CM (1995) Neural networks for pattern recognition. Oxford University Press, New York
Bishop CM (2006) Pattern recognition and machine learning. Springer, New York
Chang SI, Ho ES (1999) A two-stage neural network approach for process variance change
detection and classification. Int J Prod Res 37:1581–1599
Cheng CS (1995) A multi-layer neural network model for detecting changes in the process mean.
Comput Ind Eng 28:51–61
Cheng CS (1997) A neural network approach for the analysis of control chart patterns. Int J Prod
Res 35:667–697
10 A Model-free Approach for Quality Monitoring of Geometric Tolerances 309
Cheng CS, Cheng SS (2001) A neural network-based procedure for the monitoring of exponen-
tial mean. Comput Ind Eng 40:309–321
Cho NW, Tu JF (2001) Roundness modeling of machined parts for tolerance analysis. Precis Eng
25:35–47
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
Cottrell M, Girard B, Girard Y, Mangeas M, Muller C (1995) Neural modeling for time series:
a statistical stepwise method for weight elimination. IEEE Trans Neural Netw. 6:1355–1364
Georgiopoulos M, Fernlund H, Bebis G, Heileman GL (1996) Order of search in fuzzy ART and
fuzzy ARTMAP: effect of the choice parameter. Neural Netw 9:1541–1559
Georgiopoulos M, Dagher I, Heileman GL, Bebis G (1999) Properties of learning of a fuzzy
ART variant. Neural Netw 12:837–850
Graf S, Luschgy H (2000) Foundations of quantization for probability distributions. Lecture
notes in mathematics, vol 1730. Springer, Berlin
Guh RS (2005) A hybrid learning-based model for on-line detection and analysis of control chart
patterns. Comput Ind Eng 49:35–62
Guh RS, Hsieh YC (1999) A neural network based model for abnormal pattern recognition of
control charts. Compu. Ind Eng. 36:97–108
Guh RS, Tannock JDT (1999) Recognition of control chart concurrent patterns using a neural
network approach. Int J Prod Res 37:1743–1765
Gupta S, Montgomery DC, Woodall WH (2006) Performance evaluation of two methods for
online monitoring of linear calibration profiles. Int J Prod Res 44:1927–1942
Haykin S (1998) Neural networks. A comprehensive foundation, 2nd edn. Macmillan, New York
Ho ES, Chang SI (1999) An integrated neural network approach for simultaneous monitoring of
process mean and variance shifts-a comparative study. Int J Prod Res 37:1881–1901
Hornik K, Stinchcombe M, White H (1989) Multilayer feedforward networks are universal
approximators. Neural Netw 2:359–366
Huang J, Georgiopoulos M, Heileman GL (1995) Fuzzy ART proprieties. Neural Netw 8:203–
213
Hwarng HB, Hubele NF (1993a) Back-propagation pattern recognizers for X-bar control charts:
methodology and performance. Comput Ind Eng 24:219–235
Hwarng HB, Hubele NF (1993b) X-bar control chart pattern identification through efficient off-
line neural network training. IIE Trans 25:27–40
Jones MC, Rice JA (1992) Displaying the important features of large collections of similar
curves. Am Stat 46:140–145
Kang L, Albin SL (2000) On line monitoring when the process yields a linear profile. J Qual
Technol 32:418–426
Kim K, Mahmoud MA, Woodall WH (2003) On the monitoring of linear profiles. J Qual Tech-
nol 35:317–328
Mahmoud MA, Woodall WH (2004) Phase I analysis of linear profiles with calibration applica-
tions. J Qual Technol 46:380–391
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
Murata N, Yoshizawa S, Amari S (1994) Network information criterion-determining the number
of hidden units for an artificial neural network model. IEEE Trans Neural Netw 5:865–872
Pacella M, Semeraro Q (2005) Understanding ART-based neural algorithms as statistical tools
for manufacturing process quality control. Eng Appl Artif Intell 18:645–662
Pacella M, Semeraro Q (2007a) Using recurrent neural networks to detect changes in autocorre-
lated processes for quality monitoring. Comput Ind Eng 52:502–520
310 M. Pacella et al.
Pacella M, Semeraro Q (2007b) An unsupervised neural network approach for on-line monitor-
ing of machined profiles. Paper presented at the 7th annual conference of ENBIS (European
Network for Business and Industrial Statistic), Dortmund, Germany, 24–26 September
Pacella M, Semeraro Q, Anglani A (2004a) Manufacturing quality control by means of a fuzzy
ART network trained on natural process data. Eng Appl Artif Intell 17:83–96
Pacella M, Semeraro Q, Anglani A (2004b) Adaptive resonance theory-based neural algorithms
for manufacturing process quality control. Int J Prod Res 40:4581–4607
Pugh GA (1991) A comparison of neural networks to SPC charts. Comput Ind Eng 21:253–255
Reed R (1993) Pruning algorithms – a survey. IEEE Trans Neural Netw 4:740–747
Rossi F, Conan-Guez B (2005) Functional multi-layer perceptron: a nonlinear tool for functional
data analysis. Neural Netw 18:45–60
Rossi F, Delannay N, Conan-Guez B, Verleysen M (2005) Representation of functional data in
neural networks. Neurocomputing 64:183–210
Roy A, Kim LS, Mukhopadhyay S (1993) A polynomial time algorithm for the construction and
training of a class of multilayer perceptrons. Neural Netw 6:535–545
Schittenkopf C, Deco G, Brauer W (1997) Two strategies to avoid overfitting in feedforward
networks. Neural Netw 10:505–516
Smith AE (1994) X-bar and R control chart interpretation using neural computing. Int J Prod Res
32:309–320
Stinchcombe MB (1999) Neural network approximation of continuous functionals and continu-
ous functions on compactifications. Neural Netw 12:467–477
Vapnik VN (2000) The nature of statistical learning theory, 2nd edn. Springer, New York
Walker E, Wrigth SP (2002) Comparing curves using additive models. J Qual Technol 34:118–
129
Wang Z, Di Massimo C, Tham MT, Morris AJ (1994) A procedure for determining the topology
of multilayer feedforward neural networks. Neural Netw 7:291–300
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
Zorriassantine F, Tannock JDT (1998) A review of neural networks for statistical process con-
trol. J Intell Manuf. 9:209–224
Chapter 11
Quality Monitoring of Geometric Tolerances:
a Comparison Study
Abstract While in the previous chapters different approaches for quality moni-
toring of geometric tolerances were discussed, in the present one a comparison
study is provided in which all of these approaches are considered. The aim is to
allow practitioners to select a specific method in a given situation. A manufactur-
ing reference case study is first detailed, namely, profiles measured on machined
items subject to geometric specification (roundness). Then, two simulation scenar-
ios are considered for the comparison study, where each scenario is obtained by
perturbing the real case study. Competing approaches are ranked in each produc-
tion scenario considering as a performance index the quickness in detecting out-
of-control shapes.
11.1 Introduction
__________________________________
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
311
312 B.M. Colosimo and M. Pacella
set of data observed in a profile exceeds the control limits. 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.
Similar to a neural network, the location control chart does not require an ana-
lytical model for the statistical description of the profiles considered (model-free
approach). Differently from a neural network, the location control chart has the
main advantage of visually representing the control region around the ideal or
mean shape of profiles. This can help the practitioner to identify the potential
causes of alarms, while a neural network algorithm is a black box, which makes it
more difficult to realize why an input profile produces or does not produce an
alarm.
Among the competing approaches, the individuals control chart of the geomet-
ric error characterizing the profiles is also considered. This approach, described in
Chapter 8, consists in summarizing all the information contained in the functional
data in just one synthetic variable. This variable measures the geometric form
error as the (maximum) distance between the actual profile and the ideal geome-
try. Then the usual control charting methods can be applied to the estimated geo-
metric errors for quality monitoring. With reference to roundness, the geometric
error is summarized in a value called out of roundness (OOR). In this chapter, the
least-squares algorithm is used to estimate the OOR value for a roundness profile.
Both the location control chart and the individuals control chart of the geomet-
ric error represent approaches at a lower level of complexity, because they only
require the use of common control charting for profile monitoring. Thanks to their
inner simplicity, these approaches may be widely preferred by practitioners and, in
fact, they can be considered representative of industrial practice. On the other
hand, at a higher complexity level, approaches based on combining an analytical
model of functional data with multivariate and univariate control charting repre-
sent innovative methods which have been designed with the specific objective to
deal with functional data. At an intermediate level of complexity, a neural net-
work, which represents a general-purpose algorithm for a computer to automati-
cally learn a relationship from data, is also considered.
In order to allow practitioners to select a specific approach for monitoring func-
tional data in a given situation, this chapter provides a numerical comparison,
based on simulation, of (1) the location control chart, (2) the individuals control
chart of the geometric error characterizing the profiles, (3) the regression-based
approach, (4) the PCA-based approach, and (5) the adaptive resonance theory
neural network.
The comparison study is based on two simulation scenarios which are obtained
by perturbing a case study of machined profiles subject to geometric specification
(roundness profiles obtained by turning, where each curve is characterized by 748
evenly distributed measurements). Each measurement on a profile is the radial
distance at a specific angular position. The perturbed scenarios are designed to
represent two different although realistic productive situations.
The performances of the competing approaches are measured as the ability to de-
tect unusual patterns in the functional data during the operating phase of the control
314 B.M. Colosimo and M. Pacella
chart (known as phase II). Basically, the objective in this chapter is to investigate
situations where each approach should be preferred to the others, thus to provide
some guidelines for implementing profile monitoring in actual applications.
The remainder of this chapter is organized as follows. In Section 11.2, the ref-
erence case study of roundness profiles obtained by turning is discussed in detail.
In Section 11.3, the simulation models for each production scenario obtained from
the reference case study are presented. In Section 11.4, the out-of-control models
considered in the study are discussed. In Section 11.5, the approaches for profile
monitoring are compared with reference to phase II of profile monitoring. Sec-
tion 11.6 presents the conclusions and some final remarks.
A turning process was used to machine 100 items starting from C20 carbon steel
cylinders, which were supplied in rolled bars (original diameter 30 mm) and were
machined to a nominal diameter of 26 mm (using a cutting speed of 163 m/min,
a feed of 0.2 mm/rev, and two cutting steps of 1-mm depth each). Figure 11.1 shows
a cylindrical item before machining and one after machining by lathe turning.
Each cylindrical surface obtained by lathe turning was measured by using a co-
ordinate measuring machine (CMM). Sampling was performed by continuous
scanning of 748 generatrices on each turned specimen and extracting from each
cylindrical surface ten roundness profiles (at different distances from the spindle).
A statistical analysis of the data obtained showed that the roundness profile
changes as the distance from the spindle changes. The technological reason behind
this behavior is due to a different inflection of the workpiece while moving far
from the spindle. The case study considered in this chapter refers to the profile that
is the most distant from the spindle (represented in Figure 11.2). This profile is
extracted from each of the 100 items machined.
Figure 11.3 shows an item measured by continuous scanning of its surface on
a CMM. Each roundness profile sampled on an item consisted of 748 equally
spaced measurements of the radius. The original measurements sampled using the
CMM were scaled by subtracting the least-squares estimate of the radius and cen-
tered at the least-squares estimate of the center.
A roundness profile can be described by a polar representation where a set of
points, evenly distributed on the machined surface, represent the deviation of the
actual measurement from the nominal radius at different angle locations (Cho and
Tu 2001). Note that such deviations can be either positive or negative. A polar
representation of the whole experimental data set is shown in Figure 11.4. The
corresponding Cartesian diagram, in which the independent variable is the location
index on the part and the dependent one is the deviation from the nominal radius, is
depicted in Figure 11.5.
11 Quality Monitoring of Geometric Tolerances: a Comparison Study 315
Figure 11.1 Two cylindrical items of the reference case study: a rough piece and a final piece
after machining by lathe turning
Figure 11.2 The component machined by lathe turning (the arrow shows the roundness profile
under study)
Figure 11.3 An item of the reference case study measured by using a coordinate measuring
machine where sampling was performed by continuous scanning of 748 generatrices
316 B.M. Colosimo and M. Pacella
Figure 11.4 Polar diagram of experimental data of real roundness profiles. One hundred round-
ness profiles of 748 points each. One of the 100 profiles is depicted as a bold line
Figure 11.5 Cartesian diagram of experimental data of real roundness profiles. One hundred
roundness profiles of 748 points each. One of the 100 profiles is depicted as a bold line
From Figures 11.4 and 11.5 it seems that no systematic pattern characterizes the
roundness profiles of the reference case study. This appearance hides a common
problem of shape analysis, which consists in feature registration or alignment. In
fact, the profiles shown in Figures 11.4 and 11.5 are actually misaligned because
of the random contact angle of the turning process. Therefore, a further step is
11 Quality Monitoring of Geometric Tolerances: a Comparison Study 317
Figure 11.6 Aligned experimental data of real roundness profiles: polar diagram. One of the
100 profiles is depicted as a bold line
Figure 11.7 Aligned experimental data of real roundness profiles: Cartesian diagram. One of
the 100 profiles is depicted as a bold line
needed in order to register all the sampled profiles by minimizing the phase delay
caused by the random contact angle (Colosimo and Pacella 2007).
Figures 11.6 and 11.7 show, respectively, the polar and the Cartesian diagrams
of the aligned roundness profiles after a proper registration procedure has been
implemented on the data. From a visual inspection, it can be easily observed that
318 B.M. Colosimo and M. Pacella
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).
The following subsection gives the technical details on the registration proce-
dure implemented on the roundness profile of the reference case study. The reader
may skip these details without loss of continuity.
Variation among profile data can be considered as being composed of two compo-
nents: location (or phase) variability and amplitude variability. A usual prelimi-
nary step in profile data analysis is the registration (or alignment) of profiles,
which is performed with the aim of removing, or minimizing, phase variation so
that the analyses can focus on amplitude effects only.
Let us start with the simple case in which two profiles have to be registered.
Without loss of generality, the first profile can be used as a reference and the sec-
ond one can be registered with respect to this reference profile. When profiles
refer to closed curves, as happens for the roundness profile, registration can be
performed via circular shifts. A circular shift consists in moving one position
ahead all the data observed on the profile (where “circular” refers to the fact that
the last observation becomes the first one after the shift has been performed).
Since each profile is observed on a fixed set of equally spaced p locations, the
number of possible circular shifts is equal to the number of points measured on the
profile minus one ( p − 1) . Registration consists in determining the best number of
circular shifts that have to be performed on the current profile in order to minimize
the phase delay between the current and the reference profiles.
As an example, Figure 11.8a shows the reference profile and the current pro-
file. Both roundness profiles have a similar shape (bilobe), while they differ in
both amplitude and phase. The correlation between the two profiles is –0.715.
Figure 11.8 Polar diagrams of two ideal bilobe profiles (reference profile bold line): a mis-
aligned profiles, and b aligned profiles
11 Quality Monitoring of Geometric Tolerances: a Comparison Study 319
By performing all the possible circular shifts of the current profile, we find that
the best number of circular shifts is the one corresponding to the highest correlation
between the current profile and the reference profile. In this simple example, since
there is no additional noise, the highest correlation 1; the corresponding alignment
between the reference profile and the current profile is shown in Figure 11.8b.
As shown in this example, correlation between profiles can act as a perform-
ance criterion in the registration procedure. Note that covariance (correlation)
refers to data observed for different profiles (cross-correlation).
Given two profiles, the correlation between the i th and the j th profile is given
sij 1 p
by rij =
si s j
, where sij = ∑ [ yi (k ) − yi ] ⎡⎣ y j (k ) − y j ⎤⎦ is the covariance be-
p − 1 k =1
1 p
∑
2
tween the i th and the j th profile, s 2j = ⎡⎣ y j ( k ) − y j ⎤⎦ is the variance
p − 1 k =1
1 p
associate with the i th profile, and y j = ∑ y j (k ) is the mean value of each pro-
p k =1
file (which is equal to 0 if the least-squares estimator of the radius is subtracted
from each profile as the first step). Therefore, in order to align the i th profile with
respect to the j th one, all the possible circular shifts of the ith profile are per-
formed and the corresponding correlation indexes rij are computed. The best circu-
lar shift is the one corresponding to the maximum value of the correlation index.
This procedure allows one to register a set of profiles with respect to a given
one. In fact, if the jth profile is used as a reference, all the remaining n – 1 profiles
can be registered with respect to this reference profile. The result of this registra-
tion procedure is summarized by an n × n symmetric correlation matrix R ( j ) ,
where the subindex j is used to denote the profile which was used as a reference in
this registration step.
This procedure can be repeated by considering in turn each of the n profiles as
the reference one. Eventually, n possible registration configurations are available,
each characterized by a specific R ( j ) ( j = 1,..., n ). In order to select the best regis-
tration among this set of n possible alternatives, denote with λ1( j ) , λ2( j ) ,… , λn ( j )
the eigenvalues of R ( j ) . Since the correlation matrix is symmetric, these eigen-
values are real, nonnegative, and their sum is equal to the trace of the correlation
n
matrix (i.e., the sum of the diagonal element ∑λ
i =1
i( j) = n ). Consider two extreme
cases. When there is no cross-correlation among the profiles (worst case), the
correlation matrix R ( j ) is an identity matrix and λ1( j ) = λ2( j ) = … = λn ( j ) = 1 . On
the other hand, when ideally there is the maximum cross-correlation among the
profiles (best case), the correlation matrix is an “all-one” matrix with rank equal
to 1. In this case, all the eigenvalues but the first one are equal to zero, since the
rank of a symmetric matrix is equal to the number of eigenvalues that are greater
than zero. Furthermore, since the sum of the eigenvalue is always equal to the
320 B.M. Colosimo and M. Pacella
trace of the matrix, the first eigenvalue is equal to the trace, i.e., λ1( j ) = n and
λ2( j ) = … = λn ( j ) = 0 . Consider now the summation of the squared eigenvalues
n
∑ (λ )
2
i . In the worst case it is equal to n, while in the best case it is equal to n2.
i =1
In any other case, since the eigenvalues are nonnegative, the sum of squared ei-
n
genvalues ranges between n and n 2 , i.e., n ≤ ∑ ( λi ( j ) ) ≤ n 2 . Therefore, the best
2
i =1
configuration is the one which maximizes the sum of the squared eigenvalues.
The registration procedure is eventually summarized in the following steps:
1. Set j = 1 as a reference sample. For each of the remaining i = 1,..., n profiles,
where i ≠ j , perform all the possible p − 1 circular shifts and select the one
that maximizes rij . Denote by R ( j ) the correlation matrix obtained after align-
ing all the samples with the reference one.
2. Set j = j + 1 and repeat step 1 until all the profiles have been considered as
a reference (i.e., until j ≤ n ).
3. Among the n configurations obtained, choose the optimal one, say, j * , which
satisfies the following criterion:
n
j* = arg max j =1…n ∑ ( λi ( j ) ) ,
2
(11.1)
i =1
Chapter 9 details a SARX model of the data in the reference case study previously
described. In this model, the dependent variable y j ( k ) represents the radial de-
viation from the nominal radius measured at the angular positions θ k = ( k − 1) 2pπ ,
where j = 1, 2, … is the index of the profile and k = 1, 2, … p is the index of
equally spaced observations on each profile.
The SARX model is composed of two parts, namely, the large-scale component
and the small-scale component. The large-scale component is modeled as
a combination of two sinusoidal functions of the angular position θ k [called har-
monics, i.e., cos ( hθ k ) and sin ( hθ k ) , where h = 2,3 ]. These two sinusoidal func-
tions are exploited in order to model, respectively, the ovality and the triangularity
of roundness profiles. The small-scale component of the SARX model describes
the correlation structure as a generic spatial autoregressive of order 2.
11 Quality Monitoring of Geometric Tolerances: a Comparison Study 321
Since four parameters are used for the two harmonics and two more parameters
are needed to describe the correlation structure, each profile of index j is associ-
ated with a vector of d = 4 + 2 parameters c j = ⎡⎣c1 j c2 j c3 j c4 j c5 j c6 j ⎤⎦ . The
first four coefficients c1 j , c2 j , c3 j , c4 j are used to represent the parameters of the
harmonics ( c1 j , c2 j refer to the second harmonic, while c3 j , c4 j refer to the third
one). Coefficients c5 j and c6 j refer to the two parameters of the spatial autore-
gressive model.
By combining instances of the d-length parameter vector c j with instances of
the p-length vector ε j (of independently and normally distributed errors with zero
mean and common variance), one can use the SARX model to simulate on
a computer realistic roundness profiles (Colosimo and Pacella 2010). In this chap-
ter, the focus is on two different production scenarios.
On the one hand, the first scenario mimics the case study. In this production
scenario, the d-length parameter vector c j changes from profile to profile accord-
ing to a d-variate normal distribution c j ∼ N ( μ, Σ ) , where N represents
a multivariate normal (Gaussian) distribution of mean μ (d-length vector) and
covariance matrix Σ ( d × d symmetric matrix). The actual values of the mean
vector μ and of the covariance matrix Σ exploited for the simulation of this sce-
nario are summarized in Table 11.1.
On the other hand, the second scenario is obtained by perturbing the SARX
model of the case study with reference to the variability that characterizes the
d-length parameter vector c j . In particular, a null matrix is considered as a covari-
ance matrix ( Σ = 0 ) for the d-variate normal distribution of the model parameters.
Hence, a fixed vector of parameters is exploited for each instance ( c j = μ ).
Table 11.1 Parameters characterizing the mean μ and the variance Σ of the distribution of
coefficients c j ∼ N ( μ, Σ ) for the actual roundness data (Colosimo et al. 2008)
The first production scenario, where model parameters c j can change from
profile to profile, is referred to as random-effect model, and the second production
scenario, in which parameters c j do not change, is referred to as fixed-effect
model. Indeed, a fixed-effect model is usually assumed in traditional approaches
for profile monitoring (Woodall et al. 2004). In manufacturing, assuming that the
input material is very stable and homogeneous, profile-to-profile variability of the
small-scale component (spatial structure) can probably be neglected. Furthermore,
assuming a process is more stable and/or more controlled, also the profile-to-
profile variability in the large-scale component (harmonic structure) tends to van-
ish. In other words, while the first scenario with random effects corresponds to
a common machining process, in which natural causes of variability affect the
variability in both the parameters and the residuals of the model, the second sce-
nario, with fixed effects, can be considered representative of a process that is sta-
bler and/or more controlled in which natural causes can affect only the residuals of
the model.
2 ⎛1 ⎞
yj (k ) + δ1 sin ⎜ θ k ⎟ , (11.2)
p ⎝2 ⎠
where the second term represents the out-of-control state arising in phase II. δ1 is
the size of the shift. Values of δ1 = 0.1, 0.15, 0.2, 0.25 are considered to model
different severities of this out-of-control condition.
11 Quality Monitoring of Geometric Tolerances: a Comparison Study 323
Bilobe out of control, which can be caused by an improper setup of the work-
piece or by an increased bilobe motion error already affecting the spindle lathe.
This out-of-control condition can be simulated by increasing the amplitude of the
second harmonic in the baseline model:
2
y j (k ) + δ 2 ⎡ c1 j cos ( 2θ k ) + c2 j sin ( 2θ k ) ⎤⎦ , (11.3)
p ⎣
where c1 j and c2 j are the coefficients in the baseline model related to the second
harmonic used to simulate the in-control profile of index j . δ 2 is the increasing
factor for the amplitude of the second harmonic. Values of δ 2 = 0.1, 0.2, 0.3, 0.4
are considered to simulate different severities of this out-of-control condition.
Trilobe out of control, which can be due to an increase in the trilobe motion er-
ror already existing in the baseline model of the spindle or to an excessive force
imposed by the clamping fixture. Similarly to the previous case, this third unnatu-
ral condition can be simulated by incrementing the amplitude of the third har-
monic with respect to the baseline model:
2
yj (k ) + δ 3 ⎡c3 j cos ( 3θ k ) + c4 j sin ( 3θ k ) ⎤⎦ , (11.4)
p ⎣
where c3 j and c4 j are the coefficients related to the third harmonic in the baseline
model, used to simulate the in-control profile of index j . δ 3 is the increasing
factor for the amplitude of the third harmonic. Values of δ 3 = 0.1, 0.2, 0.3, 0.4 are
also considered in this case.
Note that the spurious harmonic can influence just one frequency (as for the
bilobe and trilobe, which influence the amplitude of the second and third harmon-
ics, respectively), or more frequencies at once (as in the case of the half frequency,
which also produces shifts in the amplitude of harmonics greater than the third).
Data obtained under these phase II models are also scaled (by subtracting the
least-squares estimation of the radius) and centered (on the least-squares estima-
tion of the center) before applying the profile monitoring method. In fact, we are
assuming that centering and scaling are standard steps applied to data when the
focus is on the out-of-roundness value only (Cho and Tu 2001).
The objective in phase II is to quickly detect any change in the process from its in-
control state. Hence, the monitoring approaches are compared in terms of the
average run length (ARL), where the run length is defined as the number of sam-
ples taken until an out-of-control signal is issued.
324 B.M. Colosimo and M. Pacella
Table 11.2 Phase II simulation results for the production scenario with the random-effect
model. Actual average run lengths (ARLs) and corresponding standard deviations in parentheses
(1,000 trials)
Bilobe 0.1 64.08 (1.92) 72.39 (2.32) 76.41 (2.42) 64.29 (1.89) 54.57 (1.82)
0.2 36.07 (1.16) 44.66 (1.42) 47.48 (1.60) 38.03 (1.22) 29.50 (0.92)
0.3 24.01 (0.77) 25.69 (0.81) 29.76 (0.92) 21.21 (0.67) 18.65 (0.56)
0.4 14.98 (0.45) 18.40 (0.57) 16.93 (0.53) 12.61 (0.38) 11.66 (0.35)
Trilobe 0.1 70.44 (2.16) 88.29 (2.80) 72.51 (2.27) 70.08 (2.19) 58.85 (1.89)
0.2 43.66 (1.44) 67.87 (2.18) 47.72 (1.50) 37.55 (1.17) 37.16 (1.17)
0.3 30.33 (0.94) 46.82 (1.49) 27.57 (0.85) 21.70 (0.67) 23.73 (0.77)
0.4 19.43 (0.60) 33.81 (1.04) 17.39 (0.54) 12.60 (0.40) 15.41 (0.46)
LOC CC location control chart, OOR CC out of roundness control chart, REG CC regression-based
approach control chart, PCA CC principal-component-analysis-based control chart, ART NN adaptive
resonance theory neural network
Table 11.3 Phase II simulation results for the production scenario with the fixed-effect model.
Actual ARLs and corresponding standard deviations in parentheses (1,000 trials)
Bilobe 0.1 72.11 (2.13) 58.40 (1.86) 6.75 (0.19) 45.89 (1.40) 49.60 (1.56)
0.2 37.70 (1.21) 21.72 (0.69) 1.20 (0.02) 7.02 (0.20) 19.87 (0.63)
0.3 21.47 (0.63) 8.87 (0.27) 1.00 (0.00) 1.71 (0.04) 8.30 (0.23)
0.4 10.63 (0.32) 4.22 (0.12) 1.00 (0.00) 1.05 (0.01) 3.81 (0.10)
Trilobe 0.1 86.85 (2.72) 82.43 (2.71) 16.70 (0.50) 64.30 (1.94) 59.63 (1.84)
0.2 60.26 (1.92) 38.99 (1.24) 2.44 (0.06) 20.18 (0.57) 31.84 (1.01)
0.3 33.59 (1.04) 20.37 (0.67) 1.11 (0.01) 5.20 (0.15) 16.86 (0.51)
0.4 21.56 (0.66) 10.42 (0.32) 1.01 (0.00) 1.87 (0.04) 9.82 (0.30)
11 Quality Monitoring of Geometric Tolerances: a Comparison Study 325
control profiles, given that all the approaches are designed to achieve the same
false-alarm probability of approximately 1% . To this aim, we also evaluated the
performance of the competing approaches in the case in which no out-of-control
condition was present in phase II, just to check the correctness of tuning for each
control chart in each simulated scenario.
Tables 11.2 and 11.3 summarize the simulation results for the two production
scenarios under study. In particular, Table 11.2 refers to the case of the random-
effect model, while Table 11.3 refers to the case of the fixed-effect model. Each
table reports the ARLs estimated by computing a set of 1,000 run lengths, given
new profiles simulated according to a specific out-of-control model.
Since in actual industrial applications the analyst is not expected to know a pri-
ori what kind of out-of-control condition will affect the production process and
how severe it will be, we consider a measure of the overall performance for each
of the five competing approaches in each production scenario. To do this, we con-
sider the mean ARL values for each competing approach in signaling a generic
out-of-control condition of any severity for that production scenario. We assume
that all out-of-control conditions previously considered are equally probable and
that the analyst knows the model for the monitored functional data (this is plausi-
ble when a retrospective phase of control charting has been accomplished).
Figures 11.1 and 11.2 graphically depict the 95% Bonferroni confidence inter-
vals of the overall ARLs presented by the three competing approaches in each
production scenario considered in our study. A discussion on the performance
comparisons is given in the following two subsections for the production scenarios
with the random-effect model and the fixed-effect model, respectively.
From the results reported in Table 11.2, graphically summarized in Figure 11.9, it
can be observed that the control chart of the geometric error has the lowest power
of detection when compared with the competing methods.
Similarly, the regression-based approach has a small power of detection in sig-
naling out-of-control conditions influencing one or more harmonics. This may be
mainly ascribed to the variability in the regression parameters which characterizes
this production scenario with random effects for the baseline model. It should be
noted that for the out-of-control conditions considered in our study (half-
frequency, bilobe, and trilobe), the majority of alarms released by the regression-
based approach are produced by the multivariate control chart of the vector of
fitted parameters. The extra variability in the regression parameters, which natu-
rally characterizes the vector of fitted parameters in this scenario, causes a lower
detection power of the regression-based control charts, in particular of the multi-
variate control chart.
It can be noted that the PCA-based approach, in many cases, outperforms the
regression-based approach. With reference to scenarios of random effects, the
326 B.M. Colosimo and M. Pacella
PCA-based approach consists of a multivariate control chart based on the first four
retained principal components and of a univariate control chart for monitoring the
residuals.
Furthermore, from Table 11.2 it can also be noted that the location control chart
has performance comparable to that observed for the PCA-based approach. In
a few cases, especially in the case of the half-frequency out-of-control conditions,
surprisingly the location control chart outperforms the PCA-based approach. This
result shows that the simple location control chart can be considered a valuable
alternative to parametric methods for profile monitoring, at least in a production
scenario with random effects.
Nevertheless, in the case of random effects, the comparison study shows that
the neural network approach should always be preferred to signal almost all the
out-of-control conditions, with the only exception of half-frequency where the
performance is equal to that of the location control chart. This result shows that
the neural network is able to model the manufacturing signature and to signal
correctly an out-of-control condition even if extra variability naturally character-
izes the parameters of the model.
70
63.5783
60
60.1653
53.5274
50
50.576
41.53 42.2532
40
39.2862 39.8588 36.4839
34.4523
30
20
LOC CC OOR CC REG CC PCA CC ART NN
Figure 11.9 The 95% Bonferroni confidence intervals of the overall average run length (ARL)
for the competing approaches for the production scenario with the random-effect model. LOC
CC location control chart, OOR CC out of roundness control chart, REG CC regression-based
approach control chart, PCA CC principal-component-analysis-based control chart, ART NN
adaptive resonance theory neural network
11 Quality Monitoring of Geometric Tolerances: a Comparison Study 327
By comparing the results summarized in Figure 11.10 with those in Figure 11.9,
one can observe that, with reference to the fixed-effect model as a reference pro-
duction scenario, i.e., a process stabler and/or more controlled than in the previous
case, each of the competing approaches has better performance in signaling any
kind of out-of-control condition. This can be simply explained by observing that
no extra variability naturally affects the in-control profiles.
Among the competing approaches, the regression-based method improves dra-
matically its performance in signaling any kind of out-of-control condition. In-
deed, the regression-based approach outperforms all of the other competing meth-
ods, even if in a few cases, especially when a high severity of the out-of-control
condition is considered, the regression-based and the PCA-based approaches may
have comparable performance.
Note that for the out-of-control conditions considered in our study (half-
frequency, bilobe, and trilobe), the majority of alarms released by the regression-
based approach are produced by the multivariate control chart of the vector of
fitted parameters, while the PCA-based approach consists of a univariate control
50
40
34.2203
30 31.9326 28.0649
23.7589
26.0435
20 22.179
13.7696
10 12.3547
3.11367
2.801
0
LOC CC OOR CC REG CC PCA CC ART NN
Figure 11.10 The 95% Bonferroni confidence intervals of the overall ARL for the competing
approaches for the production scenario with the fixed-effect model
328 B.M. Colosimo and M. Pacella
chart only (a Q control chart). When PCA is performed in the case of a fixed-
effect model, no significant principal components are identified (Colosimo and
Pacella 2007) as the PCA is performed after data centering and this first step con-
sists in subtracting the mean pattern (described by the fixed-effect model) from
each profile datum. PCA is thus performed just on the error terms and hence no
significant principal component is correctly reported. Also, note that in this case
the Q statistic is given by the sum of the squared difference between data ob-
served at each location and the average profile at that location.
From the results in Table 11.3, it is also worth noting that the neural network
has a good performance in signaling the out-of-control conditions, even if its per-
formance in this scenario is not even close to comparable with the performances of
the parametric approaches (regression-based and PCA-based). However, while the
neural network approach has a better performance than the location control chart
in signaling shifts in the second and third harmonics, the location control chart
outperforms both the neural network and the OOR control chart in the case of the
half-frequency out-of-control condition. For this specific out-of-control condition,
the OOR control chart and the neural network have a similar performance.
Nevertheless, from Figure 11.10 it can be noted that the location control chart
has the lowest overall power of detection when compared with the competing
methods. As expected, also the OOR control chart does not have better perform-
ance than the model-based control charts (regression-based and PCA-based) and
the neural network.
know a priori what kind of out-of-control condition will affect the production
process and how severe it will be.
Figure 11.11 graphically depicts the 95% Bonferroni’s confidence intervals of
the overall ARLs of the five competing approaches in signaling a generic out-of-
control condition of any severity for any type of production scenario.
It can be observed that the regression-based and the PCA-based approaches
should be preferred as these approaches have better performance. It is also worth
noting that the neural network also has an overall performance substantially simi-
lar to the overall performances of the model-based approaches, even though it is
slightly worse than the overall performances of the regression-based and PCA-
based approaches. Therefore, it is fair to conclude that the neural network ap-
proach can be considered a valuable option for profile monitoring.
As expected, the location control chart does not have overall performance bet-
ter than the overall performances which characterize the regression-based ap-
proach, the PCA-based approach, and the neural network. Nevertheless, the loca-
tion control chart outperforms the OOR control chart and, hence, when a method
of very low complexity is needed for an actual application, the location control
chart should always be preferred to the control chart of the geometric error.
50
45.496
45
43.43
40
37.5456
35 35.9389
29.8702
30 28.3513
27.7922
28.5668
26.6577 26.3259
25
20
LOC CC OOR CC REG CC PCA CC ART NN
Figure 11.11 The 95% Bonferroni confidence intervals of the overall ARL for the competing
approaches (both random effects and fixed effects as equally probable production scenarios)
330 B.M. Colosimo and M. Pacella
11.6 Conclusions
This chapter presented a comparison study between different approaches for pro-
file monitoring. Both analytical models and nonanalytical methods presented in
Chapters 8–10 were considered.
With the former kind of approach, the in-control shape of the profiles is sum-
marized by a parametric model, while profile monitoring is based on monitoring
the parameters of this model. The control charts are based on the estimated pa-
rameters of the model from successive profile data observed over time (regression-
based control charts and PCA-based control charts were considered in this work).
On the other hand, with the latter kind of approach, one can monitor one or
more measures that reflect the discrepancies between observed profiles and
a baseline profile established using historical data. In this chapter, both a location
control chart and a neural network were considered as representative of non-
parametric approaches for profile monitoring. With use of using a location con-
trol chart, a control region is implemented around a mean curve, where the bor-
ders of this control region are computed by the common Shewhart approach. In
the case of a neural network, the baseline profile model is automatically estab-
lished by means of a computer algorithm. This method allows computers to
automatically learn from historical data the relationship to represent the profiles
in space. The data set, from which the relationship is learned by the neural net-
work, consists of in-control profiles only. After training, the neural network
produces a signal when an input profile does not fit, according to a specific crite-
rion, to the learned prototype.
Finally, a common control chart of the estimated geometric errors was also in-
cluded in the comparison study, as this kind of approach is still the most represen-
tative of industrial practice.
By comparing the overall performance in phase II, which refers to the ability to
signal a generic out-of-control condition of any severity for any type of production
scenario (with a random-effect and a fixed-effect model), we can conclude that the
extra effort required by the regression-based and PCA-based approaches is
worthwhile. In fact, both the regression-based and the PCA-based approaches are
more effective in signaling a generic change in the functional data. However,
while the PCA-based approach shows superior robustness to a change of the pro-
ductive scenario, the performance of the regression-based approach may be re-
duced dramatically in signaling out-of-control profiles in the case of a production
scenario with a random-effect model.
Furthermore, though the neural network may be less effective for specific out-
of-control conditions than the analytical/parametric approaches, the performances
observed by using such a method are either comparable to or superior to those
produced by the regression-based and PCA-based approaches in several cases.
Therefore, it is fair to conclude that the neural network approach can be consid-
ered a valuable option for profile monitoring, especially when a model of the pro-
files is not available.
11 Quality Monitoring of Geometric Tolerances: a Comparison Study 331
As expected, the individuals control chart of the geometric errors is not suitable
for profile monitoring and hence this approach is not recommended for actual
applications. In fact, when a method of very low complexity is needed for actual
applications, either the location control chart or the neural network should always
be preferred.
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
Cho NW, Tu JF (2001) Roundness modeling of machined parts for tolerance analysis. Precis Eng
25:35–47
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
Ramsay JO, Silverman BW (2005) Functional data analysis. 2nd edn. Springer, New York
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
Index
333
334 Index
T Z