Computer-Aided Tolerancing

Computer-aided
Tolerancing
The 4th CIRP Design Seminar on Computer-aided Tolerancing
The University of Tokyo, Sanjo Conference Hall, Tokyo, Japan

April 5-6, 1995
Sponsored by
CIRP (International Institution for Production Engineering Research)
JSPE (Japan Society for Precision Engineers)
ASME (American Society for Mechanical Engineers)
International Program Committee
A Clement (Dassault System, France)

P.Bourdet (ENS de Cachan, France)
H.A.EIMaraghy (University of Windsor, Canada)
I.Ham (Pennsylvania State University, USA)
F.Kimura (The University of Tokyo, Japan, Chairman)
V.Srinivasan (IBM, USA)
R.Weill (Technion, Israel)
A.Wirtz (Neu-Technikum Buchs, Switzerland)
Organizing Committee
F.Kimura (The University of Tokyo, Japan, Chairman)

R.Furutani (Tokyo Denki University, Japan)
M.lnui (lbaraki University, Japan)
S.Kanai (Tokyo Institute of Technology, Japan)
H.Suzuki (The University of Tokyo, Japan)
K.Takamasu (The University of Tokyo, Japan)
Computer-aided
Tolerancing
Proceedings of the
4th CIRP Design Seminar
The University of Tokyo,
Tokyo, Japan, April 5-6, 1995
Edited by
Fumihiko Kimura
The University of Tokyo
un ~:d~~~:'~;: :~~k
I I .Tokyo· Melbo",•. Mad,.,
Published by Chapman & Hall in association with the International
Institution for Production Engineering Research (CIRP)
Published by Chapman &; Ibll, l-6l1oundll'} Row, London SEIIIHN, U K
Chapman &. HaJJ. 2-6 Boundary Ro .... LorxIon SEI IIHN. UK

Chapman &. HaJI GmbH. Pappelallee 3. 69469 Weinheim. Germany
Chapman &. HaJl USA. lIS Filib Avenue. New Y<Kt.. NY 10003. USA
Chapman &. HaJl Japan. rIl'.Japan. Kyow. Building. 3F. 2·2·1 HirakawacOO. Chiyoda·ku.
Tokyo 1m. Japan
Chapman &. HaJI AU$!ralia, 102 Dod<b SlteeI. South M,lhoume. Victoria nos. Ausuali.
Chapman &. HaJIlndia, R. Sesbadri. 32 Second Main Road. CIT East. Madras (I(l()03S. India
First edition 1996
CI 1996 Chapman & Hall

Softcover reprint oflhe hardcover 1Sl edition 1996
ISBN·13: 978·94.(110·7183·3 c.[SBN.13: 978·94.(I()9.1529.9

DOl: 10.10071978·94-009'1529·9
Apan from any fair deaJing for the purposes of research or private study, or criticism 01'
review. as pennilted undef the UK Copyright Designs and PalcnlS Act. 1983. this publicauOlll
may no! be reproduced. slORd, or tran.mitted. in Illy form or by any means. wilhout tbe prior
penniss;on in writing efthe publisheR. or in the case ofrepropapbic reproclUCtiOlll only in
accordance with tbe wms of the Jio;ell(:C$ issued by the Copyright Ucensing Agency in tbe
UK, 01' io accordance with tbe wms of licences issued by !he lppfOpriaIC Reproduction
Rights Orpniz.ation outsiok the UK. EDquiries con«ming reproo:II.Ituon oulSiok !he tennJ
Slated here should be sem to the publishers ac \be London address printed on this page.
The publisller make. no repre$Clll8lioa. "'-prUI or implied .... ith regard to the accumcy of
tbe information conu.i1lfJd in this 0001<. and cannot accept any legal responsibility or liabi lity
for Illy errors 01' omission. that may be IIlII'k
A catalogue record for this book is available from the British Libnlry
Library of Congress Catalog Card Number: 96-8432S
SPrinted on pelTTtanent acid·free text paper, manufll:tlUed in lI:eo.-dano:e with

ANSIIN1S0 Z39.48·1992 and ANSIINISO Z39.48-1984 (Permanence of Paper).
CONTENTS
International Program Committee and Organizing Committee ii

Preface vii
PART ONE Functional tolerancing 1
1 A declarative infonnation model for functional requirements
A. Clement, A. Riviere and P. Serre 3
2 Constrained dimensioning and tolerancing assistance for mechanism
A. Desrochers and R. Maranzana 17
3 Univocal expression of functional and geometric tolerances for
Design, Manufacturing and Inspection
A. Ballu and L. Mattieu 31
4 A tolerancing tool based on kinematic analogies
O. W. Salomons, H.J. longe Poerink, F. van Slooten
F.l.A.M. van Houten and H.J.J. Kals 47
5 Modelling spatial dimensional chains for CAD/CAM applications
v.T. Portman and R.D. Weill 71
PART TWO Tolerance modelling 87
6 Toleranced feature modelling by constraint of degree of freedom for
assignment of tolerance
T. Maeda and N. Tokuoka 89
7 Quality features in CAD- and CAPP- systems
H.K. Tonshoff and M. Ehrmann 104
PART THREE Modelling geometrical error 117
8 Geometrical behaviour laws for computer-aided tolerancing
p. Bourdet and E. Ballot 119
9 Configuration space based analysis of position uncertainties of
parts in an assembly
M. Inui and M. Miura 132
10 Physically based modelling for evaluating shape variations
H. Suzuki, K. Kase, K. Kato and F. Kimura 147
PART FOUR Statistical tolerancing 157
11 Towards an ISO standard for statistical tolerancing
V. Srinivasan and M. O'Connor 159
vi Contents
12 Statistical process control using vectorial tolerancing

K. Martinsen 173
13 Probabilistic analysis of geometric tolerances
A O. Nassef and AA EIMaraghy 187
14 Taguchi rules in some Japanese standardizations oftolerancing
T. Shibayama 204
PART FIVE Tolerance system 217
15 Tolerance optimization using genetic algorithms: benchmarking
with manual analysis
M.P. Iannuzzi and E. Sandgren 219
16 Optimal tolerance synthesis by genetic algorithm under the machining
and assembling constraints
S. Kanai, M. Onozuka and H. Takahashi 235
17 Holistic approach and advanced techniques & tools for tolerance
analysis & synthesis
J.M. Parks 251
PART SIX Computational metrology 271
18 Data processing method for geometrical forms with form deviations
in Coordinate Metrology
K. Takamasu, l. Fukuda, R. Furutani, J. Hong and S. Ozono 273
19 Algorithmic circularity measurement for fringe analysis and
sub-micron position sensing
J. Pegna and T.P. Hilaire 283
20 Inspection method for geometrical tolerances using coordinate
measuring machine
F. Tanaka, P. Ikonomov, H. Okamoto and T. Kishinami 298
21 A soft gaging approach for complex cases including datum shift analysis
of geometrical tolerances
C. Fortin and J.-F. Chatelain 312
22 An evaluation of geometrical errors by segmentation with fitting form
features
K. Kase, H. Suzuki and F. Kimura 328
Author index 339
Keyword index 341
PREFACE
Theory and practice of tolerances are very important for designing and manufacturing engineering
artifacts on a rational basis. Tolerance specifies a degree of "discrepancy" between an idealized
object and its physical realization. Such discrepancy inevitably comes into our product
realization processes because of practical cost consideration or our inability to fully control
manufacturing processes.
Major product and production characteristics which are affected by tolerances are product quality
and cost. For achieving high precision machines tight tolerance specification is necessary, but
this will normally increase product cost. In order to optimally compromise the conflicting
requirements of quality and cost, it is essential to take into account of the total product life cycle
throughout product planning, design, manufacturing, maintenance and recycling. For example,
in order to construct durable products under severe working conditions, low sensitivity of
product functionality with respect to tolerances is required. In future, re-use of components or
parts will become important, and tolerance synthesis with respect to this aspect will be an
interesting future research topics.
Rational tolerance synthesis and analysis are very complicated issues, especially if the scope of
consideration is expanded into the total product life cycle. Traditionally much effort has been
spent for these issues, but new progress is now expected by the introduction of computer aided
methods in design and manufacturing. By using today's CAD/CAM systems, we can fairly
easily construct models of our target products in their nominal representation. But, current
tolerancing methods are not precise enough to represent products with their allowed variations
under specified tolerances.
We require new ideas and methods for capturing, in computer understandable manner, the
intended engineering semantics of tolerance specification normally represented on engineering
drawings. One basic requirement is to establish a method to represent shape variations under
various disturbing conditions, and to relate such variations with conventional tolerancing
practices for tolerance analysis and synthesis. The other issue may be more fundamental that
product functionality specification should be maintained and should be related with tolerance
descriptions for product realization. These are new and challenging research and development
subjects. Already remarkable technological advances have been achieved, and contemporary
CAD/CAM systems begin to provide such tolerancing capabilities to facilitate designers' tasks.
Key technology for computer aided tolerancing described above will include such topics as
functional tolerance description, formal modelling of tolerancing, modelling of geometrical errors
and variations, statistical tolerancing, computer support systems for tolerancing, and
computational metrology and its relationships with tolerancing. This book is a collection of
relevant papers which are related with the above issues. It contains basic materials relevant for
the introduction to computer aided tolerancing, and it can also serve as a basis for further
advanced researches.
The contents of this book comes from the seminar on Computer-aided Tolerancing organized by
viii Preface
CIRP (International Institution for Production Engineering Research), JSPE (Japan Society for
Precision Engineers) and ASME (American Society for Mechanical Engineers), Tokyo 1995.
This seminar is a successor of the former three seminars held in USA 1989, Israel1991, and
France 1993. The seminar is intended to consolidate and advance our understanding of
tolerancing as distinct technology, and as important tools in the design and manufacturing of
quality products. For expanding the scope of tolerancing to cover the whole product life cycle,
it is necessary to have deep models and comprehensive methods for tolerancing, such as
mathematical models and computational algorithms for tolerancing. Research areas for
metrology, tolerance standards and computer aided tolerancing, mutually interact each other. It
is also an objective of this seminar to provide a forum for researchers of those different areas to
establish a proper way of future collaboration.
We expect that this book will be a good introduction for newcomers to computer aided
tolerancing, as well as CAD/CAM users who like to utilize tolerancing capability in CAD/CAM
systems, and, at the same time, will give a good starting point for advanced research activity.
Finally we would like to express our sincere appreciation to all the contributors and the members
of the International Program Committee and the Organizing Committee for their substantial effort
to make this book available in the present form.
Furnihiko Kimura
PART ONE
Functional Tolerancing
1
A Declarative Information Model for
Functional Requirements
A. Clement
Dassault Systemes,
9 quai Marcel Dassault, BP 310,92156 SURESNES Cedex, FRANCE
A. Riviere
Laboratoire d'Ingenierie Integree des Systemes Industriels (L.I.1.S.1,)
ojISMCM"' ,
3 rue Fernand Hainaut, 93407 SAINT-OUEN Cedex, FRANCE.
P. Serre
Laboratoire d'Ingenierie Integree des Systemes Industriels (L.1.IS.1.)
ojISMCM*,
3 rue Fernand Hainaut, 93407 SAINT-OUEN Cedex, FRANCE.
Phone: 33-1-49452922. Fax: 33-1-49452929.
Abstract
The tolerancing synthesis consists in proposing a part tolerancing diagram directly obtained
from the mechanism good-working conditions. However, tools that can be used to describe the
functional requirements are a few and often very specific. The objective of this paper is then to
propose a mean to define, from a univocal and sufficiently adaptable point of view, all the
functional requirement types a design engineer is brought to specify.
A declarative information model based on TTRSI will be explained using the EXPRESS
language.
Three examples will be examined and described with the EXPRESS-I language:
• Parallelism condition between two gearwheels;
• Waterproofness condition obtained by a sealing ring;
• Assemblability condition for an automatic machine assembly.
Keywords
Functional requirement, mechanism specification, TTRS, EXPRESS, tolerancing synthesis
'Institut Superieur des Materiaux et de la Construction Mecanique.

'Technologically and Topologically Related Surface.
4 Part One Functional tolerancing
1 INTRODUCTION
Design is a very iterative process. The design engineer must be able to use several tools in
order to simulate the solution's behaviour, in its « real» environment. In fact, these tools
facilitate the choice of standard components, in order to dimension them, to study the part's
behaviour under load, ... However, aid tools linked to the tolerancing of mechanical parts are
rare.
Realistic part modelling allows to consider, as soon as possible, the manufactured parts
inaccuracies. Such a model offers the main advantage to provide a solution ensuring at the
same time, the respect of the functional requirements as well as an economical manufacturing.
At the beginning of the project, the design engineer identifies a technical function group that
should be satisfied by the mechanism being defined. For each function, imposed optimal work
conditions are expressed in terms of functional requirements. For example: a two-surface gap,
the distance between the axes of two parts of revolution, a minimum distance, ...
Nowadays, there are no standards to specify these functional requirements, neither drawing
or data exchange standards. Most of the standard component manufacturers suggest a
tolerancing diagram for each part implicated in the technical function rather than the direct
expression of the functional requirement. For tolerancing synthesis, it appears important to
structure this particular domain. Therefore, the objective of this study is to develop a
declarative information model for functional requirements, based upon the TTRS theory
(Clement, 1994) and using the EXPRESS language (ISOIDIS 10303-11). This paper presents
this declarative model.
2 DECLARATIVE INFORMATION MODEL

The functional requirement list is extremely long and varies a lot. It is tempting to associate
systematically one or many functional requirements to every known technical function.
However, the designer should be able to state clearly his need all the time, even if a particular
case has never been identified.
Currently, the recent STEP standard allows the exchange of high level information between
heterogeneous CAD systems.
An information model concerning the mechanical part dimensioning and tolerancing based
on TTRS model (Charles, 1989) is presented. Dimensioning and tolerancing of mechanical
parts can be achieved by the structure of the TTRS theory. The principal concepts of this
theory can be summarised as follows (Clement, 1993) (Gaunet, 1993) (Riviere, 1993):
• The TTRS entity is the association of few surfaces of a one part;
• The MRGE 2 entity is the positioning element of the TTRS. It is made of a point andlor a
line andlor a plane;
• There are 7 classes ofTTRS;
• There are 28 cases of dimensioning;
• There are 44 cases of tolerancing.
A new entity called Pseudo-TTRS similar to the TTRS entity is used when the' surfaces
implicated in the association are from different parts. Description of the dimensioning and the
tolerancing of the Pseudo-TTRS will be used to define functional requirements.
The representation of the information model is achieved by using the EXPRESS-G language
for the graphic form and EXPRESS for the literal one. The examples will be expressed in the
literal EXPRESS-I form (Schenck, 1994).
2Minimum Reference Geometric Element.

A declarative information model 5
,
,r-~----------------
: : Pseudo_TIRS_component
~_l___ _ ___ _
component 8[1 :?1
i schema_DRS.DRS
, ,
l ______________ J
component 5[1 :?] ,
form_tolerance:
c~us~~::_I_I_
1--------
,
I
id I I id
identifier I
,L _ _ _ _ _ _ _ _ _ _I '
Mechanism
f----------------~
, '
Pseudo_ITRSyrismatic
value
dimensionJ)arameter
Figure 1 EXPRESS-G diagram of Pseudo_TTRS_schema.
The EXPRESS-G graphic form of the information model is shown in Figure 1. This structure
is based on two main entities: Pseudo_ TTRS and Pseudo_ TTRS joint; the former is supertype
of the latter. This diagram shows the relations between entities, for instance, the
Pseudo_ TTRS joint entity supports the dimensioning and the tolerancing between two
geometric elements, whereas the Pseudo _TTRS entity supports the common structure to each
one of the seven itemised surface classes. It also supports intrinsic constraints, for instance,
form tolerances.
The literal EXPRESS form of this model is given in (Clement, 1995).
3 EXAMPLES
3.1 Gearwheel functional requirement
Presentation and analysis of the standard
The "parallelism error between axis", illustrated in Figure 2, is defined in the French standard
(NF E 23-006) as follows:
It is the set of two errors which has a effect on the gearwheel axis parallelism in the range
of the tooth width. The two identified errors are the following:
• the axis deviation error;
• the axis incline error.
A tolerance is associated to each one of these errors:
• the axis deviation tolerance;
• the axis incline tolerance.
Their standardised values are given in function of the tooth width and of the chosen
precision class. The axis incline tolerance is the double of the axis deviation tolerance.
Figure 2 Parallelism error between axis from (NF E 23-006).
Description in accordance with the proposed model

These functional requirements are described with the EXPRESS-I language. Used notations
are as follows:
• SI the gearwheel geometric support surface and MRGEl, the SI's MRGE, composed of
the axis of the 1st gearwheel (referred to as MRGEl_DR) and with a plane containing this
axis (referred to as MRGEl]L);
• S2 the gearwheel geometric support surface and MRGE2, the S2's MRGE, composed of
the axis of the 2nd gearwheel (referred to as MRGE2_DR) and with a plane containing this
axis (referred to as MRGE2]L).
This functional requirement description needs the construction of two geometric support
TTRS, and of one Pseudo_ TTRS joint. The Figure 3 presents the complete Pseudo_ TTRS
structure.
TTRS 1 Pseudo TTRS joint

T..Yll..e: prismatic TYlLe: prismatic
Comgonent: Comgonent:
# I ISurface: S I # 1 TTRS_l
MRGE:I MRGEI]L #2 TTRS 2
MRGEI DR -~ Constraint :
implicated elements:
TTRS 2 ~ # 1 r-nWEI _DR
J
"\::me: prismatic ~ #2 MRGE2_DR
Component: tolerancing type:
# 1 ISUrface: S2 parallelism (I, t')
MRGE:IMRGE2]L Derived MRGE:
MRGE2 DR - #1 Plane PI
#2 Plane P2
Figure 3 Pseudo_TTRS structure of gearwheel functional requirement.
It is important to notice that the two planes PI and P2 are used to specify the orientation of the
parallelepiped tolerance zone of section t by t'. This zone and the two reference planes are
illustrated in Figure 4.
Figure 4 Parallelepiped tolerance zone for MRGE1_DR from (Chevalier, 1986).

The following paragraphs present the construction of three types of entities: i) geometric
support TTRS: TTRS_I and TTRS_2; ii) Pseudo_TTRS supertype; iii) Pseudo_TTRSjoint
related to the parallelism tolerance.
Construction of geometric support TTRS

For each geometric element involved in the functional requirement, one TTRS should be
constructed.
• TTRS 1 construction
sup geo 1 Geometric_support { surface ~ « Sl »;

MRGE ~ « MRGEl »; );
ttrs 1 TTRS{ id ~ « TTRS 1 »;

component ~ {TTRS_component{@sup_geo_l));
form tolerance ~ Nil;
SUPOF(@ttrs_prism_l); );
SUBOF(@ttrs_l); );
• TTRS _2 construction
sup_geo_2 Geometric_support { surface ~ « S2 »;

MRGE ~ « MRGE2 »; );

component ~ {TTRS component{@sup_geo_2));
SUPOF(@ttrs-prism_2); );
TTRS_prismatic{ SUBOF(@ttrs_2); );
Construction of the Pseudo_ TTRS supertype

The description of the deviation and incline tolerances is obtained by the supertype
Pseudo_TTRS of the Pseudo_TTRSjoint entity.
• Pseudo TTRS construction
Pseudo_TTRS{
id ~ « PS TTRS J »;
component ~ { TTRS component{@ttrs 1),
TTRs=component{@ttrs=2));
form tolerance ~ {)
SUPOF( @ps_ttrs-prism, @ps_ttrs_jo); );
ps_ttrs_prism pseudo_TTRS-prismatic{ SUBOF(@ps_ttrs); );

Construction of Pseudo_ TTRS.Joint related to the parallelism tolerance
• Pseudo_ TTRS joint construction

Pseudo TTRS_joint{ SUBOF(@ps_ttrs);
el_implicated_l ~ « MRGEl DR »;
el_implicated_2 ~ « MRGE2_DR »;
dimensioning ~ I};
geometric tolerance ~ parallelism(t,t'); };
The description of this gearwheel functional requirement depends on the tooth width. The
design engineer should introduce this datum to the system or it can be extracted from the
geometric CAD model.
Note: As a substitute for the tooth width, the engaging width would certainly be a better
datum concerning the axis parallelism.
3.2 Waterproofness functional requirement

Presentation and analysis of the standard
The French standard (NF E 48-371) specifies the characteristics and the tolerances of shafts
and bores, as well as the specifications ensuring the waterproofness of this subassembly. An
example of sealing ring waterproofness dimensioning is illustrated in Figure 5.
Two surfaces participate in the waterproofness of the joint: the bore with the diameter D in
which the sealing ring is positioned and the shaft with the diameter d 1.
We will especially examine and represent the "total run-out" specification of the bore with
respect to the axis of the shaft (referred to as A) and the "coaxiality" specification of the axis of
the bore with respect to the axis of the shaft (referred to as A) .
., A
:r
o
lSi
Figure 5 Sealing Ring Waterproofness from (NF E 48-371).


These functional requirements are described with the EXPRESS-I language. Used notations
are as follows:
• F 1: the bore geometric support surface and MRGE 1, the F 1's MRGE, composed of the
bore axis (referred to as MRGEl_DR);
• F2: the shaft geometric support surface and MRGE2, the F2's MRGE, composed of the
shaft axis (referred to as MRGE2_DR).
TTRS, and of two Pseudo_ TTRS joint. The complete Pseudo TTRS structure of this
example is shown in Figure 6.
Pseudo TTRS joint I TTRS I Pseudo TTRS joint 2

Component: Class: cylindrical Component:
#1 TTRS_ I Component: # 1 TTRS_ I
#2 TTRS 2 1( r - # IISUrface: FI #2 TTRS 2
Constraints: MRGE: MRGEI DR -~ Constraints:
implied element: implied element:
+---""
vl---+U2
#1 FI TTRS 2 i'- # I MRGEI DR
" MRGE2_0R+i\ CI=, "";""",,, MRGE'=OR
toleraneing type: Component: tolerancing type:
total run-out (t I) # II Surface: F2 coaxiality (0t2)
......... MRGE: MRGE2 DR-
Figure 6 Pseudo_TTRS structure of the waterproofuess functional requirement.
The following paragraphs present the construction of 4 types of entities: i) geometric support
TTRS: TTRS_l and TTRS_2; ii) Pseudo_TTRS supertype; iii) Pseudo_TTRSjoint_l related
to the "total run-out" tolerance; iv) Pseudo_TTRSjoint_2 related to the "coaxiality"
tolerance.
Construction ofgeometric support TTRS

One TTRS is constructed for each geometric element involved in the functional requirement.
• TTRS_1 construction
Geometric_support surface ~ « Fl »;
MRGE ~ « MRGEI »; };

component ~ {TTRS_component{@sup geo I}};
SUPOF(@ttrs_cyl 1); };
TTRS_cylindrical{ SUBOF(@ttrs 1); };

sup_geo_2 Geometric_support { surface ~ « F2 »;
MRGE ~ « MRGE2 »; };

component ~ {TTRS_component{@sup_geo_2}};
SUPOF(@ttrs_cyl_2); };
ttrs cyl 2 TTRS cylindrical{ SUBOF(@ttrs 2); };

The Pseudo_ TTRS supertype that describes the association of the two geometric supports,
TTRS_I and TTRS_2, is constructed.
ps_ttrs_1 = Pseudo_TTRS{
id ~ « PS TTRS 1 »;
component ~ {TTRS component{@ttrs I},
TTRS=component{@ttrs=2}};
form tolerance ~ {}
SUPOF( @ps ttrs cyl,
@ps-ttrs-jo 1,
@ps=ttrs-jo_2); };
ps ttrs cyl = Pseudo TTRS cylindrical { SUBOF(@ps ttrs 1); };
Construction of Pseudo_TTRS.Joint_l related to the " Total Run-Out" tolerance

The Pseudo_TTRSJoint that describes the "total run-out" tolerance is constructed. It
corresponds to a geometric tolerance with datum.
• Pseudo TTRSJoint 1 construction
Pseudo_TTRS_Jolnt{ SUBOF(@ps_ttrs 1);
el_implicated_1 ~ « F1 »; . . . . . . (surface)
el_implicated_2 ~ « MRGE2_DR »; (MRGE)
dimensioning ~ I);
geometric tolerance ~ « total run out (t1) »;
); -
Construction of Pseudo_ TTRS.Joint_2 related to the "Coaxiality" tolerance

The Pseudo_TTRSJoint that describes the "coaxiality" tolerance is constructed. It
corresponds to a geometric tolerance with datum.
• Pseudo_TTRSJoint_2 construction
Pseudo_TTRS joint{ SUBOF(@ps_ttrs 1);
el implicated 1 ~ « MRGE1 DR »; (MRGE)
el_implicated_2 ~ « MRGE2 DR »; .. (MRGE)
dimensioning ~ I};
geometric tolerance ~ « coaxiality (0t2) »; };
The description of this waterproofuess functional requirement is rather easy. It is realised in a

very similar way to the tolerancing of mechanical part.
It should be mentioned that the representation of functional requirements by means of
mechanical part dimensioning and tolerancing has already been used by standard associations.
The studied "total run-out" tolerance is expressed between two elements belonging to the
different parts.
3.3 Assemblability functional requirement

Presentation and analysis
The assembly operation of part #3 with a subassembly composed of parts #1 and #2 is depicted
in Figure 7. The location of this subassembly on the part fixture of the assembly machine is
ensured by the planar and the cylindrical surfaces of the part # 1.
It should be possible to assembly any part #3 with any subassembly, part #1 and #2, with
the assembly machine. This condition cannot be represented by the existing standard tools. In
consequence, we propose to describe it with the model presented in this paper.
Figure 7 Assemblability functional requirement from (Nevins, 1989).

A declarative infonnation model 13

The functional requirement is described with the EXPRESS-I language. Used notations are as
follows:
• C 1: the bore geometric support surface and MRGE 1, the C l' s MRGE, composed of the
bore axis (referred to as MRGEI_DR);
• C2: the shaft geometric support surface and MRGE2, the C2's MRGE, composed of the
shaft axis (referred to as MRGE2_DR).
TTRS and of one Pseudo_TTRSjoint. The complete Pseudo_TTRS structure of this example
is given in Figure 8.
ITRS 1 Pseudo TIRS joint

.clill: cylindrical Comoonent'
Component: # 1 TTR I
# 1 Surface: CI #2 TTRS- 2
LJ->
MRGE : MRGEI
OR-~ Constraints:
implicated element :
TTRS 2 __ # 1 MRGEI _DR
Class: cylindrical #2 MRGE2 DR
Component : tolerancing type:
# 1 Surface: C2 coaxiality (0 t) eM)
MRGE: MRGE2 DR
Figure 8 Pseudo_ TTRS structure of assemblability functional requirement.
The following paragraphs present the construction of three types of entities: i) geometric
support TTRS: TTRS_I and TTRS_2; ii) Pseudo_TTRS supertype; iii) Pseudo_TTRSjoint
related to the "coaxiality" tolerance.
Construction ofgeometric support TTRS

One TTRS for each geometric element involved in the functional requirement is constructed.
sup_geo_1 Geometric_support { surface -+ « C1 »;

MRGE -+« MRGE1 »; };
ttrs 1 TTRS{ id -+ « TTRS_1 »;

component -+ {TTRS_component{@sup_geo_1}};
form tolerance -+ Nil;
supor (@ttrs cyl 1); };
surface ~ « C2 »;
MRGE ~ « MRGE2 »; };
ttrs 2 TTRS{ id ~ « TTRS_2 »;

component ~ {TTRS_component{@sup_geo_2}};
SUPOF(@ttrs_cyl_2); };
The Pseudo_ TTRS used to describe the minimum gap that can be found between C 1 and C2
surfaces is constructed. It is a functional requirement between two surfaces and this tolerance
type is not proposed in the standard (Chevalier, 1986).

The Pseudo_ TTRS supertype that describes the association of the two geometric supports,
TTRS _1 and TTRS_ 2, is constructed.
Pseudo_TTRS{
id ~ « PS_TTRS »;
component ~ {TTRS component{@ttrs I},
TTRS=component{@ttrs=2}};
form tolerance ~ {}
SUPOF( @ps ttrs cyl,
@ps=ttrs::::jo); };
Pseudo_TTRS cylindrical{ SUBOF(@ps_ttrs); };
Construction of Pseudo_TTRSJoint related to the "coaxiality" tolerance
• Pseudo_ TTRS joint construction

ps_ttrs_jo= Pseudo_TTRS_joint{ SUBOF(@ps_ttrs);
el_implicated_1 ~ « MGREI »;
el_implicated_2 ~ « MGRE2 »;
dimensioning ~ {I;
geometric_tolerance ~« coaxiality (t)~»; };
The link between the dimensioning of each surface and its relative position is obtained for the
tolerancing, by the maximum material principle.
4 CONCLUSION
The tolerancing specification of mechanical parts has been for a long time the principal concern
of the design engineer. However, this specification is nothing more than the consequence of the
respect of technical functions that should be ensured by the mechanism.
The model described in this paper offers the possibility to enter directly into the world of the
mechanism specification.
With the TTRS and Pseudo_ TTRS entities which are both based on the same theory, two
assisting tools for tolerancing can be provided: tolerancing analysis ensures that the mechanism
specifications are satisfied by the part specifications, on the other hand, the part specifications
are obtained from the mechanism specifications through tolerancing synthesis.
The examples presented in this paper concern design and assembly. Future work will
validate the information model with other examples and will complete its structure by
integrating pertinent entities of PART 47 (ISO/CD 10303-47).
The functional requirements for manufacturing will be studied with the proposed model, in
particular:
• the « minimal-cutting» functional requirement, corresponding to the association of two
surfaces belonging to the same part but in two different states (raw and manufactured
states);
• the « adjustment of the machine» functional requirement, corresponding to the association
of the part-fixture positioning TTRS and the tool active surface;
• the « positioning dispersion» functional requirement, corresponding to the association of
the part-fixture positioning TTRS and the part positioning TTRS for a given manufacturing
activity.
5 REFERENCES
Charles, B. Clement, A. Desrocher, A. Pelissou, P. Riviere, A. (1989) « Toward a computer aided
functional tolerancing model », International Conference on CAD/CAM and AMT in Israel,
CIRP Sessions on Tolerancing for Function in CAD/CAM Environment. Proceedings, vol.
2, Jerusalem, Israel, December 11-14, 1989.
Chevalier, A. Lecrinier, 1. (1986) « Cotation et Tolerancement - Dictionnaire », AFNOR
TECHNIQUE,1986.
Clement, A. Riviere, A. Serre, P. (1995) « A Declarative Information Model for Functional
Requirements », 4th CIRP Seminar on Computer Aided Tolerancing, University of Tokyo,
Japan, April 5-6, 1995.
Clement, A. Riviere, A. Temmerman, M. (1994) « Cotation tridimensionnelle des systemes mecaniques
- Theorie et pratique», PYC Edition, 1994.
Clement, A. Riviere, A. (1993) « Tolerancing versus Nominal Modeling in Next Generation CAD/CAM
System», 3rd CIRP Seminar on Computer Aided Tolerancing, Ecole Normale Superieure de
Cachan, France, April 27-28, 1993.
Gaunet, D. (1993) « Vectorial Tolerancing Model», 3rd CIRP Seminar on Computer Aided
Tolerancing, Ecole Normale Superieure de Cachan, France, April 27-28, 1993.
Nevins, J.L. Whitney, D.E. (1989) « Concurrent Design of Product and Processus», McGraw Hill,
New York 1989.
Riviere, A. (1993) « La geometrie du groupe des deplacements appliquee it la modelisation du
tolerancement», These de doctorat, Ecole Centrale Paris, November 1993.
Schenck, D. Wilson, P. (1994) « Information Modeling: the EXPRESS Way», OXFORD
UNIVERSITY PRESS, 1994.
ISO/CD 10303-47 (1994) « Product Data Representation and Exchange - Part 47: Integrated Generic
Resources: Tolerances », ISO, 1994.
ISO/DIS 10303-11 (1992) « Industrial Automation Systems - Product Data Representation and
Exchange - Part II: The EXPRESS Language Reference Manual », ISO, 1992.
NF E 48-371 (1985) « Transmissions hydrauliques - Etancheite par bagues it levres pour arbres
toumants - Dimensions et caracteristiques du logement et de l'arbre », AFNOR, May 1985.
NF E 23-006 (1967) « Precision des engrenages paral1eles it denture it developpante », AFNOR, July
1967.
2
Constrained dimensioning and tolerancing
assistance for mechanisms
A. Desrochers and R. Maranzana

Ecole de Technologie Superieure
Department ofAutomated Production Engineering
4750, avenue Henri-Julien, Montreal, Canada, H2T 2C8
Tel.: 514-289-8930, Fax.: 514-289-9042
e-mail: alain@gpa.etsmtl.ca
Abstract
This paper presents a systematic approach for dimensioning and tolerancing mechanisms. The
method is based on a model of Technologically and Topologically Related Surfaces (TTRS) and
uses a binding graph representation of the mechanism. The complete dimensioning and tolerancing
of each part is generated by recursively associating surfaces along kinematic loops on the graph.
Technological requirements and user constrains are integrated in the system through specific
choices of kinematic loops and partial references (Minimum Geometric Datum Element). The
system developped insures that the geometric dimensioning and tolerancing be complete and non
redundant while conforming to ANSI YI4.SM or ISO standards
Keywords
Geometric dimensioning and tolerancing, Computer Aided Design (CAD), Technologically and
Topologically Related Surfaces (TTRS), mechanism, assembly.
1 INTRODUCTION
Contemporary CAD/CAM systems provide an effective mean for generating part geometry.
However they lack in providing the designer with appropriate tools for expressing functional
requirements through proper dimensioning and tolerancing of the parts (Clement, 1993).
In the system we are proposing, tolerance assistance is a process following which tolerances
are inferred from the functional requirements of the part. These, in tum, are most often linked to
the topology of the mechanism into which the part will be operating. Practically speaking,
functional requirements translate in terms of clearances and fit within the mechanism.
In any case, however, a paramount functional requirement for all mechanism is assembly.
Indeed, parts must first assemble before any other requirement can be met. Therefore, a sensible
dimensioning and tolerancing strategy must guarantee assembly above all while handling other
functional requirements as additional constrains to be dealt with (Clement, 1991). As with any
constrained system, conflict and redundancy must be detected and avoided.
In view of this, we are proposing a dimensioning and tolerancing scheme, well suited for
CAD/CAM system implementation, and based on a surface association model using a constrained
graphical representation of the mechanism.
2 THE TTRS REPRESENTATION MODEL
2.1 Mathematical basis
The work presented in this paper is built upon a model called TTRS for Technologically and
Topologically Related Surfaces, which has been the topic of several previous papers (Desrochers,
1994) and thesis (Desrochers, 1991) and more recently a book (Clement, 1994). Basically, the
model is based on successive associations of elementary surfaces.
More specifically, in this model, surfaces are classified according to their degrees of invariance
(translations and rotations under which the geometry remains unchanged). Following this
definition, seven distinct classes of surfaces where identified in table I.
Table 1 Classes of surfaces and TTRS along with their MGDE
Surface or Degrees of MGDE Physical element

TTRSclass invariance
Spherical [S] 3 rotations A point Sphere center
Plane [PI] 2 translations A plane Plane parallel or coincident with

1 rotation the actual surface
Cylindrical [C] 1 translation Aline Cylinder axis
1 rotation
Helical [H] 1 helical A line and Screw axis plus point belonging to
displacement a point it
Of revolution 1 rotation Aline and Axis of revolution plus point

[R] a point belonging to it
Prismatic [PrJ 1 translation Aline and Line in the direction of translation

a plane plus plane defining orientation
General [G] none A line, a plane Any combination of surface

and a point element
Constrained dimensioning and tolerancing assistance 19
2.2 Recursive association process
The association process leading to the generation of TTRS is a recursive one; elementary surfaces
are first associated together to form TTRS objects or entities which are themselves associated
together or to remaining surfaces. Normally, at the end of the association procedure, all functional
surfaces and all TTRS should be related in a hierarchical fashion to form an association tree or
graph as in figure 1.
Furthermore, along the association process, TTRS will be classified in one of the seven classes
of invariance according to table 2. As this table indicates it, there are 44 cases of binary
associations between elementary surfaces and/or TTRS (Riviere, 1993). For each of these
association cases, there will be a corresponding dimensioning and tolerancing scheme describing
the relative position of the two surfaces/TTRS.
The association process can be generated either manually on isolated parts or it can also be
generated using the graph theory applied to mechanisms as it will be shown later on.
Table 2 TTRS recursive classification and tolerancing instances
CLASS Spherical Plane Cylinder Helical Revolute Prismatic General

[S} [PI} [C] [H} [R} [Pr} [G}
[S] C [S]~[S] [S]U [Pl]~[R] [S] C [C]~[R] [S]U [H]~[G] [S] C [R]~[R] [S]U[Pr]~[G] [S]U [G]~[G]
Spherical
[S] U [S]~[R] [S]U [C]~[G] [S]U [R]~[G]
[S]
[Pl]P[Pl]~[Pl] [Pl]T [C]~[R] [Pl]U[H]~[G] [Pl]T[RHR] [Pl]p[Pr]~[Pr] [Pl]U[G]~[G]
Plane
[Pl]u[Pl]~[Pr] [Pl]P[C]~[Pr] [Pl]U[R]~[G] [Pl]u[Pr]~[G]
[PI] [Pl]U[C]~[G]
[C]C [C]~[C] [C]C [H]~[H] [C]C [R]~[R] [C]P[Pr]~[Pr] [C]u [G]~[G]

Cylinder
[C]P [C]~[Pr] [C]u [H]~[G] [C]u [R]~[G] [C]U[Pr]~[G]
[C] [C]u [C]~[G]
Helical [H]C[H]~[H] [H]U [R]~[G] [H]U[Pr]~[G] [H]U[G]~[G]

[H1UH1~[Gl
[H]
Revolute [R]C [R]~[R] [R]U[Pr]~[G] [R]U[GHG]
[R]U [R]~[G]
[R]
Prismatic [Pr]p[Pr]=[Pr] [Pr]U[G]~[G]
[Pr]u[Pr]~[G]
[PrJ
[G]U[G]~[G]
General
[G]
Remark: C = Concentric or Coaxial, P = Parallel, T = perpendicular, U = Other

2.3 The concept of Minimum Geometric Datum Element (MGDE)
Along with this association process, we have also devised a method to generate partial or
complete datum systems representing the various surfaces associations. We call these partial or
complete reference frames "Minimum Geometric Datum Elements" or MGDE; they describe,
visually, the degrees of invariance associated to a surface or TTRS.
Practically, three basic MGDE will be considered and combined; the point, the line and the
plane (Desrochers, 1991). Their combination will enable a user to represent the degrees of
invariance of a surface or TTRS as well as the precedence of the datum associated to a part.
Indeed, a primary datum will be represented by a plane, a secondary datum by a line and a tertiary
datum by a point (Desrochers, 1994).
2.4 Dimensioning and tolerancing TTRS
As prescribed by the ANSI and ISO standards, each elementary surface has its own form
tolerances and, in some cases, its own dimension (diameter). This is summarized in table 3.
Additionally, in our model, every surface or TTRS association has its own dimensions and
tolerances conforming to both ANSI and ISO standards. Therefore, for each of the 44 association
cases described in table 2, there is a corresponding dimensioning and tolerancing. These involve
mainly linear or angular dimensions as well as location, orientation and runout tolerances.
Table 3 Dimensioning and tolerancing of elementary surfaces
Surfaces Dimension Tolerances Symbols

Spherical Diameter Circularity 0
Planar N.A. Flatness C7
Straightness
Cylindrical Diameter Circularity 0
Cylindricity ~
Straightness (axis)
Helical Diameter ref ANSI Y14.6 N.A.
Pitch or thread
Of revolution Diameter Circularity 0
Angle (for a cone) Straightness (axis)
Prismatic
Radius (for a torus)
N.A.
Profile of a line
Profile of a surface
Straightness (axis)
"
C
Profile of a line ~
Profile of a surface 0
General N.A. Profile of a line f'.
Profile of a surface 0
2.5 Application to mechanism modeling
Mechanisms can be represented schematically by a binding graph. In such a graph, all the parts
and all the surfaces participating in contacts are represented. These surfaces are said to be
functional because they participate in the assembly and, more often than not, their dimensions are
crucial to insure proper mating and functioning of the mechanism. Contacts between functional
surfaces are represented as lines or binds on the graph.
Once parts, surfaces and contacts are represented, it becomes possible to extract numerous
closed loops or cycles on the graph. These cycles are called kinematic loops and they are, in many
ways, the three-dimensional equivalent of the familiar tolerancing chains. The association process,
or the generation of TTRS, will then be made according to a given number of selected
independant kinematic loops:
Independant loops = Number of contacts - Number of parts involved +
f
S 1: cylinder
MGDE: axis
I
Diameter
Form tolerance 1
1
TTRS 1: prismatic Distance
MGDEI: axis Sl (or S2) Location
I + plane S I1S2
S2: cylinder Diameter Parallelism
..
r MGDE: axis Form tolerance
~ S3: plane ~IForm tolerance

MGDE: plane
I TTRS2: general
MGDE2: MGDEI
+ point S I1S3
Location
tolerance
Figure 1 TTRS and MGDE creation example through the use of a binding graph.
3 DIMENSIONING AND TOLERANCING MECHANISMS
3.1 Opening level of the model
The simple example presented in figure 1 allows us to foresee two areas of relative uncertainty in
the modeling procedure involving the graph theory. Indeed, the result of the association process
will heavily depend on the selection of independant kinematic loops and the choice of relevant
MGDE.
In the case depicted in figure 1, two loops had to be selected out of a possible three (the third
going through S 1 and S3 in this instance). A different choice would have affected the association
sequence which could have, in more complex cases, changed the outcoming dimensioning and
tolerancing of the parts.
Again in figure 1, it appears that the MGDE selected for TTRS 1 could have either been the axis
of cylinder S 1 or S2 along with the plane passing through both axis. The choice of one axis or the
other would obviously have modified the generation of every subsequent MGDE. This, in tum,
would have been influencial in determining to which axis a perpendicularity tolerance would have
been applied in this particular case.
Therefore, as we can see, the binding graph modeling procedure offers a flexibility that can
advantageously be exploited to introduce technological requirements into the association process.
These technological requirements will translate as loop or MGDE selection criteria to the system.
3.2 Global association strategy
The surface association strategy will be based, firstly, on a proper selection of kinematic loops on
the binding graph and, secondly, on an appropriate choice ofMGDE.
Hence, a set of technological criteria will be used to place all potential loops on the graph in
a hierarchical fashion within a list. Once this classification will have been performed, the first n
independant loops will be retained and dealt with in the order in which they appeared in the list.
Similarly, user defined constrains such as datum surfaces or basic dimensions will dictate the
choices ofMGDE along the association procedure. This will be done by placing a priority on
every surface or MGDE chosen as datum or involved in the specification of a basic dimension.
Schematically, the complete association strategy can best be illustrated on a sequential diagram
as in figure 2.
3.3 Loop selection criteria
As depicted in figure 2, technological criteria are being used to place all potential loops of the
binding graph in a hierarchical manner. In this respect, three important criteria have been
identified for that task; the relevance ofthe smallest sub-assembly, the significance oftolerancing
chains and the relative size and location of the surfaces involved. In the following paragraphs, we
shall justify each of these criteria by showing how they reflect sound dimensioning and tolerancing
practices.
Association procedure Technological requirements
I Set of all possible cycles (loops) I

I
Loop selection criteria :
* Smallest sub-assembly
* Tolerance chains (stack up)
* Surface size and location
I Hierarchical list of cycles (loops) I-
I
* User defined constrains
* .......
l
I Selection of n independant loops I
!
I Interpretation of each loop J
! ...
I TTRS generation I
I
MGDE selection criteria:
* Datum surfaces
* Surface size
* Functional dimensions
* Functional conditions
I Selection of proper MGDE 1--
r * Other user defined constrains
!
I Selection oftolerancing case J
Figure 2 Association strategy.
Smallest sub-assembly
It is reasonable to assume that mechanisms are, more often than not, designed to be assembled
through a succession of small sub-assembly. For this reason, a designer will normally specity
dimensions and tolerances suited to ensure proper mating of the parts without the need for
tolerance transfer. In our model, this can be achieved by selecting the smallest loops, that is, those
passing through the least number or parts. Our experience has also shown that this criteria will
generate the most "natural" geometric dimensioning and tolerancing strategy.
Figure 3 illustrates a simple three part mechanism in which only four loops out of a possible ten
are independant. In this particular instance, we would therefore select three loops involving two
parts; one between part 1 and part 2 and two between part 2 and part 3. The fourth loop would
have to include the three parts in order to take contact f (between surfaces S 13 and S31) into
account.
Figure 3 Selection of shortest loops for sub-assembly.
Tolerance stack-up
This criterion is also of paramount importance for many designers since it deals with tolerancing
chains. Tolerancing chains are a tool that allows the computation of gaps, clearances and
interferences within any given mechanism. In tum, these gaps and clearances express a functional
requirement in the mechanism. For example, the value of a clearance may reflect an alignment
condition or some lubrication requirement. On the binding graph, tolerancing chains will translate
as loops involving contacts sharing a common orientation as with the basic mechanism illustrated
in figure 4.
01 _....
.....
J) D3 D2
r-- -
® ®
CD
Figure 4 Tolerance stack-up involving planar contacts.
However, contacts are not limited to pairs of planes and so the notion of orientation need to
be extended to other types of surfaces. Basically, contact orientations retlect the degrees of
freedom left by the contacts and these are, in turn, related to the MGDEs involved. Following this
reasonning we can establish contact classes according to every type of basic MGDE. This is
summarized in table 4.
Table 4 Classes of contacts according to basic MGDE types
Contact type MGDE and surface types Contact direction
Spherical MGDE: Point All direction in three

dimensionnal space
Surfaces:
Surface of revolution
Helical surface
General
Cylindrical MGDE: Axis (ligne) In a plane perpendicular to

the axis of the surfaces
Surfaces:
Cylindrical surface
Prismatic surface
Planar MGDE: Plane In a direction normal to the

surfaces in contact
Surface:
Planar surface
As it can be seen on this table, surfaces featuring a point in their MGDE will mate to provide
a contact of spherical type, restricting any movement in all three directions. For cylindrical and
prismatic surfaces, contact directions will be limited to a plane normal to the axis. Finally, planar
surfaces will yield a contact along their normal vector.
In the event of a contact between surfaces of different type, the least restrictive surface would
dictate the resulting type of contact. For example, in the case of a cylinder lying on a plane, the
outcoming contact type would be planar, retlecting the relative degrees offreedom between both
surfaces. Consequently, it becomes possible to establish a type for every contact and to look for
a common direction among the contacts involved in a loop on the binding graph.
In some rare instances, this criterion may contlict with the first one, dealing with the smallest
sub-assembly. In these occasion we shall give priority to the first criterion.
Surface type and size

Dimensionning and tolerancing is, in many ways, affected by the type, the relative size and even
the relative location ofthe surfaces of a part.
Indeed, the surfaces providing the largest areas are often used as datum for dimensionning and
tolerancing a part. In many cases, the relative size of the datums is even reflected in their order
of precedence. "A" being the primary datum, is frequently chosen to be the largest of all surfaces
while "B" and "e", the secondary and tertiary datum are traditionally smaller. Additionally, these
surfaces will, more often than not, constitute the external boundaries of the part. This is a
practical feature from a machining and control set up point of view.
From an other perspective, smaller surfaces of same type, same size and same orientation are
frequently associated together with common tolerancing information, using a common set of
datum. This is also a sensible practice since they will often be produced in the same set up.
These concepts are best illustrated on the example depicted in figure 5. It can be seen that the
external plane making up the boundaries of the part are being used as datum with the largest being
the primary datum.
Moreover, the three cylindrical surfaces have the same size and the same orientation. For this
reason they have been considered as a pattern, using basic dimensions and common tolerancing
information. Within our model, they are considered as a prismatic TTRS which is, in tum, located
with respect to the datum set of the part using, again, basic dimensions.
3X 06.4+g· 2
Figure 5 Hole pattern located using a complete set of datum (reproduced from Foster, 1994)
It should also be noted that this technological criteria, along with the two previous one, can be
applied systematically and automatically without the need for operator intervention. These criteria
however do not provide any guidance concerning the appropriate choice ofMGDE for the TTRS
generated along the association process. This is the topic of the next section.
3.4 MGDE selection criteria

In certain conditions, the MGDE can be generated automatically along the association process.
Regarding this, special rules have been established according to three typical cases of association.
Rule 1
Condition: The resulting TTRS can be classified in the same class as both of its original
components (surfaces or TTRS).
Rule: The resulting MGDE is the MGDE of one of its original components or a new one.
User constrains or guidance is therefore needed to perform a choice.
Example: Two parallel planes forming a planar TTRS. The resulting MGDE will be either
chosen between the MGDE of its original components or will be the median plane of the
association.
Rule 2
Conditon: The TTRS produced is classified into the same class as one of its original components.
Rule: The new MGDE is the MGDE of the component sharing the same class.
Example: A prismatic surface parallel to a cylinder, yielding a new prismatic TTRS. The resulting
MGDE in this case, will be the MGDE of the original prismatic surface.
Rule 3
Condition: The TTRS is classified in a class of higher rank (fewer degrees of freedom) than
any of its original components.
Rule: Both original MGDE will participate in the determination of the resulting MGDE.
Example: A plane containing the axis of a cylinder and forming a prismatic TTRS. The rank
or the invariance degree of the prismatic TTRS is higher than that of the cylinder or the plane
alone; both have more degrees of freedom. Consequently, the prismatic MGDE will include
the cylinder's axis or MGDE along with the plane itself to define the prismatic TTRS
completely.
As it can be seen, these three basic cases cover all possible situations in terms of surfaces and
TTRS associations. Rule 1 does not require any user assistance for the determination of the
MGDE. However, rule 2 calls for some kind of guidance from the operator while rule 3 will
greatly depend on the type of surfaces associated. In this last instance, user input would have been
needed if the two surfaces associated had been parallel cylinders. Then the resulting prismatic
MGDE should have been selected as one the two cylinders axis along with a plane passing through
both axis.
From this, it appears necessary to establish user defined constrains that would guide the TTRS
and MGDE generation process. Moreover, these constrains should comply with current practices
in design so that they could be applied in a most natural way by any operator.
3.5 User defined constrains
If the dimensioning and tolerancing strategy is to reflect the designers intent, it is essential to
provide him with the means to express these intentions to the system. On a drawing, this is usually
done by specifying clearances or fit, datum surfaces and basic dimensions. The tolerances
themselves serve mostly as a way to achieve the given clearance or fit. In view of this, our system
will basically accept the same information a designer would put on paper. It should also be noted
that these user defined constrains will affect both the loop and MGDE selection criteria. However,
in the absence of such specifications, the association process will proceed using solely the criteria
mentionned earlier to perform the dimensioning and tolerancing of the mechanism.
Datum specification
In the case where the designer has specified several datum, a first check will be performed to
insure that every datum removes at least one degree of freedom to the part on which it has been
specified. These datum will then be associated through a loop to form the first TTRS of the part.
At the same time, the corresponding MGDE will be generated to properly represent the nature and
order of precedence of these datum. Due to the recursive nature of our association process, every
TTRS generated thereafter should be defined and located with respect to this MGDE representing
the datum set. Additionally, a search will establish in which loop the datum chosen, first appear.
The system will then reorder any loops bearing the same" priority values" so that the one carrying
a datum be placed in front. Finally, each isolated datum will be designated as MGDE when the
association process comes accross it. This will guarantee that the datum specified by the user be
used as such.
Basic dimensions
Basic dimensions may be specified by the user if he wishes to impose them for functional reasons.
In this instance, they will be treated as constrains in quite the same way the datums were. Again,
the loop priority list will be modified to favor loops using surfaces involved in basic dimensions;
loops of equal priority value will thus be set appart. Additionally, the basic dimension surfaces will
be used as MGDE when they appear in a TTRS, if they were not already designated as datums.
In this way all basic dimensions specified by the user should be generated, provided they are not
redondant or do not contradict each other. In these cases, the problems will be detected and
highlighted by the system.
Clearances and fit

These constitute the most important and relevant way of specifying a functional condition within
a mechanism. Moreover, clearances and fit always appear in a tolerancing chain which our
procedure automatically detects and uses as a technological criterion to generate the ordered list
of kinematic loops. In other words, clearances and fit are automatically being dealt with by the
system from the start. However, as an additional precaution, a loop passing through a contact
assigned as clearance or fit will be placed before another one of equal priority value (that is
another tolerancing chain of equallenght).
Other user defined constrains

As it can be infered from the preceding paragraphs, the procedure for handling constrains follows
two basic steps; the reordering of equal priority potential kinematic loops and the assignment of
MGDE to which datum and dimensions will be attached. Using this method, any other user
defined constrain could be taken into account quite easily. Furthermore, the relative importance
of all these criteria can be modified simply by changing their order of application or by removing
some.
Constrained dimensioning and toierancing assistance 29
4 CONCLUSION
The TTRS model through the use of a binding graph has proven to be a valuable tool for
dimensioning and tolerancing mechanisms. Indeed the approach is generic enough to accomodate
various technological requirements and constrains.
In this respect, the integration of manufacturing constrains is being considered as a mean to
achieve concurrent engineering. The manufacturing constrains would then be added to the design
constrains and the resulting dimensioning and tolerancing would reflect the combination. In the
case where manufacturing and design constrains would conflict, the system would be able to point
out where and a solution could most probably be found through the use of tolerance transfer.
Additionally, kinematic loops provide an effective vehicule for computing gaps or interferences
and also for toleran~e allocation. In this instance, tolerance values would be retreived from the
CAD database and stacked up, when appropriate.
Finally, as pointed out in the title, the procedure presented provides, above all, an assistance
to the designer. In many case a system alone would be faced with choices beyond its capacity or
knowledge. Therefore, the objective was to elaborate a methodology that would conform to the
operator's intent while insuring no redundancy or uncertainty in the dimensioning and tolerancing
of the mechanism. That would leave the designer with only the most significant part of the
tolerancing process.
5 REFERENCES
Clement, A., Riviere, A. and Temmerman, M. (1994) Cotati on tridimensionnelle des systemes
mecaniques: Theorie & pratique, PVC Edition, Ivry-sur-Seine.
Clement, A. and Riviere, A. (1993) Tolerancing versus nominal modeling in next generation
CAD/CAM system, Proceedings of 3rd CIRP Seminars on Computer Aided Tolerancing,
Ecole Normale Superieure de Cachan, April 27-28, 1993, p. 97-113.
Clement, A., Desrochers, A. and Riviere, A. (1991) Theory and practice of 3D tolerancing for
assembly", Proceedings of 2nd CIRP Seminars on Computer Aided Tolerancing, Penn State
University, May, 1991.
Desrochers, A. and Clement, A. (1994) A dimensioning and tolerancing Assistance Model for
CAD/CAM Systems, International Journal ofAdvanced Manufacturing Technology, 9:352-
361.
Desrochers, A. (1991) Modele conceptuel du dimensionnement et du tolerancement des

mecanismes. Representation dans les systemes de CF AD, These de doctorat, Ecole Centrale
Paris.
Dessins techniques - Tolerancement geometrique - Generalites, definitions, symboles, indications

sur les dessins: NF E 04-552 (1983), AFNDR, France.
Dimensioning and tolerancing: ANSI YI4.5M-1982 (1982), American Society of Mechanical

Engineers, New-York.
Foster, L.W. (1994) Geo-metrics IIIm: the metric application of geometric dimensioning and
tolerancing techniques, as based upon harmonization of national and international standards
practices. Addison-Wesley, Reading, Massachusetts.
Riviere, A. (1993) La geometrie du groupe des deplacements appliquee it la modelisation du

tolerancement, These de doctorat, Ecole Centrale Paris.
6 BIOGRAPHIES
Professor Alain Desrochers earned his Doctoral degree from Ecole Centrale Paris in France, his
Master in Science from the University of California, Los Angeles and his Bachelor degree in
mechanical engineering from Ecole Polytechnique de Montreal in Canada. His research interest
is in the area of CAD/CAM and more specifically in computer aided dimensioning and tolerancing.
Professor Roland Maranzana holds a Doctoral degree from Universite de Valenciennes in France
and an Engineer diploma from Ecole Nationale d'Ingenieur de Belfort. His research efforts are in
the fields of CAD/CAM with an emphasis on Computer Process Planning (CAPP) and
manufacturing.
3
Univocal expression of functional and
geometrical tolerances for design,
manufacturing and inspection
BALLU MATHIEU
LMP - Universite Bordeaux I LURPA - ENS de Cachan
351 Cours de la Liberation 61 Avenue du President Wilson,
33405 TALENCE cedex, 94235 CACHAN cedex,
FRANCE FRANCE
Tel: 3356846616 Tel: 33 1 4740 22 21
Fax: 33 56846964 Fax: 33 1 47402220
baUu@upel.lmp.u-bordeauxfr mathieu@lurpa.ens-cachanfr
Abstract
With the systematic use of computers at all levels of product life cycle, especially at the levels of
designing, manufactUling and inspecting, a consistent model of the product functional geometry
is required.
The analysis of various models on tolerancing expression, points out the lack of "company-
wide" models. Some models integer the know-how of the standards, some develop new
concepts, some are coherent but none takes all these aspects into account at once.
On the basis of a product geometrical model, a solution is proposed that integrates the
various points of view of the entreplise actors and the standards. This model is based on
geometrical operations which are applied, not only to the ideal features, defined by the
geometrical modelers, but also to the real features. The operations are themselves defined by
constraints on the fmm and relative charactetistics of the features.
With this common language, the differences between the various existing approach and their
lacks can be pointed out. Moreover, with the model genericity, new concepts of specitication to
express a design, manufacturing or inspection intent can be defined.
To illustrate the field of applications, an ordinary mechanism which is easy to perceive has
been considered.
Keywords
Model for tolerances, tolerancing language, product model, life cycle and quality issues
1 INTRODUCTION
With the systematic use of computers at all levels of product life cycle, especially at the levels of
designing, manufactUling and inspecting, a consistent model of the product functional geometry
is required.
The CAD systems available today do not fulfill this need. They allow to define a 3D ideal
geometry indeed exploitable at various moments of the elaboration of products but insufficient
because it does not integrate functional aspects. The current solution to transmit functional
geometrical information of the products remains very conventional, it is the technical drawing.
Although very faithful to the technological culture of the world of the mechanics, this solution
presents ambiguities linked to the graphic standardized language. Indeed, tolerances which
translate the functional aspect of products do not respect the principle of univocallangage and do
not allow to express all functional requirements.
In recent years, various research on tolerancing expression has been presented. Citing only
the best-known, let us consider the following works concerning:
• Functional relations between the definition surfaces of the product expressed by a
geometrical and technological linking model.
• Specifications by size developed in "vectorial tolerancing".
• Univocal standard language with the improvements of the definitions, the mathematical
models and the exchange standards.
1.1 Functional relations

The model developed by A. Clement is based upon the concept of Technologically and
Topologically Related Sur1"aces (Clement, 1991). All features in functional relation create a new
TTRS which can itself be in relation with another feature ... and so on ... The relative situation
between TTRS is described by angular and linear parameters between the Minimum Geometric
Datum Element of each feature.
The variation of the nominal geometry is put as the variation of the parameters and are
expressed by small displacement torsors and limited by constraints This modelisation is used by
the authors to implement the analysis of mechanism dimensioning and tolerancing (Gaunet,
1994).
This model of specifications is not complete as it only takes an ideal aspect of the surfaces
into account without ever considering form deviations of surfaces.
1.2 Vectorial tolerancing

With A. Wirtz's model, named "vectorial tolerancing", the geometric features of parts are
defined with vectorial geometry. Each part-feature is described by three vectors specified in a
coordinates system. These vectors represent the position, the orientation and the size of the
feature (Wirtz, 1989), (Wirtz, 1993).
The variations of the nominal geometry are put as the variations of the components of the
three vectors. This model is coherent for all the actors of the company, but it is completely
different from the standard language. It is closer to manufacturing and inspection needs than to
design ones. Indeed, the expression of a functional requirement is not direct, but has to be
decomposed according to the system of vectors.
1.3 Univocal standard language

Presently, the CEN, at the European level, is trying to harmonize the product tolerancing and
inspection standards (Bennich, 1993). But great difficulties in using the present standards can
be noted. Despite many specification possibilities, the standards are often inexplicit and not well
structured (Ballu, 1993a).
To improve the semantics of the graphical language, the Americans have tried to formalize
the geometrical tolerances by mathematical expressions (Walker, 1993). The ANSI YI4.5.1M
standard is the result of these works. This contribution is important, but limited to the graphical
expressions of the ANSI Y 14.5 standard. So, this solution lacks genericity. Moreover, the
mathematical definition of the titting of datum to real features is not complete.
Univocal expression 33
Another manner to indicate tolerances is considered in STEP, Standard for Exchange of

Product Data (Laurance, 1993). These exchange standards, ISO 10303 intend to define a single
model of product data for all the actors of the company. They deal or will deal with the various
scopes generating product data. The part 47 : "Integrated generic resources: Shape variation
tolerances" defines the resources for the dimensioning and tolerancing expressions. This
specification language integrates most of the features of the YI4.S standard with most of its
ambiguities and limits. Nevertheless, the genel1c formalism of the EXPRESS language should
eventually integrate a complete model for the product functional geometry.
1.4 Conclusion
The analysis of these various works points out the lack of a "company-wide" model. Some
models integer the know-how of the standards, some develop new concepts, some are coherent,
but none takes all these aspects into account at once.
In the following, we wish to show, by an example, the diversity of product models
according to the point of view of actors of the enterprise and the need of a common language.
Later on, the model of definition of the geometry of products that we have developed, will be
used (Ballu, 1993a).
As this model was already presented at the 3rd CIRP seminars in 1993, just the main
concepts will be reminded.
2 GEOMETRIC SPECIFICATIONS MODEL

The proposed model is based on geometrical operations which are applied, not only to the ideal
features, defined by the geometrical modelers in a CAD system. but also to the real features. The
operations are themselves defined by constraints on the form and relative characteristics of the
features.
Conditions
Geometric features
(~! (?:
~\fX~)~X _______ X
Part real surface
Figure 1 Geometric specifications.

2.1 General concepts

The generals concepts are:
• a specification is a condition on a dimension defined from geometric features,
• these geometric features are features created from the real surface of the part by different
operations.
These concepts are illustrated by the Figure 1. The set of basic geometric features for the
specifications is the set of the real surfaces of the parts. The real surface of a part is a closed
surface which modelises the partition between the part and its surroundings. The geometric
features, assigned to express the specifications, are identified from the real surface of the parts
by the application of a set of operations.
An operation identifies one or more geometric features from one or more other geometric
features previously identified.
Finally, a specification is expressed by a condition on geometric features defined by valious
operations.
A condition defines an interval of IR inside of which the value of a dimension of geometlic
features must lie.
In the following, these concepts of geometric features, operations and conditions are
developped.
2.2 Geometric features

We distinguish different types of elements, ideal features, real features and limited features.
We name ideal features, features such as :
• A plane of the nominal model of the pmi.
• A cylinder fitted to a real feature nominally cylindIical.
The ideal features can be divided into two classes: nominal features and fitted features.
So, on the one hand, the plane of the nominal model is a nominal feature as every feature used
to define the nominal model of a part. On the other hand, the cylinders fitted to real surfaces are
fitted features.
We name real features, features such as:
• A surface parcel of the real surface of a pm·t.
• A real axis.
iGeometric features
Limited features
Ideal features Real features
x x
Nominal features
x
x x x x
x x
x x x x
x x
Fitted features
x x
x
x x x x x
x x
Figure 2 Geometric features overview.

Among the ideal and real features, we can distinguish particular features such that their
"sizes" are limited. A limited feature is a geometric feature composed of a denumerable set of
points, finite length lines finite area surfaces, finite content volumes.
A segment is a limited ideal feature, a real axis is a limited real feature. These elements have the
property to fit in a cylindrical tolerance zone. A straight line (infinity) can fit in such a zone only
if the cylindrical zone and the straight line are pelfectly parallel.
To have an overview of the geometric features, the different sets are represented on Figure 2.
This figure points out that, the geomeuic features are ideal or real features, the ideal features are
nominal or fitted features, the limited features set includes ideal and real features.
2.3 Operations
The operations used to define the specifications can be divide into four classes, extraction,
union, fitting and construction.
An extraction identifies a part of a feature with the whole feature by the expression of
contraints. These constraints are expressed on extracted points with respect to ideal features
So :
• A real surface is exu'acted from a real sUli"ace by intersection with a parallelepiped.
• A real line is extracted from a real surface by intersection with a plane, a cylinder, a cone.
• A point is extracted from a real surface by intersection with a straight line.
A union identifies a geometric feature as the union of geometlic features.

So:
• A bi-plane is fonned by the union of two planes.
• A real axis is the non-denumerable union of section centres of a nominaly cylindlic surface.
• A real median face is the non-denumerable union of centre points between sets of opposite
points of the opposite surfaces.
A fitting identifies one or more fitted features to one or more limited features by the expression
of constraints and/or an objective.
So: "
• A plane is fitted to a surface such that the sum of the squares of the distances between the
surface points and the plane is minimal.
• A cylinder is fitted to a surface such that every surface point is in the cylinder, the cylinder is
perpendicular to a datum plane and the cylinder diameter is the largest.
In some cases, the constraints and/or objective are expressed with respect to one or more ideal
features. The constraints and the objective bear on the fitted features form, on their situation
with respect to the limited features and on their situation with respect to the ideal features.
A construction identifies one or more ideal features by the expression of constraints. These
constraints are expressed between ideal features.
So:
• A plane is constructed such that it includes a datum point and is perpendicular to a datum
straight line.
• A cone is constructed such that its summit is identical to a datum point, its axis is parallel to a
datum su"aight line and its angle is equal to 45°.
A construction can be considered as a particular case of fitting for which there are no limited
features and no objective in the Clitelia.
2.4 Conditions and geometric deviations

A condition has been defined as the expression of an interval in which a dimension value has
to lie. This dimension is a characteristic of the features created by operations from the real
surface of a part.
So:
• The distance between two planes must be between two values.
• The angle between two straight lines must be inferior to a given value.
• The set of distances between a real surface and a plane must be over valued.
The respect of a condition by a characteristic value of a particular part characterizes the respect of
a specification. The characteristic value limited by the condition characterizes the greater or
lesser satisfaction to the specification. The difference between the characteristic value and its
nominal value quantifies a geometric deviation of the palt.
2.5 Characteristics
Characteristics, useful to the definition of specifications, belong to three families, form
characteristics of an ideal feature, situation characteristics between ideal features, situation
characteristics between ideal and limited features.
Form characteristics of an ideal feature are specific to its type, consequently, it is

impossible to make an exhaustive list as there is no exhaustive type list of features. Form
characteristics are real, integer or binary charactelistics.
So:
• The diameter of a cylinder is a real charactel1stic.
• The number of cylinders of a set of cylinders is an integer charactedstic.
• The sense of a helicoidal smface is a binary charactedstic.
Situation characteristics between ideal features are expressed between features of

situation of these ideal features.
The relative position of ideal features is based on the position of simple features (point,
straight line, plane or helix) attached to these ideal features. Thus the position of a cylinder is
dermed by the position of its axis (a straight line). These features are called features of situation.
Consequently, the number of characteristics is reduced and they are classified in:
• Orientation characteristics.
• Position characteristics.
• Configuration charactelistics.
• Phase charactedstics.
Remark I: phase charactedstics descl1be the angular position of the helix around its axis, with
respect to other elements.
Remark 2: for a same set of olientation and position charactedstics defining the relative situation
of features, several configurations of the shape often exist, hence the necessity to define
configuration charactelistics.
Orientation characteristics are angles between two features of types plane or straight
line. The value of the angle is contained between 0 and 90 degrees.
Orientation charactelistics:
• Angle (plane, plane).
• Angle (straight line, plan) and angle (plan, straight line).
• Angle (straight line, straight line).
Thus constraints on the angle between two features can be defined, angle (plane A, axis of the
cylinder B)= 0°, angle (plane A, axis of the cone C)= 43,75°
Position characteristics are distances between the different combinations of elements

among points, straight lines and planes.
Position characteristics:
• Distance (point, point).
• Distance (point, straight line) and distance (straight line, point).
• Distance (point, plane) and distance (plane, point).
• Distance (straight line, straight line).
• Distance (straight line, plane) and distance (plane, straight line).
• Distance (plane, plane).
Remark l: the distance (straight line, straight line) is the distance according to the common
perpendicular.
Remark 2: the distance (straight line, plane) has a meaning only if the straight line and the plane
are parallel, the distance (plane, plane) has a meaning only if the two planes are parallel.
Thus constraints on the distance between two features can be defined, distance (plane A, axis
of the cylinder B)= 10, distance (axis of the cylinder B, point C)= radius of the cylinder B
A set of constraints of the type, distance (plane A, plane B)= 10, distance (plane A, plane C)=
20 is not sufficient to define a unique situation of planes A, Band C. It is necessary to add
configuration constraints. Configuration constraints are built with configuration
characteristics.
In the example above-mentioned, the solution adopted is to define two normal vectors in the
plan A, one directed from A to B, the other from A to C. The sign of the scalar product of the
two vectors is then constrained to be positive or negative according to the desire of the designer.
The model that we propose generalizes the expression of these configuration characteristics to
points and to straight lines (Ballu, 1993b).
Characteristics between ideal and limited features are functions of the set of distances
(positive or negative) from points of the limited feature to the ideal feature.
The distance from a point M to an ideal feature G is defined as the smallest distance from the
point M to a point P of the ideal feature.
Functions of distances serving to desCI1ptions of most CUlTent specifications are:
• The sum of squares of the distances.
• The greatest distance.
• The smallest distance.
In function of needs it can be envisaged to spread the model to others functions as:
• The sum of powers p of the distances.
• The sum of distances.
2.6 Summary
The model is mainly based on operation and characteristic concepts. In the following we are
going to observe all their importance. These concepts are developed to obtain a small set of
operations and charactel1stics to descl1be the quasi totality of CUlTent specifications (standardized
or not). This model allows to communicate geometrical informations which can come from
design, manufacturing or inspection. With the consideration of the real features and the
genericity brought forth by the operations and the characteristics, the model is a "company-
wide" model. It is a univocal language, common to design, manufactUl1ng and inspection.
The presented model for the product geometry integers:
• The descl1ption of the nominal geometry.
• The design specifications to express the Val10US functions of the pan.
• The manufacturing specifications to express the Vat10US manufacturing processes.
• The inspection specifications to express the various inspection processes.
Each point of view can be expressed with this one language. With this common language, the
differences between the various existing approaches and their lacks can be pointed out.
Moreover, with the model genericity, new concepts of specification to express design,
manufacturing or inspection intent can be defined.
3 MECHANISM AND PART STUDIED

To illustrate problems of specifications expression, we have voluntarily considered a simple
mechanism. This mechanism is composed of a frame, two shafts and a spur-gear. See Figure 3.
~~-
~/r--T---'
~ ~ A-A I
I
~
( ~
-+-
I
l~
I
I
I I I
)
I
~ I
I
I
I
I
I 0
I I en
-T-
I
80
Figure 3 Mechanism
The considered functional condition is relative to the gearing of wheels.
4 SPECIFICATION FOR DESIGN

For design, this functional condition is equivalent to a requirement on the distance between the
two shafts in the gearing zone. Its respect depends obviously on deviations of each parts
intervening in the kinematic chain and especially on deviations of the two holes constituting the
shaft-beatings. Consider the nominal geometry of the frame below.
For different points of view, design, manufacturing and inspection, we give expressions of
specifications concerning the two holes E and F. First, we are going to express the functional
requirement with our specification model. Then always in our formalism, we will translate
standat·dized specifications.
A-A
o
""
o
en
o
""
100
Figure 4 Nominal geomeu)'
4.1 Expression of functional need

The solution proposed to transfer the functional need on the frame consists in expressing
variations of the distance between axes due to deviations of holes. For that, we consider the
assembly of the frame with two shafts of perfect form to the minimum material (diameter
30mm). In the zone of gearing, we give the limits of position that two points can take on the
axis of a shaft in relation to a definite ideal position by the second shaft. The expression of
deviations in shaft zone and not in gearing zone has been retained to simplify the translation of
need.
PI. ... n
PT F60n PTF100n
Figure 5 Functional requirement, definite features
For different points of view, design, manufacturing and inspection, we give expressions of
specifications concerning the two holes E and F. First, we are going to express the functional
requirement with our specification model. Then always in our formalism, we will translate
standardized specifications.
The plane used to define the position of the two points is obtained by a fitting operation.
Fitting plane PLAlsq
Constraints: none
Objective to minimize:
sum of the squares of the distances(SA,PLAlsq)
Deviations are obtained by a fitting operation to real surfaces of a constituted ideal geometry of
two cylinders, a plane, a straight line and two points.
Fitting 2 cylinders CY En et CY Fn' 2 poinl~ PTF60n et PTFloon' one plane
PLMn and one straight line SLMn
Constraints:
diameter( CYEn) = 30
diameter(CYFn) = 30
CY En inside SE
CYFn inside SF
distance (PTF60n' PL Alsq) = 60
distance (PTF60n, axis(CYFn» = 0
distance (PTFloon, PLAls,!) = 100
distance (PTFloon, axis(CYFn)) = 0
angle (PLMn, axis(CY En)} = 0°
distance (PL Mn , axis(CY En» = 0
distance (PLMn, PTF60n) = distance (PLMn' PTFloon)
angle (SLMn, axis(CYEn» = 0°
distance (SLMn , axis(CYEn» = 90
angle (PLMn' SLMn) = 0°
distance (PL Mn , SLMn) = 0
Objective to maximize:
max {distance (PTF60n, SLMn),distance (PTFIOOn, SLMn)}
Condition:
max {distance (PTF60n, SLMn),distance (PTFIOOn, SLMn)} ,,; 0.14
Cun-ently, the standardized language does not allow to express this functional requirement under
this form. Paragraphs 4.2 is going to illustrate possibilities of this language.
4.2 Classical standardized tolerance

A first solution consists in expressing the functional requirement transferred on the frame by a
specification of location. It is a simple tolerance zone specification with a dimensional
specification on the diameter of holes, without requirement of the minimum matetial condition or
the maximum matelial condition or projected tolerance.
Dimensional specifications on diameters, are defined by local sizes. The expression of these
dimensions is given by the distance between two opposite points responding to the definition
proposed by projects of standardization of the CEN and ISO. The specification is given by
limits of vmiation of these distances 30 and 30.02.
A·A +0,02
030+0
004 E
Figure 6 Classical standardized tolerance
Fitting cylinder CY EIsq Fitting cylinder CY Plsq

Constraints : none Constraints: none
Objective to minimize: Objective to minimize:
sum of the squares of the distances(SE, CYEIsq) surn of the squares of the distances(Sp, CY Plsq)
Construction straight lines SL Ei Construction straight lines SLpi

Constraints : Constraints :
angle(axis(CYEIsq), SLEi) = 90° angle(axis(CY Flsq), SLFi) = 90°
distance (axis(CYEIsq), SLEi) = 0 distance (axis(CY Flsq), SLFi) = 0
Extraction points PTEi I et PT Ei2 Extraction points PTFi I et PTFi2

PTEd = (SE n SLEi) PTPil = (SF n SLpi)
PTEi2 = (SE n SLEi) PTFi2 = (SE n SLFi)
Condition: Condition:
30:::; distance(PTEiJ,PTEi2):::; 30.02 30:::; distance(PTFil,PTFiZl:::; 30.02
The specification of location has the real axis LF of the surface SF for toleranced feature. This
axis is defined according to projects of standardization of the CEN and ISO.
Construction planes PLpi Extraction lines LF,

Constraints: Constraints:
angle(axis (CY Flsq), PLFi) = 90° LFi = (SF n PLFi)
Fitting circles Clpi Union giving line Lp

Constraints : United features:
CIFi belong to PLFi centers of (Clpi)
sum of the squares of the distances(LFi, ClF,)
Figure 7 Classical standardized tolerance, definite features
The specification of location is expressed by two fitting operations, one for the datum and the
other to define the condition:
Fitting cylinder CY Eda' Fitting straight line SLF
CYEd., inside SE angle (axis (CY Eda,),SLF) = 0°
Objective to maximize : distance (axis (CY Eda,),SLF) = 90
diameter (CY Eda,) Objective to minimize :
the largest distance (LF,SLF)
Condition:
the largest distance (LF,SLF) :5 0.03
5 SPECIFICAnON FOR MANUF ACTURING
For manufacturing, the tolerancing expression directly depends on the process. Let us consider
the realization of the two holes in machining. The features for positioning the part on the
machine tool are B, A and C. The deviations due to manufacturing are limited by a specification
of location.
Figure 8 Manufacturing specification.

Pl.8doI
.--_ _-~-..,J P\.--
'1 0 ..0 1
~e·~ri--------~~==~
Figure 9 Manufacturing specification, definite features.
New real surfaces SB and Sc have to be extracted

Extraction surface Sa Extraction surface Sc
Constraints : Constraints:
Sa corresponds to B Sc corresponds to C
The positioning surfaces are obtained by three titting operations.
Fitting plane PLBdat Fitting plane PLAdat
the smallest distance (Sa.PLBdat) ~ 0 angle (PLBdat,PLAdaU = 90°
Objective to minimize: the smallest distance (SA,PLAdat) ~ a
tile largest distance (Sa,PLBdat) Objective to minimize:
tile largest distance (SA,PLAdaU
Fitting plane PLCdat

Constraints :
angle (PLBdat,PLCdaU = 90°
angle (PLAdat,PLCdau = 90°
the smallest distance (Sc,PLCdat) ~ 0
the largest distance (SC,PLCdat)
The specification of location is translated by a construction of two straight lines SL Eman and
SL pman in ideal position. The deviations are expressed between real axes and straight lines. The
real axis LE of the surface SE is defined according to projects of standardization of the CEN and
ISO.
Construction straigilt lines SLEman et SLpman Condition:
Constraints : max {the largest distance (LE, SLEman),
angle (SLEman, PLBdaJ = 0° the largest distance (Lp, SLPman)) S 0.01
distance (SLEman, PLBdat) = 30
angle (SLEman, PLadat) = 0°
distance (SLEman, PLCdat) = 30
angle (SLpman, PLBdat) = 0°
distance (SLpman, PLBdat) = 30
angle (SLpman, PLadat ) = 0°
distance (SLpman, PLCdat) = 120
6 SPECIFICATION FOR INSPECTION

For inspection, the interpretation of the functional condition is entirely dependent on its
metrology-plant. Especially in coordinate metrology, this interpretation derives from the
software solutions. Current geometrical functions are:
• Fitting of simple features according to the criterion of the least squares.
• Construction of points, straight lines, planes and coordinate systems.
• Evaluation of angles and distances.
According to users, different processes are elaborated. We propose a process for the classical
standardized tolerance defined in paragraph 4.2.
6.1 Inspection of the classical standardized tolerance

With the plan PL AIsq defined in paragraph 4.1 and the cylinders CY EIsq, CY Fis defined in
paragraph 4.2, two points are constructed on the hole axis in the planes PL AIsq and PLoIsq-
-'L,
Figure 10 Inspection of the classical standardized tolerance, definite fea tures.
Extraction surface So Fitting plane PLOIsq

Constraints : Constraints: none
So corresponds to D Objective to minimize:
sum of the squares of the disk1nces(So,PLOIsq)
Construction point PTFO Construction point PTFA

distance (PTFD. PLDlsq) = 0 distance (PTFA. PLAIs'l) =0
distance (PTFO. axis(CYFIsq» = 0 distance (PTFA, axis(CYFIsq» = 0
The specification of location is translated by a consu'uction of one straight line constrained to be
parallel and distant of90mm to the CYElsa axis and fitting the points PTFD and PTFA.
Construction point PTM Construction plane PLM
Constraints : Constraints:
distance (PTM, axiS(CYFIsq» = 0 angle (PLM, axis(CY EIsq)) =0°
distance (PTM, PTFA) = distance (PTM, PTFO) distance (PLM, axis(CYEIsq» = 0
distance (PLM. PTM) = 0
Construction straight line SLM Condition:

Constraints : max {distance (SLM, PTFD), distance (SLM, PTPA)}
angle (SLM, axis(CYElsq» = 0° $ 0.03
distance (SLM, axis(CYElsq» = 90
angle (SLM, PLM) = 0°
distance (SLM, PLM) = 0
7 CONCLUSIONS AND FUTURE

With the application, we have pointed out the fact that several points of view can be attached to a
same functional condition: functional, standardized, inspection and manufacturing. These
different points of view involve a geometrical specification for each.
For a good communication, the existence of a common language to all actors of the entreprise
is crucial. This unique language has to be sufficiently powerful to translate precisely expression
of each.
The proposed model allows to express fully each of these aspects. Preoccupations of the
different actors can be taken in account by this unique model based on geometry, and this is
possible with the generative power of the operations and the charactelistics defined in the model.
In each of the previous cases, the specification defined by operations can be parametrized
with the constraints. As the constraints bear on characteristics of the features, the
parametrization is simply made through the constraints. Thus, in the expression of the functional
condition the distance from the beginning of the functional zone to the pm1 can be changed from
60 mm to an other value.
Once this model is defined, the different actors can communicate, but this is not the only
interest. Indeed for a good functioning, transfers of specifications between the different points
of view have to be made. The problem of specification transfers is their influence on tolerances,
also these tranfers have to be minimized. Considering this model, we have already eliminated
the transfer to the coordinate metrology since we have elaborated algorithms of calculation
allowing the solution of all used operations (Ballu, 1991), (Ballu, 1993b). Points to solve are
about the expressions of all the specifications with the standards and about the setting of
machine tool with poweltull methods (Mathieu, 1991).
8 REFERENCES:
Ballu, A. - Bourdet, P. and Mathieu, L. (1991) The processing of measured points in
coordinates metrology in agreement with the definition of standardized specifications, Annals
of the CIRP, Vol. 40/1/1991, pp 491-494.
Ballu, A and Mathieu, L. (l993a) Analysis of dimensionnal and geometlical specifications:
standars and model, Proc. of 3rd CIRP Seminars on Computer Aided Toierancing, Cachan,
France, April 27-28, 1993.
Ballu, A(1993b) Identification de modeles geometliques composes pour la specification et la
mesure par cOOl'donnees des caracteristiques fonctionnelles des pieces mecaniques,1hese de
Nancy I, Janvier 1993.
Bennich, P. (1993) Chain of standards: a new concept in tolerancing and engineeling drawing
GPS-standards, Geometrical Product Specification Standards, Proc. of the 1993 Inti Forum
on Dimensional Tolerancing and Metrology, Dem'bom, Michigan, June 1993 17-19, CRTD -
Vol. 27, pp. 269-278, 1993.
Clement, A - Desrochers, A and Riviere, A (1991) Theory and practice of 3-D tolerancing for
assembly, Proc. {)f2nd CIRP Seminars on Dimensioning and Toierancing, Pennstate,
5/1991.
Gaunet, D. (1994) Modele fOimel de tolerancement de position. ContIibution a l'aide au

tolerancement des mecanismes en CFAO. These de [,Ecole Normale Superieure de Cachan,
FevIier 1994
Laurance, N. (1993) A high level view of STEP: A formal specification of the information
content of a product design, Proc. of the 1993 Inti Forum on Dimensional Tolerancing and
Metrology, Dearborn, Michigan, June 1993 17-19, CRTD - Vol. 27, pp. 31-37, 1993.
Mathieu, L. and Weill, R. (1991) A model for machine tool setting as a fonction of positionning
errors, Proc. of 2nd CIRP Seminars on Dimensioning and Tolerancing, Pennstate, 5/1991
Walker, R.K. and Srinivasan, V. (1993) Creation and evolution of the AS ME Y14.5.1M
standard, Proc. of the 1993 Inti Forum on Dimensional Tolerancing and Metrology,
Dearborn, Michigan, June 1993 17-19, CRTD - Vol. 27, pp. 19-30, 1993.
Wirtz, A. (1989) Vectorial tolerancing, Proceedings of The Int. Con! on CAD/CAM and AMT,
Jerusalem, 1989.
Wirtz, A. (1993) From unambiguously defined geometry to the pelfect quality control loop,
Annals of the CIRP, Vol. 42fI, pp. 615-618,1993.
9. BIOGRAPHY
Ballu, A. prepared a PhD at the Laboratoire Universitaire de Recherche en Production
Automatisee of the Ecole Normale Superieure de Cachan, he is an Assistant Professor in the
Laboratoire de Mecanique Physique at Bordeaux University. He has special interest in
Mechanical EngineeIing and Metrology.
Mathieu, L. is an Assistant Professor in the Depaltment of Mechanical EngineeIing of the

Conservatoire National des Arts et Metiers at Paris. He is researcher in the Laboratoire
Universitaire de Recherche en Production Automatisee of the Ecole Normale Superieure de
Cachan. He is a corresponding member of the CIRP and an expert of ISOfTC 3-1O-57fJHG,
CENITC290. His research interests include computer-aided tolerancing and metrology.
4
A Tolerancing Tool Based on Kinematic
Analogies
O. W. Salomons, H.I. longe Poerink, F. van Slooten, F.I.A.M. van

Houten, H.l.l. Kals
University of Twente, Department of Mechanical Engineering, Labo-
ratory of Production and Design Engineering, P.o. Box 217, 7500
AE Enschede, The Netherlands, phone: X-31-53-892532,fax:
X-31-53-335612, e-mail: o.w.salomons@wb.utwente.nl
Abstract
A computer aided tolerancing tool is presented that assists the designer in functional tolerance
specification. The theoretical concepts for subsequent tolerance analysis are also provided. The
computer aided tolerancing tool is part of a feature based object oriented (re )-design support sys-
tem, called FROOM. FROOM's assembly modelling capabilities provide basic information for
functional tolerance specification. Assembly constraints are satisfied by means of degrees of
freedom (DOF) analysis. This method is based on the use of kinematic analogies. The rotations
and translations (macro-DOF's) that components are allowed to have, are inferred using this
technique. The tolerance representation in FROOM is based on the TTRS method, by Clement
et aI., which is also based on kinematic analogies. In this method, the small displacements that
are allowed in the tolerance zone can be described by a tolerance torsor or transformation matrix.
Using the tolerance torsor or transformation matrix, tolerances are described as constraints. The
small displacements that are still allowed by means of the torsor are referred to as micro-DOF's.
For tolerance analysis, the torsor approach offers a mathematically correct description of toler-
ance zones, although a lot of equations are generated. The number of parameters of these equa-
tions can often be reduced by means of a method which applies a kind of degrees of freedom anal-
ysis considering both the macro-DOF's and the micro DOF's (tolerances).
Keywords
Tolerance representation, tolerance specification, tolerance analysis
1 INTRODUCTION
An introduction into some drawbacks of existing tolerancing standards is provided in section 1.1.
The FROOM system, whose tolerancing module is the focus of the remainder of this paper, is
introduced in section 1.2. Section 1.3 provides an overview of the remainder of the paper.
1.1 Some drawbacks of existing tolerancing standards
There are some problems related to present tolerancing standards. For instance, the standards do
not provide for a theoretic tolerancing model that is consistent with today's 3D geometry models
(usually solid models). This is due to the fact that when tolerancing standards emerged, no 3D
solid models existed and a lot of freedom was left to the human interpretation of the specified
tolerances. Therefore, tolerancing standards are drawing oriented, giving rise to ambiguities
when having to be processed by computer. Another drawback of today's tolerancing standards
is that they do not provide for a method for tolerance specification. Preferably, tolerance specifi-
cation should be performed using a functional point of view in the first place: functional toleranc-
ing.
Functional tolerancing will prevent redundant or unnecessary tolerances to be specified and will
therefore support the design of manufacturable products at reasonable cost. The reason that the
tolerancing standards do not provide for a method for functional tolerancing is mainly due to the
standards being part oriented and not assembly oriented. It is the assembly where functioning as-
pects are (often implicitly) available and not in the individual parts. Therefore, current toleranc-
ing standards can be compared to a language with a large and rich syntax but with not enough
semantics and methods, at least to perform computer aided tolerancing. In Ballu (1993), Ballu
and Mathieu (1993), Koplewicz (1993) and Srinivasan (1993) more detailed overviews are pro-
vided regarding the insufficiencies in current tolerancing standards as well as current efforts to
overcome these.
1.2FROOM
FROOM is a prototype of a re-design support system, currently under development. The devel-
opment ofFROOM has been motivated by the following facts. First, for re-design tasks no proper
computer support tools are available yet. Second, CAD tools that integrate well with automated
process planning (CAPP) systems are only scarcely available. If available, they are most often
not suitable in practice, especially in situations where design and manufacturing are performed
in different companies. The CAPP system that is of particular interest to us, is the PART system
(Houten 1991), which has been developed in the authors' laboratory. FROOM is an acronym for
Feature and Relation based Object Oriented Modelling.
FROOM is a feature based system, allowing the modelling of both components and assemblies.
Features in the FROOM context can be design form features, manufacturing form features and
even abstract features. For a review of feature-based design, refer to (Salomons et al. 1993a).
FROOM employs conceptual graphs for knowledge representation. The conceptual graphs allow
re-design support to be provided and facilitate the link with CAPP (Salomons et al. 1994a). The
latter subject will not be addressed in this paper, however. Figure 1 shows Froom's system archi-
tecture.
Figure I shows that two different kinds of users are distinguished: end-users and system manag-
er users. End-users are the designers who design assemblies, components etc.; they perform the
A tolerancing tool based on kinematic analogies 49
Figure 1 System architecture of FROOM; IOD means interactive object definition.
actual design tasks. System managers customize the system to a certain application domain, com-
pany and user (group). They define the features to be used, the catalogues from which selections
can be made etc .. For these two different user groups, separate user interfaces are available. In
the modelling module ofFROOM, components, assemblies and constraints can be modelled by
the end-user. The tolerancing tool, which is the focus of this paper, is part of the constraint model-
ling module of FROOM. The modelling module has access to the kernel modeller, which offers
the basic geometry processing functionality. In FROOM, the commercially available ACISTM
kernel modeller is used for this (Spatial 1994). The application module includes the possible map-
pings to applications and the applications themselves. An example of an application might be
manufacturability evaluation. For more details on FROOM, refer to (Salomons et a1. 1993b,c,
1994a,b, 1995).
1.3 Overview of the paper
An overview of previous work in computer aided tolerancing is provided in section 2. Based on

this, the overall design of the FROOM tolerancing module is outlined in section 3. In section 4
the FROOM assembly modelling module is briefly examined as an important precondition for
automatic tolerance specification (section 6). Tolerance representation in FROOM is addressed
in section 5. Tolerance analysis is described in more detail in section 7. This section provides ex-
tensions to theories as developed by Riviere (1994) and Gaunet (1994). Apart from the overall
synthesis as provided in this paper, these extensions are the main contributions of this paper. Fi-
nally, conclusions and recommendations are provided in section 8.
2 REVIEW OF PREVIOUS RESEARCH
Research in computer aided tolerancing is subdivided in tolerance representation (2.1), tolerance

specification (2.2), tolerance analysis (2.3) and tolerance synthesis (2.4).
2.1 Tolerance representation
A lot of (possibly theoretically and mathematically correct) approaches for the computer repre-
sentation of tolerances as proposed in literature do not sufficiently comply with the international
standards. For human designers, these approaches will not be sufficient to replace the tolerancing
standards as they are not expressive enough; their "syntax and semantics" are too limited. More-
over, it will be hard to introduce a completely new and different tolerancing standard on a world
wide basis even if it has enough expressiveness to humans as well as provide for a mathematically
correct model for use in computer systems. Therefore, it seems more appropriate to look for a
theoretical model of tolerancing, or tolerance representation scheme, that is in accordance with
the international standards.
An early example of previous work in order to come to a mathematically correct model fOf toler-
ancing is the solids offset approach by Requicha (1983), extended in Requicha (1984), and Requi-
cha and Chan (1986). In this approach, nominal surfaces are given a pair of offset surfaces to de-
termine the tolerance zones. This approach differs from the tolerancing standards (Farmer and
Gladman 1986). Because the individual pairs of offset surfaces are combined to obtain a compos-
ite tolerance zone of the entire solid, the individual tolerances cease to be independent constraints.
Other early work has been described in (Johnson 1985) and (Ranyak and Fridshal1988).
Jayaraman and Srinivasan introduced the notion of virtual boundary requirement in an attempt
to redefine the notion of tolerances from the viewpoint of functional tolerances (J ayaraman and
Srinivasan 1989), (Srinivasan and Jayaraman 1989). Jayaraman and Srinivasan note that the real
purpose of tolerances is to characterize functional requirements. Two functional constraints are
cited that are believed to account for most tolerances; the maintenance of material bulk in critical
locations and spatial relationships for assembly. These requirements can be captured as virtual
boundary requirements. Virtual Boundary Requirements can be considered as a collection of
virtual half spaces. The interpretation of datums in this approach is not in accordance with the
standards.
Turner proposed a feasibility space approach (Turner and Wozny 1988), (Turner 1993) which
however does not seem to be suitable for 3D tolerance representation because of a too high com-
plexity and doubtful usefulness in 3D.
Wirtz used a vectorial approach in which each tolerance is represented as a limit on the compo-
nents of a vector that relates a given toleranced feature to a given reference feature (Wirtz 1993).
The vectorial approach by Wirtz also seems insufficient: it is not close to the standards and is ori-
ented too much towards older dimensioning and to1erancing practices. Form tolerances, for
instance, are not accounted fOf.
Building on their earlier work in the field of computer aided inspection (e.g. Bourdet (1979)
and the work by Requicha and Jayaraman and Srinivasan and Wirtz, Clement et al. arrive at a
tolerance representation model which is compatible with the standards and which seems to be
theoretically and mathematically sound (Clement et al. 1991, 1993, 1994), (Desrochers et al.
1994). Using the theory of the set of displacements by Herve (1978), Clementet al. have proven
that there are only seven elementary face types: spherical face, planar face, cylinder face, helical
face, rotational face, prismatic face and "any" face. When these 7 face types are combined, 28
cases of combination can be found. These combinations of faces are also called TIRS: Techno-
logically and Topologically Related Surfaces. There is a finite number (2: 44) of reclassifications
of TTRS which denote the theoretical number of different tolerances (cases). On this basis, a
computer system can automatically propose tolerance types employing an assembly model (geo-
metric tolerances). Also reference (datum) elements can be determined (semi-)automatically.
These elements are called MGDE: Minimum Geometric Datum Element; see Figure 2 for the
--
elementary surface MGDE element MGDEsymhol
sphere point
•
---
plane plane 0
cylinder line
helical point and line or line and plane c::::::><::J
rotational point and line
prismatic line and plane p
any point and line and plane p
Figure 2 The MGDE associated with each elementary surface (after Clement (1991».
MGDEs related to each elementary surface as discerned by Clementet al .. The same set ofMGDE
is used for determining the datum of composed TTRS. When two elementary surfaces are com-
bined in a TTRS, the resulting TTRS can be classified into one of the seven basic classes depend-
ing on the types of surfaces involved and the geometric relations between them. See Figure 3 for
the reclassifications of TTRS. In this approach, tolerances are represented vectorially, in so-
called torsors, allowing fortolerance analysis. This has been addressed in (Gaunet 1994) and (Ri-
viere et al. 1994).
.
{Ib p}
~
{E} {Tn} {R Il} {C n } {G p } {So}
p p ....----- C><J 0
p p p p p -------
p p p
p pp p p pp PP P
....----- ,p p ,p 7P ~P
C><J ©
~p ~p p P
Concentric
jPP /PP /p
@ " -L "
II Parallel
-------
D ~ Perpendicular
O/J /
. ?/
Figure 3 Cases of tolerancing between surfaces, including the resulting MGD Es (redrawn after
Clement (1991». Note that the concentric relation in some cases - e.g. in case of the surface
of revolution - sphere association - is not literally a concentric relation; it represents the inter-
section or coincidence of both MGDE.
2.2 Tolerance specification
Tolerance specification is the activity of specifying tolerances; defining the tolerance types and
related tolerance values (geometric tolerances are assumed). Tolerance specification is prefer-
ably carried out in conformance with the tolerancing standards (e.g. ISO 1101 (1983), ANSI
Y14.5M (1982». However, the standards do not give a method of how tolerances should be speci-
fied. Therefore, research into the specification of tolerances seems to be required. Nevertheless,
not a lot of research into this field has been perfonned. The main contributions are described in
a number of publications by prof. A. Clement et al..
This approach to tolerance specification as presented in Charles et al. (1989), Clement et al.
(1991,1993,1994) and Dufosse (1993) is an exception to most previous approaches towards tol-
erance specification. Clement et al. propose a method to detennine tolerance types automatically
from the assembly model, resulting in a truly functional tolerance specification. Most other ap-
proaches in computer aided tolerancing are based on manual tolerance specification, usually
starting out from single components. The tolerance specification method by Clement et al. is car-
ried out on the basis of face associations between the different components in the assembly, also
called the mating function (Briard et al. 1989). By finding kinematic loops (in the graph repre-
senting the assembly), faces are found on individual components which can be toleranced relative
to one another (Dufosse 1993).
2.3 Tolerance analysis
Tolerance analysis is a method to verify the proper functioning of the assembly after tolerances
have been specified. Most often, tolerance analysis is perfonned by verifying two aspects:
• Verifying assemblability of the assembly; the feasibility of assembly (fit).

• Verifying if specified clearances between parts are still met; the quality of assembly (clear-
ance).
Tolerance analysis can be carried out by determining the tolerance zones belonging to the speci-
fied tolerance types. Generally, two types of tolerance analysis are distinguished:
1. Statistical tolerance analysis; in this case statistical methods are used together with accompa-
nying probability distributions.
2. Worst case analysis; the study of extreme cases (like for example MMC and LMC). Examples
of previous research in this case are described in (Hillyard 1978) and (Turner and Wozny
1988).
One of the first approaches in the field of tolerance analysis has been carried out by Bj¢rke in
which tolerance chains in 2D mechanisms were used to calculate (maximum) clearance (Bj¢rke
1978). However, this work was mainly restricted to conventional plus/minus tolerances in the 2D
case.
In tolerance analysis and tolerance synthesis, a mathematically correct model of tolerances is
required to be able to compare the influence of different types of geometric tolerances, and to
verify how they propagate in 3D. In tolerance analysis literature most often the worst case ap-
proach is followed accompanied by some additional assumptions: form tolerances are often as-
sumed to be negligible or to be present within position, orientation and size defects. Little is
known towards actual 3D tolerance analysis. Gaunet seems to be on the right track, presenting
a 3D tolerance analysis framework on the basis of the tolerancing method by Clement et al. (Gau-
net 1994). However, only a 2.5 D tolerance analysis example is provided.
2.4 Tolerance synthesis
Tolerance synthesis is regarded as optimizing and completing the (functional) tolerance specifi-
cation, taking into account manufacturing and inspection concerns. In this field relatively few
research has been published. Most approaches that have been published are based on the opti-
mization of a cost function. One of the most interesting methods seems to be the method by N assef
and EIMaraghy (1993), using genetic algorithms. The general system of which this synthesis ap-
proach is a part is described in (EIMaraghy and EIMaraghy 1993). Simulated annealing has been
employed for tolerance synthesis by Zhang et al. (1993).
3 OVERALL DESIGN OF THE TOLERANCING MODULE IN FROOM
A first design iteration of the design of the FROOM tolerancing functionality has been described
in (Jonge Poerink 1994). Figure 4 shows the main function of a tolerancing module in a simplified
way. Figure 4 also shows that apart from the tolerance types and values, a lot of additional data
can be generated by a tolerancing module, e.g. on feasibility and quality of assembly, cost etc ..
Designer knowledge Manufacturing knowledge
(geometric) tolerances
datums
assembly model assembly feasibility
Functional tolerancing module assembly quality
cost
manuf. info
assembly requirements inspection info
Figure 4 Main function of the tolerancing module.
assembly
part tolerances + assembly tolerances
manufacturing
knowledge, cost
assembly assembly model (extended with

model tolerancing info)
info on assembly feasi- (geometric) tolerances
datums
bility & quality
assembly feasibility
assembly qUality
optimized part (+ assembly) tolerances, cost info cost
manufacturing info
inspection info
Figure 5 The main functions: tolerance specification, analysis and synthesis and their connec-
tions.
The areas of tolerance specification, analysis and synthesis can be seen as functions that need to
be present in a re-design support system (Figure 5). The functional decomposition of Figure 5,
was decomposed further resulting in the following sub functions: tolerance representation, toler-
ance specification, tolerance analysis, tolerance synthesis, tolerance presentation, storage and re-
trieval (Salomons 1995). For each of these functions a selection was made as to how to fulfil it:
L For tolerance representation, i.e. the computer internal mathematical model of the tolerances,
the vectorial torsor approach combined with the TIRS and MGDE concepts as described in
e.g. Clement (1993, 1994) and Riviere (1993) has been selected. The main reason for this was
that most other methods were not in conformance with the tolerancing standards. Another rea-
son for selecting the torsor approach is that it is based on a strong mathematic background in
kinematics, e.g. (Herve 1978). Using this mathematic background can offer great advantages
in tolerance analysis and synthesis; see also section 5.
2. For tolerance specification, essentially the method as proposed by Clement et al. (1991) was
adopted, merely because there are no other methods for functional tolerance specification and
because of the straightforward and logical nature of this approach. More details are provided in
sections 4 and 6.
3. For tolerance analysis, the torsor approach offers a mathematically correct description of toler-
ance zones and can be used for tolerance analysis applications. Thus, the approach as described
in (Gaunet 1994) seems favorable. However, in Gaunet's approach to tolerance analysis, a lot
of equations are generated that need to be solved. Although a matrix approach as described in
(Riviere et al. 1994) reduces some of the redundant equations, a great number of parameters
remain (Jonge-Poerink 1994). Therefore, the number of parameters that are generated in the
approach by Riviere may be reduced by applying a kind of degrees of freedom analysis consid-
ering both the macro-DOF's as studied by Kramer (1992) and the micro DOF' s (tolerances) as
described in Clement (1993, 1994). This is detailed further in section 7.
4. For tolerance synthesis, none of the methods as presented in literature seems to be appropriate.
Almost all tolerance synthesis methods employ statistical techniques and a cost function in or-
der to achieve tolerance optimization. In the cost function, for each machining operation, the
relation between the cost versus tolerance (values) is expressed. However, during design it is
often not known in advance which processes and/or machines are going to be used for
manufacturing the individual parts. Therefore, it is difficult to derive a reliable cost function.
One could therefore better try to achieve a generic indication of production cost. In this case, it
is useful to employ general knowledge of the average cost of the manufacturing processes
available in combination with a tolerance factor by which different types of tolerances can be
unified (compared). Tolerance factors as proposed by Boerma (1990) may be used for this.
Due to space limitations this is not discussed further; refer to (Salomons 1995).
5. Tolerance presentation, i.e. presenting the tolerance information to the user, can be performed
in a variety of ways. Tolerance presentation should be supported as much as possible according.
to the standards to avoid confusion. However, it will not always be possible to show specified
tolerances or present results from tolerance analysis and synthesis according to the standards,
especially in 3D views. Therefore, graphs and tables could be used in addition to a presentation
method which conforms to the standards.
6. Storage of tolerance related information such as the tolerance types and values, results oftoler-
ance analysis and synthesis, can be performed in a number of ways. However, as a relational
database was already selected for other data storage functions in FROOM, the most appropri-
ate way to store the tolerancing related information was in a relational database. The data struc-
tures needed can be based on the information needed in TIRS and MGDE elements.
7. Assembly/component information retrieval (including tolerancing information) can be per-
formed by queries on the relational database.
4 ASSEMBLY MODELLING IN FROOM
For assembly representation in FROOM, conceptual graphs are used as an internal representation
format as well as for the external presentation to the end-user (Salomons 1994a, 1995). The con-
ceptual graph representing the assembly consists of the objects that make up the assembly and
the connection relations between these objects.
Assembly constraints can be specified by means of symbolic constraints like: "against", "align!
fit", "contact", "orient" and "parallel offset". The implementation of these relations (assembly
constraint satisfaction) is based on a slight modification of the work presented by Liu and Nnaji
(1991) and Kramer (1992). This method of constraint satisfaction symbolically reasons about the
geometry and based on this, a so-called plan fragment is generated. A plan fragment contains
the actions that should be performed in terms of rotations and translations in order to satisfy the
previously specified assembly constraint(s). This approach allows for automatically determining
the kinematic degrees of freedom (DOF) of each component and for automatically determining
the functional surfaces which is important for tolerancing. More details on the assembly model-
ling module of FROOM can be found in (Salomons 1994a, 1995).
5 TOLERANCE REPRESENTATION IN FROOM
In the following the concepts of1TRS, MGDE and torsor are elaborated further as these are cen-
tral to the tolerance representation FROOM. Finally, the FROOM internal tolerance representa-
tion format containing 1TRS, MGDE and torsors is discussed.
5.1 The applied TTRS concept
In tolerancing, surfaces are often associated two by two. The 7 elementary types of surfaces that
were derived by Clement et al. can be associated two by two. The association of surfaces is central
to the definition ofTTRS (Technologically and Topologically Related Surfaces) as has been pro-
posed by Clement et al. (1991) and Briard et al. (1989):
A TTRS is defined as an assembly formed by two sUrfaces (or between a surface and a ITRS
or between two TTRS) belonging to the same solid (topological aspect) and located in the same
kinematic loop in a given mechanism (technological aspect).
A somewhat different, more generic, definition of TTRS is provided by Riviere (1993):
A ITRS is a pair of sUrfaces (or ITRS) belonging to the same solid which are associated because
ofjunctional reasons.
Clement et al. have made an extensive classification of all possible associations of the 7 elemen-
tary face types, and thus 1TRS. One would expect 7 x 7 =49 different types of association. How-
ever, only relative positions of two surfaces are of interest, so that for instance the association
of cylindrical surface - prismatic surface is equivalent to the association prismatic surface - cy-
lindrical surface. Therefore, only 28 cases of surface association remain (i.e. 7 + 6 + 5 + 4 + 3
+ 2 + 1 cases or one half of a 7 x 7 matrix including the diagonal); see Figure 3. Although surface
associations can be regarded as "symmetric", tolerances cannot. This is due to the fact that there
is both a referenced and a toleranced face. Therefore, some tolerancing cases which are "below"
the diagonal in Figure 3, and shown blank, should also be regarded (Clement 1994).
When two elementary surfaces are combined in a TTRS, the resulting TTRS can be classified
into one of the seven basic classes depending on the types of surfaces involved and the geometric
relations between them. For instance, if two cylindrical surfaces are associated in a TfRS and
if their axes are coaxial, the resulting TfRS is classified as a cylindrical surface. This is the case
because both cylindrical surfaces remain invariant to rotations around their common axis and
translations along it. Another example could be the association of two parallel cylindrical faces;
these are only invariant for translations along the direction of the (parallel) axes and therefore
the association results in a prismatic TfRS. A systematic analysis of all possible cases of object
reclassification has revealed (at least) 44 cases of tolerancing (Riviere 1993), (Gaunet 1994). This
is shown in Figure 3. It must be noted here that the number of 44 was obtained by counting the
number of cases in Figure 3 and could grow if more special cases and cases below the diagonal
are considered. Therefore, the actual number of the tolerancing cases is less important than the
fact that there is only a finite number of tolerancing cases.
5.2 The applied MGDE concept
The Minimum Geometric Datum Element, or MGDE, of a TfRS is the minimum set of points,
lines or planes necessary and sufficient to define the reference frame corresponding to the invari-
ant sub-group of that TfRS. The concept of MGDE also has been proposed by Clement et al.
(1991, 1993). The MGDE remains invariant for the displacement it is defining. The MGDE is
a set of a reference point, reference line and a reference plane, but not all these elements are al-
ways necessary to define the MGDE sufficiently. The MGDE for a cylinder, for example, is the
axis of the cylinder (i.e. the MGDE consists of a reference line only). Figure 2 shows for each
elementary surface its associated MGDE as well as the symbol used for the MGDE.
5.3 The applied torsor concept
The torsor concept has originated in the field of metrology, e.g. Bourdet (1979). For every point
M in Euclidian space, its small displacements can be described by two vectors: DM and 8. The
e
DM vector represents 3 translations along perpendicular directions, two by two. The vector cor-
responds to three small rotations around perpendicular directions, two by two. These two vectors
define a so-<:alled small displacement screw model or torsor T M,B so that (Riviere 1993):
e "[~] OM" [~l T.," [~]- m

For each tolerance related to a TfRS, the tolerance zone can be represented as a tolerance torsor,
which represents the small displacements that are possible within the tolerance zone. In general
the torsor contains three translation parameters and three rotation parameters. However, as the
elementary surfaces, and thus the TfRS, usually have some invariances, often some of the torsor
parameters reduce to zero. For instance, a cylinder has two invariances; one translation along the
axis (x direction) and and a rotation around the axis (x direction). Therefore, a cylindrical toler-
ance zone (Figure 6) obtains the same invariances:
T_ ~ [~·l ~ [f]
T~- [D:] ~ m Figure 6 Cylindrical tolerance zone (re-
drawn after (Oaunet 1994).
Instead of torsors, matrices can be used for representing small displacements (Riviere et al.
1994). Let us again consider a cylindrical tolerance zone of diameter t. The toleranced element
is the straight segment AB which could for instance correspond to the axis of a cylinder surface
(Figure 6). The coordinate system is again defined in the point 0 in the center of AB and the x
direction is the direction from B to A. Now suppose we are considering a position tolerance of
the cylinder surface with respect to a complete reference system. The group of displacements that
does not leave the segment AB globally invariant can be represented by the transformation matrix
D(v,w, ~, Y), expressed in the coordinate system (0, x,y,z) (Riviere et al. 1994):
[
cyc~ - Sy CyS~ 0]
SyC~ Cy SyS~ v
D(v, w,~, y) = _ S~ 0 C~ w' (1)
o 0 0 1
As the segment AB remains in the cylindrical tolerance zone, the following constraints hold
(Riviere et al. 1994):
j YdA 2 + ZdA2 S 2t with

~:l = D*A
j Y.. 2 +ZdB 2 S2t

~:]
OS~Slt
with = D*B OsySlt
Note that the range of the angles ~ and y can be restricted further to: 0 S ~ S 2.bA and
0 < y < t
- - 2.0A.
5.4 Tolerance representation implementation in FROOM
The data structure in which both TTRS, MODE and torsor matrix are stored in FROOM is shown
in Figure 7. This structure is stored in the Oracle™ database. How TIRS, MODEs and torsors
are determined, is elaborated as part of the tolerance specification functionality.
TIRSl
-~
!, ~
TIRS2
-
i :1 ~~
N
8 en en 0)
<.l en
N
en :s!
§ :1
I
~~
0: 0:
~
~ ~ ~~ @
S
:E .~ § 5 .~
·5
;§ :E :s
5 ~ .!l
,::
~ '::l
~ ea
~
~ :s!
I tolerancel
~
~ .9 6
:E :s :a ~
~
~ .9 6
to!.
Toler- to!. ref. ref. to!. cy!. band ref. m crea to!.
obL
ance face_id face_i obL type zom widt off- max mat.
proj.
mode factor
Next
id h set mat. tol
d id
tolerance2
to!.
ref. to!. cy!. band ref. m proj. crea to!.
.....
Toler- to!. ref.
face_i obL type zon widt off- max
obL Next
ance face_id
id mat. tol mode factor
d id h set mat.
I MGDE I I
tj,:i
reference face reference line reference point
~ face normal
I reference ve~tor I
~ reference pomt reference pomt
Figure 7 Data structure for TfRS, MODE and torsors in FROOM.
6 TOLERANCE SPECIFICATION IN FROOM
Tolerance specification includes specifying the tolerance types and values for nominal geometry
parts. In current tolerancing practice, designers have to manually specify both tolerance types and
values either on a drawing or in a CAD system. The tolerance specification in this case depends
largely on the designer's judgement and experience. Therefore, different designers will possibly
arrive at different tolerance specifications for the same nominal geometry. Thus, a true functional
tolerance specification is not guaranteed. Functional tolerance specification can be performed
semi-automatically when an assembly model is present, assuming that the assembly model con-
tains (implicit) functional information (Clement 1991). Clement et aI. have presented a method
to detect functional surfaces in an assembly graph. Having determined the functional surfaces
in an assembly, the functional surfaces on each component are associated; TTRS are built. Also,
the datum elements (MODE) are automatically created, together with the creation of the TTRS.
The type of association indicates the type of tolerance that, from a functional point of view, needs
to be specified. The user has to determine the tolerance value. In the following we will elaborate .
on the theory by which the functional surfaces (TfRS) and datum elements (MODE) can be de-
termined (section 6.1). Then, implementation in FROOM is addressed (section 6.2).
6.1 Applied tolerance specification theory
Suppose we have an assembly graph showing the components, the assembly relations like
"against" and "align/fits" as well as the surfaces of the components involved in these mating
constraints. These surfaces are considered as functional surfaces. The functional surfaces and the
assembly relations give rise to "kinematic"loops in the assembly graph. The kinematic loops in
the assembly graph in a way overconstrain the nominal positions of the components in the loop
(when expressing the positions in terms of algebraic equations). Tolerancing the functional sur-
faces involved in the loop is a way to relax the overconstraints (Dufosse 1993). Figure 8 gives
an example; the parts are represented by circles while the mating relations are expressed as arcs
and functional surfaces are shown as partial circles that are placed on their component circles.
If part I is positioned relative to part II, it can easily be observed that part I's x-position is overcon-
strained by the two align/fit relations. If the holes in part II would have had bottom faces, part
I's y-position would also have been overconstrained. Different arc types denote different assem-
Figure 8 Loops between functional surfaces in an assembly graph.
bly relations and together with the geometric features they represent different directions in which
tolerances work. Note that in Figure 8 there is only one loop working in the same direction. Thus,
the faces S 11 and S 12 on part I have to be associated as well as S22 and S21 on part II. Figure 3 is
used for determining the type of the resulting TTRS. Now on each part, one functional face re-
mains. As the remaining faces are functional, they also should be toleranced. This can be done by
relating the remaining surfaces to the previously built TTRS andlor by tolerancing the respective
MODE's (Desrochers et al. 1994).
The shortest kinematic loops one can detect are those between two parts. Two different types
of kinematic loops are then possible (Jonge Poerink 94):
1. Loops consisting of two surfaces on each part.
2. Loops consisting of one surface on one part and two on the other.
The latter case can provide some problems, however. If a loop consisting of one surface on one
part and two on the other is the first loop considered for the part with one functional face in the
loop, a "normal" TTRS cannot be built. The surface can only be toleranced by a form tolerance.
Therefore, these types of loops should be avoided as starting loops.
In the assembly graph loops can easily be detected using standard graph theory like e.g. (Even
1979) and (Noltemeier 1976). A kinematic loop in the assembly graph always consists of more
than one part and no more than two surfaces on each part in the loop. There can be more than one
loop between several parts and one surface can participate in several loops. As there can be more
than one loop, the order in which surfaces are to be associated is not trivial. In fact, it depends
on the order in which the loops are processed. Also, loops should be independent; they may not
be linear combinations of others. Therefore, once the loops have been detected, the following still
has to be resolved:
1. Determining the starting loop.
2. Determining the sequence in which the loops have to be processed.
Clement et al. have proposed the following criteria, that are mainly related to technological rea-
sons, for the detection of the starting loop (Clement 1991), (ISMCM):
1. One dimensional loops passing through contacts having their normal vectors pointing in oppo-
site directions (tolerance chain).
2. Shortest independent loops; i.e. loops involving the least number of parts.
3 Choosing loops generating the TIRS that were designated using dimensioning assistance or
through technical functions.
4. Choosing loops passing through surfaces designated as references (MODE).
The above rules apply to one direction in the assembly graph. For all directions in the graph,
all loops in the same direction have to be found and based on the above rules, the loops have to
be sorted. By doing this, TTRS trees per direction and per component can be constructed. If these
TIRS, possibly in combination with remaining functional faces, are unrelated, they can be
associated on their turn. This can be done by considering the previously built TIRS and shortest
loops containing multiple directions as well as MODE. In this way, a sensible dimensioning of
every subassembly is allowed, as well as the generation of the shortest tolerancing chain. Howev-
er, the previous criteria may not always be sufficient, and therefore users should be enabled to
select the order in which loops are to be processed (Clement 1991). Jonge Poerink added some
additional criteria for loop selection (Jonge Poerink 94):
1. Avoid loops that contain only one surface on a part.
2. Choose loops containing the simplest basic geometry.
Automatic tolerance value specification is not examined. Checking for coherence and complete-
ness of specified tolerances is also not addressed, although these can be considered as part of tol-
erance specification.
6.2 Tolerance specification as implemented in FROOM
Based upon the theory as presented above, FROOM is able to generate tolerance types automati-
cally based on a conceptual graph representation of an assembly model. The rules for loop detec-
tion, constructing MODE, TTRS etc. are implemented using C++. Figure 9 provides an example
of tolerance specification of the gearpump which will be addressed later for explaining the con-
cepts regarding tolerance analysis. Apart from an automatic tolerance specification, a manual tol-
erance specification is desired; not all tolerances should by definition be of functional impor-
tance. The manual tolerance specification in FROOM is currently under development.
7 TOLERANCE ANALYSIS IN FROOM
This section only summarizes the theory of tolerance analysis as is to be applied in FROOM; it
provides an extension to the theories on tolerance analysis by Oaunet (1994) and Riviere et al.
(1994). The implementation is currently under development and has not been completed; there-
fore it is not discussed. A distinction is made between analyzing the feasibility of assembly (7.1)
and the quality of assembly (7.2).
: . ..
: ....
~'~':."" .
".
:!-,-~
.... ,.. .,.
.. .. ~p
.t§§ ~ ";1
"~~ ............ ~k
..
';:::::.:::::::::::::::::::::::::::::::::::::::::::.:';;::::::::::::::'::.'f'~''''''''''''''''::';
-------.
?-
Figure 9 The system proposes a parallelism tolerance between the two gear holes.
7.1 Analyzing the feasibility of assembly
Feasibility of assembly can be considered by checking face associations two by two, without con-
sidering the other faces on a part or in the remainder of the mechanism. Tolerance torsors can be
used to calculate softgages for each surface. Softgages are the nominal geometry combined with
their virtual tolerance zones. Softgages can be calculated for most material condition and least
material condition. Using MMC softgages, a check for feasibility of assembly can be performed.
For each point on a surface, its virtual displacement can be calculated using the torsor related
to that surface (association). As there can be infinitely many points on the surface, usually a re-
stricted number of discrete points is chosen which are used to construct the softgage. The more
points are chosen, the better the approximation of the softgage. The comparison of the softgages
of two different surfaces then results in insight into the feasibility of assembly; interference or
not. If there is interference, either the nominal dimension values or the tolerance values can be
changed, in order to achieve non-interference.
7.2 Analyzing the quality of assembly
As in both the torsor approach as presented by Gaunet (1994) and the matrix approach by Riviere
et al. (1994), a great number of constraints and parameters appear, it seems of great interest to
reduce the number of parameters before trying to solve the constraint set. Also, in the currently
presented tolerance analysis approaches it is difficult to find the directions in each "joint" which
influence the most the clearance one is interested in. In fact, this seems to be the main obstacle
in applying the torsor approach to 3D tolerance analysis. Including an analogy to degrees of free-
dom analysis as proposed by Kramer seems to be a possible solution to both problems; by deter-
mining the relevant directions for each joint in which calculations should be made this could re-
duce the constraint/parameter set. To illustrate this idea, an example of a gearpump, similar to
(Farmer and Gladman 1986), (Gaunet 1994) and (Riviere et al. 1994) is presented. The example
includes TfRS and MGDE construction.
7.2.1 A simple example
The pump considered is an assembly of a housing and two shafts; the cover and the bolts fitting
the cover on the housing are not considered. The shafts are considered to contain the gears. Figure
10 shows the pump with only some size and fit tolerances.
01517
Figure 10 The pump (dimensions and size and fit tolerances after (Riviere et al. 1994».
Tolerance specification for the gearpump example
As has been explained previously, tolerance types can be determined by detecting kinematic
loops from the assembly graph. Figure 11 shows the assembly graph and the loops in the case of
the pump. The faces AI, BI, CI and 01 are surfaces of part 1 (the housing), faces A2 and B2
Detected loops:
(AI. BI, A3, B3)

(CI, 01, A2, B2)
(CI, AI, B2, B3)
(DI, BI, A2, B2, A3, B3)
Figure 11 The assembly graph of the pump mechanism and the loops that can be detected.
are surfaces of part 2 (the small shaft) while faces A3 and B3 are surfaces on part 3 (the large
shaft).
According to the two most important criteria for loop sequencing, i.e. shortest loops and one
dimensional loops, the loops (AI, BI, A3, B3), (CI, 01, A2, B2) and (AI, CI, B2, B3) would
be the first to be processed. However, the loop (AI, CI, B2, B3) is dependent on the other two
shortest loops. Therefore it is a dependent loop and will not be considered. As the two remaining
loops are of equal size and contain cylindrical surfaces, either loop can be used as a starting loop
for building the TfRS. Consequently, the loop that is not chosen as the first, has to be used as
second. Suppose we choose loop (CI, 01, A2, B2) as the first loop. Then loop (AI, B 1, A3, B3)
will be the second loop. The remaining loop is (01, BI, A2, B2, A3, B3). According to the main
loop criteria, the sequence would then be:
1. (CI, 01, A2, B2)
2. (AI, Bl, A3, B3)

3. (AI, Cl, B2, B3) (not considered; dependent loop)
4. (Dl, Bl, A2, B2, A3, B3)
The first loop contains 4 surfaces: surfaces Cl and Dl on part 1 and surfaces A2 and B2 on part
2. Therefore, the first TTRS of part 1 contains surfaces Cl and D1. According to Figure 3, the
resulting class is a cylinder (C 1 and D 1 are both cylindrical surfaces and are concentric). Accord-
ing to Figure 3, the MODE for this TTRS can be the axis (of surface D 1 or C 1). In this case, sur-
face Dl has apparently been selected arbitrarily (often the choice will be on the largest surface
which in this case is Cl). Surfaces A2 and B2 on part 2 also result in a cylindrical TTRS as they
are concentric cylinders as well. The second loop can be processed similar to the first. The third
loop contains only faces that are already used in previous TTRS and none of these faces is a refer-
ence face or MODE. Therefore, the third loop is a dependent loop and not relevant to building
a TTRS. In the last loop, all faces have been used in previous TTRS, but the axes of D 1 and B 1
are MODEs. Also, because of the link B2-B3, it is an independent loop. Therefore, faces D 1 and
B 1 on part 1 are related; the resulting TTRS is prismatic. According to Figure 3, the resulting
MODE is a combination of a line and a planar face. The resulting TTRS trees that can be gener-
ated from the loops of the functional faces of the parts of pump are shown in Figure 12.
CI DI Al BI A2~B2 A3~B3
ITRS~2(CYI) ITRS2.I(cyl) ITRS3.I(cyl)
TIRSl.3
(prism.)
MGDEI.I:axisofDI MGDEp.,a MGDEpart3
MGDE2.1: axis MGDE3.1: axis of A3
MGDEI.2: axis ofBI ofA2
MGDEpartl = MGDEI.3: composed of axis ofDI
and plane of axes of Bland D 1
Figure 12 The resulting TTRS trees for each part of the pump.
Based on the TTRS trees that are obtained, geometric tolerance types can be proposed automati-
cally by the system. For TTRS1.l this results in a concentricity tolerance between face Cl and
Dl and a cylindricity tolerance on each individual face. TTRS1.2 is similar to TTRS1.1.
TTRS2.1 and TTRS 3.1 are also similar to TTRS1.1. For TTRS 1.3 it is necessary to tolerance
Dl and B1. As the faces Dl and Bl are nominally two parallel cylindrical faces, a parallelism
and a position tolerance are applied to both surfaces. The proposed geometric tolerances are
shown in Figure 13. The arrows point from reference to toleranced face. Note that for reasons
of simplicity, unlike the example of Figure 8, all tolerance directions are the same. If, for instance
we would have associated the planar face of the bottom of holes Al and Cl with the correspond-
ing faces of the gears, different directions would have been introduced. Also, if the cover would
have been included in the example, this would have been the case.
Tolerance analysis in the gearpump example
Suppose we are interested in the clearance between the two gears: between face B2 and face B3.
Figure 14 shows the assembly graph once more; the clearances are now expressed as "tolerances"
in the graph. The clearance of interest is part of two of the previously determined loops. Normally,
one would have to calculate the clearances resulting from both loops independently. From this,
the clearance that is most significant should then be determined. This criterion has not been iden-
tified in previous literature. In our example, however, we will only consider the loop (Dl, Bl,
A2, B2, A3, B3), as the other loop has been shown to be dependent. Note that in this case, the
tolerance analysis problem is primarily a 2D problem due to the simplified model. In Figure 14,
i -.. displacement
.. - J associated to zone
Figure 13 The geometric tolerance types that Figure 14 The assembly graph with both the
can automatically be proposed. inter-piece contacts and the inner piece TfRS
symbolized as tolerances and displacements
(slightly adapted from Riviere et al. (1994)).
01 has been hatched to indicate that the displacements of surfaces in the loop influencing the
clearance are calculated relative to 01. The situation has been outlined in Figure 15.
y
Mak..- • - -- -- - - -- - - -- - ~-1- ~ - ------------~ MSlb
-,;.- - - - -"+
I ......................................... ~ ... ~ ... X................................. .
y 15
~t
--------- -J
""'-" , - - -- m - - -
, z x
.. 5 ..I~ ~5- _&2_ -- -- -- - MDtb
Figure 15 Tolerance zones between surface Bl and 01 (redrawn after Riviere et al. (1994)).
We will now consider the constraints involved in the displacement ofBl relative to 0RI. Sup-
pose R is the mechanism coordinate frame and Rl and R2 are those of surfaces Bl and 01. If
PR-I-R I is the transformation matrix of coordinate frame R to frame R I, then (Riviere et al. 1994):
[~(O,'"") -['~"]' [~(""';

The matrix Du representing the tolerance zone of surface B 1 with respect to OR 1 can be written
as in equation (1). The vector of displacement MM' of a point M in the tolerance zone to point
M' expressed in RI is now (Riviere et al. 1994):
The constraints to be fulfilled at the extremes MA and MB of the cylindrical tolerance zone are
(Riviere et al. 1994):
The other constraints for each tolerance are derived similarly. Using these constraints, the mini-
mum/maximum clearances can be calculated. In the example of the pump (Riviere et al. 1994),
human interpretation easily concludes that only the y-direction is of importance for clearance
calculations. However, a computer needs generic methods which also help out in more complex
cases. In the following paragraph, this is detailed further and can be seen as an extension to the
theory by Riviere et al. (1994).
7.2.2 An extension to the method by Riviere et aI. for 3D tolerance analysis
In the pump example, when calculating the clearances, two halves of the kinematic loop are
"walked through" starting from a reference face (face D 1). In the case where one aims at reducing
the parameter set, it is preferable to start out at the faces that one is interested in (i.e. the faces
forming the clearance or orientation). Based on the face types involved, it will then be possible
to determine a kind of virtual plan fragment, analogous to the plan fragments by Kramer (1992).
By virtual is meant that only directions are considered along which translations (in case of clear-
ance) or rotations (in case of orientations) can be carried out. Thus, no values are assigned to these
translations and rotations. As a matter of fact, these values related to the translations and rotations
are the items one is finally interested in as they determine the clearance or orientation one would
like to know. The two components participating in the clearance or orientation are called primary
components from now on. Once a virtual plan fragment has been calculated for each primary
component, i.e. a direction in which clearance or orientation are maximum, the two halves of the
kinematic loop can be traversed back to the reference face. During this process, the number of
parameters can be reduced by reasoning on the macro-DOF's (actual kinematic degrees of free-
dom; large displacements) in relation to the micro-DOF's (tolerances).
In the case of our pump we are interested in the clearance between the two gears (part 2 and
3). These gears are parallel cylindrical surfaces (simplified geometry), and have each one rota-
tional macro-DOF around axes that are parallel. By means of a kind of plan fragment table we
may be able to find the virtual plan fragment (direction in which the clearance is maximum). In
fact, the plan fragment table should contain all possible cases of the face associations as shown
in Figure 3. However, the faces that are associated must be part of different solids now. Figure
16 shows a preliminary version of such a virtual plan fragment table for clearances. Figure 17
illustrates some of the cases of figure 16 in some more detail. In the case of two parallel cylinders
(that do not share their axes), the virtual pan fragment can be determined by calculating the
direction of the line perpendicular to the two axes (or MODE): see figure 17.
Now, for the directions of D22 and D33 that are internal to parts 2 and 3 respectively and that
form part of the two halves of the kinematic loop, only the direction determined by the virtual
plan fragment is important. For the mating between parts 2 and 1 and parts 3 and 1, again a virtual
plan fragment is derived, using figure 16. The difference with the previous, "primary" plan frag-
ment is that in this case itis between a primary component and a secondary component: a "second-
ary" plan fragment. In general, the new plan fragment directions are compared with those of the
primary plan fragment. In this case we have two cylinder-cylinder combinations which are con-
centric. Then the direction of the secondary plan fragments may be in any plane perpendicular
to the line MODE. As the direction of the primary plan fragment lies in this plane, it can again
be selected as the direction in which to calculate. Thus, D12 and D13, which represent the clear-
ances between parts 2 and 1 and parts 3 and 1 are calculated only in the direction of the primary
plan fragment. As we now arrived on the base component, i.e. the component with 0 macro-
DOF's, no new virtual plan fragments are needed. It is sufficient to observe Dl1 in the required
direction. Thus, using an algorithmic approach, we can arrive at the same set of equations as those
~®~~~eD0
rp-
~
no redw:tion DO
poealJle
reduc- 0 reduc-
liDo poIIible ion pouible
no reduc-
tion pouible
~ reduc-
lOOn posa'ble
no recl&c-
lion poaible
no reduc-
tiaa poaible
dcpeads OIl no toduc- no reduc- depends on DO reduc- depeDds on

posilion &: lion possible tion pouib. positioo &. lion. posu'blc poIitioo &:
....... w ~, ..w
Figure 16 Preliminary virtual plan fragment table for (maximum) clearances. This table will
probably only hold in case of form closed loops and not in case of force closed loops.
sphere-cylinder
_~,,-._Any
20
cylinder-cylinder helical-helical
cylinder-plane s.of rev - surf. of rev. ~
.~~.
Figure 17 Some cases of Figure 16 in more detail.
that were manually derived in (Riviere et al. 1994). In this example, however, no directions had
to be taken into account for a certain tolerance zone in the chain for which that tolerance zone
is invariant, but for which the clearance one is interested in is not invariant. Especially in these
cases reasoning on the macro-OOF's is necessary.
Although only one simple example was shown in which a kind of OOF analysis has been used
to reduce the parameters in the sets of equations and to determine directions in which to calculate,
this seems to open the way for true 3-D tolerance analysis.
8 CONCLUSIONS AND RECOMMENDATIONS
A tolerancing tool has been presented which allows for functional tolerance specification. Toler-
ance specification is based on the method by Clement. This method has been verified by the cur-
rent research and shown to be a "sound" method although further evaluation in practice seems
desirable. Tolerance representation is based on the TIRS concept. For tolerance analysis, an ex-
tension has been presented to the methods by Gaunet and Riviere et al.. The extension for toler-
ance analysis combines the kinematics approach for small displacements (tolerances) with that
of large displacements. However, these extensions have not yet been implemented and verified.
In the future this should be done. More effort in the field of checking completeness and coherence
of tolerance specifications seems required as well. Also, tolerance synthesis should be developed
in combination with the presented tolerance specification and analysis methods.
Acknowledgements
This research has been supported by the Technology Foundation (or S1W: a Dutch foundation,
funded by the Dutch government).
9 REFERENCES
ANSI Y 14.5 (1982) Dimensioning and tolerancing, American Society of Mechanical Engineers,
New York.
Ballu, A (1993) Identification de modeles geometriques composes pour la specification et la me-
sure par coordonnees des caracteristiques fonctionelles des pieces mecaniques, in French,
PhD thesis (these de Nancy I).
Ballu, A. and Mathieu, L. (1993) Analysis of dimensional and geometrical specifications: stan-
dards and models, Proc. of3rd CIRP Seminar on Computer Aided Tolerancing, Cachan (F),
157-170.
Bjf,'!rke, 0. (1978) Computer-Aided Tolerancing, Tapir publ.
Boerma, J.R. (1990) The design of fixtures for prismatic parts, PhD thesis, University ofTwente,
Enschede.
Bourdet, P. (1979) Identification geometrique des pieces mecaniques, these de 3eme cycle, Paris
VI.
Briard, M., Charles, B., Clement, A., Pelissou, P. and Riviere, A (1989) The mating function in
CAD/CAM systems, Proceedings, Design Automation Conference Montreal, Canada,
September, 1989.
Charles, B., Clement, A., Desrochers, A., Pelissou, P. and Riviere, A (1989) Toward a computer
aided tolerancing model, Int. Conf. on CAD/CAM and AMT in Israel, Proceedings, CIRP
sessions on tolerancingfor junction in a CAD/CAM environment, vo1.2, December 11 -14,
Jerusalem, Israel.
Clement, A., Desrochers, A and Riviere, A (1991) Theory and practice of 3D tolerancing for
assembly, Proceedings, CIRP seminar on Computer aided tolerancing, Penn State Univer-
sity, USA
Clement, A. and Riviere, A (1993) Tolerancing versus nominal modelling in next generation
CAD/CAM system, Proceedings, CIRP Seminar on Computer Aided Tolerancing, ENS de
Cachan, Paris, 97 - 113.
Clement, A, Riviere, A and Temmerman, M. (1994) Cotation tridimensionelle des systemes me-
caniques, tMorie & pratique, PYC Edition, Yvry-Sur-Seine Cedex (ISBN
2-85330-132-X), in French (English version is in progress).
Desrochers, A and Clement, A. (1994) A dimensioning and tolerancing assistance model for
CAD/CAM systems, Int. J. Adv. Manuf. Techno!., 9, 352-361
Dufosse, P. (1993) Automatic dimensioning and tolerancing, Proceedings, 3rd. CIRP seminar
on Computer Aided Tolerancing, Cachan (F), 1-10.
EIMaraghy, H.A and EIMaraghy, W. H. (1993) A system for modeling geometric tolerances for
mechanical design, Proc. 3rd CIRP seminar on Compo Aided Tolerancing, Cachan (F),
11-24.
Even, S. (1979) Graph Algorithms, Computer Science Press.
Fanner, L.E. and Gladman, e.A. (1986) Tolerance techno1ogy- Computer-based analysis, An-
nals of the ClRP, Vol.35, No.1, 7 -10.
Gaunet, D. (1994) Modele fonnel de tolerancement de position, contributions a l'aide au toler-
ancement des mecanismes en CFAO, in French, PhD thesis, ENS de Cachan.
Herve, J.M. (1978) Analyse structurelle des mecanismes par groupe des deplacements, Mecha-
nism and Machine Theory, Vol.l3, 437-450.
Hillyard, R.C. and Braid, I.e. (1978) Analysis of dimensions and tolerances in computer-aided
mechanical design, Computer-Aided Design, vol. 10, no.3, 161-166.
Houten, EJ.A.M. van (1991) PART: a computer aided process planning system, PhD thesis, Uni-
versity of Twente, Enschede.
ISMCM, (year of publication unknown) Resolution Tensor, tome 1, Institut Superieur des mater-
iaux et de la Construction Mecanique, Laboratoire de Mecatronique, infonnation bulletin
on the research at ISMCM obtained via personal communication, St-Ouen (F).
ISO 1101 (1983) Technical drawings - Geometrical tolerancing - Tolerances of fonn, orienta-
tion, location and run-out - Generalities, definitions, symbols, indications on drawings.
Jayaraman, R. and Srinivasan, V. (1989) Geometric tolerancing: I Virtual Boundary Require-
ments, IBM 1. Res. & Developm, Vol. 33, No.2, 90-104.
Johnson, R.H. (1985) Dimensioning and tolerancing, final report, R84-GM-02-2, CAM-I, Ar-
lington, TX.
Jonge Poerink, H.J. (1994) A functional tolerancing module for FROOM. MSc thesis, Uni-
versity of Twente, report no. PT-516.
Koplewicz, D. (1993) Specification by tolerance zones, current situation and evolutions, Proc.
3rd ClRP Seminar on Computer Aided Tolerancing, Cachan (F), 2l3-222.
Kramer, G.A. (1992) Solving geometric constraint systems, a case study in kinematics, The MIT
Press, Cambridge, Massachussetts.
Liu, H-C. and Nnaji, B.O. (1991) Design with spatial relationships, Journal of Manufacturing
Systems, Vol.IO, No.6, 449-463.
Nassef, A. and EIMaraghy, H.A. (1993) Allocation of tolerance types and values using genetic
algorithms, Proceedings 3rd. int CIRP sem. on Computer Aided Tolerancing, ENS de ea-
chan, 147-156.
Noltemeier, H. (1976) Graphentheorie mit Algorithmen und Anwendungen, (in Gennan) de-
Gruyter Lehrbuch, Berlin.
Ranyak, P.S. and Fridshal, R. (1988) Features for tolerancing a solid model, Proceedings, Com-
puters in Engineering Conference (CIE), San Francisco, 275-280.
Requicha, A.A.G. (1983), Toward a theory of geometric tolerancing,Int. J. ofRobotics Research,
Vol. 2, No.4, 45-59.
Requicha, A.A.G. (1984), Representation ofto1erances in solid modeling: issues and alternative
approaches, in: Solid modeling by computers, from theory to applications, eds. Picket M. S.,
Boyse J.W., Plenum press.
Requicha, A.A.G. and Chan S.C. (1986) Representation of geometric features, tolerances and
attributes in solid modelers based on constructive geometry, IEEE Journal. of Robotics &
Automation, Vol. RA-2, No.3, 156-166.
Riviere, A. (1993) La Geometrie du groupe des dep1acements appliquee a1a modelisation du tol-
erancement, PhD thesis, Paris.
Riviere, A., Gaunet, D., DuM, I. and Desrochers, A. (1994) Une approche matricielle pour la
representation des zones de tolerance et des jeux, Proceedings, FORUM 1994 de la SCGM
(Societe Canadienne de Genie Mecanique), June 27-29, McGill University.
Salomons, O.w., Houten, EJ.A.M. van, and Kals, H.J.J. (1993a) Review of research in feature-
based design, Journal of Manufacturing Systems, Vol. 12, no.2, 1l3-132.
Salomons, O.w., Kappert, J.H., Slooten, Evan, Houten, EJ.A.M. van, and Kals, H.J.J. (1993b)
Computer support in the (re)design of mechanical products, a new approach in feature
based design, focusing on the link with CAPP, IFIP Transactions B-ll, Knowlegde Based
Hybrid Systems, Mezgar I., Bert6k P."(eds.), Elsevier Science Publishers B.V. (North Hol-
land), pp 91-103. (Also: internal report PT 455), http://utwpue.wb.utwente.nllstw-doc!pa-
pers/paper-knowhs. ps.
Salomons, O.W., Slooten, Evan, Houten, EJ.A.M. van, and Kals, H.I.I. (1993c) A computer
support tool for re-design, a prototype system resulting from applying a methodic design
approach, proceedings of the International Conference on Engineering Design, ICED '93,
The Hague, August 17 -19, Vol. 3, 1559 - 1570, http://utwpue.wb.utwente.nllstw-doc/pa-
pers/paper-iced. ps.
Salomons, O.w., Slooten, E van, Koning, G.W.E de, Houten, F.J.A.M. van, and Kals, H.I.I.
(1994a) Conceptual graphs in CAD, CIRP annals, Vol. 43/1, 125-128, http://utw-
pue. wb. utwente.nllstw-doc/papers/paper-<:irp94. ps.
Salomons, O.W., Jonker, H.G., Slooten, Evan, Houten, EJ.A.M. van, and Kals, H.J.J. (1994b)
Interactive feature definition, proceedings IFIP WG 5.3 Conference on Feature modeling
and feature recognition in advanced CAD/CAM systems, Valenciennes, May 24-26,
181-204, http://utwpue.wb.utwente.nllstw-doc/paperslpaper-ifd.ps.
Salomons, O.W. (1995) Computer support in the the design of mechanical products, constraint
specification and satisfaction in feature based design for manufacturing, PhD thesis, Uni-
versity of Twente, http://utwpue.wb.utwente.nll-otto/thesis/
Spatial Technology (1994) ACIS Geometric modeler, Boulder, Colorado.
Srinivasan, V. and Jayaraman, R. (1989) Geometric tolerancing: II. Conditional tolerances, IBM
J. Res. & Developm, Vol. 33, No.2, 105-123.
Srinivasan, V. (1993) Recent efforts in mathematization of ASMEIANSI Y14.5M, Proc. of 3rd
CIRP Seminar on Computer Aided Tolerancing, Cachan (F), 223-232.
Turner, J.U. and Wozny, M.J. (1988) A mathematical theory oftolerances, in: Geometric Model-
ling for CAD Applications, eds. Wozny M.J., McLaughlin J.L., Elsevier Science Publish-
ers,IFIP, 163-187.
Turner, J.U. (1993) A feasibility space approach for automated tolerancing, ASME Journal of
Engineering for Industry, Vol. 115,341-345.
Wirtz, A. (1993) Vectorial tolerancing a basic element for quality control, Proc. 3rd CIRP
seminar on Computer Aided Toierancing, Cachan (F), 115 -128.
Zhang, C. and Wang, H-P. (1993) Integrated tolerance optimisation with simulated annealing,
J. Adv. Manu! Techn., vo1.8, 167-174.
10 BIOGRAPHY
O.W. Salomons obtained a PhD degree on his work on the FROOM system at the University of
Twente. He obtained his MS degree in Mechanical Engineering in 1990 at the same university.
Presently, as an assistant professor, he is perfonning research at the laboratory in the field of de-
sign support systems as well as their link with process planning systems.
H.J. Jonge Poerink recently obtained his MS degree in mechanical engineering at the University
of Twente. His MS work was on the tolerancing module in the FROOM system. Currently he is
employed at Philips Machinefabrieken.
Evan Slooten holds a BS in Computer Science, which he obtained in 1990. Since 1991 he works
as a system analyst! technical software developer at the laboratory of Production and Design En-
gineering of the University of Twente.
EJ.A.M. van Houten is associate professor of the laboratory of Production and Design Engi-
neering at the University ofTwente. Professor Van Houten obtained his MS degree in Mechanical
Engineering at the Technical University of Eindhoven in 1976. He has been working at the labo-
ratory of Production Engineering at University of Twente since 1978. Van Houten has worked
in the field of CAD and CAPP; he has been closely involved with the development of several
CAPP systems. In 1991 he obtained a PhD degree on his work on the PART system.
H.J.J. Kals is professor and head of the laboratory of Production Engineering at the University
ofTwente. He is also part-time professor at the Technical University of Delft. Professor Kals ob-
tained his MS degree in Mechanical Engineering in 1969 at the Technical University of Eindho-
ven. In 1972 he obtained his PhD degree. In 1977 he became professor of the Laboratory of Pro-
duction & Design Engineering at the University of Twente. Professor Kals is a member of CIRP.
He is active in the fields of CAD, CAPP, CAM, workshop- and work station control. Currently
one of his main scientific interests is concurrent engineering.
11 QUESTIONS
The following questions were asked after the presentation:

Question by prof. A. Clement: Do you think working with such a VPF table will also work in
every possible case?
Answer by Otto Salomons: We are not sure about that yet The VPF approach has only been
tested on relatively simple assemblies so far such as for example the gearpump, or a conical gear.
We are now trying to make the approach generic so that it also works for more practical examples.
However, we have a very good feeling about this approach.
Question by prof. C. Fortin: Can you actually reduce the number of constraints in your ap-
proach?
Answer by Otto Salomons: In principle you can only reduce the number of parameters in certain
constraints and you cannot in general reduce the constraint set itself. However, in some cases you
can reduce some constraints. For instance in the case where you are dealing with force closed
connections.
5
Modelling spatial dimensional chains
for CAD/CAM applications
V. T. Portman
Department of Mechanical Engineering, Ben-Gurion University of the
Negev, POB 653, Beer She va, 84105, israel. Phone: 972-7-461348,
Fax: 972-7-280252, e-mailportman@menix.bgu.ac.il
R. D. Weill
Faculty of Mechanical Engineering, Technion -israel institute of
Technology, Haifa, 32000, israel. Phone: 972-4-293851,
Fax: 972-4-324533, e-mail MERIB02%TECHNION.BITNET@taunivm.tau.ac.i1
Abstract
An analytical and computational procedure for modelling spatial dimensional chains (DC) is
based on the kinematic interpretation both of nominal dimensions forming the DC and their
errors. The goal of the investigation is to assess the accuracy of the closing link of the DC,
taking into account the errors of the intermediate links of the DC. The three-stage procedure
successively models the nominal DC, the actual DC with positional errors of the links, and the
metrological estimations of the accuracy. The latter stage is undertaken using metrological
models of accuracy, which brings the accuracy estimation into conformity with a nominal form
of the boundaries of the DC's closing link and the relative positions of the boundaries. The
theoretical results apply to important design problems: evaluating closing tolerances, choosing
compensating links, and structural optimization of the DC according to accuracy criteria. As an
example, calculation of the spatial DC between two perpendicular bores is considered.
Keywords
Tolerancing, dimensional chain, compensation, accuracy evaluation, accuracy optimization
INTRODUCTION
The term "Dimensional chain (DC)" is a fundamental concept which is related to the tolerancing
of relative positions of features in a manufactured part after different stages of the
manufacturing process: design, process planning, machining, measuring, and assembling. This
concept was successfully used for the establishment and observance of tolerances based on
both probability theory and on a deterministic approach (Kirshling, 1991), computer aided
tolerancing (Fainguelernt et aI., 1986; Weill, 1988; Bjorke, 1989), application of computer
intelligence in design, in the frame of concurrent engineering (Dong, 1993), etc. The accuracy
optimization procedures for machining (Irani, et ai., 1989) and tolerancing were analyzed based
on the DC models. Fundamental notions related to the DCs was standardized in the former
USSR. The overwhelming majority of investigations in the field of DCs focuses on models of
the linear DCs; i.e., either the proper linear DCs or the angle DCs (sequence of rotations around
a common axis). The problem of the spatial DCs, i.e. the DCs in the 3D-space, remains open,
although partial approaches have been developed successfully (Charles. Clement. et al .• 1989.
Weill. et al .• 1991. Shen and Duffle. 1991).
From the point of view of tolerancing computation. the most important feature of the linear
DC is the collinearity of all the nominal dimensions (members of DC) and their errors. This
feature is invalid for spatial DC. since each nominal dimension in the space connects with six
elementary errors (three small translations along the X. Y. and Z axes and three small rotations
around these axes). As a result. a number of difficulties arise which cannot be resolved by
methods which are typical for the linear DCs. The difficulties are easily observable in a
representation of the spatial DC in the form of the set of projections on the two or three
projection planes (Yakushev. et al .• 1986) or in the form of tree branch graphs (Dong and
Soom. 1991). Namely. there is an interrelation between dimensions and errors. which are
incorporated into diverse branches of a dimension tree. This problem has been under study
(Jayararnan and Srinivasan. 1989; Turner and Wozny. 1990).
A simple example is represented in Figure 1. where features HI and H2 (two bores. the axes
of which are nominally perpendicular one to the other) of a part are slightly shifted because of
errors in the DC. such as errors of positioning of the part in its fixtures. geometric and
kinematic errors of the machine tools. etc. The interrelation of accuracy parameters for this case
is apparent. for example. from the formula that estimates the deviation Lld of the distance
between the axes of the bores HI and H2:
where cSBI and cSB3 are the linear errors of the dimensions Bland B3. respectively; and cxo and
~O are the deviations from parallelism of the axes of the bores HI and H2. respectively. with
reference to the base plane.
From the formula. we notice that the error of the dimension located along axis Z depends on
the dimensions located along two other axes. i.e .• in contradiction to the linear DCs. the Abbe
principal breaks down in the spatial Des. The so called Abbe offset systems are well known
from the theory of accuracy machine tools. robots. coordinate measuring machines. etc. This is
the reason that the methods of calculations of the accuracy of machines (Reshetov and Portman.
1988) and robots (Veitschegger and Wu. 1986) were used for calculations of spatial Des
(Portman and Shuster. 1987; Weill and Shani. 1991).
ZO,ZI, Z2 Z3 iZI . Z2 ,Z3 ZO
P2
X2.X3
XO
B4
Figure 1 The spatial dimension chain between two perpendicular bores HI and H2: XO. YO.
ZO..... X3. Y3. Z3 are the axes of the coordinate systems So •...• S3 connected with the
boundaries of the links.
Modelling spatial dimensional chains 73
In this paper, an analytical and computational procedure for the computerized synthesis of a
theoretical model of any spatial DCs is proposed, making allowance for all factors affecting the
accuracy of the closing link. The mathematical formulation is based on the kinematic
interpretation of the dimensions (links of the DC): the linear dimensions are interpreted as the
linear displacement of the coordinate systems, and the angular dimensions are interpreted as
rotations of these systems. The errors connected with dimensions will be considered as errors
of linear and angular positions of corresponding systems. This approach enables us to apply the
in-depth study of the homogeneous coordinates of points and coordinate transformation
matrices of the fourth order for modeling spatial DCs. Applying the infinitesimal matrices for
presentation of the errors (Reshetov and Portman, 1988) is a typical feature of matrix models in
the field of accuracy evaluation and tolerancing.
To solve the synthesis problem we used the following initial data: the sequence of nominal
dimensions constituting the chain, as specified on the drawing; the machining routine,
indicating all basic and ancillary locating faces needed for machining the faces forming the
boundaries of the dimension chain links; and the errors of elements forming the locating faces,
obtained from the design and production planning documentation, requirements for an accuracy
of the chain to be analyzed. The synthesis procedure generates a system of accuracy
assessments relating to the closing link of the dimension chain. Thus avoiding typical errors
encountered when calculating dimension chains (omitting individual deviations, ascribing
incorrect signs to the angular errors, and incorrectly determining transfer ratios from errors of
the intermediate links to the error of the closing link).
2 MATHEMATICAL FORMULATION
2.1 Scheme of calculation procedure

The errors estimations are based on the following transformations: the input error in the DC are
the errors of position of all the surfaces relative to their nominal position and the errors of the
links (displacements). These errors are transformed into metrological estimation Ll of the
closing link by two transformations, as shown in Figure 2: a linear factor, "Nominal DC,"
a
converts the set of the input errors into the errors Llr of the real surface to be measured, and a
non-linear factor, "Metrological Models," converts Llr into the metrological estimations Ll of
accuracy (errors of size, form, orientation, and position of the closing link).
o ar a
------>Nominal DC ------->Metrological Models ------>
K; Q Ill, ... ,IlL
Figure 2 Flow chart for calculating the accuracy of the closing link.
2.2 Modelling the nominal chain

Graph of the chain and code system
In the first stage, a vector model of the nominal DC is constructed. In each link L; of the DC,
there are two boundaries. The final boundary of the one link L; is the initial boundary of the
next link L;+ 1. These boundaries may be points, lines, or surfaces. The "links-boundaries"
graph is represented in Figure 3, in which edges and vertices correspond to the DC links and
the DC boundaries, respectively. The boundaries are modeled by Cartesian right-hand
coordinate systems Si (i = 0, 1, ... , I, where I is the number of "moving" links of the DC)
connected with the ith surfaces (points, lines).
LI L2 L3
SO <-----------> S1 <------------> S2 <------------> S3
k/;q/ k2;q2 k3; q3
1
1 Closing link 1
1<-------------------------------------------------------------- >1
Figure 3 Graph of the spatial dimensional chain consisting of three links: So, S /, S2, and S3
are the coordinate systems of the link boundaries; ki and qi (i = 0, 1, 2, 3) are the coordinate
and parameter codes of the link Li.
Modelling the links is based on the following kinematic interpretation: the ith linear link of the
DC is modeled as a translation relative to one of the axes of coordinate system Si./ connected
with the (i-l)th (i.e., the previous) link; an angular link of the DC is modeled as a rotation
around one of the axes of the system Si.j. This interpretation is valid for most of the DCs since,
for any pair of neighboring links of the DC, the directions of their translations and the axes of
their rotations either coincide or are orthogonal to each other.
The intermediate link Li is modeled by the coordinate transformations Ai./,i converting a
radius vector of a point given in the coordinate system Si into a radius-vector given in the
system Si./. The transformations are formulated by link code ki and link parameter qi, where ki
=1,2, or 3 if the ith boundary of the DC is displaced on the distance qi along the Xi·j, Yi-j, or
Zi./ axes of the previous coordinate system Si.jlinked with the (i-l)th boundary; and ki=4, 5,
or 6 if the ith boundary of the DC is rotated by an angle qi around the Xi./,Yi./ of Zi./ axes,
respectively. Succession of these codes in the order given by the graph of the DC forms the
coordinate code K and the parameter code Q (the latter is the ordering list of the dimensions of
the DC):
K= k/ k 2 ... k,; (1)

Q =q/ q2'" q,. (2)
Special case: if the codes of all links are identical, i.e., k /=k 2=... =k,=k, we have the
(proper) linear (if k = 1,2 or 3) or angular (if k = 4,5 or 6) DC.
Ident(fication of DC
The procedure of an identification of the DC includes: fixing the boundaries of the links among
the geometrical subjects on the drawing and forming the coordinate systems Si connected with
these boundaries; compiling the codes (1) and (2) to the conditions that an initial link of the DC
is given in the form of the coordinate system So. For this purpose, the pairs of the numbers ki ,
qi, where ki is an integer and qi is a real number (i= 0,1, ... , I), are subsequently fixed for all the
links. The rules of identification of code k; of the link of the spatial DC, given in the standard
drawing form (three planes of projection), are shown in Table 1 (the direction of the link is
fixed relative to the previous link).
Table 1 Codes of the link

Type Plane of projection
of the link
Lj Frontal plane ZOY Horizontal plane XOY Profile plane XOZ
Vertical linear link kj= 3 =

kj 1 kj= 3
Horizontal linear link kj= 2 kj = 2 kj= 1
Angular link (0 <qj <1tI2) kj=4 kj= 6 kj= 5
Model a/nominal closing link

The pair consisting of the initial boundary So of the first link LJ and the final boundary S/ of the
last link L/ is the closing link of the DC. In other words, the system So separates the closing
link and the first link L J, and the S/ separates the closing link and the last link L/ of the DC. A
vector equation to be derived for the final boundary in the coordinate system of the initial
boundary will be considered as the mathematical model of the nominal closing link:
(3)
where rO = (xO. YO. zOo I)T and r/ = (x/. y/, z/. I)T are radius-vectors of the point (points) of
the closing link in the two coordinate systems. SO. linked to the initial (base) boundary of the
DC, and Sblinked to the last boundary; qj. q2 • ... , q/ are the parameters of the DC given by the
code Q; and Ao,/ is the coordinate transformation matrix (4x4), which is automatically
synthesized by multiplying I elementary one-parameter factor matrices A~~l,j (qj) in the order
given by code K according to formula (1). Each of the six values kj of the code is in one-to-one
correspondence with the matrix Ak of the elementary motion along the generalized coordinate
kj; for instance:
100 q] 1000] [cos q - sinq 00]

I 0100 A 2( ) = [ 0 1 0 q .... A 6( ) = sinq cos q 0 0 .
A (q) = [ 0010 ; q 0010 ' q 0 0 10
0001 0001 0 0 01
Therefore, the matrix of transformation is given by:
(4)
The radius vector r/ in formula (3) may have three diverse forms, depending on the geometric
nature of the last boundary. For the most important case when the last boundary is a surface
(5)
where u, v are the curvilinear coordinates of the surface.

2.3 Modelling the actual chain

Vector equation of the error
The actual DC includes both nominal dimensions and their errors. According to the kinematic
approach, the errors are interpreted as deviations of the position of the coordinate system
connected with the links of the DC. The positioning errors of the final boundary of the closing
link are automatically determined in the coordinate system of the initial boundary by summing
up the position errors of the DC's links. As a first approximation, the total error is determined
by the formula (Portman, 1981; Reshetov and Portman, 1988):
I
~ro = L
;=0
Ao,;c; A;,o ro (6)
where ro is the radius-vector of the point determined by Eq. (3); AO,; is the transformation
matrix for i first links of the DC; Ao,; is determined by formula (4) by changing from I to i;
and, Ai,O is the reciprocal matrix relative to Ao,;; namely,
AO,i = (Ai,o)-l = A O,] ... Ai-l,i' if i '# 0 (if i = 0, Ao,; =Ao,o is a unit 4 x 4 matrix);
e; is the infinitesimal matrix of the position errors for the ith link:
(7)
where Dxi' Dyi, DZi, ai, Bi, Yi are elementary errors of the position of the ith link; Dxi, Dyi ' DZi
are the linear position errors, i.e., small displacements of the origin of the coordinate system
Si, tied to the ith link along the Xi, Yi, Zi axes, respectively; and, ai, Bi, Yi are angular
position errors, i.e., the small angles of ith link rotation around axis Xi, Yi, Zj, respectively.
Scalar Balance of Errors

The elements of matrix (7) may be numbered in the following generalized form, which
corresponds to numbers of the generalized coordinates in the code (1):
(8)
Metrological estimations of accuracy are usually measured in the specified direction,

particularly in the direction of measurement. For this reason, it is necessary to obtain a
projection ~n of the error-vector ~o on the direction given by unit-vector n:
I 6
~n = n T ~ro = I I aji Dii ,
;=0 j=1
(9)
where Dii is the jth elementary position error (j = 1, ... , 6) of the ith link of DC according to
formula (6); aii is its transfer coefficient for the error Dii,
(10)
and, Dj (i=1, ... ,6) is a 4x4 projection matrix determined by partial differentiation of the matrix
by formula (7) with respect to the elementary errors to be defined by formula (8):
f.
Dj = dE/dOj (11)
i.e., D/ = dE/dO x' D2 = dE/dO y ' D3 = dE/dO z' D4 = dE/da., Ds = dE/d~, D6 = dE/dY.
Note, that aji are scalar values (or functions). Thus, the error of the closing links of the DC
given by codes K and Q is obtained in the form of the sum (9), comprising 6(l+ 1) elementary
errors of all intermediate links; i.e., in the form which is best suited to practical implementation.
Interpretation of elementary errors

The errors of size, form, orientations, and positions of the DC links and boundaries are the
natural geometrical sources of the elementary errors of the intermediate links. The simplest
interpretation has the error of size. It is the error Oji of the DC link Li corresponding to the
condition j = ki' i.e., the number j of this error must coincide with the code ki of the link. The
*
other five errors (with j ki) are as a rule the errors of the orientations and positions of the
DC boundaries. An example will be considered in Section 4.2.
In the case of presence of the errors of form, it is necessary to turn from the actual form of
the boundary to its nominal form (without distortion). To do this, one of the metrological base
surfacesllines must be constructed. The most important types of bases are the superimposed
surfaces/lines and metrological zone. The latter is constructed in accordance with the Taylor's
(or minimum-maximum material) principle. Sometimes the root-mean-square base surface/line
is also used. Six errors of position of the base surface relative to its nominal position are added
as the addenda to the elements of the corresponding matrices f.i in the sum (6), i.e., we take
into account the summarized deviation of the form and position.
3 METROLOGICAL MODELS OF ACCURACY

3.1 Main Definitions and Classification
Mathematical description of the estimations of the accuracy parameters of the parts - in
particular, the accuracy of form and attitude in 3D-space - is given in terms of the errors of the
closing link of the DC. Metrological characteristics of the closing link are determined as the
linear and angular positioning errors of the final boundary of the closing link relative to its
initial boundary. These characteristics are scalar quantities, which are obtained using
metrological models of accuracy (MMA). The MMA is the formula, algorithm, or procedure for
determination of accuracy characteristics of the closing link of the DC. The MMA is calculated -
if the pair of boundaries (ro,o and rO,l) for the closing link and their positioning errors Liro,o,
and LirO,I, are known - in the single coordinate system according to formulae (3) and (6). For
the widespread case, when the vectors ro,o, rO,l, Liro,o, and Liro,! are functions of u, v or s [see
Eg. (3)], MMA is the functional 11, which brings a real number or a set of the numbers into
correspondence with four vector functions:
11 p = IIp (ro,o, ro./, Liro,o, LirO,/) , (2)

where p is the nominal parameter of size, form, or position.

The MMA depends on a nominal form of the boundaries of the DC's closing link and their
relative positions. Symbolically, a three-digit code is allocated to the closing link,
m= m] m2m3,
where digits m] and m2 express the geometric nature of the initial and closing boundaries of the
closing link, and have values from 1 to 8 as follows: 1 - point; 2 - straight line; 3 - circle; 4 -
other plane line; 5 - plane; 6 - cylinder; 7 - sphere; 8 - other surface [this classification is similar
to the TTRSs' (Technologically and Topologically Related Surfaces) system proposed by
Clement (Charles, 1989)]; digit m3 characterizes the nominal relative positions of the objects:
m3 = 0 if the distance between them is nominally zero and m3 = 1 if not.
With the exception of the MMAs, in which one of the boundaries is a point (m]=l andlor
m2=1), each of the MMAs corresponds to several linear and angular standard estimations of
accuracy; i.e., MMA may be used both directly as the assessment itself, or for calculating
accuracy assessments of size, form, and attitude required by the standards. Sets of the MMA
according to formula (12) composes the library of the MMAs.
3.2 Examples of MMAs

Error of distance from a point to a straight line
The nominal equations and the errors are specified in the form:
the point: rO,1 =r]; tlrO.I =tlr] ; (13 )

the line: ro,o = res) = r2 + sb; Aro,o = Ar(s), (14)
where r] is the radius-vector of the point; res) is the radius-vector of the line; r2 is the radius-
vector of a point belonging to the straight line; b is the unit-vector of the direction of the straight
line (lbl = 1); s is a current parameter of the straight line; Arl is the error-vector of position of
the point; and Ar(s) is the deviation of the coordinate of the points of the line from nominal,
specified as a function of parameter s. For the case, when the form of the straight line is not
distorted, ar(s)=E res), where E is a matrix of errors of the position according to formula (7).
The error M of distance d from a point to a straight lines is specified by the formulae:
!J.d = I ad x b I, with d = 0 (m=120) ; (15)

!J.d = (ad x b) (d x b)/1 d x b I, with d ~ 0 (m=121) , (16)
where d = r] - r2; ad = Ar] - Ar2(SO); and, so is the value of the parameter corresponding to
the base of a perpendicular dropped from point (13) to the line (14); i.e., so = d'b; m is the
code of MMA. Thus, formulae (15) and (16) bring the metrological assessment of error !J.d of
distance d into correspondence with four vectors, specified by formulae (13) and (14); i.e., in
the form (12): !J.d =!J.d (fl, res), Arl, Ar(s». The MMAs for the distance from a point to a
point, a line, and a surface are listed in Table 2 (the equation of the first boundary is
r=r] +ar]).
Errors of angular parameters

MMAs for the relative angular positioning of lines andlor surfaces are expressed in terms of the
unit vector a I and a2 of the corresponding directions and errors al and fl.2 of these unit vectors:
O<P=lal a2 + a2 A/i/ial x a2 I, where a/xa2~O. The most important cases are deviations from
parallelism O<ppar and from perpendicularity O<pperp, determined from the equations:
OCPpar = I( IlJ - 112 ) x aJI ; =

aJ a2 = a (17)
OCPperp= laJ 112+ a2 IlJI ; aJ' a2 = 0 (18)
Relationships linking the angle positions of straight lines, planes, and cylindrical surfaces
with oCPpar and oCPperp determined by Eqs (17) and (18) are given in Table 3.
Table 2 Errors of the distances from point to point, line, plane, and surface
Distance from Model Equation of

point to mlm2m 3 the second boundary
point 110 r=r2+llr2 IIldl11

111 d·lld 11/11ld III
straight line 120 r( s )=r2+sb+flr2(S) Illd12Xbl
121 (fld12Xb)(dxb)/ld x bl
plane 150 r(u, v)=r2+ub+vc+llr2(U, v) Ild IS'n
151 IldlS'n - (d'n)(lln'n)
surface 180 rs=r2 (us, VS)+llr2(US, vs) [llrJ - Ilr2(UO,vO)]'ns
181 (f·g·h) Ilf·g·hl
Notations:
• Ibl=l; Ibxcl=l; n=bxc; d = rt - r2;
• Ildll = Ilrl - Ilr2; Ild12 = Ilrt - Ilr2(SO), where so = d·b;
• Ild IS = Ilrt - Ilr2(UO, vo),where uo=(b'd)'( c'c)-(b'c)( c'd); vo=(b'b ).( c·d)-(b·d)(b·c);
• Iln = bxllr v - cxllru, where Ilru =dllr2(uo,vo)/du, Ilrv =dllr2(uo,vo)/dv
• g=ru+dNu IINI; h=rv+dNvIINI; and f=llr J-llr2-(d I INI)(r2uxllr2v+llr2vxr2u).
Table 3 Deviations from parallelism and perpendicularity for lines, planes, and cylinders
The boundaries III 112 OCPperp
Two straight lines (m]m2=22) Ilb] Eq. (19) Eq. (18)

Two planes (mlm2 = 55) Ilnl Eq. (19) Eq. (18)
Straight line and plane (mlm2 = 25) Ilbl Eq. (20) Eq. (17)
Two cylinders (mlm2 = 66) Ilel Eq. (19) Eq. (18)
Cylinder and plane (mlm2 = 65) Ilel Eq. (20) Eq. (17)
Notions: Ilb, Iln, and Ile are the deviations of the unit vectors of the straight line direction,
the normal direction, and the direction of the cylinder axis, respectively.
Error of a distance between two straight lines

For this case, both the boundaries are the straight lines. There are four types of their relative
nominal position: coincidence, parallelism, intersection, and skew. The relative position of
these lines may be estimated by means of the following errors:
• Deviation of the shortest distance between the lines; i.e., deviation of the distance on a
common normal to the lines from the nominal value.
• Deviation of the distance between the lines in a section of the specified plane; i.e., deviation
from the nominal value of the distance between the points of intersection of the
lines with the plane.
• Deviation of the distance between the lines over a specified length in a given direction;
i.e., the maximum value of the estimation described in paragraph 2 on a given set of
parallel planes.
• Constancy of the distance between nominally parallel lines over a given length; i.e., the
maximum value of the assessment described in paragraph 2 on a given set of planes
perpendicular to the nominal direction of the lines.
4 APPLICATIONS TO DESIGN PROBLEMS
4.1 Formal procedures
Evaluation of closing tolerance

Formula (9) gives the error of the closing link of spatial DCs as the sum of the number of the
elementary errors. From the point of view of the following computations, this representation of
the error does not differ in the least from that in the traditional linear DC. Therefore, the well-
known methods of the mathematical evaluation of the closing tolerance are valid forthe spatial
DCs, too. The case in point is calculations of the closing tolerance by means of an arithmetical
or statistical summing up. In the latter case, the tolerance is calculated on the assumption that
the individual actual sizes (links of the DC) have normal or other distributions.
On the other hand, the spatial DC essentially differs from the linear DC: (1) as a consequence
of the Abbe's offsets in the spatial DC, there is an interrelation between angular errors of the
links and nominal values of the linear links; (2) as a consequence of diversity of directions of
the spatial DC links, choosing a compensating link among all the links is quite a complicated
problem.
Choosing a compensating link

As is known, a compensating link is used in DC for providing the required accuracy of the
closing link when the accuracy of the intermediate links is inadequate for this purpose. In terms
of Eq. (9), the nominal size c of the compensating link is calculated from the condition c+D.n==O.
Therefore,
I 6
C == - D. n == - LL
i~O j~1
aj; Dj; , (19)
Physically, the compensating link is realized as an additional controlled displacement of one

or more of the intermediate links; i.e., c == L ac Dc, where Dc is the compensating displacement
of the compensating link; and, ac is the transfer factor according to formula (10) for the values i
and j corresponding to the link to be chosen. From formula (19), it follows that the
displacement of the ith link in the jth direction must be chosen as a compensating link if, and
only if,
(20)
Furthermore, the link of choice for compensation corresponds to the conditionj == k;, i.e., the
code k; of the nominal motion of the ith link, by which the compensating motion is realized,
must coincide with the number} of the error to be compensated (these numbers are defined by
formula (8); both k; and) can take the integer values between 1 and 6).
Optimization of the DC with respect to accuracy criteria

The specific features of spatial DCs enable us to transform the order of the links without
changing the nominal closing link. These features were analyzed in the theory of form-shaping
systems (Reshetov and Portman, 1988). These allowed transformations are:
Transformation 1. Rearrangement of two adjacent linear links of the DC. These
transformations correspond to the following rearrangements in the coordinate code K according
to formula (1): 12 = 21; 13 =31; 23=32.
Transformation 2. Rearrangement of two adjacent links of the DC, one of which is linear
and the other angular, if the two links have the same axis. These transformations correspond to
the following rearrangements in the coordinate code K: 14 = 41; 25 =52; 36=63.
Transformation 3. Unification of the several adjacent (with the accuracy of the allowed
transformations 1 and 2) links, if their codes coincide. The transformations correspond to the
following changes in the coordinate code K: II = 1; 22 =2; 33=3,44=4, 55=5, 66=6.
Transformation 4. Partitioning a link into several links with the same code (the formulae are
inverse to those for Transformation 3).
All these formal transformations describe the widespread actual conversions in relation with
the drawings: different versions of the dimensioning which are equivalent from the point of
view of the closing link; changing an order of the measuring or machining the part; etc. But
with equivalence from the point of view of the nominal DC, the transformations result in the
diverse accuracy of the actual closing link. The differences result from the diverse Abbe's
offsets in variants of the nominal equivalent DCs. The procedure of an optimization of accuracy
of the closing link is based on this fact. This procedure includes the following stages: (I)
constructing any first variant of the DC according to code K; (2) resetting all the variants of the
codes using the allowed Transformations 1-4; (3) calculating the errors for all the variants; (4)
an optimization according to criterion of minimum of the error of the closing link.
4.2 Example of calculating a spatial DC
Preliminary analysis
The DC between bores HI and H2 shown in Figure 1 will be analyzed. Its links LI, L2, and L3
as well as their boundaries (the holes HI and H2 and the planes PI and P2) are shown in Figure
4. The mathematical models of the DC are given in Table 4: the codes of the links, the codes of
the whole DC according to formulae (1) and (2), and the systems 5; connected with the
boundaries (naturally, it is one of the possible variants of the DC).
Since the boundaries of the closing link are the holes HI and H2, the set of the metrological
models connected with this DC characterize the relative position of the holes: tolerance zone of
perpendicularity of the axis of the hole HI with reference to the axis of the hole H2; deviation of
the distance between the axes, etc.
The initial boundary of the initial link is the cylindrical surface of hole HI. An equation of the
boundary in its coordinate system 50 has the form:
ro,o = (R, cos e, YO, R, sin e, I)T, (21)
where Ro is the radius of cylinder HI; and, e and YO are the polar angle and the axis
coordinate, respectively, of a point on cylinder HI.
The last boundary of the last link is the cylindrical surface H2. Its equation in the coordinate
system 54 has the form r3 = (X3, R2 cos <1>, R2 sin <1>, I)T, where <1> and X3 are polar angle
and axis coordinate, respectively, of a point on the cylinder. The equation of the closing link in
the coordinate system So according to formula (3) is rO,3 = A I( -B4) A3(B3-B 1) A2(B2) r3' and
after formal transformations, we have:
Table 4 The DC between bores HI and H2 in Figure 1
Link Boundary Description of the links and boundaries ki qi

Hole HI Coordinate system So (the Xo, Yo, and Zo axes)
LI Translation along the Xo axis at the length B4 -B4
Plane PI Coordinate system SI (the XI, YI, and ZI axes)
L2 Translation along the Z I axis at the length B3 - B I 3 B3 - Bl
Plane P2 Coordinate system S2 (the X2, Y2, and Z2 axes)
L3 Translation along the Y2 axis at the length B2 2 B2
HoleH2 Coordinate system S3 (the X3, Y3, and Z3 axes)
Codes of the DC: K={132}

Q = {- B4, (B3-Bl), B2}
Figure 4 Model of DC between bores HI and H2 according to Figure 1.

rO,3 = (X3 - B4, R2 cos <I> + B2, R2 sin <I> + B3 - BI, I)T, (22)
By this means both the initial and the closing boundaries of the nominal closing link are
represented in the single coordinate system So by the vectors ro,o and rO,3, respectively,
Evaluating the errors

The vector ArO,3 of the error of the vector rO,3 is calculated according to Eq, (6) taking into
account Eq, (22):
ArO,3 = [Eo + A 1(-B4) 101 A I(B4)+ A 1(-B4) A3(B3-B 1) 102 A3(-B3+B 1) AI(B4) +
+ AI(-B4) A3(B3-BI) A2(B2) 103 A2(-B2) A3(-B3+BI) A I(B4)] rO,3 ' (23)
where 10; (i = 0,1 ,2,3) are the error matrices with the structure given by formula (7); EO is the
error of the position of the base system SO; and, 10 I, 102, and 103 are the matrices of the errors of
the links of the DC, i.e" the errors of the boundaries S I, Sband S3 with reference to the
boundaries So' SI' and Sb respectively,
As an example, let us consider a possible physical meaning of the elements of the matrix 10 1,
which is the matrix of the error of the position of plane PI with reference to axis HI: 0XI is the
error of the size B4; 0yI is the error depending on the form of the plane (flatness tolerance
zone); 0zl is the deviations (with sign minus) from true position of the axis of the hole relative
to the plane D (see Figure 1); ai and "(1 are deviations (with sign minus) depending on
tolerance zone perpendicularity of the axis with reference to the plane PI in two specified
directions (relative to the X and Z axes, respectively); PI = 0 for this case.
The expression in the square brackets in formula (23) is the matrix EL of the generalized error
of the position of the closing boundary (i.e., the hole H2) relative to coordinate system So. The
matrix EL is also the infinitesimal matrix of the type in formula (7). After formal
transformations, we obtain 6 elements 0x1:, 0y1:, 0z1:, a 1:, P1:, and "(1: of this matrix in terms of
i\;, 0y;, 0z;' a;, Pi, and, "(; (i=0,I,2,3):
3
0x1:= L 0xi+(BI-B3)(P2+P3)+B2Y3;
;=0
3
0)'1: = L, Oy; - (B I-B3) (a2 + (3) + B4 ("(1 +"(2 +"(3) ;
;=0
L L
3 3 3 3
0z1:= L 0z; -B2arB4(P/+P2+P3);a1:=L ai; P1:= Pi; "(1:= "(i' (24)
i=O i=O 1=0
Consider, as an example, a deviation of the shortest distance MIM2 (see Figure 4) between
the axes of holes HI and H2. The equation of the axis ofthe hole H2 is obtained from Eq. (22),
ifR2 = 0, i.e., r ax ,3 = (X3-B4, B2, B3-BI, I)T. Then, X3 = B4 for the point M2 and direction
of the line MIM2 is given by the unit vector of the axis Z, i.e., n = (0, 0, 1, O)T. The deviation
is obtained by means of formula (9) taking into account formulae (22) and (23):
Ad = n'(EL r ax,3) = B2(aO + al + (2) - B4 (P 1+ P2 + P3) +L, 0z;' (25)

i=O
Compensating and optimization procedures

Formulae (24) and (25) enable us to choose the compensating links according to condition (20).
For example, compensating the angular error 'YJ; of the closing link may be carried out only by
means of angular displacements 'Yi of the intermediate links. Compensating the linear error I'1d
of the closing link according to formula (25) may be carried out by means both of linear
displacements aZi and/or angular displacements ao, aI, a2, ~ I, ~2> and ~ 3'
The optimization procedure may be carried out as follows. The allowed transformations of the
code of the DC in Table 4 result, for example, in additional variants of the code of the DC: K2 =
( I 2 3) and K 3 = (3 I 2). As a result of diverse order of the links in the DC, the error ,1d are
changed; namely:
3
(,1d}K2=B2(ao+al)-B4(~I+~2+~3) +L, azi, for thecodeK2;

i=O
L azi, for the code K3·

J
(,1d}K3 = B2(ao + al + a2) - B4 (~2 + ~3) +
i=O
The corresponding differences between these errors and that according to formula (25) are as
follows: ,1d -(,1dJK2 = B2 a2; ,1d -(,1d}K3 = - B4 ~ I; and(,1dJK2 -(,1d}K3 = - B2 a2 - B4 ~ I·
5 CONCLUSIONS
I. Modelling the spatial DC is an effective method for an investigation of the accuracy and a
calculation of the metrological assessments within the computer aided tolerancing problems.
2. The kinematic interpretation of the spatial DC enable us to use the well known mathematical
formulation of homogeneous coordinates and transformation matrix of the fourth order for the
calculation of the accuracy of the DC's closing link. The mathematical procedure includes
modeling the nominal DC, the actual DC, and the metrological estimations of accuracy.
3. In metrological terms, the main distinction between the spatial DC and the linear DC is the
presence of Abbe's errors in the former. As a result, there is an interrelationship between the
nominal position of the intermediate link within the DC and the transfer ratios of the positional
errors of the link.
4. The procedures for evaluating the tolerances, choosing possible compensating links, and
structure optimization are developed for the spatial DCs. As a real world example, the DC
between two bores with perpendicular axes is analyzed. The proposed methods were used for
synthesizing models of complex spatial dimension chains for process planning in FMS.
6 ACKNOWLEDGMENT
The authors are grateful to the Israel Ministry of Science and Technology (Grant No.
85397301) and Paul Ivanier Center for Robotics Research for partial support of this work.
7 REFERENCES
Bjorke, O. (1989) Computer aided tolerancing, 2nd Edition, ASME Press, New York.
Charles, B., Clement, A., et al. (1989), "Toward a computer aided functional tolerancing
Model", Int. Conf. on CAD/CAM and AMT, CIRP. Sessions on Tolerancing for function in
a CAD/CAM Environment, Proceedings, Vol. 2, Jerusalem.
Dong, Z. (1993) Design for automated manufacturing, in Concurrent Engineering:

Automation, Tools, and Techniques (ed. A. Kusiak), John Willey & Sons, Inc., New
York.
Dong, Z. and Soom, A. (1991) Some Applications of Artificial Intelligence Techniques to
Automatic Tolerances and Synthesis, in Artificial Intelligence in Design (ed. D.T. Pham),
Springer-Verlag, London, 10 1-124.
Fainguelernt, D., Weill R. and Bourdet, P. (1986) Computer aided tolerancing and
dimensioning in process planing. Annals of the CIRP, 3511, 381-6.
Irani, S. A.., Mittal, R. 0., Lehtihet, E.A. (1989) Tolerance chart optimization, International
Journal of Production Research, 27(9), 1531-52.
Jayaraman, R., Srinivasan, V. (1989) Geometric tolerances: Virtual boundary and
requirements, IBM Journal of Research and Development, 33, 90-104.
Kirshling, G. (1991) Quality assurance and tolerances. Springer-Verlag, Berlin.
Portman, V. T. (1981) Uni versal method for calculating the accuracy of mechanical devices,
Soviet Engineering Research, 1(7), 11-15.
Portman, V.T. and Shuster, V.G. (1987) Computerized synthesis of a theoretical model of a
spatial dimension chain, Soviet Engineering Research, 7(8), 57-60.
Reshetov, D.N. and Portman, V.T. (1988) Accuracy of machine tools, ASME-Press, New
York.
Shen, Y.L. and Duffle, N.A. (1991) Uncertainties in the acquisition and utilization of
coordinate frames in manufacturing systems, Annals of the CIRP, 40/1, 527-30.
Turner, J. and Wozny, M. J. (1990) The M-Space theory of tolerances, Proc. of the ASME
Advances in Design Automation Conference, Chicago, 217-25.
Veitschegger W.K. and Chi-Haur Wu. (1986) Robot accuracy analysis based on kinematics,
IEEE Journal of Robots and Automation, RA-2, No.3, 171-79.
Weill, R. (1988), in Collaboration with: Clement, A., Hocken, R., Farmer, L.E., Gladman,
C.A., Wirtz, A., Bourdet, P., Freckleton, J.E., Kunzman, H., Ham, I., Trumpold, H.,
Mattias, E., Tolerancing for function. Annals of the CIRP, 37/2, 1-8.
Weill, R., Shani B. (1991) Assessment of accuracy of robots in relation with geometrical
tolerances in robot links, Annals of the CIRP, 4011,395-9.
Yakushev, A.I., Vorontsov, L.N. and Fedotov, N.M. (1986) Interchangeability,
standardization, and technical measurements, Machinostroenie, Moscow (in Russian).
7 BIOGRAPHIES
Professor Vladimir T. PORTMAN was born in 1935 in Russia. He graduated from the
Moscow Machine Tool and Tooling Institute in 1958 and received his doctorates (PhD) in 1970
and (DSc) in 1988. He was Chief of the Department of Computer Aided Calculation of Machine
Tools and FMS in the Experimental Research Institute for Metal Cutting Machine Tools in
Moscow (ENIMS). From 1995, he serves as Associate Professor in the Department of
Mechanical Engineering, Ben-Gurion University of the Negev, Beer Sheva, Israel. He
published 88 papers, books, and patents concerned with machine tool accuracy, reliability,
precision machining, dimensional metrology, and FMS simulation.
Professor Roland D. WEILL was born in 1923 in France. He graduated from the Ecole
Poly technique and the Ecole Nationale Superieure de I'Armement in Paris. He served as the
Head of the Laboratory for Production Engineering of the French Armement Authority. He was
nominated Professor of Production Engineering at the Israel Institute of Technology in Haifa in
1976. He is the author of over 90 articles, books, and patents in the field of manufacturing
technology, robotics, and tolerancing analysis
PART TWO
Tolerance Modelling
6
Toleranced feature modeling by constraint
of degree of freedom for assignment of
tolerance
T MAEDA and N TOKUOKA

Department of Mechanical Engineering, KEIO University
3-14-1 Hiyoshi, kohoku-ku, Yokohama, Japan,
Tel.:045-563-1141,Fax.:045-563-5943,e-mail:toshi@mech.keio.acjp
Abstract
The purpose of this study is the faithful expression oftoleranced feature in machines or drawings
following the current tolerancing, and the proposition ofthe feature model that is able to evaluate
the effect of tolerances against the designer's demand such as performance, costs and so on. We
examined the tolerance zone ISO recommend and classified the relationship between tolerance
and feature. The tolerance zone is the set of allowable feature and the behavior of feature is
represented by using matrices and constraint equations that limit the degree of freedom (DOF) of
feature. And the relationship of features and tolerances were classified by the feature, datum, size
and tolerance information as unit. Every behavior oftoleranced feature is expressed by the
combination of them. Furthermore, in addition to these DOF constraint model, we proposed the
functional expression of each manufacturing lot for the current geometrical feature model. The
manufacturing group of it consists of the material, manufacturing and functional shape to
connect the manufacturing information. We also considered the assignment of tolerances that
satisfies the target cost or performance using the manufacturing cost.
Keywords
Feature modeling, degree of freedom, vector tolerancing, assignment
1 INTRODUCTION
It is suggested that one important subject for current CAD system is the problem of handling the
accuracy in the design process. Because tolerance data such as the types, values, relational
90 Part Two Tolerance modelling
features, etc., are only handled as symbolic data in the drawing, it deals only with operations
related to geometric forms, omitting operations related to accuracy information such as
tolerances and roughness. By using dimensional and geometrical tolerances to express the actual
feature data, it is possible to evaluate and analyze the accuracy of parts. It is necessary for aiding
the accuracy design to express the dimensional and geometrical tolerance zones, to establish the
calculation sequence of cumulative tolerance zone and the handling method of forms,
dimensions and tolerance data.
We tried to express the behavior of toleranced feature in parts or machines or drawings for the
faithful expression following the current tolerancing. And we considered the feature model which
is able to evaluate the affiliation of tolerances against the designer's demand such as
performance, costs and so on. This paper mainly consists of three parts; (1) expression of
toleranced feature, (2) proposition offeature modeling to combine with manufacturing
information and (3) assignment of tolerance using considerable performance and costs.
2 CONCEPT OF THIS RESEARCH
2.1 Concept of Tolerance zone and behavior of feature
Tolerance zones should be expressed as not to conflict with the definition ofISO considering the
behavior offeatures on the actual products. Figure 1 shows the concept of the behavior of
"
feature in the tolerance on the manufactured products. The actual zone is the area including the
61
feature of the products and defined as
some width with the perfect form as the
center. The position, orientation and
,, ,
width of this area depends upon the
"
. '
, Actual Zone
products, manufacturer, technology etc .. Tolerance Zone
It is necessary to consider behavior of
the perfect form and actual zone on the
L~ ~
Actual Fonn
investigation of cumulative tolerances,
but for the expression of tolerance zone, Perfect Featw"e
the behavior of the perfect form is
enough. Therefore, the actual zone is
assumed to be zero and the allowed
Figure 1 Concept of tolerance zone.
tolerance zone is expressed by the
behavior of the perfect form.
2.2 Relationship between feature and dimension
Bernstein (1989) added the degree of freedom to features such as faces, edges and vertices in the
B-reps model and represented the tolerance zone by using 4x4 matrices as the constraint
equation against the nominal position on a single part. Martinsen (1993) considered the
expression of orientation and location tolerances for all types of surfaces using vectorial
tolerancing. However, they don't correspond to the various patterns between feature and
Toleranced feature modelling 91
tolerances following current tolerancing,

because they deal only nominal features such as
face, edge and so on.
Figure 2 shows the expression of plane feature
and the behavior of feature in tolerance zone
when the straightness tolerance is specified to
the feature. In case that the dimension is
theoretically exact one and the feature has
nominal shape, the tolerance zone means the
rectangular shape has width t on the nominal
plane as shown in Figure 2 (a). The fact is, even
though it is assumed that the dimensional
tolerance or general·tolerance or general
tolerance is zero, the zone never be defined the
fixed shape without treating the corresponding
feature as the datum feature as shown in Figure
2 (b). Figure 2 Tolerance zone of plane feature.
As dimensional tolerance only means the
distance of two points, it doesn't limit the shape ofthe corresponding feature. Hence, in case that
one shape offeature can't be defined, other shape never define the allowable zone. As the
definition of shape, several methods are suggested; (1) using the datum feature as CSA B78.2,
(2) assuming the ideal feature, (3) limiting the feature by general tolerance, etc .. In this paper, we
assume that general feature indicates all features implicitly and further dimensional tolerance
limit the relative relationship among features such as distance and angle.
3 EXPRESSION OF TOLERANCES
3.1 Expression of Information by Degree of Freedom
Four types ofinforrnation is defined to

express the behavior of toler anced feature.
And Each information is expressed by limits
of its own degree of freedom (DOF). Figure ,
,
3 shows the relation among them. The left
1"
figure means the parts that has one
dimension and one geometrical tolerance as
the example and the right shows the relation
between two features.
_______ Cl
Expression of dimension and
Omstrwnl ofooF
feature's location and
orientation
In order to express the behavior of feature Figure 3 Expression of toleranced feature.
using DOF of every feature and to consider
cumulation of tolerances, it is necessary to express the location and orientation of every feature.
The location and orientation of a local coordinate against the absolute one are expressed by DOF
and named Ce .
Dimension constrains a relative location oftwo features and its direction is same as that written
on a drawing. Therefore, the DOF of dimension named C, is as the constraint of relative location
between features. the feature that never have the constraint float against other features.
Expression offeature
It is necessary to combine a feature and a length
and width of tolerance in order to express the
behavior of the actual feature in the tolerance
zone. In this case, the feature is given 6 DOF
named Cf and is expressed by restricting it.
In the same feature, the shapes of tolerance
zone change by the types of specified tolerances.
Figure 4 shows the example of tolerance zone
for the circular feature by types of specified
toleranced.
To express the various shape, all features are (a) Circularity (b) Circular runout (c) Straightness
(for Axis dir.) (for Axis)
defined by limiting own DOF as movable and
rotational range.
Figure 5 shows the example of features Figure 4 Tolerance zone of circular feature.
represented by the constraint ofDOF. For
example, vertex feature restrict as all DOF is zero and edge is expressed by limiting one DOF of
movement. Furthremore, several features such as face, cylinder, etc. are expressed by the set of
fundamental feature hierarchically.
For example, surface feature f, and f2 are constructed by limiting DOF of their fundamental
feature such as vertex and edge, and in the
same feature the condition of constraint
changes by the types and the limiting
direction of tolerances. Concretely, f, is
applicable for the flatness for the surface
feature and f2 is for the straightness as the
line in the surface feature.
And complex features such as cylinder are (a) plane
treated as the combination of multiple
fundamental features and dimensions,
because in the drawings and current
tolerancing there are various limits of
tolerances by the type of toleranced feature
and the dimensional tolerance. Figure 5 (b)
shows the example expression of cylindrical
surface feature. We assume that dimension
such as the diameter of cylinder is handled as Figure 5 Definition of feature.
the distance between the center line and the
vertex or edge that represent the cylindrical surface in the drawing. Here, each Cf means the
corresponding feature such as the center line / vertex or the element of surface feature, and both
features are connected by the dimension Cs.
Expression of tolerance
The direction of feature, of tolerance and its (I)
value are necessary to define the movable
area of the feature. And the constraint of
DOF named Ct for features is classified into 7 (2)
types. Table I shows the relations of
toleranced feature and the shape of tolerance
zone. The symbol A to G shows the result of (3 )
this classification by the difference of
constraints. Where, (D) means the constraint
(4 )
of the straight line feature in optional plane,
and restricts the behavior of the feature by the
limit of one movement and rotation in zone. (5)
The actual allowable zone of toleranced
feature is defined by the combination among
their information Cf , C, and Ct as shown in (6 )
Figure 6. This figure shows the example of
types of constraints and zones in case of Figure 6 Expression of tolerance zone of
toleranced cylindrical feature. cylindrical feature.
Table 1 Relations between features and tolerances
__ sjl]gl~___________________ ____________________ _
~_ !l
VERTEX soace • C
PLANE --~~:~-- ----- ------I~~£~ ---- --------- --- ----

__ _____________________ _____________ _
----- ______ :'__ A
r~~j~~ ~_ ~
CIRCLE l~;:;,r:.~( .0-13 -------- - ' __A____ _
CYL INDR ICAL - _t..a_djij - ------ -- -- - <? ~ !I- ----- II_~ ~- - - --

__ ~I!l!~g _ - D _ _ __ __ _ _ ___ ________
___ _ _
SURFACE axial ~ i D .~.l 2D ~! D -!,(.i. E
_JadiJ.Is__ ___ __ _ __ ___ ____ __ ____ ~~ ~ _____ _
CON ICAL __C.!l!til]g __ .:-___~ ________ _ __ ________
_ _ _ __
, A
SURFACE -
- _n~!;;fL ~-£ -Ii -lZr£ri n- -n ---- ----"':l!l-i
- ~~*~~~ -~~~ra-~~~~ ----- ----- ----- ---- --- --------- ~-I!(\ = nA
3.2 Representation by Using Matrix
To express the behavior of DOF the matrix M is defined as following,
(1 )
where M a , Mfl and Mr are the matrices for movement and M 0, M;, M If' for rotation, Furthermore,
elements of matrices are defined by the constraint equations, They are classified into two kinds,
One is to restrict each element in matrices and the other to control constraint equations, An
example of the former is Eq,(2) and of the latter Eq,(4),
A = {xl-OJ 5 x 5 OJ}, :Jt EA (2)

{tl-OJ 5 t 5 OJ} (3)
Equation (2) is rewrited as Eq,(3), For example, it defines values of elements in matrices and
means that t can take any values from -0,1 to 0,1.
(4)
Equation (4) controls the relation between tJ and t2 and means that the sum oftJ and t2 should be
equal and less than 0,1, Every tolerance zone is expressed by the combination of matrices and
constraints and indicated by T.
ee ~
3.3 Classification of Tolerances by Connecting Conditions
geom:tncalc:lim=ional
The relations between feature(s) and ~ tolerance tolerance
tolerance( s) can be classified into 4 tol tol
categories as a unit, namely, (1) feature
specified geometrical tolerance, (2) feature
related to datum, (3) dimension specified ~8
dimensional tolerance, (4) tolerance related geOlretrical
tolerance
tol
dimmsional
tolerance
to another tolerance, Those categories are (a)Tg (b)Td (c) Ts (d) Tr
shown in Figure 7, The relations among
feature, datum, size, tolerance in parts can be
Figure 7 Four categories of relations,
expressed by those 4 categories,
Tg: feature - geometrical tolerance

All geometrical tolerances applied to a single feature belongs to this category, Their tolerance
zone can be managed independently, In case that the general geometrical tolerance is introduced
to a single feature, the relation belonging to this category is applied to them,
Figure 8 shows the expression of a tolerance zone oflinearity as an example of this category,
The feature, "line", is specified tolerance t. Here, xl, Yl and Zl are local coordinates of nominal
feature, The actual feature should exist within -tl2 to tl2 for Yl direction, The feature can moves
for 13 and rotate in \If within tolerance

zone. The limit of \If becomes the IMATRIXI
function of 13. 13, \If and the relation of M=M, . M~
them are defined as the constraint IC<>l'.STRAINll

equations. In case that a single feature is L =(length oh, -dir.)
specified by plural tolerances, each Tg is (.tt-l f2SPS lf2)
defined independently.
(1I'1II1'lS tan-' (/ -~f L)l
Td: feature - datum
In case that the geometrical tolerance is
related to datum, the geometrical Figure 8 Representation of feature's behavior by Tg.
.r'
property such as pO$ition, orientation
and direction, of nominal feature is [MATRiXI
affected by datum. So it is necessary to •..•... 81 ' "
M = M •..11, · M~
fit the local coordinate first. This relation ...... •.•••• v~ '91bz,
A .' ,
[CONSTRAiNTI
is named Td. The tolerance zone is .;., ¢I , ,Ccl····· .. ~ ,
specified by the behavior of the nominal
....... C1 E) ..1'" ",, ' Ce2. {pip = Jrl l }
". ~ ...• u ' {(;, , ;,~j; , - ;,I= p}
feature (by Tg). Therefore, tolerance featurel •.•.••..~.... .4>2.,'
zone of the feature related to datum is . p ' , 'fea1ure2
{Oil}: 19, - 0,1}
expressed by Tg and Td. Figure 9 shows [datum] {IIIIIII =!Iv, - III,I}
the case of perpendicularity of center
line related to datum. The deference of Figure 9 Representation offeature's behavior ofTd.
local coordinates Cel and Ce2 are
corrected by constraints. The constraints
indicate rectangular given by two rotational component e and \If .
Ts: dimension - dimensional tolerance

The dimension is defined as the ~ (MAllUX)
distance between two points on H BJ M = M, ·M ,
features and the dimensional I [CONSTRA.lN11
tolerance limits the relative L, = (length offeature yl-dir.)
I
distance and angle between them. L, = (length offeature y2- <tir.)
Therefore, the tolerance zone is not
~,~J ~ +~=uPS~SUPS~S~
defined except for being defined
{P,rt,I2SP, St,(2)
the shape of one feature. It is
assumed that the dimension is (O,I.O,1S tan"{t,- 2l,P,11 L,))
measured in the direction specified {p,I-t,/HP, S t,l2)
in the drawing and each feature is featurel (o,I:O,l s larl'(~ - 2l,P,11L,)}
related to the general geometrical
tolerance implicitly. This condition Figure 10 Representation of feature's behavior ofTs.
is also expressed by Tg. The
dimensional tolerance is expressed by two T related to two features as shown in Fig.l o. Both
features have tolerance tl, t2 respectively and they should satisfied the relation 2t = tl + h.
Tr: tolerance - tolerance

The relations between plural features lCONSTRAINTI

specified by plural tolerances are geometrical dimensional
tolerance tolerance I , = (tolerance of Tg)
classified into 2 cases: (I) one feature I , = (tolerance of Ts)
specified by plural tolerances, (2) one
{I, 'II" S; I, 's; I,. }
geometrical tolerance gives the effect
{'...,."..,lt60>i1., = 12 '- I I }
on the other like MMP. Basically, in
] {I, 'II,'= I, + I....,}
case that plural tolerances specify a
(tolerance of Tg) = I,'
single or plural features, they can be
(tolerance of Ts) = I,'
defined independently. But in special
case that one tolerance affects on the
other like MMP, it is necessary to Figure 11 Representation ofTr.
express the relation between them as
the constraint equations. Figure II shows the constraints of tolerances related by MMP. Here,
the geometrical tolerance is 11 and dimensional tolerance h.
3.4 Relations between Information in Parts

Figure 12(a) shows an example of a part specified the perpendicularity on a center line and its
relation among feature, dimension and tolerances is shown in Figure 12(b). This relation is
expressed by a simplified
formulating Tg, Td, Ts and Tr, as Table 2 Combination by categories
shown in Figure 12(c). The
multiple
~
cumulation or the propagation in single ,
the tolerance chain can be used feature (regardless) regardless ,, restriction
re~tive
til
along the network. For the
simplification of relation, actual single Tg Tgg - ~'
,I
-
t il I
pattern of the relationship among

information are classified into II
types as shown in Table 2. In this
table, the symbol 'd', 'g' and's'
Ts
--------- ~
Tdg ~.:! Tgs
__________
dgg ~ 1: __________
T_________ (IM')
aEII ~
means datum, geometrical and

multiple
~~g-~~~-'
Tgs DEl:
_____ ~ __ , Tdgs
~
B :
dimension respectively, and the
symbol 'Tdg' means the
toleranced feature with related
Tds q
Tggs (IW)
feature such as straightness. I

Tdg=Td+Tg Tgg=Tg+Tg Tdgs = erg + Tg)+Ts I
(a) D"",;ng (b) Relabon among information (c) Connection using categories
Figure 12 Relation between elements in one parts.

4 Expression of feature model
4.1 Classification offorms and functions in parts

In order to express the geometrical feature model, we classified forms in actual machines and
parts and obtained the characteristics of forms.
Research and classification ofform

We classified about the geometrical characteristics and
" 6 0 . 0 100
"f
functions in actual parts in order to relate between every .' 0
shape in parts and manufacturing information. We
selected and researched about three machines as samples Ty p. 1 .... 1 oS! .ndnd
ON on.s...ndar
that has different number of parts, winch, pump and lathe. T yp.2 ."""- J . P""
Figure 13 shows the number of parts and the ratio of Ty pe J ~7~ r
standard parts in samples. The ratio of standard parts
increases in proportion to the increase of number of parts. so 100 1 so 200
Here, we regard general standard parts that has s a m e . . par-11

shapes like bolt as single parts. ~lgure ~3 RatIO of standard parts
Figure 14 shows the classification according to shapes of m machmes.
materials. Though the

Number
ratio becomes smaller in
o SO 100 ISO
the order, casting, 80
cylinder, tube, etc., in 60
~
case of the actual number 40
of machines, the ratio of
20
large size parts such as
casting decreases. About o
Type I Type Hype 3 ALL
the symmetry and Types of Machine,
complexity of shape of
Figure 14 Types of shapes of Figure 15 Geometrical property
material, symmetrical material.
parts exists in large of parts.
quantities as shown in Number

Figure 15. o 2040 60 80100120
umber
Figure 16 shows the o 20 40 60 80 100120 Chunrer
mechanical functions
about symmetrical and
~J{
g\t .. in&~~
ii
•~\Ildne55 !"~.'~"~~~~
Screw
Hole
unsymmetrical parts as .g Slot
.-ti§ Sub Fix I~=====~
&i\'t Gear
the reason of Figure 15. Control
"'" ~~~!'! a Non-Rotative shap ~ Taper
There are a large amount Others
Others O Rota1i\'c shape
of parts that have er::::::::::::::::::=1
functions of guide or Figure 17 Types of functional
support independency of Figure 16 Mechanical function of shapes.
symmetry of parts. parts.
The functions in parts are mainly classified as shown in Figure 17. Parts are constructed by their
functional shapes and have own functions.
Definition ofform model

As a result of above mentioned research, considering
manufacturing process, we proposed three feature
model as unit, namely, (1) material, (2)
manufacturing, (3) functional form model. Forms in
parts can be expressed by using three model. Figure
18 shows the example of assembly constructed by
IMA TERIALSI ~)
C"",,-)
plate, cylindar,box etc.
them.
(1) material form
In case of consideration of manufacturing process
and method, the initial shape of parts affect the
---
definition of manufacturing method. We defined the
initial shape as material form. It consists of simple
geometrical form such as plate, cylinder, box, etc ..
~P2
.:I
(2) manufacturing form FlJNCTIONAL
SHAPES -- ,' " "
In the removal manufacturing process such as lathe
chamfer, roUD:lness, - -' , _~ -
turning, the manufacturing surface is generated on screw,hole, slit, gear, 0
the material surface and made the targeting form. We spline,taper, others t lmnfed
defined these generated surface as manufacturing Figure 18 Types of forms in parts.
form.
(3) functional form
In addition to the general manufacturing form such as screw, hole, etc., surfaces that have
single function by the characteristics of form exist in parts named functional form.
4.2 Structure of model

Figure 19 shows the concept offeature
DOF con traint B-n:ps Manufactu ring
model and the relationship among ,-- ---------
information. Model consists of three ,
parts; (1) geometrical feature model, (2)
degree of freedom (DOF) constraint
model and (3) manufacturing group.
About geometrical feature model, we
expand the B-reps model to express the
virtual feature such as center line of plane
and to treat them as same as general
feature. We also added some categories
such as parts and block that is equal to the
primitive of CSG model to the hierarchy Figure 19 Construction of model.
of information.
4.3 Relationship between DOF constraint and B-reps model

In order to express the toleranced feature, we express the area and the behavior of feature and
tolerance zone by limiting degrees of freedom (DOF) against the feature and other information.
The constraint ofDOF is applicable to four information such as feature, dimension, tolerance and
coordinates as mentioned above. The feature that has no constraints ofDOF floats in space. For
the purpose of applying this model to the geometrical feature model that has B-reps structure, it
is necessary to combine information between each model.
Basically, the surface information ofB-reps model relate to DOF model as shown in Figure 4.
In case ofB-reps, the surface information relates to edge and vertex according to the topological
relation. DOF constraint model, in case of surface, treats surface feature as the set of
fundamental feature such as line or point that exist in it. The type of fundamental included the
higher rank feature changes by the type of tolerance specified feature such as the straightness and
the flatness. In this case, each fundamental feature is expressed by the locational information Cc
and Cf to express the size of itself and higher rank. Each information, Cf , Cc, are limited their area
of behavior by constraint equations. In case of the cylindrical feature as shown in Figure 4 (b),
there are multiple cases by the types of tolerances, in case off3 , it can be expressed by the
network among the center line and the cylindrical surface and the dimensional information C•.
Dimensional and geometrical tolerance and datum information are combined with the
corresponding information such as feature and dimension. Concretely, we expressed their
behaviors by using matrixes and constraint equations according to each constraints.
4.4 Manufacturing group
In case of assignment of tolerance using manufacturing information included tolerances, is it

necessary to correspond the manufacturing process to the geometrical model and to divide each
shapes into each manufacturing lot. Defining the manufacturing method, it is mainly affected by
the order, surface treatment, size, accuracy etc .. In this paper, we don't refer to the definition of
manufacturing order, but we think that it is Manu,ac1uriog Func tional
possible to reconstruct its order because the MII.,;,I Surf... Shape =
relation between surfaces that is formed by
phase from material to functional shapes can
be obtained hierarchically.
The manufacturing group is classified into
three group as shown in Figure 18. The
material form has solid information
constructed by simple forms and their
combinations. Each manufacturing surface
relate to the material form hierarchically.
Therefore, parts have the relevance as the
GT family by the similarity of material forms
included manufacturing and functional
forms as shown in Figure 20. The material
form has several manufacturing forms, and
the types of them are mainly classified into
features such as plane, cylindrical surface,
taper, free-for med. the toleranced feature in Figure 20 Grouping of parts by GT
methodology.
the geometrical model correspond to

this manufacturing form. Furthermore, Table 3 Types of manufacturing and functional forms
The functional form exist on each
manufacturing form.
Table 3 shows that the relationship
between the manufacturing form and J:~::e W~VJji'>;l - - - !?;~~;~S_ - --
the corresponding functional form. We _ 1~~r ______ --------------
defined the manufacturing and Curved _____ ,!oJ!l _____ _
functional form as the unit of _ P!i!.t1.L _ _ _ _ _____ ~I_i! _____ _
manufacturing lot. Using this, the _ Q"i Li.[l~[ Lc_
al _
standard parts are defined by the _I~~[ ______ ~~~~
concept of complex parts of GT. Curved
4.5 Relationship among information in parts
Figure 21 shows the example of

breaking into each model. Here,
limiting information in this parts,
there are single dimension of hole
FWlctionai
and single straightness against the
hope
center line of single hole.
It is assumed that the material ~
form of this parts is rectangular
shape and has two hole as the
manufacturing surface.
Furthermore, single plane feature is
treated as the datum feature. Two
holes are defined as functional
forms in this case. The functional
forms also have the directions and
lengths of the manufacturing
surfaces to combine with the
manufacturing information.
Finally, each geometrical feature
model is spread into DOF
constraint model.
In case of calculation of
cumulative tolerance, these relation
are replaced by using fundamental
categories to simplify the
propagation network of feature,
tolerance, dimension and datum as
shown in the bottom of Figure. Figure 21 Relation among information in parts.
5 Assignment of tolerance
5.1 Assignment considering sensitivity
We assume the set of tolerance specified against a parts and multiple parts, namely T.
(5)
Assuming that the target performance and function are replaced by the accuracy specified
parts and machines, against the target performance Pi (i =1, ... , m) and the target manufacturing
costs Cj G=I, .. .,j), considering the performance or the accuracy Pi and Ci are expressed as follows;
(6)
and in case that every performance and costs need to satisfy against the target, the relation
between pi and Pi and between Cj and Cj must satisfy the following condition.
(7)
Furthermore, in order to define every tolerance t satisfying above conditions, we used the
following sensitiveness against each performance and costs for the assignment of tolerance.
ap .
Sp= - - ' ,
ac
sc = - -
j
(k=l,., n) (8)
at k at k
We considered about the allocating

method of tolerance using the
performance and the manufacturing
cost. Figure 22 shows the concept and
assumption of the relation among
tolerance, target performance and
costs in parts. The target performance
that means the accuracy defines by the
cumulation of every tolerance, and its
computation follows the cumulation
as mentioned above. Similarly, the
sensitivity of every tolerance against Figure 22 Concept of performance and cost using
the target performance is able to be tolerance.
obtained.
5.2 Connection between feature and manufacturing cost
The manufacturing cost divide into two parts; one is the direct cost like manufacturing and the
other is the indirect such as materials, personal expenses and so on. We handled only the direct
cost and treated with the direct as the fixed one to confirm of the validity of our allocating
method.
For the purpose of the computation of the manufacturing cost, we considered about the data
Figure 23 Data flow of computation of manufacturing costs.

flow from accuracy data such as tolerance and dimension to the manufacturing cost. Figure 23
shows the flow of calculation of the manufacturing cost in case of the lathe as the example.
Using these flows, and indicating the manufacturing method and order, manufacturing costs for
each feature or each lot are calculated following them.
5.3 Flow of assignment of tolerance using manufacturing cost
Figure 24 shows the flow of assignment of

tolerance using manufacturing costs. Using
three model as the initial condition, we set
the targeting costs and performances. And
modification of targets or
against the target value we define the models
standard of evaluation such as 'greater
than', 'less than', etc .. Here, we assumed
that the target values were able to
expressed by using the accuracy in parts or
assemblies. First, the system check the
validity of current tolerances against
targets. About features that no tolerances assignment of tolerance
are specified, we defined that the general
tolerances applied to them. In case that definition of tolerances
current conditions don't satisfy the target,
the system indicates and requests the end
change the targets values or the

modification of design. In case of changing Figure 24 Flow of assignment of tolerances.
relating tolerance, the system changes the tolerance considering the sensitivity for the target.
After satisfying conditions, it assign and define tolerances satisfying each standard of evaluation
by using the sensitiveness against the target for tolerances.
6 Conclusion
We considered the feature model to faithfully express a dimensional and geometrical tolerance on
a computer against the current tolerancing, we proposed the three model; geometry, DOF
constraint and manufacturing, and considered the relation of information. DOF constraint model
express the tolerance zone defined by allowable behavior of feature. And the manufacturing
model means the expression of shapes and functions against each manufacturing lot. Using their
model, we considered the method of assignment of tolerances using the sensitiveness of
tolerances against the target.
7 References
BERNSTEIN, N.S. and PREISS, K. (1989) Representation of Tolerance Information in Solid

Models. Advances in Design Automation, ASME, 37.
Canadian Standards Association (CSA) B78.2-M86 (1986). Dimension and Tolerancing of
Technical Drawing.
CHARLES, B., CLEMENT, A, DESROCHERS, A, PELISSOU, P. and RIVIERE, A (1989).
Toward a computer aided functional tolerancing model, The International Conference on
CAD/CAM and AMT proceedings, c-2-4, 1-7.
FLEMING, A D. (1988) Geomeric relationships between toleranced features. Artificial
Intelligence, 37, 403-412.
MARTINSEN, K. (1993) Vectorial Tolerancing for All Types of Surfaces. Advances in Design
Automation, ASME, 2, 187-198.
MITTAL, R.O., IRANI, SA and LEHTIHET, EA (1991) Tolerance control in the machining
of discrete components. Journal of Manufacturing Systems, 9, 3, 233-246
REQUICHA, AAG.(1983) Toward a Theory of Geometric Tolerancing. The International
Journal of Robotics Research,2,4,45-60.
RESHETOV, D.N. and PORTMAN, v.T. (1988) Accuracy of machine tools. ASME PRESS,
New York.
WIRTZ, A (1988) Vektorielle Tolerierung zur Qualitatssteuerung in der mechanishen
Fertigung. Verein Deutsher Ingeniere, 751, 493-498.
8 Author(s) biography
Toshio MAEDA is a Ph. D Student in the department of mechanical engineering at KEIO

University. He expects to Ph. D. degree in product modeling on CAD in 1995. His research
interests include tolerance analysis, functional representation on CAD and decision making.
Naochika TOKUOKA is presently an Assistant Professor of mechanical engineering at KEIO
University. His research interests are the area of the decision making of mechanical design.
7
Quality Features in CAD - and
CAPP - systems
H. K. Tonshoff, M. Ehrmann
Institute of Production Engineering and Machine Tools
University of Hanover
Schlosswender Str. 5
30159 Hanover
Germany
Phone: + +49-511-762-5220
Fax: ++49-511-762-5115
e-mail: ehrmann@ifwsn4.ifw·uni-hannover.de
Abstract
The contributed paper describes an approach for the integration of tolerances in CAD- and
CAPP-systems. The use of quality features eases the work of the users in design, process
planning and NC-programming.
In computer systems products or components are represented as internal workpiece models,
which have a perfect shape. Tolerances for dimensions and geometry are used to state the
limits of deviations, which are allowed in manufacturing. A workpiece produced with an
accuracy within the limits of the stated tolerances will fulfil its function.
Similar to the feature approach in design and process planning accuracy information can be
stored in quality features. According to the type of accuracy data the quality features can be
divided in several groups: Dimensional tolerances, geometrical tolerances and surface
treatments. Current standards developed for two-dimensional drawings have to be extended
for their use in a three-dimensional volumetric model. A quality feature comprises in-
formation on tolerance type, tolerance value, toleranced feature, reference feature and
attributes for the tolerance zone and plane. Quality features support tolerance relationships
within one feature or between different features. New techniques allow to apply tolerances to
the resultant geometry, which may not exist in the feature templates.
Keywords
Feature based tolerancing, tolerances in CAD-/CAPP-systems
Quality features in CAD- and CAPP- systems 105
INTRODUCTION
Systems available on the market support the processing of the three-dimensional workpiece
geometry through the use of a solid modeller. In addition to the pure geometry further text
based information can be added to the workpiece in some systems. The information is not
directly related to the workpiece model and is not part of the data structure transferred to
other systems (Wozny, 1992).
To increase efficiency of CAx-systems research work is focusing on improving the
functionality through the use of features (Rudolph, 1993; Shah, 1991). Feature oriented
design and process planning systems improve the bi-directional workpiece data transfer
between the different stages of the product development. Furthermore the use of features in
the design process eases the incorporation of technological information into the workpiece
model and leads to more user friendly man-machine interfaces of CAD/CAPP-systems
(Case, 1991; Salomons, 1994). Features are units of information, which may include the
geometric description and additional data corresponding to the user's point of view
(Rieger, 1994; Tonshoff, 1994-1).
Solids Symbols
Cylinder 1 Cylinder 2 Cone Thread
OJ ;
W
I
9 ,
W i
¢
~
!
I
i
I
t"'P'J
i
I
I
Design Feature Counterbore Hole whh Thread
57/13077 © IFW
Figure 1 Example for a design feature.

2 THE FEATURE APPROACH IN PRODUCT DEVELOPMENT
In the design stage design features such as different types of holes, slots or bearing seats sup-
port the designer's functional way of thinking. Figure 1 shows the structure of the design
feature Counterbore Hole with Thread. The feature consists of two cylinders, a cone and a
thread, in which all the entities have the same axis. The feature description includes, beside
the explicit data like diameters, lengths and thread type, implicit information as follows:
• The end position of cylinder one is the start position of cylinder two.
• The end position of cylinder two is the start position of the cone.
• The maximum diameter of the cone conforms to the diameter of cylinder two.
• The thread is placed on the surface of cylinder one.
In process planning features are grouped considering technological criteria. They are called
manufacturing features and have a set of possible manufacturing processes associated with
them. The features are defined in generic templates, which are instantiated by the user during
the design or planning session.
3 PROBLEMS ENCOUNTERED IN TOLERANCING A 3D-MODEL
Quality features for stating functional requirements

The model created by a designer in a CAD-system represent the exact or nominal shape of a
workpiece. In a later stage of design the user has to state the limits allowed in manufacturing
by the specification of accuracy data (Juster, 1992 and Stark, 1994). Accuracy data may
comprise information of workpiece material, surface conditions or tolerances. According to
current standards tolerances may be divided in several groups such as dimensional and
geometric tolerances or surface treatments. Whereas design features are used to support the
designers' functional view implicitly, quality features are a mean to express functional
requirements in an explicit way.
Addressing resultant geometry level entities

As opposed to design features consisting of geometry and additional information, accuracy
features include only pointers to existing geometry. Furthermore design features which are
associated with volumetric entities contain a reference to a complete feature body, whereas
accuracy features often refer to entities of the resultant geometry level. Examples for such
entities are faces, edges or vertices, which can be part of the design feature templates or the
result of several merged features. In Figure 2 two intersecting holes and a tolerance of
roundness are shown. The area of the toleranced face depends on the moment, when the
tolerance is attached to. If the tolerance is stated before the second hole is modelled, the
whole, but in the final version intersected, face of the first hole is toleranced. If the tolerance
is applied in the end of the design, only the part of the face created after intersecting the both
holes is referenced. This example indicates that the adaptation of current standards to the
feature technology causes problems.
Various reference entities

Another difficulty is the diversity of toleranced entities. The same tolerance type may need
different parameters depending on whether an edge or a face is referenced. This leads to
problems at the implementation of tolerance templates and procedures for applying these
tolerances.
'd.: OF·'
Type: Th<ougt>hoIe Id.: 0'·2
lac-X: 120
Type: Tlvougt>hoIe
lot-V: 150 Loc-X: 250
lOC-Z: 39 Loc-Y: 150
Road:
Loc-Z: 39
Rad~
kt: EFI·,
Class: .'pOoI
Typo: A<><.<mou
Ref·' : OF·'
TVillue': 0.000
57/ 15369 0 IFYi
Figure 2 Possible resultant toleranced entities depending on the moment of tolerancing.
Adaptation to the three-dimensional model

Most standards written for the use in two-dimensional drawings limit a tolerance zone
referring to the drawing plane. For the use in a model in the 3D-world it is necessary to
develop an expanded tolerance description. The tolerance has to include parameters like a
position and a vector identifying the tolerance plane, likewise procedures have to be
implemented to help users stating these planes.
Types of quality features

Similar to the different meaning of features in design and process planning, tolerances
support the specific way of thinking of the product developer. In the design stage tolerance
features provide support in functional aspects, whereas in the process planning stage they
assist to take decisions considering technological possibilities of the workshop. Tolerances
and dimensions are attributes, which are explicitly added to the workpiece model by the
designer. Therefore quality features comprising this kind of accuracy data are called explicit
quality features.
This quality infonnation is used to detennine suited manufacturing operations in the

succeeding process planning stage. An example is given in Figure 3. Depending on an
associated surface finish and its value, a hole may be manufactured only by drilling
operations or has to be finished by reaming.
Type: Hole Type: Hole Type : Surface

ID: DF-l ID: DF-2 Finish Design
ID: QF-l
Ref-ID: DF-2
Functional
Ref-Ent: Face requirements
SF-Type: Ra
Value: 1.6
Manufacturing Manufacturing
Feature Feature
Type : Hole Type: Hole

tD: MF-l ID: MF-2
Manufacturing
Technological
requirements
Mlcrocycles: I Microcycles :
drilling
~~I~~1/,g will1 reamers
sPQldrllllng
I dniling
I
~In~illing
W;;lshing will1 reamerS
'- ./
57/15371 © IFW
Figure 3 Microcycle selection supported by quality features.
A second type of quality features are the implicit ones. The implicit quality features include
hints on interactions between different features. These interactions result of the position,
orientation and dimensions of single features and have influence on the manufacturing
process. For example a thin wall resulting from the proximity of two pockets can cause
difficulties for manufacturing.
Support of the mapping process

Further problems arise through the mapping from design to manufacturing features. For
instance a design feature of type counterbore hole is transferred into two manufacturing
features of type hole during the mapping process. An update mechanism now has to check if a
tolerance attached to the design feature has to reference one or both manufacturing features.
Furthermore it is necessary to create an implicit relationship of type concentricity, which

states the common origin of the two manufacturing features of type hole.
Data exchange between systems

A formal language is used for the exchange of quality features between different systems. The
language contains the description of the tolerances in an implicit form. The use of this
implicit description makes the language independent of the involved systems. The content is
not given in an explicit, system-specific format, but the rules for the generation of the features
are transferred.
Currently quality features are used in detailed design, process planning and NC-
programming. Future actions may include the link to early design stages such as the
conceptual design phase, which will take into account existing models like the TTRS
(Clement, 1995 and Desrochers, 1995).
In the following the described approach is examined in two research projects.
4 THE SESAME SYSTEM
SESAME (Simultaneous Engineering System for Applications in Mechanical Engineering) is

a Brite-Euram project funded by the European Union. The consortium of SESAME consists
of a software-vendor, an enduser and two universities. The project aims at the development of
an integrated system covering the areas design, process planning and NC-programming. The
efficiency of the system will be evaluated by three reference parts provided by the enduser, an
Italian company producing flexible machine centres (Tonshoff, 1994-2).
The use of feature technology in the whole SESAME system provides information for later
stages as early as possible, i.e. the designer uses design features which are transferred and
mapped to manufacturing features in process planning.
The SESAME system is subdivided into three main modules: the feature based design
system, the automatic process planning system and the interactive process planning system
with the NC-programming module.
4.1 Feature Based Design System
The feature based design system developed in SESAME allows the designer to use design
features such as slots or bearing seats, which support the functional way of thinking. A
substantial assumption is that the designer follows a specific way of modelling. He has to start
with the modelling of the raw material, e.g. a casting. The feature based design system offers
the functionality of creating solids and changing them through Boolean operations.
Alternatively any geometric model created with the ACIS modeller can be imported into the
feature based design system as non-parametric base part. Starting with the raw material the
designer can define the design features, which are subtracted from the base part analogous to
the machining process.
The features are defined in generic templates, which can be easily supplemented by user
defined features. During the design session the user creates new instances of the appropriate
feature template and integrates them with the rest of the model. Each feature is described by a
feature specific set of data including information about the name, type, position, orientation,
kind of visualisation and additional attributes. The feature type comprises three sub classes.
Volumetric design features like slots, pockets or bearing seats, which denote a subtractive
volume. The SESAME feature based design system includes a total of eighteen volumetric
design features. The second type are local design features like chamfers and blends, which
produce no new volume, but modify parts of the existing model. Furthermore some non-
geometric features are implemented, so called explicit feature interactions. Explicit feature
interactions include all remarks, which are attached by the designer in an explicit way to the
workpiece model such as tolerances or surface conditions.
,=
Human
Designe
'----,-
Using
EFI
EFI
P1.1.1 Templates I 0 II Explicit Feature
I
input
Explicit W Workpiece
Interaction Library
Feature Model Data I0 I' Design Feature

llnteraction I Workpiece Model
Data
I<- MCL+-
1
Functions MCL+-
0 Interface
1
Workpiece
Model Data
l
~T
I
P1.1.2
T
1P1.1.3 1P1.1.4
check compute represent
Explicit Explicit Explicit
Feature Feature Feature
Interaction Interaction Interaction
Data Data Data
Display Data I Workpiece
1
Model Data
P1.2
I D I, Design Feature
display Workpiece Model
Explicit
Feature
Interactions
in FBDS
57/15370@IFW
Figure 4 Detailed design of the EFI modules (Gane Sarson notation).

Explicit Feature Interactions

The software modules providing explicit feature interaction (EFI) functionality are integrated
as MCL+-procedures into the feature based design system using the MCL+-interface of the
design system. MCL+ is a system specific interpreter language based on C++. Figure 4
describes the architecture of the EFI-module. The human designer adds explicit feature
interactions such as geometric tolerances to the workpiece model using the EFI-templates.
The workpiece model data may be read or stored in design feature workpiece models. MCL+-
functions which can be addressed through the MCL+-interface support the processing of the
workpiece model data.
In order to allow a better understanding a user defined MCL+-procedure is presented used
for attaching a geometrical tolerance of roundness to a feature based workpiece model. The
procedure is structured in four key parts: Input module, check module, computing module,
representation module.
The input module

The input module contains the procedures for entering the parameters essential for associating
a tolerance to an entity of the workpiece model. The necessary data depends on the tolerance
type as well as the toleranced and referenced entity. In the example of the roundness tolerance
four parameters are needed besides its name and value. First the toleranced entity has to
specified, in this case an edge or a face. Afterwards a plane has to be stated, in which the
tolerance symbol shall be placed. The third parameter to be set as a vector gives the direction,
in which the tolerance symbol shall show. At last the symbol height has to be put in. The
setting of the described parameters is an interactive process supported by the system. For
instance in the status line a brief comment indicates the meaning of the parameter to be
specified. Identifying an entity is done using the cursor. Different cursor shapes support the
user in picking appropriate entities. Furthermore for the indication of the tolerance symbol
direction the user can accept default values suggested by the system.
The check module

After the user has instantiated a parameter, the value of this parameter is evaluated in the
check module. This process serves different purposes. A check can be performed as
consistency check verifying the correctness of stated data. For instance in stating a tolerance
of roundness a toleranced edge has to be circular, a toleranced face cylindrical or conical.
Moreover checks are used for the selection of the further proceeding. The results of this kind
of checks form the basis for the processes performed in the succeeding computing module.
An example is the check if a toleranced face has a regular shaped end influencing the way of
determining the right position to place the tolerance symbol.
The computing module

Based on the results of the check procedures the main task of the computing module is to find
out a position suitable for starting the visualisation of the tolerance symbol. In case of a
roundness tolerance this position should be related to the toleranced edge or face. The
position proposed by the system depends on the toleranced entity type and the symbol
direction given by the user.
The representation module

The representation module consists of procedures needed to provide data for the visualisation
of the attached tolerances. Up to now two types of user support have been implemented. First,
tolerance symbols similar to standard symbols used in two-dimensional technical drawings
are adapted to a presentation form in the 3D-world. Moreover a toleranced entity is marked
by a special colour which allows the simple identification of such entities.
4.2 Interactive Feature Planner
The feature planner is designed as an interactive process planning tool and is divided in two
parts. The planning module provides a graphical visualisation of the planning process and the
process plan editor supports the formation of process plan header data and the manipulation
of single process plan operations. The process plan generation is carried out using a feature
based workpiece description.
The workpiece description used for process planning consists of manufacturing features
which result of the mapping process. During the mapping process the design feature
workpiece description is transferred to the manufacturing feature workpiece description using
mapping rules. This mapping information is stored in the feature library, which is used for the
initialisation and configuration of the feature planner. The advantages and requirements of
configurable feature oriented CAPP-systems are explicitly described in (Tonshoff, 1994-1).
A few design features are mapped to manufacturing features of the same type without
having alternatives. An example for mapping with two alternatives is the design feature of
type counterbore hole. This feature may be manufactured using a special tool which is
capable to produce the feature in one operation. The alternative is to manufacture it in two
steps first using a drill for the hole and then a counterbore tool for the counterbore. If
mapping alternatives exist, the user has to decide which one is applied for manufacturing the
feature.
Based on the mapped workpiece description a machine group and a suited set-up can be
related to each manufacturing feature. To choose an alternative among several possibilities is
the interactive task of the user taking into account the minimisation of costs and set-ups. After
that the process planner has to select a suited microcycle which comprises the operation
sequence for manufacturing a feature. Each manufacturing feature has an associated set of
microcycles which represent different manufacturing technologies. Criteria for the selection
of a suited microcycle are the dimensions of a feature and additional restrictions like explicit
or implicit feature interactions. After the assignment of microcycles to features, tools and
other technological data have to be related to each operation. Information about microcycles,
tools and technological data is included in a database, which is accessible by the feature
planner.
4.3 Automatic Process Planner
The automatic process planning system used for standard parts consists of two modules, the
planning space generator and the optimiser. The planning space generator reads feature based
workpiece descriptions generated in the feature based design system. After the geometric
reasoning process where implicit feature interactions are recognised, a network of
manufacturing possibilities for the features of the component is generated by relating

technologically possible microcycles to each feature. The assignment of microcycles is per-
formed in consideration of feature parameters as well as explicit and implicit feature inter-
actions. Tables have been produced in which technologically possible microcycles may be
selected depending on feature dimensions, tolerances, anteriority constraints and implicit
feature interactions. After the selection of microcycles, tools are selected for each microcycle.
This is also done according to rules stated by the end-user. Finally cutting parameters are
related to each type of operation.
The network of manufacturing possibilities generated by the planning space generator
serves as input for the optimiser. An optimised process plan considering costs, number of set-
ups and machining time is generated by the optimiser using genetic algorithms.
4.4 NC-Code Generator
The NC-programming module reads in process plans generated from a SESAME process
planning module, checks their feasibility and generates NC files. Process plans are handed
over to the NC-programming module in the process planning language that comprises a file
describing the features to be manufactured with their parameters and information on
machines, tools and technological parameters generated in process planning. The NC-
programming module is based on a commercially available NC-programming toolkit
(STRATA). It evaluates the feasibility of the generated process plan by checking for
collisions, and calculates times and tool paths. The result of the collision check and the
manufacturing times are returned to the process planning module. In case of problems it has
to provide another process plan or has to give hints to the designer which constellations
caused the collision.
5 QUALITY FEATURES FOR THE AUTOMATIC PROCESSING IN THE

CAX-CHAIN
In a national funded project the main goal is the development of quality assurance features
supporting the generation of a quality assurance plan. First quality assurance features have to
be defined taking into account the information and the dependencies between accuracy
features and design or manufacturing features. In order to handle complex dependencies such
as tolerances with several referenced entities it is necessary to develop a method to describe
them. Likewise the interactions between quality features have to be considered. The
intersections of tolerance fields shall be examined and a concept for the description will be
worked out.
Quality features may comprise accuracy data like tolerance information providing
dependencies between several features. For example two design features may be linked
together by a quality feature. A hierarchy attribute states the position of each feature in this
feature compound. The features can have the same rank or a parent-child relationship.
Depending on the kind of hierarchy, the necessary modifications may be performed
automatically in the compound after one related feature was changed. Furthermore the
dynamic behaviour of features shall be analysed, if more than two features are combined.
Moreover quality features often refer to entities of the resultant geometry level. To address
such entities not included in feature templates, a modeller independent method will be
developed. A new concept for the definition and description of tolerance zones in a 3D-model
will be worked out, which will fit in the frame given by STEP. These three-dimensional
tolerance zones shall be part of the quality features. Based on the developed concept,
interactions between three-dimensional tolerance zones have to be examined and a
mathematical model has to be developed to describe fields and resulting intersections.
Furthermore new techniques for the visualisation of dimensions and tolerances have to be
developed. Existing standards for the representation of dimensions and tolerances refer to
planes in two-dimensional drawings and are not well suited for visualisation in a three-
dimensional model.
A quality assurance plan consists of several quality assurance features, which are built up
processing the information stored in a feature based workpiece model. Feature parameters
such as type or dimensions and the data included in quality features support the generation of
quality assurance features considering different quality assurance strategies. The proceedings
take into account both manual quality checks and the automated control by a coordinate
measuring machine.
6 CONCLUSIONS
The described approach is examined in two research projects. A Simultaneous Engineering

System for Applications in Mechanical Engineering is developed in SESAME, which is
funded by the commission of the European community. The SESAME project aims at the
integration of all aspects of prismatic mechanical part development from design to process
planning and NC-programming. Dimensional and geometric tolerances defined in
international standards are adapted to the requirements of three-dimensional modelling and
implemented as part of the feature oriented design module. The explicit quality features are
described in a system-specific programmers language, which is based on C++. Implicit
quality features like intersecting features or proximity of features are detected by means of
geometric reasoning in the process planning module. Both explicit and implicit quality
features influence the way of manufacturing a feature.
In a national funded project quality features support the automatic processing of workpiece
models in the CAD-CAPP-CAM-chain. Quality features are described in a neutral, system
independent format using a formal description language. The information transferred by
quality features support the automatic generation of process and quality assurance plans. A
current research goal of this project is to find mathematical descriptions for three-dimensional
tolerance zones. Based on the descriptions the interactions between tolerance zones can be
examined.
The prototype implementations show that the approach of handling tolerances as quality
features is well suited to improve the functionality provided by systems in the product
development.
7 ACKNOWLEDGEMENTS
The presented research project SESAME is being funded by the Commission of the European
Community (CEC) under the BritelEuRam Programme (Project BE #4539).
The presented research project "Quality features for the automatic processing in the CAx-
chain" is being funded by the National Research Foundation (DFG) under the programme
"Innovative quality assurance in manufacturing" (Project TO 56/118-3).
8 REFERENCES
Case, K.; Gao, J. (1991) Feature Technology - An Overview. Proceedings of Symposium

Feature-Based Approaches to Design and Process Planning, Loughborough, September
1991.
Clement, A.; Riviere, A.; Serre, P. (1995) A Declarative Information Model for Functional
Requirements. Proceedings of the 4th CIRP Seminar on Computer Aided Tolerancing,
Tokyo, April 1995, pp. 1-19.
Desrochers, A.; Maranzana, R. (1995) Constrained Dimensioning and Tolerancing Assistance
for Mechanisms. Proceedings of the 4th CIRP Seminar on Computer Aided Tolerancing,
Tokyo, April 1995, pp. 21-29.
Juster, N. P. (1992) Modelling and representation of dimensions and tolerances. Computer
aided design, volume 24, number 1, January 1992, pp. 3-17.
Rieger, E. (1994) Semantikorientierte Features zur kontinuierlichen Untersttitzung der
Produktgestaltung. Reihe Produktionstechnik 158, Carl Hanser Verlag, Mtinchen, Wien.
Rudolph, F. N. (1993) Konfigurierbare Technische Elemente flir Konstruktion und
Arbeitsplanung. Dissertation, UniversiUit Hannover.
Salomons, O.W.; van Slooten, F.; Jonker, H.G.; van Houten, FJ.A.M.; Kals, H.J.J. (1994)
Interactive Feature Definition. Proceedings of the IFIP international conference "Feature
modeling and recognition in advanced CAD/CAM-Systems", Valenciennes, Mai 1994,
pp. 181-205.
Shah, J.J. (1991) Assessment of Features Technology. Computer-aided design, Vol. 23, Nr. 5,
1991,pp.331-343.
Stark, R.; Weber, C. (1994) FHichenorientiertes Toleranzmodell als Grundlage der
rechnergesttitzten Tolerierung mit 3D-CAD. Tagungsband CAD '94
Produktdatenmodellierung und Prozej3modellierung als Grundlage neuer CAD-Systeme,
Paderborn, Miirz 1994, pp. 199-222.
Tonshoff, H.-K.; Aurich, J. C.; Baum, T. (1994) Configurable feature-based CAD/CAPP
System. Proceedings of the IFIP International Conference on Feature Modeling and
Recognition in Advanced CAD/CAM Systems, Valenciennes, France, May 1994.
Tonshoff, H.K., Baum. Th., Ehrmann, M. (1994) SESAME: A System for Simultaneous
Engineering. Proc. 4th Int. FAIM Con! 'Flexible Automation and Integrated
Manufacturing', 8.-11. Mai 1994, Virginia, USA, pp. 380-389.
Wozny, M. J.: Research trends in the U.S. for next generation CAD systems. CAD '92 - Neue
Konzepte zur Realisierung anwendungsorientierter CAD-Systeme, Springer Verlag, 1992,
pp.30-65.
9 BIOGRAPHY
Dr.-Ing. Dr.-Ing. E.h. H.K. Tonshoff is full professor of production engineering at the
University of Hanover, Germany. After receiving his Ph.D. degree in mechanical engineering
in 1965 he worked as a designer, technical manager and vice president in a machine tool
company. In 1970 he became director of the Institute of Production Engineering and Machine
Tools at the University of Hanover. In 1992 his work was honoured by bestowing a
Dr.-Ing. E.h. degree on him. He is vice president of the German Research Council (DFG) and
member of the International Institution for Production Engineering (CIRP).
Dipl.-Ing. M. Ehrmann studied production engineering at the Technical University of

Munich, Germany. Since 1993, he is a research engineer at the Institute of Production
Engineering and Machine Tools, University of Hanover, where he is currently working in the
CAD/CAPP group.
PART THREE
Modelling Geometrical Error

8
Geometrical Behavior Laws for
Computer-aided Tolerancing
P. BOURDET, E. BALLOT
LURPAIEcole Normale Superieure de CACHAN
61 Avenue du President Wilson 94235 Cachan Cedex FRANCE
Tel: 33 1 47402215 - Fax: 33 147402220
e-mail: bourdet@lurpa.ens-cachanfr
Abstract
The analysis of the difticulties encountered with traditional dimension chains in the description
of the behaviour of a set of parts, with variation, has led us to develop a tridimensional model
of variations. It is therefore designed to treat the problem of transfer of dimensions.
The definition of the model of the variation on surfaces of parts relies on a 2-type
classification of defaults which allows to then connect them formally.With this model, the
requirements of tolerancing have brought about a definition of two operators. The first
determines the specifications that may be found on a PUit by defining the intersection of the
domains of deviation of surfaces in relation with their nominal model. The second calculates the
union of motions of a part under the intluence of each of the mating faces with contiguous
parts.
The implementation of the model and of the two operators then allows for a definition of the
varied possible configurations of the mechanism. For each of these configurations, the
constraints of tolerancing that are exerted on the geometry of each part in a mechanism can be
expressed mathematically.
Keywords
Mechanism, tridimensional tolerance chains, geometrical model, dimensioning, mobility
degrees.
1 INTRODUCTION.
Though it is theoratically possible to obtain a perfect geometry, one has to admit that for an
optimal cost production tools make an approximate geometry. To take the real geometry of a
part into account, surface geometric variations have to be limited.
Tolerancing is the means that help designers express these variations.
120 Part Three Modelling geometrical error
Among the varied existing theories on to1erancing of parts, some have emphasized
mathematical specifications, e.g. A Raquicha's model of shifting (Raquicha, 1983). Others
have highlighted the transfer of functional conditions, such as A.Clement's method of
technologically and topologically associated surfaces (Clement, 1991) or M.Giordano's and
D.Duret's model of gap spaces (Giordano, 1993). Though the varied aspects of tolerancing are
all connected, we will concentrate here more on the choice of tolerances.
2 TOLERANCING: NATURE OF THE PROBLEM
The tolerancing of parts of a mechanism must ensure the assembly of the mechanism and the
interchangeability of parts.Tolerancing thus has to satisfy two constraints: to respect the
functional requirements that set limits for the relative positions between parts, and to take the
variations induced by the manufactUling into account.
Our analysis of the way geometric specifications are obtained will therefore start by the
analysis of the transfer of functional conditions on the parts of a mechanism.
This transfer is usually calculated with a one-directional model called calculation of tolerance
chain.
So as to explain the limits of this one-directional model, we will consider an example of a
mechanism. This study will allow to establish the characteristics necessary for a tri-dimensional
model of variations that will then be introduced.
2.1 Introductory example
The mechanism represented in figure I has several links plane/plane, shaft/hole and two
tolerance chains. The orthogonal morphology of the part lets believe that it can be processed
with a one-directional tolerance chain. Actually, there is no a-pliOli coupling between vmiations
belonging to 0I1hogonai directions. An independant processing of tolerance chains along both
directions then detelmines dimensional specifications of each part. In this modeL the geometry
of the real pm1 is compared with a set of double points or with parallel planes.
Figure 1 Writing of tolerance chains relative to functional conditions

Geometrical behaviour laws 121
Figure 2 represents the same mechanism with variations on each part, voluntarily amplified
for the sake of demonstration. The real surfaces are modelized by surfaces of the same nature,
tangent to nominal surfaces on the free side of the material. Compared with the ideal geometric
model, each surface hence shows position, orientation and individual variations (radius
variation of the shafts and holes of the example).
-+--+_C1
Figure 2 Representation of the mechanism with geometlic valiations
2.2 Limitations of one-directional models
The one-directional tolerance chain does not take OIientation valiations of surfaces into account;
the proposed example, despite its simplicity, allows to see two defects of one-directional
models.
Connection between variations from d~tlerent directions.

The study of gap, direction after direction, and the olthogonal morphology of the part result in
the independance of the specitications of both directions.
In figure 2, it is on the contrary shown that there is an interdependance between the value of
gap C 1 and the OIientation of plane 3 in relation with plane 4; this orientation does not only
depend on the tlat mating but also on the configuration of the shaft/hole mating.
This figure 2 also illustrates another phenomenon: the intluence of the relative position of
both parts in the assembly. Actually, if one considers the shaft after a half turn rotation, gap
changes value. In the case of turning parts, this implies the research of the most unfavorable
values and in that of a fixed jig, the possibility to look for a position that would minimize the
intluence of variations. One-directional methods never take this type of relationship into
account.
The example also shows that there is practically no mechanism that can be processed as a
purely one-directional problem.
Translation (~l results into normalized .\]Jecitlcations.

The results of one-directional tolerance chains can be rather easily translated into dimensional
specifications or into simple position specitications. This type of specitication nevertheless only
represents a part of the potential geometric specifications. The specitications of orientation or
localisation, through a reference system cannot be obtained with such a model.
2.3 Parts motion under mating variation.
The calculation of the value for a functional condition between two surfaces of different parts in
a mechanism requires the calculation of the position of these surfaces. This requires a
determination of the position of each part in relation with the variations of the surfaces in
contact with the other parts of the mechanism. Geometric variations are thus propagated from
part to part throughout the mechanism.
The function of calculation of the consequence of variations or assembly with variations is
essential in the problem of tolerance chains, as have stressed F.Schneider and J.Remy-Vincent
(Schneider, 1993). Numerous works have actually been dedicated to these questions and have
offered different approaches:
• Case study is the solution chosen by the S.A.T.T method of A.C1ement (Clement,
1991). The function of the calculation lies upon the research for Minimum
Reference GeometIic Element~.
• Applying rules to compute the suppression of degrees of mobility (Turner, 1992).
• K.Takahashi has imagined an algorithm to assemble parts with variations (Takahashi,
1993). The Held of application for this algOlithm is restticted to nat surfaces which makes it
perfectly suited for the study of polyhedral parts.
• An algorithm of optimization by linear programming has been suggested by
J.Turner (Turner, 1990) and completed by R. Sodhi (Sodhi, 1994). These
calculations procedures of behavior have to be applied to each case with variations,
which leads to a calculation of tolerances by sim ulation.
Nevertheless, there is no systematic fOlmal method that deals with this problem. We are then
going to rely on the existing works and our analysis of the example to propose a fOlmal model
of variations for geomeuic toleraneing.
3 MODELIZATION OF VARIATIONS ON SURFACES OF PARTS
The model developed is founded upon vaJied hypotheses:

• First, the conservation of the typology of the nominal surface to modelize the real
surface. A real surface nominally nat will then be represented by a plane.
• The "perfect" surface representing the real surface will be tangent and out of the
material in relation with the real surface of the part. This hypothesis allows for the
assembly of PaJts.
• The little amplitude of the variations regarding the nominal dimensions of the parts
allows to modelize the transformation, which associates the surface of the real
model to the nominal surface by a small displacement torsor.
• The motion of the part~ under the effect of their variations is also a small amplitude
motion and can be modelized by a small displacement torsor.
3.1 Two types of variables for variations
The surfaces traditionally used for mechanical link often have pmticular invariance propelties by
translation along an axis or rotation around an axis. A plane is globally invariant for all
translations of a colinear vector to the nominal plane and for all rotations on a colinear axis to
the nominal plane. This property of invariance for certain usual surfaces such as plane,
cylinder, sphere, is all the more easily checked for small displacements.
Let us consider the torsor that modelizes the small displacement that connects a real surface
with its nominal surface. It is expressed under its canonic form, in an orthonormed reference
that includes the nOimal of the nominal plane, according to what follows:
fa u\
ITReaUNominaA =\ f3 v f (1)
y w ()
The components of the small displacement torsor that leave the reference plane globally
invariant are called undetennined components Ind.
For example, the small displacement torsor of the considered plane is:
l
f a Ind" \
ITReallNolllinal) = \ f3 In £I,}, f (2)
\ Inil,.: w ()
The components of the transformations of the small displacement that globally leaves the
surface invariant are generally arbitrarily fixed, 0 for instance. This () value is ambiguous for it
hides the notion of invmiance in relation with tme () vatiations.
As these components of small displacements are always undetermined, in the reference
associated with the surface and in relation with the considered surface, it means that they have
the following propelties:
Va E 9\; a + Ind = Inti V a,b E 9\2; a.Ind + b.lnti = Inti (3)
In the model we are considering, these components are considered as undetermined
parameters as regards the considered displacement. Hence, they are exogenous parameters as
regards a small displacement but not necessarily exogenous as regards the whole part or
mechanism. This property will he used hy the operators of the model.
3.2 Small displacement torsors of the model
The characterization of the part variations, of the gap between parts as well as of the small
displacements of parts, under the effect of variations, and as regards their theOl'atical relative
position, depends on a set of torsors that we are going to define.
Deviation torsor
A deviation torsor expresses the variation between a nominal surface and a surface of the same
nature by modelizing the real surface. This torsor is associated with a surface the same way a
L\e is associated with each surface of a part, (Bourdet, 1973). The deviation torsor is built upon
the composition of two torsors: a small displacement torsor for the surface and a torsor of the
characteristics variation of the surface. The characteristics variation torsor takes into account the
variation of the radius of a sphere or the top angle of a cone, for instance.
In the case of a z direction cylinder, it yields:
fa u \ fO dr cos(8)
/ TRealiNollllnad = \f3 v f + \ () dr sin(8)

Indrz Indtz 0 0 o
fa u+d,. cos(8)
\
/ T RellllNolll;rwtl =\f3 v+d,. sine 8) (4)
Ind,z Ind,z 10
p is a point of the surface parametered by e
The characteristics variation torsor may be zero when the surface does not have specific
characteristics, as is the case for a plane.
One has to notice that the components of the characteristics variation tors or which are not
used, are equal to zero as the taking into account of in valiance by motion is done at the level of
the small displacement torsor of the surface.
The deviation torsors thus modelize position, orientation and specific variations of each part
surface.
Variation torsor
A variation torsor expresses the relative variation between two or more real surfaces of a one
part. In the case when it expresses the variation between two surfaces, the variation torsor can
be calculated by the composition of the deviation torsors of each surface. This is done with a
common reference on both surfaces and with the application of the property (3). To build such
a common reference, one can refer to the definitions of EGRM presented in chapter 2 of D.
Gaunet's thesis (Gaunet, 1994), or to the operator of intersection of deviation tields presented
in paragraph 4.2.
The following formula (5) is an example of a variation torsor between two nominally non
parallel planes.
/TPlaneIlPhme2} = /TPlaneIlPart} -/TPlaneIlPart}
fa(l/2) [nd/ x ( 112)

(5)
/ TPlano IIPlano2} = Ind (1I2) Indty (112)
\ 'Y
Indrz (112) Ind,z( 1/2)
Gap torsor
A gap torsor expresses the gap between two surfaces of different parts which are nominally in
contact. There will be a gap torsor associated to each couple of surfaces of a mechanism that
makes a contact. The gap torsor then only concems two surfaces that belong to different parts.
The definition of the gap torsor requires the same operator as before but in a different
application.
Actually. the part considered here is an immaterial wedge that fonnalizes the gap between the
two parts. The operator for the research of the intersection fields of deviations is applied to this
virtual part only made up of two surfaces. Then. for instance. the gap torsor of an axis x
cylinder in contact with a nOlm plane in z can be detelmined in the fOlmula (6):
Irx
(TP/andCylinderl ={ Ind,}' (6)
l11dr1
where I represent~ the components of the gap and Ind the undeterrmined components of the
mating.
Small displacement torsor per part

The small displacement tors or associated with each pan gaps is central in this model. It actually
establishes. through all the composition chains of all the small displacements. the connection
between all the deviations of a pan. It will then be used either to express the displacement of the
model of the real pan in its assembly with the nominal model. or to calculate the small motion
of a part in the mechanism. function of the surface variations of the part in contact with the
adjacent surfaces and the small displacements of the neighboming patts.
Each part will have its associated small displacement torsor in relation with a set of
references :
small displacement of nominal part

A
(7)
All these torsors yield the elements necessary to describe the mechanism. We are now going
to define the operators that handle these different torsors so as to answer the requirements of
geometric tolerancing.
4 OPERATORS LINKED WITH THE MODEL.
To solve problems of tolerancing by tolerance chains one must be able to answer two questions:
what is the set of possible dimensions for each part'! what are the significant positions of parts
in the mechanism with variations?
Before showing the answer that we can bling to these two questions, let us develop on the
composition of small displacement~ as shown by an elementary mating.
4.1 Small displacement resulting from an elementary mating
The three torsors of deviation, gap and small displacement per part are connected by a
composition relation of small displacements. For the model, an elementary mating is based on
the couple of two surfaces in contact with each other. These surfaces may be simple surfaces
such as plane or cylinder or sets of surfaces; nevertheless, a single deviation torsor will be
associated, by part, with each surface or set of surfaces. Figure 4 illustrates the respective
positions of the valied torsors that act in the modelization of a mating.
TB/R Small
Nomi ... l Surfa~.: 2 Part B ..... tlisplrocemem
b--
lorsur p.1r1 B
Re:~ surface
R.:al surface I ~
omill:~ Surface 1 ,,-
~TNR Small
~isJllacemcnl
lorsur p:U1 A
Part A
Figure 3 Small displacement torsors of a mating
The composition equations of the small displacement torsors of the elements that build up the
mating can be wlittenas follows, to express the motion of pmt B in relation with the mating
variations:
deviation torsor deviation torsor
real surface 1 -----..... real surface 2/
TBIR = TAIR + TIIA - T21B +T1/2

(8)
_ _ , 1
Small deplacement Small deplacement r
torsor part B torsor part A Gap torsor
This relation can be applied similarly for the modelization of either the contact between two
parts of a mechanism or that between a real surface and its nominal model.
4.2. Intersection between deviation fields.
The common deviation field between two surfaces or between a set of surfaces is the set of
characteristics that can be specified. Fonnula (8) allows to connect the displacement of the real
part as regards its nominal model, and that for each surface.
The characteristics that can thus be specified are the common characteristics of this set of
elementary mating. The intersection operator can be built upon the principle of the unicity of the
positioning of the part as regards the nominal and for the set of considered matings. The
following operator then can be defined:
T(Pllrt) _
PIR -
(j T(f".)
PIR (9)
i=l
where ei is the displacement due to the mating ei.
This operator is concretely defined by the unicity of the displacement torsor of the real part as
regards its nominal model. The application of the Gauss elimination method to the linear
equation system that can be formulated after this principle, shows a set of constraints which are
the required conditions for the existence of a global small displacement torsor for the whole
part. These conditions are actually the expression of the tridimensional dimension of the part.
Variation coefficients then indicate the reference of expression of the variation torsor in its
canonic shape.
The tolerancing of a part is equal to the determination of conditions that constraint the model
of the real part on the nominal model. This principle, translated by the operator of the
intersection field of deviations gives all dimensioning opportunities between n surfaces (122) of
a one part. The application of this operator to the gap between two parts calculates the
parameters of the gap torsor.
The implementation of this operator has also shown the combinatorics of the possible
specifications. There is not one dimensioning solution but a sometimes rather important set of
minimal specifications that are simultaneously possible for one part.
4.3 Union of setting chains
The principle on which the model is built is to associate a deviation torsor to each surface of
each part. The links between parts are then cut up into couples of mating faces. It is therefore a
series of local descriptions, that then have to be related to see the global behavior of the
mechanism.
Each elementary mating allows to write the composition of small displacements of the chain
relative to the mating. To calculate the displacement of the part relative to all matings, the union
of the set of small displacement5 for each mating has to be realized. An operator that can
translate the following definition has to be found:
T(Pllrt) _
PIR -
.:J T(ePIRi)
(10)
i=l
where ei is the displacement due to the mating ei.
The definition of the union operator uses the property of unicity of the relative motion of the
part. It then yields:
\..I. ··T(e.) _ T(e;)
v IJ, PIR - PIR (11)
The research for the small displacement of the part due to vUliations leads us to consider this
equality of torsors. After an expression at the same point, as a system of 6( n-l ) equations, it
will be solved in relation with the undetelmined values of these matings.
The solution for this system of linear equations of sma)) displacements is obtained by
applying the Gauss elimination method. This algorithm has been altered so as to fit the
characteristics of the systems encountered with this model, i.e systems that can be
simultaneously over and under determined.
As in the case of the intersection operator, there is a combination of possible positionings for
each part. This is directly linked to the isostatism or the hyperstatism of the positioning of the
part. The isostatic positioning of a part only yields one configuration. But the hyperstatic
positioning yields a set of configurations that can then be expressed.
The consequence of variations in connected sets has thus been formally described. The
operators have been defined in a formal mathematical programming language and applied to
numerous cases. To illustrate part of the results ohtained by the model, we are now going to
apply it to an example.
5 EXAMPLE
The mechanism in figure 4 will illustrate the concepts developed. Both parts of the mechanism
are joined to fOlm a prismatic link. The possible combinations of the mobile palt in relation with
the bar are being searched, with the union operator of displacements. A deviation torsor is
therefore associated with each surface of the mechanism, here, each plane. A sma))
displacement torsor per part completes the modelization. The description procedure is
systematic, which allows to process this link or a whole mechanism the same way; only the
time of computation vmies.
f%
1-
(part B)
h
--+_ .......- -
?C.
(part A)
mobile
I bar
-
Each part of the mechanism is defined by:
mobile = ({1.(t.O,O) , (O.O,Pi), Plan), Coordinates point angles between referential surface
of deviation torsor and referential mechanism
(2,{- t ,o,O). (O,O,O), Plan).
Surface {1, {t,O,O}, {O,O,Pi}, Plan}
(3,(O,O,h), (Pil2 ,Pil2 ,O). Plan) number t Surface nature 1
Figure 4 The mechanism and its systematic description procedure
Determining the gap torsors is the result of the application of the algorithm of intersection of
the deviations fields for each link plane/plane. Then, for instance, the gap torsor of the mating
between horizontal planes can be:
, Jrx Ind tx \
(TMobile plane/bar plane) = \ J ,y Indty ,. (12)
bulrz Jtz 0
The assembly of the two paIts is done through :I links plane/plane.

The union operator of motions due to each mating calculates the small displacement of the
mobile palt in relation with the bar by elimination of undetelmined values; it also calculates the
varied configurations of the mechanism. These configurations, systematically listed, are due to
the system of 4 in dependant linear equations with 7 unknowns:
/3{lIA)+j3(2IA) + j3( 4IB)+/3( 51B )-J,y(411)+Jry( 512)= () (13)

- j3(2IA)-j3(5IB)+J{ 3IA)-J{6IB)-J,y(512)+Jry(613)= () (14)
- J{ 1/A)-J{2IA)+J{41B )-J{ 51B )-Jrz{ 411)+1,-z( 512)= () (15)
z]1ry(411 )-z2J,y( 512)-y ]1,.z(411 )-Y21,-z(512)-Jtx!411 )+JtJ512)+u( 1/A)+ ... +u( 5IB)= () (16)
there are then 3 among 7, i.e 35 potential contigurations.

The configurations are directly obtained by substituting the () value to each set of gap
components identitied, so as to show contact. This contact can be full or pal1ial as is shown in
figures 5 and 6.
The algorithm lists 9 distinct configurations. A partial illustration of the results of the
computation of configurations is given in figures 5 and 6. The systematic enumeration of these
configurations makes sure that, when calculating the vaIiation chains due to each configuration,
one actually deals with the most unfavorable case. If all configurations do not reflect the
working of the mechanism, one only has to select amidst the built set, those that are functional.
For instance, the three configurations in which the horizontal plane lean is dominant.
Figure 5 Examples of configurations with amplified vaIiations (front view)

Figure 6 Examples of configurations with amplified vaIiations (bottom view)
The application of the operator of the union of the mating displacements also shows the
mobility degrees by the remaining of undetermined components in the small displacement
torsors per part; here, it is the axis of the ptismatic link built in the mechanism.
f -[3(3/A)-[3(6/B) z2([3(2/A)+[3(S/B» -y2('X3/A)+'X6/B»+u(2IA)-u(SIB) \
(TpartAIPanB) =\ 'X3/A)-'X6/B) Indetermi.nate f
'X2IA)-'XS/B) -u(3/A)-u(6/B) 0
Despite the complexity of the fonnal calculations at stake, the formalization of this example is
systematic and does not present any difficulty. A link with a geometric modeller is actually
being developed so as to suppress this phase of computation of the model.
6 CONCLUSION
The model presented here formally links the varied types of variables that variation chains
represent. Our approach is founded upon a local description of the behavior of each mating.
The global charactetistics of the mechanism are then searched by the operator of union of small
displacements. Establishing links between the varied torsors of small displacements is made
possible when identifying two types of deviation vaIiables.
The definitions of the tor50rs and of the operators are a first step in the formalization of
tridimensional tolerance chains. These will then be obtained by the composition of the small
displacement torsors of the surfaces that contain mating variations. The model, thanks to its
general character, can nevertheless be thought of as applicable to manufacturing dimension
chains.
7 REFERENCES
Bourdet P.(J973) Chaine de cotes de fabrication: Ie modele. L'ingenieur et Ie technicien de
l'enseignement..
Clement A., Desrochers A.and Riviere A.(1991) Theory and practice oL'l-D tolerancing for
assembly. CIRP International Working Sel11inar on Tolemnc:ing, Penn State, pp. 25-55.
Gaunet D.(1994) Modele formel de toh~rancement de position Contributions a I'aide au
tolerancement des mecanismes en CFAO. These de Doctorat. La/Joratoire de Mecatronic de
17SMCM - ENS de Cachan.
Giordano M.and Duret D.(1993) Clearance space and deviation space. Application to three
dimensional chain of dimensions and positions. (Espace jeu. espace ecart). 3rd CIRP
Seminar on Computer Aided Tolerancing. pp. 179-196.
Requicha A.A.G.(1983) Toward a Theory of Geometrical Tolerancing. The International
Journal (~lRobotics Research, Vol. 2, No.4, pp. 45-60.
Roy U., Liu C. R.and Woo T. C.(l99l) Review of dimensioning and tolerancing:
representation and processing. Computer-aided design, Vol. 23, No.7, pp. 466-483.
Schneider F.and Remy-Vincent 1.(1993) The assembly function applied to the definition,
control and manufacturing of mechanical components. 3rd CIRP Seminar on Computer
Aided Tolerancing. ENS de Cachan, pp. 129-146.
Sodhi R.and Turner J. U.(l994) Relative positioning of vatiational part models for design
analysis. Computer-aided design, Vol. 26, No.5, pp. 366-378.
Takahashi K., Suzuki H.and Kimura F.( 1(93) Motion analysis of parts with geometrical
enors based on dynamic simulation. 3rd CIRP Seminar on Computer Aided Tolerancing,
ENS de Cachan, pp. 85-96.
TurnerJ. U., Srikanth S.and Gupta S.(l9!)2) Constraints representation and reduction in
assembly modeling and analysis. IEEE Journal oj"Ro/Jotics and Automation, Vol. 8, No.
6, pp. 741-750.
Turner J. U.(1990) Relative positioning of parts in assemblies using mathematical
programming. Computer-aided design, Vol. 22, No.7, pp. 394-400.
8. BIOGRAPHY
P. Bourdet is Professor. ex-head of the Department of Mechanical Engineering, and

Director of the Laboratoire Universitaire de Recherche en Production Automatisee of Ecole
Normale Superieure de Cachan. France. He has authored and co-authored over 50 research
papers. He is a member of the CIRP and he has special interest in tridimensional metrology,
tolerancing and the designing of advanced manufacturing systems.
E. Ballot holds a Ph. D. from the Ecole Normale Superieure de Cachan. He is a member of
the Laboratoire Universitaire de Recherche en Production Automatisee of Ecole Normale
Superieure de Cachan, France. He is a former student of the Ecole Normale Superieure de
Cachan and holds an Agregation in Mechanical Engineering. His research interests include
process planning, computer-aided tolerancing synthesis and analysis.
9
Configuration space based analysis
of position uncertainties of parts in
an assembly
Masatomo Inui and Masahiro Miura
Department of System Engineering, Ibaraki University
Naka-narusawa-cho 4-12-1, Hitachi-shi, Ibaraki 316, JAPAN
Phone: +81-294-38-5204, Fax: +81-294-32-1546
Email: inui@hit.ipc.ibaraki.ac.jp
Abstract
A configuration space based method for analyzing position uncertainties of two parts in
an assembly is proposed if certain shape variations of individual parts are allowed by
tolerances. Proposed method is applicable to 2 dimensional polygon models of machine
parts, for example sections of polyhedron parts in an assembly. The analysis is achieved by
calculating a variation bound of the configuration of the parts if nonnominal parts allowed
by the tolerance satisfy the positioning condition in an assembly. Different from the prior
approach, our method does not assume any specific deviation types of the nonnominal
part. An algorithm for computing the variation bound is developed and the configuration
uncertainty of some sliding joints in the positioning mechanism is analyzed. Computa-
tional experiments suggest that the visualized image of the bound of the configuration
variation is helpful for understanding the positioning characteristic of the joint.
Keywords
Tolerance modeling, tolerance analysis, configuration space
1 INTRODUCTION
Any machine parts have geometric deviations due to their production. The deviation
affects position and orientation of parts in an assembly. It consequently causes various
functional uncertainties of mechanical products. Machine designers must control such un-
certainties by specifying ranges of allowable deviations of individual parts as tolerances.
Since the accuracy is of primary concern in machine design, tolerances should be ideally
as close to zero as possible. However, tolerances given too tightly inevitably raise the
manufacturing cost and reduce product competitiveness in the market. Tolerance opti-
mization is thus important to ensure the acceptable functional and economical quality of
the product. Despite the importance of the proper specification of the tolerance, machine
designers in general still rely on their experience and costly trials and errors. Much efforts
Configuration space based analysis 133
are needed in developing appropriate technologies for computer assistance of the toler-
ance optimization. Position analysis of assembled parts is one of such technologies, which
usually means the determination of the configuration of parts in an assembly if they have
deviations allowed by tolerances. Because of geometric uncertainties of individual parts,
the configuration of the assembled parts also has some uncertainties.
Although many algorithms have been developed for positioning machine parts based on
their assembling constraints, most of them assume the parts having their nominal shape
(for example (Ambler, 1975), (Lee, 1985), (Liu, 1990)). Relatively small number of stud-
ies has been reported for positioning parts if they have deviations. Fleming proposes a
method for deriving inequality constraints on configuration parameters of assembled parts
based on the analysis of extreme configurations of two parts in contact (Fleming, 1987).
Symbolic SUP-INF method solves the constraints to derive a bound of the parameters.
Takahashi et al. consider only vertex-face contacts are possible between actual nonnom-
inal parts (Takahashi, 1991). They determine the assembled configuration of two parts
by trying to realize I-point contact, and then successively moving one component part
to realize additional greater point contacts. Turner formulates the positioning problem of
polyhedron parts as a mathematical optimization problem (Turner, 1990), (Sodhi, 1994).
Based on a vertex-face contact between nominal polyhedrons in an assembly, a noninter-
ference constraint of a vertex of a nonnominal polyhedron is derived as a condition of the
vertex being in the outside half-space of its mating face. Within the feasible space of the
part configuration defined by the constraints, other conditions are maximally satisfied by
using the mathematical programming method and part positions are determined. Geo-
metric deviations of nonnominal parts are simulated by position and orientation variations
of planar faces of the nominal part in his study. Acceptable types of deviations must be
extended to more precisely evaluate the effect of various manufacturing errors.
In this paper, the authors explain a configuration space based method for analyzing
relative position uncertainties of two parts in an assembly. In most mechanical assemblies,
part positioning is carried out in a sequential manner (Wilson,1994). Our method thus
can be applied to more complex mechanical products if their assembling process is de-
composable to a set of primitive positioning of two parts or subassemblies. This method
is currently applicable to 2 dimensional polygon models of machine parts, for example
sections of polyhedron parts in an assembly. This might not be a severe restriction of
the method because tolerances are usually specified to 2 dimensional projection figures of
the part. The authors' algorithm is based on the same concept of Turner's, so that part
positions are determined by optimizing a certain objective function within the feasible
space of the part configuration. Different from the prior approach, our method does not
assume any specific deviation types of the non nominal part.
2 BACKGROUND
As a preparation of the following explanation, some background knowledge is given.
2.1 Configuration space

Consider an assembly of two polygons M and 13 in a Euclidean workspace W = R2. M
and 13 are models of nonnominal parts obtained by giving arbitrary but small geometric
.-k-Mn
M
Figure 1 Two assembled polygons M and 13 with small geometric deviations and their
original nominal polygons Mn and 13 n.
deviations to edges of their corresponding nominal polygons Mn and 13 n (see Figure 1).
The local coordinate frame of M is assumed to be identical to the coordinate frame
F Mn embedded in the nominal polygon Mn. The local coordinate frame of 13 is similarly
fixed at the local coordinate frame FBn of 13 n. Because of the above placement of the
coordinate frames, positions of M and 13 in the global coordinate frame Fw of Ware
uniquely determined when positions of the nominal polygons Mn and 13 n in Fw are given.
Therefore, the positions of Mn and 13 n are only considered hereafter.
Since our interests are relative positions of two polygons in an assembly, either one of
Mn and 13 n can be fixed at arbitrary place in Fw without losing the generality. 13 n is fixed
in our case as its local coordinate frame F Bn becomes identical to Fw. A configuration
of Mn with respect to Fw (= FBn) can be specified by 3 parameters x, y, and o. x and
y represent the position of the coordinate frame F Mn in Fw. 0 indicates the rotation of
F Mn relative to Fw. Since allowable geometric variations of individual parts are small,
variations of their position and orientation in an assembly are expected to be small also.
Configuration parameters, x, y, and 0 thus satisfy the following inequalities; Xn - Llx :::;
x :::; Xn + Llx, Yn - Lly :::; y :::; Yn + Lly, and On - LlO :::; 0 :::; On + LlO where Xn> Yn,
and On represent the ideal position and orientation of Mn in F w , and Llx, Lly, and LlO
mean sufficiently small values. By introducing new parameters 5x, 5y, and 50 satisfying
15xl :::; Llx, 15yl :::; Lly, and 1501 :::; LlO, x, y, and 0 can be expressed as x = Xn + 5x,
y = Yn + 5y, and 0 = On + 50. Instead of x, y, and 0, new parameters 5x, 5y, and 50 are
used for specifying the configuration of Mn in the following discussion. The subset of W
occupied by Mn at configuration q = [5x, 5y, 50] is denoted by Mn( q). The configuration
space of Mn means the space C of all the possible configuration q of Mn. According to the
definition, C can be represented as a small rectangloid [-Llx, Llx] x [-Lly, Lly] x [-LlO, LlO].
2.2 Tolerance model
In machine manufacturing, deviations of nonnominal objects such as M and 13 are con-

trolled by a set of tolerance assertions. The basic idea of tolerancing is to define a range
of allowable shape variations of nonnominal objects with respect to the nominal geometry
M(Mn(q»
L(Bn)
M(Bn)
Figure 2 Spatial relationships between maximum-material polygons M(Mn(q» and

M(Bn), least-material polygons C(Mn(q» and C(Bn), and nonnominal polygons M and
B with allowable deviations.
(ISO, 1983). Based on the tolerance model proposed in (Requicha, 1983), the authors
define the spatial condition of the allowable variation in the following manner.
Let M(Mn(q» and C(Mn(q» respectively bea polygon in the maximum-material
condition and a polygon in the least-materi~l condition allowed by tolerances in asso-
ciation with the nominal polygon Mn at q. The maximum-material polygon means a
polygon to which any material cannot be added without violating the tolerance. The
least-material polygon corresponds to a polygon from which any material cannot be re-
moved. The form, position, and orientation of M(Mn(q» and C(Mn(q» are assumed
to be fixed with respect to the local coordinate frame of Mn( q). Based on the nomi-
nal polygon Bn and its associating tolerance assertions, the maximum-material polygon
M(Bn) and the least-material polygon L.:(Bn) are defined in the same manner. Geometric
deviations of nonnominal polygons M and B are judged acceptable if they can satisfy the
following spatial conditions (see Figure 2);
M(Mn(q)) ~ M(q) ~ C(Mn(q»

(1)
M(Bn) ~ B ~ C(Bn)
2.3 Positioning conditions

Machine parts in an assembly are considered to be placed by maximally satisfying some
constraints relating to the machine function within the feasible configuration space in
which the parts do not interfere. The configuration of nonnominal polygons M and B in
an assembly thus must satisfy the following two conditions;
Noninterference condition: M and 13 do not collide, in other words M and 13 do not

occupy the same point in W.
Optimization condition: The configuration of M and 13 must optimize a certain ob-
jective function concerning the assembling method, for example the maximal proximity
of attached contacts.
Since the configuration of M and 13 and the configuration of their original nominal poly-
gons Mn and 13 n are identical in our definition, the configuration of Mn and 13 n is opti-
mized instead of the configuration of M and 13. The optimization condition is thus defined
as the minimization of a certain algebraic function z( q) on the configuration parameter
q = [5x, 5y, 5B] of Mn with respect to 13 n . Since 5x, 5y, and 5B are small, z( q) is linearized
by ignoring its second and higher order terms.
3 POSITION UNCERTAINTY ANALYSIS METHOD

Position uncertainty analysis is formulated in our study as a problem of computing a vari-
ation bound of the configuration q of Mn if some nonnominal polygons M and 13 allowed
by the tolerance condition (1) also satisfy the positioning conditions in an assembly. The
following two theorems are derived concerning the variation bound of q. See (Inui, 1995b)
for proofs of the theorems.
Theorem 1 Let q be the configuration of Mn. If q is within the following set of config-
urations Sni;
(2)
then, some nonnominal polygons M and 13 allowed by the tolerance do not collide.
Figure 3 illustrates the whole configuration space C. Union of the light gray region and the
white region in the figure corresponds to the configuration subspace Sni defined above.
Theorem 2 Let qmin be the configuration of Mn giving the minimum value of the linear
objective function z( q') in the following configuration subspace St;
(3)
The configuration q of Mn must be contained within the following set of configurations

So to satisfy the optimization condition;
(4)
The white region in Figure 3 represents a set of configurations St in C. The configuration

qmin is depicted by a small circle in the figure. Configuration subspace corresponding
to So is illustrated as a half-space sectioned by a plane (a straight line in the figure)
representing z( q) - z( qmin) = O. Small arrow means the normal vector of the plane.
Figure 3 Configuration subspace Sni, St, and So in C.
Because of the above definitions, the intersection of the configuration subspace Sni and
So corresponds to an allowable variation bound of the configuration q of Mn. Precise
algebraic computation of the configuration space is known hard, therefore some approx-
imations are usually needed in the configuration space construction (see for example
(Latombe, 1991)). In the next section, the authors explain an algorithm for computing
the configuration subspace Sni n So in the octree representation. The algorithm is im-
plemented and applied for computing the configuration uncertainty of some assemblies
if certain shape variations of component parts are allowed by tolerances. Computation
results are given in the fifth section.
4 POSITION UNCERTAINTY ANALYSIS ALGORITHM

The objective configuration subspace Sni n So is computed in the following three steps;
1. Computation of the configuration subspace Sni in the octree representation.

2. Search of the optimum configuration qmin in the configuration subspace St.
3. Intersection of the configuration subspace Sni and So = {qJz(q) :S Z(qmin)}.
4.1 Computation of Sni in octree representation
Sni is a set of configurations q E C satisfying the condition C(Mn(q)) n C(Bn) = 0. The

following recursive algorithm derives Sni in the octree representation.
(a)
E§2f L (B .)
(b)
Figure 4 Cell classification by using the outer swept area OS.c(M)[lI:j and the inner swept
area IS.c(M)[K].
The whole configuration space C is first decomposed into 8 rectangloid cells of equal
size and their edges are parallel to the axes of the space. Let II: be a decomposed cell in
the form [8x},8x2j x [8Y1,8Y2] x [881 ,88 2] £;; C. II: is labeled either full, empty, or mixed
according to the following rules (see also Figure 4);
(1) If polygons .c(Mn(q)) for all q E II: do "not" collide .c(13n ) then II: is labeled full. This
condition is satisfied if the outer swept area OS.c(M)[lI:j = UqE".c(Mn(q)) does not
collide .c(13 n)(case (a) in Figure 4).
(2) If polygons .c(Mn(q)) for all q E II: collide .c(13n ) then II: is labeled empty. This condi-
tion is satisfied ifthe inner swept area IS.c(M)[lI:j = nqEK .c(Mn( q)) collides .c(13n)(case
(b) in Figure 4).
(3) Otherwise II: is labeled mixed.
The cells labeled mixed are further decomposed into smaller cells with appropriate
labels. The decomposition process continues until all the cells are labeled either full or
empty, or all mixed cells become smaller than some prespecified size. Such smaller cells are
labeled full in our study so that the computed configuration subspace contains the correct
subspace Sni within. An approximated representation of the configuration subspace Sni
is obtained consequently as a set of cells with full label. These cells are hierarchically
arranged in the octree structure.
Some algorithms have been reported for computing the swept area OS.c(M)[lI:j and
IS.c(M) [lI:j. For example, Zhu and Latombe develop a method which traces the contour
of OSC(M)[lI:j and IS.c(M)[lI:j in time O(n 2 ) where n is the number of edges of .c(Mn(q))
(Zhu, 1990). Xiao proposes another method for computing OS.C(M) [I>:] (Xiao, 1993). In our
configuration uncertainty analysis problem, the size of the configuration space C is very
small. Based on this characteristic of the problem, the authors develop a bounding/bounded
polygon based algorithm for hierarchically approximating the configuration space. This
algorithm can determine the label of a cell more efficiently than the swept area based
algorithm mentioned above. See (Inui, 1995a) for details of the algorithm.
4.2 Spatial search algorithm of qmin
qmin is the configuration giving the minimum value of the linear objective function z( q)
in the configuration subspace St = {qIM(Mn(q)) n M(Bn) = 0}. The structure of the
search algorithm of qmin is analogous to the algorithm for constructing Sni. The whole
configuration space C is again decomposed into 8 rectangloid cells of equal size. In this
time, the label of a cell I>: is determined by the following rules;
(1) If polygons M(Mn(q)) for all q E I>: do not collide M(Bn) then I>: is labeled empty.
This condition is satisfied if the outer swept area OSM(M)[I>:] = UqEK M(Mn( q)) does
not collide M(Bn).
(2) If polygons M(Mn( q)) for all q E I>: collide M(Bn) then I>: is labeled full. This condition
is satisfied if the inner swept area ISM(M) [I>:] = nqEK M(Mn(q)) collides M(Bn).
(3) Otherwise I>: is labeled mixed.
The condition M(Mn(q)) n M(Bn) = 0 is satisfied if the configuration q corresponds

to an interior point or a point on the boundary of an empty cell 1>:. Because z( q) is linear
and I>: is convex, the configuration giving the minimum value of z in I>: must be one of eight
"vertices" of 1>:. For example, if coefficients a, b, and c of z( q) = a 8x + b 8y + c 80 + d are all
positive, then a vertex [8Xl' 8Yl, 801 ] of a cellI>: = [8Xl' 8X2] X [8Yl' 8Y2] X [801 , 8(J2] gives the
minimum value in 1>:. The optimum configuration qmin in St can be detected by visiting
empty cells and mixed cells of C in the depth-first search manner. If the visited cell is
labeled empty then the objective function is evaluated with the configuration minimizing
z in the cell. If the function evaluation at the current cell gives a smaller value than the
minimum value obtained in the prior cell visit then the configuration referred to in the
evaluation is (temporally) recorded as qmin' When a mixed cell is encountered, the cell is
decomposed into smaller cells, and the search procedure is recursively applied to them.
By using the spatial hierarchy of C, some subtrees of C can be excluded from the visit.
When a mixed cell is encountered, the objective function is evaluated with the optimum
configuration in the mixed cell before the decomposition. If the evaluation result is equal or
worse than the optimum value obtained in the prior cell visit, then the cell decomposition
is not necessary because decomposed cells never give a better solution. Figure 5 illustrates
exclusions of some subtree visits with 2 dimensional configuration space C in the quadtree
representation. The objective function to minimize is z(8x,8y) = 28x + 8y. Because all
coefficients of z( 8x, 8y) are positive, a configuration corresponding to the left-bottom
vertex of an empty cell gives the minimum value of z in the cell. Visiting of a subtree
enclosed by a rectangle B in Figure 5(b) is not needed because the left-bottom vertex of
the mixed cell corresponding to the subtree (a square region B surrounded by bold lines in
Figure 5(a)) does not give a better solution than the minimum value previously obtained
["lli]
[ill]
visiting
order
A B
(b)
(a)
Figure 5 Exclusions of some subtree visits in searching the optimum configuration qmin'
at the empty cell A. Because of the same reason, another subtree enclosed by a rectangle
C is not visited also.
4.3 Intersection of Sni and So

The configuration subspace So is a set of configurations q E C satisfying the linear in-
equality constraint z( q) ::::: z( qmin)' Sni n So is thus computed by selecting configurations
q E Sn; which also satisfy the condition z( q) ::::: z( qmin)' Based on the hierarchical octree
structure of Sni, this selection process is efficiently achieved by visiting and relabeling
octree cells of Sni according to the following rules;
3. The label of a cell K is not necessary to change if z( q) ::::: z( qmin)' Vq E K. Because K is

convex, this condition is satisfied if eight vertices of K are all within the below side or
on the plane z(q) - z(qmin) = O.
2. The label of K is changed to empty if eight vertices of K are all within the above side
of the same plane.
3. Otherwise,
3.1 If the label of K is empty , do nothing.

3.2 If the label of K is either full or mixed, decompose K into smaller cells and apply the
same relabeling procedure to the resulting cells.
The objective configuration subspace Sni ns o is derived consequently as a set of full cells.
M(Mn(q))
t
L(Mn(q))
r======d
t;w.x ,r======u
L(Bn)
t
M(Bn)
Figure 6 Maximum-material polygons and least-material polygons allowed by tolerances.
5 COMPUTATIONAL EXPERIMENTS
The algorithm is implemented by using C language and several computational experiments
are done.
5.1 Position uncertainties of assembled parts
Figure 6 illustrates maximum-material polygons M(Mn( q)) and M(Bn), and least-material
polygons L(Mn(q)) and .c(Bn) based on our example nominal polygons Mn and Bn in
an assembly (nominal polygons are not illustrated in the figure.). The maximum-material
polygon is defined by externally offsetting supporting lines of segments of the nominal
polygon by 0.3. The least-material polygon is defined similarly by internally offsetting the
supporting lines by 0.3. Rather large tolerance values are specified to make the shape dif-
ference between the maximum-material polygons and the least-material polygons visible.
The local coordinate frame F Mn is specified at the center of gravity of Mn. z( q) = by
is used as the optimization function for positioning the parts so that the configuration
giving the minimum y coordinate of the origin of F Mn in Fw is searched. The solution
is considered to be the most stable position of allowable polygons in the assembly if the
force is statically applied to them in the negative y axis direction of Fw.
The configuration space C is set to be -1.6 S 8x S 1.6, -0.8 S 8y S 0.8, and -0.025 S
8() S 0.025 in radian. The maximum depth of the cell decomposition is specified to be 7.
Figure 7 visualizes a derived variation bound Sni n So ofthe configuration q = [8x, 8y, 80]
&: H I--- - " " '· Sni r. So .~- - -+ -1
..,
09 + ~!-.n>--~-~I·"'--~M'--~ITo"--'I~..r~
+ ~~------~--------~r,
(b)
(a)
Figure 7 Derived bound of position uncertainties of parts in the example assembly.
of Mn in the octree representation. The bound Sn; n So has a polygonal shape with 7
flat faces as illustrated in Figure 7(a) . Depth (oy) value of the bottom of the polygon
at [oO,oxj is illustrated by a gray-scale in Figure 7(b). Totally 4323.6 CPU seconds are
needed for the computation by using an SGI Indigo workstation with R3300 CPU (30
Mips, 4.2 Mflops) and 48 MB memory. Figure 8(a) shows two extreme configurations of
M(Mn(q)) and L:(Mn(q)) if they have the maximum Ox value and the minimum ox
value allowed in the bound. Figure 8(b) shows other extreme configurations with respect
to 80 in the bound.
5.2 Positioning characteristics of mechanisms

The visualized image of the bound of the configuration variation is helpful for under-
standing the positioning characteristic of the mechanism if small deviations of component
parts are allowed. Figure 9(a) illustrates nominal polygons Mn and Bn corresponding to a
section of a sliding joint with a V-slot. This joint is well used in the workpiece positioning
mechanism of conventional milling machines. The maximum-material polygons and the
least-material polygons are respectively defined by externally and internally offsetting the
Maximum Ox
Minimum Ox
(a)
Maximum {;e
Minimum {;e
(b)
Figure 8 Some extreme configurations allowed in the bound.

(a) (b)
Figure 9 Derived bound of the configuration uncertainty of the joint with a single V-slot.
nominal polygons by 0.05 (they are not illustrated in the figure.). The whole configuration
space C is set to be -0.2 ~ bx ~ 0.2, -0.2 ~ by ~ 0.2, and -0.005 ~ 88 ~ 0.005 in
radian. The same as our prior example, the local coordinate frame :FMn is specified at
the center of gravity of the upper part Mn. z( q) = by is again used as the positioning
function. A variation bound of the configuration is derived as shown in Figure 9(b).
Figure 10(a) illustrates other assembled polygons corresponding to a section of another
sliding joint with double V-slots. This joint is specially used if very precise positioning of
the workpiece is required. Using the same positioning function and the same configuration
space definition, a variation bound of the configuration is derived as shown in Figure
10(b). Because of the downward force of the gravity, the configuration of these sliding
joints tends to have smaller by coordinate in the configuration space. Dark gray regions in
the visualized variation bounds correspond to such configurations. The dark gray region
in Figure 10(b) (double V-slots) has a small rectangle shape. On the other hand, the same
color region in Figure 9(b) (single V-slot) has a vertically longer triangle shape. This shape
difference means that the configuration of the single V-slot joint can have larger position
uncertainties especially in the x axis direction. The double V-slots joint is thus verified
to be suitable than the single V-slot joint for precisely positioning the workpiece.
6 SUMMARY AND DISCUSSION

A configuration space based method for analyzing position uncertainties of two parts
in an assembly is proposed if certain shape variations of individual parts are allowed by
tolerances. The analysis is achieved by calculating a variation bound of the configuration of
the parts if nonnominal parts allowed by the tolerance satisfy the positioning condition in
(a) (b)
Figure 10 Derived bound of the configuration uncertainty of the joint with double V-
slots.
an assembly. An algorithm for computing the variation bound in the octree representation
is developed and applied for visualizing the configuration uncertainty of some sliding
joints. Computational experiments suggest that the visualized image of the variation
bound is helpful for understanding the positioning characteristic of the joints.
The following two questions must be addressed as our future research subjects;
• Required computation time is too much to use the method in the practical application.
The authors are currently developing an improved algorithm which approximates the
variation bound as a union of some convex polyhedrons in the configuration space.
• Derived bound of the configuration variation is still too large and contains some physi-
cally unstable configurations. Additional positioning conditions of assembled parts are
currently under study for bounding the configuration variation more tightly.
7 REFERENCES
Ambler, A.P. and Popplestone, R.J. (1975) Inferring the Positions of Bodies from Specified
Spatial Relationships, Artificial Intelligence, 6, 2, 157-174.
Fleming, A. (1987) Analysis of Uncertainties and Geometric Tolemnces in Assemblies of
Parts, Ph.D. Thesis, University of Edinburgh.
Inui, M., Miura, M., and Kimura, F. (1995a) Relative Positioning of Assembled Parts
with Small Geometric Deviations by Using Hierarchically Approximated Configuration
Space, Proc . of IEEE Conf. Robotics and Automation , 1605-1612.
rnui, M., Miura, M., and Kimura, F. (1995b) Analysis of Position Uncertainties of Parts
in An Assembly Using Configuration Spaces in Octree Representation, Proc. of ACM

Symp. Solid Modeling and Applications, 73-81.
ISO (1983) Technical drawings - Geometrical tolerancing - Tolerances of form, orienta-
tion, location and run-out - Generalities, definitions, symbols, indications on drawings,
ISO 1101-1983.
Latombe, J. -C. (1991) Robot Motion Planning, Kluwer Academic Publishers.
Lee, K. and Andrews, G. (1985) Inference of the Positions of Components in an Assembly:
Part 2, Computer-Aided Design, 19, 1,20-24.
Liu, Y. (1990) Symmetry Groups in Robotics Assembly Planning, Ph.D. Thesis, University
of Massachusetts.
Requicha, A.A.G. (1983) Toward a Theory of Geometric Tolerancing, Int. Journal of
Robotics Research, 2, 4, 45-60.
Sodhi, R. and Turner, J.U. (1994) Relative Positioning of Variational Part Models for
Design Analysis, Computer-Aided Design, 26, 5, 366-378.
Takahashi, K., Suzuki, H., and Kimura, F. (1991) Tolerance Analysis in Machine Assembly
by Classifying Part Contact State, Proc. CIRP Int. Working Seminar on Computer-
Aided Tolerancing, 57-76.
Turner, J.U. (1990) Relative Positioning of Parts in Assemblies Using Mathematical
Programming, Computer-Aided Design, 22, 7, 394-400.
Wilson, R.H. and Latombe, J.-C. (1994) Geometric Reasoning about Mechanical Assem-
bly, Artificial Intelligence, 71, 2, 371-396.
Xiao, J. (1993) Automatic Determination of Topological Contacts in the Presence of
Sensing Uncertainties, Proc. of IEEE Conf. Robotics and Automation, 65-70.
Zhu, D. and Latombe, J.-C. (1990) Constraint Reformulation in a Hierarchical Path
Planner, Proc. of IEEE Conf. Robotics and Automation, 1918-1923.
8 BIOGRAPHY
Masatomo Inui has been an associate professor in the Department of System Engineer-
ing of Ibaraki University, Japan, since 1993. He received a Dr.Eng. degree in precision
machinery engineering from the University of Tokyo in 1991. His interests include solid
modeling, machining and process-planning automation, manufacturing process modeling,
and product modeling. Masahiro Miura is a master course student in the Department of
System Engineering of Ibaraki University.
10
Physically Based Modelling for
Evaluating Shape Variations
H. SUZUKI, K. KASE, K. KATO and F. KIMURA
The University of Tokyo
Department of Precision Machinery Engineering, Hongo 7-3,
Bunkyo-ku, Tokyo 113, JAPAN.
Telephone: +81-3-3812-2111, ext. 6490, Fax: +81-3-3812-8849
email: suzuki@cim.pe.u-tokyo.ac.jp
Abstract
In the designing of mechanical products, it is important to analyze the causal relation-
ship between variations in part's shapes and their effects on the product's functionality.
Computer simulations are promising design tools for evaluating those effects. This paper
describes our attempt to apply physically based modelling methods to implement sim-
ulations, in particular, those for calculating the contact states of an assembly of simple
parts and its motion. These simulations adopt simplified physical models, since their real
physical phenomenon involves numerous complex factors. In this sense, the simulations
do not necessarily reflect real physical behaviors, but the results of the simulations are
considered helpful for designers to qualitatively grasp the relationship between behaviors
of an assembly and variations in part's shapes.
Keywords
Computer Aided Tolerancing, Variational Modelling, Physically Based Modelling, Contact
State, Rigid Body Motion.
1 INTRODUCTION
Demands for computer aided tolerancing (CAT) systems are increasing and a number
of studies have been conducted on this subject. One of the fundamental factors that
has become clear through those studies is causal relationship, from geometrical errors in
product's shape to product's functionality, interchangeability and other product life cycle
properties. This factor is crucial for implementing CAT systems.
Of course, experienced designers usually have knowledge for this factor to some extent.
They know how much tolerance must be given in order for a product to work properly
with longer duration, etc. They also know how much accuracy can be achieved by what
kind of manufacturing methods and their costs. Thus one straightforward approach for
implementing CAT may be a knowledge based approach in which such knowledge of expert
designers is applied. This kind of CAT is very useful, however, it has some drawbacks.
For instance, it does not work well for very novel designs.
Another promising approach which is more generic for implementing CAT is a simu-
lation based approach where the physical behaviors of products with geometrical errors
are evaluated so that designers can estimate effects on behaviors caused by the geomet-
rical errors. For instance, several articles have been devoted to the study for analyz-
ing cumulative geometrical errors of part's tolerances in an assembly based on assembly
simulations[6, 6, 6].
In this paper, we are also interested in such assembly tolerances as well as more kine-
matic behaviors related to interaction( contacts) between parts. For this purpose, tech-
niques of physically based modelling are used, which have been developed in the field of
computer graphics in order to create realistic animations by employing physical laws. In
particular, theories and methods for rigid body motion are introduced.
In order to evaluate effects of geometrical errors, it is necessary to represent these
geometrical errors as computational models. In the following sections, we first briefly
discuss this issue. Then two kinds of simulations will be shown. The first one is to calculate
contact states of two rigid objects. Figure 1 provides an example of this simulation,
showing a kinematic bearing way comprising of the platen and the bed of a lathe. This
simulation calculates contact states of the platen and the bed when their shapes have
geometrical errors. In the second simulation, the motion of the platen in a contact state
is calculated. Free and contact motions of a rigid body are also simulated. Those are a
brief review of our recent research works.
Figure 1 Double Vee Type Kinematic Bearing Way.
2 MODELLING GEOMETRICAL ERRORS

Research on the modelling of tolerance information has been conducted by a considerable
number of researchers [6, 6]. Requicha[6] classified the tolerance models into two classes,
the parametric semantics model and the zone semantics model. The zone semantics model
represents a "tolerance zone" defined in the industrial standards as a volume or a proper
subset of E3. The volume is basically constructed by offsetting a nominal shape of an
object. In this sense, this semantics is mathematically faithful to the original definition of
the standards. However, to my view, it is very hard to implement simulations, specifically
for dealing with characteristics in assemblies such as contact and fit, with this semantics
model, because the boundary of a part's shape is not fixed and also because it is needed
to handle complex systems of inequalities[6].
Evaluating shape variations 149
On the other hand, the parametric semantics model is based on parameterization of

pose and form of toleranced features of an object. The features change their poses and
forms according to the parameters. A set of objects whose features satisfy tolerances is
called a variational class.
If features of an object in a variational class maintain their nominal forms, the tol-
erance model is called "perfect form semantics", otherwise "imperfect form semantics."
The perfect form semantics is unrealistic, because no manufacturing process can produce
perfect shapes. Even with imperfect form semantics, it is impossible to represent all the
possible shape deviation by a finite number of parameters.
However, the parametric semantics model is very useful in terms of tolerance simula-
tions, because each member of the variational class provides a shape representation. It
is also easy to assign probability functions to each of the parameters and to do a statis-
tical tolerance analysis by repeating simulations with those different parameter values.
To produce a shape representation for each member of the variational class, the so called
variational geometry or parametric modelling technologies can be used.
Tolerance specification defines a variational class for a nominal shape. This variational
class consists of shapes similar to the nominal shape and thus their differences from the
nominal shape are considered small. However, it is interesting that the mathematical
framework of the parametric semantics model does not require small differences. The
variational class can therefore be applied to other sets of shape variations than those de-
fined by the tolerance. For instance, the variational class can represent a set of alternative
designs that have different values for major dimensions.
Consequently the parametric semantic model with perfect forms is used for our research.
We assume that a nominal shape of a part is represented by a solid (r-set) S C E3. Shape
errors are usually local to portions of part's surfaces, which are generally referred to as a
feature. The boundary of S, or as is divided into a collection of nominal features Fi cas.
Fi is a closed subset of a primitive surface that is a connected, irreducible and non-trivial
algebraic surface. The actual shape of the part is denoted by P and its actual surface
features by G i . As P exhibits very complicated shape, G i is replaced with a surrogate
feature Ei that is also a closed and connected subset of a primitive surface. This surrogate
feature is used to represent shape errors.
There are three types of shape errors; form error, orientation error and position error. In
order to discriminate these three kinds of errors, the surrogate feature should be defined
by the parameters of its position, orientation and sizes. A canonical form of a surface
equation is adopted, in which surfaces are represented by its reference frame and size
parameters. For instance a cylindrical surface is represented by its radius value and a
reference frame of which z-axis defines the cylinder's axis. By changing these parameters
of the canonical form, the surrogate feature is transformed to obtain geometrical errors
from its nominal shape. By combining surrogate features, we can also generate geometrical
errors for complex features.
3 CONTACT STATE SIMULATION

In this section, a method is described to calculate contact states of parts whose shapes
have geometrical errors. In Figure 3, A and B represent an object or a part's shape.
Without loss of generality, we assume A is fixed and B is movable. We represent the
Figure 2 Geometric Models for Shape Errors
position and orientation of B with parameter q, which is called configuration parameter.

We denote B at the configuration of q by B(q). When B contacts with A, its configuration
is called contact configuration. The problem is to find this contact configuration. Similar
problems have been solved in the field of assembly modeling and tolerance analysis[6, 6].
However, those attempts are unfortunately limited for our application.
Figure 3 Objects and Configurations Figure 4 Potential Field
The basic idea to calculate the contact configuration is a potential field in which the
potential energy of B decreases, as B reaches to A and finally becomes zero when they
contact[6]. To define such a potential field, the work space and objects are discretized to
a regular array of small square cells (v). Then a distance function D t is defined between
A and a cell to generate the potential field. Figure 4 shows equipotential energy contours
of th~ potential field. Consequently, the potential energy of B(q) is given by: U(q)
L Dt(V)2.
VeB(q)
Another remaining issue is to find configurations q at which U (q) is minimum. The do-
main of configuration of B is also discretized to a set of a finite number of configurations,
from which we can find minimum energy configurations. This problem is a combinato-
rial optimization problem. We employ simulated annealing method[6] which is a kind of
stochastic relaxation method to solve the combinatorial optimization problem. The algo-
rithm of the simulated annealing is a large number of iterations of a simple procedure to
stochastically move in the configuration space and to possibly reach a globally minimum
energy state. Figure 5 shows the process of simulated annealing and the contact states of
the objects with some geometrical errors.
Figure 5 Contact States
4 MOTION SIMULATION
By finding contact states, contact forces and motion constraints can be calculated. We
propose a method that computes possible paths of the movable part without interference[6,
6). The simulation algorithm is based on the laws of Newtonian dynamics for the systems
of rigid objects[6, 6, 6, 6). Some basic conditions taken in this system are as follows;
1. Parts are rigid polyhedra.

2. The initial contact state is given.
3. The gravitation acts at the center of gravity of a movable part. External forces except
contact forces, are restricted to those which act at fixed positions, and their magnitudes
and directions are unchangeable.
4. The distributed forces (normal and tangent) acting at a contact region are modeled as
concentrated forces acting at the vertices of that region.
5. Impulse forces due to collision are not considered.
6. Dynamic friction follows Coulomb's law. Static friction is not considered.
Under these assumptions, the motion of a part is regarded as a series of contact states
(quasi-static kinematic motion). Each contact state makes a smooth transition to the next
contact state. Suppose the movable part moves from one contact state to another under
constant forces, some contact points will maintain their contacts and others will lose.
There are no good methods for determining which contact points are maintained or lost.
This lack of information makes the system of motion equations indeterminate, that is, the
number of unknown parameters is more than the number of equations. But by guessing
for the contact points with a certain procedure, Newtonian equations of motion can be
set up and a motion path can be generated by numerically integrating the equations.
Figure 6 shows an example of the kinematic bearing way in Figure 1. In this case,
positions of the male vee way are horizontally inclined as shown by the arrows in the
figure. Figure 6 shows two major contact states. This "double vee" typed mechanism is
known to be sensitive to the geometrical errors, and as shown in Figure 7, the motion of
the platen is distorted by these geometrical errors. The arrow Fp shows the direction of
the applied driving force.
This simulation method is extended to deal with impact forces due to collision and
static friction forces. Figure 8 shows a pipe through which a part is brought by air flow.
The design problem is to determine a shape of the pipe junction. For smooth part's
transfer, the inner shape of the pipe junction should be slightly tapered, otherwise the
part becomes stacked at the junction. The left figure shows an example of where the part
becomes stacked, while the right shows a smooth example.
L-x
Figure 6 Dominant Contact States for Horizontal Error Model
5 CONCLUSIONS
In the designing mechanical products, it is important to analyze the causal relationship
between variations of product's shape and their effects on the product's functionality. We
propose physically based modelling methods for implementing simulations to calculate
the contact state of an assembly of simple parts and its motion.
Since real contact phenomena involve many complex physical factors such as friction,
elastic deformation and surface properties, an accurate physical model for contact is in-
' " II
Figure 7 Motion Path for Horizontal Error Model
feasible. Thus we adopt simplified physical models. In this sense, the simulations do not
necessarily reflect real physical behaviors. We also make assumptions, approximations,
simplifications and exaggeration for making the simulation feasible. Strictly speaking, it
should be called pseudo simulation.
The results of the pseudo simulations are however considered helpful for designers to
qualitatively grasp relationship between behavior of an as~embly and shape variations.
REFERENCES
D. Baraff (1991) Coping with Friction for Non-penetrating Rigid Body Simulation, Com-
puter & Graphics Vol 25, no.4. July.
D. Baraff (1993) Issues in computing contact forces for non-penetrating rigid bodies,
Algorithmica.
D. Baraff (1994) Fast Contact Force Computation for Non-penetrating Rigid Bodies, ACM
Computer Graphics Proc. of Annual Conference.
J. Barraquand et al. (1989) Numerical Potential Field Techniques for Robot Path Plan-
Figure 8 Air-flow Parts Transferring Mechanism.
ning, Stanford Univ. Report No. STAN-CS-89-1285, Oct ..

B. R.Donald and D. K.Pai (1993) The Motion of Planar, Compliantly Connected Rigid
Bodies in Contact, With Applications to Automatic Fastening, Journal of Robotics
Research Vo1.12. noA. August.
A. D. Fleming (1989) Geometric Relationships between Toleranced Features, Geometric
Reasoning, MIT Press, ppA03-412.
S. Hirai (1991) Analysis and Planning of Manipulation Using the Theory of Polyhedral
Convex Cones, Ph.D Thesis, Kyoto Univ. Japan.
R. Jayaraman and V. Srinivasan (1989) Geometric Tolerancing: I. Virtual Boundary Re-
quirements, IBM. J. RES. DEVELOP. Vo1.33, No.2.
F. Kimura et al. (1992) Product design evaluation based on effect of shape errors for part
assembly, Annals of the CIRP, Vol. 41, No.1.
S. Kirkpatrick et al. (1983) Optimization by Simulated Annealing, Science, 220, 4598.
A. A. G. Requicha (1983) Toward a Theory of Geometrical Tolerancing,Int. J. of Robotics
Research, Vo1.2, NoA.
A. A. G. Requicha (1993) Mathematical Meaning and Computational Representations of
Tolerance Specifications, ASME International Forum on Dimensional Tolerancing and
Metrology, CRTD-Vol.27.
D. N. Rochaleau and K. Lee (1987) System for Interactive Assembly Modeling, Computer-
Aided Design, 19, 2.
H. Suzuki et al. (1992) An Approach for Estimating Accuracy Measures based on Object
Modeling, Proc. of IFIP PROLAMAT 92.
K. Takahashi et al. (1993) Motion Analysis of Parts with Geometrical Errors Based on
Dynamic Simulation, CIRP Int. Working Seminar on Computer-Aided Tolerancing,

1993, Paris.
J. U. Turner (1990) Relative Positioning of Parts in Assemblies using Mathematical Pro-
gramming, Computer-Aided Design, 22, 7.
BIOGRAPHY
Hiromasa SUZUKI He is an associate professor in the Department of Precision
Machinery Engineering at The University of Tokyo. His research interests include geo-
metric modeling and reasoning, and their applications to mechanical CAD /CAM sys-
tems. Suzuki received his doctor degree in precision machinery engineering from the
University of Tokyo in 1986. Suzuki is a member of JSPE(Japan Society for Precision
Engineers), JSME(Japan Society for Mechanical Engineers), IEEE and ACM.
Kiwamu KASE He is a researcher at RIKEN, Inst. of Physical & Chemical Research,
Japan. Kase received his doctor degree in precision machinery engineering from the
University of Tokyo in 1994. His research interests include mathematical modelling of
products and its application to rapid prototyping.
Kazunori KATO He is an engineer at Chubu Electric Power Co. Japan. He received
MSc in precision machinery engineering from The University of Tokyo in 1995. His
research interest includes physical based modelling and its application to CAE.
Fumihiko KIMURA He is a professor of Department of Precision Machinery Engineer-
ing at The University of Tokyo. Ando received BSc from The University of Tokyo in
1995. His research interest is in geometric reasoning for layout constraint solving. He
is a member of IFIP WG5.2/5.3, ISO TC184/SC4 and CIRP.
QUESTIONS AND ANSWERS

Question from Prof. Portman: Along the motion, the object must make contact at
multiple points. I do not see, for instance, five point contact in your example. Don't you
have this? How do you deal with this problem?
Speaker: Actually, the multiple point contact happens during the motion. Our method
deals with this situation by solving a simulateneous equation to estimate the contact
forces. It has been observed that such multiple contact situation is transient and usually
does not last so long.
PART FOUR
Statistical Tolerancing
11
Towards an ISO Standard for
Statistical Tolerancing
Vijay Srinivasan
IBM T.J. Watson Research Center, Yorktown Heights, NY 10598, USA.
Telephone: 914-945-3411. Fax: 914-945-3242.
email: vasan@watson.ibm.com
and
Columbia University, New York, NY 10027, USA.
Michael A. O'Connor
IBM T.J. Watson Research Center, Yorktown Heights, NY 10598, USA.
Telephone: 914-945-3442. Fax: 914-945-3242.
email: oconnor@watson.ibm.com
Abstract
Although statistical tolerancing has been practiced in industry for a long time, standards
have paid only limited attention to it. This is changing now because of economic reasons,
and ISO is investigating how to standardize statistical tolerancing. In this paper we trace
the history and document the current state of the knowledge in standardizing statistical
tolerancing. Our objective is to disseminate this information as widely as possible so that
the international community is aware of, and has an opportunity to influence, the thinking
behind the future ISO standard.
Keywords
ISO, tolerancing, standards, statistics.
1 INTRODUCTION
Both classical and statistical tolerancing are currently practiced in industry. Classical
tolerancing is popularly known as the worst-case tolerancing. Given an actual feature, it
defines the test to decide whether the feature is acceptable. Statistical tolerancing, on the
other hand, deals with a population of features. It defines the test that decides whether
a given population of actual features is acceptable.
All international and most national standards have codified only the classical toler-
ancing. Company specific internal standards have so far sustained the industrial practice
of statistical tolerancing. The main reason for the popularity of statistical tolerancing in
industry is economy. The same economic forces are now pushing for national and interna-
160 Part Four Statistical tolerancing
tional standards for statistical tolerancing. The latest ASME Y14.5M-1994 standard on
dimensioning and tolerancing (ASME, 1994a) has yielded to this pressure by providing
a symbol to indicate in drawings those dimensions that should be controlled statisti-
cally. Realizing that it is not sufficient, researchers in the U.S.A. have started developing
mathematically sound definitions of the syntax and semantics of statistical tolerancing.
Meanwhile, at the Florida meeting in January of 1994, the ISO Joint Harmonization
Group approved the formation of a Task Group on Statistical Tolerancing so that an ISO
standard may emerge. Since there are few national standards on statistical tolerancing,
it was felt that we had an excellent opportunity to develop an international standard
first, which could then be simply adopted by the national standards organizations. Those
familiar with ISO standardization procedures know that it is usually the other way around.
In this paper we summarize our current state of knowledge towards standardizing sta-
tistical tolerancing. Our objective is to disseminate this information as widely as possible
so that the international technical community is aware of, and has an opportunity to
influence, the thinking behind the future ISO standard. Section 2 briefly documents prior
attempts at standardizing statistical tolerancing. In Section 3 we show how to derive the
distribution of one or more random variables from a population of actual parts. A family
of population of parts then lead to a family of distributions. Section 4 explains three ways
of specifying the acceptable subset of distributions from this family. Using a simple pin-
in-hole example, Section 5 illustrates a statistical tolerance analysis using the semantics
of statistical tolerancing developed thus far.
2 PRIOR ATTEMPTS AT STANDARDIZATION

Company standards were developed to promote the use of statistical techniques in tol-
erancing as early as 1960 (IBM, 1960). These standards explained design methodologies
when parts are stacked up to form an assembly. The cumulative effect of the individual
variations was predicted using simple rules of statistics, such as the resulting standard
deviation being the square root of the sum of the squares of individual standard devi-
ations. The economic advantage of this over the worst-case stack-up analysis, in which
the individual part variations were combined linearly, was strongly emphasized by these
documents.
The first serious national standard for statistical tolerancing was issued in 1974 by
West Germany (DIN, 1974) It focused on size tolerancing and illustrated how statistical
principles can be used in tolerance analysis involving linear dimensional chains. Moreover,
it provided an expanded syntax (see Figure 1) to enable designers to indicate in a drawing
what statistical distributions are acceptable for a particular dimension. The caption in
Figure 1 explains the semantics per the same standard. A second part of the standard
expanding the statistical aspects and calculation methods was circulated in a draft form
in 1980 (DIN, 1980). The draft was, however, withdrawn because the German industry
felt that the content was too complicated for designers to follow (Grode, 1994).
A recent member of company standards is the Motorola six-sigma methodology popu-
larized in late 1980s (Harry and Stewart, 1988). It is notable for its adoption of process
capability indices to specify a set of populations of parts that are statistically acceptable.
By adopting popular terms from SPC (statistical process control) the Motorola method-
ology made it easier to communicate design specifications to manufacturing engineers.
ISO standard for statistical tolerancing 161
I' 17.6 ± 0.06 ± 0.03P86% ,I
-[--------f-3
Figure 1 Specification of statistical tolerancing according to DIN 7186. The interpreta-
tion is that at least 86% of the population of this part must have the indicated dimension
fall within 17.6 ± 0.03. In addition, at most 7% of the population may lie within 17.6=g:g~,
and at most 7% of the population may lie within 17.6:tg:g~.
A note such as the following must be placed on the drawing:
FEATURES IDENTIFIED AS STATISTICALLY TOLERANCED ® SHALL BE

PRODUCED WITH STATISTICAL PROCESS CONTROLS.
Figure 2 AS ME Y14.5M-1994 designation of statistical tolerancing.
The U.S. national standard (ASME, 1994a), issued in 1994, defines statistical to1eranc-
ing as " ... the assigning of tolerances to related component items of an assembly on the
basis of sound statistics." Further, on applications, it says " ... the tolerances assigned to
component items of an assembly Me determined by arithmetically dividing the assembly
tolerances among the individual components of the assembly. When tolerances assigned
by arithmetic stacking are restrictive, statistical tolerances may be used for increased in-
dividual feature tolerances." It, however, cautions " .. the increased tolerance may reduce
manufacturing cost, but shall only be used when the appropriate statistical process con-
trol will be used." Then, it proceeds to define a symbol as shown in Figure 2 to designate
statistical tolerancing in a drawing. It provides no further semantics.
This brings us to the current ISO interest in standardizing statistical tolerancing. There
is a general consensus that this area is important, and the time is ripe for internationally
codifying what is considered the best practice in industry. This involves collecting available
information, processing this information to extract the essential principles, and proposing
these principles for standardization. Work has begun in all these three areas.
shape/pose
Gaussian Fit
/
parameter( s)
actual variable( s )
feature of interest
\ Chebyshev Fit
actual
value(s)
Figure 3 Extracting parameters or actual values by fitting.
3 DERIVING A DISTRIBUTION
An actual part is a (model of a) manufactured part. Its boundary can be partitioned
into many actual surface features. These actual surface features do not have perfect-form.
However, we can fit a perfect-form feature to an actual feature. The fitted perfect-form
feature is known as the substitute geometry. As shown in Figure 3, two types of fits are
popular.
A Gaussian fit produces the perfect-form feature from a collection of similar features
by minimizing the sum of the squared distances to a set of points on the actual feature.
The shape and pose parameters of the fitted surface then become the values of interest. In
classical parametric tolerancing each of these parameters is permitted to lie only within an
interval specified for it. Vectorial tolerancing is a special case of parametric tolerancing.
A Chebyshev fit produces a perfect-form feature or a tolerance zone for the actual feature
using min-max criterion. This results in one or more actual values as defined in the AS ME
Y14.S.1M-1994 standard (ASME, 1994b). Actual values are crucial for providing statistical
semantics for geometric tolerancing.
As actual values appeared in standards only recently, they deserve some explanation.
For an actual cylindrical· pin, the actual mating size is the diameter of the minimum
circumscribing cylinder, that is, it is the minimum value of the diameter of a right circular
cylinder that contains the actual pin. For an actual cylindrical hole, the actual mating
size is the diameter of the maximum inscribing cylinder. The notion of actual mating
size can be extended to cases where the mating features are constrained by datums. In
Figure 4 we show the actual mating size of a pin at basic orientation. It is the diameter
of the minimum circumscribing cylinder oriented perpendicularly to the datum plane. If
the position, not just orientation, of the pin were constrained then an actual mating size
actual mating envelope

at basic orientation
'----t---'
~
AS SPECIFIED AS PRODUCED
Figure 4 Actual mating size of a pin at basic orientation is the diameter of the actual
mating envelope at basic orientation.
of the pin at true position can be similarly defined. To complete the description, and to
help us later in statistical tolerance analysis, the actual mating size of a hole at basic
orientation is illustrated in Figure 5.
Now, consider an actual surface feature of interest in a manufactured part. If we hy-
pothesize infinitely many instances of the manufactured part, we have a population of the
actual surface features which, through the process of fitting, result in the distribution of
one or more random variables. The function of the part dictates whether the Gaussian or
the Chebyshev fitting should be used. If assembleability is the major functional consid-
eration, then Chebyshev fit is preferred. This process of mapping a population of actual
features to the distribution of one or more random variables is illustrated in Figure 6.
Thus far we have used the fitting process to derive the distribution of one or more
random variables from a population of actual features. In some cases, it is possible to
construct the same distribution by a set containment argument without invoking the
Chebyshev fitting process directly. For this, consider an actual feature G and a tolerance
zone Zt that is parameterized by t. Now consider the outcome of the test "Is G c Zt?"
as a binary event. Then the probability P( G c Zt) is the fraction of the population of
G that passes this test. This probability is directly related to the cumulative distribution
function of actual values derived from Chebyshev fitting. This alternate method of deriving
a distribution from a population of features seems to be conceptually more appealing.
actual mating envelope

at basic orientation
AS SPECIFIED AS PRODUCED
Figure 5 Actual mating size of a hole at basic orientation is the diameter of the actual
mating envelope at basic orientation.
a population of distribution of
Fitting
actual features random variables( s)
Figure 6 Mapping a population of actual features to distribution of random variable(s).
4 ACCEPTABLE DISTRIBUTIONS
Having mapped a population of actual features to a distribution, we can define an ac-
ceptable class of populations of actual features (or, equivalently, actual parts) by defining
which distributions are acceptable. Currently, three ways of specifying acceptable class of
distributions have been proposed. They have been described in detail elsewhere (Srini-
vasan and O'Connor, 1994) and we summarize them here.
The first uses process capability indices. Let x be a random variable, and LSL and
USL be the lower and upper specification limits. If 11- and (1 are the mean and standard
deviation of x, then
C _ USL- LSL C . {11- - LSL USL - 11-}

p - 6(1 and pk = mm 3(1 , 3(1 (1)
are known as the process capability indices. In addition, it is useful to define
111- - LSLtuSL 1
Cc = USL L~L (2)
to quantify the mean shift from the target value of (LSL+ USL )/2. We then specify three
numbers P, K, and F, so that Cp :2 P, Cpk :2 K and Cc :::; F. It can be shown that this
specifies a statistical tolerance zone in the Cpk-Cp plane as illustrated in Figure 7(a). The
Pr(x :::; X)
(1
1 2 X
(a) (b) (c)
Figure 7 Three ways of defining statistical tolerance zones. (a) A tolerance zone in the
Cpk-Cp plane. (b) A tolerance zone in the 11--(1 plane. (c) A CDF tolerance zone plotted in
a normal probability paper.
triplet (P, K, F) also defines a tolerance zone in the 11--(1 plane as shown in Figure 7(b).
This is the second way of specifying acceptable distributions. It is often more useful for
design purposes to refer to the tolerance zone in the 11--(1 plane. These two specifications
are completely equivalent. The third, and last, way of specifying acceptable distributions
deals with the cumulative distribution function (CDF) of x. A tolerance zone is defined
by two CDFs, G and H, by accepting all random variables whose CDF F(X) satisfies
G(X) :::; F(X) :::; H(X) for all X. The zone can be visualized in the graph of X vs
Pr(x :::; X) as shown in Figure 7(c), either in the usual linear scale or in a special
probability paper popular among statisticians. The statistical tolerance zones of Figure 7
are shaded to illustrate that they are indeed zones. From now on we will drop the shading,
and show the zones by drawing only their boundaries.
Pr(x ~ X)
0.93
0.50
0.07
17.54 17.57 17.63 17.66 X
Figure 8 A CDF tolerance zone for the DIN 7186 example, plotted in a normal proba-
bility paper.
The CDF tolerance zone for the DIN 7186 example of Figure 1 is shown in Figure 8. It is
instructive to look at the same DIN specification as a statistical tolerance zone in the I-HT
plane. For this, it is first assumed that the actual value associated with the indicated size
dimension varies according to the normal distribution. The specification is then relaxed
to mean that at most 7% of the population can be below 17.57, and similarly at most 7%
of the population can be above 17.63, to account for infinite tails of a normal distribution.
At the same time, we observe that the DIN specification as interpreted as a CDF tolerance
u
0.0203
17.54 17.57 17.63 17.66
Figure 9 A tolerance zone for the DIN 7186 example in the p,-u plane under the as-
sumption of normality. This zone can also be obtained by assuming LSL = 17.57, USL =
17.63,P = K = 0.4917, and F = 1.
zone in Figure 8 has portions that cannot be reached by a normal distribution. Taking
these two observations into account, and still under the assumption of normality, the
statistical tolerance zone for the DIN example in the /.HJ' plane is shown in Figure 9. It is
an easy exercise to plot the corresponding statistical tolerance zone in the Cpk-Cp plane.
There are several geometric tolerances for which a normal distribution for the actual
value is never realized. All form tolerances, for example, come under this category. It
has been proposed (Henzold, 1994) to handle them using an extension of the DIN 7186
syntax, and a CDF zone based semantics, as shown in Figure 10. This example illustrates
I/ I 0.025:!:g:g~~ ± 0.025P75% A
Pr(x ::; X)
1.00 1---------.
0.75
0.50
0.25
,
0.08.000 0.025 0.~50 0.075 0.100 X
Figure 10 Syntax and semantics of runout tolerance according to an extension of the

DIN 7186 standard. The random variable x is the actual value for the runout tolerance.
Any acceptable CDF of x must lie within the CDF tolerance zone, plotted in a usual
linear scale.
the power and flexibility of specifying acceptable distributions using a CDF tolerance
zone.
We are now ready to define the semantics of statistical tolerancing as a theoretical
inspection procedure first defined by Requicha (1983) for classical geometric tolerancing.
Figure 11 illustrates the semantics. Given a population of parts, we first derive the distrib-
ution of one or more random variables using the fitting process. Then, depending on which
specification of acceptable distributions is chosen we decide whether the derived distrib-
ution (and hence the population of parts) is acceptable. It should be emphasized that in
tolerancing semantics we are concerned only with the theoretical inspection procedure.
In an actual inspection procedure, one has to consider sampling errors and measurement
uncertainties. These issues do not affect our definition of the semantics of statistical tol-
erancing.
Is (Cpk,Cp) in the
yes
tolerance zone? accept
the population
distribution of a Is (/L, IT) in the

random variable tolerance zone?
reject
Is CDF in the no the population
tolerance zone?
Figure 11 Semantics of statistical tolerancing as a theoretical inspection procedure.
5 STATISTICAL TOLERANCE ANALYSIS

How can a designer use the statistical tolerancing semantics developed so far in a tol-
erance analysis? We will answer this question using the pin-in-hole example previously
encountered in Figures 4 and 5. These two figures illustrate the classical tolerancing under
which assembleability is ensured even under the worst ca.se. To see this, let us define
(3)
where d ho1e and drnn are, respectively, the actual mating sizes of the hole and the pin at the
basic orientation. For parts in-tolerance, dpin varies between 11.7 and 12.0, which represent
the resultant and virtual conditions, respectively. Similarly, for parts in-tolerance, dhole
varies between its virtual and resultant conditions, 12.0 and 12.3, respectively. Therefore,
in classical tolerance analysis, gap is found to vary between 0 and 0.6 and assembleability
is ensured for all in-tolerance parts.
In statistical tolerancing we will relax allowable variations in the pin and the hole by
choosing, say, LSL = 11.55 and USL = 12.15 for drnn, and LSL = 11.85 and USL = 12.45
for d ho1e . In addition, let us choose P = 2, K = 1.5, and F = 0.25 for both drnn and d ho1e •
These choices for P, K, and F lead to simple statistical tolerance zones in the /L-lT plane,
as we will see soon. Assuming that dpin and d ho1e are distributed independently, we can
derive the mean and variance of the gap in terms of the means and variances of dpin and
dho1e as
/Lgap = /Lhole - /Lpin, (4)

and
(5)
Using these two relationships, we can compose the statistical tolerance zones for drnn and
in the j.I.-17 plane to obtain the tolerance zone for gap as illustrated in Figure 12.
dhol e
17 17
0.0707
0.05 0.05
12.07512.225 11.775 11.925 0.15 0.45
gap
Figure 12 Composing statistical tolerance zones for the pin and the hole to obtain the
tolerance zone for the gap.
The next task is to perform a risk analysis using the statistical tolerance zone for gap
obtained earlier. For this we need an additional assumption about the distribution of
gap. If we assume that gap is normally distributed, and this is a strong assumption, we
can perform the risk analysis as follows. The risk in this case is a failure to assemble,
and can be quantified as the probability Pr(gap < 0). Then, every pair (j.I.,l7) can be
easily mapped to the risk Pr(gap < 0), and level curves of a graph of this mapping are
straight line segments in the J.I.-17 plane as shown in Figure 13 as iso-risk contours. As
the (J.I., (7) point moves away from the "worst" corner, the risk of non-assembly decreases
dramatically.
This simple example illustrates a few common issues to be addressed in any statistical
tolerance analysis. We first need mathematical rules for combining part level variations
to produce the assembly level variation. In the example considered, the assumption of
statistical independence simplified this mathematical rule. In addition, specifying indi-
vidual part tolerances statistically as axis-parallel rectangular zones in the J.I.-17 plane led
to a simple evaluation of this mathematical rule. In general cases, such as the triangu-
lar statistical tolerance zone shown in Figure 9, the rule of composition even under the
assumption of independence involves computing the Minkowski sum of objects bounded
by curve segments. If we choose to work only with CDF tolerance zones then the rules of
composition are different. Such problems will be addressed in future publications.
17 parts per thousand
11 parts per million
0.0707 0.146 parts per billion
~ 0 parts per trillion
0.15 0.45
Figure 13 Iso-risk contours for the pin-in-hole assembly under the assumption of nor-
mality. The risk is marked in the popular "ppm" (parts per million) terminology, and
indicates the relative number of instances in which the pin and the hole will not assemble
maintaining the datum contact.
Next, we need tools for performing the risk analysis. The assumption of normality in the
evaluated function simplifies this analysis. If the "loop closure" function, a generalization
of the gap function encountered in the example above, is a linear function of many random
variables then the central limit tendency often justifies assumption of normality in the
loop closure function. That is, even if the individual random variables are not normal,
their sum tends to normal quickly as the number of variables increases. Currently, the
risk analysis merely presents the projected risk for all the acceptable distributions. It does
not predict which of these risks is more likely to occur.
6 SUMMARY
We have presented the current state of the knowledge in statistical tolerancing syntax,
semantics, and their potential use in tolerance analysis. Company standards have been
promoting the use of statistics in design and manufacturing internally within their or-
ganizations and by their vendors. Recent efforts in this area. make clever use of process
capability indices borrowed from statistical process control, thereby increasing their ac-
ceptance by manufacturing engineers. The DIN 7186 standard issued in 1'974 has been
the first and the best of the national standards in statistical tolerancing, though its follow
up received a disappointing reception. The latest ASME Y14.5M-1994 standard provides
partial syntax and no semantics.
A population of features (equivalently, parts) was given a statistical distribution through
the process of fitting. Using Gaussian and Chebyshev fits, a unified semantics for paramet-
ric and geometric tolerancing was developed. A key concept that enables this unification
is the actual value of geometric tolerancing as defined and promoted by the new ASME
ISO standardfor statistical tolerancing 171
Y14.5.1M-1994 standard on the mathematical definition of dimensioning and tolerancing

principles.
The task of statistical tolerancing is to specify a theoretical inspection procedure to
decide which of the multitude of statistical distributions (equivalently, populations of
parts) are acceptable. Here we summarized three ways of specifying statistical tolerance
zones to test whether a distribution is acceptable. The tolerance zone in the Cpk-Cp plane
might be more appealing to process and quality control engineers. The tolerance zone in
the p.-a plane, which is completely equivalent to the Cpk-Cp specification, makes it easier
to perform statistical tolerance analysis. Finally, the CDF tolerance zone has greater
flexibility and generality; it is closer to the DIN 7186 semantics and can accommodate
non-symmetric distributions easily. The standards community will have to make the choice
among the three, or come up with easy means of coping with multiple choices.
Proper methodologies and tools for statistical tolerance analysis based on the proposed
semantics need to be developed Iurther. We outlined one approach that showed how this
could be done. In addition, actual inspection procedures that account for sampling errors
and measurement uncertainties must be developed to ensure proper conformance tests.
These are subjects of active research and will be addressed in future publications. Clearly,
more work needs to be done and we believe that we are headed in the right direction.
7 ACKNOWLEDGMENT AND A DISCLAIMER

We gratefully acknowledge our colleagues in the AS ME and ISO standards committees
for many helpful suggestions. However, the opinions expressed in this paper are our own
and they do not represent the official position of the ASME, ISO, or any of its member
bodies.
REFERENCES
AS ME (1994a) ASME Y14.5M-1994, Dimensioning and Tolerancing. The American So-
ciety of Mechanical Engineers, New York.
ASME (1994b) ASME YI4.5.1M-1994, Mathematical Definition of Dimensioning and
Tolerancing Principles. The American Society of Mechanical Engineers, New York.
DIN (1974) DIN 7186 Baltt I, Statistische Tolerierung: Begriffe, Anwendungsrichtlinien
und Zeichnungsangaben, Deutsche Normen, Berlin 30, BRD.
DIN (1980) Entwurf DIN 7186 Teil2, Statistische Tolerierung: Grundlagen fur Rechen-
verfahren, Deutsche Normen, Berlin, BRD.
Grode, H.-P. (1994) Personal Communication.
Harry, M.J. and Stewart, R. (1988) Six Sigma Mechanical Design Tolerancing. Motorola
University Press, Schaumburg, IL.
Henzold,G. (1994) Handbook of Geometric Tolerancing. John Wiley, New York.
IBM (1960) Statistical Dimensioning - IBM Corporate Standard 0-0-1017-0, International
Business Machines Inc.
Requicha, A.A.G. (1983) Toward a theory of geometric tolerancing. International Journal
of Robotics Research, 2, 45-60.
Srinivasan, V. and O'Connor, M.A. (1994) On interpreting statistical tolerancing. Manu-

facturing Review, 7, 304-311.
BIOGRAPHY
Vijay Srinivasan is a research staff member at the IBM Thomas J.Watson Research
Center in the Manufacturing Research Department. He is also an adjunct professor in
the Department of Mechanical Engineering at Columbia University, New York, where he
teaches courses in engineering design and manufacturing. He is a member of the Ameri-
can Society of Mechanical Engineers, the AS ME YI4.5.1 subcommittee on Mathematical
Definition of the Y14.5 standard, and the U.S. Technical Advisory Group of the ISO
TC 10/SC 5. He is also the Chairman of the AS ME Research Committee on Dimensional
Tolerancing and Metrology.
Michael A. O'Connor is a research staff member at the IBM Thomas J. Watson Re-
search Center in the Manufacturing Research Department. He is a mathematician with
research interests in geometrical representations and algorithms and their application in
geometrical modeling and tolerancing. He is a member of the American Mathematical
Society and the Mathematical Association of America.
QUESTIONS AND ANSWERS AFTER THE PRESENTATION

Prof. van Routen: Of the three statistical tolerance zones you presented, which do you
prefer?
Dr. Srinivasan: I personally like the CDF tolerance zone, because of its generality. The
DIN standard has also chosen that interpretation.
Prof. Fortin: Wasn't it Prof. Deming who introduced the Cp and Cpk process capa-
bility indices?
Dr. Srinivasan: I don't know. I have heard that these indices originated here in Japan.
12
Statistical Process Control using Vectorial
Tolerancing
Kristian Martinsen
University of Trondheim, Department of Production and Quality
Engineering
Richard Birkelandsvei 2b, N-7034 TRONDHEIM, NORWAY
Phone: +4773593800 Fax: +4773597117,
E-mail: kristian@protek.unit.no
Abstract
There will always be a certain amount of systematic and random errors in a manufacturing
process. Quality loss is kept at a minimum when the systematic error is minimal and the
process is centred at nominal value. To achieve this, controlling the manufacturing process is
vital. In Vectorial Tolerancing surface location and orientation is described with vectors in a
Workpiece Coordinate System. Vectorial Tolerancing provides, as opposed to conventional
tolerances, a clear distinction between the size, form, location, and orientation deviations. The
magnitude and direction of each feature is therefore known. Deviations on these different
features will have different causes and must be controlled separately. This paper shows how
orientation and location of a surface can be controlled. Correction to the systematic process
errors was done by mirroring the calculated mean values onto the nominal plane, and making a
new NC-code for the milling machine with the mirrored planes as new nominal. Multivariate
Statistical Process Control charts were used to keep the process in control. The results showed
that Statistical Process Control of surface location and orientation using Vectorial Tolerancing
is possible.
KeyWords
Vectorial Tolerancing, Milling, Multivariate Statistical Process Control
1 INTRODUCTION
A manufacturing process will always contain a certain amount of systematic and random
errors. According to the Taguchi loss function (Aune, 1994). The quality loss is kept at a
minimum when a process with a given dispersion is centred at nominal value. To achieve this,
controlling the manufacturing process is vital. One way to do this is by Statistical Process
control. The conventional ISO standards for tolerancing and dimensioning are, however, in
some cases not suited for process control. This is because ISO-tolerances is sums of the
features location, orientation, size and form, where the magnitude and direction of each feature
are unknown. These different features will have different causes, and in many cases it is
impossible to control the manufacturing process without dealing with each feature separately
(Wirtz, 88). SPC based on the ISO-definition of tolerances can only make sense when the form
deviations are negligible, and the measuring task is truly one dimensional. Process control on a
distance tolerance can only work when form deviations and orientation deviations are
negligible small. When the orientation deviations, such as parallelism, is defined according to
the ISO standard, process control is impossible because the direction of the deviation is
unknown.
Best fit
Fonn
substitution
deviation Parale II ism plane Orientation
Figure 1. Orientation deviation. Comparison of ISO and Vectorial Tolerancing. In the ISO
definition, form deviation is enclosed and direction of the orientation deviation is unknown.
Farmer and Gladman (1986) and Wei! (1988) writes that tolerance definitions based on
tolerance zones that covers all types features such as the ISO tolerances, takes no account of
functional, manufacturing, assembly and measuring requirements. Wei! (1991) points out that
in the case of robot links, the largest errors are caused by orientation deviations. The direction
of the orientation deviation, however, is unknown in the ISO orientation tolerances.
Vectorial Tolerancing is a method that provides a clear distinction between size, form, loca-
tion, and orientation (Wirtz 88) . The aim of the experiments described in this paper was to
prove how Vectorial Tolerancing is suited for Statistical Process Control of both location and
orientation.
1.1 Vectorial Tolerancing

A workpiece is material limited by surfaces. The surfaces are described by the geometrical
features form, size, location, and orientation. Vectorial Tolerancing deals with these different
features separately, and describes the position of a surface with a location vector and an
orientation unit vector in a Workpiece Coordinate System, WCS. (Wirtz, 88) This is based on
the six degrees of freedom. Generally, the location vector fixes the three translation degrees of
freedom and orientation vector fixes two rotational degrees of freedom of a surface
(Martinsen, 93). However, a given surface can have a number of open degrees of freedom. A
Statistical process control 175
plane, for example, has three open degrees of freedom; two rotations and one translation.
Hence, there are three quality characteristics describing the position of a plane; one for location
and two for orientation. Similarly, a cylinder has two open degrees of freedom, and therefore
four position quality characteristics. In addition to orientation and location, there are quality
characteristics on size and form. (Controlling size and form is not covered in this paper).
Sub coordinate systems

In some cases, for example when a surface is not aligned with the WCS, a sub coordinate
system has to be defined. This coordinate system can be defmed by the location point of the
nominal surface as origin, and nominal surface orientation vector defines the principal axis. The
deviations of this surface are described in this sub coordinate system.
Figure 2 Vectorial deviations of a plane.
2 DESCRIPTION OF EXPERIMENT METHOD
48 24 24
Figure 4 Milling of plane

Figure 3 Test workpiece (workpiece 1). E4 and planes E2 and E3.
The equipment available was a Willemin W400 5-axis milling machine and a Wentzel
Coordinate measuring machine with a Renishaw PH9 Measuring probe. To inspect the
temperature during the milling sequences, a temperature sensor was fixed on the main spindle
bearing. The workpiece material was aluminium. There was a constant flow of coolant during
cutting.
The workpieces where mounted on palettes with 4 screws at surface E8. E8 is therefore not
machined. The Workpiece coordinate system is defined this way: Plane E2 defines by its
normal the principal axis. The intersection line between El and E2 defines the secondary axis,
and the intersection point between El, E2 and E6 is the origin. The deviations on plane E3 and
E4 was studied in the experiments described in this paper. Both planes are manufactured in one
cut (Se Figure 4). E4 has three quality characteristics; location 20, orientation Ex and
orientation E y. E3 has two quality characteristics, location Yo and orientation Ex. The reason
E3 has only one orientation is that E3 is a side plane to plane E2. This means that E2 and E3 is
manufactured in one cut, E2 is face milled and E3 hub milled, and the angle between them is
decided by the milling tool. E2 defines the principal direction in the coordinate system.
Deviations to orientations on E2 are therefore by definition zero.
3 DECREASE SYSTEMATIC ERROR

The task is to decrease the systematic error after a small sample of manufactured workpieces.
A number of samples of the same workpiece were manufactured, and measured. Correction to
the process was done by mirroring the calculated mean values onto the nominal plane, and
making a new NC-code for the milling machine with the mirrored planes as new nominal
(Wirtz, et al. 93). Roughly estimates of the process parameters mean and variance is calculated
from the sample. The estimated mean and variance will also be an estimate of the systematic
and random error of the process. If the systematic error is significant larger than the random
error, a sample of 5-10 workpieces can be enough to make the first correction of the process.
This correction is validated by the confidence interval of the estimated mean (Montgomery,
91). If the nominal value is outside the confidence interval, the process is corrected. If the
nominal value is inside the confidence interval, the sample size has to be increased before a
decision can be made. The chosen confidence level (I-a) is dependent on the cost of making a
correction. If the cost of making a correction is high, the confidence level must be high, for
example 95%. This is to prevent unnecessary corrections of the process.
It was chosen to look at each quality characteristic independently although this approach
can be misleading. If this is critical, multivariate confidence regions should be used in addition
to the individual plots (Johnson, 1992).
3.1 Description of experiment and results

In the experiments described here, plane E3 and E4 were corrected. At first a series of 6
workpieces was manufactured (Series 1.1). Table 1 shows the results from series 1.1. U and L
are upper and lower limits in a confidence interval with a = 0.05 (95% confidence level). The
corrections made after the first series is listed in the last row. The results of the corrections are
shown in Figure 5 to Figure 9.
Table 1 Results from first series, work2iece 1

Series 1.1 E3 Yo (mm) E3 Ex E4 Zo (mm) E4Ex E4E~
Nominal 24 0 20 0 0
Mean 24,06498333 0,001192 19,993425 0,0001 0
Mean deviation -0,06498333 -0,00119 0,006575 -0,0001 0
Variance 7,79787E-05 3,64E-08 3,813E-05 4E-09 1,3E-08
U 24,07425042 0,001392 19,999905 0,000166 0,00012
L 24,05571625 0,000991 19,986945 3,36E-05 -0,00012
Correction 0,06498333 0,00119 No corr. 0,0001 No corr.
These figures also shows the corrections made after the third and fifth series. The correction
made after the fifth series was on the Y-component of the orientation. In order to make this
correction, the milling paths on plane E4 had to be turned 90 degrees and with multiple cuts
instead of one cut. This was due to kinematic limitations on the milling machine.
Location YO, Plane E3, Series 1.1 - 1.6, Workpiece 1
24.1
24.08
24.06
24.04
24.02
E
E
24
23.98
23.96
23.94
Series 1 Series 2 Series 3 Series 4 Series 5 SeMes 6
23.92
Workpleces
Figure 5 Plane E3, Location Yo, Series 1.1-1.6, Workpiece 1.
Orientation EX, Plane E3, Series 1.1 - 1.6, Workpiece 1
Correction:
-0.00119
0.001
0.0005
8in(T)
3 5
-0.0005
Series 1 Series 2 Series 3 Series 4 Series 5 Series 6

~.OOI~==~----~~--~~~----~~--------------------~
workpiaces
Figure 6 Plane E3, Orientation Ex. Series 1.1-1.6, Workpiece 1.

Location ZD, Plane E4, Series 1.1 -1.6, Workpiece 1
20.01 r----------------------..,
20
19.99
~ 19.98
19.97
19.96
eries 1 Series 2 Series 3 Series 4 Series 5 Series 6

19.95
Workpieces
Figure 7 Plane E4, Location Zo, Series 1.1-1.6, Workpiece 1.
Orientation Ex, Plane E4, Series 1.1 - 1.6, Workpiece 1
0.001...--------------------,
0.0008 ... Correction
0.0001
0.0006 . . E:orrection .
-0.0001
:~
J:.
0.0004
sin(ll)
0.0002
Series 6
Figure 8 Plane E4, Orientation Ex, Series 1.1-1.6, Workpiece 1.

Orientation Ey, Plane E4, Series 1.1 -1.6, Workpiece 1
0.0008,----------------------.
Correction of Ey
0.0006
\.
.. alieredttie· p·rocess· .
0.0004 ..... .
0.0002
sin(a)
4 2 4 6
-0.0004 Series 1 Series 2 Series 3 Series 4 Series 6

Workpieces
Figure 9 Plane E4, Orientation Ey, Series 1.1-1.6, Workpiece 1.
3.2 Discussion of results

Plane E3
Table 1 shows both location and orientation of E3 the nominal value outside the 95%
confidence interval. Both quality characteristics are therefore corrected. Figure 5 and Figure 6
shows successfully correction of both location and orientation of plane E3. After series 2 and 3
when more workpieces where present, the orientation was found to be over-corrected. (Figure
6). The orientation is therefore given a new correction.
Plane E4
On plane E4, The orientation component Ey was corrected after series 5. In order to make this
correction, the process had to be altered, and the effect of this alteration is totally new
deviations. The effect can be seen on all quality characteristics of plane E4. Although the
nominal value is just outside the confidence interval after the first series, the location was
corrected after series 3. The correction was successfully (when we leave out series 6). The
orientation component Ex was corrected even though the nominal value was inside the
confidence interval. The deviation was too small compared with process dispersion to justify a
correction after the first series. Ex is therefore corrected back to original values after series 3.
4 MULTIV ARIATE PROCESS CONTROL

Now the task is to keep the process in control, and SPC charts are found to be useful. To
establish the control limits of the control chart, there must be enough samples to get a good
estimation of the process parameters. Since there are more than one quality characteristic for
each surface in Vectorial Tolerancing, multivariate SPC has to be used. Rotelling (1947)
introduced the TZ-control charts as a technique for testing bomb sights. Alt (1985) shows how
the F-distribution is used to find control limits of the control charts when a small number of
preliminary samples are present. In the experiments in this paper, it was chosen to use the
method described in Alt (1985).
One important assumption to the control charts is that the quality characteristics are jointly
distributed according to the p-variate normal distribution. The results from the experiments
validated this assumption. However, Bjl'lrke (1991) shows that when a converges to 0, this
assumption is not longer valid. (a=P{type I error}=P{reject HoIHo is true} that is: concluding
the process is out of control when it is really in control) This is also the case when the process
is skew distributed.
4.1 Control of process mean

Assume there are p quality characteristics, jointly distributed as a p-variate normal. The pro-
cess in control means IJ;JT = [!Ll,!L2, ... ,!Lp] and the (p x p) process covariance matrix 1:0 are
unknown and have to be estimated from preliminary samples. Let Xi be the (p x n) data matrix
of the ith sample. There are m samples of size n available. For sample i the (p xI) vector XI of
the sample means and the sample covariance matrix Sj can be calculated. According to
Alt(1985), the plot of sample i is:
(1)
These charts are known as Hotelling's TZ-charts. When a large number of preliminary samples
are present( m >20 - 25), it is common to use control limits UCL = X!,(l However, if this is not
the case the T Z-distribution has to be used. Alt et al. (1976) shows that TZ is related to the F
distribution and that (XI - X)TS-l(XI - x) is distributed as:
p(m-l)(n-l) F (2)
x p,mo-m-p+!
n (mn-m-p + 1)
•
The control limits to the TZ - charts are: UCL = c(m,n,p) x Fp,mn_m_p+!,(l and LCL = 0 where
F",v"a denotes the corresponding F-percenti1e with VI and Vz degrees of freedom. This UCL
should be used to analyse whether the initial samples are in control. Alt (1985) call this stage 1.
When future samples are analysed to determinate whether the process remains in control
c\m,n,p) is:
p(m+l)(n-l) F (3)
x p,mo-m-p+!
n (mn-m-p + 1)
•
This is called stage 2 control limits. Alt (1985) suggests that if one or more of the initial
samples are out of control, these samples are left out and a new UCL is calculated. This
procedure is repeated until all remaining initial samples are in control. When this occurs the
control limits for future samples, UCL', can be calculated using the remaining samples. The
original control limits can be revised when new samples are present.
One difficulty encountered with the use of multivariate control charts is; which of the p
quality characteristics are responsible for an out of control signal? Jackson(1980) recommends
using control charts based on the p eigenvalues of the covariance matrix. This approach is not
applicable in the case of Vectorial Tolerancing because of the large difference in magnitude of
the quality characteristics. Alt suggest using x -charts with Bonferroni-type control limits by
replacing ZoJ2 with ZoJ(2P). This means in the case of two quality characteristics and with P {type
I error} = ex = 0.0027, ZoJ(2p) = 3.2. The type I error is then al2.
4.2 Control of process variability

The approach for control of process variability is based in the sample generalised variance lSI,
which is the determinant of the covariance matrix. According to Montgomery (1991), most of
the probability distribution of lSI is contained in the interval E(lSI) ± 3.JV(ISI). E(ISI)
and V(ISI) are mean and variance of lSI. According to Montgomery (1991) and Alt (1985):
(4)
where
1 P (5)
bl = - - p I1(n-i).
(n-1) ;=1
1 p [P n P]
b2 = (n-1)2 P U(n-i)X UC - j+2)UCn - j) .
The control limits of the lSI-control chart are therefore:
(6)
CL=lsl and
4.3 Multivariate experiments

The experiments described here is an example of multivariate process control of plane E3. The
NC-code is in this case the original NC-code with no corrections. The experiment lasted two
days. The first day, 4 samples with a sample size of 6 workpieces was taken. I.e.: m=4, n=6
and p=2. The results are shown in Table 2 and Figure 10 to Figure 13.
Table 2 Results from day 1

Sample Mean Yo Mean Ex VarYo VarEx Covar lSI
1 7,8E-05 3,6E-08 -7,8E-07
2 0,0001 4,8E-08 -I,5E-06
3 4,IE-05 1,6E-08 5,7E-07
4 24,0856 0,00079 5,4E-05 5,2E-08 -4,3E-07 2,98038 2,7E-12
5 24,0758 0,00062 7,IE-05 2,2E-08 6,8E-08 0,09929 l,5E-12
6 24,067 0,00072 8,2E-05 3,9E-08 3,2E-07 0,77943 3,IE-12
7 24,0676 0,00053 2,5E-05 8,8E-09 2E-07 1,4418 1,8E-13
Averages 24,074 0,00066 6,5E-05 3,2E-08 -2,2E-07 2E-12
Sample 1,2 and 3 was present from previous experiments, and the variations from these
samples were added to get a better estimate of the covariance matrix. It was decided to let a
be equal 0.001. The calculated VCL for T2 was therefore 2,6729. The VCL for lSI was
8,34* 10. 12 •
3 UCL =2,6729 16011
2,5 UCI.. = 8,346012
86012
2
N 66012
< 1,5
I-
!1l
46012
0,5
26012
0
4 5 6 7 0
4 5 6 7
Samples
Samples
Figure 10 Stage 1 HotellingT 2control
Figure 11 Sample generalised variance
chart.
control chart.
0.0012
~ 0.001 .b-------------------------.
:: 0.0008
~ 0.0008
~ 0.0004 1-_-_-_-_-_-_-_-_-_-_-_-_--.
ell 0.0002
0~---+----4---~
4 5 6 7 4 5 6 7
Samples
Samples
Figure 12 x -chart of Yo. Outmost Figure 13 x -chart of Ex.

control limits is Bonferroni type.
As the control chart in Figure 10 shows, the first sample (sample 4) is out of control. The
individual x -chart for Yo shows that the first sample is outside the Zo.OO27l2limits, but inside the
Zo.OO27/4 (Bonferroni) limits. The cause of this out of control signal was probably the temp-
erature gradient. Although the milling machine was pre-heated and the first workpieces was
left out, the process was not yet stabilised when sample 4 was manufactured. The out of
control sample (sample 4) was eliminated and the stage 1 T2 control limits were revised. The
three remaining subgroups plotted in control with new control limits as well.
Stage 2 control limits were calculated on the basis of the three reaming samples. Now a
was chosen to be 0.0027. Since m =3, p =2 and n =6, the stage 2 T2 DCL = 4.4260. DCL for
generalised variance lSI is still 8,34*10. 12 •
Table 3 Results from day 2

Sample Mean Yo Mean Ex VarYo VarEx Covar r lSI
8 24,1019 0,00068 5,7E-05 2,lE-08 -5,lE-07 16,4424 9,3E-13
9 24,0909 0,00053 5,7E-05 1,9E-08 -3,2E-07 6,62313 9,7E-13
10 24,0888 0,00055 8,6E-05 2,4E-08 -I,3E-06 5,34433 2,4E-13
11 24,0806 0,00062 0,00015 4,3E-08 -2,2E-06 1,73743 2E-12
12 24,0646 0,00059 6,8E-05 6,9E-08 -1,4E-06 0,55577 2,9E-12
13 24,0616 0,00062 0,00015 5,2E-08 -2E-06 1,15726 3,9E-12
14 24,0567 0,00061 8,8E-05 1,2E-08 -9,IE-07 2,89321 2,lE-13
18 1E-11
16 UCL = 8.34E-12
8E-12
14
12 6E-12
... 10
< !il
I- 8 4E-12
6
4 2E-12
2
0
0 <Xl Ol
<Xl Ol o ,.... C\I ~ ;:!
~ Samples~
Samples
Figure 14 Stage II Hotelling T2 control Figure 15 Control chart for generalised

chart. variance, lSI.
0,0012
24,11
24,1 0,001
f-------
.,24,09 I-"'=:"-!!!!!!o"",~~~ i 0,0008
:24,08
~.. 0,0006
::24,07
~24,06 I-~~~~""_r-- a
~ 0,0004
til
.)il24,05
0,0002
24,04
24,03 +--+---+--+--+-+--1 0+--+---+-+--+---+-----.,
co 0) ~ ;: ~ ~ :!
Samples Samples
Figure 16 x -chart of location Yo Figure 17 x -chart of orientation Ex.

When the manufacturing was started on day two, the milling machine was cold. The task
was to study the effect of increasing temperature. A temperature sensor was fixed on the
bearing of the main spindle. Table 3 and Figure 14 to Figure 17 shows the results of stage 2
control, with the control limits calculated from the results at stage 1 on day 1.
With a total of 11 samples after day 2, the stage 2 control limits for future samples can be
updated for a possible stage 2 control. In a Stage 1 control chart, sample 8, 9 and 10 had to be
removed before all samples plotted inside the control limits with a = 0.001. For future samples
(stage 2) the UCL =2,77405 when a =0.0027.
3
2,5 +----,4..-c:....::."'-..:;:; 24,09
2
1,5
•
c
24,08
24,07
II
TZ 1 ~ 24,06
0,5 co
>- 24,05
O+--+--f-_f---'f--+---i
24,04 +--+--f-----'-+--+---i
5 6 7 11 12 13 14 tn N "II:t
Samples
Q) ...... ...... ('I)
~ ~
Samples
Figure 18 Stage 1 T2 control chart with Figure 19 New Stage 1 x -chart of Yo,
new samples. Sample 8-10 removed. samples 8-10 removed.
4.4 Discussion of results

In the multivariate experiments, the task was to keep the process in control with multivariate
control charts. The results shows that the mean Yo-location of plane E3 is not in control. A
trend in Yo is caused by an increasing temperature in the machine tool during manufacturing.
The trend forms a "U-shaped" curve in the T2 control chart. To get only in-control plots on
day two, the three first samples of day two had to be removed. The orientation component (Ex)
on the other hand, was stable. The variability of the process was stable throughout the
experiments. This shows how the cause for the process being out of control is isolated, and
corrective actions can be made.
In the experiments described here, there where two quality characteristics describing a
single surface, plane E3. On other workpieces there might be several surfaces of interest, and
these surfaces might be correlated. If all surfaces on workpiece 1 (except the fixing-surface,
E8) should be controlled, there would be 14 quality characteristics. Seven planes, that means a
total of 21 quality characteristics is possible for workpiece 1. Six is used to define the
Workpiece Coordinate system, and since E3 is a side plane to E2, there are two and not three
quality characteristics on E3. This leads to a total of21 - 6 - 1 = 14.
5 CONCLUSIONS
The results from the experiments can be summarised as the following:
• Statistical Process Control of surface location and orientation using Vectorial Tolerancing is
possible. Correction is made by mirroring the calculated sample means onto the nominal
plane.
• The quality characteristics on one surface are correlated. Different surface on the same
workpiece can be correlated. Multivariate Process Control is therefore used.
• The first corrections can be made after a few workpieces are manufactured.
6 FURTHER WORK
6.1 Multiple cuts

All surfaces controlled in these experiments was milled in one cut. In the case of surfaces
milled with multiple cuts, each individual cut should probably be treated as one surface. There
was made some experiments on this, but no conclusions could be made. This is an interesting
topic for further studies.
6.2 Correction of cylinders

There where also made an experiment on correcting the four quality characteristics on a boring
(two locations and two orientations). The results showed a successful correction of the
cylinder. The experiments are, however, not included in this paper because there was not
enough readings to draw a certain conclusion. This will also be a topic for further studies.
7 QUESTIONS AND ANSWERS

This are the questions from conference participants and the answers given by the author after
the presentation:
Question 1: "How do you handle form deviations?"

Answer: Controlling form deviations is not covered in this paper. The form deviations is
filtered out by the Gaussian best fit.
Question 2: "Is there compatibility between ISO tolerances and Vectorial Tolerancing?"
Answer: It is possible to calculate from Vectorial Tolerances to ISO tolerances but not
reverse. In Vectorial Tolerancing the Gaussian best fit algorithm is used to obtain
a substitute surface when measuring a real surface. The ISO tolerances is using
tangential surfaces. To get full compatibility the form deviations needs to be
expressed as the location, orientation and size deviation between the best fit and
the tangential surfaces. When the form deviation is expressed as the maximum
point deviation from best fit surface, full compatibility is not possible because of
the different way to express the form deviations.
8 ACKNOWLEDGEMENTS
I would like to thank professor Wirtz, professor Uiuger and Arnold Patt at Neu Technikum
Buchs in Switzerland for letting me using their equipment and for valuable help. In addition I
will thank to the Norwegian Research Council for making this research possible by giving me
financial support.
9 REFERENCES
Alt, F.B. 1985, Multivariate Quality Control, In Encyclopedia of Statistical Sciences, Vol 6,
edited by N.L. Johnsen and S. Kotz, John Wiely, New York
Alt, F.B., Goode, J. J., and Wadsworth, H. M. 1976, Annual Tech. Conf. Trans. ASQC, p.
170-176
Aune, A. 1994, Kvalitetsstyrte bedrifter, Ad Notam Gylendal AS, 2. ed. (In Norwegian)
Bj~rke, 0. 1991, Computer Aided Tolerancing, ASME press, New York
Farmer L.E. , Gladman C.A. 1986, Tolerance Technology - Computer based analysis, The
Annals of the CIRP Vol 3511
Hotelling, H. 1947, Multivariate Quality Control, In Techniques of Statistical Analysis,
Eisenhart, Hastay, and Wallis, eds., McGraw-Hill, New York.
Jackson J.E. 1980, Principle Components and Factor Analysis, Journal of Quality Technology
voll2.
Johnson R. A., Wichern D.W. 1992 Applied mulitivariate statistical analysis, 3. ed. Prentice-
Hall International Inc.
Martinsen K. 1993, Vectorial Tolerancing for all types of surfaces, Advances in Design
Automation, vol. 2, p. 187-198. ASME
Montgomery, D.C. 1991, Introduction to Statistical Quality Control, 2nd ed., John Wiley &
Sons.
Wei! R. 1988, Tolerancing for Function, Keynote Paper Annals of the CIRP Vol. 37/2
Weil R. 1991, Assessment of Accuracy of Robots in Relation with Geometrical Tolerances in
Robot Links, Annals of th CIRP, Vol 4011
Wirtz, A. 1988, Vektorielle Tolerierung zur Qualitiitssteuerung in der mechanischen Fertigung,
Annals of the CIRP, Vol. 37/1 p. 493-498
Wirtz, A. Giichter, c., Wipf, D., 1993, From Unambigously Defined Geometry to the Perfect
Quality Control Loop, Annals of the CIRP, Vol 4211 p. 615-618
10 BIOGRAPHY
Kristian Martinsen is finishing his Dr. Ing. (PhD) thesis at University of Trondheim. His
Supervisor is Professor Dr. Ing. 0yvind Bj~rke. A part of the research was done in the lab of
Professor Dr. Ing. Adolf Wirtz in Neu Technikum Buchs, Switzerland.
13
Probabilistic Analysis of Geometric
Tolerances
A. O. Nassef
PhD Candidate, Mechanical Engineering Dept.
McMaster University, Ontario - Canada
Email: ashra.f@ims.uwindsor.ca
H. A. EIMaraghy
Prof. and Dean, Faculty of Engineering
University of Windsor, Ontario - Canada
Email: hae@ims.uwindsor.ca
Abstract
This paper describes a procedure for the statistical analysis of geometric tolerances. The proposed
procedure assumes that a manufactured surface lies between two ideal offset surfaces positioned
at equal distance from the nominal surface. These surfaces do not represent a tolerance zone, but
rather the volume which has the highest probability of containing a point on the generated
surface. The generated surface is represented by a set of points, which are assumed to be random
variables having a multinormal distribution. Using the generated points, the minimum deviation
zone of each geometric deviation in each set is compared with the tolerances specified for the
feature. Genetic algorithms are used to conduct these checks to ensure reaching the global
optimum value of the minimum deviation zone. If the set of points is acceptable the Monte Carlo
simulation is updated. To ensure that the probability of rejection of the feature due to the violation
of the specified tolerances is calculated with a low variance of error, two methods of variance
reduction techniques were used during the simulation. These are, Latin Hypercube Sampling and
Antithetic Variates. An example for simulating a cylindrical feature is given at the end of the
paper and the results of the algorithms using the proposed variance reduction techniques is
compared with those using simple Monte Carlo simulation.
1 INTRODUCTION
Machining processes produce dimensions and geometries which deviate from the nominal
values. At the same time designers specify ranges of values for critical dimensions or assembly
clearances that should not be exceeded to ensure the assembly's proper functioning. The ranges
of allowable deviations specified on dimensions and geometries are defined according to
functional requirements. In statistical tolerance analysis, each dimension is assumed to be a
random variable with a probability distribution. Hence, the functional requirements becomes a
function of random variables. Statistical tolerance analysis techniques are concerned with the
calculation of the probability of violating the functional requirements (referred to in the paper
as the rejection probability).
Previous methods of tolerance analysis include the root sum square method for linear function
requirements (WU et. al. 1988), the reliability index method (Lee 1988), the method of moments
(Chase & Greenwood 1988) and the Monte Carlo simulation (Chase & Greenwood 1988 and Wu
et. al. 1988). While a Monte Carlo simulation is not restricted to a specific type of probability
distribution, it requires a large number of samples to converge to the solution, with low variance.
The higher the variance, the larger the error in the calculated rejection probability. Variance
reduction techniques like the antithetic variates (Rubinstein 1988) and the Latin hypercube
sampling (Rubinstein 1988) have been used to reduce the variance of probabilities calculated by
the Monte Carlo simulation. Unlike dimensional tolerances, the statistical analysis of geometric
tolerances has not been investigated sufficiently. An approach based on the concept offeasibility
space, where a method of representing boundaries on model variables was proposed by Turner
(Gupta & Turner 1993 and Turner 1993). The work was limited to the evaluation of worst case
tolerance analysis and did not include the statistical analysis of geometric tolerances.
The presented paper introduces a procedure which in essence simulates the manufacturing and
inspection processes of a feature of size. Geometric tolerances are viewed as values of
constraining functions of points lying on the manufactured surface.
2 TOLERANCE ANALYSIS
5.00 ± 0.Q2
I... .1
I I
.. x
4.98 5.00 5.02
Figure 1 Probability Distribution of a Single Dimension.
Figure 1 shows a dimensional tolerance with a probability distribution representing the

variability of the machining process creating the dimension. The probability of rejecting the part
due to the violation of the dimensional tolerance is equal to the shaded area under the curve and
Probabilistic analysis of geometric tolerances 189
the probability of satisfying the imposed tolerance is equal to the hatched area. Both probabilities
can be calculated easily. The calculation of the probability of rejection (or acceptance) for
features with geometric tolerances is not as straightforward as that for dimensional tolerances.
Geometric tolerances control surface variations and require an appropriate representation for the
manufactured surface. Figure 2 shows a cylindrical feature in a block along with a set of
geometric tolerances specified to control its geometric variations. The objective of the work
presented in this paper is to evaluate the probability of satisfying the geometric tolerances shown
in Figure 2 as well as fmding tolerance values which satisfy a desired probability of rejection.
The ftrst step is to fmd an appropriate surface representation of the manufactured surface as
detailed in the following section.
6.250 ± 0.125
-$- qxU50@ A
.l 0.075@ A
~
[±]~I
~6.250 ± 0.400
! !
Figure 2 A Toleranced Cylindrical Feature.
Figure 3 Feature Generation.

2.1 Feature Representation
The manufactured part is represented by a set of points generated from the machining process
probability distribution (Figure 3 ). The feature shown in Figure 3 is represented by 8 x 3 points
positioned at equal angles and measured from the datum reference frame A,B and C. The radii
vector R of the points are generated from the probability distribution representing random
variation of the machined surface. A parametric surface based on NURBS (Non-Uniform
Rational B-Splines) is interpolated to the generated points representing the manufactured
surface. The NURBS surface equation is given by:
n m
I I Wij Ni,i u ) Nj,q(v) Pij
i=Oj=O
P(u, v) n m (1)
I I Wij Ni,i u ) Nj,q(v)
i=O j=O
where u and v are surface's independent parameters, Pij i E {O) ,.. ,n}, jE {O) , ..,m} are the control
points (points of a polygon approximating the surface), n is the number of control points in the
u direction, m is the number of control points in the v direction, Wjj are the weights of the control
points, and Nj,p are the B-Spline functions of the pth order given by the equation:
_ u - tj ti+p+I-U
Ni(u) - t,
I+P
tNip-l(u)
I'
+ t
i+p+1
t
i+1
Ni+1p_l(u)
'
(2)
N,O(u) = {I if u 2: Ii & U < li+l
t, 0 otherwise
where tj j E {O,l ,.. ,n} is the knot vector of the B-Spline function. The parameters of equation (I)
are evaluated as follows: (piegI1991)
Given a set of generated points Q (The X,Y,Z form of points whose radii R are generated by
Monte Carlo simulation) a NURBS surface is sought which, for certain parameter values Ii; and
Vj, passes through Oij.
n m
I I Wij Nj,iUi)
i=Oj=O
Nj,/v) Pij
(3)
n m
I I
i=O j=O
Wjj Ni,iUi) Nj,q(v)
The above equation is solved for P jj but certain parameters must be known at first. Figure 4
shows a circle represented by a 2nd order NURBS curve.
w=05
8
I Wi Ni.iU) Pi
i=O
W=] P(u) (4)
w=l
LW
8
i Ni,iU)
i=O
....-----=::-.e-~'------. w=05
w=l
Figure 4 A NURBS Circle.
Assume that the circle represents a nominal curve in a design drawing and the points D; are the
points at which measurements are to be taken after machining. The nominal points D; will deviate
to a set of points M; on the machined circle. These deviations are minute compared to the
circumferential length ofthe circle and consequently the parameter values u; at the points M; and
the weights W; of the control points of a NURBS curve passing through M; can be taken equal
to those of the ideal circle. The same knot vector can also be used for the new curve. Figure 4
shows that nine control points (The fIrst and last points coincide) are needed for four points on
the machined circle, and thus the number of control points needed for N points on the machined
line is equal to 2 XN+1. The coordinates of the control points are left to be evaluated by solving
the set of linear equations:
n
I W; N;.2(U;) Pi
i=O (5)
n
I
i=O
Wi N;.2(U;)
Figure 5 A Machined Surface.
The coefficient matrix of the set of linear equations is proved to be non-singular (piegl 1991),
and the above system of equations can be solved easily. The above process can be extended to
interpolate a NURBS surface to points machined for a nominal cylindrical surface. Figure 5
shows an exaggerated example of a machined cylindrical surface represented by NURBS.
2.2 Machining Processes Probability Distribution

Previous work in tolerance analysis considered that probability distributions representing the
variation in manufacturing processes to be independent from each other. Hence, the notion of
statistical independence is valid. The parameters of the probability distribution of the
manufacturing process are usually obtained by taking several measurements of points on the
actual parts and estimating the population parameters. Thus if the manufactured feature is
represented by M points and N parts were manufactured the variance of the measured radii would
be equal to:
(6)
where R;,j is the jth generated radius of the ith manufactured part, and
N M
IIRiJ
_ i=lj=1
(7)
R= NxM
The same measurements can be used to calculate the covariance matrix of the generated points
a 1.1 a l ,2
a a · . a 2,n
· . al,n]
C = [ ~,l ~,2 (8)
an,1 an,2 · . a n•n
where
I
N
(Ri k - R;) (RA - R)
k= I (9)
and
N
'" R.'k
L
- k=1 (10)
Ri=~
The mu1tinorma1 distribution can then be used to represent the random nature of the machining
process, thus the random variables dependence is taken into account.
_ 1
f,{r) - (21l)n/2ICl/21
[1 -
exp - 2(r - R) C
T -1 -
(r - R)
] (11)
where
r = (rl' r2' ·..r n) is the random vector of the feature's radii (12)
R = (R 1,R2 , ...Rn) is the vector of distribution means (13)
2.3 Random Variate Generation

The simulation of the manufactured parts depends on the type of the distribution representing the
variation in the process. If the points on the manufactured surface are considered to be
independent random variables, then inverse transform method (Rubinstein 1981) for the
generation of the machined points radii can be used. If the variation is represented by a
multinormal distribution (11), then the following procedure is used.
Let:
(14)
be the lower triangular matrix of the covariance matrix C such that:
(15)
then the vector R of the simulated points is equal to:
R=cZ+R (16)
where Z is an independent generation of a random vector from the standard normal distribution.
Thus the algorithm (Rubinstein 1981) for generating points from the mu1tinormal distribution
is :
1. Generate Z = (ZI' Z2, ... ,:z" ) (17)

2. Calculate
j-l
a·· '" c·c·
lJ - L I, h
k=1
(18)
c;J = [ j-l ]1/2
aJJ
.. - 'L" c·h 2
k=1
where:
o
'" c·c·
L lot h: = 0 '
l~j~i~n
k=1
3. R=cZ+R (19)
2.4 Probabilistic Analysis
Si
Figure 6 Probability Region.
Figure 6 shows the probability region for the radii of two points on a machined surface. The
geometric tolerances imposed on the feature divide the probability region into two: (a) the
feasible region Sr(represented by the shaded area) and (b) the infeasible region Sj. The objective
of the probabilistic analysis is to fmd the feasible volume under the multivariate distribution of
the random variables. The volume is represented by the integral
I = f f···· f
rCS,
Nr) drldr2···drn
(20)
Since the boundaries of the feasible region are hard to represent analytically, Monte Carlo
simulation can be used to evaluate this integral. An estimator of the integral after Nt genemtions
of the random vector r is:
1= Na (21)
""N:-a-+:=-:-N"-r
where Nais the number of generated samples within the feasible region Sf and Nr is the number
of genemtions falling in the infeasible region Sj. Standard Monte Carlo simulation has two main
problems. First as the number of points increases the variance of the calculated probability of
acceptance becomes larger, thus the expected error in the calculated probability increases. Second
the standard Monte Carlo simulation tends to genemte instances of the random vector away from
the tails of the marginal probability distributions, thus the extreme cases in which the tolerance
constraining functions are violated will be missed. This may result in a probability of satisfying
the imposed tolerances equal to one. Therefore variance reduction techniques are used with
Monte Carlo simulation. The following sections describe the proposed approach and techniques
used.
3 VARIANCE REDUCTION TECHNIQUES
3.1 Latin Hypercube Sampling

Latin Hypercube Sampling is a constrained sampling method, which covers the range of the
sampled random variable (Rubinstein 1981) by breaking it into intervals, and distributing the
sampling process over these intervals.
3.1.1 Sampling for One Variable

Consider a random variable x whose probability distribution function j(x) and probability
function F(x) are shown in Figure 7 . The probability function is divided equally into M
non-overlapping intervals. All intervals have the same probability of occurrence, and each
interval can be used to generate M values of the random variable for each Monte Carlo simulation
cycle.
Random Variable x
F(x) ~jv;---=::=:::==I
Random Variable x
Figure 7 Latin Hypercube Sampling.
,---
xI
,
/
I R4,1 R 4, R 4,
~
R44 ~
'
I
,--/ I
~
, R 3,1 R 3 , R 3.3 R3,4 I
,
,
R 2,1 R 2, R 2 , R 3,4
I
l
-
RI,I R I, R I, R I ,4 I
,./
/ 1\
Figure 8 Latin Hypercube Sampling for Two Variables.
The M values can be generated by generating M random numbers from a uniform distribution
over the range [0.1] and allocating a random number Ui to the ith interval. thus for each random
number U E [0.1] generated for the ith interval. there is a value Ui such that:
U. = U + (i - 1) i E {1.2..... M} (22)
I M M
Equation (22) shows that Uiranges from (i-1)/M to ilM. The random numbers Ui represent a set
of M probability values for the random variable x. The equivalent values of the random variable
can be produced using the inverse transform method
Xi = F-1(U) (23)
where p--l is the inverse probability function.
3.1.2 Sampling for More than One Variable

The Latin Hypercube Sampling method delivers a set of M values in each Monte Carlo simulation
cycle. For more than one random variable. the method divides the probability space into Mn
probability subregions. where n is the number of random variables. Figure 8 shows the division
of the probability space for two random variables into 16 intervals. The sampling method delivers
in each cycle xlj and x2j ,j EO {1.2•...• M}. which is equivalent to generating random variates in
the regions Rl.l • R2,2 ..... RM,M' To eliminate the generation bias to the diagonal regions. the
generated random variates are randomly assigned to different regions. This is done by
rearranging the variables Xl and X2 such that:
(24)
where P is a random permutation of the set {1.2 .......M}.
3.1.3 Advantages of Using Latin Hypercube Sampling
Acceptance
Region
Rejection
Region
Figure 9 Latin Hypercube Sampling for Two Variables.
Latin Hypercube Sampling is one of the stratified sampling methods. There are two main
advantages to using them over the standard Monte Carlo simulation random sampling. The first
advantage is that the variance of the calculated probability is less than that of the probability
calculated using standard Monte Carlo. In tolerance analysis problems the calculated probability
of rejection is very small. and the limit functions tend to lie near to the tails of the probability
distributions. If the number of simulated variables is large. the standard Monte Carlo method
tends to generate samples in the central regions of the probability space thus missing the rejection
regions and giving a zero probability of failure (Figure 9 ). The second advantage of Latin
Hypercube Sampling is that it constrains the sampling within regions such that for each sample
a value is generated at the tail of the distribution.
3.2 Antithetic Variates
fix)
x
F(x)r-----t1t-=======9
l-U ~-----+---tJ'
0.5 ~-----+--A'
Figure 10 Antithetic Variates.
Figure 10 shows the probability distribution function and the probability function of a random
variate x. Let a random number U be used to generate an instance of X of the random variate x
using the inverse transform method and let another number l-U be used to generate X'. If the
process is repeated for N times the covariance between X and X' will be
N
I(X i - XXX'i - X)
COV(X, X') = _i=...:..l_~_:--_ _ (25)
N-l
Since (X; - X) and (X'; - X) can never have the same sign, the covariance is strongly negative.
If a probability of rejection is calculated then let U be a random number drawn from a uniform
probability distribution and let Pr' be the probability of rejection calculated from N random
generations of U and let Pr" be the probability of rejection calculated from N generations of
(I-U). The probability of rejection has a mean equal to:
Pr = Pr' + Pr" (26)

2
and a variance equal to:
VAR(Pr) = VAR(Pr') + VAR(Pr") + COV(Pr',Pr") (27)

4
and since COV(Pr' ,Pr") is strongly negative, then the variance of Pr is reduced (Rubinstein
1981).
4 EXAMPLE
Figure 11 Generation of example's points.
The example shown in Figure 2 is used to test the proposed probabilistic analysis method. Points
on the feature are generated from a multinonnal probability distribution having a mean radius
of 3.125 mm measured from a centreline lying 18.75 mm from the datum Band 6.25 mm from
datum C. Eight points are used on the feature's circumference at three levels along the feature's
length (Figure 11 ). The radii of the points are assumed to have a normal probability distribution
with a mean value equal to the nominal radius (3.125 mm), and thus the covariance matrix has
zero non-diagonal elements, while the diagonal variances are equal to the square of the largest
imposed tolerance (0.25) divided by 6.
4.1 Generation of the Radii Random Vector

The random vector of the generated points radii is simulated using the following procedure:
1. The feature is represented by 8 X 3=24 points. Thus each generated vector R has 24
radii.
2. Latin Hypercube Sampling from the Standard Normal Distribution:
For each random variable (radius) a standard probability distribution is divided into
M=1O equal probability regions, from which 10 values of a standard normal
distribution variable z are obtained. These values are rearranged in a random
permutation producing 10 normal random vectors {ZJ, Z2,'" ,ZJO }.
3. Antithetic Variates:
If a probability P is used to generate the vector Zj, another probability (I-P) is used
to generate the vector Z/.
4. Given the values of the mean radii and the values of the covariance matrix, the vectors
Z and Z' are used to calculate the values of the generated radii (R and R') using the
procedure defmed by equations (l7), (18) and (19).
5. Each generated vector (R or R") is checked whether its points are within the specified
geometric tolerances or not. If the geometric deviations are within the specified
tolerances, a counter of the acceptable instances Na (Or N'a for acceptable instances
of R") is incremented, otherwise a counter of the rejected instances Nr (Or N'r for
rejected instances of R") is incremented.
6. The process is repeated for 500 times and two probabilities of rejection are calculated
whose average is taken as the probability of rejection
(28)
Pr' (29)
N'a + N'r
P . = Pr + Pr' (30)
reJ 2
4.2 Evaluation of Geometric Deviations
2 x max(min(r(U,v,PIP2))) 2 x max(min(r(p, u, v)))

Pl,P2 U,V •
p u,v
(a) (b)
Figure 12 Evaluation of Size tolerance.
Each generation of the vector of generated radii is checked for the imposed tolerances. The
checked deviations are size, perpendicularity and position deviations. A NURBS surface is
interpolated to the generated points and is used to evaluate the geometric deviations of the
machined surface. Size deviation is checked for two criteria according to the ANSI Y14.5M
standards. The first criterion is that the perfect form envelope of the feature should have a
diameter (or a cross sectional size in case of a pair of parallel planes) less than the maximum
material condition of the specified size tolerance. This means that the maximum inscribing
perfect cylinder to the machined feature should have a diameter less than the MMC of the size
tolerance. This deviation is found by optimizing the objective function shown in Figure 12 (a).
The second criterion used for evaluating the size tolerance is the local size. The centre point or
the sphere is a point lying on the actual axis of the feature. The centre points at the feature's
entrance and end are used for the evaluation of perpendicularity and position tolerance. The
largest sphere inscribed in the feature at any cross-section should be between the limits of size
tolerance. The objective function used for the evaluation of local size at a given cross-section
is shown in Figure 12 (b). Genetic algorithms (Goldberg 1989) are used to evaluate both
deviations. Figure 13 shows the start and end points of the feature's actual axis. The projection
of these points on a datum plane is used to evaluate perpendicularity and position deviation. The
distance between both projected points plus any difference between the perfect form size
deviation and the size tolerance is equal to the perpendicularity deviation. Position deviation is
checked by examining whether the projected points lie within the position tolerance zone
established from the specified datum reference frame.
E5---
Datum Plane
Figure 13 Evaluation of Perpendicularity and Position Deviation.
0.2
c
~
~ 0.15
.0
8
0..
!:i
.9
U 0.1 ...... :, ......... , .. :. ..
<l.l
.~
0::: i I I I I jStandardM.C.(Il=O.14, V=2.3x10-4)

o 9 9 9 9 Q A.V. (Il=O.139, V=1.8 X 10-4)
0,05 .............
;.....-----;: L.H.S. (Il=O.139, V=1.9 x 10-5)
~ * * * * ~LHS+AV (Il=O.123, V=7.5x10-6)

o.----~--~---~---~---~
o 100 200 300 400 500
Number of Simulations
Figure 14 Tolerance Analysis Results.
4.3 Results
Figure 14 shows the simulation results of combining Monte Carlo simulation and variance
reduction techniques for up to 500 generated samples. In addition, the figure shows the mean
value and the variance of the probability of rejection for 100 runs of the different simulations.
Results show that a combination of antithetic variates and Latin hypercube sampling converged
after 200 samples with a low variance.
5 CONCLUSIONS
The presented paper investigated the statistical tolerance analysis and tolerance value
optimization of parts with geometric tolerances. The presented work simulated in essence the
manufacturing and inspection of the part to evaluate the part's probability of rejection. Methods
of variance reduction techniques within Monte Carlo simulation were included for the calculation
of the probability of assembly rejections. Results showed that a combination of Latin hypercube
sampling and antithetic variates with Monte Carlo simulation leads to an accurate calculation of
the probability of rejection with a low variance. The proposed method of tolerance analysis does
not require the knowledge of the exact boundaries of the tolerance zones impos~d on the feature,
which are difficult to defme when form deviations are considered.
REFERENCES
American Society of Mechanical Engineers, (1982), "Dimensioning and Tolerancing", ANSI

Y14.5M.
Chase, K. W. and Greenwood, W.H., (1988), "Design Issues in Mechanical Tolerance Analysis",
Manufacturing Review, ASME, Vol. I, No.1, pp. 50-59.
Goldberg, D., (1989), "Genetic Algorithms in Search, Optimization and Machine Learning",
Addison Wesley,
Gupta, S. and Thrner, J. U., (1993), "Variational Solid Modeling for Tolerance Analysis", IEEE
Computer Graphics and Applications, Vol. 13, No.3, pp. 64--74.
Lee, Woo-Jong, (1988), "Tolerancing: Computations on Geometric Uncertainties", Ph.D.
Dissertation, University of Michigan, Ann-Arbor, Michigan.
Nassef, A. and EIMaraghy H., (1994), "Evaluation of Actual Geometric Deviations using
NURBS Surface Interpolation and Genetic Algorithms", Proc. of the 3rd IEASTED
Conference on Computer Applications in Industry, Cairo, 1994.
Nassef, A. and EIMaraghy H., (1993), "Allocation of Tolerance Types and Values Using Genetic
Algorithms",Proc. ofthe 3rd CIRP Seminars on Computer Aided Tolerancing, CACHAN,
France, 1993.
Piegl, L., (1991), "On NURBS: A Survey", IEEE Trans. Computer Graphics and Applications,
pp.55-71.
Rubinstein, Y.R., (1981), "Simulation and the Monte Carlo Method", John Wiley & Sons Inc.,
New York.
Turner, J. U., (1993), "A Feasibility Space Approach for Automated Tolerancing", Transactions
of the ASME, Journal of Engineering for Industry, Vol. 115, pp. 341-346.
Wu, Zhang, ElMaraghy, W.H. and EIMaraghy, H.A., (1988), "Evaluation of Cost Tolerance
Algorithms for Design Tolerance Analysis and Synthesis", Manufacturing Review, ASME,
Vol. I, No.3, pp. 168-179.
7 BIOGRAPHY
Mr Ashraf Nassef is a PhD candidate in the Mechanical Engineering Department, McMaster

University, Hamilton, Ontario, Canada. His PhD research is conducted in the areas of tolerance
analysis and synthesis. He obtained his B.Sc. and M.Sc. from Cairo University in Egypt
Dr. Hoda A. ElMaraghy is a Professor of Mechanical Engineering, Dean of Faculty of
Engineering at the University of Windsor and Director of the Integrated Design and
Manufacturing Centre, University of Windsor, Windsor, Ontario, Canada. Her current research
interests include Intelligent Automation, Robotic Assembly, Flexible Manufacturing,
Feature-Based Design, Expert Task Planning and Production Control. She is a Principal
Investigator of the Manufacturing Research and Corporation of Ontario (MRCO) and a project
leader in the Institute for Robotics and Intelligent Systems (IRIS).
14
Taguchi rules in some Japanese

standardizations of tolerancing
T. Shibayama
JAPAN*
Abstract
Three Japanese Industrial Standards and a Standard Draft are issued on tolerancing, that
incorporating Taguchi rules. Reference Table of the standard 11S K 7109-1986 tabulates
actual operation tolerances for plastic parts as compared to the tolerance by the standard.
Careful review shows that Taguchi loss function shall be a higher order one than a quadratic
original, also that Taguchi tolerance is independent of the part population but guarantees
the part quality not enough. The supplier's expense for care of the customer's objection,
the supplier's dole, and the customer's duty depend on the population. It shall be centered
and narrow far more than the tolerance. The Standard Draft applies the rules to any value
other than dimensions, precise data less available. The other two standards, 11S K 0971-1989
for microvolumeter and 11S Z 9090-1991 for measurements, describe principles of mean
square errors or so, incorporating Taguchi rules only auxiliaril y. Before standardization, the
thought and the mathematics shall be reviewed more accurately and more precisely.
Keywords
Tolerancing, Taguchi loss function, Taguchi methods, Taguchi quality engineering, plastic
part tolerance, calibration, measurement, Japanese Industrial Standard
1 INTRODUCTION
Three Japanese Industrial Standards (K 7109-1986, K 0971-1989 and Z 9090-1991) and

a Standard Draft incorporate the rules and the way that are of Taguchi,G., the practitioner of
his quality engineering. Agendas and discourses are available but papers or treatises are rare.
A few people promoted the standardization calling the name of the great practitioner for help
'Present address (home): Sunshine Kanayama 502,14-13 Furuwataricho, Naka-ku, Nagoya, 460 JAPAN.
Telephone-Facsimile +81 (0)52-331-8295.
Taguchi rules 205
as a member or a chief of the committee, but failed in legible description. Perhaps because
unfamiliar with mathematics, whereas other people are busy in other problems. Careful review
and clarification are needed. The specification and the background of the standard shall be
restated concisely in a form mathematically consistent and engineeringly conformant.
2 TAGUCHI LOSS FUNCTION
• JIS K 7109-1986 General Tolerancing Rules for Plastics Dimensions.
Specification and background

A supplier shall be aware of the intrinsic diversions of a product from the target that may be
described in terms of a population density function as known, whereas a customer will be aware
of the allowable diversions as described by an objection fraction function as below.
If parts monodispersed having a common diversion, a part may incur an objection another
not. Objection fraction is the fraction of the number objected in the supplied. Therefore,
UU = NB/Nl' consequently NB = NT*uu where (1)
UU for the objection fraction,

NB the objected number of parts, and
Nl' the supplied number.
The fraction is zero if the off-target diversion is zero, whereas larger as the diversion larger
and attains to unity at the full objection diversion, so a function is suggested as follows,
uu(y) =1 for lyl.<1 , and = lylW for Iylsl , on the argument y such that (2)
y = Y¥.zzz / Y¥.NMZ provided Y¥.NMZ = Y¥.FUZ where (3)
uu(y) for an objection fraction function

giving the value uu in eqn (1),
w the exponent of the power type function.
y the normalized diversion, defined above in eqn (3),
Y¥.zzz any the part diversion,
Y¥.NMZ the normalization diversion : = Y¥. FUZ, and
Y¥.FUZ the full objection diversion.
With a screened shipment of a polydispersed population, it follows that
NB=NN* ~:UU(Y)*ff(Y)*<:iy, NT=NN* ~:ff(Y)*dy, and OC=NN-NT, where (4)
NN for the total number of parts produced, and

OC the rejected number besides Nl' and NB in eqn (1) ,
yAL for the lower diversion of screening ( -yAZ , say) ,

yAR the higher diversion (yAZ , say) , ( with
yAZ the absolute value of yAL or yAZ , ) and
ff(y) the population density of the parts as produced.
Income and expense are calculable of the supplier and the customer as follows
UE Supplier's dole = Income - Expense (for NN parts)

= NN*EE (EE : dole per a part manufactured )
= NI'*TI' (Tl' : dole per a part shipped )
= NI'*OC - NB*BB - OC*RR - NN~ (5)
WD Customer's duty =NB*DD + ~ (for NI' parts) , and (6)
SQ Social loss =NB*OO + OC~ (for NN parts) , where (7)
OC Price of a part as supplied to the customer,

BB Supplier's expense per an objection, for the care, and
03 Cost of a part as produced,
00 Social loss per an objection, = DO + 88, provided that
RR Supplier's expense on rejecting a part by screening and
DO Customer's duty per an objection, i.e., the loss,
and others are the same as with eqn (4). (8)
Equation (4) is substituted into eqn (5) or (7). Differentiation with respect to either the
diversion yAL or yAR gives an equation for the critical tolerance yA (= yAL or yAR ),
d UE /dyA = 0 or d SQ /dyA = 0 , respectivel y, where (9)
UE for the supplier's dole in eqn (5), as maximized,

SQ the social loss in eqn (7), as minimized, and
yA the screening diversion, yAL or yAR in eqn (4) .
The function ff(y) in eqn (4) is fixed. The values OC, BB, 03 and 00 in eqn (5), too.
The rejection expense RR in eqn (5) is neglected. Otherwise, the numerator OC in eqn (10)
below is replaced by OC+RR. Equation (2) is substituted into eqn (9) and solved with respect
to yA, giving the critical tolerance yC or yO , respectively, as follows:
yC = (OC/BB)l/w yy.crz = yC * yy.~ ) or (10)

yO = (03/00) l/w YY.c;Jrz = yO * yy.~ ), where (11)
w for the exponent of the function uu(y) in eqn (2),

yC the tolerance maximizing the dole UE in eqn (5),
yO the tolerance minimizing the loss SQ in eqn (7),
yy.crz the tolerance yC in ordinary unit,
YY.c;Jrz the tolerance yO in ordinary unit,
yy.~ the normalization diversion, = YY. FUZ, in eqn (3),
and others are the same as with expn (8) and below.
Taguchi rules 207
The tolerance yO in eqn (11) is primarily the tolerance as specified by the standard. The
exponent w is taken as equal to 2. Taguchi quadratic loss function is the function uu(y) in
eqn (2) (w=2) as multiplied by the loss QQ, as tacitly postulated in the standard and in all
the literatures - Taguchi agendas and discourses (1985-1986), or Taguchi (1985-1986).
With a population monodispersed at the diversion yC in eqn (10), the fraction uu is equal
to the ratio ex::/BB, all the sales Nl'*<X: lost in the expense NB*BB - cf. eqns (1) and (2).
Additional Remarks
The tolerance yC in eqn (10) and yO in eqn (11) are independent of the population ff(y)
in eqn (4): but the dole lIE in eqn (5), the duty WD in eqn (6) and the loss sa in eqn (7) all
depend on the population ff(y) strongly, as proved easily. Furthermore, the density shall be
not 0 at the diversion yC or yO : the value ff(yC) or ff(yO) be not o. Otherwise, the
equation (10) or (11) loses the meaning because of a division by 0 from eqn (9). Either optimum
is only with a fixed population ff(y), but completely of no meaning if both the bounds are
within the tolerances -yC and yC, or -yO and yO, respectively: A population centered
more and narrow more gives the dole lIE more in eqn (5) and the loss SQ less in eqn (7).
The expense 88 and the loss QQ are distinguished each other often vaguely in the standard
and the related literatures. The price ex:: and the cost Q;, too. So, the tolerances yC in
eqn (10) and yO in eqn (11), too. Multipliers for a precision may be defined as follows,
lB = QQ / 88 and IC = ex:: / Q; so that yC/yO = (lB*IC) l/w, where (12)
lB for the ratio of the loss QQ to the expense 88 ,

IC the ratio of the price ex:: to the cost Q;, and
others are the same as with expns (10), (11) and (8).
Instead of the diversion W.FUZ often is used the half objection diversion W.HFZ just
like LD50 in pharmacy as told - Article (1) of Example in Article 3.1 of the standard, and
uu(yH)=y!iW=1/2 , so, yH = (1/2)1/w W.HFZ = yH*W.l'l'1Z , and (13)

uu(yU)=yUW=l so, yU = ll/w = 1 .- W.FUZ = yU*W.l'l'1Z . Consequently (14)
W.HFZ = (1/2)I/w * W.FUZ : where (15)
uu(y) for the objection fraction function in eqn (2),

w the exponent of the function in eqn (2),
yH the normalized half objection diversion,
yU the normalized full objection diversion,
W.HFZ the half objection diversion, as defined,
W.FUZ the full objection diversion in eqn (3),
W.l'l'1Z the normalization diversion in eqn (3), and furthermore,
W.OfZ = yC * W.HFZ and W.QHZ = yO * W.HFZ, where (16)
yC for the normalized tolerance in eqn (10),

yO the normalized tolerance in eqn (11),
YY.CliZ as defined, and convertible from YY. C1'Z by eqn (17) below,
YY.(JiZ as defined, and convertible from YY. QI'Z by eqn (17) below, and
YY.HFl the half objection diversion in eqn (13). Then,
YY.CliZ = (1/2)1/w*yy.C1'Z and YY.(JiZ = (1/2)1/w*yy.QrZ by eqn (15), (17)
where ... all are the same as with eqns (10) and (11).
In the standard and the literatures, the tolerances YY.OiZ above and YY.C1'Z in eqn (10) are
distinguished vaguely, again. The tolerances YY.(JiZ above and YY.QrZ in eqn (11), too.
The tolerance YY.OiZ may be converted by eqn (17) to the tolerance YY .CI'Z in eqn (10),
as to be related immediately to the half objection diversion YY.HFl as follows,
YY.CI'Z/yy.HFl = 2 1 /w * (OCjBB)l/w where (18)
YY.C1'Z the tolerance in eqns (10) and (17),

YY.HFl the half objection diversion in eqns (13) and (15), and
others are the same as with eqns (10), and see expn (8).
Operation Tolerance
In Reference Table of the standard, are tabulated, for thirtynine actual cases,
* YY.OPN (,say) the operation tolerance, empirical, by the 'assembler' ( * ),

AO QQ the unit loss in eqn (11), or 88 in eqns (10) and (18),
A Q} the part cost in eqn (11), or OC in eqns (10) and (18),
AD YY.HFl the half objection diversion in eqn (18), and
!:J. YY.(JiZ the tolerance calculated by eqns (11) and (16).
In the standard, as mentioned of eqn (12) above, the loss QQ is distinguished only vaguely
from the expense BB , and the cost Q} from the price OC , too. The loss QQ in the
Table may be reinterpreted as if to be the expense 88 ,and the cost Q} to be the price OC .
If this to be the case, the tolerance YY. (JiZ in the Table shall be reinterpreted really to be
the tolerance YY.CliZ in eqn (16). With the reinterpretation, the effect estimable by eqn (12),
the tolerance YY.C1'Z as reinterpreted is compared to the operation tolerance YY.OPN .
The logarithmic ratio log (YY. crz/yy • HFl) is taken on the ordinate and plotted against the
logarithmic ratio log (OCjBB) on the abscissa, then a line_a (w=2) or _b (w=20) with a slope l/w
is obtained through the point -log 2 on the abscissa, as shown in Figure 1. In the figure the
ratio log (YY. OPN/yy. HFl) is taken on the ordinate and plotted against the ratio log (OCjBB)
on the abscissa. An operation point above the line_a (or _b) indicates that the operation
tolerance YY .OPN is larger than the critical tolerance YY .CI'Z . A supplier loses the dole as
mentioned of eqn (10), if operating with a monodispersed population with such the diversion.
The popUlation shall be suppressed far below the line. A polydispersed population, too, shall be
suppressed far lower than the tolerance YY .CI'Z ,lest much of the income NI'*OC in eqn (5)
should be lost in the expense N81r88, even a negative dole UE in eqn (5) - a debt. Operation
points shall lie below the line. The points are actually fixed and independent of the critical
line, therefore, the line_b is preferable, actually. The exponent w is far larger than 2.
Taguchi rules 209
1
N
I I I I "I I I II" I
~
:Ii I" " I
1-13 Outer dia. I
~ 14-30 Inner dia. <b
31,32 Width
"1
§
•.-1 33,34 Length
Ul 35-37 Pitch
~
•.-1
• 38,39 Deflection
'tl
,?
§
...,
•.-1
U
<1l
·n
.D
0
.... 013
7iJ
:r:
••
20
'-
z
f:l)
•
31 28
~ CD
33 26
fl 90 e
~
<1l
.1
Each point referred
rl
...,0 a to data in the Ref.
Table, JIS K7109,
§ by the id'n number.
...,
..-1 37
e ~
& .001
I I I IIII
.01
I I "I III
.1
I I I III
Part price CC / Objection expense BB
Figure 1 The ratio YY.OPNffl.HFZ related to the ratio OC/88 - cf. eqn (18) and around.
The data are tabulated in Reference Table of the standard 1IS K 7109-1986. The operation
tolerance YY. OPN is equal to the critical tolerance YY. crz as calculated of the ratio OC/88 ,
if an operation point is on the line_a. The calculation is by eqn (10) with the exponent w
as taken equal to 2. If the operation point is on the line_b, the critical tolerance YY .crz
is that as calculated with the exponent w taken as equal to 20. ?'I/f - The point 32 shall be
shifted along the arrow if the operation tolerance of the data 32 in the Reference Table shall
read O.OS mm instead of O.SO mm or the half diversion (~O) 0.08 mm instead of 0.8 mm.
The real ratio OCjBB would be lC*lB times the plotted ratio - eqn (12), because the ratio
00/00 might have been plotted. If it be converted to the ratio OC/88 , the operation points
are shifted by log (lC*lB) rightwards, not so largely with the value lC*lB less than S- 6.
As suggested by the discussion above, many the actual tolerances YY.OPN would be larger
than the critical tolerance YY.QHZ by the standard - as calculated by eqns (16) and (11)
provided the exponent w taken as equal to 2. This is confirmed by a direct comparison
- cf. Figure 2_a of Shibayama (199Sa) p.236, or Figure 7 of Kamoshita and Matsuda (1986)
essentially the same. Operating actually on such the operation tolerance, a supplier loses the
dole completely or even bears a debt, as discussed in the foregoing with eqn (18).
With the direct comparison, again, the situation is essentially changed if the exponent W
is taken fairly larger, e.g., as equal to 20 - cf. Figure 2_b of Shibayama (1995a) p.236.
As detailed there, some ambiguities in the specifications of the standard are of no effect.
Population quality
If any population is screened by an arbitrary tolerance yAZ , as mentioned of eqn (4), then,
Nr = ff. TIT*NN, NB = uu. TIT*NN, OC = ff .RRR*NN, and NN = ff .UUU*NN, (19)
where Nr for the shipped, i.e., the supplied number,

NB the objected number,
OC the rejected number,
NN the produced number,
ff.TIT the integrated fraction, shipped,
uu.TIT the integrated fraction, objected,
ff.RRR the integrated fraction, rejected, and
ff.UUU the integrated fraction, produced, all such as follows
ff. TIT = kF, uu. TIT = yAZW*kU , ff • RRR = I-ff. TIT, and ff . UUU = 1 , (20)
fTl fTl
kF= -1 ff&(y&)-kdy&, and kU= _l ly&lw*ff&(y&)-kdy&, provided y&:= y/yAZ, (21)
where kF for the shipment integral as defined in eqn (21),

kU the objection integral as defined in eqn (21),
yAZ a screening diversion, in eqn (4),
W the exponent in eqns (2) and (4),
ff&(y&) the renormalized population density,
uu&(y&) the renormalized objection function,
y& the renormalized diversion as defined in eqn (21),
ff(y) the population density in eqn (4),
uu(y) the function in eqns (2) and (4), and
Y the normalized diversion in eqn (3): and consequently,
(NB*88)/(Nr"kCX:) because NB/Nr = uu.TIT/ff.TIT = yAZW*k, (23)
where k kU/kF as defined, (24)
and others are the same as with eqns (20) to (22) and
with expn (8) of the quantities such as 88 and ex:
If the screening tolerance yAZ in eqn (23) is equal to the tolerance yC in eqn (10), the
wth raised power yAZW in eqn (23) is equal to the ratio ex:/88. Then, the coefficient k is
equal to the fraction of the objection expense NB*88 in the sales NI'*<l: of the population.
Taguchi rules 211
(al (a)
Figure 2 Approximate contours of the coefficient k - cf. eqn (24) and below, with the
screening tolerance yAZ in eqn (21) as taken equal to yC in eqn (10): _a for a rectangular
population in eqn (25), and _b for a normal population in eqn (26). The contours are more
accurate near the origin. For the details, see Shibayama (1995a) p.239, eqn (39) and around.
And (a) w=2, k=l; (b) w=2, k=O.01 ; (c)w=20, k=l ; (d) w=20, k=O.01 ; respectively.
A typical population may be rectangular or normal such as renormalized as follows,
ff&(y&)=l/(zV3s&) for m&-V3s& s y& s m&+V3s&, and = 0 otherwise, or (25)
ff&(y&)=(VzvJts&)-lexp( -(y&-m&)2/2s&2) , respectively, provided that (26)
m& = m/yAZ and s& = s/yAZ where (27)
yAZ for screening diversion in eqn (22) or (23),

ff&(y&) the renormalized population in eqn (22),
y& the renormalized diversion in eqn (21),
m& the renormalized mean in eqn (27),
s& the renormalized standard deviation in eqn (27),
ff(y) the population density in eqn (4),
y the normalized diversion in eqn (3),
m the mean of the population ff(y), and
s the standard deviation of the population ff (y) ,
and the population ff(y) supposed as a rectangular or a normal.
Then, the fraction k in eqn (24) is a function of the mean m& and the standard deviation s&
both as defined above. The integrals kU and kF in eqn (21), and the integrated fractions in
eqn (21), too. The integrations are simplified if the values m& and s& are near the origin,
so that the integration bounds may be replaced by infinities. The fraction k is traced by
the approximation in Figure 2_a or _b, as the equifraction contours with the mean m& as
taken on the abscissa and the standard deviation s& on the ordinate.
The coefficient k is equal to the expense/sales fraction for a screened shipment by the
tolerance yC (or Y¥.crZ) - eqn (24) and below: yAZ=yC. If the fraction to be 4 % (,say)
or less, and if the exponent w =2, both the mean m& and the standard deviation s& in
eqn (27) shall be 0.2 or less - Figure 2. The mean m and the standard deviation s before
the renormalization shall be 0.2 times the diversion yC or less. The situation is relaxed, if
the exponent w =20. The exact value k is calculable semianalytically, and is equal to c. 30 %
for a rectangular population such as the half width as equal to the tolerance yC whereas
c. 10 % for a normal population such as three times the standard deviation as equal to yC,
provided that w=2. If w=20, c. 0.5 % or 0.2 %, respectively. -Shibayama (1995a,b, 1994).
Taguchi quadratic loss function shall be modified to a higher order one, obviously. The
order shall not be fixed a priori, but found a posteriori - empirically for case to case. Though
the critical tolerance is defined of such a function, it is not enough to guarantee the quality of
part. More centered and more narrow population should be prefered. Furthermore, even the
cost BB or the price cc may vary with the diversion or the population. The part quality,
the expense and the dole, shall be traced continuantly. No standardization is required.
3 OTHER TAGUCHI RULES

.1IS DRAFr Plastics Characteristic Values. - Infinitely targeted values
Part dimensions are targeted finitely, i.e. nominal the best (Taguchi), whereas some values
such as strength, dielectric loss, or others are infinitely targeted, i.e. larger or smaller the
better (Taguchi). If targeted upward to 00, it may be characterized by an objection fraction
function such as =1 for ys1 whereas =yW for y<!:1 with a negative exponent w,
provided that the normalized value y is ratio of the part value to the limit value such as all
the parts incur customer's objection. If a part value is targeted downward to 0, the function
is defined in the sense as reversed with a positive exponent w . Any example of precision
is hardly available, yet. Other matters are all the same as with a finitely targeted value .
• 1IS K 0971-1989 Micro-Volumetric Apparatus. - Orthogonal design
Seemingly a purpose of this standard is determination of 'error variance' under fluctuations of

environmental factors, as explained in Article 5.2.2 of Informative Reference 1 of the standard.
Orthogonal design experiment is recommended strongly, however, for such the purpose, an
ordinary stepwise procedure is legible and adaptable for situations more -Shibayama (1995c).
The recommendation is confusing and meaningless. Anyway, the numerical procedures shall
be written more concisely. All physical rules necessary and more important are nicely defined
in 1IS K 0970-1989 Piston Operated Micro-Volumetric Apparatus, and so wisely .
• 1IS Z 9090-1991 Calibration System. - Repeated adjustment
Beside principles popular in metrology, this standard tells, in Article 4, Procedure 7, that
calibration period and calibration precision shall be optimized as detailed in the Annex 1.
The underlying rule is simple, supposing that the diversion of a part value as produced of
or as measured by a system varies monotonically with the cycles of operation, as follows.
Taguchi rules 213
or equivalently, where (28)
DJ for diversion of a part value,

uJ number of the parts as processed,
T drift constant, or equivalently
v drift number := (MT)2 , with 11 as defined of eqn (29) below.
The system is readjusted so the diversion DJ to be equal to 0 when this attains to and above
a prescribed limit D, once per u cycles, so the diversions DJ ranging between 0 and D .
The money L lost per a part is equal to sum of the measurement cost B/n , the adjustment
cost C/u and the objection loss (E+n*F/u)(D/l1)W.A, all per a part so that
L = B/n + C/u + (E+n*F/u)(D/I1)W.A , all per a part, where (29)
L for the loss of money per a part, caused of

B/n Cost of one measurement B averaged per a part,
measured once per n parts,
C/u Cost of adjustment C of the system averaged per a part,
adjusted once per u parts;
E(D/I1)W.A .- Average objection loss per a part, as caused
of u parts between neighboring two adjustments; and
.- Average objection loss per a part, as caused
of n*F parts near about an adjustment;
D .- the adjustment diversion,
11 .- the full objection diversion,
W .- the exponent of a power type objection fraction function,
A .- the money loss, per an objection,
n*F .- the number of the cycles until an adjustment realized,
after the diversion DJ having attained to and above
the diversion D; as detected by one measurement per n parts;
F .- a constant around a half and up to unity, and
E .- 2/(w+2) , or simply l/(w+l) in Taguchi, as below.
The diversions DJ range from 0 to D as with the number uJ from 0 to u according to

eqn (28), the objection fraction (DJ/l1)w as averaged over the number uJ gives the fraction
E(D/l1)W where E=2/(w+2). In originals of Taguchi, that is estimated more approximately.
In spite of eqn (28), the diversions DJ are supposed to range linearly with the number uJ,
giving the value E as equal to l/(w+l), done as enough or on intention of taking finely up
a complicated reality by muffling all the reality in a delicate balance of many the facets.
Equation (29) as differentiated gives the period n that minimizing the loss L as follows,
n=(B/FA)1/2(T/I1)(D/I1)-(w+2)/2 , whereas, (30)
if differentiated with respect to the diversion D , it gives the adjstment diversion D that
minimizes the loss L . Dependency of the adjustment period u on the adjustment diversion D
is taken into account by eqn (28). Supplemented by eqn (30), an equation follows such that
wFA(D/MW+2+(w-2)(FAB)1/2(T/M(D/M(W+2)/2_2C(T/M2=O . (31)
It looks complicated but is essentially a quadratic equation as to be solved easily.

The equations (30) and (31) may be nonnalized in tenns of a number y. and a diversion fj, ,
say, either derived of v or 11 as being modified by a power of AjC , respectively. The
solutions may be plotted in simple diagrams to cover various situations - Shibayama (1995c).
If every part measured ( n=l), the measurement cost B/n is large ( =B ) whereas the objection
loss (n*F/u)(D/I1)W.A is small and =(F/u)(D/Mw.A. Ifapart once per n parts, the
measurement cost is less (=B/n ) whereas the objection loss more and =n*(F/u)(D/I1)W.A .
The period n is optimized to minimize the loss L . The adjustment diversion D , too. As the
diversion (and the period u ) larger, the objection loss more whereas the cost C/u less.
A good picture, well without any standardization. In application, again, trials and errors.
• The other rules miscellaneouslly incorporated. - Associated rules
• Least Square Fitting of Equations, Evaluation of Residues, Analysis of Variance,

• Signal-to-Noise Ratios, Conversion of Variables,
• Algebraic Estimation on Abridged Experiment, Parameter Design, and Others.
Not closely related with tolerancing, the rules are incorporated in illustrative examples in
the standards. The mathematics may be edited as fonnulas, diagrams or procedures legibly,
as to be employed by people without any assistance - Shibayama (1995c,1989,1988).
4 CONCLUSION
Taguchi tolerancing, coming of intuition as other Taguchi ideas, affords a good picture, though
sometimes lacking precision. The quadratic loss function is the lowest even order term of
a series, but this does not apply for a function y4 (or y20 ) or so! - Shibayama (1995b,1994).
The exponent as presumed equal to 2 gives a simple picture but the precision shall be tested
a posteriori, if necessary. The tolerancing coming of the balance of losses before and after
the shipment, and trying to cover all the losses as possible, but the meaning shall be analyzed
if necessary -Shibayama (1987). Some more reasons all suggest Taguchi tolerance design.
Microlocality shall be covered by docile flexibility as to modify the way to meet with every
the concrete fact by trials and errors, if otherwise a confusion -Shibayama (1995b,1994).
As developed of the problems-solving activity of the great practitioner who being busy in
actual problems, the rules and the way shall be salvaged in agendas and discourses, many in
Japanese, but even for a Japanese profession not easy to read or to refer to. But the profound
thought and the fine mathematics are combined in a vivid and matured intuition, as oriented to
actual problems and around, mathematically, technologically, engineeringly and economically.
The way works well without any standardization, providing very general pictures.
The thought and the mathematics shall be reviewed carefully, accurately and precisely,
and reinterpreted and rearranged, as to be restated legibly and concisely. People might be afraid
of impairing nice unification and amalgamation of ideas in the original, but it would clarify the
foundations of the method behind the highly implicated assertions.
Taguchi rules 215
5 REFERENCES
e Japanese Industrial Standards (available in English) or Draft (only in Japanese).
JIS K 0970-1989 Piston Operated Micro-Volumetric Apparatus.

JIS K 0971-1989 Calibration Methods for Micro-Volumetric Apparatus (Inc!. Infve Refs 1-3).
JIS K 7109-1986 General Tolerancing Rules for Plastics Dimensions (Inc!. Appendixes 1 and 2).
JIS Draft (since c.1987) General Tolerancing Rules for Characteristic Values of Plastics.
JIS Z 9090-1991 Measurement - General Rules for Calibration System (Inc!. Annexes 1 to 4).
eTaguchi agendas and discourses (Japanese).
Standardizat'n and Quality Control (Hyoujunka To HinshitsuKanri ) (1986), 52 (5), Taguchi in USA.
The Plastics (GouseLlushi) (1985),31(12), Dimensional Tolerancing and Improvement.
Instrum 'n Control (KeiryouKanri ) (1986),35 (10), New Method of Tolerancing and Application.
e Expository papers (Japanese).
Kamoshita,T. and Matsuda,J. (1986), Proposal of Determining Dimensional Tolerance of Plastics

Parts Using Loss Function. J. Japan Soc. Precision Engng 52 (8),1321-1325.
Taguchi,G. (1985-1986), Design of Experiments for Quality Engineering.J. Japan Soc.Precision
Engng 51(4),762-767;-; 52(3),463-468.
e Presentation papers, abstracts or manuscripts.
Shibayama,T. (1995a), Taguchi Rules in Some Japanese Standardizations of Tolerancing.

Proceedings 4th CIRP Seminar Computer Aided Tolerancing, p.229-244. (And the handout).
Shibayama,T. (1995b), Meaning of Tolerance as Defined by JIS K 7109-1986. Japan Soc.
Precision Engng, 1995 Spring Ann. Con!, J80. Papers p.1061-1062. (And English handout).
Shibayama,T. (1995c), Taguchi Rules Formulated in Engineering Physical Mathematics.Japan
Soc. Precision Engng 1995 SpringAnn.Con[, C46. Papers p.411-412. (And English handout).
Shibayama,T. (1994), Tolerance Design in Terms of Objection Fraction Function. Japan Soc.
Quality Control 24th Ann. Conf., 2-3. Abst. p.33-36. (And English handout).
Shibayama,T. (1989), Quality Control Engineering in Thermoplastics Injection Molding. The
Polymer Processing Soc. Internat?, PPS-5(Kyoto), IP-lO, Abst.p.269. (And English handout).
Shibayama,T. (1988), Application of Quality Control Engineering for Better Injection Molding
Operation. Polymer Preprints ,Japan (English edn) 37(7), E 544, 27Plb35.
Shibayama,T. (1987), Effect of Order of Loss Function for Tolerancing Plastic Parts. Japan
Soc. Precision Engng 1987 Autumn Ann. Con!, G61. Papers p.763-764. (Japanese).
6 BIOGRAPHY
Tadao SHIBAYAMA. Graduated 1953 from Dept of Physics, Faculty of Sci., Nagoya Univ.,
employed in Nagoya Municipal Industrial Research Institute as an engineer being involved
in thermoplastics injection molding and others, and retired of the age limit 1989.
PART FIVE
Tolerance System
15
Tolerance Optimization Using Genetic
Algorithms:
Benchmarking with Manual Analysis
M. P. Iannuzzi
Variation Systems Analysis, Inc. (VSA)
300 Maple Park Blvd.
St. Clair Shores, MI48081 USA
Phone: 810.774.2640 Fax: 810.778.6470
Internet: miannuzzi@vsa.com
E. Sandgren
TRW, Sterling Heights, MI USA
Abstract
A computational design methodology is presented which permits the optimal allocation of tol-
erances for mechanical assemblies. A nontraditional genetic optimization method is coupled to
a Monte Carlo based tolerance analysis. The objective is to determine the maximum tolerance
zone value for each nominal feature while simultaneously meeting all critical dimensional and
functional constraints imposed upon the design. A discretization of possible tolerance zone
values is performed and a global search conducted using a genetic algorithm. Prior work dem-
onstrated the feasibility of the approach. This paper seeks to examine the performance of the
algorithm on industrial level problems in comparison to the leading alternative technique
- iterative manual analysis. Three industrial problems are considered and the best solution gen-
erated by a manual method for each problem is compared to the solution generated by the opti-
mizer. The results demonstrate the current capability of the approach and provide direction for
future development. The goal is to develop a comprehensive software tool which can support
dimensional variation decisions from concept through design, manufacturing, and inspection.
This work was supported by a SBIR grant from the National Science Foundation.
Keywords
Genetic optimization, tolerance analysis, tolerance synthesis, simultaneous engineering
220 Part Five Tolerance system
INTRODUCTION
Simultaneous engineering is a critically important concept for any product oriented organiza-
tion seeking to enhance its competitive position in today's global marketplace. The successful
implementation of simultaneous engineering requires a direct linking of design, manufacturing
and inspection. Many CAD companies are striving to embrace this concept by transforming
their systems to support a comprehensive product database. To realize the potential of simulta-
neous engineering it is necessary to locate key communication channels which link the product
development process and improve the quality of the information flowing through these chan-
nels. Certainly dimension and tolerance allocation form a vital common ground in the product
development process. Yet few tools are available which support the decision making process
between the diverse organizational elements.
Tolerancing is generally practiced as an art rather than a science. Tolerance values are often
based on historical practice rather than as a well thought out and integrated part of the product
development process. The issue of high quality is often associated with a reduction in tolerance
zones which emphasizes a "less is better" philosophy. This is often an expensive way to
achieve quality objectives. A means of specifying "optimal" tolerance levels can have a tre-
mendous positive impact on the process as well as fostering a close link between the design,
manufacturing and inspection functions.
The traditional approach to optimal tolerancing is to formulate the problem as a linear or
nonlinear programming problem. The design variables become the component or assembly tol-
erance range values and the objective function to be minimized is directly related to manufac-
turing or production cost [Chase, 1986, Spotts, 1973, Speckhart,1972]. Solution algorithms
include gradient based nonlinear programming [Michael, 1981], an integer branch and bound
method [Lee and Woo, 1987], simulated annealing [Zhang, 1992] and other discrete optimiza-
tion techniques [Loosli, 1987]. The coupling of a genetic optimization algorithm with a Monte
Carlo simulation approach was applied to several small scale problems by Lee [Lee, 1993].
In all previously reported work, the scale of the problem considered was very small. How-
ever, industrial applications may have several hundred or thousand dimensions and tolerance
callouts. Complex geometric conditions such as out of roundness, true position of surfaces and
three dimensional skewing effects must be considered. Assemblies of components have to be
dealt with as well as production and assembly fixturing. Functional and quality measurements
must be specified and evaluated. In essence, the optimal tolerance allocation problem is
beyond the scope of research conducted to date. The problem to be solved is extremely large,
nonlinear, multi-modal and often discontinuous. This makes the solution to any real world
problem by conventional means unlikely.
A genetic optimization algorithm was coupled to a sophisticated Monte Carlo based simula-
tion and modeling environment and a wide assortment of problems were formulated and
solved [Iannuzzi and Sandgren, 1994]. The performance was tested in relation to both the fine
and coarse grain characteristics. The fine grain performance was evaluated on simple models
where the optimal distribution was known beforehand. The combined algorithm handled all of
these problems and even generated solutions for several cases which exceeded published
results. The coarse grain performance refers to the common operating environment for an opti-
mizer wherein the optimal distribution is not known and the focus is on the ability of an algo-
rithm to satisfy the constraints. A set of five validation models including two commercial
Tolerance optimization using genetic algorithms 221
applications were examined and the algorithm converged to reasonable solutions in each case.
However, the character or quality of the solution was not known as compared to alternative
methods, primarily an iterative solution generated .by an experienced user.
A characteristic of the manual, iterative optimization process is that a local view of the
design space is taken. This is somewhat analogous to a gradient based nonlinear programming
algorithm operating in a multi-modal solution space. A local downhill search is conducted
about the initial design point by increasing tolerance range values when measurements are
over-achieved and by decreasing tolerance range values when measurements fall short of spec-
ification. Little if any consideration is given to other, perhaps far more productive design
regions.
In contrast, the genetic optimizer takes a global view and tends to adjust all tolerance range
values. This would seem to be an advantage. Considering the complex trade-offs which must
be made and potenti'al interaction among the design variables, it is unlikely that even a knowl-
edgeable user can approach an optimal solution. On the other hand, the manual process can
apply experienced based knowledge which is difficult to build into an automated optimizer.
Given that the true global solution for a industrial application problem is not known exactly,
the manual approach provides the only available estimate of the solution with which to bench-
mark the performance of the optimizer.
In order to define the capabilities of the optimizer versus a solution generated manually, a
benchmarking study was undertaken to better understand the performance of the optimization
algorithm and to locate areas which require further development. Three problems taken
directly from industrial applications ranging from 100 to 500 tolerance range values are con-
sidered herein. Solutions are generated both manually and using the genetic algorithm. A
description of each problem and a detailed comparison of the solutions are presented, but first
a brief review of the tolerance analysis and optimization algorithm is presented.
IDLERANCE ANALYSIS
The tolerance modeling and analysis procedure implemented in this research was the Varia-
tion Simulation Analysis (VSA) software. VSA is a Monte Carlo based simulation program
capable of analyzing three dimensional mechanical and nonmechanical systems. It compre-
hends component variation in accordance with the ANSI and ISO GD&T standards as well as
handling assembly variation and procedural logic. The modeling environment supported by
VSA operates on a geometric feature set of three-dimensional modeling points. The model is
generated so that it specifies how the modeling points are to be constructed and can handle the
assembly of multiple parts, including fixturing. Toleranced parameters are imbedded in the
construction and assembly procedure and the output defines the influence of each of these
parameters on user defined measurements. These characteristics enable VSA to solve a wide
spectrum of "real world" problems reliably and accurately [VSA, Inc., 1993].
The VSA simulation process operates as follows:
I. A geometrically varied instance of each component is created by perturbing each feature in

its inherent degrees of freedom within the zone described by the tolerance callout. The spe-
cific degrees of freedom and magnitude of deviation are determined by mapping a random
number to an associated probability distribution. These distributions represent design intent
or manufacturing capability.
2. The component instances are assembled according to specific assembly logic. Assembly
variation is introduced based on a probability distribution mapping to assembly characteris-
tics.
3. For each assembled instance, measurements are computed for critical characteristics con-
cerning quality and functional requirements. A rich array of measurements are possible
including clearance, kinematic motion and indirect geometric quantities such as load.
4. The simulation process is repeated and measurement data is statistically evaluated until the
desired statistical confidence is achieved. The data is fit to a statistical distribution to
enhance "tail" information.
5. An imbedded designed experiment capability is performed to identify and rank key contrib-
utors to the variation of each functional requirement. These indicators provide valuable
insight for design change scenarios.
The modeling and analysis capability forms a vital element of the tolerance synthesis pro-
cess. The results from this analysis may be used by an optimizer or alternatively by a skilled
user in an iterative fashion.
GENETIC OPTIMIZATION
Genetic optimization methods emulate the natural selection process of nature and operate on a
principle of survival of the fittest. The goal of a genetic algorithm is to discover the basic build-
ing blocks of a good design and to combine them in such a way as to create the best possible
design. The approach relies on the randomness present in natural selection, but quickly
exploits information gathered in order to produce an efficient design procedure. A design to a
genetic algorithm is an abstract representation which is directly analogous to a chromosome in
a living organism. The chromosome is composed of a number of genes, each of which may
assume one of a number of possible values or states. In the specific example of tolerance allo-
cation, each gene value represents a choice of tolerance range value for a specific geometric
feature. The genetic algorithm operates by manipulating the coding of the set of gene values or
tolerance range values.
Moreover, a genetic algorithm operates on a population of designs rather than a single
design point and the rules which govern the transition from one set of designs to the next are
probabilistic rather than deterministic. The overall suitability of a chromosome, that is the per-
formance of a specific design, is termed its fitness. The fitness determines the probability of the
chromosome becoming a parent for the next generation of designs. The chromosomes possess-
ing the greatest fitness have the highest chance of being selected as parents. The selected par-
ents are combined using various genetic operators to produce new designs for the next
generation. The offspring contain genetic information from both parents. The process is con-
tinually repeated with the expectation that the fitness of the individual designs as well as the
average fitness of the population will gradually improve until the optimal design(s) are pro-
duced. A flowchart depicting the major operations taking place within a basic genetic algo-
rithm is pictured in Figure I. Additional detail on genetic algorithms is available from several
sources [Goldberg, 1989 and Davis, 1991].
Figure 1 Genetic Algorithm Flow
BENCHMARKING
The purpose of this benchmarking study was to substantiate the viability of combining
genetic optimization and Monte Carlo tolerance analysis. A previous study did prove the feasi-
bility of the approach, however the goal sought is a commercially robust application. Accord-
ingly, the true test for the tool is its being effective and practical on "real world" problems. To
do so it is necessary to understand and reveal the performance, quality and character of the
solution of this algorithm in comparison to the leading alternative techniques. At this time, the
only alternative technique capable of solving commercial grade problems is optimization
through iterative manual analysis. Since the true or global optima is unknown for complex
problems, the manual analysis becomes the gauge point.
The problem statement for the optimizer, in this study, is to distribute the tolerance range
value for each nominal feature to maximize the fitness function while satisfying all measure-
ments to the required level. The fitness is a cost related function which penalizes tight toler-
ance ranges as opposed to loose tolerance ranges. In addition, a penalty term is added to the
fitness function to account for violation in any functional requirement. The specific fitness
form selected may be expressed as:
N, a. N,
Fitness = - L-'-b+RL
(x)'
i=1
[g. (x)].
j=1 ]
(I)
Where:
Xi = ith tolerance range value (normal distributions)
ai' b i = cost coefficient and exponent (ai=l, b i=2 for this study)
Nt =number of tolerance values
R = penalty factor
g/x) = the jth measurement yield constraint ~ 0
Ne = number of constraints
The fitness could be an exact cost function which would provide scaling to the various lev-
els of importance of the tolerances and constraints. The challenge here is that considerable
effort is required to develop cost data which relates directly to every tolerance range for every
possible tolerance callout. The function given in Equation (1) is more general and requires no
specific cost information. As such, Equation (I) was deemed satisfactory for the purposes of
this study. A drawback is that the function can be dominated by one or more tolerances when
they assume very small values or a disparity in the units of the parameters. Although the form
used for this benchmarking was not an exact cost function, the quantification of the reciprocal
of the tolerance squared will be referred to as the cost of the design for comparison purposes.
It is noted that this benchmarking study was a casual experiment. That is, we asked the
expert user to "do what you would normally do" in manually achieving the performance
requirements of the design. No emphasis or incentive was placed on the user to "beat" the opti-
mizer. However, in "working" the optimization algorithm there was the motivation to improve
upon the manual results. That is, the optimization program has a number of parameter settings
which control the behavior of the algorithm as follows:
• NSIMS: Monte Carlo sample size;

• NPOP: population size;
• NGEN: number of generations;
• R: penalty factor;
• PFACT: penalty factor multiplier;
• NUP: number of generations between penalty updates;
• PMUT: probability of mutation;
• NSAVE: Number of best members to pass from generation to generation.
Various runs of the optimizer where performed on each test problem to identify and under-
stand the effect of the parameter settings, until satisfactory results were achieved. The values
used for the results are listed in Table 1. Finally, the solution time values given for each prob-
lem are provided to give afeel for the amount of time spent and are approximate values. The
manual timing numbers are wall time and the optimizer numbers are CPU time.
Table 1 Algorithm Parameter Settings

Problem NSIMS NPOP NGEN R PFACT NUP PMUT NSAVE
Doors 100 400 250 100000 2.0 5 0.03 5
Valve 500 200 105 200000 2.0 5 0.03 5
Retractor 100 425 200 100 1.5 5 O.oI 5
The Manual Approach
The manual optimization process has the following steps:
1. Gather Information
Conducting a sophisticated tolerance analysis involves interfacing with many elements within the
organization to understand all sources of variation and its impact. Discussions are held with
design engineering to understand functional requirements and obtain CAD models or blue prints.
Manufacturing engineers are interviewed to comprehend the manufacturing process and capabil-
ity. Finally, inspection plans are reviewed with quality control engineers. This body of interaction
truly reflects the simultaneous or cross-functional characteristic of dimensional analysis.
2. Model Construction
Constructing a correct model is the cornerstone of the process. Without good information,
without a correct model - the results would be of dubious value. VSA models can be pro-
grammed by hand or generated automatically from various CAD systems. In all cases the
model is processed as described in the previous section titled Tolerance Analysis.
3. Nominal Build
A key verification of the model is performed by setting all sources of variation to zero and
examining all dimensional values. If the values differ from that of the as designed condition
corrective action is taken to adjust the model or determine if the design is in error. Many prob-
lems are identified and resolved at this point.
4. Parameter Optimization
With the nominal condition verified, the tolerances are activated and a simulation of the assem-
bly is performed. A wealth of information is available to the user after the completion of a sim-
ulation run in the VSA software. A first report considered is the process report. It provides a
look at several characteristics of the simulation including the output distribution for each
imposed measurement, an estimate of the number of components or assemblies which are non-
conforming and estimates of indices such as Cp and Cpk.
Based upon the results and if the performance of the design requires adjustment, many non-
tolerance design changes are possible. These changes include: nominal shifts in dimensions,
assembly locators, fixturing, gauging, assembly sequence, etc. Adjustments to any of these are
possible to improve the robustness, quality and cost of a design and are the first objective in the
process.
5. Tolerance Optimization
If the design is not in conformance by this point, reallocation of tolerances is considered.
Employing the main-effects design of experiment (HLM) report, the engineer identifies toler-
ances for adjustment. This report may be used to see which tolerances are contributing most
significantly to the variation of a single or group of measurements. By selecting the currently
violated measurements as the targeted group, a list of the key contributing tolerances can be
generated. These tolerance range values are then reduced and the simulation is re-executed.
Repeating this process, a design can generally be produced which meets the measurement
specifications. When the manual approach is successful, usually a limited number of tolerance
values are reduced.
Note, at this time, all of the steps outlined above are the same for the automated optimiza-
tion process except step number five where the optimizer is employed. Future work in auto-
mated optimal variation reduction is targeted at step four; parameter optimization. The
following benchmarks and conclusions reveal the significance of parameter optimization.
The Automotive Door System Problem
Many design objectives must be considered simultaneously for the door assembly of a sedan
style vehicle. There are gap and flushness objectives from door-to-door, door-to-quarter panel
and door-to-roof. In addition, there are seal margins between the doors and body. A typical
CAPIO
GAPI2
CAPI6
GAPIS
Problem Attributes
Components: 20
Tolerances: 414
Functional Measurements: 42
Parameter Optimization:
Seal Margin
Adjustment to
Fixture Locators
Figure 2 Automotive Door System Problem.
sedan style door configuration is shown in Figure 3. This model contains a door installation
fixture for both front and rear doors. Each fixture has a set of three locators which position the
door in the cross car direction. Gage pins in the fixture are used to set the door in the fore/aft
and up/down directions. As flushness was the most demanding specification, the manufactur-
ing plan is to first enhance flushness, then gap and seal margin. Output definitions are devel-
oped for gap, flushness and primary seal margin around the periphery of the doors.
Examination of the initial manufacturing plan simulation by the experienced user revealed
larger amounts of variation in the door header areas. It was determined that this was due to all
cross-car gaging and installation fixture locators occurring at or below the up/down midpoint
of the vehicle. Thus, the initial manual iteration (parameter optimization) revised the manufac-
turing plan by enlarging the controlling planes of the door-to-fixture and fixture-to-body inter-
face. This strategy did reduce the cross-car variation in the header specification.
As mentioned, modifying assembly locators and assembly sequence is not within the cur-
rent scope of the optimization software. Accordingly, the starting point of the head-to-head
comparison began after all pertinent modifications to these non-tolerance parameters were
made.
A summary of results for the initial condition as well as the result from the manual and opti-
mization solutions are listed in Table 2. Several interesting points can be made from the infor-
mation contained in this table. At the starting point, virtually every measurement constraint is
violated, some as much as 30 percent (Note: a constraint is considered violated for this prob-
lem if its estimated yield falls below 95.0% or in other words 5.00% out of specification.)
After twenty manual iterations, employing percent contribution as the guide, the results
from the manual approach was successful in achieving the yield requirement with little nega-
tive impact to the objective function. (The term objective function is used to describe the quan-
tity fix»~. In this context, the goal is to minimize the objective function. Note, in equation (I)
fix) is negated to satisfy the maximization convention of the fitness function.
Table 2 Automotive Door System Problem - Analysis Summary (500 Samples)
M I0 . . . Genetic
Item Initial Condition anua ptlmlzatlOn 0 . . .
ptlmlzatlOn
fix) =L 1/x2 40735 40775 27305
Min Tolerance (mm) 0.03 0.03 0.02
Max Tolerance (mm) 2.00 2.00 4.08
Average Change (%) -9.70 -10.0
Max out of spec. (%) 30.12 4.91 4.93

(Target:::; 5.0)
Average out of spec. (%) 12.31 1.52 0.87
Min out of spec. (%) 0.11 0.05 0.00

(Target:::; 5.0)
Solution Time (Hrs) 8 24
In contrast, the optimizer also successfully achieved the yield requirements and did so while
simultaneously decreasing the cost of the design from the starting condition! This is a dramatic
difference from the traditional manual paradigm. In the traditional paradigm bringing a design
into specification is associated with tolerance tightening which can only increase the fabrica-
tion cost of a design. The optimizer, since it is operating on all tolerances simultaneously, was
able to find a solution to the tolerance allocation which decreased the cost in this example3.
Legend
.
• • Genetic Optimizer
• Manual Solution
•
•
• •
• • • ••
. ....• . • ....
•
• •
-
•
• ... • • •• I'
-
• 1' •
III •• "II, • • _;---.-
.- ,.
:::~ ~
••
--.....- ••••- _.."'. ..
••
.!... -Fl'.
• ......
... ... . . . . :I'
.• . . . . . • .
=. -, ·
rI' .
.. . -
• , .... • • ...:..' •• :
A7~""-'..
... ••••
rI".... - •• •• ... ....
••
rt:' III ~
__
.,~~~--~~~~-f~~~~~~_'~
,~~~T-~-r~--~~~~~--~~-.~~
20 40 60 80 '00 120 140 160 '80 200 220 240 260
ToLERANCE
Figure 3 Automotive Door System Problem - Fraction of Initial Value.
An indicator of the character of the solution is illustrated in Figure 3. In this figure the per-
cent change of the tolerance value from its initial condition is plotted. A value of zero means
no change, negative means the tolerance value was reduced and accordingly, a positive value is
an increase from the initial condition. This graph depicts two feasible designs, but with com-
pletely different character. The manual solution adjusts a limited number of tolerances and
when doing so decreasing the value only. Whereas the optimizer produces a "shotgun" pattern
of change.
Seat Belt Retractor Problem
This example models a seat belt retractor mechanism which is pictured in Figure 4. The objec-
tive of this study was to insure that the locking gears will engage securely and without interfer-
ence when activated due to sudden stops. The model had 151 tolerances and two functional
requirement outputs were specified. The measurements consisted of a distance between the
pilot lever tooth to the ratchet wheel outer diameter and the distance from the assembly lever
tip to the ratchet wheel surface.
For the initial run of the optimizer, a very interesting effect occurred. One of the measure-
ments had a significant offset in the distribution (Cpk was much smaller than Cp). The opti-
mizer attempted to bring the design into specification by altering the shape of the distribution
curve to that of a nonnormal curve. It was evident that this was not the best way to address the
nonconformance. Upon examining the results, the expert user identified a nominal shift in the
size of one of the dimensions. Rerunning the model showed that the nominal dimension change
did indeed shift the output distribution and balanced the Cpk and Cp values. Once again, a non-
tolerance parameter adjustment was made to improve the variation characteristics. Accord-
ingly, the head-to-head tolerance allocation benchmark was made after the nominal design
change.
Both the manual method and the optimizer located valid final designs. A constraint is consid-
ered violated for this problem if its estimated yield falls below 99.73% or in other words 0.27%
out of specification. Table 3 lists the results of this study. With only two measurements and a
continuous design space, the manual approach was able to quickly produce a feasible design.
Again, the optimizer produced a superior design with an objective function value one-third
that of the manual solution. Unlike the automotive doors problem though, the optimized solu-
tion was at a greater cost than the initial condition, but less then the manual solution. Figure 5
Problem Attributes
Components: II
• Tolerances: lSI
Functional Measurements: 2
• Parameter Optimization:
Nominal Shift
Pilot Lever---1~!::;f!:~L~=:",,-_ _ _~~~9~:22::r.~~_R_

to Rachet
...
Wheel
Distance
Figure 4 Seat Belt Retractor Problem.

Table 3 Seat Belt Retractor Problem - Analysis Summary (500 Samples)
Item Initial Condition Manual Optimization Genetic Optimization
f(x) = L 1/x2 18596 75440 26823
Min Tolerance (mm) 0.025 0.005 0.02
Max Tolerance (mm) 0.50 0.50 0.375
Average Change (%) -13.89 3.10
Meas. I - % out of spec. 18.72 0.24 0.26

(Target :s; 0.27)
Meas. 2 - % out of spec. 4.29 0.27 0.27
(Target :s; 0.27)
Solution Time (Hrs) 2 8
depicts the percent change in tolerance value from the initial condition and illustrates the a true
tolerance allocation process occurring for the optimizer.
Valve Tappet Problem
The final example concerns the clearance between the camshaft and the tappet in an overhead
camshaft engine. The model includes primary dimensions and tolerance values for components
including the valve, valve stem seal, valve retainer cap, valve retainer lock, valve tappet, tappet
shim and camshaft. The gap is measured for both the initial assembly and for the burned-in
condition. A series of shims (35) are available during assembly to bring the gap within specifi-
2
Legend
• Genetic Optimizer
• • Manual Solution
• • • • • • • • ••
• •• • • • • • •
• •
•
•• • •
..
• • ••
••
•
•
.. .. ... .. ...
_ A . . _•_ _•. _ • •• • • •• •
•rI' • • • •• • •• • •
.
.. ...... • • •• •
••
• •• •
•
••• • • • • • •• •
., •
......
••
• • •
o 10 20 30 40 50 60 70 80 90 100 110
ToLERANCE
Figure 5 Seat Belt Retractor Problem.
Problem Attributes Shim Use Output

• Components: 8
• Tolerances: 82
• Functional Measurements: 2
• Parameter Optimization:
Nominal Shift - Cpk « Cp
Figure 6 Valve Tappet Problem.
cation. The model includes approximately 80 tolerances and the two measurements previously
mentioned. A graphic view of the assembly is provided in Figure 6.
Table 4 summarizes values for the initial design, the manual solution and the optimized
solution. At the starting point, both constraints were violated. This study also included a nomi-
nal shift in one of the dimensions to center one of the outputs. Head-to-head comparison began
from this point.
Using the percent contribution analysis to guide the manual process, after 15 iterations no
further tolerance tightening improvements could be found. The manual approach was not able
to bring both constraints into specification. The likely reason is attributed to a discontinuity in
the model due to the shim selection process. This discontinuity confounded the gradient-like
behavior of the percent contribution analysis.
On the other hand, the optimizer was able to produce a design with both measurement con-
straints within specification. The ability of the optimizer to address this problem is due to its
global view of the design space and nongradient based transition rules, identifying a com-
Thble 4 Valve Tappet Problem - Analysis Summary (500 Samples)
Item Initial Condition Manual Optimization Genetic Optimization

1.0e+12 1.0e+12 2.6e+13
Min Tolerance 1.0e-6 1.0e-6 2.0e-7
Max Tolerance 0.50 0.50 0.56
Average % Change -9.20 -42.0
Meas. 1 - % out of spec. 5.69 0.09 0.00

(Target :s; 0.27)
Meas.2 - % out of spec. 10.1 7.32 0.23
(Target :s; 0.27)
Solution Time 3 Hrs 6Hrs
3
• Legend
• Genetic Optimizer
• Manual Solution
•
• •
• ••
.....•....•...... ...... ...-

• •
.• ..-...• •........ .....
••••••••• . . ••ItI.ItI.ItI.ItI.ItI.ItI.ItI.ItI.ItI.ItI.ItI.ItI.1tI.1tI.14I.1tI.1tI.1tI.1tI.14I.1tI.1tI.1tI.14I.I4I.MI.MI.I4I."
. . . . . . . . . . . . ..
•
-1
10 20 30 40 50 60
lbLERANCE
Figure 7 Valve Tappet Problem.
pletely different allocation of tolerance values than present at the starting point. The informa-
tion generated by the manual approach is valid at the initial design, but, as in this case, may
change considerably over the design space.
Although the value of the objective function is less for the manual solution, since a feasible
design (no greater than 0.27% nonconformance) was not produced, the manual solution is
judged to have failed. If a cost basis had been developed for both the tolerances and con-
straints, then it may have been entirely feasible for the best cost to not meet the target specifi-
cation (scrap cost less than manufacturing cost), but this was not the intent of the
benchmarking formulization objectives.
Examining the fraction of initial value plot, Figure 7, the optimizer reduced most of the tol-
erance range values. This may have been necessary to meet the functional constraints. But, it is
easy to see that scaling may have played an important role in this process and illustrates the
weakness of the functional form of the fitness used. In this case the optimizer tightened one of
the shim tolerances, which was an already small number, to an even smaller number. In the
functional form of the fitness used, a disparity in the magnitude of the parameters will "blind"
the optimizer, meaning increases to noncritical tolerances will have a negligible impact. Using
this formulation a fixed percentage change in a variable is treated equally. That is whether one
is dealing with units of length or angles, the effect is the same. This issue would be resolved if
cost were minimized as all units would then be uniform and all scaling issues of tolerance to
cost benefit would be resolved in the formulation. While this is an attractive concept, it is desir-
able during the latter phases of the product development process. What this seems to suggest is
that the selection of a fitness function form can have an important impact on the solution char-
acter. It is also clear that additional research is needed to locate even better forms of a noncost
based fitness function.
~
Vi' -20
0H ~~~~~~~~:~5~~;a_;:::~ 0.02
0
6 -40 -0.02 ~
z
o -60 ~
i= -0.04 ~
z -80 (0-
rr,
;;,
I>:. -0.06 ~
1.;l-100 :;J u
c: 1EEo -0.08
G-120 ~ 0
\>l
;-140 ~ -0.1
o -160 L--_ _ _ _ _ __ _ _ _ _ _ __ _ _ _ _ _----" -0.12
o 20 40 60 80 100 120
GENERATIQ S
I-e-F(X) *P(x) -CON(1) 0 CON(2) I

Figure 8 Valve Tappet Problem Convergence Plot.
The convergence property of the genetic algorithm for this problem is shown in Figure 8
and is typical of the all the problems studied. Recall the fitness function as currently imple-
mented (Equation (1)). The function is to be maximized and that constraints must be greater-
than or equal-to zero. The general trend is for the penalty to become increasingly more impor-
tant which will drive the solution to a feasible region. The reciprocal squared tolerance contri-
bution tends to provide this movement toward feasibility by keeping the smaller values as large
as possible. The specific example shows that the un-penalized fitness function fix) and the
penalized fitness function p(x) remaining fairly stable as the penalty increases (R is updated at
the end of each 5 generations), until about halfway through the process. This reflects a very
desirable characteristic in that improvements to constraint satisfaction are being made without
the expense of carte-blanc tightening of tolerances. As the process continues the final con-
straint satisfaction could only be accomplished by reducing tolerance range values which
accounts for the large decrease (negative increase) in the objective function. The algorithm is
able to make up for some of this effect by altering various range values to increase the critical
ones which is indicated by the gradual increase in fitness function toward the end of the pro-
cess. In fact, Figure 8 seems to indicate that additional progress might be possible if the algo-
rithm were allowed to run longer.
SUMMARY AND CONCLUSIONS
This paper posed the question as to the performance, quality and character of tolerance optimi-
zation using genetic algorithms and Monte Carlo tolerance analysis. Specifically, is the com-
bined algorithm capable of handling real world industrial problems? To answer these
questions, a benchmark with manual analysis was conducted as the only competitive alterna-
tive for recognizing and comparing solutions.
The perfonnance of the algorithm on the test set proves that industrial problems involving
several hundred variables can be solved successfully. In each case examined, the algorithm
moved the design from an infeasible point to an optimum point which was consistently supe-
rior in quality to that of the solution generated by conventional manual practice. Although the
solution time was significant, it is important to note that enhancing the computational perfor-
mance of the algorithm was not an objective at this time. Results examined, such as percent
change in tolerance value, give insight to the character of a globally based optimization solu-
tion. The character of the solution sought is one that exhibits the ability to make complex trade-
offs in searching the design space. The tolerance allocations arrived at by the optimizer demon-
strated this character very well.
The challenge of the manual method is that it is extremely tedious to simultaneously bring
all constraints into specification. As the problem size increases, the complexity to make the
proper trade-offs to achieve conformance is difficult enough let alone a perhaps impossible
task of complete optimal tolerance allocation.
An issue of great interest in this study is what exactly is involved in the tolerance synthesis
process? It is clear from the benchmarking activity that parameter optimization is integral and
required element of tolerance synthesis. The significance and impact of this fact is seldom con-
sidered in other investigations. Yet understanding this is essential in correctly performing toler-
ance optimization. In each of the three problems, nonnegligible effort in parameter
optimization was necessary. For two problems a mean shift was needed and, on another, varia-
tion reduction was achieved through improvements in fixture locators.
There are many areas identified for continued research. The first issue is in extending the
problem domain to include parameter optimization. Nominal dimensions will be included in
the process. A multi-phased approach will allow for dimensions and tolerances to be consid-
ered simultaneously as well as individually. This approach should open up new potentials in
robust design. Next, it is evident that the perfonnance of the algorithm is sensitive to the for-
mulation of the fitness function in its ability to represent the desired character of the design.
Both a scaled noncost and cost based fonnulization will be pursued. In addition, a key factor in
the successful outcome for the optimizer was the proper manipulation of parameter settings
which controlled the operation of the algorithm. For commercial success it will be necessary to
identify automated methods to set these controlling parameters. This research provided valu-
able insight to this goal. Finally, the computational performance of the algorithm has many
enhancement scenarios, including the incorporation of directed intelligence, multi-phased!
hybrid methods and a reordering of the design variables between the optimizer and the model
to facilitate more productive genetic operations.
The tolerance synthesis is broad and complex. Currently the optimizer has been shown to be
an effective tool in its domain, but the user's skill plays a key role in the success of the out-
come. Some of these skills have now been identified and effort will be place on trying to cap-
ture, implement and automate this process in the optimization fonnulation.
REFERENCES
Chase, K.W. and Greenwood, W.H., "Computer Aided Tolerance Selection: CATS User
Guide," ADCATS Report No. 86-2, Brigham Young University, 1986.
Davis, L., Handbook of Genetic Algorithms, Van Nostrand Reinhold, New York, 1991.
Goldberg, D.E., Genetic Algorithms in Search, Optimization and Machine Learning, Addison
Wesley, 1989.
Iannuzzi, M and Sandgren, E, "Optimal Tolerancing: The Link Between Design And Manufac-
turing Productivity" ASME, Design Theory and Methodology, DE-Vol. 68 1994.
Lee, I., "Tolerance Optimization Using Genetic Algorithm and Approximated Simulation,"
Ph.D. dissertation, Department of Mechanical Engineering, The University of Michigan,
1993.
Lee, WI. and Woo, T.e., "Optimum Selection of Discrete Tolerances," Technical Report No.
87-34, Department of Industrial and Operations Engineering, The University of Michigan,
1987a.
Lee, WI. and Woo, T.e., "Tolerancing: Its Distribution, Analysis and Synthesis," Department
of Industrial and Operations Engineering, The University of Michigan, Aug., 1987b.
Loosli, B.G., "Manufacturing Tolerance Cost Minimization Using Discrete Optimization for
Alternate Process Selection," Master's thesis, Department of Mechanical Engineering,
Brigham Young University, 1987.
Michael, Wand Siddall, J.N., "Optimization Problem with Tolerance Assignment and Full
Acceptance," ASME Journal of Mechanical Design. 103, No.4, pp842-848, 1981.
Speckhart, EH., "Calculation of Tolerance Based on a Minimum Cost Approach,", AS ME
Journal of Engineeringfor Industry 94, No.2, pp447-453, 1972.
Spotts, M.E, "Allocation of Tolerances to Minimize Cost of Assembly," AS ME Journal of
Engineering for Industry 95, No.3, pp762-764, 1973.
VSA-3D, Variation Systems Analysis, Inc., St. Clair Shores, MI, 1993.
Zhang, e. and Wang, H.P., "Simultaneous Optimization of Design and Manufacturing - Toler-
ances with Process (Machine) Selection," Annals of the CIRP, Vol. 41, No.1, 1992.
BIOGRAPHY
Mark Iannuzzi
Mark Iannuzzi is the Vice President of Software Development at Variation Systems Analysis,
Inc. (VSA). Mark is responsible for all software development activities and has produced six
major releases of VSA-3D, VSA-2D and VSA-SQC. Most recently he has created and intro-
duced a new product, VSA-GDT, for analyzing the completeness and well-forrnedness of com-
ponent GDT as compared to ANSI and ISO standards. Mark has served as Manager of the
Design Optimization Laboratory at the University of Missouri where he was responsible for
the creation, distribution and support of a wide range of optimization software.
Dr. Eric Sandgren

Dr. Eric Sandgren received his BS, MS, and Ph.D. degrees in mechanical engineering from
Purdue University. His industrial experience includes positions at IBM, General Motors and
his current position as Director of Design Engineering for the Steering and Suspension Divi-
sion of TRW in Sterling Heights, Michigan. His academic experience includes appointments at
the University of Missouri and Purdue. Dr. Sandgren has conducted research in the areas of
computer aided design, optimization and computer aided manufacturing. He has published in
excess of fifty technical papers and was awarded the Presidential Young Investigator Award in
1984.
16
Optimal Tolerance Synthesis by
Genetic Algorithm under the
Machining and Assembling Constraints
S.KANAI, M.ONOZUKA, and H.TAKAHASHI
Tokyo Institute of Technology
2-/2-10-okayama, Meguro, Tokyo, JAPAN
Tel/Fax: +81-3-5734-2162, e-mail: kanai@mech.titech.ac.jp
Abstract
The purpose of this research is to develop a computer aided method of tolerance synthesis. The
assembly was represented by solid model, and dimensional and geometrical tolerances were
formulated as a set of inequalities constraining substitute features. Differential coordinate
transformation and linear programming were used to analyze the tolerance stack-up. The cost
data base described the machining, inspectional and assembling costs for tolerances. Tolerance
synthesis was represented as the combinatorial optimization problem under the stack-up
conditions. Genetic Algorithm (GA) was applied to solve the problem. The coding method and
genetic operators were discussed. The algorithm was evaluated through a gear box example.
Keywords
Tolerance synthesis, tolerance optimization, genetic algorithms, geometrical dimensioning &
tolerancing, differential coordinate transformation, CAD, cost-tolerance model
INTRODUCTION
In the tolerance synthesis, designers have to allocate the suitable values to the tolerances on
components by considering the trade off between functions and manufacturing cost. Therefore
many researchers regarded the tolerance synthesis as optimization problems. In these problems,
tolerance values on components were the control variables, and manufacturing cost was the
objective function to be minimized. Tolerance stack-up limit conditions were regarded as
constraints on the variables.
As for the optimal tolerance allocation, Bj0rke (1978) proposed the linear programming
method for linear cost-tolerance functions. Spotts (1973) and Chase et al (1988) applied the
Lagrange multipliers method to the nonlinear cost-tolerance model. Lee et al (1990,1993) also
proposed the non-linear programming methods. If cost-tolerance relation is expressed as a
discrete function, the tolerance allocation becomes the combinatorial optimization problem.
Ostwald et al (1977) solved it by zero-one search algorithm, and Lee (1989) solved it by the
branch & bound algorithm. Chase et al (1990), and Wu et al (1988) surveyed and evaluated
several discrete optimization schemes for the tolerance allocation.
The first problem with these researches is that they focused on the optimization only for the
dimensional tolerance. However, when we deal with the geometrical tolerance synthesis, we
have to do the 3D mathematical model for tolerances. Elmaraphy et al (1993) pointed out this
problem, and proposed the 3D tolerance synthesis package. Guilford et al (1993) also proposed
the system in which they used the feasible space and linear programming approach for 3D
tolerance analysis.
The second problem with the previous works is that they only tried to minimize the
machining cost. Actually, the inspectional cost and assembling cost are also affected by the
tolerance values, and we have to consider them as the object of the minimization.
The third problem is that designers usually select tolerance values from several discrete
standard values. So it is realistic to define the cost-tolerance model as discrete functions. In case
of the geometrical tolerance, the cost varies in types of features, types of tolerances even for the
same machine tool and same tolerance value. It causes the tolerance synthesis to be a complex
combinatorial optimization problem.
And the last problem is the robustness and efficiency of the optimization. If we consider the
inspectional and assembling cost in the discrete objective function, the function becomes non-
monotonous. Moreover, not only the tolerance value but also the type of machine tools and
inspectional devices become control variables. It makes the number of control variables
remarkably larger than the one in the dimensional tolerance optimization.
Based on the discussion, the purposes of our research on the tolerance synthesis are to
develop the method of 3D stack-up analysis for the geometrical tolerances, to estimate the cost
factors besides machining cost affected by the tolerance values, and to develop the robust and
efficient algorithm for the large-scale combinatorial tolerance optimization problem. In order to
realize these purposes, we developed the computer-aided tolerance synthesis software, and
made the following approaches in this software:
1 Assembling structure and geometry of the product were expressed by solid modeler.
2 Worst case stack-up analysis for geometrical tolerances were done by linear programming.
3 Machining, assembling, and inspectional cost were described as discrete functions of
tolerance values, type of machine, type of feature and type of tolerance. Sum of these costs
was evaluated as a objective function.
4 Genetic algorithm (GA) was applied to solve the tolerance optimization problem.
2 OVERVIEW OF DEVELOPED TOLERANCE SYNTHESIS SOFTWARE
Figure I shows the overview of developed tolerance synthesis software. It consists of two
modules; tolerance analysis module and synthesis module. The modules were implemented by
C on the Unix EWS. We used solid modeler toolkit (PARASOLID) to represent the assembly
structure and to retrieve the data from it effectively.
The tolerance analysis module accepts the assembly structure model of nominal dimensions,
and the types and values of tolerances for features on each component. It also accepts the data of
the pair of features on two different components where we have to evaluate the stack-up values.
Optimal tolerance synthesis by genetic algorithm 237
Pair of Features for

Stack-up Evaluation
Search for
Feature Chain
Cost-Tolerance
Objective Feature Chain Model Database
Function
& E ualities for Stack-up
~~--~--~~~ Machining Inspectional Assembling
Linear Programming Tolerance
Analysis Module Cost-Tol. Cost-Tol. Cost-Tol.
Model Model Model
Generation of
Initial Allocation
Tolerance Synthesis Module
Figure 1
The module derives the feature-chains which affect the stack-up values between the pair of
features. and converts the tolerance specifications to the inequalities or equalities conditions.
Finally. the module outputs the worst-case stack-up values by the linear programming.
The tolerance synthesis module accepts the pre-specified limit of stack-up values. and it
accepts the stack-up values and simplex multipliers from the analysis module. To evaluate the
manufacturing cost. the module refers the machining. inspectional and assembling cost-tolerance
model. The module searches the optimum tolerance values for every component by using GA.
3 REPRESENTATION OF TOLERANCED FEATURES
3.1 Mathematical model for geometrical and dimensional tolerances
Approximation to the substitute feature

We firstly prepare the mathematical model to express geometrical tolerances for features as
shown in Figure 2. The nominal feature has an ideal form. position. and orientation. while the
real feature has the complex form with errors. We approximate the real feature to the substitute
feature. The substitute feature has a perfect form but has some positional and orientative
deviations from the nominal feature.
Consider the several typical vertices Vk ( k = I.2 •... N v ) which have homogeneous
~(N~'~'i~'~iF~~i~'r~"""sy~=:~~~e"""""""'" ~ "j""
{ll + I VI
, ,
{A : : :... ;...... ... . .....• ~"i~;;~ ,..£;:;"e .lV'...'.

Figure 2 Substitute Feature and Typical Vertex
(a) Parallelism (b) Positional Tolerance

Figure 3 Typical Vertices and Tolerance Zones for Related Features.
coordinates Vk with respect to the nominal feature as shown in Figure 2. Typical vertices are
selected on the substitute feature so as to deviate farthest from the nominal feature. When a
substitute feature deviates, the position of Vk changes to
[ A (ox, oy, Oz, dx, dy, dz) + I ] V k (1)

A (ox,oy,Oz, dx, dy, dz) Rot (x,ox) Rot (y,O y) Rot ( z,oz )
Trans( dx, dy, dz) (2)
where I denotes 4x4 unit matrix, A denotes a 3D differential transformation matrix (Paul,1981)
with respect to the nominal feature, and oX,oy,Oz denotes differential rotation angles about x,
y, z axis, and dx, dy, dz differential translations along x, y, z axis of the nominal feature.
Model for Related Feature

We can regard the geometric tolerance as the zone where the substitute feature must exist. This
is the extension of the model of shape errors by Inui (1991). In case of the tolerance of
parallelism and perpendicularity, arbitrary pair of typical vertices V k ,VI (k~l) on substitute
feature must exist inside of the tolerance zone. We can define this geometrical condition A by
A= {A( &, by, Oz, dx, dy,dz) I A

l::;i::;M, l::;k</::;Nv
[ /;( ox, oy, Oz, dx, dy,dz, Vk , ¥t, to) ::; 0 ] }
(3)
where e denotes the size of the tolerance zone, and the scalar function Ii and integer Mare
uniquely determined by the shape of the feature and the zone. If we express the tolerance of
coaxiarity or positional tolerance, every typical vertex Vk must exist inside of the tolerance zone.
This condition can be obtained only by setting k=l in eq(3).
For instance, for the rectangular feature of a by b shown in Figure 3(a), we select four comer
points as the typical vertices. Then we can express the tolerance of parallelism Apar as
(4)
Similarly, for the cylindrical feature of length h shown in Figure 3(b), we select two end
points of the center line as the typical vertices. Then we can express the positional tolerance Apos
of this line with cylindrical tolerance zone as
(5)
In case of planer feature, Ii in eq(3) is the linear function of ox, oy, oz, dx, dy, dz, while in
case of cylindrical feature, Ii becomes the quadratic function. Similar to the method by Turner
(1990), we approximate the quadratic inequality Ii ~ 0 in eq(3) to the set of linear inequalities.
Modelfor Single Feature

When we express the tolerance for a single feature, we cannot assume that the substitute feature
has a perfect form. In this case, we divide a substitute feature into Ns pieces of partial features
Fs 1, FS2, ..... , FSNs where contacts with other components occur as shown in Figure 4. Similar
to eq(3), arbitrary pair of typical vertices Vik, Vj/ on the two partial features FS i , FSj must
exist in the tolerance zone. Therefore, if we put the differential transformations of partial
features as b.l, b.2,... ,b.Ns , and put the numberoftypical verticesonFs i and FSj as Nvi and Nvj,
the tolerance for the single feature A can be expressed by
A= 1~l "'~i"'~ j"'~N I ,

/\
15,m5,M,l5,i.j5,N,
[ fmC oXi,··,dzi,Oxj, .. ,dzj , l'ik, Vj/, e) ~ 0 1)
15,k5,N" ' 15,/5,N'j (6)
.-~ __ F
E\- Vik Ft_---,------- Nominal Feature

_- - - - - - - - Tolerance Zone
Figure 4 Tolerance Model for Single Feature.
I/> (dh - de)

Figure 5 Contact between Planar Features. Figure 6 Contact between Cylindrical Features.
3.2 Contacts between Toleranced Features

We express the condition of the contact between two planer or cylindrical substitute features as
the set of linear inequalities. As shown in Figure 5, consider the contact between two planar
substitute features F] and F2 on different components. We can consider the differential
transformation flc for F] to make F/ coplanar with F2. Because M typical vertices on F/
must exist outside of F2, this condition r can be expressed by
(7)
On the other hand, consider the fit between two cylindrical feature F] (peg) and F2 (hole) as
shown in Figure 6. If it is the clearance fit, axes of F/ and F2' are not coincident, and two
typical vertices VI /, V 2/ on the axis of F] / must exist inside of the small cylindrical zone whose
diameter equals to the clearance. If we put the upper deviation of the hole diameter as dh, and
lower deviation of the peg diameter as de> this condition r can be expressed by
(8)
If this contact is the tight fit, the condition can be reduced to r = { flc
= 0 }. We approximates the non-linear condition in eq(8) to the set of linear inequalities.
4 TOLERANCE STACK-UP ANALYSIS BY LINEAR PROGRAMMING
Consider the stack-up problem to evaluate the deviation of relative position and orientation of the
target feature on one component with respect to the datum feature on another component. We
express this deviation of the target feature as the differential transformation matrix fl a . From the
relation of the coordinate transformations among features, the six differential elements ox a, oYa,
OZa, dx a, dYa, dZ a in fla can be respectively expressed as the linear function of elements in
differential transfonnation 11 of the toleranced features and I1c of contacts between the features.
We have to convert the maximum deviations ox a, oYa, OZa' dxa, dYa' dZ a to the size of the
zone where the target feature deviates. The size of the zone can be regarded as the stack-up
value. Generally we have to analyze Na tolerance stack-up values esj (j = 1, 2, .. , N a) in an
assembly. Similar to the eq(3), e sj can be evaluated by
(9)
where V kj, l'tj are the coordinates of typical vertices on the j-th target feature, function g ij is
the linear function of OXaj - dZaj, and Mj is the number of the functions to evaluate e sj .
Based on the discussions before, we can define the worst case tolerance stack-up analysis as
the following standard fonn of the linear programming problem;
maximize gij = Cij x t (i=I,2, ... ,Mj, j=1,2, ... ,Na ) (10)
subject to Ax = b (11)
x ~ 0 (12)
where cij is the vector of coefficients of the linear function gij defined in eq(9).
The control variable vector x consists of 1) the elements of differential transfonnations of
toleranced features, 2) the elements of differential transfonnation of the deviations in contacts
between toleranced features, 3) the additional variables introduced by expressing maximum
material conditions, and 4) the slack variables. The linear constraints in eq(11) corresponds to
I) linear inequalities of tolerance conditions A in each feature, and 2) linear inequalities or
equalities in the contact conditions r between two features.
We solve these problems by applying the simplex methods (MJ + M2 + ...... + MNa ) times,
and evaluate stack-up values tsj by substituting the maximum values of gij into eq(9).
5 COST-TOLERANCE MODEL
5.1 Machining Cost Tolerance Model
The machining cost tolerance model describes the relationship among the setup cost, types of
features, types of geometric tolerances, precedence conditions of machining operation, and
discrete cost-tolerance functions. They are described by the following generic fonn :
[MmA, CmAset (Nto t ),

[ftypea [ttypea, CAaa (e), MpAaa], [ttypefJ, CAaj3 (e ), MpAaj3 ] .... ]
[/rypeb [ttypeS, CAbS(e), MpAbs], [ttypeTJ, CAb'1 (e ) , MpAb'1] .... ] ...... ]
where M mA is a name of the machine tool in the factory, (]nAset ( Nto t ) is the set-up cost of this
machine which depends on the lot size Nto t • ftypea, /rypeb' .... are the shapes of features which
can be machined by MmA. ttypea ttype f3 ,....

denote the type of geometric tolerance which can f type =Plane. ttype = Flatness
125
be specified on each feature. CAaa (e) denotes
o
the discrete cost-tolerance function for the feature
100
P o
End Milling
Face Milling
type f'typea and tolerance type ttypea . The Oparating
Cost
<><> • Slab Milling
domain of the tolerance value is divided into e" • Planer
•• <>
75
CAaa (E) Surf. Grinder
e2, ...... , e n based on the IT standard tolerance
• •
0
grades. M l.aa denotes the other machine tool
50
required to achieve the tolerance type ttypea on

feature !typea before the machine MmA.
25 • 0
0 •
o
o o
We established the machining cost-tolerance a
a 0.1 0.2 0.3 0.4 0.5
model for several kinds of machine tools such as Tolerance Value ei [mm]
turning, milling, planer, grinder and jig borer etc.
Figure 7 Discrete Cost-Tolerance
Figure 7 shows an example of the discrete cost-
Function.
tolerance functions in the case of !type=plane and
ttype = flatness.
5.2 Inspectional Cost Tolerance Model

If the tolerance value becomes tight, time and inspectional device becomes more expensive.
Similar to the machining cost tolerance model, the inspectional cost-tolerance model is expressed
as the following generic form:
[ M tA , Ct Asel ( Niol ),
[f'typea [ttypea, (£Aaa I, £Aaa u), CAaa], [ ttype/3, (£AajJI, eAajJ u), CAajJ ] .... ]
[f'typeb [ttypeS, (e Ab~ [, e Ab~ u ), CAb~ ], [ t lype 1J, (e AbT//, e AbT/ u) , CAbT/ ] .... ] ... ]
where MIA is a name of the measuring device. CtAsel ( Nlot) means the set-up cost function of
M tAo (e Aaa I, e Aaa u) denotes the possible measuring range of M tAo CAaa denotes constant
inspectional cost for feature type ftypea and tolerance type ttypea .. We established this model for
CMM, roundness measuring machine, dial gauge, micrometer calliper, plug gauge etc.
5.3 Assembling Cost Tolerance Model

We only evaluate the assembling cost of the insertion operation of peg into hole because the
difficulty of the insertion especially depends on the tolerance of fits. Consider the features of a
peg F p and a hole F h in fit relation, and the fit condition ( tight-.fit or clearance-.fit or
transition Jit ) between them has been pre-specified from the functional requirement. If we put
the dimensional tolerances of Fp and Fh as ep and eh, and fit condition as ft..P h , the set of
feasible pair of tolerance classes suitable to the fit condition can be found as {(tcp"tchl), (tcp2,
tch2), ....... } where (tcpi, tchi) means a grade of fit such as (H6, h6). If we select one grade of
fit (tcpj, tchj) from it , we can distribute tolerance values cp ,Ch into the upper and lower
deviation for the diameters of the peg and hole. Then the assembling cost Ca ph to insert the F p
into Fh is evaluated by the
Caph = C asm (fryh, tCpj, tchj, Apph) (13)
where Apph denotes whether the insertion is done automatically by the assembling machine or
manually. The function C asm was modeled based on the Boothroyd's work (Boothroyd,1986)
where the cost of insertion was estimated for fit condition.
6 OPTIMUM TOLERANCE SYNTHESIS BY GENETIC ALGORITHM
6.1 Tolerance Synthesis Problem

The tolerance synthesis problem in this research can be formulated as the following
combinatorial optimization problem;
minimize (14)
subject to ( j = 1,2, ... , N a ) (15)
where C is the total manufacturing cost which are the sum of all machining, inspectional and
assembling cost for a product. N is the number of tolerance values to be determined. Cj denotes
of the tolerance value, Mmj denotes the machine tool used in the final operation, and Mtj
denotes the inspectional device. The Csj is the function used to calculate the tolerance stack-up
value from eq(9), and Caj denotes the pre-specified limit of the stack-up deviations.
The properties of this optimization problem are: 1) the objective function is non-monotonous
for the control variables, 2) there are ten to one hundred tolerance values only for one assembly,
and 3) the domains of the variables are numerical values and symbols. Therefore, we adopted
the genetic algorithm(GA) in order to solve the tolerance synthesis problem. It is suitable for the
large-scale combinatorial optimization, and it avoids the solution to fall into the local minimum
by controlling the stochastic operations. The application of GA to the tolerance synthesis has
been reported by Lee (1993) and Nassef (1993). However, Lee only dealt with the synthesis of
the dimensional tolerances. And in both works, they only minimized the machining cost.
6.2 Coding and Fitness Value of Tolerance Allocation
We code the control variables in eq(14) as the genotype of an individual. An individual

corresponds to one plan of tolerance allocation. The genotype of the individual I is defined by
I = [ G 1 Gz ...... G j ..•.. G N ], where G j = (Cj ,Mm j , Mtj)t denotes a gene
corresponding to the combination of tolerance value Cj, name of the machine tool Mmj , and
name of the inspectional device Mt j •
We defined the fitness functions!v for an individual as
(Csj ::; Caj)
(csj > Caj)

(16)
where the scalar function q imposes the penalty when the stack-up conditions are not satisfied,
and w is the positive coefficient of the penalty. Csj is the stack-up value, and we have to repeat
this evaluation many times. So we firstly find the simplex multipliers, and do the sensitivity
analysis in order to quickly re-evaluate these values.
6.3 Genetic Operators for the Optimum Tolerance Synthesis
Creation of the initial population

We initially create Qindividuals{ll, 12, ...... , IQ}, (h=[ Gkl Gk2 ..... Gk; .... GkN])
in the gene pool corresponding to the Q different plans of the tolerance allocation. The initial
values in each Gk! are randomly selected from the cost tolerance model.
Selection operator
Selection operator determines the set of individuals which remain at the next generation. If we
put the fitness value for the individual h as !vk, and positive factor as r, the probability Psk that
the individual Ik will be selected from the current generation is calculated by
(17)
Crossover operator
In the crossover, we generate the new individuals la' and Ib' from parent individuals la and h
with the probability Pc. The la and h are selected from the population with the probabilities
P sa and P sb. They are divided into sub-parts at multiple points of crossover, and the new
individuals are obtained by swapping them between these individuals.
We use two types of crossover. In the first type, the points of cross over and the number of
them are entirely selected at random. In the second type, the points of cross over are determined
so that sub-parts belonging in one component are not divided after crossover. We execute 70%
of first type crossover and 30% of the second type.
Mutation operator
In the mutation, we randomly select and change the contents of genes randomly selected with a
probability Pm. By considering trade-off between the premature problem and the convergence,
Pm gradually decreases as the generation increases by
Pm = Pm-init ( 1.0 - 0.5Jt1T ) (18)
where t denotes the generation, T denotes the generation of convergence where we stop the
search, and P m.init is the mutation probability at an initial generation (t=O).
By considering the technological aspects, we apply three types of mutation operators:
Mutation for tolerance value in one machine tool
In this type of mutation, we randomly increase or decrease tolerance values in the genes but
keep the selected machine tool. This operation aims to improve the machining cost close to the
current solution. If we represent the ratio of Pm2 for the Pm at t =T as P m2.end, the possibility of
this mutation Pm2 is defined by
Pm2 = Pm P m2-end (t / T)2 (19)
Mutation to decrease the number of machine tools

In this type of mutation, we randomly increase or decrease the tolerance values so that the
number of machine tools used for a component becomes minimum. The possibility of the
mutation Pm3 has to increase as the generation increases, and is defined as same as eq(l9) .
Full random mutation
In this type of mutation, a gene is randomly selected from one individual and the tolerance
value, name of machine tool and inspectional device are entirely changed at random with the
probability Pm 1. We imposed the constraint P ml = Pm - P m2 - P m3 on these probabilities.
7 EXAMPLES OF TOLERANCE SYNTHESIS
7.1 Problem Definition

As an example, we consider the tolerance synthesis for the simplified gear box assembly shown
in Figure 8. This assembly consists of ten components, and four of them are standard ball
bearings. Only the types of the tolerance are pre-specified as shown in Figure 9. We have to
determine 29 geometrical or dimensional tolerance values £1 - £29 in this figure.
From the functional requirements of the gear box, we consider the next three conditions on
the limit of tolerance stack-ups, and find the optimum tolerance allocation for each condition;
Condition 1: deviation of the distance between two axes of pitch cylinder of gear 5 and
gear 6 must be within ±0.15mm.
Condition 2: parallelism of the axis of pitch cylinder of gear 6 for the datum axis of pitch
cylinder of gear 5 must be within 0.20 mrn.
Condition 3: condition 1 and 2 must be satisfied simultaneously.
7.2 Optimization Process and Results
We determined the genetic parameters as Q= 100, W= 10 3 , r= 1.0, T =550, Pc = 1.0, P m-init =

0.05 and Pm2-end=PmJ-enJ= 0.4 from the pre-experiment. Figure 10 shows the improvement
40~-------------------------,
(X102)
35
mean of C'
c' among
individuals
minimum of C'
among individuals
15
10
0 100 200 300 400 500 550
Number of Generation t
Figure 8 The Gear-box for Tolerancing. Figure 10 Cost Optimization Process by GA.
CD £7
B 38£8
~,~, 1--
__- __
-f--_-.--.- ~.L. !-= .J- l
L.:==-:
25
Figure 9 The Values of Tolerances to be determined on the Components.
of the total cost C' ( = lI!v) in case that condition 3 was imposed. We can find that the
optimization quickly proceeds from t=O to 100 and reach the near-optimum tolerance allocation
at t=550. It was shown that the proposed GA based optimization worked well.
Table 1 Optimum Tolerance Values for Different Stack-up Conditions [mmj

vompo
Type of Tolerance Condition 1 Condition 2 Condition 3
nent
Parallelism
CD Parallelism
e1
e2
0.5
0.5
0.1
0.1
0.1
0.1
e3
Positional Tol. 0.04 0.04 0.04
Dimensional e4 0.02 0.02 0.02
® Tol.
Pos~ional Tol.
e5
e6
0.0
0.04
0.0
0.04
0.0
0.04
Tol. e8 0.0 0.0 0.0
Pos~ional Tol.e9 0.04 0.04 0.04
® Tol. en
Pos~ional Tol.
e12
0.0
0.04
0.0
0.04
0.0
0.04
Tol. e14 0.0 0.0 0.0
Coaxiality e15 0.02 0.01 0.01
Tol. e17 -0.01 -0.005 -0.005
@) Dimensional e18 0.005 0.012 0.012
Tol. e19 -0.005 0.002 0.002
Dimensional e20 0.005 0.012 0.007
Tol. e21 -0.005 0.002 0.002
Coaxiality e22 0.01 0.01 0.01
® Dimensional e23
Tol. e24
0.02
0.0
0.01
0.0
0.01
0.0
Coaxiality e25 0.01 0.01 0.01
Dimensional e26 0.005 0.012 0.005
® Tol. e27
Dimensional e28
-0.005 0.002 -0.005
0.012
0.005 0.005
Tol. e29 -0.005 -0.005 0.002
Total Manufacturing Cost 1261.4 1337.4 1342.4
Table.1 shows the optimum tolerance values for the conditions 1,2, and 3. Table.2 shows
the machine tools and inspectional devices selected for the optimum tolerance allocation. From
Table1, the tolerance values of parallelism el and e2 in condition 2 are smaller than the ones in
condition 1. This is because the distance between two gears are not so sensitive to the el and
e2. At the dimensional tolerances e16 ,e17, e23, ande24, the tolerance values in condition 2 are
smaller than the ones in condition I. This is because the parallelism between two axis of gears
are very sensitive to the clearance of the fit between component 4 and 5. From these results, it is
shown that the system can distinguish the sensitivity of the tolerance stack-up, and the system
can allocate the loose tolerance values with low sensitivities to reduce the machining cost.
On the other hand, we have the same tolerance allocations for e3-e14 in components 2 and
3, even if the stack-up conditions are different. For these components, the machining process
were uniformly done by the drilling and reaming as shown in Table2. This result reflects the
proposed mutation operators to unify the number of machine tools in one component.
Table 2 Machine Tools and Inspectional Devices selected

for the Optimum Tolerance Allocation
Compo Tolerance Machine Tools Inspeclional Devices
nent
£1
CD
Condition 1 : End_Milling
Dial Gauge
£2 Condition 2, 3 : Slab Milling
£3
£4 £5 Drilling Reaming Positional Tol. : 3DCMM
® £6
£7 £8
Machine
~ Machine Dimensional Tol. of Hole: Plug Gauge
£9
£lO£n
® £12
Drilling
Machine ~
Reaming
Machine
Positional Tol.: 3DCMM
Dimensional Tol. of Hole: Plug Gauge
£ 13 £ 14
£ 15
0 £ 16 £ 17
£ 18 £ 19
Rough
Turning ~
Fine
Turning
Micrometer Calliper
£ 20 £ 21
£ 22 Gear Hobbing Machine 3DCMM
® £23 £ 24 Rough Turning ~ Fine Turning Plug Gauge
£ 25 Gear Hobbing Machine 3DCMM
® "26 £ 27
£ 28 £ 29
Rough Turning ----.. Fine Turning Micrometer Calliper
7.3 Computational Efficiency of the Optimization
Under the condition 3, it took 356sec for the tolerance analysis, 140sec to calculate the simplex
mUltipliers, and 665sec to find the optimum tolerance allocation by HP9000/720. The program
checked about 55x103 different tolerance allocation plans by GA. The number of feasible
tolerance allocation plans is about 1012 in this case. It suggests that if we take the exhaustive
search, we will have to take 3.8x102 years to find the optimum allocation. From this estimate,
GA is efficient to find the optimum tolerance allocation without combinatorial explosion.
8 CONCLUSIONS
1. For the purpose of geometrical tolerance analysis, tolerance of features and contacts between
features could be uniformly expressed by inequalities for the elements of 3D differential
transformation. Stack-up values could be expressed by the linear functions of these elements.
We could evaluate the worst-case stack-up value by applying the linear programming.
2. For the purpose of realistic cost estimation, machining, inspectional, and assembling cost
were respectively evaluated as a discrete function of the type of machine, features and tolerance,
and tolerance value in our cost-tolerance model.
3. We formalized the geometric tolerance synthesis as combinatorial optimization problem, and
solved the problem by the genetic algorithm (GA). The crossover and mutation were
customized for the property of tolerance synthesis. It was confirmed that the proposed GA based
method could find the optimum tolerance allocation plan for a rational period.
ACKNOWLEDGEMENT
This research was supported by the Ministry of Education of Japan as Grand-in-Aid for Science
and Research (C05650lll) ,and by THOMAS research body in T.I.T.
REFERENCES
Boothroyd,G. and Dewhurst,P. (1986) Design for Assembly. Boothroyd Dewhurst Inc.,
Wakefield, RI.
Bj0rke, 0. (1978) Computer-aided Tolerancing. Tapir Publisher, Trondheim.
Chase, K.W. and Greenwood, W.H. (1988) Design issues in mechanical tolerance analysis.
Manufacturing Review, 1(1), 50-59.
Chase, K.W., Greenwood, W.H., Loosli,B.G. and Hauglund,L.F. (1990) Least cost
tolerance allocation for mechanical assemblies with automated process selection.
Manufacturing Review, 3(1),49-57.
EIMaraghy, H.A. and EIMaraghy,W.H. (1993) A system for modeling geometric tolerances for
mechanical design. Proc.of 3rd CIRP Seminars on Computer Aided Tolerancing, 11-24.
Guilford). and Turner,J.U. (1993) Advanced tolerance analysis and synthesis for geometric
tolerances. ASME Int. forum on dimensional tolerancing and metrology, CRTD 27, 187-
198.
Inui,M. and Kimura,F. (1991) Algebraic reasoning of position uncertainties of parts in an
assembly. Proc. of ACM Symp. on Solid Modeling Foundations and CAD/CAM
Applications, 419-428.
Lee, J. and Johnson, G.E. (1993) Optimal tolerance allotment using a genetic algorithm and
truncated monte carlo simulation. Computer-Aided Design, 25(9),601-611.
Lee,W.J. and Woo, T.e. (1989) Optimum selection of discrete tolerances. Trans. of ASME
Journal of Mechanisms, Transmiss.and Automation in Design, 111, 243-251.
Lee, W.J. and Woo, T.C. (1990) Tolerances: their analysis and synthesis. Trans. of ASME
Journal of Engineering for Industry, 112, 113-121.
Lee, W.J., Woo, T.C. and Chou, S.Y. (1993) Tolerance synthesis for nonlinear systems based
on nonlinear programming. IIE Trans., 25(1),51-61.
Nassef, A.O. and EIMaraghy,H.A. (1993) Allocation of tolerance types and values using
genetic algorithms. Proc. of3rd CIRP Seminars on Computer Aided Tolerancing, 147-159.
Ostwald, P.F. and Huang, J. (1977) A method for optimal tolerance selection. Trans. of
ASME Journal of Engineering for Industry, 99B(3), 558-565.
Paul, R.P. (1981) Robot Manipulators, MIT Press, Massachusetts, 85-118.
Spotts.,M.F. (1973) Allocation of tolerances to minimize cost of assembly. Trans. of ASME
Journal of Engineering for Industry, 95, 762-764.
Turner,J.U. (1990) Relative positioning of parts in assemblies using mathematical
programming. Computer Aided Design, 22(7), 394-400.
Wu,Z., EIMaraghy,W.H.and EIMaraghy,H.A.(1988) Evaluation of cost-tolerance algorithms
for design tolerance analysis and synthesis. Manufacturing Review, 1(3), 168-179.
BIOGRAPHY
Satoshi KANAI is an associate professor of mechanical engineering and science at Tokyo

Institute of Technology. He also holds the post of an associate professor of system and
information engineering at Hokkaido University. He received his PhD from Hokkaido
University. His research interests include modeling and simulation of manufacturing process,
factory automation, and solid modeling.
Mamoru ONOZUKA is an engineer of plant design in Toyo Engineering, Inc. (Tokyo). He
received his MS in mechanical engineering from Tokyo Institute of Technology. His research
interests include the computer-aided assembly planning and tolerance design.
Hidetomo TAKAHASHI is a full-time research associate of mechanical engineering &
science at Tokyo Institute of Technology. He received his MS in precision engineering from
Hokkaido University. His research interests include an application of Virtual Reality to the
CAD/CAM.
Question and Answers after the presentation (Speaker: Satoshi KANAI )
Question from V.T.Portman: I think the values in your machining cost-tolerance model
are not accurate, because you assumed the uniform cost-tolerance tables for one kind of machine
tool. How do you consider the uncertainty of the accuracy of machine tools and machining
process in your cost-tolerance model?
Speaker: We primarily focused on establishing the general formulation of cost-tolerance
models more sophisticated than the ones proposed in other researches. So we considered the cost
factors besides the machining cost, and established the model dependent on types of features and
geometrical tolerances even for the same machine and same tolerance values. In this paper, we
assumed the standard values on the discrete cost-tolerance tables for the machine tools by
referring several machinability hand-books. But, in our software, we can easily modify and
adjust the values on the cost-tolerance tables by editing the external files, and prepare this table
for machine by machine. Usually, every machine shops or factories quantitatively manage the
uncertainty of their process as a set of process limits or process capabilities, and they also has a
set of standard cost. So we can fit our cost-tolerance tables with them, and can reflect the real
situation to our models.
Question from M.P.IANNUZZI: In your generic algorithm, you represented one gene
as a vector including a real-value component and symbol-value components. How did you
crossover the genes including these different types of components?
Speaker: As you pointed out, we couldn't use the one-dimensional and single-string type
expression for the genes, because if we use this type of expression, essential information may be
destroyed at the crossover operation. So we used multi-dimensional expression of genes and
applied multi-dimensional crossover. We swapped the tolerance values only among the set of real
values, and did the names of machines only among the set of symbols.
17
Holistic Approach and Advanced Techniques &
Tools for Tolerance Analysis & Synthesis
Jean M. Parks
Xerox Corporation
800 Phillips Road, Webster, NY 14580, USA
Phone: (716)422-7702; Fax: (716)422-8548; jmp.wbst147@xerox.com.usa
Abstract
This paper presents a holistic Methodology on Design and the associated Toolset developed
to enable its implementability; the approach is based on comprehensively addressing
effects of Variability. In Design, Tolerancing for Fit and for Function is one of the most
crucial and interlinked efforts during technology and product development. Tolerancing
for Function, especially, has been little addressed, and at best rudimentarily, by the engi-
neering community. This is because to do so appropriately and adequately requires local
and global approaches to enable decision making of nominals and tolerances. For this, one
not only needs techniques and tools for properly conducting Contribution Analysis and
Variability Analysis for any given relationship and set of input variabilities, but one must
be able to quantitatively propagate the thousands of variabilities to a few resulting
performance variabilities. Lacking these capabilities, as a user-friendly computational
toolset that is 'complete', has been the fundamental level barrier to holistic Tolerancing.
This has now been provided by our breakthrough Toolset. Quantification and decision
making require yet more; the Methodology addresses those and putting relevant Design
Practices in context. The Methodology has been implemented.
KeyWords
Tolerance Analysis, Tolerance Synthesis, Tolerancing, Variability Analysis, Reliability,
Sensitivity Analysis, Contribution Analysis, Latitude, Design Methodology, Robust
Design, Optimization, Taguchi Methods, Probabilistic Design, Probability Distributions.
1 INTRODUCTION
1.1 Background
For many corporations, it is without question that Time-to-Market is heavily driven by
their competency in attaining Design Latitude for products. For example, in Xerox a
feasibility machine that performs excellently for nominal customer input can be put
together very quickly, yet it could take us hundreds to thousands of person-years to
productize it, i.e., to deliver latitude. What is the missing fundamental level capability? It
is in enabling probabilistic analysis be part of the natural work process. (Note: Definitions
will be given in § 1.3)
Think of it, in designing electro-mechanical products, virtually every entity is probabi-
listic, yet the engineering community has traditionally been schooled in dealing in the
Deterministic realm, only to add onto typical work processes some ways of addressing
Variability, for example, conduct Geometric Tolerance Analysis to analyze Fit, develop
latitude (typically experimentally), or conduct Taguchi's Parameter Design to optimize
nominals of the control factors. Do these comprise an adequate set? Even if they did,
where are interconnects? Do user-friendly tools and techniques exist? Are there asso-
ciated unforeseen shortfalls? Are the engineers equipped to execute these, and do so with
adequate understanding?
While the mentioned practices have been recognized as difficult, the subject ofProba-
bility has been typically perceived by engineering management as somewhat simple. After
all, most engineers understand something about Tolerance stack-up, R88, and some even
Monte Carlo. But appropriate application of Probability is far from simple. In fact, limita-
tions associated with even how one uses R88 are not generally understood. And, worse yet,
some existing tools & techniques, and even concepts, contain some erroneous aspects of
implementing basic probability. We will show in §3 that when one ventures into proba-
bilistic space, without rigor, one is in dangerous ground! The net effect is, the subject and
practice of Tolerance Analysis and Synthesis has been in relative infancy. Indeed, the
following portray the situation:
'Research in Tolerancing is greatly needed.' (Tipnis, 1989)
'Tolerance Design ... is not terribly well understood.' (Cutler, 1993)
'I'd go to the other end of the building to use a Tolerance Analysis tool. .... We've got
to know how to avoid functional surprises from every build.' (Parks, 1990)
1.2 Basics
Clearly a big hole needed to be filled. To do so required focusing on providing the capabi-
lity for 'Functional Tolerancing'; in turn this meant one would need to develop a holistic
methodology on designing against Variability. Thus we started the effort; such a methodo-
logy has now been developed and implemented. Till recently the approach was named 'A
Unified Methodology on Design for Latitude' (UMDFL) (Parks, 1991, 1992). With en-
hanced concepts incorporated, and for easier referencing, it was recently renamed as 'A
Holistic Methodology on Probabilistic Design', or simply 'Holistic Probabilistic Design'
(HPD) (Parks, 1995). In this paper, 'the Methodology: or 'HPD', is used to stand for HPD
Methodology. The Methodology shows how to relate the thousands of piece part variabi-
lities and other noise to a few performance characteristics, how to narrow down the problem,
and how to make enhanced decisions in Synthesis.
To enable implementing the Methodology, we developed a 'complete' suite of tools for
the Umbrella Area of Variability Analysis (which includes Tolerance Analysis and Sensi-
tivity or Contribution Analysis). Labeled as the 'HPD Toolset', its capabilities were driven
by the quantification needs of the Methodology. The Toolset goes much beyond utilizing
Monte Carlo, Linearization & R88, and other techniques in existence. In fact, we devel-
oped a technique (the 8MB technique) for 'exact' computation of resulting distributions.
Holistic approach and advanced techniques & tools 253
This latter capability, together with being able to (i) accept any type of input information,
(ii) handle any type of relationship, and (iii) quantitatively propagate variability, are
what contributed to 'completeness' of the suite. Completeness and user-friendliness of the
Toolset are the key factors for executability of the Methodology. Additionally, these attri-
butes enable it to be considered as a stochastic calculator, obviating one's need for possess-
ing probability expertise, hence enabling probabilistic analysis be part of the natural work
process. Thus, when one sees a function, say, Z = g (x, y), one can think of it as a function
of random variables instead, and think naturally of the distributions of x, y, and z.
One other aspect of our advanced techniques and tools will be briefly discussed in §3:
We have been able to show through the 8MB technique the adequacy or sometimes
unexpected gross inadequacy of prevailing probabilistic techniques used; one example is
that for Contribution Analysis!
Beyond the HPD Toolset, it is apparent that implementing the Methodology may
require some advanced levels of analysis and modeling skills. It will be pointed out that
stochastic modeling is more than a simple extension of deterministic modeling, and we'll
briefly discuss approximations.
Due to length restriction, this paper is in very compact form. Only the basics of the
Methodology presented in detail earlier (Parks, 1991, 1992) are recapped here. These are
integrated with the concept extensions of the Methodology and the implementation
specifics which are new. Reference is made wherever a discussion would be significantly
enhanced by reading the earlier works. Since we have now gone much beyond the original
intent of the Toolset, and its development recently completed, this aspect is a primary
focus. But in this paper, there is only space to outline and minimally describe the content;
a companion paper (Parks, 1995) focuses more on the Toolset. Brief discussion of some
application examples concludes the paper.
1.3 Definitions/Acronyms
The definitions below are (a) to aid understanding the extension from geometric Toler-
ancing to Functional Tolerancing and (b) to aid clarification of the many practices relating
to Variability such that they can be put in context within the Synthesis activity.
• Stochastic: Used interchangeably with Probabilistic.
• Tolerancing: Synonymously stands for Tolerance Analysis & Synthesis.
• Analysis & Synthesis: Example: Let z = g (x, y). Analysis is when x & yare given, one
obtains z. Synthesis is when z is given, one makes decisions (adequate or best) on x & y
such that they would enable meeting the target z ; Optimization is making the best
decisions (vs. adequate). The Synthesis process requires Analysis; Tolerance Synthesis
requires Variability Analysis (more below).
• Product Design: For electromechanical products, Product Design is actually the
activity of synthesizing in the Stochastic realm. Attaining Fit & Function, the latter
within cost & performance constraints, comprises the heart of the Design engineers'
hardcore work content. Achieving Function with wide Latitude IS achieving Robust-
ness. (Fit alone is the simpler part.)
• Latitude: Used at the product level, latitude refers to the amount of manufacturing
and usage variability the product can tolerate with negligible impact on its targeted
level of performance; the larger the latitude, the more robust the product. Similar
meaning holds for describing subsystems, or lower level entities. Throughout techno-
logy and product development, latitude is to be expanded through selection and control
of critical parameters. In general, latitude is associated directly with the quantified
effect of variability; in §5, we'll propose a quantitative definition for latitude.
• Robust Design: Synonymous (by our definition) with 'Design for Latitude'. HPD is, in
fact, a comprehensive Robust Design methodology.
• Critical Parameters: Those parameters which are determined to be critical to the
product's function. Typically, they are at a higher level than that of piece part dimen-
sion, and are controllable during technology and product development (e.g., forces,
temperature, relative positions). (Often, some of these are the 'control factors' in
Taguchi Methods.) The 'art & science' of technology development is to minimize the
number of Critical Parameters such that many others thought to be critical would be
noises against which these will be robust.
• Tolerances; Variability Spread: Tolerances refer to allowable variability limits from
the nominal of a characteristic of not only a geometric part but also of any part or
material (Tipnis, 1989). The equivalent term for the characteristic of any higher level
entity is ± Variability Spread about the mean. Variability Spread is sometimes simply
referred to as Spread, it's associated interval is the Range of Variability.
• Variability Analysis (VA); Tolerance Analysis (TA): Common usage of conducting
VA and TA: Typically, when analyzing resulting variability of geometric entities or of
Critical Parameters, one uses the term TA; when doing so for characteristics of higher
level entities, one uses the term VA. But VA can be used for both. Indeed the term VA
will also be used to mean the Umbrella Area of Variability Analysis, see §2.3. (Guilford
& Turner state: 'Tolerance Analysis is the process of taking known tolerances and com-
puting their effect on a design function' (Guilford, 1993). Thus researchers also use TA
to mean both as well; however, this is not a common usage when the design function is
of a higher level.)
• FOV Flow of Variances
• 5MB Stochastic Modeling Based
• RV A Random Variable (Where notation distinction is needed, RV s are in bold.)
• Cp; Cpk = Manufacturing Process Ca~ability indices
(Note: Conventions we use are (i) capitalizing first letters of terms that need to stand out in the
text (e.g., Design, Synthesis) and (ii) quotes being used on terms with their special meaning (e.g.,
'model'); continued use is situational.)
2 THE HOLISTIC APPROACH

2.1 Introduction
The holistic approach addresses all aspects relating to Variability and how to quanti-
tatively deal with those in technology and product development. Since there are existing
'Best Practices' or concepts (e.g., Taguchi Methods, QFD, 60 Quality) that overlap some
parts of the holistic approach, or are even included, we will show such relationships. That
is part of 'putting all in context'. See §2.3.
As is often the case, the practice is easier than reading the description. The simplest
way to start reading is to assume one has a simple (and local) problem at hand and one
needs to see if the resulting variability is too much, what to do about it, and what if this
variability is a contributor to a bigger system one is analyzing.
2.2 Basic Assumption

The basic assumption is that we are dealing in product segments where product cost is a
major competitive factor; thus one must achieve highest performance with as low cost
tolerances as possible. This means achieving the narrowest performance variability with
the largest possible manufacturing tolerances while handling the widest possible
customer noise is the objective of the technologist and Design engineer. (Recognize that
Taguchi Methods aims at the same objective.) Since a complex product has thousands of
parts and a number of performance attributes, Synthesis requires best techniques for
comprehensively addressing effects of variability and making Tolerancing tradeoffs.
Note that even leaving aside associating cost to tolerances, synthesizing would still
require comprehensively analyzing effects of all variabilities on performance quality.
2.3 The Methodology

Overview
The essence of HPD is that one must articulate the Flow of Variances (FOY) from the
lowest piece part level to ultimately the performance level, covering the variabilities due
to manufacturing, customer noise, and usage degradation; this is given as the generic FOY
shown in Figure 1. Part of the implementation process is to flesh out a detailed level FOY
(Figure 2) and, if necessary, one or more at higher levels (Figure 3). These flow relevantly
towards product performance attributes. How the FOY relates to 'modeling' is explained.
Having articulated, one is to quantify the articulated flow. Wise approximations and
experimental information can replace quantifying some partial FOYs. The Methodology
recommends segmenting the total problem at hand and making decisions/trading off
locally, and then globally. These drove the requirements for the HPD Toolset.
Schematic depiction of the Analysis and Synthesis activities in the dotted box in the
upper right-hand corner of Figure 1 shows the coupled solid & dotted arrows for Synthesis.
This simply means: (D The information flow is from Manufacturing to Quality Loss for
Analysis and vice versa to enable synthesizing; (ii) aside from direct Optimization, much
decision making, including tradeoffs, during Synthesis can be executed with powerful
WHAT-IF Analysis, e.g., for Tolerance Design. Another key element in Synthesis is
correctly conducting Sensitivity/Contribution Analysis, thus enabling properly narrowing
the problem to addressing the higher contributors.
Problems engineers encounter at anyone time range widely in complexity; likewise,
the associated FOYs. Indeed, many tolerancing or variability decisions are made at the very
local levels and do not need FOYs, but those tolerances/variabilities could feed into FOYs.
Evaluating via the Toolset and tracking and utilizing such variability information for
more global level analysis and tradeoffs are part of the Methodology.
Articulating the Flow of Variances (FOV)

GENERIC FOV: The generic FOV is shown in Figure 1. It serves to remind the
engineer to not forget all types of contributing elements and relationships. (Other arrows
may exist (Parks, 1992).) Interpretation of some elements in this FOV are according to the
problem at hand, e.g., if one is working at the subsystem level, then 'Performance' means a
performance characteristic of that subsystem and 'Component' may be an electro-
mechanical component or another subsystem that is interrelated to this subsystem. The
.......................
j Analysis _ ~
!Synthesis ~ ~
rMt·iF;-m~i;ti;n~~/
/
/
r - - - - - __ L~,
L .!~~~i~~ .J I Design Effects on I
I Contaminants Generation I
I (o!::!:r:s~':ge I
--l~_~_~_~: ________________________ J--::---.,--r

.......................
j Need:
L -- -- -- • Tolerance/ContriblVariabi-
lityl Performance Analyses
• Relationships I ... I
V = Variance ; • Experiments/Testing ( .)
:::: : J == Additional factors which affect variances L...~ ..?~~.~~~~.~::· )
Figure 1 Generic Flow of Variances (FOV).
variabilities are linked by the arrows in the flow. The solid arrows are not simply quali-
tative process flow arrows; they have meanings. They represent relationships that in some
way, empirically, analytically, or via approximations, one must express if one were to
quantify the flow. Think of these relationships as 'models' in the broad sense; i.e., they can
be equations, logical algorithms (continuous or not) including computer programs, or lab
data.
DETAILED LEVEL FOV (FOV-d): Using the generic FOVas the guideline, the next
step is to flesh out an FOV -d for the problem at hand for a characteristic to be quantified.
(At the product level, the characteristic would be a customer dissatisfier; at lower levels, it
could be, say, the charge density of the photoreceptor or even the torque of a clutch.) An
FOV -d is actually part B of a 3 part process as shown in Figure 2 where we have excerpted
a few representative entities in the FOV-d for an image-to-paper Registration problem.
Constructing the FOV-d is related to how one would 'model' the problem; iterating the
modeling process would possibly rearrange the flow. The step before articulation, as for
modeling, is often defining a coordinate system. In the FOV-d, 'level' means hierarchical
level in the Flow; it is indicated also by the indentation of the entity. Having defined the
symbols for the entities, it is simple next to represent the dependencies as functional
relationships, but only in symbolic form for now, as shown in Part C. The purpose is to list
what needs to be understood; one will then state in the right side column where or how to
obtain the infonnation. If approximations are made for some, they are tracked for record if
no further work is needed, or for signaling supplementary work content to substantiate
the approximations. Here we have not shown addressing the link to Quality Loss and
Total Cost to Customer in the FOV-d; for those, see our earlier paper (Parks, 1992). The
above 'modeling' process is far more detailed than for typical Detenninistic modeling since
all variabilities must be accounted for and later dispositioned; many will not be of signi-
ficance, and many will be made robust against.
P. 1: • A schematic ofthe subsystem being analyzed

PART A • Coordinate system defined
• Performance characteristic (and target variability range - optional) specified
PP. 2 -4: Hierarchical Dependencies of Variabilities

~ (A quite incomplete part of an FOV for an image transfer to paper registration problem)
level DEPENDENCIES Symbol Attributed Var. Cateq'y

~ Cycle Time ( ... time for P/R LE to reach pt F ... ) tF sub-FOV not shown here
....~ .....~ .. !!'!!.~.?~.~~~~~.'!.~.~~~!':'.~p..~ f. .......................... .... h ....................................... ..

... ~ ........ ~ ...~~.~~!.~!'!!~ ~~~~'!.~~. ~~.'!!~ .'~.~'!!. ~!~ .~!~ .t.~ .~~.~.......~~~ ......~.~~~:. ~~~.~. ~F..~~ ........ .
... ~ ............. ::. ~~.~ .~'? ~.~!!!~. ~,?~!~~.~~~.~i~~....................... ?<.~'.:.: ~.~ ..~~.r.t.~ i.. ~~~~~b.ly' ............ .
... ~ .............~:. ~~~!.: .~~.!!I.~ .~~~~~'?~ .~~~!.f.i~!~~~.................. PP.t! .... .................................. .
... ~ ..................:. ~~?~!. !~~.~.\~J.............. :.. '!.IJ............... ...'!.IJ: .~.....~.s.~.~~!.~~t. ~.i.x. ~.i.x.......... .
... ~ ........ ~ .. ~h.~~.~ .~~. ~~;~!~!.~~~!!!'?~ r?~':.~ ~'?~~ ~~.Nl?? ..... .... !:.f! ....................................... .
... ~ .............~:. ~~~:.~~.~ .:~~.~~??.'.t.~?~.~~.~!~~......................... r:!:r:. .....~.~?~~.t.'X: :.~i~. ~.~~!':.::: ... .
... ~ ........ ~ .. !.~~~.!~ ~~~'J.;~~.~.~'!. ~'?!!~............................ .... ~!.I ..... .................................. .
... ~ .............~: ..~~~.~~.I.t? ~.t.~~~.~~~.I.I':.~. ~.~!! .............................~9. ......~!?~~.~~~!~~.i~~:. ~~( ....... ..
3 -- t4
:: :~::: :::::::: :::: ::::: !~~~~!~~! :~~~:~~i~:::: :::::::::::::::::::::::::: : : :~~::::: :::::::::::::::: : : : : : : : : : :

Drive Shatt Actuation Time
... ~ ...................... :..~!~.~~:.~~ .~!. ~?I! ........ :.. ~.~~~.?f. ~~I.I..... .I?~t .'!':!I. .. .................................. .
. .. ~ ..................:.~!~:.c.~.~~~.~~~?~}!~.~ .............................. ~? ....................................... .
5 - Torque T
6 - Effect of Wear/contamination
4 - Gears Backlash B Mfg process (vendor)
PP.5-6: Hierarch. Var. Dependencies - Reiationshipsiinfo

PARTe Relationship / Item How Obtain Relat'nshpNar'blty Info
ts = ts (tss, Lo, tIl,''') Need to define/approximate
tss =tss (Xl>'" ) Need to define/approximate
Xl From vendor & assembly function
jtPB =
jtPB ( Rh, k) Experimentally
Figure 2 A Detailed Level FOV - for a specific perfonnance characteristic.

Example is excerpted from a Registration (image-to-paperJ problem.
HIGHER LEVEL FOVs: One may construct higher level FOVs as needed. Though part
of a detailed FOV, doing so can simplify conceptual understanding. In such cases, a total
FOV-d can be treated as a flow of sub-FOVs. Examples: A copier's xerographic marking

module is a system of sequential subsystems. Given an image to be copied, these
subsystems operate on the original image, e.g., illumination, charging, development,
transfer, and fusing. The generic FOV applies to each of these as well as to the sequence.
Similarly for a printer, the imaging module embodied as a laser/scanner/optics design is a
sequential system that passes the light beam from the source to ultimately the photo-
receptor surface. Figure 3 depicts a sequential module. Combining the above examples for
a printer, the 'illumination' part for the copier is replaced by the printer's imaging module,
i.e., the latter's sequential module feeds into one element of the printer's sequential
marking module. Thus for evaluating a resulting Variability, there are sequential
subprocesses, there are multiple-to-one subprocesses, and there are the linkages of those. In
short, there is an articulatable hierarchy.
:0···..... • Here Input shown as deterministic (need'nt be) • SIS", subsystem charac.
:. .; • Solid means nominallmean of characterisic • Not shown: Charac. after each processing
... .... • Dotted means variability of charac. • Each SIS may need a detailed FOV
Figure 3. A Higher Level FOV; a sequential process FOV
COMPLEXITY/EASE: Complexity of the FOVs depends on the system and charac-

teristic being analyzed. One may only need partial FOVs, or none at all for a simple
characteristic for which there is a model or one has confident experimental variability
data. Even for a relatively complex problem, such as the one shown partially in Figure 2,
preparing the FOV is a simple process.
Evaluating Resulting Variability

It is clear from a detailed FOV that one can break a problem down to manageable segments
and that these segments flow from lower to higher hierarchical levels. Assume for now one
is able to establish relationships between entities of different levels. Then for segments for
which the random variables are independent sets, computation is easy with the HPD
Toolset; simply compute the resulting distribution for each segment, then roll up to the
next hierarchical level. In the Registration example, assuming that what is shown is all
the information needed, one would compute the distributions in the following order: (i) T,
(ii) IR and t2, (iii) t4, (iv) tIl. Next, the distributions for tss and Lo are computed, and then
the distribution for ts. The final step is dealing with the performance characteristic, E, ,for
this customer dissatisfier. E, is expressible in terms of tF and ts , thus its distribution is
computable. If a target Variability Range for E, had been specified, that is used to compare
with the computed one and a failure rate can also be computed; if not, the resulting
variability range will be used in some decision process, see §3.2. There is more on how to
use the Toolset's output information in Synthesis.
In actual practice, a 'model' sometimes spans a few hierarchical levels, thus one may
not need to go through the strictest of hierarchical computations. The FOV's clear depiction
of the hierarchy, however, serves to remind one to not forget effects of one on another and
on those up the chain. For example, for the Registration problem, the first analysis had not
examined the effect of clutch wear. That then had a holding place. Note that there is a
great deal of vendor and in-house experimental information that needs to be coordinated
with the Variability Analysis (shown schematically in Figures 4 & 5).
Think Naturally in Stochastic Realm

THINK NATURALLY MUCH BEYOND NORMAL DISTRIBUTIONS: The crux of the
capability to propagate variabilities is that since the relationships can be of any type, then
even if the lowest set of variabilities were given as Normal distributions, the resulting
distributions at any hierarchical level could be of any type; and these wouldn't have any
name associated with them as those we encounter in probability literature. Thus for most
of the propagation, one is handling any form of distributions; all that matters is that they
can be used for decision making or for computing the next level resulting variability.
Identifying to type is unnecessary, just as in the Deterministic realm, one does not require
identifying all functions to types to work with them.
VARIABILITY SPREAD ABOUT THE MEAN VS. TOLERANCES: In Design, the terms
'nominal & tolerances' are constantly used; the convenience is that one doesn't need to
specify any distribution. But suppose one did, as one in fact should. Then 'Variability
Spread' is more general than 'Tolerances' since it applies to any characteristic and type of
distribution, and refers to inputs and outputs. Using the general terms 'mean & Varia-
bility Spread about the mean' enables one to think more broadly.
FOV IS ACTUALLY 'FLOW OF DISTRIBUTIONS': In the FOV, the term 'Variances' was
used because it conveys a quantitative meaning to engineers (due to their familiarity with
Normal distributions) who can then envision a flow of Variances; 'Flow of Variabilities',
more appropriate, would have sounded abstract to them. If, however, one thinks of nomi-
nals & Tolerances or means & Spreads at the same time, one could easily transition to
thinking naturally of probability distributions. In so doing, 'Flow of Variances' becomes
'Flow of Distributions'! Indeed, that one could quantitatively flow the variabilities
required using the distribu-tions to compute the next level distribution. Thus the FOV
carries a much fuller and natural meaning in the Stochastic realm! Tolerancing is
Analysis & Synthesis in the Stochastic realm.
Analysis & Synthesis

TOP LEVEL VIEW: The total problem of designing against Variability is complex;
general practice has been at local levels. A top level holistic picture of Tolerance Analysis
& Synthesis is given by Figure 4. Figures 5 & 6 are to aid understanding the work process.
These show a broader scope of Synthesis techniques and requirements as well as the
beginnings of peeling the onion towards utilizing the specific techniques of HPD for
Synthesis.
Since we are operating in the Stochastic realm, Analysis here means Variability
lab Data Think both

Nominals & Tolerances!
Means &Var. Srreads;
Think DISTRIBUTIONS in
Cumpetency
.... t
Example:
~ An .• lysis: Knowx&y; FindZ

~ SynlMsis: Know Z, Find Rest (or
Adequate) Combo of X & y
Figure 4. Top level holistic picture of what it takes for Tolerancing
Analysis (VA), a term that will also be used to mean the Umbrella Area of Variability
Analysis which includes
• Sensitivity Analysis/Contribution Analysis
• Tolerance Analysis of geometric features, assemblies, or Critical Parameters
• Variability Analysis of Characteristics (e.g., time of arrival, charge density)
• Performance Analysis (e.g., failure rates)
Using Figure 4 to clarify Synthesis techniques is that one often refers to all of Synthesis as
Optimization without distinguishing between it and the strict sense of the latter. This
causes great confusion in techniques to use. In fact, much of the time, one is not able to
conduct direct Optimization; much of Synthesis is done by using Analysis (i) to aid making
adequate decisions, (ii) as part of the process towards enabling direct Optimization, or (iii)
to aid indirect Optimization. The foundation enabler, however, is Variability Analysis.
ANAL YSIS & "MODELS"; Quantitative Analysis of causes & effect always requires
knowing or approximating the relationship, or 'model'. For the Registration example, we
showed analyzing resulting variability was easy when relationships were assumed
known. In fact, practicing engineers often find it hard to 'model'; they readily go to
experimental techniques. But realize that datasets between experimental setpoints of the
causal variables and output responses in a Design of Experiment (DOE) comprise
relationships, thus are 'models' in the broad sense. The HPD Toolset can directly treat the
datasets as 'models'. The other two types of models are the typical ones of equations and of
algorithmic representations which include computer programs. To analyze in the stochas-
tic realm also requires input distributions (the link to supplier data, in-house experi-
mental data, or sub-FOV results). To analyze correctly in the stochastic realm requires
having used rigorous probabilistic techniques; this is an area that is not at the finger tips of
engineers. These are all reasons why the HPD Toolset was provided. Finally, competency
to 'model' stochastically and appropriately to enable true predictability is needed for
adequate VA. Modeling is an art and cannot be all taught; further, much of it depends on
advanced analysis competency coupled with skills in representing the phenomenon being
modeled. The HPD's FOV does give a structure to how to represent both the local and
global relationships without showing how to specifically model. The structure would be
helpful for designing DOEs, at some levels, for empirical models as well.
Points on models and stochastic models: (i) One should be aware that in DOEs, the
results are typically obtained from one hardware fixture for which, except for the
experimental setpoints and the input conditions that are varied, all else are or are con-
sidered fixed, Le., there is much variability not addressed! To accommodate those, one
might impose some variability on the fitted coefficients as well; or one could construct sub-
FOVs that relate noise that could affect the control parameters of the DOE. (ii) Many small
scale models are readily transformable into stochastic models since the deterministic
expressions can simply be viewed instead as functions of RVs. (iii) There are many
situations which can't be expressed deterministically and then simply extend those to
functions of RVs; examples include some types of self-adjusting systems. (iv) Some
situations do not lend themselves to empirical modeling or DOE techniques; an example is
where there are many input RVs with small variabilities that affect system performance.
The Registration problem was one such system level example; it had a large number of
RVs, and most with small variabilities that couldn't be controlled in experiments.
In supplying all relationships in the FOV for analyzing, there are many that may
appear too difficult within time constraint. Providing wise approximations for those
characteristics as distributions to conservatively cover those may be a temporary solution
to enable executing the next level VA now. Later, that 'Variability Envelope' needs to be
verified. All this is trackable by Part C of a detailed FOV.
Aside from providing the resulting distribution or failure rate as the analysis output,
the HPD Toolset's Contribution Analysis is especially useful for Tolerance Synthesis.
There are other techniques of using the Toolset's VA capabilities to aid Synthesis.
SYNTHESIS/DECISION MAKING: As shown in Figure 4, synthesizing in the Sto-
chastic realm means making decisions on nominals & Tolerances or means & Variability
Spreads. It ultimately must comprehend value to the corporation and to the customer
(recall the right-hand end of the FOV in Figure 1 and our brief explanation of the relation-
ship between Synthesis & Analysis. Contribution Analysis plays an important role in both
local levels of tolerancing decisions as well as enabling the selected higher contributors to
play in more global level decision making for tradeoffs. Figures 5 & 6 give a broader
picture yet. More discussion on how to use customer data from the field for global
Optimization is given in (Parks, 1992); that is related to Dr. Taguchi's Quality Loss
concept. Also discussed there is Stochastic Operating Window Optimization (shown there
is a 1-0 case; but the technique can be for any dimension); these Operating Windows
include having captured effects of variability and utilize the failure rate computation
technique ofHPD. Strictly speaking, that belongs in the category of indirect Optimization.
Indeed, there are very few situations where one truly can make the best decision by direct
Optimization. Such is usually only possible at a local level. Anytime when tradeoffs are
made, one is actually using indirect Optimization, thus actually yielding adequate, not
necessarily the best, decision. Note: DOE-based Optimization techniques are local types
and they include Taguchi Methods' Parameter Design.
When it is necessary, HPD does sometimes call for using DOE-based Optimization
techniques, such as Taguchi Methods. The latter is for Optimizing nominals against
Quantitative Synthesis I Decision Making Requires Facility in Comprehensively Dealing with Variability
Figure 5 Roadmap ofHPD Work Process
customer noise only (though Taguchi Methods calls for using manufacturing variability as
part of the outer array noise, as generally practiced, however, it is not implemented; part
of the reason is that it is too difficult to implement). The dataset from the DOE, however,
Piece Part
Tolerance I Cost
Verification;
Cpo Cpk
Failure Rate
Prediction
Optimization/Decision Making (O/OM):

Failure Rate-Based Stochastic Operating Window
Local to Global Variability Analysis-Based O/OM
r--------------------------------------------------------------------------------------,
: Understanding effects of variability enables myriad detailed design decisions including :
! - establishing hardware lat various levels), and :
: - finally simplifying decision making for the ~few" manageable critical parameters :
! HPO ties analysis (simple to complex) and experimental work content.
,, _______________________________________________________________________ !
... ______________ .l
Figure 6 Product Synthesis & HPD -- a simplistic view
can be used for processing other VA results with the HPD Toolset. Leading to the stage of
designing the DOE and the hardware is much VA work content. Having the DOE datasets,
it is then possible to analytically optimize for the nominals of the setpoint at which the
resulting variability against manufacturing noise is minimized - this is provided by our

second HPD Toolset; it is a direct Optimization technique which will appear as a later
publication. The primary HPD Toolset is focused on the VA capability, so that Synthesis is
enabled by using it for powerful WHAT-IF analysis and indirect Optimization, and
especially by coupling with the ability to correctly identify higher contributors that then
would allow (i) easily iterating for the required tolerances/Spreads to be specified, (ii)
reducing complexity of higher order direct Optimization, and (iii) simplifying higher level
tradeoffs. This Toolset is also the enabler for our direct Optimization technique.
One should realize that when highly fractionalized factorial DOEs, such as Orthogonal
Arrays (OA), are used, what information is one able and not able to obtain. When a result
from one experiment of an OA is used in all possible ways of averaging with the others,
there are consequences on the resulting information in the small and in the large. One
should then realize under what limitations would that level of fractionalization yield
correct results; we have done this (to be published later). Thus HPD only utilizes DOE-
based information and Optimization with that realization.
Other "Best Practices" incorporated in HPD; Some Clarifications: (i) Identifying cus-
tomer dissatisfiers (refer to Figure 5) can certainly come from having used QFD's first
house. But one needs to judiciously come from the engineering side as well, of starting
from the perspective of Soft & Hard failures (see (Parks, 1992)) of the system, and working
that down to a level that is perceivable by the customer. In this way, one can identify the
right level of performance characteristics to articulate FOVs for. (ii) Another Synthesis
activity called for is Manufacturability Analysis (see Figure 6) which includes the
comparison of the specified tolerances to Manufacturing's Process Capability, the CpkS.
Note that unless the required performance Variability, and Quality Loss, are quantifiable
from the specified tolerances, simply specifying those and having the suppliers deliver
certain levels of Cpk would not guarantee the expected performance level from the
hardware built. When one speaks of 6a Quality, it is moot unless we have the Variability
Propagation capability (more on this in a later publication). (iii) Databases of manufac-
turing process variability would be an aid to implementing HPD. Today, much focus is on
GD&T callouts and standardization to serve the needs of Tolerancing for Fit, but for
Functional Tolerancing, during Design we have very inadequate information on the
expected distributions of the features and properties of supplied parts! - some are far from
Normal. (iv) Since part ofHPD is to clarify existing practices that relate to Variability, let
us clarify the following; The term 'Robust Design' has often been considered by some as
synonymous with Taguchi Methods. In fact, that should not be. The term and its principle
are indeed Dr. Taguchi's. Corporations, however, have expanded on what is needed to
enable Robust Design because they found it requires more. Examples are AT&T and
Motorola, the latter's 6a Quality is part of their Robust Design program. HPD defines
Robust Design to be synonymous with Design for Latitude (DFL). That is because practices
to enable DFL require understanding comprehensively how to deal with Variability, thus
it requires HPD; that is why initially we had named HPD a Unified Methodology on DFL
(UMDFL). HPD is, in fact, a comprehensive Robust Design methodology.
On Implementation
The HPD has been presented to show the unifying total perspective. As stated earlier, it is
doable (see applications, §4) and is quite simple once focused on one's immediate objective.
Many are working at the local level, so that applying WHAT-IF analysis with the Toolset
would enable making decisions on Tolerances/Spreads, so long as cost allocations are met.
Today, one builds a safety factor with the CpkS, but that is related to cost; further, as
already stated, attaining certain levels of Cpk values does not guarantee adequacy if no
appropriate Variability Analysis had been executed. Much improvement can be made on
prediction of cost and performance and total Optimization via HPD.
For the local level of analyzing a customer perceivable performance characteristic,
however, typically today the engineer is not extending one level down to the piece part
characteristic level and flowing the lower level result (from a sub-FOV for a Critical
Parameter, say) to the higher computationally. An example is that for the parts and
assembly that enable providing a normal force (a Critical Parameter) for a subsystem; this
force flows into the FOV for the subsystem characteristic. Note that this relates physical
variability to functional variability. Indeed, to tolerance parts, one must relate a perfor-
mance characteristic down to the part level. Segmenting and the HPD work process
simplifies large scale problems.
Next, we will see why the Toolset makes it easy for non-experts to apply probabilistic
analysis as a part of the natural work process and thus implement the Methodology.
3 THE HPD TOOLSET
3.1. Introduction
Relation to Companion Paper
A keynote address on HPD (Parks, 1995) was delivered at the 3rd International Confer-
ence on Stochastic Structural Dynamics in January, 1995. The paper is being written
almost concurrently with this one. Since the interest of that audience is more on the
analysis content of the Toolset, we decided that the two papers would form a perfect
companion set, with this paper focusing on clarifying the Methodology, relevance to other
practices, and implementation. Therefore, this section on the Toolset is in synoptic form.
Improvement on Earlier Concept of Capabilities Needed

Since we have referred the reader to our earlier work (Parks, 1992) in several areas
because there some concepts were more fully explained, to avoid causing confusion, we
need to point out that the concepts for the capabilities and the features of the Toolset
expanded significantly since that was published, e.g., 'completeness' and user-friendliness.
Overview
(i) The capabilities of the HPD Toolset were driven by the needs of the Methodology. (ii)
Developing user-friendliness was a primary requirement as well. (iii) The Toolset is based
on rigorous techniques. In the process, we have advanced Stochastic methods significantly
beyond what exists and have been able to show inadequacies of prevailing techniques that
can lead to gross errors (see Figure 7); some warnings on those are given in §3.3.
Combining (O-(iiO led to providing a 'complete' suite of tools. (iv) A very simple work
process of how to use the Toolset is shown.
3.2. Capabilities of the Toolset

For the Umbrella Area of Variability Analysis
Let ~ denote a customer-perceivable characteristic, and Z denote any other type of output
characteristic in the FOV.
• TOLERANCE ANALYSIS I VARIABILITY ANAL YSIS FOR z: The Toolset computes the
distribution, Fz , the mean, Ji-z, and the 30 equivalent bounds of Fz , the latter are labeled as
Zmin and Zmax, (these points bound the Range of Variability and are points corresponding to
Fz =.00135 and Fz =.99865, these being associated with the typical notion of 3u Tolerances); the
density distribution, fz , is also computed. Zmin and Zmax thus are the TolerancelVaria-
bility bounds of Z . Other distribution attributes are computed as well. This is followed by
a Sensitivity/Contribution Analysis.
• PERFORMANCE ANALYSIS FOR {: The Toolset computes the failure rate of { based on
its set offailure criteria. One may treat ~ like Z in that one may wish instead to compute
its distribution without utilizing the failure criteria.
• SENSITIVITY I CONTRIBUTION ANAL YSIS: One may choose to run this first to reduce
the problem and then run a TA / VA. This is the preferred process.
For Implementing HPD

• ACCEPTS ANY FORM OF INPUT DISTRIBUTIONS: This is crucial, especially to enable
next capability.
• VARIABILITY PROPAGATION: The output density distributions from any level of the
computation in the FOV are used as (a partial set 00 input distributions for the compu-
tation in the next level FOV. These input distributions are of any type, of course.
• ACCEPTS ANY 'MODEL': The model can be represented by equations, algorithms/
computer programs, or data sets (e.g., from DOE or ... ); these, in fact, encompass all of an
engineer's types of relationship information.
• ACCURACY CHOICE: Several analysis techniques are offered to enable accuracy/run-
time tradeoffs, or accuracy selection due to knowing certain types of functions. Realize
that some engineering problems require high precision.
• SPRINGBOARD for OPTIMIZATION TECHNIQUES: This had been briefly discussed
earlier; the Toolset is the foundation enabler.
3.3 Techniques Incorporated; Rigor; Warnings

Techniques incorporated are: Our 8MB technique (see (Parks, 1995)), Monte Carlo,
Pearson System fit, Linearization but allowing non-Normal outputs, R88, and combina-
tions of any of these. For example, every tool element, based on any of these techniques,
includes conducting Contribution Analysis that is based on the 5MB technique and thus
takes into account input distributions. Also, determination of whether the problem is
linearizable is part of the output from some of the relevant tools elements.
It is the 8MB technique that has enabled determining adequacy or not of other
techniques; we have been able to show shortfalls with, e.g., Monte Carlo, Pearson System
fits, Linearization, and prevailing Contribution Analysis techniques. With respect to the
latter two, Figure 7 shows the unexpected gross errors that could arise, and one wouldn't
realize it if one employed those techniques! Thus any tool or technique that utilizes 'simple'
Linearization as part of its technique has this major unsuspecting problem ('simple' here
means not properly justifieci). Similarly, if one were to depend on any prevalent Contribu-
tion Analysis tool or technique to simplify or narrow the problem down, one would also be
on dangerous ground. Note: There are situations where it is not possible to attribute %
: .................................................................... ": : ...................................................................:

Example 1 Example 2
z=x2 Z=(X-Y)(W-V)
x: U (0.7, 1.3) x, y: N (10, 0.5); W: N (10, 1); v: N (9.5,1)
Monte Carlo
. -.I._! f I Pearson System

[fused Linearization,
r reduced problem with
~
j fit of an MCrun
l.---_"°_'°_,°+-;_=-_ (2000 Sims)
Ordinary Contribution Analysis
'0.0·,;°:"·°';'1
"
,,,"" ,,,
AnotherMC "
run gave quite
i a different fit.
It was very
,,
,'
highly peaked
on the left side. ,,
, \....- 5MB
I '
,w.
5MB
An Me run w. PS fit :
one Aggregation ;
Interval size** :
: ....................................................................,; : ...................................................................:
Figure 7 Very simple examples - shows problems with Prevailing Techniques

•• Histograms of some aggregation sizes were closer to 5MB distribution, but fits differed little from the one
shown. More on these in Reference 5.
contribution to individual RVs though one can identify their importance; the example in
Figure 7 is one such! These and other unsuspecting types of shortfalls are discussed in the
companion paper (Parks, 1995). Other mathematically rigorous techniques besides sto-
chastic techniques were incorporated into the Toolset as well.
3.4 User-Friendliness; "Completeness"

The User Interface is simple and unifies all of the offerings in a user-friendly manner.
Some users required no instructions. Incorporated into the codes are mathematical
techniques that enabled the 8MB tool to be user-friendly, but this did require more
computational time for the non-typical case of large numbers of RVs for which the 'model'
cannot be segmented or narrowed by Contribution Analysis, or handled by our pre-
processing techniques. Thus other stochastic analysis techniques were incorporated as
possible alternatives (though in our applications, we haven't needed these). Also included
as part of user-friendliness are the usual amenities of commercial toolsets.
That the Toolset is described as 'complete' is due to having anticipated every need the
user might encounter and provided the best rigorous techniques available and any lesser
ones as necessary alternatives but for which the output gives information on how appli-
cable those are. We also prescribe work-around processes and supplementary interlink-
able tools for seemingly impossible situations. Rigor to enable correct analysis contributed
to 'completeness'. Of course, nothing is truly complete, that is why the quote marks.
3.5 The Toolset & HPD Work Process

Much has been briefly explained already. Here let us just look at two small aspects of using
the Toolset in the simplest way in the work process. (i) At some local level of the FOV, if
there are a small # of input RVs, conduct VA, iterate on tolerances of higher contributors
till resulting Variability meets requirement. If the # of RV s is large and the model is not
easily segmentable, conduct Contribution Analysis first, select the higher contributors
and conduct VA. Iterate on Tolerances of these till output requirement is met; adjustment
is made for the very small contributors omitted. In some cases, one should re-run
Contribution Analysis for the full set to obtain the re-ordered set of higher contributors for
next level tradeoff purposes (since tightening tolerances for some would change the
relative weights of the total set). (ii) Again, at some local level, if there are some input RVs
for which the distributions are yet unknown, such as those to be obtained from vendors,
experiments, or a lower level VA, then wise/conservative approximations (Variability
Envelopes) can be used for now. If any turns out to be a high contributor and the resulting
variability is not meeting requirement, one would need to do further work to determine
that input distribution. Else one could go onward; the wise approximation was adequate,
but one should keep track of all such information in the detailed level FOV. Bear in mind
that in all analytical work, some form of verification testing should be conducted, but this
is far simpler than having to resort to all experimental techniques for Tolerancing.
4 APPLICATIONS
HPD has been implemented, but not yet broadly; the Toolset's full completion is relatively
recent. Without giving any detail, nor exposing proprietariness of actual applications, the
brief content here is simply to show doability, broad applicability, and value.
The applications have ranged from the very simple to the very complex. Simple ones
include electro-mechanical component selection and simple geometric parts tolerancing
coupled with assembly modifications to enable meeting required Ranges of Variabilities of
Critical Parameters. Recent very complex applications include:
• The Registration problem -- The prediction, far smaller system registration error than
had been feared with simple R88, was verified months later when the (very large)
system build was completed and system testing conducted. There were no usable sub-
models. Our analysis had enabled the product program's determination of latitude and
tolerancing adequacy without needing to use more costly components.
• A post-fusing Curl problem -- A perennially impossible problem attempted for over two
decades; involved an 'integrative' modeling process that linked diverse types oflab data,
analytical models, and innovative approximations. The VA required a few levels of
segmented computation. The results enabled product program's decision to include and
develop a critical subsystem, at no small labor and unit costs, to alleviate the problem.
• A sensitive Motion Quality/Drives System Design problem -- Involved using a very time
intensive computer program as a model, and also a DOE-based model.
Earlier, much failure rate prediction analysis and Stochastic Operating Window
Optimization had been applied to several product programs (Parks, 1980).
In actual applications, there have been cases where results from our correct Contri-
bution Analysis have contrasted with incorrect results from prevalent techniques used,
and thus have enabled correct solutions.
5. CONCLUDING REMARKS
The HPD and its associated Toolset comprise the Cart & Horse for Functional Tolerancing.
Providing the Stochastic Analysis capability as a user-friendly toolset, based on advanced
and rigorous techniques, and covering almost all needs, is a significant step towards
removing a fundamental level barrier in Design Synthesis, for this enables probabilstic
analysis be part of the natural work process. The 'exact' computation of distributions and
correct Contribution Analysis capabilities are fundamental breakthroughs in computa-
tional probability. Those, coupled with mix & match of all other capabilities incorporated,
are extremely powerful and have enabled lifting the veil on unsuspecting inadequacies of
prevalent techniques. The nature of Tolerancing for Fit and Tolerancing for Function
work content differs, the latter requires HPD to better structure and represent the
relationships needed for Tolerancing, with techniques for segmenting large scale problems
for simpler local and subsequent higher level treatments and with a process for user-
involved quantitative decision making on the variabilities.
One last effort at putting things in context and relating concepts. We propose the
following quantitative definition for Latitude: Latitude is given by the pair,
where fz is the resulting density distribution and the set of f mS ,where m = 1,2, ... , M,
is the set of input distributions. This can be interpreted at any local level of an FOV, or at
the total system level! Thus when one says that a subsystem has Latitude, or is Robust,
one knows exactly how much input variability from what sources can it sustain and what
is the associated quantified variability of its performance characteristic. One could also
loosely express the pair in terms of the associated Variability Spreads of the distributions.
Note that the proposed definition means Latitude represents the quantified FO V.
ACKNOWLEDGEMENTS:
The author owes gratitude to Ray Ashton, Maurice Holmes, Charlie Sie, and Vijay Tipnis
for valuable discussions over the years on technology and product development, design
practices and methodologies, and variability. Special gratitude is owed Chun Li for his
software development expertise and John Browne for many discussions on probability that
have been of mutual benefit.
CREDITS:
The Toolset development effort has been an intensive joint effort of Chun Li, Chi-Cheng
Chen, and the author.
REFERENCES
Guilford, J. and Turner, J. (1993) Advanced Tolerance Analysis & Synthesis for Geometric
Tolerances, in Proceedings of the 1993 International Forum on Dimensional Toler-
ancing and Metrology (ed. V. Srivinasen and H. B. Voelker). ASME CRTD-27, NY.
Parks, J. M. (1980-1982) Miscellaneous Xerox Internal Technical Reports and Memos.
Parks, J. M. and Bailey, J. M. (1990) Consolidated Findings from the 'Application Transfer
Study' (an intensive internal Xerox study on CAE needs and supporting enablers).
Parks, J. M. (1991) Unified Methodology on Design for Latitude and the Role of Stochastic
Modeling, in Proceedings of the Design Productivity International Conference (ed. K.
Ragsdell and T. Holt). Hawaii.
Parks, J. M. (1992) The Unified Methodology on Design for Latitude as A Concept and
Capability for Tolerance Analysis & Synthesis, in Selected Case Studies in the Use of
Tolerance and Deviation Information During Design of Representative Industrial
Products (ed. V. Tipnis). ASME CRTD-15-1, New York.
Parks, J. M. (1995) THE Two Enablers for Holistic Probabilistic Design: Methodology and
A Complete Suite of Tools for Stochastic Analysis, in Proceedings of the 3rd Interna-
tional Conference on Stochastic Structural Dynamics (to be published in 3Q95), Puerto
Rico.
Tipnis, V. A., (1989) (chairman) Research Needs and Technological Opportunities in
Mechanical Tolerancing -- Results of An International Workshop, ASME CRTD-15,
New York.
Voelker, H. and Cutler, N. (1993) Discussion session from the 1993 International Forum
on Dimensional Tolerancing and Metrology, in Proceedings (ibid, Guilford).
BIOGRAPHY
The author is a Research Fellow at Xerox Corporation. She received her M.S. from
Stanford University and her Ph.D. in mathematics from the University of Rochester in
1979. She has had 22 years of industrial experience: 4 years at Lockheed Missiles & Space,
2 years at SRI International, and 16 years at Xerox Corporation. Having modeled exten-
sively, managed large organizations in technology development and design methodo-
logies, and advanced probabilistic techniques, in 1990, she saw a great need to develop and
provide the HPD Methodology and Toolset to enable effective technology and product
development.
PART SIX
Computational Metrology
18
Data processing method for
geometrical forms with form
deviations in coordinate metrology
Kiyoshi Takamasu*, Iturou Fuk'uda*, Ryousyu Furutami , Junhee Hong+
and Shigeo Ozono'
• The University of Tokyo, t Tokyo Denki University, :j: Chungnam
University
• Department of Precision Machinery Engineering,
Graduate School of Engineering, The University of Tokyo,
Hongo 7-3-1, Bunkyo-ku, Tokyo 113, Japan.
Telephone: +81-3-3812-2111. Fax: +81-3-3812-8849.
email: takamasu@pe.u-tokyo.ac.jp
Abstract
In coordinate metrology, extracted features (Gaussian substitute features) are normally
calculated from measured data sets of CMM (Coordinate Measuring Machine). Then,
the extracted features are compared with the nominal features which are indicated on
the drawings. The extracted features are calculated using a least squares method and
expressed without form deviations. The novel data processing method for geometrical
forms with form deviations has heen developed. In the Inethod, every geometrical form
can be processed with its form deviation which is calculated from the measured data set
using the Gaussian (least squares) method as a standard variations <7. Using the method,
we can calculate the form deviations of 2-D geometrical features. This directly implies
that these values and these calculations can be used for the evaluations of measurement
uncertainties and a computational tolerancing.
Keywords
Coordinate metrology, coordinate measuring machine, geometrical form, form deviation,
Gaussian substitute feature, extracted feature
1 INTRODUCTION
In coordinate metrology, extracted features (Gaussian substitute features) are normally
calculated from measured data sets of CMM (Coordinate Measuring Machine) (ISO TR
10360-1, 1995). Then, the extracted features are compared with the nominal features
which are indicated on the drawings (see Figure 1) (ISO 10360-2,1994). We should note
274 Part six Computational metrology
Drawings
(Nominal features)
Figure 1 Data processing in coordinate metrology.
that the extracted features are calculated using a least squares method and expressed
without form deviations. This is mainly because the accuracy of CMM is too low to
measure the form deviations. However, the accuracy of CMM is increasing rapidly and
the data accusing time of CMM is decreasing (Takamasu, 1985). Therefore, we can process
form deviations, such as flatness, cylindricity, straightness and so on, positional deviation,
and surface roughness in coordinate metrology (Furutani, 1988).
Thus, a novel data processing method for geometrical form with form deviations has
been developed. In the method, every geometrical form can be processed with its form de-
viation which is calculated from the measured data set using the Gaussian (least squares)
method as standard variations 0'.
Generally, machine parts have three dimensional surfaces such as flat planes, cylindrical
planes, conical planes, spherical planes and toroidal planes which can be measured directly
by CMM. However, 2-D intersections of the 3-D surfaces such as an intersection line,
an intersection point, an ellipse and a circle can not be measured directly by CMM
(Takamasu, 1984). Therefore, these 2-D geometrical features are calculated from the 3-D
substitute features. In these calculations, we proposed a novel data processing method
based on statistical calculations using the relationship between the form deviations of
3-D geometrical features and these of 2-D geometrical features. Using the method, we
can obtain the form deviations of 2-D geometrical features (Henzold, 1994 and Waldele,
1995). This directly implies that these values and these calculations can be used for the
evaluations of measurement uncertainties and a computational tolerancing.
2 EXTRACTED FEATURE
Least squares features are normally used as the extracted features without form deviations.
We should emphasized that some other types of data structures can be considered to use as
the extracted features (ISO 1101, 1983, ISO 5459, 1981). Let us show four data structures
in Figures 2 (a), (b), (c) and (d). Figure 2 (a) shows a least squares plane without a
form deviation, and Figure 2 (b) shows that with a form deviation 0' which indicates
the geometrical form. This data type is the simplest way of using form deviations in
Data processing method for geometrical forms 275
(a) Flat plane (b) Flat plane with thickness
(c) Low frequency surface with thickness (d) Free surface
Figure 2 Expressions of the extracted feature of flat plane.
Table 1 Expression of extracted features

Extracted Location Orientation Form devia.tion Expression
feature
Plane Positional vector Normal vector Flatness Plane with thickness
(see Figure 3 (a))
Cylinder Positional vector Axis vector Cylindricity Cylinder with thickness
(see Figure 3 (b))
coordinate metrology. Figure 2 (c) and (d) indicate the other data structures with low
and high frequency ranges of forms.
From these data types, we choose the least squares form with the single value of the
form deviation (Figure 2 (b)) as the extracted features. The reason is that this data
structure is the simplest and easy to start of data processing with the form deviations in
coordinate metrology.
Figure 3 displays examples of expressions of 3-D extracted features, a flat plane and
a cylindrical plane with form deviations. The expression of the flat plane (Figure 3 (a))
consists of the vectorial expression of the Gaussian (least squares) plane, the location
vector L and the normal vector 0, and the form deviation value (flatness) (J which is
the standard deviation from the Gaussian plane. The expression of the cylinder (Figure
3 (b)) also consists of the location vector L, the axis vector 0, the radius r and the form
deviation (cylindricity) (J.
We conclude that every feature can be handled as the least squares feature and the
form deviations in Table 1.
o Cvlinder
" (ieas! squares)
~.--"""'."..- • ~ Actual feature
Measured data
(a) Expression of plane (b) Expression of cylinder
Figure 3 Expressions of geometrical forms with geometrical deviations.
3 INTERSECTION LINE OF TWO PLANES
3.1 Distribution function of point from plane

Now we consider the intersection line of two planes as one typical example of 2- D substitute
feature. The distance d1 of the length between the intersection line L and the least squares
plane M 1 , and the distribution of Dl (dd are defined. Hereafter Dl (dd is assumed as
the Gaussian distribution in equation (1). However, if the distribution function is an
even function, the relationships of integration (equations (2) and (3)) can be used in the
following analyses.
(1)
(2)
(3)
3.2 Distribution of intersection line

Figure 4 (a) shows the calculation concept of relationship between the intersection line
L and two planes Ml and M2 which cross at the angle Ci (the angle between two normal
vectors of planes). The least squares position of the extracted line L is easily calculated
as the intersection of planes Ml and M2. Then, the distribution of the intersection line
is also calculated from the distributions Dl(dd and D2 (d 2 ) of distances from two planes.
When the intersection line is positioned at (x, y) and the distance from the center of the
y
rr/2- cr.
(a) Calculation of intersection line L (b) Position (x,y) of intersection line
Figure 4 Relationship between planes and intersection line.
intersection line is the length s (Figure 4 (b)), the relationship between db d2 , x and y
is shown on equation (4). Therefore, the probability of position of the intersection line
D(x,y) is calculated in equation (.S) .
d2 -x sin a + y cos a (4)

D(x,y) DJ(x)· D 2(-J'sina + ycosa)
1 (y2
- - - exp - - - -
(-XSina+ycoso<)2)
(5)
21rUl U2 2U12 2U22
Figure 5 displays the 3-D image of the distribution D(J',y), when Ul = 1.0, U2 = 1.5,
a = 75° and Figure 6 indicates contour maps of the distribution when Ul = 1.0, U2 = 1.5,
a = 75° (Figure 6 (a)) and a = 30° (Figure 6 (b)). These distributions may be expressed
as an elliptic cylinders. However, we choose the simplest data type until now, the elliptic
cylinder is expressed as the geometrically exact cylinder.
3.3 Standard deviation of intersection line

Let us consider the radius of the geometrical exact cylinder which expresses the line with
its straightness. We assumed that the radius of the cylinder is defined as the standard
deviation of the position from the least squares line. The distribution of the intersection
line L can be defined by the intersection angle a, the distribution functions Dl (dIl, D 2 ( d2 )
of planes Ml and M 2 .
From Figure 4 (b), the relationship between s, d 1 , d2 , and a is derived as equation (6).
By the integration of equation (6), the average value of s which is the form deviation of
the line is calculated in equation (7). In other ward, we can obtain the form deviation of
0.1
D
o
Figure 5 3-D image of the distribution D(J:,y) (0"1 = 1.0, 0"2 = 1.5, 0' = 75°).
±2cr, ±2cr,
y
-6 o 6
x
(a) 0' = 75° (b) 0' = 30°
Figure 6 Contour maps of distribution (0"1 = 1.0, 0"2 = 1.5).
the intersection line in equation (8).
(6)
0" 2
s (7)
VO"1 2 + 0"2 2 (8)
\ sin 0'\
Figure 7 Intersection point of three planes ..
4 INTERSECTION POINT OF THREE PLANES

The position of the intersection point of three planes is also calculated geometrically
(Figure 7), and the distribution r from the center is shown in equation (9). Where the
displacements d l , d 2 and d:l are distances from planes M l , M2 and M3 , respectively, and
the angle a, 13 and, are the angles of the normal vectors of two planes between planes
Ml and M 2, Ml and M 3 , and M2 and M 3 , respectively. From equation (9), the standard
deviation a,. is also derived in equations (10) and (11).
1 - cos 2 a - cos 2 13 - cos 2 , + 2 cos a cos 13 cos,

2d l d2 ( cos 13 cos, - cos a) + 2d l d3 ( cos a cos, - cos (3)
(9)
+~~--7-~~~~~~~----~--~~
2d 2 d3 ( cos 13 cos, - cos a)

+ 1 - cos 2 a - cos 2 13 - cos 2 , + 2 cos a cos 13 cos,
al 2 sin 2 , + a2 2 sin 2 13 + a3 2 sin 2 a
a,. 2 (10)
al 2 sin 2 , + a2 2 sin 2 13 + a3 2 sin 2 a
(11 )
1 - cos 2 a - cos 2 13 - cos 2 , + 2 cos a cos 13 cos ,
If the relationship fJ = 1r /2, , = 1r /2 and a3 = 0 are assumed, equation (8) is derived

from equation (11).
5 EXAMPLE OF CALCULATION
Now we can get the calculation method to evaluate the form deviations of lines and points
in equations (8) and (11). Let us show an example of these calculations. Figure 8 indicates
an example of a polygonal work which has 6 planes (Ml ~ M6)' The normal vectors, the
Table 2 Position, direction and flatness of planes (Ml ~ M 6)

Name of plane Normal vector Positional vector (mm) Flatness (mm)
Ml (-0.01,-0.99, 0.12) ( 0.00, 0.00, 0.00) 0.014
M2 ( 0.90, 0.11, 0.43) (49.85, -0.18, 0.22) 0.029
M3 ( 0.02, 0.09, 1.00) ( 0.23, 0.22, 30.48) 0.019
M4 ( 0.01, 0.92, 0.39) ( 0.01, 40.23, 0.22) 0.043
Ms (-0.99, 0.12,-0.08) ( 0.00, 0.00, 0.00) 0.014
M6 ( 0.01,-0.01,-1.00) ( 0.00, 0.00, 0.00) 0.012
Table 3 Start and end points and straightness of lines (Ll ~ L 12 )

Name of line Start point and end point Straightness (mm)
Ll PI, P4 0.047
L2 PI, P2 0.080
L3 P2, P3 0.114
L4 P3, P4 0.047
Ls P4, Ps 0.051
L6 PI, Ps 0.082
L7 P2, P6 0.138
Ls P3, P7 0.115
L9 Ps , Ps 0.060
LID Ps , P6 0.099
L11 P6, P7 0.135
L12 P7 , Ps 0.060
positions and the deviations (flatness) of each plane are measured by CMM in Table 2.
From Table 2, the positions of intersection lines and their deviations (straightness) can
be calculated in Table 3, and the positions of intersection points and their deviations
(positional deviation) are calculated in Table 4.
Figure 9 (a) displays the intersection lines with the straightness (x 20), and Figure 9
(b) displays the intersection points with the positional deviations (x 20).
Table 4 Position and positional deviation of points (PI ~ Ps )

Name of points Positional vector (mm) Positional deviation (mm)
PI (49.747, -0.442, 0.502) 0.088
P2 (45.042, 39.812, 0.052) 0.139
P3 ( 4.928, 40.420, -0.355) 0.119
P4 ( 0.000, 0.000, 0.000) 0.059
Ps (35.442, 3.218, 29.506) 0.104
P6 (33.429, 28.396, 27.280) 0.164
P7 ( 1.196, 28.476, 27.918) 0.141
Ps (-1. 995, 3.682, 30.213) 0.070
Figure 8 Polygonal work of 6 planes (Ml ~ M2)'
P,
P,
(a) Lines with straightness (b) Points with positional deviations
Figure 9 Intersection lines and points with form deviations.
6 CONCLUSIONS
We developed the novel data processing method to calculate the form deviations of the
intersection lines and points. We have reached the following conclusions.
• Some types of data structures for the extracted feature in coordinate metrology are
displayed.
• The simplest data structure which has the form deviation consists with the least squares
feat ure and the form deviations (one value).
• The novel calculation methods which derive the form deviation of 2-D features from
the deviations and the geometrical relationship of 3-D features are obtained .
• The example of these calculations has been displayed .
REFERENCES
Furutani, R. et al. (1988) Estimation method for three dimensional forms, J. Soc. Precision
Eng., 54, 890-895.
Henzold, J. (1994) Vectorial dimensioning and tolerancing, ISO JHG/TG 3 Nl Draft
proposal.
ISO 1101 (1983) Technical drawings - Geometrical tolerancing - Tolerancing of form, orien-
tation, location and run-out-Generalities, definitions, symbols, indications on drawings.
ISO 5459 (1981) Technical drawings - Geometrical tolerancing - Datums and datum sys-
tems for geometrical tolerances.
ISO TR 10 360-1 (1995) Co-ordinate metrology - Part 1: Definitions and applications of
the fundamental geometric principles.
ISO 10 360-2 (1994) Co-ordinate metrology - Part 2: Performance assessment of co-
ordinate measuring machines (CMM).
Takamasu, K. and Ozono, S. (1984) Graphical analysis of space coordinate data to deter-
mine the spatial position and form of a three-dimensional object, Precision machinery
medical eng. mechaoptoeiectronics, 1, 150-155.
Takamasu, K. and Ozono, S (1985) Determination of datum plane by least squares
method, J. Soc. Precision Eng., 51, 563-568.
Waldele (1995) Comparison of vectorial tolerancing and conventional tolerancing, ISO
JHG/TG 3 N2.
19
Algorithmic Circularity Measurement
For Fringe Analysis and Sub-Micron
Posi tion Sensing.
J. Pegna, Assistant Professor
C.Guo, Doctoral Student
Department of Mechanical Engineering, Aeronautical Engineering
and Mechanics, Rensselaer Polytechnic Institute, Troy, New York
12180-3590, USA, Tel:(518) 276-6030, Fax: (518) 276-2623,
emails: pegnaj@rpi.edu, guoc@rpi.edu
T. P. Hilaire, Research Engineer

Renault Corporate R&D, 9-11 Avenue du 18 Juin 1940, 92500 Rueil
Malmaison, France, Tel.: (1) 47 7793 42, Fax: (1) 47 77 92 04, email:
hilait@ipm.dmi.rpi.edu
Abstract
Circle (or sphere) and circularity (respectively sphericity) measurements are a com-
mon occurrence in many fields of physics and engineering. In metrology, circles and
spheres are often representative features of points to be measured. This paper pro-
ceeds from the need for a fast circularity and two-dimensional center position mea-
surement. Central to this process is a circle fitting algorithm in the sense of least L-
infinity norm, also known as Chebishev or MinMax fit. The problem at hand can be
formulated as follows: Given a set of points in the plane, find the pair of concentric
circles with minimum radial gap enclosing all the points. Applications of this algo-
rithm to fringe pattern analysis for alignment and sub-micron position sensing will
also be presented.
Keywords
Computational Metrology, Computational Geometry, Metrology, Soft Gauging, Image
Metrology, Fringe Pattern Analysis, Interferometry, Precision Engineering.
1 INTRODUCTION
The circle and the sphere are routinely used as surface features to characterize a point
in many fields of engineering and physics. In metrology applications, the most com-
mon measurement on spheres and circles are center position and sphericity (circular-
ity for problems in the plane.) Those specific measurements are central to the work
presented in this paper, which stems from experimental developments in the design
of precision machinery. A 2-D position sensor based on a modified Michaelson inter-
ferometer was developed and demonstrated an accuracy of 80nm transverse displace-
ment over a depth of field of 1 meter. Calibration experiments revealed that
parallelism error between the CCD plane and interferometer's reference mirror was
critical to accuracy. It also became apparent that this orientation error could be easily
measured and self-calibrated by a circularity measurement on the fringes.
This paper reports on a fast algorithm used to find the center of circular fringe pat-
terns and measure their circularity. Center position measurements are used to track
translation errors in machine carriages. Circularity measurements are used to control
the position of the reference mirror by means of piezo-electric actuators.
Sections 2 through 6 present a marching algorithm to find the pair of concentric cir-
cle enclosing a data point set with minimal radial gap. Convergence properties, run
time and accuracy are discussed in section 7. Application of this algorithm to fringe
pattern analysis and design of a sub-micron position sensor are discussed in section 8,
where specific advantages of the proposed algorithm as compared to previous
approaches are also discussed.
2 COMPUTATIONAL METROLOGY OF THE CIRCLE
Circular fit is an old problem that has stimulated a large body of literature, often
duplicated among the fields of metrology, computer vision and computational geom-
etry. The dominant share of previous works uses the least square fit method or a mod-
ified least square fit method to find the center and radius of the fitting circle. Less
common are techniques to find the best fitting circle in the sense of L-infinity or Che-
byshef norm.
Landau [Lan 87] introduced an iterative algorithm for estimating the center and
radius of a least square fit circle and the solution may lead to a good approximation of
the Chebyshev circular fit. Thomas and Chan [Tho 89] proposed a modified mean
square error function to obtain a simple solution. The error function was defined as
area difference instead of the normal distance to the circular fit. Takiyama and Ono
[Tak 89] extended the problem to finding a common center for multiple arcs and their
respective radii.
Algorithmic circularity measurement 285
Lai and Wang [Lai 88] introduced the first computational geometry-based method
to find the Chebyshev circular fit to a point set. They proposed using the intersections
of the farthest Voronoi diagram and the medial axes from a points set as the center for
the pair of concentric circles. Even though the algorithm converges, it doesn't guaran-
tee that the radial gap found is minimum. To enclose all the points between a pair of
concentric circles with the minimum radial gap, Roy and Zhang [Roy 92] proposed an
exhaustive search method by comparing all the gaps with centers located at the verti-
ces of nearest Voronoi diagram, farthest Voronoi diagram, and intersections of the two
diagrams. This exhaustive method will give the location of the center and radii with
the minimum gap but it also requires computation time that are incompatible with the
needs of real-time processing (see Table 1).
3 PRELIMINARY DEFINmON
Nearest Voronoi diagram (NVD): Let S denote a set of n points (called sites) in the
plane, the nearest point Voronoi region associated with site Pi is the set of all the
points that are closer to Pi than to any other sites. This region is always convex [For
87]. The intersection of the nearest Voronoi region of all the sites of S is the nearest
Voronoi diagram.
Farthest Voronoi diagram (FVD): Let S denote a set of n points in the plane, the
farthest Voronoi region associated with site Pi is the set of all the points that are farther
from Pi than from any other sites. The intersection of the farthest Voronoi region of all
the sites of S is the farthest Voronoi diagram.
Voronoi edge: A Voronoi edge is a line segment bisecting exactly two sites of a
Voronoi diagram, each point on a Voronoi edge is equidistant from these two sites.
Voronoi vertex: A Voronoi vertex is the common intersection of the Voronoi edges of
the diagram. Each vertex is equidistant from at least three sites.
For a given set of points, any center of a minimally constrained pair of concentric
circles enclosing all the points will be such that at least (;me point lies on the inner cir-
cle and at least one point lies on the outer circle. If there are exactly two points on the
outer (respectively inner) circle then the center of the two concentric circles is located
on an edge of FVD (respectively NVD.) Note that in case the center is at an intersec-
tion of a nearest Voronoi edge and a farthest Voronoi edge, then the two circles are
fully constrained. If there are three points on the outer (respectively inner) circle, the
center is located on a vertex of FVD (respectively NVD). Similarly, if there are three
points on the inner circle, the center is located on a vertex of NVD. Hence the two con-
centric circles are fully constrained if their center is on a vertex of NVD, a vertex of
FVD, or an intersection of a FVD edge and a NVD edge.
4 THE 3-POINTS CIRCULAR ENCLOSURE PROBLEM
In this section we will discuss the variation of the gap between the two concentric cir-
cles when the two circles are constrained by three points. Note that two points must
be located on one circle and that the last point lies on the other circle.
First let us define our reference frame: Assume PI, P2 are located on the same circle
and P3 lies on the other circle as shown in Figure 1. The X axis is defined as the bisec-
tor of PI and P2' If PI and P2 are on the outer circle, the X axis will be an edge of FVD;
the X axis will be an edge of NVD. Conversely, if PI and P2 are on the inner circle then
the center must be located on an edge of the FVD. The origin of the X axis is taken as
the center of the circle which passes through P v P2, and P3. The positive orientation of
the X axis is defined such that PI and P 2 are on the exterior circle when the center
moves along the X axis. Hence this provides us with the reference frame (see Figure 1)
for further analysis.
The location of center 0 can be found from the following two equations,
plo e P I P 3 = OP3 e P I P 3 (1)
P 2 0eP2 P I = OPl eP2 P I , (2)
and the unit vector of the X axis can be found as:

P 2 P o+ PIPo
(3)
u = IIP2P O + PIPoll'
By converting all the points to this new coordinate system of X axis and its corre-
sponding y axis with origin at 0, we can compute the radius of the outer circle as
R 2 = Cp2 2 2
I = (XI - x) + Yl' (4)
and the radius of the inner ci!cle as 2
r = CP3 = (x3-x)2+y~ (5)
Thus, the gap between the two circles, I R - r I, can be easily found from Equations 4
and 5. Since Xv Yv x3, and Y3 are known, the gap only depends on the location of cen-
ter C and is just a function of x as shown in Figure 2. The horizontal axis represents
the location of point C and the vertical axis represents the gap between the two con-
centric circles.
The maximum gap occurs at point M when P v P3, and center C are co-linear. Point
M can be found from the intersection of the two lines PIP 3 and OC. As seen from
Figure 2, if the direction of OC and OM are opposite, moving the center C toward
point 0 will shrink the gap between y

the two circles. If OC and OM are
pointing to the same direction and
IIOMII > lIoell, the gap will also shrink
when we move center C towards the
origin. Thus, we need to consider two
different cases when minimizing the
gap.
x
Case1: When OM. oe < 0, the gap '-+-I--I-~-------:::::OO~C""'---I~
will shrinks as the center C moves
toward point O. Therefore, to obtain
the smallest gap, we need to minimize
the distance between center C and
point 0 as much as possible. This is the Figure 1: Definition of the reference frame
same when OM. oe > 0 and for the 3-point enclosure problem.
IIOMII > Iloeli.
Case2: If OM • oe > 0 and 10
IIOMII < lIoell, the center C needs to
move away from the origin in order to
get a smaller gap. 5
The objective of gap minimization

now can be easily achieved by relying - - o f - - t t - - - - '. .~----+_-4t_
on local monotonicity of the gap func- -100 M C
tion IR - r I [WIL 78], either minimiz- Figure 2: Plot of the gap function.
ing or maximizing the distance from _____
--l;~ _ _ _ _ _ __
-J.l~
the center C to the origin based on their relative positions. Using local monotonicity
analysis techniques allow us to replace the non-linear gap function by a linear func-
tion (the elongation x) that varies in the same direction as our objective.
5 THE 4-POINTS CIRCULAR ENCLOSURE PROBLEM
Assume now that at lease one more point is included in the ring. The movement of
the center C towards the origin will stop when a new point is touched by either the
inner or the outer circle. Hence, the center C is now on a Voronoi vertex, either NVD
or FVD, or on a common intersection of a farthest Voronoi edge and a nearest Voronoi
edge. The position of the center is then fully constrained by 4 points. Assuming Pl and
P2 are on the outer circle and P3 is on the inner circle, any point Pi' i =4 to n, should
satisfy the follOWing two constraint conditions,
(6)
and
Expanding and rearranging Equation 6, we shall have the following results:

2 2
CPi -CP I ~O, (8)
(CPi-CP I ). (CPi + CP I ) ~O, (9)
PIPi • (CPi+CP I ) ~O. (10)
Hence, we have the first constraint equation as follows,
(11)
Similarly, from Equation 7, we have the second constraint equation as

(12)
If PI and P2 are on the inner circle and P3 is on the outer circle, any point Pi should
satisfy the following two constraint conditions,
(13)
and
(14)
Therefore, the two constraint equations are as follows,

PIP,. (OP i +OP I -2xX) ~O,i=4ton; (15)
P 3 P i • (OPi +OP3 -2xX) ~O,i=4ton. (16)
As seen from Equations 11, 12, 15, and 16, the constraints are simply linear. There-
fore, the process to find the new center for a smaller gap is a Linear Programming
problem at each iteration. Since it is an one-dimensional LP problem, we can simply
solve it by sorting all the constraints and find out which one will be the active con-
straint. In this paper, a quick sort [Pre 89] is used for sorting all the constraints.
6 CHEBYSHEV CIRCULAR FITS
We now tum our attention to the more general problem of finding a Chebyshev fit to
an arbitrary set of data points and we present an algorithm to that end. While this
algorithm is not guaranteed to converge under all circumstances, its speed and high
statistical rate of convergence when the starting point is already a least square fit cen-
ter make it a good candidate for real time implementation.
Input: A point set 5 with n points, and a starting point.
Output: The minimum gap of a pair of concentric circles to enclose all the n points.
A list of four points, P v P2, P3, and P4, is maintained to keep tracking the current
constraint points. Among these four points, PI is the oldest constraint point found.
Point P2 is the second oldest constraint point, P3 is the third oldest, and P4 is the
youngest constraint point found.
(i) Find the first pair of concentric circles. The center of these two circles is a vertex of
NVD [For 87] nearest to the starting point. Set P v P2, and P3 as the three points on
the inner circle and P4 as the one on the outer circle.
(ii) Repeat steps iii and iv while a smaller gap is found.
(iii)(a) If P2, P3, and P4 are not on the same circle, find the new constrain point and the
associate two circles by relaxing PI from 5, i.e., temporarily remove PI from 5. The
center of the new circles and the gap between their radii are found from the subrou-
tine GAP(P2, P3, P4, PI)'
(b) IfP v P3t and P4 are not on the same circle. Call GAP(P v P3t P4, P2) again and
relaxing P2 to find another pair of circles and the associate gap.
(c) If P2, P3, and P4 are on the same circle at (3a) or P v P3, and P4 are on the same
circle at (3b). Relax P3 and call GAP(Pv P2, P4t P3) to find a new pair of circles and
the gap between their radii.
(iv)If the smallest gap between 3a to 3c is smaller than the original minimum gap,
drop the constrain point we relaxed when finding this new minimum gap from the
constrain point list. Reset the remaining three constrain points as P v P21 and P3. Set
the new constrain point as P4 and replace the original minimum gap with the new
minimum gap.
(End of step 2 loop)
(v) End.
GAP(Tl' T 21 T31 T4) : This routine is used to find the new constrain point and compute
the center of the new pair of circles and the gap between them.
Input: The four points Tv T21 T3, and T4 are the points locate on the current two con-
centric circles. Point T4 is the one we are relaxing for finding a new pair of concentric
circles.
Output: A new constraint point touched by either inner or outer circle. The center and
gap between the new pair of concentric circles.
(i) Rearrange Tv T21 and T3 for TI and T2 to be the points on the same circle and T3 is
the one on the other circle.
(ii) Find the center 0 that passes through Tv T2, and T3'
(iii)Find the unit vector, X, of OC, which is the bisector of TI and T2'
(iv)For i = 5 to n, find all the constraints from the following two equations if Tl and
T2 are on the outer circle.
TlTj e (OTj + OTl - 2xX) ::; 0 (17)
T3Tje (OT j +OT3 -2xX) ~O (18)
If T1 and T2 are on the inner circle, the above two equations are substituted by the
following two equations.
TlTje (OT j +OT l -2xX) ~o (19)
T3Tj e (OTj + OT3 - 2xX) ::; 0 (20)

(v) Sort all the constraints obtained at step iv.
(vi)Compute the location of point M . Determine which procedure is needed, a mini-
mization or a maximization.
If (OM e OC < 0) minimize = TRUE
else if (IIOMII > IIOCII) minimize = TRUE
else minimize =FALSE
If (minimize = TRUE) minimize the distance, x, between the current center C and
point 0
else maximize the distance x
(vii)Compute the new center and the new gap, return.
7 SOFfWARE CALIBRATION
The procedure designed to test the algorithm convergence and accuracy consisted of
applying it to a set of point randomly generated inside a ring centered at the origin (0,
0). For the example shown in figures 3 through 6, the outer circle's radius is set to 100
and the inner circle's radius is set to 90. The starting point for these figures is set at a
radius of 100 and polar angle equals of 100 degrees.
In all instances, the result of our algorithm was calibrated against the solution pro-
vided by the exhaustive method of [Roy 92]. This was designed to ensure that the
algorithm converges to the actual minimal solution. Figures 3 through 6 show the
march of the algorithm superimposed to the NVD and FVD. For all of the data sets we
have examined, the answers found by the proposed algorithm were the same as those
found by the exhaustive methods of [Roy 92] or [Lai 88]. Moreover, the proposed
algorithm is significantly faster than the exhaustive methods (see Table 1). An exam-
ple for n =50 is illustrated in Figures 3-6, and the results are shown as follows:
The proposed method yields the minimum radial gap between a pair of concentric
arcs enclOSing all the data points, therefore measuring the circularity error of the set.
The method may however fail to find the global optimum if the spanning angle for
the two arcs is smaller than Tr/6. When the spanning angle of the two arcs is small, the
result of the algorithm. will be

affected by the position of the
starting point. For best results,
we choose the starting point as
the center of a least square cir-
cular fit [Tak 89]. Furthermore,
the total number of steps to
reach to the global optimum is
also reduced by the assistance
of a least square method.
Figure 3: NVD with the Figure 4: FVD with the
trace of the center of the trace of the center.
While convergence of the two concentric circles.
algorithm. to the actual global
mInimum is not guaranteed, Monte-Carlo simula-
tions show a remarkable statistical rate of success. Points Tlme(s)
When the center of a least-square fitting circle is used Exhaustive Marching
n
as starting point, no failure was recorded. A typical method algorithm
convergence map of the algorithm. is shown in Figure !>U U.Ul U.U1
7 100 0.17 0.03
300 0.86 0.12
600 3.48 0.32
800 6.04 0.36
8 APPLICATION TO FRINGE 1000 8.89 0.62
METROLOGY FOR SUB-MICRON Table 1: Comparative run

PosmON SENSING time between the marching
algOrithm and exhaustive
methods
The algorithm presented in this paper was specifi-
cally cfeveloped to support image metrology of interference fringe patterns in the
design of a sub-micron position sensor. Two types of measurements are needed for
that purpose: center position and circularity.
8.1 Sensor Design Overview

Figure 8 shows the basic design of an optical interferometer designed to track the two
components of displacement of a characteristic point of a machine carriage that are
perpendicular to the axis of translation. A spherical mirror is used to characterize a
point of the carriage and a collimated light is used as a reference axis. The collimated
light is reflected off the spherical mirror as spherical wavefronts near the optical axis,
as per Gauss approximation [Das 91]. The reference beam is reflected off a plane mir-
ror as planes wavefronts. Recombination of the measurement and the reference beams
produces Newton's rings in a plane perpendicular to the optical axis.The concentric
circular patterns are then captured by a vision system.
8.2 Center Detection:

Application to a 2-D
nanometric position sensor
Figures 9 and 10 summarize the

measurement principle. If the dis-
placement of the spherical mirror is
perpendicular to the optical axis then
the circular fringe pattern will be
translated physically by the same
amount. Since the fringe pattern is
captured by a vision system, the phe-
nomenon will be seen as a translation
of the center of the two dimensional
pattern by an amount proportional to
the actual displacement of the spheri-
cal mirror. The proportionality coeffi-
cient is a function of the design of the
interferometer and of the pixel
dimensions. If there is no longitudinal
displacement, then the radii of the
pattern will stay the same, as shown
in Figure 9.
A motion parallel to the axis corre-

sponds to a longitudinal translation
along the railing. As shown in Figure
10, if a very small displacement
occurs, of the order of 10 nanometers,
the fringes will change in diameter
but the position of the common center
will remain unchanged.
Actual results are a combination of

the two previous motions. There is a
change in fringes diameter, expansion
or decrease depending on the dis-
Figure 6: Enlargement in the cen er area of placement of the sphere with respect
F.i~e 5, showing the final steps of the algo- to the camera. The displacement of
.:..r1:..:.t:.=.:...________________ the center of the pattern will replicate
exactly the radial displacement away from the optical axis.
8.3 Circularity Detection:

Application to Self-
Calibration of Parallelism
Experimental calibration of the device is

reported in [Hil 93]. It revealed that mea-
surements were extremely sensitive to
parallelism errors between the CCD plane
and the reference mirror. Indeed, if the
camera plane and the reference plane mir-
ror are strictly parallel concentric circles
will be recorded. If not, then the pattern
consists of "near ellipses". A simulation of
this phenomenon is shown in Figure 11 for
greatly exaggerated parallelism error.
Figure 7: Convergence map of the
Detection of the parallelism error can be marching algorithm. Randomly gener-
achieved by two means: (i) measure of the ated points are colored cyan (light) when
circularity error of the fringes, or (ii) mea- the algorithm converges to the actual
sure of the cocentricity error of the fringes. minimum. Points are colored blue (dark)
points when the algorithm stops at a local
The latter is illustrated by the simulations minimum.
shown in figures 12 and 13. Figure 12
shows the change in center location for
successive fringe orders as the angular
error increases. Figure 13 traces the
motion of a fringe center for a horizontal beamspliuer
motion of the mirror for various angular
errors. Note that the response of the sys-
tem is still linear, but that the offset
changes with the angular error. Figure 8: Interferometer for transla-
tional error measurements.
This simulation has shown that if the
computation procedure kept the same set
of fringes then the center location will move along a straight line whatever the angu-
lar defects are, and most importantly along a line with the same slope that if there was
not any angular defect. Hence the calibration results after correction and the conver-
sion coefficient is measured as:
M = 800 nanometers/pixel ± 10% (21)
Since the repeatability of the system was found to be 0.1 pixel, and that the conver-
sion coefficient was 800 nanometers, then the resolution of the system is:
resolution = 80 nanometers. (22)
Most of the circle measurements exposed in this section were originally performed
using the resolution tensor approach

exposed in [Hil 93]. These results
'~~1
original position
were further improved by up to a fac-
~.
,$:: 2});'" tor of 5 using the marching algorithm
presented in Section 6. Repeatability
, ~~~
measurements performed on multiple
..
,', - sample measurements of the same
position are shown in Figure 14.
Idv = a,dxl
Experimental results have also
Fi~ure 9: A displacement of the spherical revealed that fringe circularity mea-
mlrror in the plane normal to the optical axis
will result in a proportional displacement of surement is a poor indication of paral-
the center of the interference pattern. lelism error. However the cocentricity
error among fringes, and in particular
its orientation and magnitude trans-
late directly into an angular measure-
ment and can be used in the feedback
control of piezo-electric actuators
mounted in the back of the reference
mirror. This remark provides an excel-
lent opportunity for self-calibration of
parallelism prior to performing posi-
original position after longitudinal
displacement tion measurement. Once the parallel-
ism error is compensated for, then
experimental results showing an
80 nanometer resolution are vali-.
dated.
Fi~ure 10: A displacement of the spherical

muTOr along the axis will create a cliange in 9 CONCLUSION
the fringe radii, but the center remains in the
same place.
This paper introduced a novel algo-
rithm applicable to the computa-
tional metrology of the circle. This
algorithm was developed in the con-
text of precision machine design and
led to many "first" in metrology. First
(b) and foremost, the marching algorithm
Figure 11: Simulation of the effect on the cap- constitutes the backbone of a new
tured fringe pattern when angular defects image processing approach to fringe
occur between the reference plane mirror and pattern analysis. This approach was
the camera plane. a) No angular defect, the
Eattern is circular concentric. b) Angular used in tum in the design of a preci-
aefect = 45' about the vertical axis, the pattern sion 2-D position sensing interferome-
is no longer circular. ter with a resolution of 80 nanometers
over a depth of field X center coor-dln.te function of tn. .ngula,. ctefet Y center coordinate function of
___
tn. "'lIula... ct.fec
of 1 meter. The reso- O"r-~-~-------'

O~~-~- ~-,
lution actually may 0.08 '~"'~.~, , 1st fringe

-
be as fine as 10 nm
but could not be ver-
0.06
~ 0.04
-0.5 ...............
" ----
"'" " - _.2nd fringe
t t
',..... ...... .....
ified as this is below 0.02 ~ ~ -1
,, ""2:~ fri~ge-
the resolution of our ,~----------------~,
o
··
calibrating appara- d~ -0,02 ... -1.5
. "-..., ....
d
tus. ~ -0.04
"
·
t -0,06 t -2
4th fringe
-0.08
We also reported
that the major stum- -0"0
Anaula,. I»fect tcleirM)
bling block to this
work was the preci- Figure 12: Variation in the center position for four successive
sion of parallelism fringe radii, angular defect around the vertical axis only.
between the CCD
plane and the reference mirror of the ,.;, _u. ,-.~ _. -,,~ ~ - -, ',-' ,-... ~ _. ,~"~,,,~ _ ..
. t f
mer A s'It turns out, the
erometer. xT=IOOpixels
oT=150pixels
I
same a gon ·thm h as proven use ful .
m: ,.- • T=200 + T=220 pixels .,' ,.,:.
::::'
pixels
the measurement of angular errors ; ,.~
! ~
' : .. ..
and could be used in the self-calibra- ;~.~
I I : : = t ; I t I I I : I I I I
tion of parallelism. :~._

~.-
Running time of this algorithm are ~.", ,., ,.. ,.. ,.•
such that it represents 1-2 order of ....,- "..w " . ' _ ' ....."
magnitude improvement over alter- Figure 13: Simulation on a theoretical image

native approaches. Using a least of the effect of the detector plane angular
square method to generate the start- defects on the center location for different
ing point not only can reduces the _thre __sh_o_l_d_s_T_._ _ _ _ _ _ _ _ _ _ __
total time required but also can avoid
the potential failure to reach the global optimum when processing arcs.
10 ACKNOWLEDGEMENTS
This work was initiated as a joint project with Prof. Andre Clement, of the Institut
Superieur des Materiaux et de la Construction Mecanique in Saint-Ouen, France. It
was made possible in part through a National Science Foundation grant DDM-
9057059, the ATT foundation, the Newport Corporation of Fountain Valley, California
and the Department of Mechanical Engineering, Aeronautical Engineering and
Mechanics at Rensselaer.
Resolution Tensor Marching Algorithm

235.·················································· ......................... 235
230 -I
.. . . 230
. .
...........................................................
~:
.. ." 225
.
225- ••• ·.~t. .
220 220
.~ .'
215 215
210 210+---~----~--~----~--~
302 304 306 308 310 31 302 304 306 308 310 312
(a) 40 Sample Measurement
Resolution Tensor Marching Algorithm

250 250 ..........................................................
240+························ •........•........................................... 240
230+················································ " ••....................

lr.
230
220 ......... .............................. ..~.

'.
210-/············································ .................................. '.
. 220
210
• ••
200~------------------ __--~ 200
280 290 300 310 320 330 280 290 300 310 320
(b) 73 Sample Measurement

Figure 14: Comparative repeatability study of the Resolution tensor versus the march-
ing algorithm performed on 40 (a) and 73 (b) experimental images taken for the same
position. The results show a factor of 5 improvement in the repeatability.
11 REFERENCES
[ANS 85] ANSI standards, Methods for Performance Evaluation of Coordinate Measuring
Machines, ANSI/ASME. , B89.1.12M-1985.
[BOT 79] Bottema, 0., and Roth, B., Theoretical Kinematics, North Holland Series in
Applied Mathematics and Mechanics, Lauwerier, H.A., and Koiter W.T., Ed., (1979)
[BUR 77] Burch, J.M., and Williams, D.C., Varifocal Moire Zone Plates for Straightness
Measurement, Applied Optics, Volume 16, N° 9 (1977) pp. 2445-2450.
[CEL 88] Celaya, M., Hologram Interferometer to Calibrate and Measure the Straightness
in Micropositioning Equipment, PROCEEDING SPIE, Precision Instrument DesilW,
Volume 1036 (1988) pp. 44-51.
[CLE 91] Clement,A., The Resolution of Positioning Solids, CIRP Annals. Vol. 40/1
(1991) pp. 511-514
[DAS 91] Das, P.K., Laser and Optical Engineering, Springer-Verlag, (1991)
[DYS 55] Dyson, J., "An Interferometer for "Straightness" Measurement", Nature, Vol-
ume 175, N° 4456, (1955), pp. 559-560.
[FOR 87] Fortune, S., "A Sweepline Algorithm for Voronoi Diagrams," Algorithmica,
Vol. 2, 1987.
[GON 87] Gonzalez, R.c., and Wintz, P., Digital Image Processing, Addison-Wesley,
(1987)
[HIL 93] Hilaire, T.P., Optical Sensing and Fringe Pattern Analysis of Translational Errors
in Machine Carriages, Ph.D. Thesis, Rensselaer Polytechnic Institute, Dept. of Mech.
Eng., Aero. Eng. and Mechanics (1993)
[LA188] Lai, K., and Wang, J., "A Computational Geometry Approach to Geometric Toler-
ancing," 16th North American Manufacturing Research Conference, 1988.
[LAN 87] Landau, U. M., "Estimation of Circular Arc Center and Its Radius," Com-
puter Vision Graphics Images Processes. Vol. 38, 1987.
[NI 88] Ni, J., and Wu, S.M., "Laser Alignment Techniques for Simultaneous Machine Tool
Geometric Error Detection," SPIE Optical Testing and Metrology IT, (1988) pp. 694-
701.
[N193] Ni, J., and Wu, S.M., "On-line measurement technique for machine volumetric
error compensation," Transactions of ASME, TournaI of Engineering for IndustIy. v
115, n I, (Feb 1993) pp. 85-92
[PRE 89] Press, W. H., Flannery, B. P., Teukolsky, S. A., and Vetterling, W. T.,
"Numerical Recipes in C: The Art of Scientific Computing," Cambridge University
Press. New York, 1989.
[ROY 92] Roy, U., and Zhang, X., "Establishment of a Pair of Concentric Circles with the
Minimum Radial Separation for Assessing Roundness Error," Computer Aided
DesilW, yol. 24, No.3, March 1992.
[TAK 89] Takiyama, R, and Ono, N., "A Least Square Error Estimation of the Center and
Radii of Concentric Arcs," Pattern Recognition Letters, Vol. 10, 1989.
[THO 89] Thomas, S. M., and Chan, Y. T., "A Simple Approach for the Estimation ofCir-
cular Arc Center and Its Radius," Computer Vision Graphics Images Processes, Vol.
45,1989.
[WIL 78] Wilde, D.J., "Globally Optimal Design," Wiley-Interscience Pub .. (1978)
[WU 89] Wu, S.M., and Ni, J., "Precision Machining Without Precise Machinery",
Annals of CIRP, Vol 38/1 (1989) pp. 533-536.
[YIN 91] Yin, c., Two-Dimensional Automatic Straightness Measurement System Based on
Optical Activity, Optical Engineering, Volume 30, N° 4 (1991) pp. 480-482.
20
Inspection Method for
Geometrical Tolerances
using Coordinate Measuring Machine
F. Tanaka, P. Ikonomov, H. Okamoto and T. Kishinami
Division of Systems and Information Enginnering, Hokkaido Univ.
Kita-13, Nishi-8, Kita-ku, Sapporo, Hokkaido, 000, JAPAN
TEL :+81-11-706-6449 FAX:+81-11-706-7836
E-mail: tanaka@hupe.hokudai.ac.jp
Abstract
The method of inspecting geometrical tolerances is important from the economic and qualitative
viewpoints in Coordinate Measuring Machine (CMM) measurement. Much work has been done
on the evaluation of the geometrical deviation of form for a single feature from a measured data
set. But, on the other hand, the inspection of geometrical deviation for related features has not
been well developed. In this paper, we first discuss the following topics: l)Problems of CMM
inspection based on ISO Geometrical Tolerance, 2)Formal representation of ISO geometrical
tolerance using EXPRESS, and 3)Necessity of a virtual gauge as a criterion for deriving
geometrical deviation. Secondly, we propose a virtual gauge as a criterion for deriving the
geometrical deviation of the toleranced feature from a measured data set. For the mathematical
representation of the virtual gauge, we introduce mathematical representations of geometrical
elements and the relationship between datum and toleranced feature. We extend the small
displacement screw method in order to find out the minimum geometrical deviation based on the
virtual gauge. We also show that the proposed method is very effective by applying it to some
examples.
Keywords
Geometrical Tolerance, Computational Metrology, Coordinate Metrology.
INTRODUCTION
In the manufacturing field, it is recognized that a form evaluation method that could calculate the
definition parameters of substitute geometrical elements, such as position, orientation and size
from the measured data set of real features in Coordinate Measuring Machine (CMM)
measurement, using the least square method, min-max method, or some other, would be very
Inspection method for geometrical tolerances 299
important. The method of inspecting geometrical tolerances is also important from the economic
and qualitative viewpoints in coordinate measuring. Much work has been done on the
evaluation of the geometrical deviation of form (planarity, cylindricity, etc.) for a single feature
from a measured data set [ISO TR I0360-1,Wirtz (1993)]. But, on the other hand, the
inspection of geometrical deviation for related features (angularity, positional deviation, etc.)
has not been well developed [Ikonomov (1993), Ballu (1991)]. In this paper, we discuss the
problems of ISO geometrical tolerance inspection, the representation of ISO geometrical
tolerance using formal specification language EXPRESS [ISO 10303-11 (1994)], and propose
a method of evaluating geometrical deviation for related features by means of a virtual gauge.
2 PROBLEMS OF INSPECTION BASED ON ISO GEOMETRICAL

TOLERANCE
To inspect the required geometrical tolerance using CMM, first we must derive the geometrical
deviation for the geometrical feature from the measured data set. Second, we must check
whether the derived geometrical deviation is within the given geometrical tolerance or not.
However, explicit representations (such as mathematical form) of the geometrical feature and
the relationship between datum and toleranced features are not given by ISO geometrical
tolerance [ISO 1011 (1983)]. Therefore, we cannot inspect a given geometrical tolerance by
computer in a straightforward manner.
To solve these problems, we represent the structures of related features (the geometrical
elements and the relationships between them), tolerance, measured data set, and functional
gauge using formal specification language in order to make them clear. Also, we propose a
computer model (the virtual gauge) to serve as the functional gauge or a criterion for
establishing the datums, and for deriving the deviation of the features. This computer model can
represent the geometrical constraints of the position and orientation of features related to a
datum.
3 FORMAL REPRESENTATION OF ISO GEOMETRICAL

TOLERANCE USING EXPRESS
We represent the relationships between tolerance specification, deviation, feature and datum
using formal specification language EXPRESS. The purpose of using EXPRESS for our
representation is to clarify the structure of objects and to construct the virtual gauge computer
model.
3.1 Formal representation of datum using EXPRESS

A datum is a, theoretically exact, established geometrical reference for controlling the tolerance
zone when specifying a geometrical tolerance for a related feature [1IS B 0022 (1984)]. A
datum feature is a real feature of the considered object used for establishing a datum. Figure 1
shows the relationship between datum and datum feature. In actual measurement, we usually
use a simulated datum feature that is a real surface of sufficiently precise form to be used in
contact with the datum feature for establishing a datum.
4L1 oierance specification
Figure 1 Illustration of datum inspection procedure.
Figure 2 shows the representation of the relationships between datum and geometrical element,
and datum and measured data set using EXPRESS-G notation[ISO 10303-11 (1994)]. In
EXPRESS-G notation, a rectangular enclosure, normal line and thick line indicate a class of
object, attribute, or sub/supertype relationship, respectively. As shown in Figure2, the Datum
Feature is defined at the design stage and has a geometrical element that represents the nominal
shape. At the inspection stage, we calculate the geometrical element of a datum from the
measured data set of that datum.
_ - - - - - - - - - - - - - - - - D e s i g n Stag
Parameter
Parameter
' - - - - - - - - - - - - - - - - i n s p e c t i o n Stag
Figure 2 EXPRESS-G notation of datum specification and inspection.

3.2 Formal representation of related feature using EXPRESS

There are 11 varieties of geometrical deviations of features [1IS B 0621 (1984)], but we restrict
the scope of our research to only those related features that can be derived from a measured data
set directly, Geometrical deviations of axis are outside our scope,
Figure 3 shows the ISO Tolerance [JIS B 0021 (1984)] definition of parallelism and the shape
of real features. Parallelism shall refer to the amount of deviation of a straight line feature or
plane feature, which is supposed to be parallel to the datum straight line or datum plane, from
the geometrical straight line or geometrical plane parallel to the datum. Thus, the deviation of a
measured point is the distance from the plane with exact directions in relation to the datum. The
maximum deviations will be compared to the tolerance zone. Positional deviation also shall
mean the amount of deviation of a point, straight line feature, or plane feature from the
theoretically exact position specified in relation to the datum or other features. In both cases, we
must establish the geometrically exact element whose position and orientation are exact in
relation to the datum. The substitute element proposed in Coordinate Metrology is the exact
shape and the criteria for deriving the geometrical tolerances.
datum
Figure 3 Illustration of related feature inspection procedure.
Figure 4 shows the representation of ISO geometrical tolerance using EXPRESS-G notation.
The product model has datum features and toleranced features specified at the design stage.
Both features have a geometrical element. The geometrical tolerance is defined as the relation
between datum and toleranced feature in which the toleranced feature is related to the datum. It
also has the attributes labeled Toleranced Type and Tolerance Zone. In tolerance inspection
using CMM, first we calculate the geometrical deviation of each measured point. Second, we
compare the maximum deviation of a measured point set and the tolerance zone. However, the
spatial relationships between datum and feature, that is, the geometrical constraints of position
and orientation, are not specified explicitly as computational or mathematical models as a
standard. Therefore, we need to supplement the Deviation Type constraints and establish the
exact feature as the basis for deriving the deviations.
Design Stag
relating r···········:·············:····, related
: GeometncaLConstramt :
'--------------0--------------
1 Tolerance_Type Tolerance_Zone 1
11
y Y
1 GeometricaLTolerance 1
relating related
I Parameter 1 1 Parameter 1
Y ?
L(j GeometricaLElement 1 ·1 Geometrical_Element J:>-
nominalY (') () Ynominal
Datum_Feature Toleranced_Feature
Measured_Data_Set 1 1 Measured_Data_Set 1
~ Geometrical_Element 1 1 Geometrical_Element p-
O ~) \,) n
1
Parameter I I Parameter I
datum exacLfeature
1 GeometricaLDeviation 1
Deviation_Type
n n
MAX_Deviation
1 1
()
relatina related
Geometrical Constraint
Inspection Stag
Figure 4 EXPRESS-G notation of related feature specification and inspection.
3.3 Formal representation of related feature under Maximum

Material Condition using EXPRESS
Thirdly, we analyze the evaluation method of related features applying Maximum Material
Principle (MMP) [JIS B 0023 (1984)]. The MMP gives the relationships of mutual dependence
between dimensional tolerances and geometrical tolerances based on the maximum material
condition. Figure 5 shows a hole feature where perpendicularity tolerance is specified under
MMP. In an inspection procedure under MMP, the maximum material condition of a real
feature is compared with the virtual condition, which is the limit of tolerance related to the
datum. In this procedure, a virtual size functional gauge is often used.
Maximum material condition

Functional gauge
Figure 5 Illustration of related feature inspection under Maximum Material Condition.
Figure 6 shows the representation of ISO geometrical tolerance under MMC using EXPRESS-
G notation. In coordinate measuring, we simulate the functional gauge in a computer. The
inspection procedure is as follows: First, we determine the position and orientation of the
functional gauge from the measured data set using the contact element method. Second, we
calculate the minimum and maximum distances between the gauge and the measured data set.
Third, we compare the parameter of the functional gauge and distances to the tolerance
specifications in order to evaluate tolerance.
4 VIRTUAL GAUGE AS CRITERION FOR DERIVING

GEOMETRICAL DEVIATION
For deriving the geometrical deviation of the toleranced feature from a measured data set, we
propose the virtual gauge: this consists of geometrical elements representing the datum and
toleranced feature, and geometrical constraints representing the relationship between datum and
toleranced feature, in mathematical form. Like a gauge, the virtual gauge gives us a criterion for
deriving the geometrical deviation for the toleranced feature from a measured data set in
computer software form. We explain our computer model of the virtual gauge and the
derivation method for geometrical deviations in the following subsections.
4.1 Computer model of virtual gauge
The virtual gauge is a basis for deriving geometrical deviation in computer software form, and
its parameters are calculated under the fitting criteria (Min-Max method, or Contact Element
method) from a measured data set.
Design Stage
relatino r········· .. ~·············:····. related
: Geometrical_Constraint •
··----·--------0--------------
1 Tolerance_Type Tolerance_Zone 1
11
y Y
I Geometrical_Tolerance I
relating related
I Parameter I Parameter I
y Y
LQ Geometrical_Element 1 GeometricaLElement p.
nominal Y () 0 Ynominal
Datum_Feature Toleranced_Feature
Measured_Data_Set I I Measured_Data_Set
--c:l Geometrical_Element I I Geometrical_Element ..p-

1'\ () 1..) 1'\
I Parameter I Parameter I
datum gauge
1 Functional_Gauge 1
A A
I Gauge_Type MINIMAX_Distance I
()
relating related
Geometrical Constraint
Inspection Stag
Figure 6 EXPRESS-G notation of related feature specification and inspection

under Maximum Material Condition.
As shown in Figure7, the virtual gauge has two subtypes, the Datum Virtual Gauge and Related
Feature Virtual Gauge. The Datum Virtual Gauge is a simulated datum feature, and it has some
attributes which are the Geometrical Element, representing the datum feature, and the Fitting
Criteria.
The Related Feature Virtual Gauge also has some attributes, which are the Datum Virtual
Gauge, Geometrical Element, Deviation Type, Fitting Criteria, and the MAX Deviation and
Geometrical Constraint between datum and feature. It is a criterion for deriving geometrical
deviation and simulates a functional gauge for deriving the deviation of related features under
MMC.
L -_ _ _ _ _ _ _--:~ GeometricaLConstraint 1----1
Figure 7 EXPRESS-G notation of virtual gauge model.
We introduce plane, cylinder, sphere, and cone as the geometrical elements of the virtual gauge.
Figure 8 shows the representation of these geometrical elements and the defined parameters of
the virtual gauge using EXPRESS-G notation These elements have a local coordinate system
[ISO 10303-42 (1994)] that defines their position vector P and orientation vector N, as shown
in Figure 8. Cylinder and sphere also have a radius, and cone has a semi angle.
GeometricaLElement
Figure 8 EXPRESS-G notation and vector representation of geometrical element of virtual

gauge.
4.2 Derivation Method for Geometrical Deviations

In the derivation procedure for geometrical deviations using the virtual gauge, we must find out
position so that the distance from a measured data set to the virtual gauge is minimized. This
procedure can be formulated as a constrained optimization problem. To solve it, we use the
small displacement screw (SDS) method, which was proposed by P. Bourdet [Bourdet
(1988)]. In Figure9, Mi is the measured point on a real surface, Mgi i~ the gauge point lying on
the unit normal vector (nj), and the deviation (~i) from Mi is MgiMrni. The deviation after
optimization (ej) is given by Equation (I) :
ei = ~i - {DA·ni + (~ixni).R} (I)
where, D A = [DAx DAy DAz] is the displacement of reference point A on the virtual gauge,
and R =[ cpu I CPU2 cpU3] is the rotation vector whose rotation axis is [u I U2 U3].
Figure 9 Small Displacement Screw Method.
The geometrical constraints in the virtual gauge consist of position and orientation constraints.
The position constraint to the reference point P(Px Py Pz) of the geometrical element is given
by Equation (2) :
(2)
Figure 10 shows the position constraint plane (Ax + By + Cz + D =0) for the reference point P
of the geometrical element in the virtual gauge. Equation (2) means a restriction where the
translation displacement of the virtual gauge is only allowed in the specified plane. The
representation of this restriction is given by Equation (3) :
(3)
where, [A B C] is a normal vector of the specified plane.

The orientation constraint to the orientation vector N(N x Ny Nz ) of the geometrical element is
given by Equation (4) :
[A B C]. [N x Ny Nz Y= cos 9 (4)

where e is the angle between the orientation vector and the constraint vector. Figure 11 shows
the virtual gauge plane, defined with its position vector P and orientation vector N, and the
constraint vector CD[A,B,C] for orientation. Equation (4) means a restriction where rotational
displacement of the virtual gauge is only allowed around the specified axis. This restriction is
represented by Equation (5) :
R =[ <l>A <l>B <l>C 1 (5)
where, [A B C] is the unit direction vector of a specified axis, and <I> is the variable of the
rotation angle.
Virtual Gauge
Figure 10 Position constraint. Figure 11 Orientation constraint.
We extend the SDS method in order to handle geometrical constraints that are constraints of
position and orientation. These geometrical constraints are fonnulated as linear equations and
used in the linear programming in this method.
5 EXAMPLE
We show two examples of inspection of a related feature. The angularity for a plane is specified
such as that two parallel planes inclined at a specified angle to the datum plane, will have been
chosen so as to have the minimum width and contain between them all the measured data set, as
shown in Figure 12. The width between the two parallel planes is the angularity deviation. It is
equivalent to finding the minimum deviation, rotating the nominal plane about the normal vector
of the datum plane with the orientation constraint.
Figure 12 The angularity for a plane.

Figure 13 shows the angularity tolerance for an axis of a cylinder as specified under MMC. In
this case, the virtual gauge needs to incline exactly 75 degree against the datum and the fitting
criteria is Maximum Inscribed. Thus, preserving the orientation constraint between the axis
vector of the virtual gauge and the normal vector of the datum, and rotating the gauge around
the normal vector CD, we derive the maximum diameter of the cylinder under Maximum
Inscribed. The derived diameter is compared to virtual size in order to inspect the tolerance
specification.
Figure 13 The angularity for a cylinder under MMC.
6 COMPUTER SIMULATION
In computer simulation, the measured data set for the test piece data is calculated by the
following procedures: First, we establish a boundary cylinder that has the parameter value of
Table. 1 and a diameter the same as virtual size. Second, we calculate 6 measured points on that
cylinder, and 14 points outside of the cylinder. Figure 14 shows the measured point set of the
test piece. Then, we fit the virtual gauge that has the parameter of Table. 2 to the test piece data.
Figure 15 shows the virtual gauge before fitting and the measured data set. Figure 16 shows the
virtual gauge after fitting and the measured data set; and Table 3 shows the parameter of the
resulting position of the gauge. From Tables 1 and 3, we can see that the proposed method is
very effective.
x
5 2.5 0 -2.5 -5
, -2
10 •
. 0
2Y
6
10
• • 7.5
5z
2.5
0
Figure 14 Test piece data for the angularity for a cylinder under MMC.
Inspection method/or geometrical tolerances 309
Table 1 Parameter of test piece data.
Axis Vector [0,0.258819,0.965926]

Diameter 9.85
Reference Point [0.0,0.0, 0.0]
x
5 0 -5
-2.5
1
0
Y
2.5
•
fZS. 5
Figure 15 Initial position of gauge.
Table 2 Parameter of initial position of gauge.
Axis Vector [0, -0.258819, 0.965926]

Diameter 10.0
Reference Point [3.0, 3.0, 0.0]
x
5 2.5 0 -2.5 -5
-5
f-....JIo......:::o.....- - I 15
10
5 Z
'--_ _ _ _ _..... 0
Figure 16 Result position of gauge.

Table 3 Parameter of result position of gauge.
Axis Vector [1.64099 X 10- 12,0.258819,0.965926]

Diameter 9.85
Reference Point [-6.94797 X 10- 12, -2.57471 X 10- 12,0.0]
7 CONCLUSIONS
In this paper we discussed the following three topics:
1) Problems of CMM inspection based on ISO Geometrical Tolerance

No explicit representation of the relationship between the toleranced feature and datum systems,
makes it impossible to evaluate the geometrical deviation for a related feature from a measured
data set given by CMM.
2) Formal representation of ISO geometrical tolerance using formal specification language:

. We specify the representation of datum, toleranced feature, and the relationship between them
in ISO geometrical tolerance by formal specification language. The roles of bases in deriving
geometrical deviations and the functional gauge are also specified.
3) Virtual gauge as criterion for deriving geometrical deviation:

We propose the virtual gauge as a criterion for deriving the geometrical deviation of the
toleranced feature from a measured data set. For the mathematical representation of the virtual
gauge, we introduce mathematical representations for geometrical elements and the relationship
between datum and toleranced feature. We extend the small displacement screw method in order
to find out the minimum geometrical deviation based on the virtual gauge. We also prove that
the proposed method is very effective by applying it to some examples.
8 REFERENCES
ISO TR 10360-1, Coordinate Metrology - Part1: Definitions and applications of the
fundamental geometric principles. , International Organization for Standardization, Geneve.
Wirtz,A. et.al (1993), From Unambiguously Defined Geometry to the Perfect Quality Control
Loop, Annals of CIRP, 42,615-618.
Ikonomov, P. et. al (1993), Evaluation Method for Geometrical Forms in Coordinate
Metrology, in Robotics, Mechatronics and Manufacturing Systems (ed.Takamori and
Tsuchiya), 711-716, Noth-Holland,Amsterdam.
Ballu A. et. al (1991), The Processing of Measured Points in Coordinate Metrology in
Agreement with the Definition of Standardized Specification, Annals ofCIRP, 40,491-494.
ISO 10303 -11 (1994), Industrial Automation Systems and Integration - Product Data
Representation and Exchange - part-I/, International Organization for Standardization,
Geneve.
ISO 1101 (1983), Technical Drawing-Geometric Tolerancing, International Organization for

Standardization, Geneve.
ANSI Y 14.5 M (1982), Dimensioning and Tolerancing, The Amercan Society of Mechanical
Engineers, New York.
JIS B 0022 (1984), Datums and Datum-system for Geometrical Tolerances, Japan Standards
Association,Tokyo.
JIS B 0621 (1984), Definitions and Designations of Geometrical Deviations, Japan Standards
Association,Tokyo.
JIS B 0021 (1984), Indications of Geometrical Tolerances on Drawings, Japan Standards
Association,Tokyo.
JIS B 0023 (1984), Maximum Material Principle, Japan Standards Association,Tokyo.
ISO 10303 -42 (1994), Industrial Automation Systems and Integration - Product Data
Representation and Exchange - Part-42, International Organization for Standardization,
Geneve.
Bourdet P. et. al (1988), A Study of Optimal-Criteria Identification Based on the Small-
Displacement Screw Model, Annals of CIRP, 37, 503-506.
Author biographies
Fumiki Tanaka received the M.Eng. degree from Hokkaido University in 1987. He is research
assistant of the division of systems and information engineering, Hokkaido University. His
research interests include geometric modeling, instection of mechanical products, information
modeling of manufacuturing for mechanical products. He is a member of JSPE (Japan Society
for Precision Engineering).
Pavel Ikonomov received the D.Eng.degree from Hokkaido University in 1993. His research
interests include nstection of mechanical products. He is a member of JSPE.
Hideaki Okamoto is manager of production enginnering development department of NIKON

Corporation. He is currently pursuing the D.Eng. degrees under Professor Kishinami in the
division of systems and information engineering, Hokkaido University. His research interests
include instection of mechanical products, information modeling of manufacuturing for
mechanical products. He is a member of JSPE.
Takeshi Kishinami received the D.Eng. degree from Hokkaido University in 1971. He is
Professor of the division of systems and information engineering, Hokkaido University. His
research interests include geometric modeling, instection of mechanical products, rapid
prototyping, information modeling of manufacuturing for mechanical products. He is a member
of JSPE and JSMB (Japan Society for Mechanical Engineers).
21
A Soft Gaging Approach for Complex Cases
Including Datum Shift Analysis of
Geometrical Tolerances
C. Fortin and J-F. Chatelain

Ecole Polytechnique of Montreal
Department of Mechanical Engineering
Box 6079, Station "Centre-ville", Montreal, Quebec, Canada H3C 3A7
Phone: (514) 340-4907 Fax: (514) 340-5867
E-mail: clementjortin@mail.polymtl.ca
Abstract
A soft-gaging approach using solid tolerance zones and three-dimensional alignment algorithms
is proposed to verifY various geometric and dimensional tolerances. The method is particularly
well suited for the verification of position tolerances involving a possible datum shift which still
requires traditional metrology techniques to be solved. The totally CAD based approach uses
constrained and unconstrained alignment techniques in order to bring closer the measurement
points relative to the toleranced features and the corresponding tolerance zones. The process can
either deal, with structured sparse CMM data or unstructured dense data files generated by a
manufactured part inspection process. A typical example is presented to illustrate the developed
methodology applied to a datum shift problem, which is solved using a non-linear constrained
optimization method.
Keywords
Geometric tolerancing, datum shift, tridimensional alignment, measurement of points, solid
modelling
A soft gaging approach for complex cases 313
INTRODUCTION
Following the inspection of a part, which can be done through traditional or software based
metrology techniques, one or more features can be outside the prescribed tolerances. For those
specified at Maximum Material Condition (MMC) and related to a datum feature of size also
specified at MMC, a datum shift can further be considered. Such analysis is generally processed
through functional gaging or paper gaging which is limited to two-dimensional cases. For complex
tridimensional cases, dedicated functional gages are generally used by the manufacturing industry.
The development cost and inflexibility of these gages are however serious dawbacks of the
technique. To overcome these limitations, this paper introduces a computer aided technique for
such tolerancing analysis as well as various other geometric and dimensional tolerance verification.
The technique using optimization aligmnent algorithms and a solid tolerance zone definition can
be integrated within a CAD/CAM system from which the required solid model of the part can be
generated. The verification technique respects the ANSI YI4.SM-(1982) standard.
Best fitting of points used for the calculation of substitute features is a common practice for
the geometric tolerance verification of analytically defined features. Different techniques based on
best fit are proposed in the litterature for the calculation of size and form deviations. Lotze (1982)
presented three possible formulations for the best fit of points based on the minimum zone
principle: the minimization of the squared distances, the unilateral minimization of distances and
the bilateral minimization of the maximum distances. Murthy and Abdin (1980) evaluated the
Simplex method, the Monte Carlo method and the spiral technique to minimize the deviation zone
related to various form tolerances such as sphericity, circularity and planeity tolerances. Fukuda
and Shimokohbe (1984) focused on the comparison ofthe Least Squares and the Minimax criteria
used for the verification of different form tolerances such as straightness, planeity, circularity and
cylindricity. Much more work exists in the tolerance verification area using fitting procedures
which are all different from each other in their formulation and optimization methods used for the
calculations. For a general survey concerning the existing fitting algorithms, one can refer to Feng
and Hopp (1991). Unlike the analytically defined features, sculptured surfaces involve the
calculation of a 6 degrees of freedom rigid transformation to adjust the measurement data in
relation to the CAD related features, and then to verify their corresponding tolerances. Menq and
al. (1992) and Bardis and al. (1991) completed some work suggesting a tridimensional alignment
concept to align a measurement point data set with a corresponding feature. This concept is rather
concerned about the aligmnent of the points with respect to their related features instead of the
calculation of a substitute feature fitted through the set of points. Although these methods have
been found to be efficient and consistent with respect to the minimum zone principle specified in
most of the geometric tolerancing standards, some complex cases have not yet been completely
solved. Our technique proposes to solve various tolerance verification problems using a
tridimensional constrained alignment technique rather than a best fitting method used for the
calculation of substitute features. The technique does not aim to be an alternative to the existing
best fitting techniques but a complementary approach for the verification of various geometric
tolerances.
The alignment algorithms, based on Least Squares and Minimax optimization, are applicable
for datum computation as well as for best alignment of points over their corresponding nominal
and tolerance zone surfaces. The verification technique is described and applied through an
application example considering the analysis of four hole position tolerances, including a possible
datum shift.
2 THE TOLERANCE VERIFICATION ALGORITHM
As shown in figure 1, the geometric tolerance verification algorithm is a systematic process

partitioning the complete inspection of a part into a number of phases. Each phase is defined from
the part tolerancing specified through a CAD database. A phase includes a maximum of four
processing steps: the datums computation, if any specified, the solid tolerance zones generation,
the best fit alignment of data points over their related toleranced feature or tolerance zone
definition, and finally the data point verification within the generated tolerance zones, when
applicable. The number of phases within an inspection process is based on the number of partial
and complete datum reference frames to be computed, and the number of toleranced features not
referenced to any datum. Three types of phases are then defined: the totally constrained phase, the
semi-constrained phase and the unconstrained phase.
For the inspection of a part using a Coordinate Measuring Machine (tactile or optical
probing), the measurement data obtained is intrinsically structured since the inspection is planned
from a program based on the nominal model of the part. Then, each 3D measurement point is
associated to a specific feature of the part. Far from this, some of the non contact inspection
equipment, like laser range finders, generate unstructured dense data files. If such unstructured
data is used for tolerance control, a preprocessing mechanism including reduction and recognition
of features is necessary prior to the tolerance verification of the part.
2.1 The phase partitioning
The phase partitioning is based on the datum reference frame associated to each toleranced
feature. This reference frame involves a datum precedence relationship which restricts a number
of degrees of freedom (DOF) of the part in space for an eventual fit of the points within a
tolerance zone. A degree of freedom is considered to be one of the 3 rotations ($,8,<\» or 3
translations (Tx, Ty, Tz) possible movements of a part in a Cartesian space.
A totally constrained phase involves the calculation of a complete datum reference frame; all
6 DOFs being constrained by the datum definition. A constrained phase could be defined for the
verification of certain position, profile and size tolerances. As shown in figure 1, no further
alignment is required for the tolerance verification included in such a phase. After the computation
of the reference frame and the generation of the solid tolerance zones, the points are verified to
be within the corresponding zones.
A semi-constrained phase involves the calculation of a partially defined datum reference frame;
up to 5 DOFs being constrained by the datum definition. Like a constrained phase, more than one
tolerance verification can be included in the same phase. In addition to the partial constraint set
which must be satisfied, one best fit alignment of measurement points relative to their toleranced
feature is required for each tolerance to be controlled. This best fit alignment procedure considers
as variable parameters, a few selected DOFs based on the type of tolerance verified and on the
partial reference frame previously computed. These parameters, along with the optimization
criteria, the number of iterations and the convergence criterion, are all specified in an optimization
chart associated with each toleranced feature. A semi-constrained phase could be defined for the
verification of orientation, and some profile and position tolerances.
---- [CAD
r;::==========::;l J~~ tolerances
DO
___--__..--------. ~om~ ~~~m
I
Cooslra:ined or unconstrained
alignment: < 6 ddl
I
,_ _ _ _ _ _ _ _ _ _ _ Pomls VI solid zones I
Figure 1 Tolerance verification algorithm.

As an example, one can see a parallelism tolerance defined with respect to a datum surface "A"
which is parallel to the X-Y plane of a cartesian coordinate frame. After computing the datum
feature "A", 3 DOFs are still available for further alignment (<I>, Tx and Ty). However, in order to
align the data points related to their corresponding toleranced feature, the translation along the Z
axis (T z) can be included in the variable parameter set (DOFs) since it is not affecting the
parallelism relationship between the datum feature and the toleranced feature itself The variable
parameters included in the corresponding optimization chart for the best fit alignment process are
then <1>, Tx, Ty, and Tz, the first three DOFs being redundant for this particular example
considering infinite planes. In real applications, where the solid zones are of similar size and shape
as the toleranced feature itself, these 3 DOFs would not be considered as redundant parameters.
Finally, an unconstrained phase involves a totally free state of the part in space with no datum
computation. For form and profile tolerances (with no datum specification), an unconstrained
phase could be defined for each tolerance to be verified. The alignment process then finds an
optimal rigid transformation (3 rotations and 3 translations for the 3D space) from the best fit of
the measurement points relative to the CAD toleranced feature. The tolerance zone is then
generated and the points verified to be within the zone.
The procedure for datum computation considers an alignment algorithm, using a Least Squares
best fit, for the optimal position and orientation of each datum feature constrained or not by a
precedence relationship. As an example, the reference frame of a part, defined with a primary (A),
secondary (B) and tertiary (C) datum, is calculated from the measurement points with a specific
precedence relationship between each datum feature. The datum plane A is aligned from the
computation of all 6 DOFs as variable parameters (if plane A is not simulated from the resting
position of the part surface on a "perfect" reference surface for inspection), while the secondary
and tertiary datum planes Band C are respectively aligned from the 3 and 1 remaining DOFs
which do not affect the previous datum computation.
The tolerance zones are generated with a kinematic formulation (Rivest et al 1993, 1994)
which includes some DOFs needed to initialize the optimization charts used within the alignment
algorithms. These algorithms are used for the best fit of data related to specific toleranced features
or tolerance zones. They are based on a Least-Squares or Minimax scheme, involving up to six
degrees of freedom.
2.2 The tridimensional alignment process
Both, the unconstrained and constrained alignment processes are based on an iterative adjusting
fit, bringing closer the data cloud with the related features, with respect to an appropriate objective
function. The problem is formulated as follows :
I- MINIMIZE F(x), subject to deoi > NEAR_ZERO, with,
F(x) = [Id.lr12 or F(x) = MAX ( doj ), where
doj = I R . M; - Di - Bri·Pi I is the euclidean distance separating M; from Di,
d""i = +/- ! R . M. - D; - C; ! is the oriented euclidean distance separating

M; fromD i,
and,
~ = (11", My, MJ, is a measurement point,

D; = (Dx> Dy> DJ, is a corresponding calculated design point,
Br; is the ball probe radius, when applicable,
p; is the direction of the ball radius compensation,
(; is a vector corrector taking into account various compensations,
x = (1\1, e, <1>, Tx> Ty, TJ is the design variable vector,
ccj>cB (ccj>scj>sl\l- scj>cl\l) (ccj>sBcljr + scj>sljr) Tx

scj>cB (scj>sBsljr +ccj>cljr) (scj>sBcljr -ccj>sljr) Ty
R = (1)
-sB cBsljr cBcljr Tz
o o o
The rigid transformation is expressed in terms of the "Roll, Pitch and Yaw" notation, with 1\1,
e and <1>, being the three successive rotations about the X, Y and Z axis of the appropriate
reference frame, while Tx, Ty and Tz are the translational parameters about each axis. The design
points are computed as to be the nearest points belonging to the CAD surface related to the
appropriate corresponding measurement points. Both the design and the measurement points are
involved in the computation of the euclidean distance dci and the oriented euclidean distance deo;.
The orientation of the latter distance is required to distinguish inside points from outside points
with respect to a particular feature or tolerance zone boundary. A solid modelling approach is then
required to calculate this type of information.
The nonlinear constrained optimization problem is solved using an indirect method which
transforms the suggested formulation to an unconstrained problem, considering an artificial penalty
objective function. The Simplex method of direct search is selected to solve this new formulation:
11- MINIMIZE UCx) = Ureal + Uart ith,
Ureal = F(x),
(2)
where K is a constant increasing the objective function value U(x) when one or more
constraints are unsatisfied.
Therefore the artificial term of the objective function is only active when one or more
constraints are unsatisfied (nviol > 0). This happens when the oriented euclidean distance dcoi is
lower than the minimum prescribed value NEAR_ZERO. The orientation of this distance is
calculated depending on the geometry associated to the design point Di (tolerance zone surface,
toleranced feature surface). For an unconstrained problem, the objective function degenerates to
the real term F(x).
The iterative alignment process is responsible at each iteration for the successive improvement
of the solution, converging at the end of the process, to a feasible solution if the toleranced
features are found to be within their tolerance zones. For efficient datum computation, an
unconstrained Least-Squares formulation is solved using the Newton-Raphson optimization
technique.
The calculation of the oriented euclidean distance refers to the geometry of the related feature
associated to the design point. For planar, cylindrical and spherical features, an exact formulation
is developed, while a conservative spherical approximation is considered for the computation of
the oriented distances associated to sculptured surfaces. As examples, the formulations developed
for cylindrical and spherical surfaces are as follows :
Cylindrical sUrfaces
Based on figure 2,
Co =Di - rey!' ni
So = [ ( R ' M; - Co ) , t ] ,t + Co
with rey! < 0 for concave surfaces (hollow cylinders) or

rey! > 0 for convex surfaces (solid cylinders)
For concave cylindrical surfaces:

dcoi = SIGN [ 1rCYII- I R ' M; - So - (i I ] , I dci I
For convex cylindrical surfaces:
dcoi = SIGN [ I R 'M; - So - (;I -I rey!1 ] , I dci I
where ni is the normal of the surface at point Di, t is the axial vector of the cylinder and (i is
the corrector vector.
Spherical surfaces
Based on figure 3,
with rsph < 0 for concave surfaces (hollow spheres) or

rsph> 0 for convex surfaces (solid spheres)
For concave spherical surfaces:

dcoi = SIGN [ 1rsph 1- I R ' M; - Co - (i I ] , I dci I
For convex spherical surfaces:

dcoi = SIGN [ I R . M; - C. - Cd -I r,ph I ] . I d.d
Figure 2 Oriented distances for cylindrical surfaces.
RoM·1 - CC - (.1
Figure 3 Oriented distances for spherical surfaces.
Sculptured surfaces
For sculptured surfaces (figure 4), a conservative spherical approximation is considered for the
oriented distance calculation of a particular design point belonging to a convex or concave surface
area. A conservative approximation insures that a calculated negative distance, associated to a
particular point, is truly inside its corresponding feature (outside its tolerance). This can be
obtained from an appropriate selection of the sphere radius approximating the sculptured surface
area surrounding the calculated design point Di . For a concave area, the smallest sphere inscribed
within the surface gives the appropriate conservative approximation. The sphere radius then
corresponds to the minimum negative principal curvature value. Similarly, for a convex surface
area, the largest sphere inscribing the surface gives the appropriate conservative approximation.
The sphere radius for this case is then calculated using the minimal positive value of the principal
curvatures (table 1).
A minimum admissible value of curvature Pmin can be specified, from which a planar
approximation is considered for points associated to principal curvatures being inscribed inside the
interval: -Pmin" PhP2 " Pmin. The saddle areas are approximated as concave areas for the spheric
radius calculation. These areas, which the approximations are also conservative, are identified
when PI or P2 < o.
Correction of the alignment process
Within the oriented euclidean distance expressions, a specific corrector vector ( is defined to
consider two types of compensation variables Bri and Ai, both being applied along a specific I'-i
direction and orientation as shown below. This direction is calculated according to the geometry
of the feature (table 2).
(3)
Figure 4 Oriented distances for sculptured surfaces

Table 1 Approximated spheric radius for a sculptured surface
PRINCIPAL SURFACE SPHERE RADIUS

CURVATURES
If:PI < -Pmin or P2 < -Pmin Concave surface rsph = 1 / MIN(p h p:z)
Else if: rsph = I / P
PI> Pmin or P2 > Pmin Convex surface P = MIN(pI> P2) and P > Pmin
Else:
-Pmin S PI , P2 S Pmin ~ Plane surface not applicable
For planar surfaces, the direction vector is equal to the normal plane. For spherical and
cylindrical surfaces, Pi corresponds to the direction separating the transformed measurement point
R . Mdi with the center or the central axis of the design geometry (normal of the surface at the
projected transformed measurement point). All directions are positive when pointing outside the
solid model.
Table 2 Direction for a , corrector
Planar surfaces
Cylindrical surface
concave Pi = (S< - R . MJ / I S< - R . ~ ~
Spherical surface
concave Pi =(R . ~ - Q / I R . ~ - C< I
Sculptured surfaces Spherical approximation
The first compensation refers to the ball radius of the touch probe used for the inspection
process of the part. When appropriate, a compensation Bri corresponding to the probe radius can
be applied to transform the centered ball coordinate measurement data to the normal contact point
coordinates between the surface and the ball probe of the inspection device. The second type of
compensation refers to the alignment of a restricted number of discrete points relative to a
corresponding feature of size, both being of slightly different sizes. In this case, the Ai corrector
is responsible for this compensation during the alignment process. As shown in figure 5 for the
cylindrical feature, the feasible solution resulting from the alignment ofthe data set with respect
to the feature leads to a truly unfeasible solution when considering the actual cylindrical feature
instead of its associated points. In fact for this particular case, a negative deviation of A·cos(6/2)
really results from the alignment instead of the zero deviation calculated by the process. A specific
corrector Ai is then introduced to align the data set relative to a shifted corresponding feature
instead of the nominal one.
For cylindrical and spherical surfaces, the Ai correctors are expressed in terms of the radius and
the maximum distance separating two successive measurement points, assuming geometries of
perfect form. For planar surfaces, such a correction is irrelevant while for sculptured surfaces a
spherical approximation gives the appropriate corrector values in all cases. Referring to figure 5,
r
I
'cos(~}
2
- rl-[r
1 I
'sin(~)ll
2
+ rl - r l
A, = -----=----~-------=---------- (4)
cos(.!)
2
One can see that the corrector leads to zero when the number of points goes to infinity as well
as when radius r2 is equal to r\. In fact, the compensation is a function of the actual feature size
and the point distribution related to the inspected feature.
nominal cylinder
0.57735 . S .
~ circular section of the
. "; ... ~ nominal sphere of radius r 2
! .. SIN -1 (0.57735' SIr 1)

Cylindrical features Spherical features
Figure 5 Compensation associated to CUlVed features

3 APPLICATIONS
In order to illustrate the proposed verification methodology, let's consider the position tolerances
verification associated to the four holes of the plate shown in figure 6. These tolerances, specified
at MMC, are related to a datum feature of size D, which is also specified at MMC. This means a
shift is permissible when the holes depart from their maximum material condition. In certain cases,
for a functional gaging approach, the parts can benefit from a possible shift to avoid any
interference with the dedicated gage. The size of the gage features correspond to the virtual
condition associated to the toleranced features. The proposed technique simulates the functional
gage by calculating the best position of a subset of points with respect to the solid zones generated
from the virtual condition of the toleranced features. Figure 7 presents the cylindrical zones
associated to the position control of the 5 hole plate. All points related to hole 1 to 4, should be
outside their corresponding cylindrical surface defined from the virtual condition of each hole
(cylinder of 0.500 inch (12.7 mm) diameter for all 4 holes). Similarly, the points related to the
datum feature of size D should also be outside their corresponding virtual cylinder of 0.580 inch
(14.73 mm) diameter.
4 x 0 0.535
0.515
---x
0.490~
0.510
-A-
All dimensions in inches
Figure 6 Position tolerances with possible datum shift.

... z
. .... @ .....
. • o. ~
;{
. ..
.
§.. .. ... ~
.. ..
• .
Figure 7 Points to be fitted outside their related cylindrical zone.
An artificial measurement data set has been generated, from which the size and location of each
of the five holes have been established (table 3). The bonus of each feature of size is calculated
from the actual diameter of each hole, while the diametral error of position of the holes are
calculated from the X,Y location of their central axis. As shown in table 4, the total tolerance for
each ofthe holes equals the position tolerance value from which the bonus tolerance of each hole
is added. From these values, every hole is inside its tolerance except hole 2 which has a diametral
position error of 0.0027 inch (0.0689 mm) larger than its tolerance. Without any recourse to a
.datum shift analysis, the part must be rejected. From a paper gage analysis, one possible solution
for the acceptance of the part is a shift of +0.0013 inch (0.0330 mm) about the X axis and a shift
of -0.0012 inch (0.Q305 mm) about the Y axis.
The developed technique has been applied to align the tridimensional data outside each of their
corresponding feature boundary zones (figure 6). To illustrate the role of the Ci corrector used for
the compensation associated to a feature defined from a restricted number of points, only 10
uniformly distributed points are generated for each of the five holes of the plate. Table 5 shows
the results of the iterative constrained alignment process converging to a feasible solution after
only 3 iterations. For this case, the Ai are calculated from the actual radius r1 and the virtual radius
r2 of each hole considering an angle e of36 degrees. Instead of considering the actual size of each
feature, one could specify the LMC radius, which would lead to conservative Ai values.
From these results, all holes are inside their prescribed total tolerance value (table 4) after a
0.0033 inch (0.0838 mm) shift of the datum feature. This shift satisfies the minimum permissible
distance specified through the Ci corrector d..,>A 2=O.00048 inch (0.0122 mm) for all points related
to hole 2. Since the nominal points are calculated from the normal projection of the measurement
points on the CAD related feature, all the oriented euclidean distances in table 5 equal the normal
deviation at the end of the process as well as at the beginning of each iteration.
Table 3 Actual size and Eosition of the holes
POSITION in (mm) SIZE in(mm)

FEATURE
X Y £\(0) 0 BONUS
-1.008 +1.002 0.0165 0.518 0.003

HOLE 1
(-25.603) (25.451) (0.4191) (13.157) (0.076)
+0.991 +1.004 0.0197 0.517 0.002
HOLE 2
(25.171) (25.502) (0.5004) (13.132) (0.051)
+1.004 -1.002 0.0089 0.516 0.001
HOLE 3
(25.502) (-25.451) (0.2261) (13.106) (0.025)
-0.996 -1.006 0.0144 0.519 0.004
HOLE 4
(25.298) (-25.552) (0.3658) (13.183) (0.102)
0.595
HOLE 0 0.000 0.000 0.015
(15.113)
Table 4 Results from EaEer gage
POSITION ERROR (0) in (mm)

TOTAL TOLERANCE (0) without SHIFT with SHIFT
0.018 0.0165 0.012

HOLE 1
(0.457) (0.4191) (0.305)
0.017 0.0197 0.016
HOLE 2
(0.432) (0.5004) (0.406)
0.016 0.0089 0.013
HOLE 3
(0.406) (0.2261) (0.330)
0.019 0.0144 0.0175
HOLE 4
(0.483) (0.3658) (0.4445)
0.Dl5 0.000 0.0035
HOLE 0
(0.381) (0.0889)
Table 5 Results from the constrained alignment algorithm
T: = 0.00154 in
(0.0391 mrn) POSITION ERROR 0
Aj in (mm) (deo)MIN in (mm)
Ty' = -0.00061 in in (mrn)
(-0.0155 mm)
0.00051 0.00243 0.0132

HOLE 1
(0.013) (0.0617) (0.7925)
0.00048 0.00048 0.0164
HOLE 2
(0.0122) (0.0122) (0.4166)
0.00045 0.00193 0.0123
HOLE 3
(0.0114) (0.0490) (0.3124)
0.00054 0.00114 0.0173
HOLE 4
(0.01372) (0.0290) (0.4394)
0.00042 0.00085 0.0033
HOLE 0
(0.0106) (0.0216) (0.0838)
The same alignment has been performed without the use of the (j correctors (Aj=O). The
translational shift then obtained corresponds to T:=0.00116 (0.0295 mrn) and T/=-0.00033 (-
0.0083 mrn), resulting in a O.oI 73 (0.4394 mrn) diametral position error for hole 2 which exceeds
the prescribed total tolerance of 0.0017 (0.0432 mrn) for this hole. For 20 points associated to
each hole, the uncorrected alignment process converges to a feasible solution corresponding to
a shift ofT:=0.00176 (0.0447 mrn) and Ty·=0.00034 (0.0086 mm). These results illustrate the
effect of the number of points as well as of the corrector on the alignment process.
Eventhough this basic example is a two-dimensional datum shift problem, the technique is also
applicable to three-dimensional problems.
4 CONCLUSION
The proposed tridimensional alignment technique based on a constrained optimization process has
been shown to be well suited for the soft gaging of complex mechanical parts. The technique relies
on a solid model of the part and can work with sparse as well as dense metrology data. The
proposed method can solve tolerance analysis cases including a datum shift in space, problems
which currently cannot be solved by any other mean. The optimization technique, which is robust,
can also be applied for the alignment of forgings, castings and rough machined parts with their
corresponding final part geometry.
5 REFERENCES
ASME, Dimensioning and Tolerancing ANSI YI4.5M-1982, New York, 1982.
BARDIS, L., JlNKERSON, R.A, PATRIKALAKIS, N.M., "Localization for Automated
Inspection of Curved Surfaces.", Proc. of the First Int. Offshore and Polar Engineering
Conference., p.16-24, 1991.
CHATELAIN J.-F., FORTIN C., DUPINET E., RIVEST L., MOREL C., Tolerance Verification
Using Three-dimensional Alignment within Solid Tolerance Zones, IMS International
Conference on Rapid Product Development, Stuttgart, Germany, (January 1994).
FENG, S.C., HOPP, T.H., "A Review of Current Geometric Tolerancing Theories and
Inspection Data Analysis Algorithms", NISTIR 4509 National Institute of Standards and
Technology, Gaithersburg, MD, 1991.
FORTIN, C., RIVEST, L., CHATELAIN, J.-F., MOREL, C., DUPINET, E., "Tolerance
Modeling for Rapid Product Development", IMS International Conference on Rapid Product
Development, Stuttgart, Germany, (January 1994).
FUKUDA, M., SHIMOKOHBE, A, "Algorithms for Form Error Evaluation - Methods of
the Minimum Zone and the Least Squares", Proc ofthe Int. Symp. on Metrology for Quality
Control in Production, Tokyo 1984.
LOTZE, W., "Generalized fitting Algorithms in the Coordinate Measuring Techniques in Quality
Control", ACTA IMEKO, YoU, p.279-286, 1982.
MENQ, C.-H., YAU, H.-T., LA!, G.-Y., "Automated Precision Measurement of Surface
Profile in CAD-Directed Inspection", IEEE Transaction on Robotics and Automation,
Vo1.8, No.2, 1992.
MURTY, T.S.R., ABDIN, S.Z., "Minimum Zone Evaluation of Surfaces", IntJournal Mach.
tool des.res., Vo1.20, p.123-136, 1980.
RIVEST, L., Modelisation et analyse tridimensionnelles de tolerances dimensionnelles et
geometriques, Ph.D. thesis, Ecole Polytechnique de Montreal, Canada, 1993.
22
An Evaluation of Geometrical
Errors by Segmentation with
Fitting Form Error Features
Kiwamu Kase, Hiromasa Suzuki and Fumihiko Kimura
Rapid Prototyping System Development Team,
Instrumentation Center, The Institute of Physical and
Chemical Research (RIKEN)
Hirosawa 2-1, Wako-shi, Saitama 351-01, Japan.
Telephone: 81-48-462-1111 ext. 5834. Fax: 81-48-462-4639.
email: kiwamu@ludwig.riken.go.jp
Abstract
Geometrical errors after manufacturing processes have been evaluated by the maximum
width of tolerance zones which envelop the actual features. Details of error shapes \yithin
such zones should however be investigated when the designer studies relation between
errors and functionality of a product and compares and examines manufacturing methods
to realize high accuracy for machined surfaces.
V,Te propose a new method for the evaluation of form errors using a series of points such
as measuring points on the product surfaces. This evaluation consists of two parts: l.
Form error features which consists of error models (Bezier Template) and error conditions
2. Segmentation by fitting of form error features using the Simulated Annealing Method.
Form error models are defined as a series of cubic Bezier curves. A set of Bezier control
points can be used for classifying or characterizing typical geometrical errors. Segmen-
tation and fitting of these form error features are realized by the Simulated Annealing
Method, a statistical optimization method which can be used to search for the global
minimum combinational solution. Our method allows the designer to estimate the effects
of errors on functionality and simulate the behavior of products by using fitting Form
Error Features as a substitute. This leads to functional tolerancing.
Keywords
Geometrical Error Evaluation, Segmentation and Fitting for Scattered Data, Bezier Tem-
plate, Simulated Annealing Method
An evaluation of geometrical errors by segmentation 329
1 INTRODUCTION
Geometrical errors of machined surfaces have been evaluated by the maximum width of
tolerance zones which envelop the actual features. Details of error shapes within such
zones should however be investigated when the designer studies the relation between er-
rors and functionality of a product and compares and examines manufacturing methods
to realize high accuracy for machined surfaces. Especially, in rapid product manufacturing
environments, it is essential for designers and process engineers to look deeply into the
causes of errors and improve nominal shapes and machining conditions swiftly. Transi-
tion from quantitative to qualitative evaluation of errors is desired. The following two
researches are performed in this way.
vVirtz (1993) showed Vectorial Tolerancing to be effective for the total quality control
loop. He used the vectorial tolerance corresponding to each model parameter which con-
sists of surface geometry to control machining factors. He also expanded the vectorial
tolerance model to represent systematic form deviations occuring in manufacturing pro-
cesses such as turning or cutting. Portman (1993) showed the necessity of second or higher
order approximations for machine accuracy calculation in the estimation of machine tool
set-up error and machine setting displacements.
\Ve propose a new method for the qualitative evaluation of form errors using a series
of points such as the measuring points on the product surfaces. This evaluation consists
of two parts:
1. Form error features which consists of error models(Bezier Template) and error concli-
tions
2. Segmentation by fitting of form error features using the Simulat.ed Annealing Method.
Form error models are defined as a series of cubic Bezier curves. A set of Bezier control
points can be used for classifying or characterizing typical geometrical errors. Segmen-
tation and fitting of t.hese form error features are realized by the Simulated Annealing
Method, a stat.istical opt.imizat.ion method which can be used to search for the global
minimum combinational solution.
Our method enables the classification of geometrical errors directly linking up with
working processes, and the resulting evaluation proves useful for improving of machining
methods and for suggesting robust shapes for machining errors.
2 FORM ERROR FEATURES
2.1 Requirements
Systematic errors are generally caused by certain factors, which are likely to be reflect.ed
by the consequent error shapes characteristically. Form Error Features are introduced
to use such characteristics suggestive of error fact.ors for classification, from the scattered
data such as measured points including random errors. The form error features consist of
geometrical error models and collateral conditions. The requirements of the geometrical
error model part are the following two contradicting items:
1. Closeness to the original data

2. Conciseness of representation
To obtain a closer explanation for the original data. more complicated models are neces-
sary at the cost of conciseness. consequently classification becomes difficult. One solution
to satisfy these conflicting requirements is as follows: By selecting proper segmentation
positions in the original scattered data for better fitting, concise models which are suit-
able for classification may be able to represent the whole feature of the given data with
considerable accuracy. The method of segmentation will be described in 3 and 4. In 2.2,
the concise error model suitable for classification is explained.
2.2 Bezier Template and Collateral Conditions

\Ve adopted a series of cubic Bezier curve segments as the approximate error model. The
continuity between two neighboring segments is only Co, the cusps can be represented
easily. A cubic Bezier curve is uniquely represented by four control points (Farin, 1990).
and due to its Affine invariance property. similarity of those curves can be translated to
the similarity of the Bezier control polygon (closeness of the inner angles) equivalently.
In other words, if two Bezier curves are similar to each other, then each corresponding
inner angle of the two control polygons for those curves has close value. Figure 1 shows
two similar cubic Bezier curves and control polygons.
Figure 1 Two similar cubic Btizier curves and control polygons
Collateral conditions include nominal shape. the processing method such as 3-axis
milling, or turning, etc. and machining conditions. These information is necessary for
classification and is applicable to the reasoning of error factors.
The calculation of approximate Bezier curves will be explained in 4.4
3 SIMULATED ANNEALING METHOD

The Simulated Annealing Method (SAM) is a global optimization method using the anal-
ogy of statistical mechanics( the behavior of systems with a high degree of freedom in
thermal equilibrium at a finite temperature). This method has successfully been applied
to some combinatorial optimization problems such as "traveling salesman problem" or

optimal layout design of VLSI device (Kirkpatrick, 1983),
The requirements of this method is as follows:
1. System configuration and neighboring relationship

2. Objective function (system energy)
3. Control parameter T(analogical to temperature in physical annealing), and reducing
(annealing) schedule
Reducing T gradually, one system configuration transits to a new configuration ran-

domly selected from neighboring configurations. with less energy E than before or with
probability exp( -~E). \Vhen T is high, the system can raise the higher energy configura-
tion to avoid local minimums and can also determine the global minimum at a sufficiently
slow pace.
Our problem to be solved is translated to the multi variable combinatorial minimization
problem in search for the set of positions and number of segmentations which minimize
the gap between the given scattered data and error models described in 2. The Simulated
Annealing Method is suitable for our problem.
4 SEGMENTATION AND FITTING METHOD
Figure 2 shows our error feature extraction method. By changing the weight value('\) of
the segmentation number in the system energy, optimized fitting, even with simple curves,
can be obtained in various degrees of fineness .
• 0-
.......
••••••••••••••••••••
' .
original data
•
small •• , ' ...........
+
~ ~ ~
•• .. •• ,. ~
!1 "~#~#~"" ...... -.,

....,~
./\
\
large'..........···..·..···............................, .......~J\
Fitting curves
Figure 2 Fitting of different segmentations with multi-resolutional parameter ,\

4.1 System Configuration and Transition

System configuration( q) is represented as a pair of a segmentation number( S egiV 0) and an
orderly index( i) list of the measured data which shows the segmentation positions( S eg List).
For example, when the total number of the measured points(N) is 255, the system con-
figuration is described as: q = (SegNo, SegList) = (2, (69.213)) and so on. In this case,
the index set of the measured data is divided into three parts: (1, ... ,68),(69, ... ,212),
(213, ... ,255).
Transition between two neighboring configurations is classified into the following three
cases:
Case 1 - no change of SegNo - Transition from q = (SegNo, (i l , ... , iSegNo)) to q' =

(SegN 0, (il + dl, ... ,iSegNo + dS€gNo), where for every j E (1, ... , SegN 0): elj is ran-
domly selected from (-1,0,1).
Case 2 - increment of SegNo by 1 - Transition from q (SegNo, (i l , ... , iSEgNo ))
to q' = (SegNo + 1, (i l , ...• iSegNo+ll), where iSegNo+l IS randomly selected form
(1, ... , N). Put indices in order after that.
Case 3 - decrement of SegNo by 1 - Transition from q = (SegNo,(i1, ... ,iSegNo)) to
q' = (SegN 0-1, (i 1 ,. " , ij-l, ij+),' ..• iSegNo)), where popping out index j is randomly
selected from (l, .... SegNo).
4.2 Objective Function; System Energy
The objective function in our SAM, named energy of system(E) in the analogy of statis-
tical mechanics of molecules, is expressed as the sum of two terms which corresponds to
the two requirements discussed in 2 as follows:
1. The sum of distance between a fitting Bezier curve segment and a measured point
2. Segmentation number(SegNo) with multi-resolutional weight parameter A.
The system energy(E) is expressed in the equation(l). In the equation(l). N j is the

number of measured points, in the jth part. bj(t) is a point on the fitting Bezier curve
segment at parameter t corresponding to the jth part of the measured point set. mjk is
the kth point in the jth part.
SegNo N) k
E = ( L L Ilbj ( N.) - mjkl1 2 ) + A x SegNo (1)
j=O k=O - )
A is set to a lower value when finer segments are needed to obtain more precise fitting
results. Conversely, if more concise and abstract fitting is necessary, A is set to a higher
value.
4.3 Algorithm
An overview of our algorithm is described below:
STEP1: Set T to the initial Temperature To

STEP2: Set system configuration q to the initial configuration; (0, ())

STEP3: Repeat STEP3-1 to STEP3-4 until T < EPS, where EPS is an adequately
small positive value
STEP3-1: Set T to T x 0.9

STEP3-2: Repeat STEP3-3 to STEP3-4 for L times
STEP3-3: Transit q to q' according to the case, randomly selected from Case 1 to
Case 3(See 4.1)
STEP3-4: If q 1= q'
if E(q') < EPS then return q'
else if E(q') < E(q) then set q to q'
else if exp( -(E(q',J.-Elq))) > random(O, 1) then set q to q'
STEP4: Return q
In STEP3-4, random(O, 1) is a function which produces random number from (0,1).
4.4 Approximation by Cubic Bezier Curves
In each step of SAM, the fitting Bezier curve bj(t) is calculated in the following manner:
The first and end control points( bo, b 3 ) are fixed to the the first and end points of the di-
vided part from the measured points, respectively. Intermediate two control points( b 1 , b2 )
are calculated by the following Least Square Method.
Let the divided measured point set be (m s , ... , me), and four Bezier control points be
bo = (:r"ys), b 1 = (xl,yJ),b 2 = (X2,Y2), b3 = (:r"Ye)' bo = (x"Ys) and b3 = (xe,Ye) are
given, b 1 = (:rl,yJ) and b2 = (X2,Y2) are unknown.
A point on the fitting Bezier curve at the parameter t is represented as b(t). The square
sum of the distance between the fitting Bezier curve and measured ponts (S) is defined
by the equation(2), where e and s are the indices of the starting point and the ending
point of the measured points, respectively.
S = ~ Ilb(-J-' ) - m e _ s +JI1 2 (2)

j=O e- s
As S is quadratic, consequently the solutions of the linear simultaneous equations(3)

are what we require here.
(3)
Using MathLink( Mathematica called from C), analytical( symbolical) solutions for Xl ,Yl ,X2,Y2
are obtained. These have rational forms including the 12th degree of N term. General nu-
merical C programs cannot handle them due to numerical errors.
Example
Figure 3 Measuring of sample R==30 cylinder
5 EXAMPLE
Our method was applied to the measured data for three different radius cylinder steel
samples after 3-axis and 5-axis milling. Figure 3 shows the measurement of the sample
workpiece of R=30 cylinder. Figure 4 and 5 show magnified radial differences from the
mean arc by the Least Square Method applied for the measured points at. two different
locations of the R=30 cylinder respectively, at the top of each figure. The middle and
bottom figures of Figure 4 and 5 show the fitted Bezier curves and control polygons
with different parameter values of >., respectively. These control polygons are used for
classification as Bezier Template. Comparing Figures 4 and 5, the two Bezier Templates
with>. = 1 appear different , while the two Bezier Templates with>' = 5 appear more or
less similar and are classified into the same category. Figures 6 and 7 show t.he results of
R=60 cylinder at different measuring positions. When>. is increased, the global structure
shows the same results (two big mountains appear on the both sides of data) . The local
features , however, reflect and change the shape of the middle part of the Bezier Template.
Similarly, Figures 8 and 9 show the results of R=120 cylinder,respectively. The Bezier
Template with>' = 10 at the bottom of Figure 8 and the Bezier Template with >. = 8
at the bottom of Figure 7, have common characteristic, where both radii data look very
similar. In Figure 9, the resulting Bezier Template does not change its structure and shape
in spite of the change of >., suggesting that our SAM found global minimum combination
here.
Figures 10 ,...., 12 show the results for 5-axis milling. Comparing with 3-axis milling cases,
each error feature has less wavings and more smooth curves were fittcd.
';,/ ,"
, -.
.",. -: . . . : ...... . ..,',-: ...... ... ',., .... ., ......
:,.'
. .....-....
" #
..:. ' • •t" .,' . ....:.",.:','--' ' ../....:..:.';.....-.. ~··~v:...· ":~'."

Magnified Difference Data of Radii Magnified DiffereDce Data of Radii
Fitted Bezier Template (lambda-l)

Fitted Bezier Template (lambda-I)
Fitted Bezier Template (lambda-S)

Fitted Bezier Template (lambda-8)
Figure 4 Fitting Results for R=30
Cylinder(3-axis) Figure 6 Fitting Results for R=60
Cylinder( 3-axis)
',":.. ... ..........., :.... '.. ....,:...:. . . . ..·::·t·. .:...-.
..
_',t'", .: :'
.:
Magnified Difference Data of Radii
Ma gnified Differellce Data of Radii
Pitted Bezter Template (lambda-I)

Fitted Bezier Templete (lambda-l)
Fitted Be.ier Template (lambde-8)

Figure 7 Other Results for R=60
Figure 5 Other Results for R=30 Cylinder(3-axis)
Cylinder(3-axis)
:~ y;:'\
,-?':"".
. j:~' .:;..........~.......
~ . ......~/ . - .~'-O:
Magnified Difference Data of Radi i Mag nified Difference Data of Radii
Fitted Bezier Template (lambda-I)
A
Fitted Bezier Template (lambda - I)
Fitted Bezier Template (lambda- IO)

V Fitted Bezier Template (lambda- S)

Figure 8 Fitting Results for R=120 Cylinder{5-axis)
Cylinder{.'1-axis)
Magnified Difference Data of Rad ii
. f>\ I~
V ~\
"
\ f
'v
/V/v,
Fitted Bezier Template (lambda- S)
Fitted Bezier Template (lambda-I) f'
r
V
Fitted Bezier Template (lambda- 8)
Fitted Bnier Template (lambda-IO) Figure 11 Fitting Results for R=60

Cylinder(5-axis)
Figure 9 Other Results for R=120
Cylinder{3-axis)
REFERENCES
Magnified Difference Data of Radii
Wirtz,A. (1993) Vectorial Tolerancing A Basic
Element For Quality Control, Proc. of 3rd
CIRP Seminars on Computer Aided Toler-
ancing, 115-128.
Portman,Y.T.and WEILL,R. (1993) Higher or-
der approximation in accuracy computations
for complex mechanical systems, Proc. of 3rd
CIRP Seminars on Computer Aided Toler-
ancing, 197-212.
Farin.G. (1990) Curves and Surfaces for Com-
puter Aided Geometric Design A Practical
Guide Second Edition, Academic Press.
Kirkpatrick,S. et aI. (1983) Optimization by
Simulated Annealing, Science. 220(4598).
671-680.
Fitted Bezier Template (lambda-lO)

BIOGRAPHY
Kiwamu KASE He is a researcher at RlKEN,
lnst. of Physical & Chemical Research.
Japan. Kase received his doctor degree in
Cylinder( 5-axis)
precision machinery engineering from the
University of Tokyo in 1994. His research
interests include mathematical modelling
of products and its application to rapid
6 CONCLUSION product manufacruring, tolerancing. He
is a member of JSPE( Japan Society for
An evaluation method of geometrical errors Precision Engineers).
by extracting features from scattered data
has been developed. This method is imple-
mented with the Bezier Template suitable QUESTIONS AND ANSWERS
for classification and with the Simulated An-
nealing Method which is able to find the Question from Prof. Kanai: What kind
best combination of positions for segmenta- of applications is supposed for your re-
tion. search?
The Bezier Template is a set of control Speaker: Optimizing in manufacturing pro-
points of cubic Bezier curves which fit the cesses now.
measured data in the Least Square Method. Question from Prof. Fortin: Why did you
The resulting Form Error Features are not use a single and high orderd Bezier
collected as the library for each machining curve for fitting the measurement data?
process and used for the reasoning of error Speaker: Because of simplicity and easi-
factors. It is also applicable to the reasoning ness for classification.
of behavior of products with the use of Form
Error Features as the substitute features of
geometrical errors.
AUTHOR INDEX
Ballot, E. 117 Jonge Poerink, H.J. 47 Pegna, 1. 281

Ballu, A. 31 Portman, V.T. 71
Bourdet, P. 117 Kals, H.J.J. 47
Kanai, S. 233 Riviere, A. 3
Chatelain, J .-F. 310 Kase, K. 145,326
Clement, A. 3 Kato, K. 145 Salomons,O.W. 47
Kimura, F. 145,326 Sandgren, E. 217
Desrochers, A. 17 Kishinami, T. 296 Serre, P. 3
Shibayama, T. 202
Ehrmann, M. 104 Maeda, T. 89 Srinivasan, V. 157
ElMaraghy, H.A. 185 Maranzana, R 17 Suzuki, H. 145, 326
Martinsen, K. 171
Fortin, C, 310 Mathieu, L. 31 Takahashi, H. 233
Fukuda, I. 271 Miura, M. 130 Takamasu, K. 271
Furutani, R 271 Tanaka, F. 296
Nassef, A.O. 185 Tokuoka, N. 89
Hilaire, T.P. 281 Tonshoff, H.K. 104
Hong, J. 271 O'Connor, M. 157
Okamoto, H. 296 van Houten, F.J.A.M. 47
Iannuzzi, M.P. 217 Onozuka, M. 233 van Siooten, F. 47
Ikonomov, P. 296 Ozono, S. 271
Inui, M. 130 Weill, RD. 71
Parks, J .M. 249
KEYWORD INDEX
Accuracy Genetic optimization 217

evaluation 71 Geometric dimensioning and tolerancing 17,
optimization 71 233
Assignment 89 Geometrical
error evaluation 326
Bezier template 326 form 271
model 117
Calibration 202 tolerance 296
Compensation 71 tolerancing 310
Computational
geometry 281 Image metrology 281
metrology 281,296 Interferometry 281
Computer Aided Design (CAD) 17, 233 ISO 157
Computer aided tolerancing 145
Configuration space 130 Japanese Industrial Standard 202
Contact state 145
Contribution analysis 249 Latitude 249
Coordinate Life cycle and quality issues 31
measuring machine 271
metrology 271,296 Measurement 202
Cost-tolerance model 233 of points 310
Mechanism 117
Datum shift 310 assembley 17
Degree of freedom 89 specification 3
Design methodology 249 Metrology 281
Differential coordinate transformation 233 Milling 171
Dimensional chain 71 Mobility degrees 117
Dimensioning 117 Model for tolerances 31
Multivariate statistical process control 171
EXPRESS 3
Extracted feature 271 Optimization 249
Physically based modelling 145
Feature Plastic part tolerance 202
based tolerancing 104 Precision engineering 281
modeling 89 Probabilistic design 249
Form deviation 271 Probability distributions 249
Fringe pattern analysis 281 Product model 31
Functional requirement 3
Reliability 249
Gaussian substitute feature 271 Rigid body motion 145
Genetic algorithms 233 Robust design 249
342 Keyword index
Segmentation and fitting for scattered data 326 optimization 233

Sensitivity analysis 249 representation 47
Simulated annealing method 326 specification 47
Simultaneous engineering 217 synthesis 217, 233, 249
Soft gauging 281 Tolerances in CAD-/CAPP-systems 104
Solid modelling 310 Tolerancing 71,157,202,249
Standards 157 language 31
Statistics 157 synthesis 3
Tridimensional
Taguchi alignment 310
loss function 202 tolerance chains 117
methods 202, 249 TIRS 3
quality engineering 202
Technology and Topologically Related Surfaces Variability analysis 249
(TIRS) 17 Variational modelling 145
Tolerance Vector tolerancing 89
analysis 47,130,217,249 Vectorial tolerancing 171
modeling 130

Computer-Aided Tolerancing

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Computer-Aided Tolerancing

Uploaded by

Copyright:

Available Formats

Computer-aided

The University of Tokyo, Sanjo Conference Hall, Tokyo, Japan

International Program Committee

A Clement (Dassault System, France)

F.Kimura (The University of Tokyo, Japan, Chairman)

Chapman &. HaJJ. 2-6 Boundary Ro .... LorxIon SEI IIHN. UK

First edition 1996

CI 1996 Chapman & Hall

ISBN·13: 978·94.(110·7183·3 c.[SBN.13: 978·94.(I()9.1529.9

Library of Congress Catalog Card Number: 96-8432S

SPrinted on pelTTtanent acid·free text paper, manufll:tlUed in lI:eo.-dano:e with

International Program Committee and Organizing Committee ii

12 Statistical process control using vectorial tolerancing

'Institut Superieur des Materiaux et de la Construction Mecanique.

2 DECLARATIVE INFORMATION MODEL

2Minimum Reference Geometric Element.

Figure 1 EXPRESS-G diagram of Pseudo_TTRS_schema.

Figure 2 Parallelism error between axis from (NF E 23-006).

Description in accordance with the proposed model

TTRS 1 Pseudo TTRS joint

Figure 3 Pseudo_TTRS structure of gearwheel functional requirement.

Figure 4 Parallelepiped tolerance zone for MRGE1_DR from (Chevalier, 1986).

Construction of geometric support TTRS

sup geo 1 Geometric_support { surface ~ « Sl »;

ttrs 1 TTRS{ id ~ « TTRS 1 »;

sup_geo_2 Geometric_support { surface ~ « S2 »;

ttrs 2 TTRS{ id ~ « TTRS 2 »;

Construction of the Pseudo_ TTRS supertype

ps_ttrs_prism pseudo_TTRS-prismatic{ SUBOF(@ps_ttrs); );

Construction of Pseudo_ TTRS.Joint related to the parallelism tolerance

• Pseudo_ TTRS joint construction

3.2 Waterproofness functional requirement

Figure 5 Sealing Ring Waterproofness from (NF E 48-371).

Description in accordance with the proposed model

Pseudo TTRS joint I TTRS I Pseudo TTRS joint 2

Figure 6 Pseudo_TTRS structure of the waterproofuess functional requirement.

Construction ofgeometric support TTRS

ttrs 1 TTRS{ id ~ « TTRS 1 »;

TTRS_cylindrical{ SUBOF(@ttrs 1); };

ttrs 2 TTRS{ id ~ « TTRS 2 »;

ttrs cyl 2 TTRS cylindrical{ SUBOF(@ttrs 2); };

Construction of the Pseudo_ TTRS supertype

ps ttrs cyl = Pseudo TTRS cylindrical { SUBOF(@ps ttrs 1); };

Construction of Pseudo_TTRS.Joint_l related to the " Total Run-Out" tolerance

Construction of Pseudo_ TTRS.Joint_2 related to the "Coaxiality" tolerance

The description of this waterproofuess functional requirement is rather easy. It is realised in a

3.3 Assemblability functional requirement

Figure 7 Assemblability functional requirement from (Nevins, 1989).

Description in accordance with the proposed model

ITRS 1 Pseudo TIRS joint

Construction ofgeometric support TTRS

sup_geo_1 Geometric_support { surface -+ « C1 »;

ttrs 1 TTRS{ id -+ « TTRS_1 »;

ttrs 2 TTRS{ id ~ « TTRS_2 »;

TTRS_cylindrical{ SUBOF(@ttrs 2); };

Construction of the Pseudo_ TTRS supertype

Pseudo_TTRS cylindrical{ SUBOF(@ps_ttrs); };

Construction of Pseudo_TTRSJoint related to the "coaxiality" tolerance

• Pseudo_ TTRS joint construction

A. Desrochers and R. Maranzana

2 THE TTRS REPRESENTATION MODEL

2.1 Mathematical basis

Table 1 Classes of surfaces and TTRS along with their MGDE