Statistical and Computational Techniques

in Manufacturing
J. Paulo Davim (Ed.)

Statistical and Computational

Techniques in Manufacturing

ABC
Editor
J. Paulo Davim
University of Aveiro
Campus Santiago
Department of Mechanical Engineering
Aveiro
Portugal

ISBN 978-3-642-25858-9 e-ISBN 978-3-642-25859-6

DOI 10.1007/978-3-642-25859-6
Springer Heidelberg New York Dordrecht London
Library of Congress Control Number: 2011946096

c Springer-Verlag Berlin Heidelberg 2012

This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of
the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation,
broadcasting, reproduction on microfilms or in any other physical way, and transmission or information
storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology
now known or hereafter developed. Exempted from this legal reservation are brief excerpts in connection
with reviews or scholarly analysis or material supplied specifically for the purpose of being entered
and executed on a computer system, for exclusive use by the purchaser of the work. Duplication of
this publication or parts thereof is permitted only under the provisions of the Copyright Law of the
Publisher’s location, in its current version, and permission for use must always be obtained from Springer.
Permissions for use may be obtained through RightsLink at the Copyright Clearance Center. Violations
are liable to prosecution under the respective Copyright Law.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication
does not imply, even in the absence of a specific statement, that such names are exempt from the relevant
protective laws and regulations and therefore free for general use.
While the advice and information in this book are believed to be true and accurate at the date of pub-
lication, neither the authors nor the editors nor the publisher can accept any legal responsibility for any
errors or omissions that may be made. The publisher makes no warranty, express or implied, with respect
to the material contained herein.

Printed on acid-free paper

Springer is part of Springer Science+Business Media (www.springer.com)
Preface

In recent years, has been increased interest in developing statistical and computational
techniques for applied in manufacturing engineering. Today, due to the great com-
plexity of manufacturing engineering and the high number of parameters used con-
ventional approaches are no longer sufficient. Therefore, in manufacturing, statistical
and computational techniques have achieved several applications, namely, modelling
and simulation manufacturing processes, optimisation manufacturing parameters,
monitoring and control, computer-aided process planning, etc.
The chapter 1 of the book provides design of experiment methods in manufac-
turing (basics and practical applications). Chapter 2 is dedicated to stream-of-
variation based quality assurance for multi-station machining processes (modelling
and planning). Chapter 3 described finite element modelling of chip formation in
orthogonal machining. Chapter 4 contains information on GA-fuzzy approaches
(application to modelling of manufacturing process) and chapter 5 is dedicated of
single and multi-objective optimization methodologies in CNC machining. Chapter
6 described numerical simulation and prediction of wrinkling defects in sheet metal
forming. Finally, chapter 7 is dedicated of manufacturing seamless reservoirs by
tube forming (finite element modelling and experimentation).
The present book can be used as a research book for final undergraduate engi-
neering course or as a topic on manufacturing at the postgraduate level. Also, this
book can serve as a useful reference for academics, manufacturing and computa-
tional sciences researchers, manufacturing, industrial and mechanical engineers,
professional in manufacturing and related industries. The interest of scientific in
this book is evident for many important centers of the research, laboratories and
universities as well as industry. Therefore, it is hoped this book will inspire and
enthuse others to undertake research in this field of statistical and computational
techniques in manufacting.
The Editor acknowledges Springer for this opportunity and for their enthusias-
tic and professional support. Finally, I would like to thank all the chapter authors
for their availability for this work.

Aveiro, Portugal J. Paulo Davim

February 2012
Contents

1 Design of Experiment Methods in Manufacturing: Basics and

Practical Applications ....................................................................................3
Viktor P. Astakhov
1.1 Introduction..............................................................................................1
1.1.1 Design of Experiment as a Formal Statistical Method..................1
1.1.2 Short History.................................................................................3
1.1.3 What Is This Chapter All about?...................................................5
1.2 Basic Terminology...................................................................................5
1.3 Response ..................................................................................................7
1.4 Levels of the Factors................................................................................8
1.5 Experimental Plan – Factorial Experiments...........................................10
1.5.1 Full Factorial Design ..................................................................10
1.6 Resolution Level ....................................................................................22
1.7 More Specialized Designs .....................................................................22
1.7.1 Orthogonal Array and Taguchi method ......................................22
1.7.2 Sieve DOE ..................................................................................25
1.7.3 Split-Plot DOE............................................................................34
1.7.4 Group Method of Data Handling (GMDH) ................................44
1.8 Strategy and Principal Steps in Using DOE...........................................48
References .....................................................................................................52

2 Stream-of-Variation Based Quality Assurance for Multi-station

Machining Processes – Modeling and Planning.........................................55
J.V. Abellan-Nebot, J. Liu, F. Romero Subiron
2.1 Introduction ............................................................................................55
2.2 3D Variation Propagation Modeling.......................................................58
2.2.1 Fundamentals ...............................................................................58
2.2.2 Definition of Coordinate Systems ................................................62
2.2.3 Derivation of the DMVs...............................................................64
2.2.3.1 Fixture-Induced Variations.............................................67
2.2.3.2 Datum-Induced Variations ............................................69
2.2.3.3 Geometric, Kinematic and Thermal-Induced
Variations .......................................................................72
2.2.3.4 Spindle Thermal-Induced Variations .............................75
2.2.3.5 Cutting Force-Induced Variations ..................................76
2.2.3.6 Cutting-Tool Wear-Induced Variations..........................77
2.2.4 Derivation of the SoV Model .......................................................78
VIII Contents

2.3 Process Planning .....................................................................................81

2.3.1 Process Plan Evaluation ...............................................................82
2.3.2 Process Plan Improvement ...........................................................83
2.3.2.1 Sensitivity Indices Related to Fixtures ...........................85
2.3.2.2 Sensitivity Indices Related to Machining Operations ....86
2.3.2.3 Sensitivity Index Related to Stations..............................87
2.3.2.4 Sensitivity Indices Related to Manufacturing Process ...87
2.4 Case Study ..............................................................................................88
2.4.1 Process Plan Evaluation ..............................................................90
2.4.2 Process Plan Improvement ..........................................................91
2.5 Conclusions ............................................................................................93
Appendix 2.1 ..................................................................................................94
Appendix 2.2 ..................................................................................................96
References ......................................................................................................97

3 Finite Element Modeling of Chip Formation in Orthogonal

Machining....................................................................................................101
Amrita Priyadarshini, Surjya K. Pal., Arun K. Samantaray
3.1 Introduction ..........................................................................................101
3.2 Basics of Machining .............................................................................103
3.2.1 Orthogonal Cutting Model .........................................................104
3.2.2 Cutting Forces ............................................................................106
3.2.3 Cutting Temperature ..................................................................109
3.2.4 Chip Morphology .......................................................................110
3.3 Basics of FEM ......................................................................................110
3.3.1 Generalized Steps in FEM..........................................................111
3.3.2 Modeling Techniques................................................................112
3.3.2.1 Type of Approach........................................................113
3.3.2.2 Geometry Modeling ....................................................114
3.3.2.3 Computation Time.......................................................115
3.3.2.4 Mesh Attributes ...........................................................115
3.3.2.5 Locking and Hourglassing...........................................116
3.3.2.6 Solution Methods ........................................................117
3.3.2.7 Mesh Adaptivity..........................................................118
3.4 Brief History of FEM in Machining .....................................................119
3.5 Formulation of Two-Dimensional FE Model for Machining................120
3.5.1 Formulation Steps ......................................................................120
3.5.1.1 Geometric Model, Meshing and Boundary
Conditions ....................................................................120
3.5.1.2 Governing Equations....................................................122
3.5.1.3 Chip Separation Criterion.............................................123
3.5.1.4 Chip-Tool Interface Model and Heat Generation.........125
3.5.1.5 Solution Method...........................................................126
3.5.2 ABAQUS Platform ....................................................................127
Contents IX

3.6 Case Studies in ABAQUS Platform .....................................................130

3.6.1 Case Study I: FE Simulation of Continuous Chip
Formation While Machining AISI 1050 and Its
Experimental Validation ............................................................131
3.6.1.1 Simulation Results........................................................132
3.6.1.2 Model Verification .......................................................132
3.6.2 Case Study II: FE Simulation of Segmented Chip Formation
While Machining Ti6Al4V ........................................................134
3.6.2.1 Role of FE Inputs .........................................................134
3.6.2.2 Significance of FE Outputs ..........................................138
References ....................................................................................................140

4 GA-Fuzzy Approaches: Application to Modeling of Manufacturing

Process .........................................................................................................145
Arup Kumar Nandi
4.1 Introduction ...........................................................................................145
4.2 Fuzzy Logic ...........................................................................................147
4.2.1 Crisp Set and Fuzzy Set ..............................................................147
4.2.2 Fuzzy Membership Function.......................................................148
4.2.2.1 Some Key Properties of Fuzzy Set................................148
4.2.2.2 Various Types of Fuzzy Membership Function
and Its Mathematical Representation ............................149
4.2.3 Fuzzy Set Operators ....................................................................151
4.2.4 Classical Logical Operations and Fuzzy Logical Operations......151
4.2.5 Fuzzy Implication Methods.........................................................153
4.2.6 Decomposition of Compound Rules ...........................................153
4.2.7 Aggregation of Rule....................................................................154
4.2.8 Composition Technique of Fuzzy Relation.................................154
4.2.9 Fuzzy Inferences .........................................................................155
4.2.10 Fuzzification and De-fuzzification ............................................159
4.2.11 Fuzzy Rule-Based Model ..........................................................160
4.2.11.1 Working Principle of a Fuzzy Rule-Based Model.....160
4.2.11.2 Various Types of Fuzzy Rule-Based Model..............161
4.2.11.2.1 Mamdani-Type Fuzzy Rule Based
Model.....................................................162
4.2.11.2.2 TSK-Type Fuzzy Rule-Based Model.....162
4.3 Genetic Algorithm .................................................................................163
4.3.1 Genetic Algorithms and the Traditional Methods .......................164
4.3.2 Simple Genetic Algorithm ..........................................................165
4.3.2.1 Coding and Initialization of GA ....................................166
4.3.2.1.1 Binary Coding..............................................166
4.3.2.1.2 Real Coding .................................................166
4.3.2.1.3 Initialization of GA ......................................166
4.3.2.2 Objective Function and Fitness Function ......................166
4.3.2.3 Selection........................................................................167
4.3.2.4 Crossover (Recombination)...........................................168
X Contents

4.3.2.5 Mutation ........................................................................169

4.3.2.6 Termination of the GA ..................................................169
4.3.3 Description of Working Principle of GA ...................................170
4.4 Genetic Fuzzy Approaches ....................................................................172
4.5 Application to Modeling of Machining Process ....................................178
4.5.1 Modeling Power Requirement and Surface Roughness
in Plunge Grinding Process .........................................................178
4.5.2 Study of Drilling Performances with Minimum Quantity
of Lubricant [14] .........................................................................180
References .....................................................................................................185

5 Single and Multi-objective Optimization Methodologies in

CNC Machining ......................................................................................... 187
Nikolaos Fountas, Agis Krimpenis, Nikolaos M. Vaxevanidis,
J. Paulo Davim
5.1 Introduction .......................................................................................... 187
5.2 Modeling Machining Optimization....................................................... 188
5.2.1 Definition of Optimization ......................................................... 188
5.2.2 Objective Functions in Machining Problems ............................. 189
5.2.3 Mathematical Modeling of Process Parameters
and Objectives............................................................................ 191
5.2.4 Single-Objective and Multi-objective Optimization .................. 192
5.2.4.1 Globally and Locally Pareto Optimal Sets ................... 192
5.2.5 Choosing Optimization Philosophy............................................ 192
5.3 Building Meta-Models for Machining Processes Using
Artificial Neural Networks ................................................................... 193
5.3.1 Basics of Artificial Neural Networks (ANNs) ........................... 193
5.3.2 Machining Data Sets for Training, Validation and
Test of ANNs ............................................................................. 195
5.3.3 Quality Characteristics Predictions Using ANN
Meta-models............................................................................... 196
5.4 Genetic and Evolutionary Algorithms .................................................. 196
5.4.1 Basic Genetic Algorithm Structure ............................................ 197
5.4.1.1 Encoding ...................................................................... 197
5.4.1.2 Selection....................................................................... 198
5.4.1.3 Genetic Operators......................................................... 199
5.4.1.3.1 Crossover .................................................... 199
5.4.1.3.2 Mutation...................................................... 201
5.4.1.4 Advanced Reproduction Models .................................. 203
5.4.1.4.1 Generational Reproduction Model.............. 203
5.4.1.4.2 Steady-State Reproduction Model .............. 203
5.4.2 Evolutionary Algorithms............................................................ 203
5.4.2.1 Evolution Algorithm Structure ..................................... 203
5.4.2.2 Operators in Genetic Algorithms.................................. 206
5.4.2.2.1 Parallelism .................................................. 206
5.4.2.2.2 Migration .................................................... 207
Contents XI

5.5 Variations of Evolutionary Algorithms................................................. 207

5.5.1 Particle Swarm Optimization ..................................................... 207
5.5.2 Simulated Annealing .................................................................. 209
5.5.3 Tabu Search................................................................................ 210
5.5.4 Ant-Colony Optimization........................................................... 211
5.5.5 Tribes ......................................................................................... 212
5.5.6 Hybrids of Evolutionary Algorithms.......................................... 213
5.6 Conclusions .......................................................................................... 214
References .................................................................................................... 215

6 Numerical Simulation and Prediction of Wrinkling Defects

in Sheet Metal Forming..............................................................................219
M.P. Henriques,T.J. Grilo, R.J. Alves de Sousa, R.A.F. Valente
6.1 Introduction and State-of-the-Art ..........................................................219
6.2 Constitutive Isotropic and Anisotropic Models .....................................224
6.3 Benchmarks’ Numerical Analyses.........................................................229
6.3.1 Free Forming of a Conical Cup...................................................229
6.3.2 Flange Forming of a Cylindrical Cup..........................................233
6.4 Results and Discussions – Free - Forming of a Conical Cup.................234
6.5 Results and Discussions – Flange-Forming of a Cylindrical Cup .........245
6.6 Conclusions ...........................................................................................249
References .....................................................................................................250

7 Manufacturing Seamless Reservoirs by Tube Forming:

Finite Element Modelling and Experimentation......................................253
Luis M. Alves, Pedro Santana, Nuno Fernandes, Paulo A.F. Martins
7.1 Introduction ..........................................................................................253
7.2 Innovative Manufacturing Process .......................................................255
7.2.1 Tooling Concept.........................................................................255
7.2.2 Preforms and Mandrels ..............................................................257
7.2.3 Lubrication .................................................................................259
7.3 Mechanical Testing of Materials ..........................................................259
7.3.1 Flow Curve.................................................................................259
7.3.2 Critical Instability Load .............................................................260
7.4 Theoretical and Experimental Background...........................................261
7.4.1 Finite Element Flow Formulation ..............................................261
7.4.2 Friction and Contact...................................................................265
7.4.3 Experimental Development........................................................267
7.5 Mechanics of the Process......................................................................268
7.5.1 Modes of Deformation ...............................................................268
7.5.2 Formability.................................................................................269
7.5.3 Forming Load.............................................................................271
XII Contents

7.6 Applications..........................................................................................271
7.6.1 Performance and Feasibility of the Process................................272
7.6.2 Requirements for Aerospace Applications .................................276
7.7 Conclusions ..........................................................................................279
References ....................................................................................................279

Author Index ......................................................................................................281

Subject Index......................................................................................................283
Contributors

Viktor P. Astakhov Surjya K. Pal

(Chap.1) (Chap.3)–Corresponding author
General Motors Business Department of Mechanical Engineering
Unit of PSMi, 1255 Beach Ct., Saline Indian Institute of Technology,
MI 48176, USA Kharagpur- 721 302, West Bengal, India
Email: [email protected] Email: [email protected]

J. V. Abellan-Nebot Arun K. Samantaray

(Chap.2)-Corresponding author (Chap.3)
Department of Industrial Systems Department of Mechanical Engineering,
Engineering and Design, Indian Institute of Technology,
Universitat Jaume I, Av. Vicent Sos Kharagpur- 721 302, West Bengal, India
Baynat s/n. Castelló, C.P. 12071. Spain Email: [email protected]
Email: [email protected]
Arup Kumar Nandi
J. Liu (Chap.4)
(Chap.2) Central Mechanical Engineering
Department of Systems and Industrial Research Institute (CSIR-CMERI)
Engineering Durgapur-713209, West Bengal, India
The University of Arizona. Tucson, Email: [email protected]
AZ 85704. USA
Email: [email protected] Nikolaos Fountas
(Chap.5)
F. Romero Subiron Department of Mechanical Engineering
(Chap.2) Technology Educators,
Department of Industrial Systems School of Pedagogical and
Engineering and Design, Technological Education (ASPETE),
Universitat Jaume I, Av. Vicent Sos GR-14121, N. Heraklion Attikis, Greece
Baynat s/n. Castelló, C.P. 12071. Spain Email: [email protected]
Email: [email protected]
Agis Krimpenis
Amrita Priyadrashini (Chap.5)
(Chap.3) Department of Mechanical Engineering
Department of Mechanical Engineering, Technology Educators,
Indian Institute of Technology, School of Pedagogical and
Kharagpur, Kharagpur- 721 302, Technological Education (ASPETE),
West Bengal, India GR-14121, N. Heraklion Attikis, Greece
Email: [email protected] Email: [email protected]
XIV Contributors

Nikolaos M. Vaxevanidis R. A. F. Valente

(Chap.5) – Corresponding author (Chap.6) – Corresponding author
Department of Mechanical Engineering GRIDS Research Group,
Technology Educators, Centre for Automation and
School of Pedagogical and Mechanical Technology (TEMA),
Technological Education (ASPETE), Department of Mechanical Engineering,
GR-14121, N. Heraklion Attikis, Greece University of Aveiro,
Email: [email protected] 3810-193 Aveiro,
Portugal
J. Paulo Davim Email: [email protected]
(Chap.5)
Department of Mechanical Engineering, Luís M. Alves
University of Aveiro, (Chap.7)
Campus Santiago, 3810-193, IDMEC, Instituto Superior Técnico,
Aveiro, Portugal Universidade Técnica de Lisboa,
Email: [email protected] Av. Rovisco Pais s/n,
1049-001 Lisboa,
M. P. Henriques Portugal
(Chap.6) Email: [email protected]
GRIDS Research Group,
Centre for Automation and Mechanical Pedro Santana
Technology (TEMA), (Chap.7)
Department of Mechanical Engineering, OMNIDEA, Aerospace Technology
University of Aveiro, and Energy Systems,
3810-193 Aveiro, Portugal Tv. António Gedeão,
Email: [email protected] 9, 3510-017 Viseu,
Portugal
T. J. Grilo Email: [email protected]
(Chap.6)
GRIDS Research Group, Nuno Fernandes
Centre for Automation and (Chap.7)
Mechanical Technology (TEMA), OMNIDEA, Aerospace Technology
Department of Mechanical Engineering, and Energy Systems,
University of Aveiro, Tv. António Gedeão,
3810-193 Aveiro, Portugal 9, 3510-017 Viseu,
Email: [email protected] Portugal
Email:[email protected]
R. J. Alves de Sousa
(Chap.6) Paulo A. F. Martins
GRIDS Research Group, (Chap.7) – Corresponding author
Centre for Automation and Mechanical IDMEC, Instituto Superior Técnico,
Technology (TEMA), Universidade Técnica de Lisboa,
Department of Mechanical Engineering, Av. Rovisco Pais s/n,
University of Aveiro, 1049-001 Lisboa,
3810-193 Aveiro, Portugal Portugal
Email: [email protected] Email: [email protected]
1

Design of Experiment Methods in Manufacturing:

Basics and Practical Applications

Viktor P. Astakhov

General Motors Business Unit of PSMi, 1255 Beach Ct., Saline MI 48176, USA
[email protected]

This chapter discuses fundamentals and important practical aspects of design of

experiments (DOE) in manufacturing. Starting with the basic terminology in-
volved in DOE, it walks a potential reader through major types of DOE suitable
for manufacturing testing. Particular attention is paid to the pre-process deci-
sions and test preparation. The most common DOE as the full and fractional fac-
torial including response surface analysis as well as special DOE as the Taguchi
DOE sieve DOE, split-plot DOE, and group method of data handling (GMDH)
are discussed. The adequate examples are given for some special DOE. The
chapter ends with detailed description of the strategy and principal steps in
DOE.

1.1 Introduction

1.1.1 Design of Experiment as a Formal Statistical Method

Genomics researchers, financial analysts, and social scientists hunt for patterns
in vast data warehouses using increasingly powerful statistical tools. These tools
are based on emerging concepts such as knowledge discovery, data mining, and
information visualization. They also employ specialized methods such as neural
networks, decisions trees, principal components analysis, and a hundred others.
Computers have made it possible to conduct complex statistical analyses that
would have been prohibitive to carry out in the past. However, the dangers of
using complex computer statistical software grow when user comprehension and
control are diminished. Therefore, it seems useful to reflect on the underlying
philosophy and appropriateness of the diverse methods that have been proposed.
2 V.P. Astakhov

This could lead to a better understanding of when to use given tools and me-
thods, as well as contribute to the invention of new discovery tools and refine-
ment of existing ones. This chapter concentrates on one of the most powerful
statistical method known as design of experiments (hereafter, DOE) or experi-
mental design.
DOE is an statistical formal methodology allowing an experimentalist to estab-
lish statistical correlation between a set of input variables with a chosen outcome
of the system/process under study under certain uncertainties, called uncontrolled
inputs. The visualization of this definition is shown in Figure 1.1 where (x1,
x2,...xn) are n input variables selected for the analysis; (y1, y2,...ym)) are m possible
system/process outputs from which one should be selected for the analysis; and
(z1, z2,...zp) are p uncontrollable (the experimentalist has no influence) inputs (of-
ten referred to as noise).
The system/process is designated in Figure 1.1 as black box1, i.e. it is a device,
system or object which can be viewed solely in terms of its input, output, and
transfer (correlation) characteristics without any knowledge of its internal work-
ings, that is, its implementation is "opaque" (black). As a result, any model estab-
lished using DOE is not a mathematical model (although it is expressed using ma-
thematical symbols) but formal statistical or correlation model which does not
have physical sense as a mathematical model derived using equations of mathe-
matical physics with corresponding boundary conditions. Therefore, no attempt
should normally be made to derive some physical conclusions from the model ob-
tained as statistical correlation can be established between an input and the output
which are not physically related.
For example, one can establish a 100% correlation between the rate of grass
growth in his front yard and the level of water in a pond located in the neighbor-
hood by carrying our perfectly correct statistical analysis. However, these two are
not physically related. This physically misleading conclusion is obtained only be-
cause the amount of rainfall that affects both the rate of grass growth and the level
of water was not considered in the statistical analysis. Therefore, the experimen-
talist should pay prime attention to physical meaning of what he is doing at all
stages of DOE in order to avoid physically meaningless but statistically correct re-
sults. In the author’s opinion, this is the first thing that a potential user of DOE
should learn about this method.

1
The modern term "black box" seems to have entered the English language around 1945.
The process of network synthesis from the transfer functions of black boxes can be traced
to Wilhelm Cauer who published his ideas in their most developed form in 1941. Al-
though Cauer did not himself use the term, others who followed him certainly did de-
scribe the method as black-box analysis.
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 3

Fig. 1.1. Visualization of definition of DOE

1.1.2 Short History

For centuries, the common experimentation method was conducted using OFAT
(one-factor-at-a-time) experimentation. OFAT experimentation reached its zenith
with the work of Thomas Edison’s “trial and error” methods [1]. In OFAT, a sin-
gle variable is varied at a time, keeping all other variables in the experiment fixed.
The first factor is fixed as a “good” value, the next factor is examined, and on and
on to the last factor. Because each experimental run considers only one factor,
many runs are needed to obtain sufficient information about the set of conditions
contributing to the problem. This consumes a hefty amount of time and money,
along with running a high risk of error. Another limitation is that when factors
change, they generally change together, so it is impossible to understand the best
solution by pointing to a single, isolated factor. Traditional OFAT experimentation
frequently reduces itself to no methodology whatsoever—just trial and error and a
reliance upon guesswork, experience, and intuition of its users. As a result, OFAT
experiments often unreliable, inefficient, time consuming and may yield false op-
timum condition for the process [2].
Time has changed with growth of manufacturing. Industrial researches no long-
er could afford to experiment in a trial-and-error manner, changing one factor at a
time, the way Edison did in developing the light bulb. A need was felt to develop a
better methodology to run tests where many factors and their combination affect
the outcome.
A far more effective method is to apply a systematic approach to experimenta-
tion, one that considers all factors simultaneously. That approach is called design
of experiments (DOE). A methodology for designing experiments was proposed
by Ronald A. Fisher, in his innovative book The Design of Experiments [3] the
first addition of which was published in 1935 [4]. He was described by Anders
Hald, a Danish statistician who made contributions to the history of statistics, as "a
genius who almost single-handedly created the foundations for modern statistical
science." [5].
4 V.P. Astakhov

Although Fisher is regarded as the founder of the modern methods of design

and analysis of experiments, it would be wrong, however, to image that there had
been no development of such a technique before Fisher. One of the very first
scientific papers to relate to the topic of design of experiments was published in
the statistical journal Biometrika in 1917 [6]. The author, who used the famous
pseudonym "Student", was also responsible for developing “Student's t-test” in
1908. His true identity was Willian Sealey Gosset.
In the 1920s, Fisher developed the methods into a coherent philosophy for ex-
perimentation and he is now widely regarded as the originator of the approach. He
made a huge contribution to statistics in general, and to design of experiments in
particular, from his post at the Rothamsted Experimental Station in Harpenden,
UK. Fisher initiated the principles of DOE and elaborated on his studies of "analy-
sis of variance". He also concurrently furthered his studies of the statistics of small
samples.
Perhaps even more important, Fisher began his systematic approach of the
analysis of real data as the springboard for the development of new statistical me-
thods. He began to pay particular attention to the labor involved in the necessary
computations performed by hand, and developed methods that were as practical as
they were founded in rigor. In 1925, this work culminated in the publication of his
first book, Statistical Methods for Research Workers [7]. This went into many edi-
tions and translations in later years, and became a standard reference work for
scientists in many disciplines.
Although DOE was first used in an agricultural context, the method has been
applied successfully in the military and in industry since the 1940s [8]. Besse Day,
working at the U.S. Naval Experimentation Laboratory, used DOE to solve prob-
lems such as finding the cause of bad welds at a naval shipyard during World War
II. W. Edwards Deming taught statistical methods, including experimental design,
to Japanese scientists and engineers in the early 1950s at a time when “Made in
Japan” meant poor quality. Genichi Taguchi, the most well known of this group of
Japanese scientists, is famous for his quality improvement methods using mod-
ified fractural DOE presently well-known as Taguchi method [9]. One of the
companies where Taguchi first applied his methods was Toyota. Since the late
1970s, U.S. industry has become interested again in quality improvement initia-
tives, now known as “Total Quality” and “Six Sigma” programs [10, 11]. DOE is
considered an advanced method in the Six Sigma programs, which were pioneered
at Motorola and General Electric and nowadays becoming common in the automo-
tive industry [12].
Classical full and fractional factorial designs [13] were already in use at the be-
ginning of the 20th century, while more complex designs, such as D-optimal [14],
came with the arrival of modern computers in the 1970s. It was also around that
time that more advanced regression analysis and optimization techniques were de-
veloped [15].
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 5

Today, corporations across the world are adopting various computer-enhanced

DOEs as a cost-effective way to solve serious problems afflicting their operations.
In the author's opinion, the highest level of DOE is the so-called group of induc-
tive learning algorithms for complex system modeling [16, 17]. The most power-
ful method in this group is the Group Method of Data Handling (GMDH). It is the
inductive sorting-out method, which has advantages in the cases of rather complex
objects, having no definite theory, particularly for the objects with fuzzy characte-
ristics. Although the algorithms of this method are well developed, they are not a
part of standard DOE computer programs.

1.1.3 What Is This Chapter All about?

A great number of papers, manuals, and books have been written on the subject in
general and as related to manufacturing in particular. Moreover, a number of spe-
cial (for example, Satistica), specialized (for example, Minitab) and common (for
example, MS Excel) computer programs are available to assist one to carry out
DOE with endless examples available in the Web. Everything seems to be known,
the terminology, procedures, and analyses are well developed. Therefore, a logical
question why this chapter is needed should be answered.
The simple answer is that this chapter is written from the experimentalist side
of the fence rather than the statistical side used in vast majority of publication
on DOE. As the saying goes “The grass is always greener on the other side of
the fence,” i.e. “statisticians” often do not see many of real-world problems in
preparing proper tests and collecting relevant data to be analyzed using DOE.
Moreover, as mentioned above, the dangers of using complex computer statis-
tical easy-to-use software with colorful user interfaces and looks-nice graphical
representation of the results grow while user comprehension, and thus control
are diminished.
This chapter does not cover basic statistics so that a general knowledge of sta-
tistics including probability concept, regression, and correlation analysis, statisti-
cal distributions, statistical data analysis including survey sampling has to be re-
freshed prior to working with this chapter. Rather it presents the overview of
various DOE to be used in manufacturing commenting its suitability for particular
cases.

1.2 Basic Terminology

Terms commonly used in DOE are [1, 2, 8, 18]:
6 V.P. Astakhov

• Response: A response is the dependent variable that corresponds to the out-

come or resulting effects of interest in the experiment. One or more re-
sponse variables may be studied simultaneously.
• Factors: A factor is a variable contribute to the response.
• Levels: The levels are the chosen conditions of the factor under study.
• Orthogonality can be thought of as factors independence. It in an experi-
ment results in the factor effects being uncorrelated and therefore more eas-
ily interpreted. The factors in an orthogonal experiment design are varied
independently of each other. The main results of data collected using this
design can often be summarized by taking differences of averages and can
be shown graphically by using simple plots of suitably chosen sets of aver-
ages. In these days of powerful computers and software, orthogonality is no
longer a necessity, but it is still a desirable property because of the ease of
visualizing, and thus explaining results.
• Blocks/Blocking: A block is a homogenous portion of the experimental en-
vironment or materials that bears certain variation effects on the re-
sponse(s). A block may be a batch of material supplied by a vendor or
products manufactured in a shift on a production floor. Blocking is a me-
thod for increasing precision by removing the effect of known nuisance fac-
tors. An example of a known nuisance factor is batch-to-batch variability.
In a blocked design, both the baseline and new procedures are applied to
samples of material from one batch, then to samples from another batch,
and so on. The difference between the new and baseline procedures is not
influenced by the batch-to-batch differences. Blocking is a restriction of
complete randomization, since both procedures are always applied to each
batch. Blocking increases precision since the batch-to-batch variability is
removed from the “experimental error.”
• Randomization is a method that protects against an unknown bias distorting
the results of the experiment. An example of a bias is instrument drift in an
experiment comparing a baseline procedure to a new procedure. If all the
tests using the baseline procedure are conducted first and then all the tests
using the new procedure are conducted, the observed difference between
the procedures might be entirely due to instrument drift. To guard against
erroneous conclusions, the testing sequence of the baseline and new proce-
dures should be in random order such as B, N, N, B, N, B, and so on. The
instrument drift or any unknown bias should “average out.”
• Replica/replication is a result/process of a complete repetition of the same
experimental conditions, beginning with the initial setup. Replication in-
creases the signal-to-noise ratio when the noise originates from uncontroll-
able nuisance variables common in real-world manufacturing. Replication
increases the sample size and is a method for increasing the precision of the
experiment. However, the ‘size’ of experiment (cost, time) increases at
the same rate, more difficult to make a great number of test at the same
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 7

conditions. A special design called a Split Plot can be used if some of the
factors are hard to vary.
• Treatments: A treatment is the condition or a factor associated with a spe-
cific level in a specific experiment.
• Experimental units: Experimental units are the objects or entities that are
used for application of treatments and measurements of resulting effects.
• Factorial experimentation is a method in which the effects due to each
factor and to combinations of factors are estimated. Factorial designs are
geometrically constructed and vary all the factors simultaneously and
orthogonally. Factorial designs collect data at the vertices of a cube in
k-dimensions (k is the number of factors being studied). If data are collected
from all of the vertices, the design is a full factorial, requiring 2k runs
provided that each factor is run at two levels.
• Fractional factorial experimentation includes a group of experimental de-
signs consisting of a carefully chosen subset (fraction) of the experimental
runs of a full factorial design. The subset is chosen so as to exploit the spar-
sity-of-effects principle to expose information about the most important
features of the problem studied, while using a fraction of the effort of a full
factorial design in terms of experimental runs and resources. As the num-
ber of factors increases, the fractions become smaller and smaller (1/2, 1/4,
1/8, 1/16, …). Fractional factorial designs collect data from a specific sub-
set of all possible vertices and require 2k-q runs, with 2-q being the fractional
size of the design. If there are only three factors in the experiment, the
geometry of the experimental design for a full factorial experiment requires
eight runs, and a one-half fractional factorial experiment (an inscribed te-
trahedron) requires four runs.

To visualize DOE terminology, process, and its principal stages, consider the fol-
lowing diagram of a cake-baking process shown in Figure 1.2. There are five as-
pects of the process that are analyzed by a designed experiment: factors, levels of
the factors, experimental plan, experiments, and response.

1.3 Response

The response must satisfy certain requirements. First, the response should be the
effective output in terms of reaching the final aim of the study. Second, the re-
sponse should be easily measurable, preferably quantitatively. Third, the re-
sponse should be a single-valued function of the chosen parameters. Unfortu-
nately, these three important requirements are rarely mentioned in the literature
on DOE.
8 V.P. Astakhov

Fig. 1.2. Visualization of DOE terminology, process and its principal stages

1.4 Levels of the Factors

Selections of factors levels is one of the most important yet least formalized, and
thus discussed stage. Each factor selected for the DOE study has a certain global
range of variation. Within this range, a local sub-range to be used in DOE is to be
defined. Practically it means that the upper and lower limits of each included fac-
tors should be set. To do this, one should use all available information such as ex-
perience, results of the previous studies, expert opinions, etc.
In manufacturing, a selection of the upper limit of a factor is a subject of
equipment, physical, and statistical limitations. The equipment limitation means
that the upper limit of the factor cannot be set higher than that allowed by the
equipment used in the tests. In the example considered in Figure 1.2, the oven
temperature cannot be selected higher than that the maximum temperature availa-
ble in the oven. In machining, if the test machine has maximum spindle rpm of
5000 and drills of 3 mm are to be tested that the upper limit of the cutting speed
cannot be more than 47.1 m/min. The physical limitation means that the upper
limit of the factor cannot be set higher than that where undesirable physical trans-
formation may occur. In the example considered in Figure 1.2, the oven tempera-
ture cannot be selected higher than that under which the edges of the cake will be
burned. In machining, the feed should be exceed the so-called breaking feed (the
uncut chip thickness or chip load) [19] under which the extensive chipping or
breakage of the cutting insert occur. The statistical limitation is rarely discussed in
the literature on DOE in manufacturing. This limitation sets the upper limit to the
maximum of a factor for which the automodelity of its variance is still valid. This
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 9

assures much less potential problems with the row variances in the further statis-
tical analysis of the test results.
There is another limitation on the selection of the upper and lower limits of the
factors chosen for DOE. This limitation requires that the factor combinations
should be compatible, i.e. all the required combinations of the factors should be
physically realizable on the setup used in the study. For example, if a combination
of cutting speed and feed results in drill breakage, then this combination cannot be
included in the test. Often, chatter occurs at high cutting regimes that limits the
combinations of the regime parameters. In the example considered in Figure 1.2, it
could happen that for a certain number of eggs and the low limit of the oven tem-
perature, the consistency of the baked cake cannot be achieved no matter what is
the baking time is. Although it sounds simple, factors compatibility is not always
obvious at the stage of the selection of their limits. If some factors incompatibility
is found in the tests, the time and resources spent on this test are wasted and the
whole DOE should be re-planned from scratch.
Mathematically, the defined combination of the selected factors can be thought
of as a point in the multi-dimensional factorial space. The coordinates of this point
are called the basic (zero) levels of the factors, and the point itself is termed as the
zero point [1, 20].
The interval of factor variation is the number which, when added to the zero
level, gives the upper limit and, when subtracted from the zero level, gives the
lower limit. The numerical value of this interval is chosen as the unit of a new
scale of each factor. To simplify the notations of the experimental conditions and
procedure of data analysis, this new scale is selected so that the upper limit cor-
responds to +1, lower to −1 and the basic level to 0. For the factors having conti-
nuous domains, a simple transformation formula is used

xi − xi −0
xi = (1.1)
Δxi

where xi is a new value of the factor i, xi is its true (or real) value of the factor
(the upper or lower limit), xi −0 is the true (real) value of the zero level of the fac-
tor, Δxi is the interval of factor variation (in true (real) units) and i is the number
of the factor.
In the example considered in Figure 1.2, assume that the maximum oven
temperature in the test is selected to be x A− max = 200o C and the minimum
temperature is x A− min = 160o C . Obviously, x A− 0 = 180o C so that Δx A = 20o C .
Than the maximum of factor A in the scale set by Eq. (1.1) is calculated
as x A(max) = ( 200 − 180 ) / 20 = +1 while its minimum is calculated as
x A (min) = (160 − 180 ) / 20 = −1 . The other factors in the test shown in Figure 1.1 are
10 V.P. Astakhov

set as follows: xB − min = 1cup , xB − max = 2 cup ; xC − min = 2 cup , xC −max = 3cup ;
xD − min = 1egg , xD − min = 3eggs .
Equation (1.1) is used in practically any type of DOE for direct and reverse
transformations of the factors. The latter is needed to convert the correlation mod-
el obtained as a result of using DOE in the real scale of factors. Unfortunately, this
is routinely forgotten step in the representation of DOE results.

1.5 Experimental Plan – Factorial Experiments

A very useful class of designed experiments is a range of factorial experiments, in

which the treatments are combinations of levels of several factors. These are used
in many applications of experimental design, but especially in technological expe-
riments, where the factors might be, for example, time, concentration, pressure,
temperature, etc. As mentioned above, it is very common to use a small number of
levels for each of the factors, often just two levels, in which case a design with k
treatment factors has 2k treatments and is called a 2k factorial design. As an exam-
ple, in a experiment, if there were 10 systematic parameters then a full 210 fac-
torial might have each systematic parameter set at ±1σ; of course in this case it
would be usual as well to have one or more runs at the central 'mean value' or
'best guess' of all the systematics.
Factorial designs, including fractional factorials, have increased precision over
other types of designs because they have built-in internal replication. Factor ef-
fects are essentially the difference between the average of all runs at the two levels
for a factor, such as “high” and “low.” Because each factor is varied with respect
to all of the factors, information on all factors is collected by each run. In fact,
every data point is used in the analysis many times as well as in the estimation of
every effect and interaction. Additional efficiency of the two-level factorial design
comes from the fact that it spans the factor space, that is, puts half of the design
points at each end of the range, which is the most powerful way of determining
whether a factor has a significant effect.

1.5.1 Full Factorial Design

Referring to Figure 1.2, Table 1.1 gives the settings for a 24 factorial experiment;
usually the order of the runs would be randomized, but the structure of the expe-
riment is easier to see in the un-randomized form. The real variables are added to
the table (called the design matrix which defines all the runs and levels in the ex-
periment) for clarity – normally the design matrix includes only the code variables
for simplicity.
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 11

Table 1.1. A 24 factorial design of 16 runs, with the response labeled according to conven-
tional notation for the factor levels. Real variables are added for clarity.

Coded variables Real variables

run response
A B C D Temp Sugar Flour Egg
1 −1 −1 −1 −1 160 1 2 1 y(1)
2 −1 −1 −1 +1 160 1 2 3 yd
3 −1 −1 +1 −1 160 1 3 1 yc
4 −1 −1 +1 +1 160 1 3 3 ycd
5 −1 +1 −1 −1 160 2 2 1 yb
6 −1 +1 −1 +1 160 2 2 3 ybd
7 −1 +1 +1 −1 160 2 3 1 ybc
8 −1 +1 +1 +1 160 2 3 3 ybcd
9 +1 −1 −1 −1 200 1 2 1 ya
10 +1 −1 −1 +1 200 1 2 3 yad
11 +1 −1 +1 −1 200 1 3 1 yac
12 +1 −1 +1 +1 200 1 3 3 yacd
13 +1 +1 −1 −1 200 2 2 1 yab
14 +1 +1 −1 +1 200 2 2 3 yabd
15 +1 +1 +1 −1 200 2 3 1 yabc
16 +1 +1 +1 +1 200 2 3 3 yabcd

The run called '1', for example, has all four factors set to their low level, whe-
reas the run called '2', has factors A, B, and C set to their low level and D set to
its high level. Note that the estimated effect in going from the low level of A,
say, to the high level of A, is based on comparing the averages of 8 observations
taken at the low level with 8 observations taken at the high level. Each of these
averages has a variance equal to 1/8 the variance in a single observation, or in
other words to get the same information from an OFAT design one would need
8 runs with A at -1σ and 8 runs with A at +1σ, all other factors held constant.
Repeating this for each factor would require 64 runs, instead of 16. The balance
of the 24 design ensures that one can estimate the effects for each of the four
factors in turn from the average of 8 observations at the high level compared
to 8 observations at the low level: for example the main effect of D is estimated
by
12 V.P. Astakhov

⎛ yabcd − yabc + yabd − yab + acd − yac + yad − ya + ybcd − ⎞

⎜⎜ ⎟⎟ 8 (1.2)
y + ybd − yb + ycd − yc + yd − y(1)
⎝ bc ⎠

and similar estimates can be constructed for the effects of B and C.

Note that by constructing these four estimates, four linear combinations of 16
observations were used. One linear combination, the simple average, is needed to
set the overall level of response, leaving 11 linear combinations not yet used to es-
timate anything. These combinations are in fact used to estimate the interactions
of various factors, and the full set of combinations is given by the set of signs in
Table 1.2.

Table 1.2. The 24 factorial showing all of the interaction effects

run A B C D AB AC AD BC BD CD ABC ABD ACD BCD ABCD

1 −1 −1 −1 −1 +1 +1 +1 +1 +1 +1 −1 −1 −1 −1 +1
2 −1 −1 −1 +1 +1 +1 −1 +1 −1 −1 −1 +1 +1 +1 −1
3 −1 −1 +1 −1 +1 −1 +1 −1 +1 −1 +1 −1 +1 +1 −1
4 −1 −1 +1 +1 +1 −1 −1 −1 −1 +1 +1 +1 −1 −1 +1
5 −1 +1 −1 −1 −1 +1 +1 −1 −1 +1 +1 +1 −1 +1 −1
6 −1 +1 −1 +1 −1 +1 −1 −1 +1 −1 +1 −1 +1 −1 +1
7 −1 +1 +1 −1 −1 −1 +1 +1 −1 −1 −1 +1 +1 −1 +1
8 −1 +1 +1 +1 −1 −1 −1 +1 +1 +1 −1 −1 −1 +1 −1
9 +1 −1 −1 −1 −1 −1 −1 +1 +1 +1 +1 +1 +1 −1 −1
10 +1 −1 −1 +1 −1 −1 +1 +1 −1 −1 +1 −1 −1 +1 +1
11 +1 −1 +1 −1 −1 +1 −1 −1 +1 −1 −1 +1 −1 +1 +1
12 +1 −1 +1 +1 −1 +1 +1 −1 −1 +1 −1 −1 +1 −1 −1
13 +1 +1 −1 −1 +1 −1 −1 −1 −1 +1 −1 −1 +1 +1 +1
14 +1 +1 −1 +1 +1 −1 +1 −1 +1 −1 −1 +1 −1 −1 −1
15 +1 +1 +1 −1 +1 +1 −1 −1 −1 −1 −1 −1 −1 −1 −1
16 +1 +1 +1 +1 +1 +1 +1 +1 +1 +1 +1 +1 +1 +1 +1

For example, the interaction of factors A and B is estimated by the contrast giv-
en by the fourth column of Table 1.2

⎧⎪ yabcd + y abc + yabd + yab − ybcd − ybc − ybd − yb − ⎫⎪

⎨
( ) ⎬ 8 (1.3)
⎪⎩ yacd + yac + yad + y a − ycd − y c − yd − y(1) ⎭⎪

which takes the difference of the difference between responses with A at high lev-
el and A at low level with the difference between responses with B at high level
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 13

and B at low level. The column of signs in Table 1.2 for the interaction effect AB
was obtained simply by multiplying the A column by the B column, and all the
other columns are similarly constructed.
This illustrates two advantages of designed experiments: the analysis is very
simple, based on linear contrasts of observations, and as well as efficiently esti-
mating average effects of each factor, it is possible to estimate interaction effects
with the same precision. Interaction effects can never be measured with OFAT de-
signs, because two or more factors are never changed simultaneously.
The analysis, by focusing on averages, implicitly assumes that the responses are
best compared by their mean and variance, which is typical of observations that
follow a Gaussian distribution. However the models can be extended to more gen-
eral settings [1].
The discussed factorial DOE allows accurate estimation of all factors involved
and their interactions. However, the cost and time need for such a test increase
with the number of factors considered. Normally, any manufacturing test includes
a great number of independent variables. In the testing of drills, for example, there
are a number of tool geometry variables (the number of cutting edges, rake angles,
flank angles, cutting edge angles, inclination angles, side cutting edge back taper
angle, etc.) and design variables (web diameter, cutting fluid hole shape, cross-
sectional area and location, profile angle of the chip removal flute, length of the
cutting tip, the shank length and diameter, etc.) that affect drill performance.
Table 1.3 shows the number of runs needed for the so-called full factorial DOE
considered above. In laboratory conditions, at least three repetitions for the same
run are needed while if the test is run in the shop floor conditions then at least 5
repetitions should be carried out in randomized manner. The poorer the test condi-
tions, the greater the number of uncontrollable variables, the greater number of
repetitions at the same point of the design matrix is needed to pass statistical
analysis of the obtained experimental data.
Therefore, there is always a dilemma. On one hand, to keep the costs and time
spent on the test at reasonable level it is desirable to include into consideration
only a limited number of essential factors carefully selected by the experts. On the
other hand, if even one essential factor is missed, the final statistical model may
not be adequate to the process under study. Unfortunately, there is no way to jus-
tify the decisions made at the preprocess stage about the number of essential vari-
ables prior the tests. If a mistake is made at this stage, it may show up only at the
final stage of DOE when the corresponding statistical criteria are examined. Obvi-
ously, it is too late then to correct the test results by adding the missed factor. The
theory of DOE offers a few ways to deal with such a problem [20, 21].
The first relies on the collective experience of the experimentalist(s) and the re-
search team in the determination of significant factors. The problem with such an
approach is that one or more factors could be significant or not, depending on the
particular test objectives and conditions. For example, the backtaper angle in drills
is not a significant factor in drilling cast irons, but it becomes highly significant in
machining titanium alloys.
14 V.P. Astakhov

A second way is to use screening DOE. This method appears to be more prom-
ising in terms of its objectivity. A screening DOE is used when a great number of
factors are to be investigated using a relatively small number of tests. These kinds
of tests are conducted to identify the significant factors for further full factorial
analysis. The most common type of screening DOE is the so-called the fractional
factorial DOE.

Table 1.3. Two-level designs: minimum number of runs as a function of number of factors
for full factorial DOE

Number of Number of Number of repetitions

factors runs 3 5

1 2 6 10
2
2 4=2 12 20
3 8 = 23 24 40
4 16 = 24 48 80
5 32 = 25 96 160
6
6 64 = 2 192 320
7 128 = 27 384 640
8 256 = 28 768 1280
9 512 = 29 1536 2560
10
10 1024 = 2 3072 5120

Very often, especially in manufacturing settings, the factors correspond to un-

derlying quantitative variables, and the levels, denoted ±1, are codes for particular
numerical values: temperatures at 80 and 100 degrees, for example. In such cases
the choice of factor levels involves both subject matter expertise, and at least in
the early stages, considerable guesswork. As well, the goal of the experiment
might not be to compare responses at different factor settings, but to find the com-
bination of factor settings that leads to minimum or maximum response.
Factorial designs adapted to these conditions are called response surface de-
signs (RSM). The basic idea is that the response y is a continuous function of
some input variables x1, x2, and so on, and factorial designs are used sequentially
to explore the shape of the response surface. Sequential experimentation in the re-
levant range of x−space usually begins with a screening design, to quickly assess
which of several factors have the largest effect on the response. Then second stage
is a factorial design at new values of the underlying variables, chosen in the direc-
tion of increasing (or decreasing) response. Near the maximum additional factor
levels are added, to model curvature in the response surface.
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 15

The first goal for RMS is to find the optimum response. When there is more
than one response then it is important to find the compromise optimum that does
not optimize only one response. When there are constraints on the design data,
then the experimental design has to meet requirements of the constraints. The
second goal is to understand how the response changes in a given direction by ad-
justing the design variables. In general, the response surface can be visualized
graphically. The graph is helpful to see the shape of a response surface; hills, val-
leys, and ridge lines.
A plot showing how the response changes with respect to changes in the factor
levels is a response surface, a typical example of which is shown in Figure 1.3a as
a three-dimensional perspective plot. In this graph, each value of x1 and x2 gene-
rates a y−value. This three-dimensional graph shows the response surface from
the side and it is called a response surface plot.
Sometimes, it is less complicated to view the response surface in two-
dimensional graphs. The contour plots can show contour lines of x1 and x2 pairs
that have the same response value y. An example of contour plot is as shown in
Figure 1.3b.
In order to understand the surface of a response, graphs are helpful tools. But,
when there are more than two independent variables, graphs are difficult or almost
impossible to use to illustrate the response surface, since it is beyond 3-dimension.
RSM methods are utilized in the optimization phase. Due to the high volume of
experiments, this phase focuses on a few highly influential variables, usually 3 to
5. The typical tools used for RSM are central composite design (CCD) and the
Box-Behnken design (BBD) [1]. With the aid of software, the results of these
complex designs are exhibited pictorially in 3D as ‘mountains’ or ‘valleys’ to illu-
strate performance peaks.

Fig. 1.3. Graphical representation of the outcome of RSM: (a) response surface plot, (b)
control plot
16 V.P. Astakhov

A class of two-level factorial designs which allow reduction of the number of

runs while still allowing estimation of important effects is the family of two-level
fractional factorial designs. The construction of two-level fractional factorial de-
signs is based on a structure of two-level full factorial designs.
In practice, interactions involving three or more factors are often negligible.
One can take advantage of the fact that some of the effects estimable in the full
factorial are negligible to reduce the number of runs required for the experiment.
In the above-considered example, one may assume that the 3-way and 4-way inte-
ractions are negligible. This information can then be used either to study addition-
al variables without increasing the number of runs or to reduce the number of runs
in the present study.
For example, a fifth factor E, for example butter (measured in cups being
0.5 cup – low level, 1.5 cup – the upper level) can be introduced to the 24 fac-
torial study (Table 1.2). A full factorial design with 5 variables requires 25 = 32
runs. Instead of adding a fifth column into the design matrix represented by
Table 1.2 which would require twice as many runs, the new fifth variable is as-
signed to the column for ABCD interaction to obtain its setting. That is, in the
first run E would be set to its high level, in the second run to its low level, in
the third run to its low level, and so on, following the pattern of +1 and −1 in
the last column of Table 1.2. The resulting contrast ( y( ) − y
1 a )
− yb + yab ± ... 8
is estimating the main effect of factor E (i.e. the difference between responses
on the high level of E to the low level of E), but it is also estimating the ABCD
interaction: these two effects are completely aliased. The working assumption
is that the ABCD interaction is likely to be very small, so any observed effect
can be attributed to E. The main effects of A, B, C, and D are estimated as be-
fore, and one now has information on 5 factors from a 16 run design. However
all the main effects are aliased with 4 factor interactions: for example A is
aliased with BCDE, B with ACDE, and so on. Further, all 2 factor interactions
are aliased with 3 factor interactions. Again, the working assumption is typical-
ly that any observed effect is more likely to be due to a 2 factor interaction than
a 3 factor interaction.
Table 1.4 gives the new design matrix. This design is called a half-fraction of
32-run full factorial design. Because it has five variables, with one of them de-
fined in terms of other 4, it is a 25-1 fractional factorial design. Note that the num-
ber of runs is equal to 25-1 = 24 = 16, half the number required for a full 5-factor
factorial test.
Table 1.5 shows the complete aliasing pattern for the 25-1 fractional factorial
design from setting E = ABCD. I is used to represent the identity element and con-
sists of a column of 16 plus signs. Such a column is normally is a part of any de-
sign matrix and used to estimate the response average in the regression analysis of
the data.
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 17

Table 1.4. Design matrix for 25-1 fractional factorial DOE

Coded variables Real variables

run
A B C D E Temp Sugar Flour Egg Butter
1 −1 −1 −1 −1 +1 160 1 2 1 1.5
2 −1 −1 −1 +1 −1 160 1 2 3 0.5
3 −1 −1 +1 −1 −1 160 1 3 1 0.5
4 −1 −1 +1 +1 +1 160 1 3 3 1.5
5 −1 +1 −1 −1 −1 160 2 2 1 0.5
6 −1 +1 −1 +1 +1 160 2 2 3 1.5
7 −1 +1 +1 −1 +1 160 2 3 1 0.5
8 −1 +1 +1 +1 −1 160 2 3 3 0.5
9 +1 −1 −1 −1 −1 200 1 2 1 0.5
10 +1 −1 −1 +1 +1 200 1 2 3 1.5
11 +1 −1 +1 −1 +1 200 1 3 1 1.5
12 +1 −1 +1 +1 −1 200 1 3 3 0.5
13 +1 +1 −1 −1 +1 200 2 2 1 1.5
14 +1 +1 −1 +1 −1 200 2 2 3 0.5
15 +1 +1 +1 −1 −1 200 2 3 1 0.5
16 +1 +1 +1 +1 +1 200 2 3 3 1.5

Table 1.5. Aliasing pattern for 25-1 fractional factorial DOE

I = ABCDE
A = BCDE
B = ACDE
C = ABDE
D = ABCE
E = ABCD
AB = CDE
AC = BDE
AE = BCD
BC = ADE
BD = ACE
BE = ACD
CD = ABE
CE = ABD
DE = ABC
18 V.P. Astakhov

It follows from Table 1.5 that each main effect and 2-factor interaction is
aliased with a 3 or 4-way interaction which was assumed to be negligible. Thus,
ignoring effects expected to be unimportant, one can still obtain estimates of all
main effects and 2-way interactions. As a result, fractional factorial DOE is often
used at the first stages of the study as a screening test to distinguish obtain signifi-
cant factors, and thus to reduce the number of factors used the full factorial tests.
The process of fractionalization can be continued; one might for example as-
sign a new factor F, say, to the ABC interaction (which is aliased with DE), giving
a 26−2 design, sometimes called a 1/4 fraction of a 26. This allows one to assess
the main effect of 6 factors in just 16 runs, instead of 64 runs, although now some
2 factor interactions will be aliased with each other.
There are very many variations on this idea; one is the notion of a screening de-
sign, in which only main effects can be estimated, and everything else is aliased.
The goal is to quickly assess which of the factors is likely to be important, as a
step in further experimentation involving these factors and their interactions.
Table 1.6 shows an 8 run screening design for 7 factors. The basic design is the 23
factorial in factors A, B, and C shown in the first 3 columns; then 4 new factors
have been assigned to the columns that would normally correspond to the interac-
tions BC, AC, AB, and ABC.

Table 1.6. A screening design for 7 factors in 8 runs, built from a 23 factorial design

A B C D E F G
1 −1 −1 −1 +1 +1 +1 −1
2 −1 −1 +1 −1 −1 +1 +1
3 −1 +1 −1 −1 +1 −1 +1
4 −1 +1 +1 +1 −1 −1 −1
5 +1 −1 −1 +1 −1 −1 +1
6 +1 −1 +1 −1 +1 −1 −1
7 +1 +1 −1 −1 −1 +1 −1
8 +1 +1 +1 +1 +1 +1 +1

Any modern DOE software (for example, Statistica, Design-Ease® and Design-
Expert®, Minitab, etc.) can generate design matrix for fractional factorial DOE
and provide a table of interactions similar to that shown in Table 1.5.
As mentioned above, the objective of DOE is to find the correlation between
the response and the factors included. All factors included in the experiment are
varied simultaneously. The influence of the unknown or non-included factors is
minimized by properly randomizing the experiment. Mathematical methods are
used not only at the final stage of the study, when the evaluation and analysis of
the experimental data are conducted, but also though all stages of DOE, i.e. from
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 19

the formalization of priory information till the decision making stage. This allows
answering of important questions: “What is the minimum number of tests that
should be conducted? Which parameters should be taken into consideration?
Which method(s) is (are) better to use for the experimental data evaluation and
analysis?” [20, 21]. Therefore, one of the important stages in DOE is the selection
of the mathematical (correlation) model. An often quoted insight of George Box
is, “All models are wrong. Some are useful”[22]. The trick is to have the simplest
model that captures the main features of the process.
Mathematically, the problem of DOE can be formulated as follows: define the
estimation E of the response surface which can be represented by a function

E { y } = φ ( x1 , x2 ,..., xk ) (1.4)

where y is the process response (for example, cutting temperature, tool life, sur-
face finish, cutting force, etc.), xi , i = 1, 2,...k are the factors varied in the test (for
example, the tool cutting edge angle, cutting speed, feed, etc.).
The mathematical model represented by Eq.(1.4) is used to determine the gra-
dient, i.e., the direction in which the response changes faster than in any other.
This model represents the response surface, which is assumed to be continuous,
two times differentiable, and having only one extreme within the chosen limits of
the factors.
In general, a particular kind of the mathematical model is initially unknown due
to insufficient knowledge on the considered phenomenon. Thus, a certain approx-
imation for this model is needed. Experience shows [21] that a power series or po-
lynomial (Taylor series approximations to the unknown true functional form of the
response variable) can be selected as an approximation

p p p p p
y = β0 + ∑∑ βij xi x j + ∑∑∑ βijk xi x j xk + ...
i =1 j =1 i =1 j =1 k =1
(1.5)
i≠ j i ≠ j ≠k

where β0 is the overall mean response, βi is the main effect of the factor (i = 1,2, ...
, p), βij is the two-way interaction effect between the ith and jth factors, and βijk is
the three-way interaction effect between the ith, jth, and kth factors.
A general recommendation for setting the factor ranges is to set the levels far
enough apart so that one would expect to see a difference in the response. The use
of only two levels seems to imply that the effects must be linear, but the assump-
tion of monotonicity (or nearly so) on the response variable is sufficient. At least
three levels of the factors would be required to detect curvature.
Interaction is present when the effect of a factor on the response variable
depends on the setting level of another factor. Two factors are said to have interac-
tion when the effect of one factor varies under the different levels of another fac-
tor. Because this concept is very important in DOE, consider an example to clarify
20 V.P. Astakhov

the concept. Suppose that Factor A is the silicon content is a heat-treatable alumi-
num. Suppose that factor B is the heat treating temperature. In this case the re-
sponse values may represent the hardness. Table 1.7 shows the response depend-
ing upon the factors' setting. A plot of the data (Figure 1.4a) shows how response
is a function of factors A and B. Since the lines are almost parallel, it can be as-
sumed that little interaction between the variables occurs.
Consider another set of data as shown in Table 1.8 In this instance, the response
is the relative wear rate, factor A is the shaft speed, and factor B may be the type
of lubricant. Again, a plot of the data (Figure 1.4b) shows how response depends
on the factors. It can be seen that B2 is a better lubricant at low speed but not at
high speed. The crossing of the lines indicates an interactive effect between the
factors. In this case, factors A and B are interrelated, i.e. not independent.

Table 1.7. A factorial experiment with no factors interaction

Factor B1 B2
A1 20 30
A2 40 52

Fig. 1.4. Concept of interaction: (a) no interaction of factors A and B, (b) interaction of
factors A and B

Table 1.8. A factorial experiment with factors interaction

Factor B1 B2
A1 1.5 2.0
A2 3.0 0.8
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 21

The βij terms in Eq.(1.5) account for the two-way interactions. Two-way inte-
ractions can be thought of as the corrections to a model of simple additivity of the
factor effects, the model with only the βi terms in Eq.(1.5). The use of the simple
additive model assumes that the factors act separately and independently on the
response variable, which is not always a very reasonable assumption.
The accuracy of such an approximation would depend upon the order (power)
of the series. To reduce the number of tests at the first stage of experimental
study, a polynomial of the first order or a linear model is sufficiently suitable.
Such a model is successfully used to calculate the gradient of the response, thus,
to reach the stationary region. When the stationary region is reached then a
polynomial containing terms of the second, and sometimes, the third order may
be employed.
Experience shows [23, 24] that a model containing linear terms and interactions
of the first order can be used successfully in metal cutting. Such a model can be
represented as

y = β0 + ∑β x + ∑β
i
i i
ij
ij i x xj (1.6)

The coefficients of Eq.(1.6) are to be determined from the tests. Using the experi-
mental results, one can determine the regression coefficients b1 , bi , and bij ,
which are estimates for the theoretical regression coefficients β1 , β i , and βij .
Thus, the regression equation constructed using the test results has the following
form

E { y} = y = b0 + ∑b x + ∑b x x
i
i i
ij
ij i j
(1.7)

where y is the estimate for E { y} .

The problem of mathematical model selection for the object under investigation
requires the formulation of clear objective(s) of the study. This problem occurs in
any study, but the mathematical model selection in DOE requires the quantitative
formulation of the objective(s) called the response, which is the result of the
process under study or its output. The process under study may be characterized
by several important output parameters but only one of them should be selected as
the response.
As mentioned above, the response must satisfy certain requirements. First, the re-
sponse should be the effective output in terms of reaching the final aim of the study.
Second, the response should be easily measurable, preferably quantitatively. Third,
the response should be a single-valued function of the chosen parameters. Unfortu-
nately, these three important requirements are rarely mentioned in the literature on
DOE.
22 V.P. Astakhov

Most statistical software (for example, Statistica, Design-Ease® and Design-

Expert®, Minitab, etc.) can estimate the significant coefficients of the chosen
model and verify the model adequacy using standard statistical procedures using
ANOVA (analysis of variance methods) [1]. Moreover, they can deduce from the
specification of the model which effects are aliased and in some packages and plus
it is relatively easy to produce a graphical display of the estimated effects that al-
lows one to assess which effects are nonzero, at least in part so that the 'nearly
zero' effects can be pooled to estimate the error.

1.6 Resolution Level

Experimental designs can be categorized by their resolution level. A design with
a higher resolution level can fit higher-order terms in Eq.(1.5) than a design with
a lower resolution level. If a high enough resolution level design is not used,
only the linear combination of several terms can be estimated, not the terms sep-
arately. The word “resolution” was borrowed from the term used in optics. Res-
olution levels are usually denoted by Roman numerals, with III, IV, and V being
the most commonly used. To resolve all of the two-way interactions, the resolu-
tion level must be at least V [8]. Four resolution levels and their meanings are
given in Table 1.9.

Table 1.9. Resolution levels and their meaning

Resolution level Meaning

II Main effects are linearly combined with each other (βi + βj).

III Main effects are linearly combined with two-way interactions (βi + βjk).

IV Main effects are linearly combined with three-way interactions (βi + βjkl) and
two-way interactions with each other (βij + βkl).

V Main effects and two-way interactions are not linearly combined except with
higher-order interactions (βi + βjklm) and (βij + βklm).

1.7 More Specialized Designs

1.7.1 Orthogonal Array and Taguchi Method

The 8 run screening design illustrated in Table 1.6 is a 27−4 fractional factorial, but
is also an example of an orthogonal array, which is by definition an array of sym-
bols, in this case ±1, in which every symbol appears an equal number of times in
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 23

each column, and any pair of symbols appears an equal number of times in any
pair of columns. An orthogonal array of size n × (n − 1) with two symbols in each
column specifies an n-run screening design for n − 1 factors. The designs with
symbols ±1 are called Plackett-Burman designs and Hadamard matrices defining
them have been shown to exist for all multiples of four up to 424.
More generally, an n × k array with mi symbols in the ith column is an ortho-
gonal array of strength r if all possible combinations of symbols appear equally
often in any r columns. The symbols correspond to levels of a factor. Table 1.10
gives an orthogonal array of 18 runs, for 6 factors with three levels each.

Table 1.10. An orthogonal array for 6 factors each at 3 levels

run A B C D E F
1 −1 −1 −1 −1 −1 −1
2 −1 0 0 0 0 0
3 −1 +1 +1 +1 +1 +1
4 0 −1 −1 0 0 +1
5 0 0 0 +1 +1 −1
6 0 +1 +1 −1 −1 0
7 +1 −1 0 −1 +1 0
8 +1 +1 −1 +1 0 −1
9 +1 +1 −1 +1 0 +1
10 −1 −1 +1 +1 0 0
11 −1 0 −1 −1 +1 +1
12 −1 +1 0 0 −1 −1
13 0 −1 0 +1 −1 +1
14 0 0 +1 −1 0 −1
15 0 +1 −1 0 +1 0
16 +1 −1 +1 0 +1 −1
17 +1 0 −1 +1 −1 0
18 +1 +1 0 −1 0 +1

Orthogonal arrays are particularly popular in applications of so-called Taguchi

methods in technological experiments and manufacturing. An extensive discussion
is given in [9-11]. For simplest experiments, the Taguchi experiments are same as
the fractional factorial experiments in the classical DOE. Even for common expe-
riments used in the industry for problem solving and design improvements, the
main attractions of the Taguchi approach are standardized methods for experiment
designs and analyses of results. To use the Taguchi approach for modest experi-
mental studies, one does not need to be an expert in statistical science. This allows
working engineers on the design and production floor to confidently apply the
technique.
24 V.P. Astakhov

While there is not much difference between the two types of DOE for simpler
experiment designs, for mixed level factor designs and building robustness in
products and processes, the Taguchi approach offers some revolutionary concepts
that were not known even to the expert experimenters. These include standard me-
thod for array modifications, experiment designs to include noise factors in the
outer array, signal-to-noise ratios for analysis of results, loss function to quantify
design improvements in terms of dollars, treatment of systems with dynamic cha-
racteristics, etc.
Although the Taguchi method was developed as a powerful statistical method
for shop floor quality improvement, a way too many researchers have been using
this method as a research and even optimization method in manufacturing, and
thus in metal cutting studies (for example, [25-29] that, in the author's opinion un-
acceptable because the Taguchi methods suffer the same problems as any frac-
tional factorial DOE.
Unfortunately, it became popular to consider the used of only a fraction of the
number of test combinations needed for a full factorial design. That interest spread
because many practitioners do not take the time to find out, or for other reasons
never realized the “price” paid when one uses fractional factorial DOEs including
the Taguchi method: (1) Certain interaction effects lose their contrast so know-
ledge of their existence is gone, (2) Significant main effect and important interac-
tions have aliases – other ‘confounding’ interaction names. Thus wrong answers
can, and often do come from the time, money, and effort of the experiment.
Books on DOE written by ‘statistical’ specialists add confusion to the matter
claiming that interactions (three-factor or higher order) would be too difficult to
explain; nor could they be important. The ideal gas law (1834 by Emil Clapeyron),
known from high-school physics as

PV = nRT (1.8)

(where P is the pressure of the confined gas, V is the volume of the confined gas, n
is the number of moles of gas, R is gas constant, T is the temperature) plots as a
simple graph. It depicts a three-factor interaction affecting y (response) as pres-
sure, or as volume. The authors of these statistical books/papers may have forgot-
ten their course in physics?
The problem is that the ability of the Taguchi method is greatly overstated by
its promoters who described Taguchi orthogonal tables as Japan’s ‘secret super
weapon’ that is the real reason for developing an international reputation for quali-
ty. Claim: a large number of variables could now be handled with practical effi-
ciency in a single DOE. As later details became available, many professionals
realized that these arrays were fractional factorials, and that Taguchi went to greater
extremes that other statisticians in the degree of fractionating. According to the
Taguchi method, the design is often filled with as many single factors for which it
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 25

has room. The design becomes “saturated” so no degrees of freedom are left for
its proper statistical analysis. The growing interest in the Taguchi method in the
research and optimization studies in manufacturing attests to the fact that manu-
facturing researches either are not aware of the above-mentioned ‘price” paid for
apparent simplicity, or know of no other way to handle more and more variables at
one time.

1.7.2 Sieve DOE

Plackett and Burman [30] developed a special class of fractional factorial experi-
ments that includes interactions. When this kind of DOE (referred to as the Plack-
ett–Burman DOE) is conducted properly using a completely randomized se-
quence, its distinctive feature is high resolution. Despite a number of
disadvantages (for example, mixed estimation of regression coefficients), this me-
thod utilizes high-contrast diagrams for the factors included in the test as well as
for their interactions of any order. This advantage of the Plackett–Burman DOE is
very useful in screening tests.
This section presents a simple methodology of screening DOE to be used in
manufacturing tests [31]. The method, referred to as the sieve DOE, has its foun-
dation in the Plackett–Burman design ideas, an oversaturated design matrix and
the method of random balance. The proposed sieve DOE allows the experimental-
ist to include as many factors as needed at the first phase of the experimental study
and then to sieve out the non-essential factors and interactions by conducting a
relatively small number of tests. It is understood that no statistical model can be
produced in this stage. Instead, this method allows the experimentalist to deter-
mine the most essential factors and their interactions to be used at the second stage
of DOE (full factorial or RSM DOE).
The proposed sieve DOE includes the method of random balance. This method
utilizes oversaturated design plans where the number of tests is fewer than the
number of factors and thus has a negative number of degrees of freedom [32]. It is
postulated that if the effects (factors and their interactions) taken into considera-
tion are arranged as a decaying sequence (in the order of their impact on the va-
riance of the response), this will approximate a ranged exponential-decay series.
Using a limited number of tests, the experimentalist determines the coefficients of
this series and then, using the regression analysis, estimates the significant effects
and any of their interactions that have a high contrast in the noise field formed by
the insignificant effects.
The initial linear mathematical model, that includes k number of factors
(effects), has the following form

y = a0 + a1 x1 + ... + a k xk + a12 x1 x 2 + ... + a k −1,k xk −1 x k + δ (1.9)

26 V.P. Astakhov

where a0 is the absolute term often called the main effect, ai (i = 1,k) are the coef-
ficients of linear terms, aij (i = 1,...,k −1; j = i+1,...,k, i ≠ j) are the coefficients of
interaction terms and δ is the residual error of the model.
The complete model represented by Eq. (1.9) can be rearranged as a split of a
linear form considering that some xi designate the iterations terms as

y = a0 + a1 x1 + ... + ak −l ,k xk −l + b1 z1 + b2 z2 + ...bl zl + δ =
(1.10)
a0 + a1 x1 + ... + ak −l ,k xk −l + Δ

where

Δ = b1 z1 + b2 z2 + ...bl zl + δ (1.11)

and

σ 2 {Δ} = b12σ 2 {z1 } + b22σ 2 {z2 }... + bl2σ 2 {zl } + σ 2 {δ } (1.12)

In the construction of the split model represented by Eq. (1.10), (k - l) signifi-

cant effects were distinguished and l effects were assigned to the noise field.
Naturally, the residual variance σ 2 {Δ} is greater than the tests variance σ 2 {δ }
so that the regression coefficients in Eq.(1.10) will be estimated with greater
errors and, moreover, the estimates of the coefficients of this model are mixed.
Therefore, the sensitivity of the random balance method is low so that the re-
sultant model has a little significance, and thus should not be used as a valid
statistical model. However, this method is characterized by the great contrast of
essential effects, which could be distinguished easily on the noisy fields formed
by other effects. The latter makes this method the simplest yet highly reliable
screening method that can be used at the first stage of testing to distinguish the
significant factors and interaction to be used in the full factoring DOE includ-
ing RSM.
To demonstrate the simplicity of the discussed test, an example of determina-
tion of critical factors and their interaction is considered for grinding parameters
of a gundrill [31]. The terminal end of a gundrill is shown in Figure 1.5. It is
formed with the approach cutting edge angles φ1 and φ2 of the outer 1 and inner 2
cutting edges, respectively. These cutting edges meet at the drill point P. The lo-
cation of P (defined by distance md in Figure 1.5) is critical and it can be varied
for optimum performance depending on the work material and the finished hole
specifications. The geometry of the terminal end largely determines the shape of
the chips and the effectiveness of the cutting fluid, the lubrication of the tool, and
removal of the chips. The flank surface 3 having normal flank angle αn1 is applied
to the outer cutting edge 1 and the flank surface 4 having normal flank angle equal
to αn2 is applied to the inner cutting edge 2. The rake angle γ is formed by the chip
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 27

removal flute face 5. To assure drill free penetration, i.e. to prevent the interfe-
rence of the drill’s flanks with the bottom of the hole being drilled, the auxiliary
flank 6 known as shoulder dub-off is ground at certain angle φ4. Their location of
the shoulder dub-off is defined by distance b. The detailed description of the gun-
drill geometry and importance of the parameters selected for the sieve test are pre-
sented by the author earlier [33].

Fig. 1.5. Grinding parameters of the terminal end of a gundrill selected for sieve DOE

Eight factors have been selected for this sieve DOE and their intervals of varia-
tion are shown in Table 1.11. The design matrix was constructed as follows. All
the selected factors were separated into two groups. The first one contained factors
x1 , x2 , x3 , x4 , form a half-replica 24 −1 with the defining relation I = x1 x2 x3 x4 . In
this half-replica, the factors’ effects and the effects of their interactions are not
mixed. The second half-replica was constructed using the same criteria. A design
matrix was constructed using the first half-replica of the complete matrix and add-
ing to each row of this replica a randomly selected row from the second half-
replica. Three more rows were added to this matrix to assure proper mixing and
these rows were randomly selected from the first and second half-replicas. Table
1.12 shows the constructed design matrix.
As soon as the design matrix is completed, its suitability should be examined
using two simple rules. First, a design matrix is suitable if it does not contain two
identical columns having the same or alternate signs. Second, a design matrix
should not contain columns whose scalar products with any other column result in
28 V.P. Astakhov

a column of the same (“+” or “−“) signs. The design matrix shown in Table 1.12
was found suitable as it meets the requirements set by these rules. In this table, the
responses yi i = 1...11 are the average tool life calculated over three independent
tests replicas obtained under the indicated test conditions.
Analysis of the result of sieve DOE begins with the construction of a correla-
tion (scatter) diagram shown in Figure 1.6. Its structure is self-evident. Each factor
is represented by a vertical bar having on its left side values (as dots) of the re-
sponse obtained when this factor was positive (the upper value) while the values
of the response corresponding to lower lever of the considered factor (i.e. when
this factor is negative) are represented by dots on the right side of the bar. As such,
the scale makes sense only along the vertical axis.

Table 1.11. The levels of te factors selected for the sieve DOE

Factors Approach Approach Flank Drill point Flank an- Shoulder Rake an- Shoulder
angle angle angle offset gle αn2 (o) dub-off an- gle dub-off
φ1 (o) φ2 (o) αn1 (o) md (mm) gle γ (o) location
φ4 (o) b (mm)
Code
x1 x2 x3 x4 x5 x6 x7 x8
designation
Upper
45 25 25 3.0 12 45 5 4
level (+)
Lower
25 10 8 1.5 7 20 0 1
level (-)

Table 1.12. Design matrix

Average
Factors tool life Corrections
Run (min)
x1 x2 x3 x4 x5 x6 x7 x8 y y c1 yc 2
1 +1 +1 −1 −1 +1 −1 +1 −1 7 18.75 11.11
2 +1 +1 +1 +1 −1 +1 +1 −1 16 11.50 11.50
3 −1 +1 −1 −1 +1 +1 −1 +1 11 11.00 11.00
4 −1 −1 +1 +1 +1 −1 −1 +1 38 21.75 16.61
5 +1 −1 −1 +1 −1 −1 +1 −1 18 13.50 13.50
6 +1 −1 +1 −1 +1 +1 −1 −1 10 21.75 14.61
7 −1 −1 −1 −1 −1 +1 −1 +1 14 14.00 14.00
8 −1 +1 −1 +1 −1 −1 +1 +1 42 25.75 16.61
9 +1 −1 −1 −1 −1 +1 −1 +1 9 20.75 20.75
10 −1 +1 +1 +1 −1 +1 −1 −1 32 15.75 15.75
11 +1 −1 +1 −1 −1 −1 +1 −1 6 17.75 10.61
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 29

y
40

30
16.4

5.77 3.57 19.7 10.6 6.9 1.2 4.9

20

10

+
x1
+
x2
+
x3
+
x4
+
x5
+
x6
+
x7
+
x8
x
Fig. 1.6. Correlation diagram (first sieve)

Each factor included in the experiment is estimated independently. The

simplest way to do this is to calculate the distance between means on the left and
right side of each bar. These distances are shown on the correlation diagram in
Figure 1.6. As can be seen, these are the greatest for factors x1 and x2 , and thus
these two factors are selected for the analysis. The effects of factors are calculated
using special correlation tables. A correlation table (Table 1.13) was constructed
to analyze the considered two factors. Using the correlation table, the effect of
each selected factor can be estimated as

y1 + y3 + ... + yn y2 + y4 + ... + yn −1
Xi = − (1.13)
m m

where m is the number of y in Table 1.13 for the considered factor assigned to
the same sign (“+” or “−”). It follows from Table 1.13 that m = 2.
The effects of the selected factors were estimated using data in Table 1.13 and
Eq.(1.13) as

y1−1 + y1−3 y1− 2 + y1− 4 17 + 9.3 37.3 + 12.5

X1 = − = − = −11.75 (1.14)
2 2 2 2
y1−1 + y1− 2 y1−3 + y1− 4 17 + 37.3 9.3 + 12.5
X4 = − = − = 16.25 (1.15)
2 2 2 2
30 V.P. Astakhov

Table 1.13. Correlation table (first sieve)

Estimated Estimated
+x1 −x1 +x1 −x1
factor factor
7 11
16 38
10 14
18 42
9
32
11
+x4 −x4

∑y = 34 ∑y = 112 ∑y 1− 4 = 25
1−1 1− 2
∑ y1−3 = 37
y1− 4 = 12.5
y1−1 = 17 y1− 2 = 37.3
y1−3 = 9.3

The significance of the selected factors is examined using the Student’s t-

criterion, calculated as

( y1−1 + y1−3 + ... + y1− n ) − ( y1− 2 + y1− 4 + ... + y1− ( n −1) )

t= (1.16)
si2
∑ i ni

where si is the standard deviation of i-th cell of the correlation table defined as

2
⎛ ⎞
∑i y ⎜⎝ ∑i yi ⎟⎠
2
i
si = − (1.17)
ni − 1 ni ( ni − 1)

where ni is the number of terms in the considered cell.

The Student’s criteria for the selected factors are calculated using the results
presented in the auxiliary table (Table 1.14) as follows

( y1−1 + y1−3 ) − ( y1− 2 + y1− 4 ) (17 + 9.3) − (37.3 + 12.5)

t X1 = = = −6.68 (1.18)
s 2 3.52
∑ i
i
ni
( y1−1 + y1− 2 ) − ( y1−3 + y1− 4 ) (17 + 37.3) − (9.3 + 12.5)
tX = = = 9.23 (1.19)
4
s2 3.52
∑ i
i
ni
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 31

A factor is considered to be significant if t X > tcr where the critical value, tcr for
i

the Student’s criterion in found in a statistical table for the following number of
degrees of freedom

fr = ∑ n − k = 11 − 4 = 7
i
i (1.20)

where k is the number of cells in the correlation table.

For the considered case, t99.9 = 5.959 (Table 5.7 in [34]) so that the considered
factors are significant with a 99.9% confidence level.

Table 1.14. Calculating t-criterion

∑y (∑ y ) ∑y
2 2
Cell # 1− i 1− i 1−i
ni s2 s 2 / ni
1 34 1156 580 2 2.00 1.00
2 112 12544 4232 3 25.33 8.44
3 37 1369 351 4 2.92 0.73
4 25 625 317 2 4.50 2.25

The discussed procedure is the first stage in the proposed sieve DOE and thus is
referred to as the first sieve. This first sieve allows the detection of the strongest
factors, i.e. those factors that have the strongest influence on the response. After
these strong linear effects are detected, the size of “the screen” to be used in the
consecutive sieves is reduced to distinguish less strong effects and their interac-
tions. This is accomplished by correction of the experimental results presented in
column y1 of Table 1.12. Such a correction is carried out by adding the effects
(with the reverse signs) of the selected factors (Eqs. (1.14) and (1.15)) to column
y1 of Table 1.12, namely, by adding 11.75 to all results at level “ + x1 ” and –16.25
to all results at level “ + x4 ”. The corrected results are shown in column yc1 of Ta-
ble 1.12. Using the data of this table, one can construct a new correlation diagram
shown in Figure 1.7, where, for simplicity, only a few interactions are shown al-
though all possible interactions have been analyzed. Using the approach described
above, a correlation table (second sieve) was constructed (Table 1.15) and the in-
teraction x4 x8 was found to be significant. Its effect is X48 = 7.14.
After the second sieve, column yc1 was corrected by adding the effect of X48
with the opposite sign, i.e. –7.14 to all results at level +x48. The results are shown
in column yc 2 of Table 1.12. Normally, the sieve of the experimental results con-
tinues while all the remaining effects and their interactions become insignificant,
at say a 5% level of significance if the responses were measured with high accura-
cy and a 10% level if not. In the considered case, the sieve was ended after the
third stage because the analysis of these results showed that there are no more
32 V.P. Astakhov

significant factors or interactions left. Figure 1.8 shows the scatter diagram of the
discussed test. As can be seen in this figure, the scatter of the analyzed data reduc-
es significantly after each sieve so normally three sieves are sufficient to complete
the analysis.

25

20
4.54 6.82
5.07

15

10
+
x1
+
x2
+
x3
+
x4
+
x5
+
x6
+
x7
+
x8
+ + + +
x3 x 7 x 1 x 6 x 4 x 8 x2 x3 x5 x

Fig. 1.7. Correlation diagram (second sieve)

Table 1.15. Correlation table (second sieve)

Estimated
+x3x7 − x3x7 +x4x8 − x4x8
factor
11.5 18.75 18.75 11.5
11 25.75 25.75 11

+x2 ∑y 2 −1 = 22.5 15.75 ∑y 2−3 = 44.5 15.75

y2−1 = 11.25 ∑y 2−2 = 60.25

y2−3 = 22.25 ∑y 2−4 = 38.25

y2 − 2 = 20.08 y2 − 4 = 12.75
Estimated
+ x3x7 − x3x7 + x4x8 − x4x8
factor
14 21.75 21.75 13.5
20.75 13.5 21.75 14
17.75 21.75 17.75 20.75
−x2
∑y 2 −5 = 52.5 ∑y 2−6 = 57 ∑y 2−7 = 61.25 ∑y 2 −8 = 48.25

y2−5 = 17.5 y2− 6 = 19 y2− 6 = 20.42 y2−5 = 16.08

Design of Experiment Methods in Manufacturing: Basics and Practical Applications 33

40

Tool life (min)

30

20

10

y yC1 yC2
Fig. 1.8. Scatter diagram

The results of the proposed test are summarized in Table 1.16. Figure 1.9
shows the significance of the distinguished effects in terms of their influence on
tool life. As seen, two linear effects and one interaction having the strongest ef-
fects were distinguished. The negative sign of x1 shows that tool life decreases
when this parameter increases. The strongest influence on tool life has the drill
point offset md.

Table 1.16. Summary of the sieve test

Stage of sieve Distinguished factor Effect t-criterion

Original data x1 −11.5 6.68
x4 16.25 9.23
First sieve x48 7.14 4.6
Second sieve - - -

While the distinguished linear effects are known to have strong influence on
tool life, the distinguished interaction x4 x8 has never before been considered in
any known studies on gun drilling. Using this factor and results of the complete
DOE, a new pioneering geometry of gun drills has been developed (for example
US Patent 7147411).
The proposed sieve DOE allows experimentalists to include into consideration
as many factors as needed. Conducting a relatively simple sieve test, the signifi-
cant factors and their interactions can be distinguished objectively and then be
used in the subsequent full DOE. Such an approach allows one to reduce the total
34 V.P. Astakhov

number of tests dramatically without losing any significant factor or factor interac-
tion. Moreover, interactions of any order can be easily analyzed. The proposed
correlation diagrams make such an analysis simple and self-evident.

18

12
Effect

6
-
0
x4 x1 x4 x 8 x5 x6 x2 etc.

Fig. 1.9. Significance of the effects distinguished by the sieve DOE (Pareto analysis)

1.7.3 Split-Plot DOE

All industrial experiments are split-plot experiments. This provocative remark has
been attributed to the famous industrial statistician, Cuthbert Daniel, by Box et al.
[35] in their well-known text on the design of experiments. Split-plot experiments
were invented by Fisher [36] and their importance in industrial experimentation
has been long recognized [37]. It is also well known that many industrial experi-
ments are fielded as split-plot experiments and yet erroneously analyzed as if they
were completely randomized designs. This is frequently the case when hard-to-
change factors exist and economic constraints preclude the use of complete ran-
domization. Recent work, most notably by Lucas and his coworkers [38], Ganju
and Lucas [39], Ju and Lucas [40],Webb et al. [41] has demonstrated that many
experiments previously thought to be completely randomized experiments also
exhibit split-plot structure. This surprising result adds credence to the Daniel’s
proclamation and has motivated a great deal of pioneering work in the design and
analysis of split plot experiments [42].
In simple terms, a split-plot experiment is a blocked experiment, where the
blocks themselves serve as experimental units for a subset of the factors. Thus,
there are two levels of experimental units. The blocks are referred to as whole
plots, while the experimental units within blocks are called split plots, split units,
or subplots. Corresponding to the two levels of experimental units are two levels
of randomization. One randomization is conducted to determine the assignment of
block-level treatments to whole plots. Then, as always in a blocked experiment, a
randomization of treatments to split-plot experimental units occurs within each
block or whole plot.
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 35

Split-plot designs were originally developed by Fisher [36] for use in agricul-
tural experiments. As a simple illustration, consider a study of the effects of two
irrigation methods (factor A) and two fertilizers (factor B) on yield of a crop, us-
ing four available fields as experimental units. In this investigation, it is not possi-
ble to apply different irrigation methods (factor A) in areas smaller than a field, al-
though different fertilizer types (factor B) could be applied in relatively small
areas. For example, if we subdivide each whole plot (field) into two split plots,
each of the two fertilizer types can be applied once within each whole plot, as
shown in Figure 1.10. In this split-plot design, a first randomization assigns the
two irrigation types to the four fields (whole plots); then within each field, a sepa-
rate randomization is conducted to assign the two fertilizer types to the two split
plots within each field.

Fig. 1.10. Split plot agricultural layout. (Factor A is the whole-plot factor and factor B is
the split-plot factor)

In industrial experiments, factors are often differentiated with respect to the ease
with which they can be changed from experimental run to experimental run. This
may be due to the fact that a particular treatment is expensive or time-consuming to
change, or it may be due to the fact that the experiment is to be run in large batches
and the batches can be subdivided later for additional treatments. Box et.al. [35] de-
scribed a prototypical split-plot experiment with one easy-to-change factor and one
hard-to-change factor. The experiment was designed to study the corrosion resis-
tance of steel bars treated with four coatings, C1, C2, C3,and C4, at three furnace
temperatures, 360oC, 370oC, and 380oC. Furnace temperature is the hard-to-change
factor because of the time it takes to reset the furnace and reach a new equilibrium
temperature. Once the equilibrium temperature is reached, four steel bars with ran-
domly assigned coatings C1, C2, C3, and C4 are randomly positioned in the furnace
and heated. The layout of the experiment as performed is given in Table 1.17. No-
tice that each whole-plot treatment (temperature) is replicated twice and that there
36 V.P. Astakhov

is just one complete replicate of the split-plot treatments (coatings) within each
whole plot. Thus, DOE matrix has six whole plots and four subplots within each
whole plot.

Table 1.17. Split-plot design and data for studying the corrosion resistance of steel bars

Temperature Coatings
Whole-plot
(oC) (randomized order)
C2 C3 C1 C4
1. 360
73 83 67 89
C1 C3 C4 C2
2. 370
65 87 86 91
C3 C1 C2 C4
3. 380
147 155 127 212
C4 C3 C2 C1
4. 380
153 90 100 108
C4 C1 C3 C2
5. 370
150 140 121 142
C1 C4 C2 C3
6. 360
33 54 8 46

The analysis of a split-plot experiment is more complex than that for a com-
pletely randomized experiment due to the presence of both split-plot and whole-
plot random errors. In the Box et al. corrosion resistance example [35], a whole-
plot effect is introduced with each setting or re-setting of the furnace. This may be
due, e.g., to operator error in setting the temperature, to calibration error in the
temperature controls, or to changes in ambient conditions. Split-plot errors might
arise due to lack of repeatability of the measurement system, to variation in the
distribution of heat within the furnace, to variation in the thickness of the coatings
from steel bar to steel bar, and so on.
In the author’s experience, ADX Interface for Design of Experiments
(http://support.sas.com/documentation/cdl/en/adxgs/60376/HTML/default/overvie
w_sect1.htm) by SAS company is most suitable for the split-plot DOE. Following
its manual [43], consider split-plot DOE for tablet production.
The tablet production process has several stages, which include batch mixing,
where ingredients are combined and mixed with water, and pellet production,
where the batch is processed into pellets that are compressed to form tablets. It is
more convenient to mix batches and randomize the treatments within each batch
than to mix a new batch for each run. Thus, this experiment calls for a standard
two-stage split-plot design.
The moisture content and mixing speed for the batch constitute whole-plot fac-
tors, while the factors that control the variety of ways that pellets can be produced
from a single batch are subplot factors. The responses of interest are measured on
the final tablets. Table 1.18 shows all the variables involved and the stage with
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 37

which they are associated. The goal of the experiment is to determine which ef-
fects are significant. The researcher is interested in both whole-plot factors and
subplot factors.

Table 1.18. Factors and their responses in the tablet formulation experiment

Level Description
Variable Name
Low High
Whole Plot FLUID 90 115 Moisture content (%)
MIX 15 45 Mixing time (min)
Split Plot EXTRUDER 36 132 Extruder speed (rpm)
SCREEN 0.6 1.0 Screen size (mm)
RESID 2 5 Residence time (min)
DISK 450 900 Disk speed (rpm)
Response MPS Mean particle size (micron)

I. Task List

1. Create the split-plot design.

2. Explore the response data with scatter plots and box plots.
3. Estimate all main effects and as many two-factor interactions as possible
between the factors.

II. Start the program

A split-plot design can be created by selecting File Create New Design

Split-plot from the ADX desktop. The main design window will appear as shown
in Figure 1.11

Fig. 1.11. Selecting the split-plot design

38 V.P. Astakhov

III. Defining Variables

To define the variables in this experiment, one should do the following:

1. Click Define Variables. The Define Variables window will appear, and
the Whole-Plot Factor tab will already be selected (Figure 1.12).
2. Define the whole-plot factors.
2.1. Create two new factors by clicking Add and selecting 2.
2.2. Enter the factor names given in the whole-plot section of
Table 1.12.
3. Define the subplot factors.
3.1. Click the Sub-plot Factor tab (Figure 1.13).
3.2. Create four new factors by clicking Add and selecting 4.
3.3. Enter the factor names given in the subplot section of
Table 1.12
4. Click the Block tab (Figure 1.14). ADX will assign a unique block level
to each whole plot when it generates the design, so you do not need to
specify the number of block levels. Change the block name to BATCH,
since each whole plot is a batch of material.
5. Enter the response information.
5.1. Click the Response tab (Figure 1.15).
5.2. Change the name of the default response to MPS and its label to
Mean Particle Size (microns).
5.3. Click OK to accept the variable definitions.

Fig. 1.12. Defining variables, the whole plot

Design of Experiment Methods in Manufacturing: Basics and Practical Applications 39

Fig. 1.13. Defining variables, the sub-plot

Fig. 1.14. Assigning block levels

Fig. 1.15. Entering the response information

40 V.P. Astakhov

IV. Exploring Response Data

Before fitting a mixed model, click Explore in the main design window. Both box
plots and scatter plots are available. Box plots help you visualize the differences in
response distribution across levels of different factors, both random and fixed.
Scatter plots show the individual responses broken down by factor level or run and
can also be used to investigate time dependence.
The box plot is the first graph that is displayed (Figure 1.16). One can explore
the distribution of the response broken down by each of the factors or batches that
make up the whole plot.

Fig. 1.16. Explore data

Close the window to return to the main design window.

V. Specifying Fixed and Random Effects

ADX does not generate a default master model for a split-plot design, so you must
do so before fitting. Click Fit in the main design window. The Fit Details for MPS
window will open (Figure 1.17). It has sections to define fixed and random effects
and classification variables. The fixed effects in a split-plot design are, as usual,
the main effects and interactions of interest between the factors. The split-plot
structure of the experiment determines the choice of random effects, which in turn
determines the proper error term for each fixed effect.
The modeling objectives of a split-plot design are in principle the same as those
of a standard screening design. You want to estimate as many effects as possible
involving the factors and determine which ones are significant. When this design
was analyzed as a standard full factorial, the 64 runs provided enough degrees of
freedom to estimate effects and interactions of all orders. However, FLUID and
MIX are whole-plot effects and apply only to batches. Therefore, with respect
to these two factors, the experiment has only four runs (the four batches). The
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 41

interaction between the whole-plot effects will estimate the whole-plot error, so
this interaction is not included as a fixed effect.

Fig. 1.17. Specifying the fixed effects

1. Click Fit in the main design window. Select Model Change master
model to open the Specify Master Model window. On the Fixed Effects tab
(Figure 1.17), specify the following as fixed effects:
• All main effects. Click and drag or hold down the CTRL key and click to
select the FLUID, MIX, EXTRUDER, SCREEN, RESID, and DISK va-
riables. Then click Add.
• All two-factor interactions except FLUID*MIX. Select all six factors and
click Interaction. Double-click FLUID*MIX in the list on the right to
remove it from the master model. (You might have to scroll to find it.).
2. Next, click the Random Effects tab (Figure 1.18). Here you specify the
whole-plot error, which is BATCH. Select BATCH and click Add.

Fig. 1.18. Specifying the random effects

42 V.P. Astakhov

3. Examine the Classification Variables tab Figure 1.19. BATCH is automati-

cally entered because it is a random effect. If there were categorical fixed ef-
fects, you would enter them on this tab.

Fig. 1.19. Examining the Classification Variables tab

Click OK to close this window and fit the master model.

VI. Fitting a Model

ADX uses the MIXED procedure to estimate the fixed and random effects and va-
riance components of the model. After specifying the model, you will see the Fit
Details for MPS window. There are three tabs:
• The Fixed Effects tab (Figure 1.20) is similar to the Effects Selector for
other designs. The table lists Type 3 tests of the fixed effects. Click an
effect to toggle the selection of significant effects.

Fig. 1.20. Fixed effects tab

Design of Experiment Methods in Manufacturing: Basics and Practical Applications 43

• The Variance Components tab provides estimates of the variance of

each random effect, including the residual error. The approximate confi-
dence limits for variance components can be calculated easily from this
information.

Fig. 1.21. Variance components tab

• The Random Effects tab lists the random effects broken down by level
(for interactions, these are broken down by each combination of levels).
These are empirical best linear unbiased predictors (EBLUPs) of the ob-
served random effects. You can use them to screen for unusual whole
plots and to assess the normality assumption of the model.

Fig. 1.22. Random effects tab

Choosing a “best” split-plot design for a given design scenario can be a daunting
task, even for a professional statistician [42]. Facilities for construction of split-
plot designs are not as yet generally available in software packages (with
44 V.P. Astakhov

SAS/ADX being one exception). Moreover, introductory textbooks frequently

consider only very simple, restrictive situations. Unfortunately, in many cases, ob-
taining an appropriate design requires reading (and understanding) a relevant re-
cent research paper or it requires access to software for generating the design. To
make matters even more difficult, construction of split-plot experiments is an ac-
tive area of research and differences in opinion about the value of the various re-
cent advances exist. Jones and Nachtsheim [42] showed how to make the proper
choice of split-plot designs in five areas: (1) two-level full factorial designs; (2)
two-level fractional factorial designs; (3) mixture and response surface designs;
(4) split-plot designs for robust product experiments; and (5) optimal designs.

1.7.4 Group Method of Data Handling (GMDH)

The important invention of Heisenberg's uncertainty principle [44] from the field
of quantum theory has a direct or indirect influence on later scientific develop-
ments. Heisenberg's works became popular between 1925 and 1935. According to
his principle, a simultaneous direct measurement between the coordinate and mo-
mentum of a particle with an exactitude surpassing the limits is impossible; fur-
thermore, a similar relationship exists between time and energy. Since his results
were published, various scientists have independently worked on Heisenberg's un-
certainty principle.
In 1931, Godel published his works on mathematical logic showing that the
axiomatic method itself had inherent limitations and that the principal shortcoming
was the so-called inappropriate choice of "external complement." According to his
well-known incompleteness theorem [45], it is in principle impossible to find a
unique model of an object on the basis of empirical data without using an "exter-
nal complement" [46]. The regularization method used in solving ill-conditioned
problems is also based on this theorem. Hence "external complement" and "regu-
larization" are synonyms expressing the same concept.
In regression analysis, the root mean square (RMS) or least square error deter-
mined on the basis of all experimental points monotonically decreases when the
model complexity gradually increases. This drops to zero when the number of
coefficients n of the model becomes equal to the number of empirical points N.
Every equation that possesses n coefficients can be regarded as an absolutely ac-
curate model. It is not possible, in principle, to find a unique model in such a sit-
uation. Usually experienced modelers use trial and error techniques to find a
unique model without stating that they consciously or unconsciously use an "ex-
ternal complement," necessary in principle for obtaining a unique model. Hence,
none of the investigators appropriately selects the "external complement"—the
risk involved in using the trial and error methods.
In complex systems modeling statistical probability distributions, like normal
distribution, cannot be used if only a few empirical points are available. The im-
portant way is to use the inductive approach for sifting various sets of models
whose complexity is gradually increased and to test them for their accuracy. The
principle of self-organization can be formulated as follows: When the model com-
plexity gradually increases, certain criteria, which are called selection criteria or
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 45

objective functions and which have the property of "external complement," pass
through a minimum. Achievement of a global minimum indicates the existence of
a model of optimum complexity (Figure 1.23).
The notion that there exists a unique model of optimum complexity, determina-
ble by the self-organization principle, forms the basis of the inductive approach.
The optimum complexity of the mathematical model of a complex object is found
by the minimum of a chosen objective function which possesses properties of ex-
ternal supplementation (by the terminology of Godel's incompleteness theorem
from mathematical logic). The theory of self-organization modeling [47] is based
on the methods of complete, incomplete, and mathematical induction [48].
This has widened the capabilities of system identification, forecasting, pattern
recognition, and multicriterial control problems.

Fig. 1.23. Variation in least square error ε(A + B) and error measure of an "external com-
plement" Δ(B) for a regression equation of increasing complexity S; Sopt is the model of op-
timal complexity

Group Method of Data Handling (hereafter, GMDH) developed as a practical

modeling tool based on the theory of self-organization modeling was applied in a
great variety of areas for data mining and knowledge discovery, forecasting and
systems modeling, optimization, and pattern recognition [47, 49-51]. Inductive
GMDH algorithms provide a possibility to find automatically interrelations in da-
ta, to select the optimal structure of model or network and to increase the accuracy
of existing algorithms. This original self-organizing approach is substantially dif-
ferent from deductive methods used commonly for modeling. It has inductive na-
ture - it finds the best solution by sorting-out possible alternatives and variants. By
sorting of different solutions, the inductive modeling approach aims to minimize
the influence of the experimentalist on the results of modeling. An algorithm itself
can find the structure of the model and the laws, which act in the system. It can be
used as an advisor to find new solutions of artificial intelligence (AI) problems.
GMDH is a set of several algorithms for different problems solution. It consists
of parametric, clusterization, analogs complexing, rebinarization, and probability
algorithms. This self-organizing approach is based on the sorting-out of models of
46 V.P. Astakhov

different levels of complicity and selection of the best solution by minimum of ex-
ternal criterion characteristic. Not only polynomials but also nonlinear, probabilis-
tic functions, or clusterizations are used as basic models.
In the author’s opinion, the GMDH approach is the most suitable DOE method
for experimental studies in manufacturing because:
1. The optimal complexity of model structure is found, adequate to level of
noise in data sample. For real problems solution with noisy or short data,
simplified forecasting models are more accurate than with any other
known methods of DOE.
2. The number of layers and neurons in hidden layers, model structure,
and other optimal neutral networks (NN) parameters are determined
automatically.
3. It guarantees that the most accurate or unbiased models will be found -
method does not miss the best solution during sorting of all variants (in
given class of functions).
4. Any non-linear functions or features can be used as input variables are
used, which can influence the output variable.
5. It automatically finds interpretable relationships in data and selects effec-
tive input variables.
6. GMDH sorting algorithms are rather simple for programming.
7. The method uses information directly from the data samples and
minimizes influence of priory researcher assumptions about results of
modeling.
8. GMDH neuronets are used to increase the accuracy of other modeling
algorithms.
9. The method allows finding an unbiased physical model of object (law or
clusterization) - one and the same for all future samples.

There are many published articles and books devoted to GMDH theory and its ap-
plications. The GMDH can be considered as a further propagation or extension of
inductive self-organizing methods to the solution of more complex practical prob-
lems. It solves the problem of how to handle the data samples of observations. The
goal is to obtain a mathematical model of the object under study (the problem of
identification and pattern recognition) or to describe the processes, which will take
place at the object in the future (the problem of process forecasting). GMDH
solves, by means of a sorting-out procedure, the multidimensional problem of
model optimization

(
g = arg min CR ( g ) , CR ( g ) = f P, S , z 2 , T1 , V
g ⊂G
) (1.21)

where G is a set of considered models, CR is the external criterion of model g

quality from this set, P is the number of variables set, S is model complexity, z2 is
the noise dispersion, T1 is the number of data sample transformation, V is the
type of reference function.
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 47

For the definite reference function, each set of variables corresponds to definite
model structure P = S. Problem transforms to much simpler one-dimensional

CR ( g ) = f ( S ) (1.22)

when z2 = const, T = const, and V = const.

The method is based on the sorting-out procedure, i.e. consequent testing of
models, chosen from set of models-candidates in accordance with the given crite-
rion. Most of GMDH algorithms use the polynomial reference functions. General
correlation between input and output variables can be expressed by Volterra func-
tional series, discrete analogue of which is Kolmogorov-Gabor polynomial

M M M M M M
y = b0 + ∑ bi xi + ∑∑ bij xi x j + ∑∑∑ bijk xi x j xk (1.23)
i =1 i =1 j =1 i =1 j =1 k =1

where X(x1,x2,...,xM) is the input variables vector, M is the number of input va-
riables, A(b1,b2,...,bM) is the vector of coefficients.
Components of the input vector X can be independent variables, functional
forms or finite difference terms. Other non-linear reference functions, such as dif-
ference, probabilistic, harmonic, logistic can also be used. The method allows the
finding of simultaneously the structure of model and the dependence of modeled
system output on the values of most significant inputs of the system.
GMDH, based on the self-organizing principle, requires minimum information
about the object under study. As such, all the available information about this ob-
ject should be used. The algorithm allows the finding of the additional needed in-
formation through the sequential analysis of different models using the so-called
external criteria. Therefore, GMDH is a combined method: it is used the test data
and sequential analysis and estimation of the candidate models. The estimates are
found using relatively small part of the test results. The other part of these results
is used to estimate the model coefficients and to find the optimal model structure.
Although GMDH and regression analysis use the table of test data, the regres-
sion analysis requires the prior formulation of the regression model and its com-
plexity. This is because the row variances used in the calculations are internal cri-
teria. A criterion is called an internal criterion if its determination is based on the
same data that is used to develop the model. The use of any internal criterion leads
to a false rule: the more complex model is more accurate. This is because the
complexity of the model is determined by the number and highest power of its
terms. As such, the greater the number of terms, the smaller the variance. GMDH
uses external criteria. A criterion is called external if its determination is based on
new information obtained using “fresh” points of the experimental table not used
in the model development. This allows the selection of the model of optimum
complexity corresponding to the minimum of the selected external criterion.
Another significant difference between the regression analysis and GMDH is
that the former allows to construct the model only in the domain where the num-
ber of the model coefficients is less than then number of point of the design matrix
48 V.P. Astakhov

because the examination of the model adequacy is possible only when f ad > 0 ,
i.e. when the number of the estimated coefficients of the model, n is less than the
number of the points in the design matrix, m. GMDH allows much wider domain
where, for example, the number of the model coefficients can be millions and all
these are estimated using the design matrix containing only 20 rows. In this new
domain, accurate and unbiased models are obtained. GMDH algorithms utilize the
minimum input experimental information. This input consists of a table having 10-
20 points and the criterion of model selection. The algorithms determine the
unique model of optimal complexity by the sorting out of different models using
the selected criterion.
The essence of the self-organizing principle in GMDH is that the external crite-
ria pass their minimum when the complexity of the model is gradually increased.
When a particular criterion is selected, the computer executing GMDH finds this
minimum and the corresponding model of optimal complexity. As such, the value
of the selected criterion referred to as the depth of minimum can be considered as
the estimate of the accuracy and reliability of this model. If the sufficiently deep
minimum is not reached then the model is not found. This might take place when
the input data (the experimental data from the design matrix) are: (1) noisy; (2) do
not contain essential variables; (3) the basic function (for example, polynomial) is
not suitable for the process under consideration, etc.
The following should be clearly understood if one tries to use GMDH:
1. GMDH is not for casual used as, for example the Tagichi method, i.e. it can-
not be used readily by everyone with no statistical and programming back-
ground. Therefore, it should be deployed in complex manufacturing research
programs of high cost and high importance of the results obtained.
2. GMDH is not a part of modern statistical software packages although its al-
gorithms are available, and thus can be programmed.
3. One a research team is gained some experience with GMDH and corres-
ponding algorithms are developed for the application in a certain field of
manufacturing studies, GMDH becomes a very powerful method of DOE
that can be used at different stages of DOE and, moreover, can be combined
with other important DOE methods, for example with split-plot DOE, to ad-
just GMDH to particular needs.

1.8 Strategy and Principal Steps in Using DOE

The overall strategy of any DOE is presented graphically in Figure 1.24. The basic
approach consists of five phases: Conceptual, Discovery, Breakthrough, Optimiza-
tion, and Validation. The appropriate phase is dependent on the researcher’s goals,
resources, timing, and subject matter knowledge. Note however, a user could se-
quence through all five phases within a single project.
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 49

Each DOE phase offers different results. For instance, in the discovery phase a
researcher’s primary goal would include ‘screening’ for vital input variables.
However, the investigator must be aware of the inability to generate a prediction
equation using low resolution screening arrays, if any variable interactions are at
work in the system. The discovery phase typically focuses on two level, fractional-
factorial arrays to identify the “vital few” variables. Fractional-factorial arrays
range from resolution III to VI. The lower resolution arrays (III and IV), for in-
stance 23-1, and 24-1 are limited in application. These arrays are used primarily to
screen input variables because of their limited capability to quantify two-factor
interactions.

Fig. 1.24. Flow chart of five phases of DOE

50 V.P. Astakhov

Full-factorial and resolution V and VI fractional arrays are more commonly

used in the breakthrough phase. All fractional-factorial arrays are subsets of full-
factorial arrays. The higher resolution fractional arrays (V and VI), such as 26-1,
offer the opportunity to screen larger numbers of variables with a minimal number
of experiments. These designs allow for the identification of some two-factor inte-
ractions with a minimal volume of additional experiments. The breakthrough
phase is capable of generating first order prediction equations which includes
main effects and interaction coefficients. From such data, higher order models
from response surface designs can be pursued.
The objective of DOE is to find the correlation between the response (for ex-
ample, tool life, surface finish, etc.) and the factors included (in metal cutting, for
example, the parameters taken into consideration: the cutting speed, feed, depth of
cut, etc.). To achieve this goal, the following steps should be followed [1, 8, 19]:
1. Select clearly defined problems that will yield tangible benefits. An
unplanned experiment often ends up in total loss of time and money.
2. Specify the required objective. To design an effective experi-
ment/process, one must clearly define the experimental objective(s). Ob-
jectives may include: screening to identify the critical variables; identifi-
cation of critical system interactions; or optimization of one or more
quality characteristics at several levels. Failure to set good objectives
may lead to excessive experiments; failure to identify meaningful quality
characteristics; loss of valuable time and resources; unclear results; and
poorly defined prediction equations.
To set good objectives consider: (1) overall project funding and tim-
ing, (2) the number of quality characteristics to be monitored, (3) the ar-
ray design to clarify the need for basic ‘main effects’ and ‘interaction ef-
fects’ or the need for more detailed quadratic models requiring factor
analysis at multiple levels.
3. Define how the solution delivered is going to be used.
4. Understand as much as possible about the problem and the data set (the
domain). Do not rely solely on statistical experts/commercial DOE soft-
ware manuals: Interaction with subject matter specialists makes profes-
sional insight invaluable.
5. Determine the response variable(s) of interest that can be measured.
6. Determine the controllable factors of interest that might affect the re-
sponse variables and the levels of each factor to be used in the experi-
ment. It is better to include more factors in the design than to exclude
factors, that is, prejudging them to be nonsignificant.
7. Determine the uncontrollable variables that might affect the response va-
riables, blocking the known nuisance variables and randomizing the runs
to protect against unknown nuisance variables. Replicating an experiment
refers to the number of times the combination of input variables are set-
up and performed independently. In other words, the defining operations
of the experiment must be set-up from scratch for each replicate and not
just repeated measurements of the same set-up.
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 51

The more times you replicate a given set of conditions, the more pre-
cisely you can estimate the response. Replication improves the chance of
detecting a statistically significant effect (the signal) in the mist of natural
process variation (the noise). The noise of unstable processes can drown
out the process signal. Before doing a DOE, it helps to assess the signal-
to-noise ratio. The signal-to-noise ratio defines the power of the experi-
ment, allowing the researcher to determine how many replicates will be
required for the DOE. Designs reflecting low power require more repli-
cates.
8. Determine if the data can be measured (collected) with required accura-
cy. Do not start statistical analysis without first challenging the validity
of the data: What can you expect out of garbage input?
9. Design the experiments:
9.1. Do a sequential series of experiments. Designed experiments
should be executed in an iterative manner so that information
learned in one experiment can be applied to the next. For example,
rather than running a very large experiment with many factors and
using up the majority of your resources, consider starting with a
smaller experiment and then building upon the results. A typical se-
ries of experiments consists of a screening design (fractional fac-
torial) to identify the significant factors, a full factorial or response
surface design to fully characterize or model the effects, followed
up with confirmation runs to verify your results. If you make a mis-
take in the selection of your factor ranges or responses in a very
large experiment, it can be very costly. Plan for a series of sequen-
tial experiments so you can remain flexible. A good guideline is not
to invest more than 25 percent of your budget in the first DOE.
9.2. Remember to randomize the runs.
9.3. Make the model as simple as possible—but no simpler.
9.4. Determine the total number of runs in the experiment, ideally using
estimates of variability, precision required, size of effects expected,
etc., but more likely based on available time and resources. Reserve
some resources for unforeseen contingencies and follow-up runs.
10. Perform the experiment strictly according to the experimental design, in-
cluding the initial setup for each run in a physical experiment. Do not
swap the run order to make the job easier.
11. Analyze the data from the experiment using the analysis of variance me-
thod. Do not throw away outliers without solid reasoning: Every piece of
data stores a hidden story waiting to be opened. Let the problem drive the
modeling (i.e., tool selection, data preparation). Stipulate assumptions.
12. Obtain and statistically verify the model. Refine the model iteratively.
13. Define instability in the model (critical areas where change in output is
drastically different for small changes in inputs).
14. Define uncertainty in the model (critical areas and ranges in the data set
where the model produces low confidence predictions/insights). Do not
52 V.P. Astakhov

underestimate the importance of randomization: The adverse effect of

systematic errors can be of vital importance.
15. Always confirm critical findings. After all the effort that goes into plan-
ning, running and analyzing a designed experiment, it is very exciting to
get the results of your work. There is a tendency to eagerly grab the re-
sults, rush out to production and say, "We have the answer!" Before
doing that, you need to take the time to do a confirmation run and verify
the outcome. Good software packages will provide you with a prediction
interval to compare the results within some degree of confidence. Re-
member, in statistics you never deal with absolutes, i.e. there is always
uncertainty in your recommendations. Be sure to double-check your re-
sults. In other words, do not blindly follow statistical conclusions without
taking into account their physical significance, practical implementation,
and economic considerations.
16. Pay particular attention to the presenting of the results. Do make full use
of graphical presentations: A picture can be worth more than hundreds of
words. Do avoid statistical jargon in conclusion and report writing: Prob-
lem language is the only thing that is universal in a corporation.
17. Do not think/hope that one iteration of a time experiment can solve the
problem once and for all: The outcome of one experiment often provides
a direction for further iterations of exploration.

References
1. Montgomery, D.S., Kowalski, S.M.: Design and Analysis of Experiments, 7th edn.
John Wiley & Sons, New York (2009)
2. Antony, Y.: Design of Experiments for Engineers and Scientists. Butterworth-
Heinemann, Oxford (2009)
3. Fisher, R.A.: The Design of Experiments, 9th edn. Macmillan, New York (1971)
4. Fisher, R.A.: The Design of Experiments. Oliver and Boyd, Edinburgh (1935)
5. Halt, A.: A History of Mathematical Statistics. Wiley, New York (1998)
6. Student, Tables for estimating the probability that the mean of a unique sample of ob-
servations lie between any given distance of the mean of the population from which
the sample is drawn. Biometrika 11, p p. 414–417 (1917)
7. Fisher, R.A.: Statistical Methods for Research Workers, 14th edn.(1st edn. 1925).
Hafner Press, New York (1973)
8. Telford, J.K.: A brief introduction to design of experiments. Johns Hopkins Apl. Tech-
nical Digest 27(3), 224–232 (2007)
9. Roy, R.K.: Design of Experiments Using the Taguchi Approach. Wiley-IEEE, New
York (2001)
10. Shina, S.G.: Six Sigma for Electronics Design and Manufacturing. McGraw-Hil, New
York (2002)
11. Ross, P.J.: Taguchi Techniques for Quality Engineering. McGraw-Hill, New York
(1996)
12. Phadke, M.S.: Quality Engineering Using Robust Design. Pearson Education, Upper
Saddle River (2008)
Design of Experiment Methods in Manufacturing: Basics and Practical Applications 53

13. Mukerjee, R., Wu, C.-F.: A Modern Theory of Factorial Designs. Springer, London
(2006)
14. Kotsireas, I.S., Mujahid, S.N., Pardalos, P.N.: D-Optimal Matrices. Springer, London
(2011)
15. Paquete, L.: Experimental Methods for the Analysis of Optimization Algorithms.
Springer, London (2010)
16. Farlow, S.J.: Self-organising Methods in Modeling. Marcel Dekker, New York (1984)
17. Madala, H.R.I., Ivakhnenko, A.G.: Inductive Learning Algorithms for Complex Sys-
tems Modeling. CRC Press, Boca Raton (1994)
18. Wang, H.S., Chang, C.-P.: Desing of experiments. In: Salvendy, G. (ed.) Handbook of
Industrial Engineering. Technology and Operations Management, pp. 2225–2240. John
Willey & Sons, New York (2001)
19. Astakhov, V.P.: Tribology of Metal Cutting. Elsevier, London (2006)
20. Mason, R.L., Gunst, R.F., Hess, J.L.: Statistical Design and Analysis of Experiments
with Application to Engineering and Science. John Wiley and Sons, New York (1989)
21. Montgomery, D.C.: Design and Analysis of Experiments, 5th edn. John Wiley & Sons,
New York (2000)
22. Box, G.E.P., Draper, N.R.: Empirical Model Building and Response Surfaces. Wiley,
Hoboken (1987)
23. Astakhov, V.P., Osman, M.O.M., Al-Ata, M.: Statistical design of experiments in met-
al cutting - Part 1: Methodology. Journal of Testing and Evaluation 25(3), 322–327
(1997)
24. Astakhov, V.P., Al-Ata, M., Osman, M.O.M.: Statistical design of experiments in met-
al cutting. Part 2: Application. Journal of Testing and Evaluation, JTEVA 25(3), 328–
336 (1997)
25. Gopalsamy, B.M., Mondal, B., Ghosh, S.: Taguchi method and ANOVA: An approach
for process parameters optimization of hard machining while machining hardened
steel. Journal of Scietific & Indistrial Research 68, 659–686 (2009)
26. Nalbant, M., Gökkaya, H., Sur, G.: Application of Taguchi method in the optimization
of cutting parameters for surface roughness in turning. Materials & Design 28(4),
1379–1385 (2007)
27. Lin, T.R.: The use of reliability in the Taguchi method for the optimization of the po-
lishing ceramic gauge block. The International Journal of Advanced Manufacturing
Technology 22(3-4), 237–242 (2003)
28. Yang, W.H., Tarng, Y.S.: Design optimization of cutting parameters for turning opera-
tions based on the Taguchi method. Journal of Material Processing Technology 84,
122–129 (1998)
29. Ghani, J.A., Choudhury, I.A., Hassan, H.H.: Application of Taguchi method in the op-
timization of end milling operations. Journal of Material Processing Technology 145,
84–92 (2004)
30. Plackett, R.L., Burman, J.P.: The design of optimum multifactorial experiments. Bio-
metrica 33, 305–328 (1946)
31. Astakhov, V.P.: An application of the random balance method in conjunction with the
Plackett-Burman screening desing in metal cutting tests. Journal of Testing and Evalu-
ation 32(1), 32–39 (2004)
32. Bashkov, V.M., Katsev, P.G.: Statistical Fundamental of Cutting Tool Tests. Machino-
stroenie, Moscow (1985) (in Russian)
33. Astakhov, V.P.: Geometry of Single-Point Turning Tools and Drills: Fundamentals
and Practical Applications. Springer, London (2010)
54 V.P. Astakhov

34. Holman, J.P.: Experimental Methods for Engineers, 6th edn. McGraw-Hill (1994)
35. Box, G., Hunter, W., Hunter, S.: Statistics for Experimenters: Design, Innovation, and
Discovery, 2nd edn. Wiley, New York (2005)
36. Fisher, R.A.: Statistical Methods for Research Workers. Oliver and Boyd, Edinburgh
(1925)
37. Yates, F.: Complex experiments, with discussion. Journal of the Royal Statistical So-
ciety, Series B2, 181–223 (1935)
38. Anbari, F.T., Lucas, J.M.: Designing and running super-efficient experiments: Opti-
mum blocking with one hard-to-change factor. Journal of Quality Technology 40, 31–
45 (2008)
39. Ganju, J., Lucas, J.M.: Randomized and random run order experiments. Journal of Sta-
tistical Planning and Inference 133, 199–210 (2005)
40. Ju, H.L., Lucas, J.M.: Lk factorial experiments with hard-to-change and easy-to-
change factors. Journal of Quality Technology 34, 411–421 (2002)
41. Webb, D.F., Lucas, J.M., Borkowski, J.J.: Factorial experiments when factor levels are
not necessarily reset. Journal of Quality Technology 36, 1–11 (2004)
42. Jones, B., Nachtsheim, C.J.: Split-plot designs: What, why, and how. Journal of Quali-
ty Technology 41(4) (2009)
43. Getting Started with the SAS® 9.2 ADX Interface for Design of Experiments. SAS In-
stitute Inc., Cary, NC (2008)
44. Heisenberg, W.: The Physical Principles of Quantum Theory. University of Chicago
Press, Chicago (1930)
45. von Neumann, J.: Theory of Self Reproducing Automata. University of Illinois Press,
Urbana (1966)
46. Beer, S.: Cybernetics and Management, 2nd edn. Wiley, New York (1959)
47. Madala, H.R., Ivakhnenko, A.G.: Inductive Learning Algorithms for Complex System
Modeling. CRC Press, Boca Raton (1994)
48. Arbib, M.A.: Brains, Machines, and Mathematics, 2nd edn. Springer, New York
(1987)
49. Ivachnenko, A.G.: Polynomial theory of complex systems. IEEE Transactions on Sys-
tems, Man, and Cybernetics SMC-1(4), 284–378 (1971)
50. Ivakhnenko, A.G., Ivakhnenko, G.A.: Problems of further development of the group
method of data handling algorithms. Pattern Recognition and Image Analysis 110(2),
187–194 (2000)
51. Ivakhnenko, A.G., Ivaknenko, G.A.: The review of problems solvable by algorithms of
the group method of data handling. Pattern Recognition and Image Analysis 5(4),
527–535 (1995)
2

Stream-of-Variation Based Quality Assurance

for Multi-station Machining Processes – Modeling
and Planning

J.V. Abellan-Nebot1, J. Liu2, and F. Romero Subiron1

1
Department of Industrial Systems Engineering and Design,
Universitat Jaume I, Av. Vicent Sos Baynat s/n. Castelló, C.P. 12071. Spain
[email protected]
2
Department of Systems and Industrial Engineering,
The University of Arizona, Tucson, AZ 85704, USA
[email protected], [email protected]

In the effort of quality assured product design and implementation, a reliable 3D

manufacturing variation propagation model for multi-station machining processes
(MMPs) is a key enabler to evaluate the output of geometric and dimensional
product quality. Recently, the extension of the stream-of-variation (SoV)
methodology provides quality engineers with a tool to model the propagation of
machining-induced variations, together with fixture- and datum-induced
variations, along multiple stations in MMPs. In this chapter, we present a generic
framework of building the extended SoV model for MMPs. Its application in
manufacturing process planning is introduced and demonstrated in detail through a
3D case study.

2.1 Introduction
Conventionally, product design has been separated from manufacturing process
design in the product development cycle. This product-oriented approach is often
referred to as the over-the-wall design due to its sequential nature of the design
activities. It prevents the integration of design and manufacturing activities, and
causes the increase of production ramp-up time, product change cost and the
degradation of product quality, which is inversely related to the geometrical and
dimensional variations of the key product characteristics (KPCs). In order to
overcome the limitation of this approach, manufacturers have begun to investigate
means of simultaneously evaluating product designs and manufacturing processes
in an attempt to proactively address the potential quality problems in
manufacturing phase, reduce ramp-up times and ensure geometrical product
56 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

quality. For this purpose, research efforts have been conducted to develop reliable
three-dimensional (3D) variation propagation models that integrate both product
and manufacturing process information. Such models can be applied in a variety
of fields, such as manufacturing variation sources diagnosis, process planning,
process oriented tolerancing.
To illustrate the importance of 3D manufacturing variation propagation models,
consider the part design shown in Fig. 2.1 and the given manufacturing process
plan (with its corresponding fixture layouts, locators specifications, cutting-tool
used, thermal conditions of machine-tools, etc.) shown in Fig. 2.2. We consider
two KPCs, i.e., the distance between surfaces S0 and S3, named KPC1, and the
distance between surfaces S6 and S8, named KPC2. In order to ensure the quality
of these two KPCs, the engineers should formulate the following questions:

250

KPC2
0
30 Z
X Z
Y S8 Z
Y

200
S2 Z

KPC1
X S3
Y S5 Y X Y X
Z
S4 S6 Z
Z
ºD Y Y
X
X X X Y
S7
X S1
X
Y
Z
Y
Z S0 KPC2 = 50 ± 0.125 mm
Z KPC1 = 180 ± 0.125 mm

Fig. 2.1. Part geometry and KPCs to be manufactured in a 4-station machining process.
Dimensions in mm

• Is this multi-station machining process (MMP) able to produce KPCs whose

dimensions are within specifications?
• Which are the critical components of the process, named key control
characteristics (KCCs), that have significant impacts on the quality of the
KPCs?
• How can we improve the manufacturing process to generate acceptable KPC
variations? Or in other words, how can we modify the process to be more
robust?
• How can we assign the tolerance of KCCs to ensure the deliver of acceptable
KPC variations?
SoV Based Quality Assurance for MMPs – Modeling and Planning 57

Fig. 2.2. 4-machining station process to manufacture the part

In order to answer these questions, engineers should be able to define

analytically how each potential source of variations affects the KPCs during
machining. In MMPs, the mathematical models that describe the relationships
between dimensional KPCs and KCCs are commonly named 3D manufacturing
variation propagation models. These models provide the engineers with a great
tool for a large number of applications such as manufacturing and product
tolerance allocation, geometric product validation, fault diagnosis, process
planning evaluation and selection, etc. In this chapter, a 3D variation propagation
model for MMPs, named the stream-of-variation (SoV) model, will be presented,
together with its application in process planning. The organization of this chapter
is as follows. Section 2.2 introduces the SoV model in its extended version, where
datum-, fixture- and machining-induced variations are included into the variation
and propagation model. Section 2.3 shows the application of the extended SoV
model in process planning. We present the SoV-based methodology that improves
the robustness of a given process plan. In Section 2.4 previous questions related to
the example shown in Figs. 2.1 and 2.2 are discussed through the application of
the extended SoV model. Finally, Section 2.5 concludes the chapter.
58 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

2.2 3D Variation Propagation Modeling

2.2.1 Fundamentals
Manufacturing variability in MMPs and its impacts on part quality can be modeled
by capturing the mathematical relationships between the KCCs and the KPCs.
These relationships can be modeled with non-linear functions. For instance, a
function f1 can be defined such that y1=f1(u), where y1 is the value of a KPC and
u=[u1, u2, …, un]T are the KCCs in a MMP. By the assumption of small variations,
the non-linear function can be linearized through a Taylor series expansion and the
value of a KPC can be defined as
δf1 (u) δf (u)
y1 = f1 ( u ) + ⋅ (u1 − u1 ) + … + 1 ⋅ (u n − u n ) + ε 1 , (2.1)
δu1 u= u δu n u = u

where ε1 contains the high-order non-linear residuals of the linearization, and the
linearization point is defined by u = [u1 , u2 ,… , un ]T . This linear approximation can
be considered good enough for many MMPs [1]. From Eq. (2.1), the dimensional
variation of a KPC from its nominal value is defined as
δf1 (u) δf (u)
Δy1 = ⋅ Δu1 + … + 1 ⋅ Δu n + ε 1 , (2.2)
δu1 u= u δu n u= u
where Δy1 = y1 − f1 (u) , defines the variations of the KPC, whereas Δu j = u j − u j ,
for j = 1,…, n , defines the small variations of the KCCs in a MMP. Considering that
there are M KPCs in the part whose variations are stacked in the vector
Y = [Δy1 , Δy 2 ,… , Δy M ]T , Eq. (2.2) can be rewritten in a matrix form as
Y = Γ ⋅ U + ε, (2.3)
where U = [Δu1 , Δu 2 ,…, Δu n ]T , ε is the stacked vector of the high-order non-
linear residuals defined as ε = [ε 1 , ε 2 ,…, ε n ]T and Γ is the matrix
T
⎡ ⎡ δ f (u) δ f (u) ⎤
T
⎡ δ f M (u) δ f (u ) ⎤
T
⎤
⎢⎢ 1 ,… , 1 ⎥ ;…; ,… , M ⎥ .
⎢ ⎥
⎢⎢
⎣ ⎣ δ u δ un u = u ⎦⎥ ⎢⎣ δ un u = u δ un u = u ⎥⎦ ⎥
1 u = u
⎦
For MMPs, the derivation of Eq. (2.3) is a challenging task. At the end of the
nineties, researchers from the University of Michigan proposed the adoption of the
well-known state space model from control theory [2, Chapter 11] to mathematically
represent the relationship between the variation sources and the variations of the
machined surfaces at each station, including how the variation of the surfaces
generated at upstream stations influence the surfaces generated at downstream
stations when the upstream surfaces are used as locating datums. In this
representation, dimensional variations of the machined surfaces from nominal values
at station k are defined by a series of 6-by-1 vectors named differential motion
vectors (DMVs), in the form of x k ,i = [(d iR ) T , (θ iR )T ]T , where d iR = [d ixR , d iyR , d izR ]T is
SoV Based Quality Assurance for MMPs – Modeling and Planning 59

a vector of translational variation and θ iR = [θ ixR ,θ iyR ,θ izR ]T is a vector of orientation

variation of the coordinate system (CS) of the ith part surface with respect to (w.r.t.)
a reference CS (R), as it is shown in Fig. 2.3.
According to this representation, ºi denotes the nominal CS and i to the current
CS. The variation of all part surfaces at station k are stacked in the vector
x k = [xTk ,1, … , xTk , i ,…]T . By the adoption of the state space model, the variation of
all part surfaces is related to the potential sources of variation at each machining
station. In a machining station, three main sources of variation are distinguished:
datum-induced variations, fixture-induced variations and machining-induced
variations. These sources of variation produce random deviations between the
cutting-tool paths and the workpiece location on the machine-tool table and thus,
random deviations of the machined surfaces occur.

d iR = [d ixR , d iyR , d izR ]T

zi
θ iR = [θ ixR , θ iyR , θ izR ]T
x k ,i = [(diR )T , (θiR )T ]T z ºi xi i
xºi xk,i yi
z ω ºRi t ºRi ºi
R y ºi
y
x t ºRi ω ºRi
Fig. 2.3. Dimensional variation of a plane surface at station k using a DMV

To illustrate how these three main sources of variation influence on part

quality, we consider an N-station machining process shown in Fig. 2.4 and the kth
machining station with the workpiece and the fixture device shown in Fig. 2.5. At
this kth station, the following variation sources exist.

Fig. 2.4. Manufacturing variation propagation in a MMP

60 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

First, the variations of datum surfaces used for locating the workpiece deviate
the workpiece location on the machine-tool table. This term can be estimated as
x dk +1 = A k ⋅ x k , where x k is the vector of part surfaces variations from upstream
machining stations and A k linearly relates the datum variations with the
machined surface variation due to the locating deviation of the workpiece.
Secondly, the fixture-induced variations deviate the workpiece location on the
machine-tool table and thus, a machined surface variation is produced after
machining. This term can be estimated as x kf +1 = B kf ⋅ u kf , where u kf is the vector
that defines the KCCs related to fixture-induced variations and B kf is a matrix
that linearly relates locator variations with variations of machined surface.
Thirdly, the machining-induced variations such as those due to geometrical and
kinematic errors, tool-wear errors, etc., deviates the cutting-tool tip and thus, the
machined surface is deviated from its nominal values. This term is modeled as
k +1 = B k ⋅ u k , where u k is the vector that defines the KCCs related to machining-
xm m m m

induced variations and B mk is a matrix that linearly relates these KCCs with the
machined surface variations.

xdk +1 = Ak ⋅ x k x kf +1 = B kf ⋅ u kf k +1 = B k ⋅ u k
xm m m

[
xk +1 = Ak ⋅ xk + Bkf Bm ][
k ⋅ (uk ) ( uk )
f T m T
] T

xk +1 = Ak ⋅ xk + Bk ⋅ uk
Fig. 2.5. Sources of variation and state space model formulation for station k

Therefore, for an N-station machining process the derivation of the state space
model can be defined in a generic form as
x k +1 = A k ⋅ x k + B k ⋅ u k + w k , k = 1,…, N , (2.4)
where B k ⋅ u k represents the variations introduced within station k due to the
KCCs and it is defined as [B kf B mk ] ⋅ [(u kf )T (u m T T
k ) ] ; and w k is the unmodeled
system noise and linearization errors.
SoV Based Quality Assurance for MMPs – Modeling and Planning 61

Eq. (2.4) shows the relationship between KCCs and KPCs along a MMP.
Considering that an inspection station is placed after station k − 1 in order to
verify if the workpiece/part is within specifications, then, following the state space
model formulation from control theory, the KPC measurements can be expressed as
y k = Ck ⋅ x k + v k , (2.5)
where y k represents the variations of the inspected KPCs; C k ⋅ x k are the linear
combinations of the variations of workpiece features after the kth station that
define the KPCs; and v k is the measurement noise of the inspection process.
Similar to x k , the vector y k is defined as [ y k ,1 ,…, y k ,q ,…, y k , M ]T , where y k , q is
the inspected variation of the qth KPC and M is the number of KPCs inspected.
For a MMP, Eqs. (2.4) and (2.5) form the generic math-based state space
representation, named in the literature as the SoV model, which allows for integration
of KPCs and KCCs through product and process information such as fixture layout,
part geometry, sequence of machining and inspection operations, etc. Based on this
model, the use of advanced control theory, multivariate statistics and Monte Carlo
simulations enables a large number of applications along product-life cycle (Fig. 2.6).
In the literature, interesting research works can be found about manufacturing fault
identification [3-6], part quality estimation [1, 7], active control for quality variation
reduction [8-11] and process planning and process-oriented tolerancing [12-16].
The SoV model can be presented in its conventional version [1, 7, 17] and in its
extended version [18]. The conventional SoV model includes datum-, fixture- and
machining-induced variations but defines the machining-induced variations as a
generic cutting-tool path variation defined by three translational and three
orientation deviations. The extended SoV model expands the conventional model
by including specific machining-induced variations due to geometric, kinematic
errors, thermal distortions, cutting-tool wear and cutting-tool deflections, etc. In
the next subsections, the derivation of the extended SoV model will be introduced.

xk+1 = Ak ⋅ xk + Bk ⋅ uk + wk
y k = Ck ⋅ x k + v k

Fig. 2.6. Diagram of the SoV model derivation and its applications
62 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

2.2.2 Definition of Coordinate Systems

As explained previously, Eqs. (2.4) and (2.5) define the SoV model in MMPs. The
derivation of the SoV model is achieved by analyzing the KCCs that may be
present in the MMP and relating them with the resulting KPC variations. In order
to define the relationships between KCCs and KPCs, it is necessary to identify the
elements that compose each machining station component (machine-tool, fixture
device and workpiece) and their corresponding CSs. For this purpose, we consider
the kth machining station of a MMP, composed of a 3-axis vertical machine-tool,
a fixture device and a workpiece, as it is shown in Fig. 2.7. For the machining
station k, the following CSs can be defined:

Design Coordinate System (DCS). The nominal DCS, denoted as D , define the
reference for the workpiece features during design. The definition of D usually
depends on the nominal geometry of the part and it is usually defined at an
accessible corner. As this CS is only used in design, this CS cannot be deviated.

Reference Coordinate System (RCS). The nominal and true RCS, denoted as
Rk and Rk , respectively, define the reference for the workpiece features (Fig.
2.7d) at station k. To facilitate the model derivation, the Rk is defined as the local
coordinate system of the primary datum feature at station k. In a 3-2-1 fixture
layout, the primary datum is the main workpiece surface used to locate the part at
the machine-tool table [19, Chapter 3]. The Rk is defined similarly according to
the actual part geometry.

Fixture Coordinate System (FCS). The nominal and true FCS at station k,
denoted as Fk and Fk , respectively, define the physical position and orientation
of the fixture device according to the fixture layout. Fig. 2.7d shows the FCS for a
fixture layout based on the 3-2-1 principle.

Local Coordinate System (LCSj). The nominal and true LCSj at station k,
denoted as Lkj and Lkj , define the physical position and orientation of the jth
nominal and actual machined surface of the part respectively (Fig. 2.7c). For
planar surfaces the Z-axis of Lkj is commonly defined normal to the surface.

Machine-Tool Coordinate System (MCS). The MCS at station k, denoted as

M k , define the physical position and orientation of the reference CS for
machine-tool movements. The origin of the M k is located at the locating origin
of the nominal machine-tool table, with its Z-axis normal to the table and pointing
SoV Based Quality Assurance for MMPs – Modeling and Planning 63

upward, its X-axis parallel to the long axis of the table and pointing to its positive
direction, and its Y-axis defined according to the right hand rule, as shown in
Fig. 2.7a. In this chapter, it is assumed that M k serves as the reference at station
k and thus will not deviate.

Axis Coordinate System (ACSi). The nominal and true ACSi of the i-axis used at
station k, denoted as Aki and Aki , respectively, define the physical position and
orientation of the i-axis of the machine-tool. The origin of the Aki is located at
the geometrical center of the joint of the i-axis. For prismatic joints, the axes of
Aki have the same orientation as that of the M k . An example of Aki 's of a 3-
axis vertical machine-tool is shown in Fig. 2.7a. The Aki is similarly defined for
an actual axis.

Spindle Coordinate System (SCS). The nominal and true SCS at station k,
denoted as S k and S k , respectively, define the physical position and orientation
of the spindle during machining. The origin of the S k is located at the
geometrical center of the spindle and the orientations of axes are identical to that
of the Z-axis of the machine-tool, as shown in Fig. 2.7b. The S k is defined
similarly for the actual spindle.

Cutting-Tool Coordinate System (CCS). The nominal and true CCS at station k,
denoted as C k and C k , respectively, define the physical position and orientation
of the cutter tip center during machining. The origin of the C k is located at the
cutter tip center and the orientations of its axes are identical to that of a S k , as
shown in Fig. 2.7c. The C k is defined similarly for the actual cutting-tool.

Cutting-Tool Tip Coordinate System (TPCS). The nominal and true TPCS at
station k, denoted as Pk and Pk , respectively, defines the physical position and
orientation of the cutting-tool tip. The origin of the Pk is located at the center of
the cutting edge that is used to generate feature j, and the orientations of its axes
are identical to that of Lkj . Please note that, when machining feature j at station k,
the cutting-tool tip removes material generating the machined feature which is
defined by the Lkj . Thus, the position and orientation of the Pk defines the
position and orientation of Lkj , as shown in Fig. 2.7c.
64 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

z
y

Head Azk x

y Spindle

ºMk x
y Work Column
z
space

Axk Table
x
z Saddle z

y
y
Ayk x Ak i
x
is
Bed ax
Y-
X-ax Prismatic joint
is

(a) (b)
z z

Sk y Sk y
Spindle Spindle
x x
z
Workpiece

z
z z ºD y
Cutting-tool x Cutting-tool

Pk y
y
Pk z y
x Rk x
y
x Ck x Ck y z
x j x
z Lk j
Lk Fk x

z
y
y
x
Frontal machining Peripheral machining
operation operation Fixture
y z

(c) (d)
Fig. 2.7. Example of CSs involve in a 3-axis vertical machine-tool

2.2.3 Derivation of the DMVs

In previous subsection, it has been shown the different elements with their
corresponding CSs that are involved during machining feature j. In this
subsection, the extended SoV model is derived by defining the relationships
among these CSs with their deviations from nominal values. For a 3-axis vertical
machine-tool, a generic illustration of these relationships is shown in Fig. 2.8.
From this figure, two main subchains of CSs can be identified. The first chain,
defined from M k to Pk , represents how the machining-induced variations
SoV Based Quality Assurance for MMPs – Modeling and Planning 65

x y
H ºº AA ykk H ºº AA z kk x AA z kk
y

xºAMx k
k
HººM
Ax
k
z
k
H ºº SAk k z
x SAk k

H ºº M
Fk
k
x CP k
k
H ºº CS k
x ºFM k k
k

H ºº FRk x CS k
k Hºº RDk H ºº CP k k

x FRk k
k

Fig. 2.8. Relationships between the different CSs in a 3-axis vertical machine-tool

deviate the cutting-tool tip w.r.t. the machine-tool CS. The most common
machining-induced variations are due to geometric and kinematic errors of
machine-tool axes, thermal distortions, cutting-tool deflections and cutting-tool
wear, which induce deviations of the CSs Aki , S k , C k and Pk from their
respective nominal values. The second chain, defined from M k to Rk ,
represents how fixture- and datum-induced variations deviate the workpiece
location w.r.t. the MCS. Note that in order to represent one CS w.r.t. another CS, a
homogeneous transformation matrix (HTM) is used. How to derive the HTMs is
explained in Appendix 2.1.
In order to derive the machined surface variation as a function of the variations
of all CSs involved in these chains, we consider the following Corollary from [1].
Corollary 1: Consider the CS R , 1 and 2 as shown in Fig. 2.9. Consider now
that CS 1 and CS 2 are deviated from nominal values. Noting the variation of CS
1 and CS 2 w.r.t. R as x1R and x 2R , respectively, then, the variation of CS 2
w.r.t. 1 in vector form can be formulated as
⎛ ⎛ 1 ⎞ T ⎛ 1 ⎞T ⎛ 1 ⎞ ⎞
⎜ − ⎜ R ⎟ ⎜ R ⎟ ⋅ ⎜ tˆ ⎟ I 3×3 0 ⎟ ⎛ R⎞
x12 = ⎜⎜ ⎝ ⎟ ⋅ ⎜ x1 ⎟,
2⎠ ⎝ 2⎠ ⎝ 2⎠
T ⎟ ⎜ xR ⎟
− ⎛⎜ R 1 ⎞⎟ 0 I 3×3 ⎟⎟ ⎝ 2 ⎠
(2.6)
⎜⎜ 0
⎝ ⎝ 2⎠ ⎠
= D12 ⋅ x1R + x 2R ,
66 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

1
where R is the rotation matrix of 2 w.r.t. 1 , I 3×3 is a 3× 3 identity matrix
2

and tˆ 1
is the skew matrix of vector t 1
(see Appendix 2.2). The proof of this
2 2
Corollary can be found in [1].
Applying Corollary 1, the machined surface variation at station k defined as the
R
variation of CS Lkj (denoted as Lk hereafter) w.r.t. Rk , denoted as x L k , can be
k
obtained as
R R Mk Mk
x Lk = DLk ⋅ x R + xL , (2.7)
k k k k

Mk
where x R represents the variation of the position and orientation of the
k
workpiece w.r.t. the MCS due to fixture- and datum-induced variations, and
Mk
xL represents the overall cutting-tool path variation due to machining-induced
k
variations when manufacturing feature j. Note that, according to the previous CS
Mk Mk
definition, x L ≡ xP .
k k
Mk
As shown in Fig. 2.8, x R will depend on the CSs variations that define the
k
Mk
chain from M k to Rk . And similarly, x P will depend on the variations of the
k

CSs from M k to Pk . In order to express these DMVs as functions of the DMVs

of all the CSs for each chain, the following Corollary from [1] is applied.

1
x1R
H ºR1
º1
H ºº12

H ºº12

H ºR2
º2
x12
R x 2R 2

Fig. 2.9. Differential motion vector from CS 2 to CS 1 if both CSs deviate from nominal
values
SoV Based Quality Assurance for MMPs – Modeling and Planning 67

Corollary 2: Consider the CSs R , 1 and 2 as shown in Fig. 2.9, with CSs 1 and
2 deviating from their nominal positions and orientations. Noting the variation of
CS 1 w.r.t. R as x1R and the variation of CS 2 w.r.t. CS 1 as x12 , then, the
variation of CS 2 w.r.t. R can be formulated as

⎛ ⎛ 1 ⎞T T ⎞
⎜⎜ R ⎟ − ⎛⎜ R 1⎞
⎟ ⋅ ⎛⎜ tˆ 1⎞
⎟ I 3×3 0 ⎟ ⎛ R⎞
x 2R = ⎜⎜ ⎝ ⎟ ⋅ ⎜ x1 ⎟,
2⎠ ⎝ 2⎠ ⎝ 2⎠
T ⎟ ⎜ x1 ⎟
⎜⎜ 0 ⎛R 1 ⎞ I 3×3 ⎟⎟ ⎝ 2 ⎠
(2.8)
⎜ 2⎟ 0
⎝ ⎝ ⎠ ⎠
= T21 ⋅ x1R + x12 .

The proof of this Corollary can be found in [1]. Applying repeatedly the Corollary 2,
Mk Mk
the DMVs x R and x P can be respectively rewritten as
k k

Mk F Mk F
xR = TR k ⋅ x F + x Rk , (2.9)
k k k k

Mk C S Az Mk Az S C
xP = TP k ⋅ (TC k ⋅ (TS k ⋅ x + x S k ) + xCk ) + x Pk . (2.10)
k k k k Akz k k k

Substituting Eqs. (2.10) and (2.9) in Eq. (2.7), the variation of the feature Lk
w.r.t. Rk can be expressed as a function of the DMVs of all the CSs that are
Mk Az S C
involved between Lk and Rk CS, such as x , x S k , xCk and x Pk . These
Akz k k k

DMVs are functions of different sources of variation such as fixture locator

variations, machine-tool inaccuracy, thermal state of the machine-tool, cutting-
tool wear, cutting-tool deflections, etc. In the next subsection, the derivation of the
DMVs for each CS will be presented in detail.

2.2.3.1 Fixture-Induced Variations

Mk
The DMV x F refers to the variations of the FCS due to fixture inaccuracies,
k
Fk
and it can also be described as the DMV x F since M k is not deviated. In
k
Fk
order to derive x F as a function of fixture layout and fixture variations, the
k
following assumptions are considered: i) locating surfaces are assumed to be
perfect in form (without form errors) and there is no deformations of locators; ii)
locators are assumed to be punctual and distributed according to the 3-2-1
workholding principle [19, Chapter 3]; iii) only variations based on small
displacements from nominal values are considered.
68 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

Let us consider a 3-2-1 fixture device where ri defines the contact point
between the ith locator and the workpiece, and ni defines the normal vector of the
workpiece surface at the ith contact point, all expressed w.r.t. Fk . Following the
Fk
research work in [20], a small perturbation in the location of Fk ( x F ) due to a
k
small variation in a fixture locator in the direction of movement constraint
(direction normal to the locating surface defined by ni ), denoted as Δli , can be
mathematically expressed as

F
Δli = w iT ⋅ x F k , (2.11)
k

where

w iT = [nTi , (ri × n i )T ], (2.12)

and × is the cross product operator. Δli is defined as

Δli = nTi ⋅ Δri , (2.13)

where Δri is the position variation of the locator i. The deterministic localization
Fk
condition (a unique solution of x F ) requires that Eq. (2.11) is satisfied for all
k
locators. Thus, considering all locators, Eq. (2.11) becomes
F
Δl = G T ⋅ x F k , (2.14)
k

where Δl = [Δl1 , Δl2 ,… Δl6 ]T and G = [ w1 , w 2 ,… w 6 ]T . Matrix G is called locator

matrix in [20]. In order to ensure a deterministic localization, the locator matrix
G must not be singular, indicating that all {w1 ,…, w 6 } are linearly independent
[20]. Thus, for a deterministic localization, matrix G is invertible, and the
variation of Fk is related to the variation of fixture locators by the following
expression

= (GT ) −1 ⋅ Δl.
Fk
xF (2.15)
k

Eq. (2.15) can be rewritten in the form

Fk Mk f
xF = xF = B k1 ⋅ [Δl1 , Δl 2 ,…, Δl6 ]T . (2.16)
k k
SoV Based Quality Assurance for MMPs – Modeling and Planning 69

f
For a generic 3-2-1 locating scheme shown in Fig. 2.10a, matrix B k1 can be
straight forward derived following the methodology explained above, resulting in
the matrix:
⎛ (l 2 y − l 3 y ) ⋅ l 5 z − (l1 y − l 3 y ) ⋅ l 5 z (−l 2 y + l1 y ) ⋅ l 5 z − l5 y l4 y ⎞
⎜ 0⎟
⎜ C C C ( −l 5 y + l 4 y ) (−l 5 y + l 4 y ) ⎟
⎜ − (l − l ) ⋅ l (l1x − l 3 x ) ⋅ l 6 z − (−l 2 x + l1x ) ⋅ l 6 z l6 x − l6x ⎟
⎜ 2x 3x 6z
1⎟
⎜ C C C ( −l 5 y + l 4 y ) (−l 5 y + l 4 y ) ⎟ (2.17)
⎜ (l l − l l ) − (−l 3 x l1 y + l 3 y l1x ) (−l1 y l 2 x + l 2 y l1x ) ⎟
⎜ 3 y 2 x 2 y 3x 0 0 0⎟
f ⎜ C C C ⎟
B k1 = ⎜ − (l − l ) (l1x − l 3 x ) − (−l 2 x + l1x ) ⎟,
2x 3x
⎜ 0 0 0⎟
⎜ − (l C− l ) C
(l1 y − l 3 y )
C
− (−l 2 y + l1 y ) ⎟
⎜ 2y 3y
0 0 0⎟
⎜ C C C ⎟
⎜ −1 1 ⎟
⎜ 0 0 0 0⎟
⎜ ( −l 5 y + l 4 y ) (−l 5 y + l 4 y ) ⎟
⎜ ⎟
⎝ ⎠

where C = l3 x l1 y − l1 y l 2 x + l3 y l 2 x + l 2 y l1x − l 2 y l3 x − l3 y l1x . Note that the position of

locator i is described by the vector l i = [lix , liy , liz ] expressed in the FCS.

Tertiary
datum Workpiece
Secondary
datum ZC
LE XC
CSC
ZB YC
CSB ºD
YB
XB CSA R XA
LF
YA
ZA Primary
LD datum

(a)
(b)
Fig. 2.10. a) 3-2-1 fixture layout based on locators; b) Workpiece datums (primary, second-
ary and tertiary). CSs centered on each face.

2.2.3.2 Datum-Induced Variations

For 3-2-1 fixture layouts based on locators (Fig. 2.10a), the deterministic location
of the part is ensured when workpiece touches the six locators. Due to datum
inaccuracies, workpiece location can be deviated from its nominal values.
Assuming prismatic surfaces, the influence of the datum variations on workpiece
location can be obtained as follows [1].
70 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

Considering a 3-2-1 fixture device and a prismatic workpiece as shown in

Fig. 2.10a and 2.10b, respectively, the following constrains apply:

• Locators l 4 and l5 touch workpiece surface B , and thus the third component
of the coordinate of l 4 and l5 w.r.t. the CS B are zero. Mathematically, this is
expressed as p [ ]
~B
l = 0 and p
4 (3)
~B
l [ ] ~ B = [l B , l B , l B ,1]T and []
= 0 , where p
5 (3) l ix iy iz i
⋅ (3)

[ ]
denotes the third component of the vector.
~C
• Similarly, locator l6 touches workpiece surface C , so p = 0.
l 6 (3)

According to these constrains, the contact points can be mathematically

expressed as
~B = HB ⋅ HA ⋅ p
p ~F , (2.18)
l
4 A F l 4

~B = HB ⋅ H A ⋅ p
p ~F , (2.19)
l
5 A F l 5

~C = HC ⋅ H A ⋅ p
p ~F . (2.20)
l 6A F l 6

According to [1], H BA is defined as:

( )
H BA = δH BA
−1
·H B
A
,
(2.21)
= (I 4× 4 − Δ BA ) ⋅ H B
,
A

where δH BA is the HTM that defines the small translational and orientational
deviations of CS B w.r.t. A due to the variation from nominal values of B and A;
and Δ BA is the differential transformation matrix (DTM) of B w.r.t. A (see
Appendix 2).

H FA it is defined as

H FA = H A
·δH FA ,
F
(2.22)
=H A
·(I 4× 4 + Δ FA ).
F

Substituting Eqs. (2.21) and (2.22) in Eq. (2.18),

~ B = (I ~F .
4× 4 − Δ B ) ⋅ H ⋅H ⋅ (I 4× 4 + Δ FA ) ⋅ p
A B A
p l l (2.23)
4 A F 4
SoV Based Quality Assurance for MMPs – Modeling and Planning 71

[ ]
By neglecting the second-order small values and considering the contact
~B
between surfaces ( p = 0 ), the following equation applies
l 4 (3)

[p~ ]
B
l4
(3)
= ⎡⎢(−Δ BA ⋅ H
⎣
B
F
+H B
F
⋅ Δ FA + H B
F
~F
)⋅p l4
⎤ = 0.
⎥⎦ (3) (2.24)

As the X coordinate of the location of locator 4 w.r.t. the CS F is zero, through the
HTM H B , the term ⎡⎢H B ⋅ p ~ F ⎤ becomes zero and thus, Eq. (2.24) is rewritten
l4 ⎥
F ⎣ F ⎦ (3)
as
⎡(−Δ A ⋅ H B
+H B ~ F ⎤ = 0,
⋅ Δ FA ) ⋅ p (2.25)
⎢⎣ B F F l4 ⎥
⎦ (3)

and thus,
⎡( H B ~ F ⎤ = ⎡Δ A ⋅ H
⋅ Δ FA ) ⋅ p B ~F ⎤ .
⋅p (2.26)
⎢⎣ F ⎦ (3) ⎢⎣
l4 ⎥ B F l4 ⎥
⎦ (3)

Following the same procedure, Eqs. (2.27) and (2.28) can be derived for locator
l5 and l6 , respectively.

⎡(H B ~ F ⎤ = ⎡Δ A ⋅ H
⋅ Δ FA ) ⋅ p B ~F ⎤ ,
⋅p (2.27)
⎢⎣ F ⎦ (3) ⎣⎢
l5 ⎥ B F l5 ⎥
⎦ (3)

⎡( H C ~ F ⎤ = ⎡Δ A ⋅ H
⋅ Δ FA ) ⋅ p C ~F ⎤ .
⋅p (2.28)
⎢⎣ F ⎦ (3) ⎣⎢
l6 ⎥ C F l6 ⎥
⎦ (3)

Note that for Eqs. (2.26-2.28), we are interested in evaluating the DTM Δ FA ,
which shows the effect of datum variations on workpiece location. Furthermore,
note that from the six parameters of Δ FA (three translational and three orientation
deviations), only three are unknown since datum variations of surface B and C
only influence on the X and Y positioning of the workpiece and on the rotation
about the Z-axis, all expressed from CS A . Thus, the three unknown parameters
that depend on datum fixture variations of surfaces B and C can be obtained
solving Eqs. (2.26-2.28). After solving, the variation of FCS w.r.t. CS A can be
rewritten by the effects of each datum feature in a vector form as

x FA = A1 ⋅ x BA + A 2 ⋅ xCA . (2.29)

As the primary datum (CS A ) also defines CS Rk , Eq. (2.29) is rewritten,

considering the station subscript, as
R R R
x Fk = A1k ⋅ x Bk + A k2 ⋅ xCk . (2.30)
k k k
72 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

For the 3-2-1 locating scheme based on locators shown in Fig. 2.10a, Eqs.
(2.26-2.28) were solved obtaining the following matrices A1k and A 2k . These
results can also be found in [1].
⎛ l5 y ⋅ (l5 z − l4 z ) ⎞
⎜0 0 − 1 LE /2 l5 z + LF /2 + 0⎟
⎜ (l4 y − l5 y ) ⎟
⎜ l6 x ⋅ (l4 z − l5 z ) ⎟
⎜0 0 0 − l6 x 0⎟
⎜ l4 y − l5 y ⎟ (2.31)
Ak1 = ⎜ 0 0 0 0 0 0 ⎟,
⎜ ⎟
⎜0 0 0 0 0 0⎟
⎜0 0 0 0 0 0⎟
⎜ − (l4 z − l5 z ) ⎟
⎜0 0 0 1 0⎟
⎜ l4 y − l5 y ⎟
⎝ ⎠

⎛0 0 0 0 0 0⎞
⎜ ⎟
⎜0 0 − 1 − l6 z − LF /2 l6 x − LD /2 0 ⎟
⎜0 0 0 0 0 0⎟
Ak2 = ⎜ ⎟. (2.32)
⎜0 0 0 0 0 0⎟
⎜ ⎟
⎜0 0 0 0 0 0⎟
⎜0 0 ⎟⎠
⎝ 0 0 0 0

2.2.3.3 Geometric, Kinematic and Thermal-Induced Variations

Machine-tools are typically composed of a kinematic chain of multiple

translational and rotational axes that are designed to have only one degree of
freedom of moving. However, due to geometric inaccuracy, misalignments and
thermal effects, all these axes have actually three translational and three rotational
deviations. Intensive research has been conducted to estimate the positioning
errors of a cutting-tool in multi-axis machine-tools using rigid-body kinematics
and HTMs [21-23]. With these methods, each axis of a machine-tool, relative to
each other or to a reference CS, is modeled using HTM. After determining the
HTMs of all the movable links, they are multiplied successively from the last axis
to the reference CS of the machine-tool, the MCS, resulting in the actual position
of the CS of the last axis w.r.t. the MCS. Considering a 3-axis vertical machine-
tool, the final position of Akz w.r.t. M k can be defined as

Mk Mk Mk
H =H ⋅ δH ,
Akz Akz Akz
y y
(2.33)
Mk Mk Akx Akx A A
=H ⋅ δH ⋅H ⋅ δH ⋅H k ⋅ δH k .
Akx Akx A
y
A
y Akz Akz
k k
SoV Based Quality Assurance for MMPs – Modeling and Planning 73

With the assumption of rigid-body kinematics and small-angle approximation, any

j
A
δH k
i can be generally defined as a product of three HTMs [21-23], i.e.,
Ak

⎛ 1 − ε zi ε yi δ xi ⎞ ⎛ 1 − ε z (i ) ε y (i ) δ x (i ) ⎞
⎜ ⎟ ⎜ ⎟
Aj ⎜ ε 1 − ε xi δ yi ⎟ ⎜ ε z (i ) 1 − ε x (i ) δ y (i ) ⎟
δH ki = ⎜ zi ⋅ ⋅
Ak −ε ε xi 1 δ zi ⎟ ⎜ − ε y (i) ε x (i ) 1 δ z (i ) ⎟
⎜ yi ⎟ ⎜ ⎟
⎜ 0 1 ⎟⎠ ⎜⎝ 1 ⎟⎠ (2.34)
⎝ 0 0 0 0 0
⎛ 1 − ε zt (t , T1 ,..., Tm , i) ε ty (t , T1 ,..., Tm , i ) δ xt (t , T1 ,..., Tm , i ) ⎞⎟
⎜
⎜ ε zt (t , T1 ,..., Tm , i ) 1 − ε zt (t , T1 ,..., Tm , i ) δ yt (t , T1 ,..., Tm , i) ⎟
⋅⎜ t ⎟.
⎜ − ε y (t , T1 ,..., Tm , i ) ε xt (t , T1 ,..., Tm , i ) 1 δ zt (t , T1 ,..., Tm , i ) ⎟
⎜ ⎟
⎝ 0 0 0 1 ⎠

The first HTM describes the mounting errors of the i-axis w.r.t. the previous j-
axis. The mounting errors are position and orientation errors due to assembly
errors and they are not dependent on the carriage position. Mounting errors can be
represented by three possible angular variations, ε xi (rotation around the X-axis),
ε yi (rotation around the Y-axis) and ε zi (rotation around the Z-axis), and three
offsets ( δ xi , δ yi , δ zi ), as it is shown in Fig. 2.11 for the X-axis carriage. The
second HTM represents the motional variations, which include the terms δ p (q)
and ε p (q ) . δ p (q) refers to the positional variation in the p-axis direction when
the prismatic joint moves along the q-axis and is a function of the position of the
q-axis. ε p (q ) refers to the angular variation around the p-axis when the q-axis
moves and it is also a function of the position of the q-axis. The third HTM
describes the geometrical variations due to thermal effects, whose components are
defined as δ tp (t , T1 ,..., Tm , q) and ε tp (t , T1 ,..., Tm , q) for position and angular
variations around the p-axis when the q-axis moves, respectively, and include
scalar thermal components and position-dependent thermal components [23].
Mathematically, δ tp (t , T1 ,..., Tm , q ) and ε tp (t , T1 ,..., Tm , q) are generally defined as

δ pt (t , T1 , … , Tm , q ) = f 0pq (T1 , … , Tm , t ) + f1pq (T1 , … , Tm , t ) ⋅ q + f 2pq (T1 , … , Tm , t ) ⋅ q 2 + … ,

(2.35)
ε tp (t , T1 , … , Tm , q ) = g 0pq (T1 , … , Tm , t ) + g1pq (T1 , … , Tm , t ) ⋅ q + g 2pq (T1 , … , Tm , t ) ⋅ q 2 + …

The term f 0pq (T1 , …, Tm , t ) and g 0pq (T1 , … , Tm , t ) are scalar thermal components
that model the position variation on the p-axis when the q-axis moves and are
function of operation time, t, and the temperatures T1 , …, Tm at different locations
on the machine-tool structure. The position-dependent thermal components are
defined by the terms f1pq (T1 , … , Tm , t ) ⋅ q + f 2pq (T1 , …, Tm , t ) ⋅ q 2 + … and
g1pq (T1 , …, Tm , t ) ⋅ q + g 2pq (T1 , … , Tm , t ) ⋅ q 2 + … and they model the position
variation on the p-axis when the q-axis moves.
74 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

From Eq. (2.34), it is shown that geometrical variations due to kinematic and
thermal effects may present non-linear relationships. In order to include these
sources of variation into the SoV model, a linearization should be conducted based
on three important assumptions. Firstly, it is assumed that the geometric-thermal
variations are modeled when the machine-tool is warmed-up adequately and thus,
the effect of the time on the thermal variations can be neglected. Secondly, it is
assumed that the workpiece is repeatedly placed in the same region inside the
allowable work space of the machine-tool table, so it is only expected small
variations in the placement of the workpiece. Thirdly, it is assumed that
geometric, kinematic and thermal variations do not change drastically along the
travels at any i-axis for the region where the workpiece is repeatedly placed on the
machine-tool table (the experimentation in [23] holds this assumption), so the
geometric-thermal variations in the machine-tool axis can be linearized without
significant loss of precision. Under these assumptions, the motional variations
δ (δ p (q))
δ p (q) and ε p (q ) are linearized as δ p (q0 ) + ·Δq and
δq q=q 0
δ (ε p (q))
ε p ( q0 ) + ·Δq , respectively, where q0 is the nominal placement of
δq q = q0

the workpiece on the q-axis, and Δq is the admissible variation range of the
workpiece placement along the q-axis. The thermal-induced variations
δ pt (t , T1 ,..., Tm , q) and ε tp (t , T1 ,..., Tm , q) from Eq. (2.35) can be linearized as

δ pt (ΔT1 ,..., ΔTm , Δq) = C0pq + C1pq ⋅ ΔT1 + … + Cmpq ⋅ ΔTm + Cmpq+1 ⋅ Δq,
(2.36)
ε tp (ΔT1 ,..., ΔTm , Δq) = D0pq + D1pq ⋅ ΔT1 + … + Dmpq ⋅ ΔTm + Dmpq+1 ⋅ Δq,

⋅) and D(⋅) are constants, ΔTc is the variation of the cth temperature at
where C(pq pq

the machine-tool structure from its nominal values Tc where c = 1,…, m .

0
As a conclusion, the resulting position and orientation deviation defined by the
Mk
matrix δH , considering the linearization, can be obtained from Eq. (2.33).
Akz

This matrix will be defined in the form of

⎛ Mk Mk Mk ⎞
⎜ 1 −θ θ d ⎟
⎜ Akz z Akz y Akz x⎟
⎜ M M M ⎟
θ k −θ z k
= ⎜ Akz z ⎟.
Mk 1 d zk
δH (2.37)
Akz ⎜ Ak x Ak y
⎟
⎜ Mk Mk M ⎟
⎜ − θ Az y θ
Akz x
1 d zk
Ak z ⎟
⎜ k
⎟
⎝ 0 0 0 1 ⎠
SoV Based Quality Assurance for MMPs – Modeling and Planning 75

For a generic 3 -axis vertical machine-tool, Eq. (2.37) can be rewritten in a

DMV as
Mk m
x = B k 1 ⋅ [ΔT1k ,…, ΔTmk , Δr k ]T . (2.38)
Akz

In Eq. (2.38), Δr k refers to the admissible variation range of workpiece

placement at the three axes and thus, Δr k = [Δx k , Δy k , Δz k ] . Please note that we
discard the systematic variations due to linealization such as the terms C0pq , D0pq ,
δ p (q0 ) , ε p (q0 ) , since they can be eliminated by an initial calibration.
For practical purposes, geometric and thermal parameters related to geometric-
thermal induced variations can be estimated from experimental data. Research
works [23-26] show in detail the derivation of these parameters.

δzx
Z1
X1
εzx δxx
δyx εxx
Y1 εyx
O1

Z
X
Y

Fig. 2.11. Position and orientation deviations of a machine-tool carriage system due to
mounting errors

2.2.3.4 Spindle Thermal-Induced Variations

The spindle thermal variations are an important contributor to the total thermal-
induced variations during machining due to the large amounts of heat generated at
high-speed revolutions [27]. Spindle thermal expansion produces three
translational and two rotational drifts to the spindle CS [23]. This variation is
represented as the deviation of S k w.r.t. Akz and is proportional to the increase
of the spindle temperature, denoted as ΔTs , from nominal conditions. At station k,
this variation is defined by the DMV
Az
k
[
x S k = f1k (ΔTsk ) f 2k (ΔTsk ) f 3k (ΔTsk ) f 4k (ΔTsk ) f 5k (ΔTsk ) f 6k (ΔTsk ) ] (2.39)
T

≈ [Cf ]
T m
x
k
Cf yk Cf zk Cfαk Cf βk 0 ⋅ ΔTsk = B k 2 ⋅ ΔTsk ,
76 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

where f i k (⋅) is a function that relates position and orientation deviations of S k

with the variation of the spindle temperature from nominal conditions. Cf i k are
proportional coefficients linearizing the f i k (⋅) functions and the f 6k (ΔTsk ) is
considered as 0 since the Z-axis is the cutting-tool rotation axis. These
coefficients can be obtained through experimentation [23].

2.2.3.5 Cutting Force-Induced Variations

The geometric variations of the machined workpieces due to cutting force-induced

variations can be modeled considering the cutting-tool deflection occur during the
machining process (Fig. 2.12). Various methods for cutting-tools with different
complexity have been applied to model the deflection variation [28-30]. To
represent these variations in the SoV model, it is necessary to describe position and
orientation deviations of the cutting-tool due to deflection. Assuming finishing
operations, where the depth of cut is small and it can be considered insignificant in
comparison to the length of the cutting-tool overhang, and simplifying that the
cutting force acts at the tool tip, the cutting-tool can be modeled as a unique
cantilever beam defined by the equivalent tool diameter [29]. Then, the
displacement of the cutting-tool perpendicular to the cutting-tool axis is proportional
to the cutting force according to the following equation [31, Chapter 7]

F ⋅ L3 64 ⋅ F ⋅ L3
δr = = , (2.40)
3⋅ E ⋅ I 3⋅π ⋅ E ⋅ D4

where E is the Young's Modulus for the material tool; L3 /D 4 is the tool
slenderness parameter, where D is the equivalent tool diameter [29] and L is the
overhang length; and F is the cutting force perpendicular to the tool axis.
Furthermore, the rotation of the tool tip around the θ -axis perpendicular to the
cutting-tool axis is defined as [31, Chapter 7],

F ⋅ L2 64 ⋅ F ⋅ L2
δθ = = , (2.41)
2 ⋅ E ⋅ I 2 ⋅π ⋅ E ⋅ D4
where F is the force applied at the tool tip perpendicular to the plane defined by
the θ -axis and the cutting-tool axis. As a conclusion, C k is deviated due to the
cutting force-induced deflection. The variation of C k w.r.t. S k can be expressed
by the DMV
S m
xCk = B k 3 ⋅ [ΔFxk ΔFyk ]T , (2.42)
k

m
where B k 3 = [C1 0; 0 C1; 0 0; 0 C2 ; C2 0; 0 0] , and C1 and C2 are
64 L3 3C
defined as C1 = and C2 = 1 . ΔFxk and ΔFyk are the variation of the
3πED 4 2L
cutting force in X and Y direction from nominal conditions, respectively.
SoV Based Quality Assurance for MMPs – Modeling and Planning 77

The parameters required to model the force-induced variations can be obtained

directly by the geometry of the cutting-tool. In order to obtain the Young's
modulus, an experimentation may be required since the manufacturers of cutting-
tools do not usually provide this value. In [29], two experimental methods are
proposed to obtain the value of Young's modulus for each cutting-tool.

Fig. 2.12. Example of cutting force-induced deviation in a ball end-milling operation

2.2.3.6 Cutting-Tool Wear-Induced Variations

Cutting-tool wear is another important factor affecting part dimensional quality.

Different types of wear can be defined in cutting-tools, such as crater wear, flank
wear, etc. Among them, the flank wear ( VB ) is the most common tool wear
measurement in industry to keep parts within dimensional specifications [32].
However, the impact of VB on part dimensional quality highly depends on the
type of machining operation and the geometry of the cutting-tool.
According to the CSs defined previously, the VB modifies the geometry of the
tool tip and causes the variation of Pk from Pk . The wear of the cutting-tool tip
will result in a loss of effective radial and axial depth of cut that generates
dimensional variations (Fig. 2.13). In order to model the effect of VB on part
quality, the following assumptions are considered: i) flank wear is homogeneous;
ii) there is no other factors on the cutting-tool edge such as the generation of a
built-up edge or similar; and iii) the cutting-tool edge is a sharp edge. Under these
assumptions, the variation of the machined surface in its normal direction (and
thus, in Z direction of CS Lk ) from its nominal value is formulated as follows

tan (α )
δz = ⋅ VB , (2.43)
(1 − tan (γ ) ⋅ tan (α ))
78 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

where α is the clearance angle and γ is the rake angle of the cutting inserts.
According to Eq. (2.43), dimensional variations are proportional to the flank wear
magnitude and thus, the dimensional quality variation can be described by a
proportional coefficient that relates the influence of tool flank wear with the
dimensional variation of a manufacturing feature for a specific cutting operation
and cutting-tool geometry.
Assuming that tool flank wear remains constant during the same cutting
operation of one workpiece, the cutting-tool tip presents a constant variation
which is modeled as the DMV of Pk w.r.t. Ck by the expression
C m
x P k = B k 4 ⋅ VBk , (2.44)
k ij

m
where B k 4 = [0 0 CfVk 0 0 0]T , VBk refers to the flank wear of the ith
Bij ij

cutting edge of the jth cutting-tool at the kth machining station and CfVk is the
Bij

proportional coefficient.

Rake angle
(γ )
Rake face

Chip
Cutting speed
Clearance
Workpiece Crater Worn tool angle
wear
(α )
Built-up edge
Flank wear Flank face
adhesion
(VB)

Nominal surface Actual surface

Non-worn tool Dimensional
deviation
Fig. 2.13. Example of cutting-tool wear-induced deviation in machining

2.2.4 Derivation of the SoV Model

Considering the derivation of the DMVs presented above, the resulting DMV that
defines the dimensional variation of part features after station k can be obtained.
SoV Based Quality Assurance for MMPs – Modeling and Planning 79

The resulting DMV depends on two components: the feature variation from
previous stations, and the variation added in the station k itself, which can be due
to variations in datum surfaces, fixture locators and machining operations.
The first component is named the relocation term. If there is no variation added
R
at station k, the resulting DMV after station k, denoted as x k k , is the same as the
R
resulting DMV from station k-1 but expressed w.r.t. Rk , denoted as x k k −1 . For a
single feature S i , the relationship between feature variation i w.r.t. Rk −1 and the
same variation but w.r.t. Rk is defined by applying Corollary 1 as
R R R R
x S k = D S k ⋅ x Rk −1 + x S k −1 . (2.45)
i i k i

If one considers all features of the workpiece, the relationship is defined as

R R R R R R
x k k = [(x S k −1 )T ,…, (x S k −1 )T ]T + [DS k ,…, DS k ]T ⋅ x Rk −1 , (2.46)
1 M 1 M k

and rewritten Eq. (2.46),

⎛I Rk
… 06× 6 ⎞⎟
⎜ 6× 6 … 06× 6 … D S1
⎜ … … … … ⎟
R ⎜ R ⎟ R R
xk k = ⎜ 0 6× 6 … I 6× 6 … D S k … 06× 6 ⎟ ⋅ [(x S k −1 )T , … , (x S k −1 )T ]T ,
⎜ i ⎟ 1 M (2.47)
⎜ … … … … ⎟
⎜⎜ 0 Rk
… I 6× 6 ⎟⎟
6× 6 … 0 6× 6 … D S
⎝ M ⎠
R
= A 3k ⋅ x k k −1 .

On the other hand, the second term for deriving the resulting DMV for dimen-
sional variation of part features is related to the variation added at station k due to
the machining operation itself and thus, they only affect on the features machined
at station k. As it was described above, the variation of a machined feature is de-
R Mk Mk
fined as x L k , which depends on the DMVs x P and x R , as it is shown in Fig.
k k k
2.8. Applying Corollary 1, the total variation added to the workpiece at the station
k due to datum-, fixture- and machining-induced variations is defined as
R R Mk Mk
x L k = D Lk ⋅ x R + xP . (2.48)
k k k k

Mk
The DMV x R defined in Eq. (2.9) can be rewritten as
k

Mk F f R R
xR = TR k ⋅ B k1 ⋅ [Δl1 , Δl2 ,…, Δl6 ]T − A1k ⋅ x Bk − A k2 ⋅ x Ck ,
k k k k (2.49)
f2 R
= Bk ⋅ u kf − A 4k ⋅ x k k ,
80 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

Rk R R
where A 4k = [01× 6 , … , A1k ,01×6 , … , A 2k ,01× 6 , …] ; x = [01×6 ,…, ( x Bk )T ,01×6 ,…, ( x Ck )T ,
k k k
R
T
,01×6 , …] , and = [Δl1, Δl2 ,…, Δl6 ] . The DMV
u kf T
defines the variations of xk k
the workpiece w.r.t. the reference CS Rk at the station k . Since features may be
deviated after machining at station k − 1 , the input of feature variations at station
k is the resulting feature variations after station k − 1 but expressed in Rk CS. As
R R
the relationship between x k k −1 and the current x k k is defined in Eq. (2.47) by a
relocation matrix, Eq. (2.49) can be rewritten adding the feature variations from
station k − 1 as
Mk f R
xR = B k 2 ⋅ u kf − A k4 ⋅ A3k ⋅ x k k −1 . (2.50)
k

Mk
On the other hand, the DMV x P defined in Eq. (2.10) can be rewritten to
k
include the machining sources of variation as
Mk C S Az m C S m C m m
xP = [TP k ⋅ TC k ⋅ TS k ⋅ B k 1 TP k ⋅ TC k ⋅ B k 2 TP k ⋅ B k 3 Bk 4 ] ⋅
k k k k k k k

⋅ [ΔT1k … ΔTmk Δx k Δy k Δz k ΔTsk ΔFxk ΔFyk VBk ]T

ij

= Bm
k
5
⋅ um
k.

(2.51)
R
Substituting Eqs. (2.50) and (2.51) into Eq. (2.48), x L k is rewritten as
k

R R f R
x L k = D L k ⋅ [B k 2 ⋅ u kf − A 4k ⋅ A 3k ⋅ x k k −1 ] + B m
k ⋅ uk
5 m
k k (2.52)
R f
= A 5k ⋅ x k k −1 + B k 3 ⋅ u kf + B m
k ⋅ uk ,
5 m

R f R f R
where A 5k = [−D L k ·A k4 ·A 3k ] and B k 3 = [D L k ·B k 2 ] . As x L k refers only to the
k k k
feature machined, Eq. (2.52) can be rewritten in a more general form to include all
the features of the part as
R R f
x k k = A 6k ⋅ [ A 5k ⋅ x k k −1 + B k 3 ⋅ u kf + B m
k ⋅ u k ],
5 m
(2.53)

where A 6k = [06×6 ,…, I 6×6 ,…,06×6 ]T is a selector matrix that indicates the feature
machined at station k.
Finally, the resulting DMV after station k that defines the dimensional variation
of part features is obtained summing up both components: the feature variation
SoV Based Quality Assurance for MMPs – Modeling and Planning 81

from previous stations, and the variation added at the station k itself. Thus, the
resulting DMV after station k is defined as
R R R f
x k +k1 = A 3k ⋅ x k k −1 + A 6k ⋅ [ A 5k ⋅ x k k −1 + B k 3 ⋅ u kf + B m
k ⋅ u k ],
5 m
(2.54)
and reorganizing terms
R R f
x k +k 1 = [ A 3k + A 6k ⋅ A 5k ] ⋅ x k k −1 + A 6k ⋅ B k 3 ⋅ u kf + A 6k ⋅ B m
k ⋅ uk ,
5 m
(2.55)
which becomes the so-called SoV model in its generic version

x k +1 = A k ⋅ x k + B kf ⋅ u kf + B m
k ⋅ uk .
m
(2.56)

Besides Eq. (2.56), the SoV model is also composed of the observation
equation, which represents the inspection of KPCs. This equation is formulated as
follows

y k = [C1k ·A 3k ]·x k + v k ,
(2.57)
= Ck ⋅ x k + v k ,

where matrix Ck defines the features that are inspected after the station k which
define the KPCs and v k is a vector that represents the measurement errors during
inspection. The inspection station can be viewed as a special machining station
where only a relocation is conducted since there is no machining operation.
Therefore, C1k is the selector matrix similar to A 6k to indicate the features
inspected, and A 3k is the relocating matrix that relates the primary datum surface at
station k with the primary datum surface at the inspection station, placed after
station k.

2.3 Process Planning

An important application of the extended SoV model can be found in the field of
process planning. The integration of the KCCs and the KPCs let the engineers
evaluate the manufacturing capability of a process plan and detect the critical
KCCs that could be modified in order to reduce manufacturing variability. To
describe this application, the SoV model defined by Eqs. (2.56) and (2.57) can be
rewritten in an input-output form at the Nth station as
N N
yN = ∑
i =1
C N ·Φ (N•),i ·B if ·u if + ∑C
i =1
( •) m m
N ·Φ N , i ·B i ·u i + C N ·Φ (N•),0 ·x 0
(2.58)
N
+ ∑
i =1
C N ·Φ (N•),i ·w i + vN ,
82 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

where
⎧A ⋅A ·…·A i , i < N
Φ (N•), i = ⎨ N −1 N − 2 (2.59)
⎩ I i=N

and I is the identity matrix. The vector x0 represents the original variations of
workpiece features in its raw state. These original variations are generated by
previous manufacturing processes (i.e. bulk forming processes). Without loss of
generality, it is assumed that the impact of initial workpiece variations on part
quality is negligible in comparison with fixture- and machining-induced variations
and other unmodeled errors. Thus, Eq. (2.58) can be rewritten as

YN = Γ Nf ·U Nf + Γ mN ·U mN + Γ wN ·WN + v N , (2.60)

where

Γ Nf = [[C N ·Φ (N•),1·B1f ] [C N ·Φ (N•),2 ·B 2f ] [C N ·Φ (N•), N ·B Nf ]], (2.61)

Γ mN = [[C N ·Φ (N•),1·B1m ] [C N ·Φ (N•),2 ·B m

2] [C N ·Φ (N•), N ·B mN ]], (2.62)

Γ wN = [[C N ·Φ (N•),1 ] [C N ·Φ (N•),2 ] [C N ·Φ (N•), N ]], (2.63)

and U Nf = [(u1f )T , … , (u Nf )T ]T , U mN = [(u1m )T , … , (u mN )T ]T , WN = [(w1 )T , … ,

(w N )T ]T , YN = y N .
Eq. (2.60) can be used by the process planner to predict the part quality
variation of a process plan according to the fixtures and machine-tools capabilities
and to detect critical KCCs whose elimination could notably improve the final part
quality. In the next subsections, the evaluation of a process plan and how to
improve its robustness will be described in detail.

2.3.1 Process Plan Evaluation

Assuming that U Nf , U mN , WN and v N are independent of each other, Eq. (2.60)
can be analyzed in terms of covariances as

Σ Y = Γ Nf ·Σ f ·(Γ Nf )T + Γ mN ·Σ ·(Γ mN )T + Γ wN ·Σ ·(Γ wN )T + Σ v , (2.64)

N UN Um
N WN N

where Σ Y is the covariance matrix of the KPCs, Σ f is the covariance matrix

N UN

of fixture variations; Σ is the covariance matrix of machining variations;

Um
N
ΣW is the covariance matrix of unmodeled source of variations; and Σ v is the
N N
covariance matrix of measurement noise. Note that, since process planners are
more focused on the variation propagation during process planning, it is assumed
that there is no mean-shift in any source of variation.
SoV Based Quality Assurance for MMPs – Modeling and Planning 83

By applying Eq. (2.64), the process planner can estimate the expected part
quality variation given a manufacturing process plan. In practice, the evaluation
and selection among different process plans is conducted by analyzing a capability
ratio that indicates the capability of the manufacturing process to manufacture the
parts within specifications. Since products possess multiple, rather than a single,
KPCs, process capability analysis is based on a multivariate capability index. A
variety of multivariate capability indices, such as those presented in [33-37], have
been proposed for assessing capability.
In this chapter, the multivariate process capability ratio proposed by Chen [34]
is adopted for evaluating the manufacturing capability of a MMP. This capability
ratio was successfully implemented by previous authors [17, Chapter 13] in
similar MMPs with the use of the SoV methodology. Focusing only on the
variations in the KPC measurements, YN , the multivariate process capability
index for the process plan χ can be expressed as

1
MC pχ = , (2.65)
r0

where r0 is a value that has the probability,

c
Pr (max(| Yi , N | /ri , N , i = 1,…, M ) ≤ r0 ), equal to 1 − α , and α is the allowable
expected proportion of non-conforming products from the process. Yic, N is the c th
component vector of the ith KPC measured at station N , and ri , N is the tolerance
value for this component. The r0 value can be obtained by Monte Carlo
simulations of Eq. (2.60), where the fixture-induced variations ( U Nf ) are assumed
normally distributed with the standard deviation defined from the fixture
tolerances and the machining-induced variations ( U mN ) are distributed according
to the nature of the machining error. The unmodeled errors ( WN ) and the
measurement errors ( v N ) are set according to process planner's knowledge about
the process and measurement devices, respectively.
Through this formulation, an index value of MC pχ = 1.0 corresponds to an
expected proportion of conforming products of exactly 1 − α . A larger MC pχ
indicates a lower expected proportion of non-conforming products or a more
capable process plan. Therefore, these interpretations enable the MC pχ ratio to
have a similar interpretation as the univariate ratio C p [17, Chapter 13].

2.3.2 Process Plan Improvement

In order to improve a MMP and reduce the final part quality variability, the
critical KCCs that influence the most on part quality should be identified. For this
purpose, a series of sensitivity indices can be evaluated from Eq. (2.60), detecting
84 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

the fixture component or the machining source of variation that contribute

significant amount of variations to KPCs. Then, these critical KCCs' effects can be
minimized or eliminated to improve the robustness of the process.
The process plan improvement can be accomplished through evaluating
different sensitivity indices at different levels following the flowchart shown in
Fig. 2.14. According to this flowchart, at phase I, the process sensitivity index
considering fixture-induced variations, S Yf , process , and the process sensitivity
index considering machining-induced variations, S mY , process , are evaluated to
determine the main source of variation (fixturing or machining) in the process
plan. At phase II, additional sensitivity indices of individual KPCs are evaluated
and compared with their corresponding process sensitivity index scaled by a factor
Cb to identify whether there is a critical KPC whose variability is notably higher
than that from other KPCs. Finally, at phase III, it is further investigated which
KCCs are the causes of the high variation of KPCs, detecting critical locators or
critical machining sources of variation. Such information can lead the process
planner to redefine the process plan and increase its robustness.

S Yf , process S mY , process

S Yf , process
S mY , process

max (S fyi, process) max ( Smyi, process)

i =1,..,M i =1,..,M

Cb S Yf , process Cb S mY , process

S fyi,P = max (S fyi,k ) S Yf , P = max ( S Yf ,k ) S mi, P = max ( S mi, k ) S mY , P = max ( S mY , k )

y y
k =1,..,N k =1,.., N k =1,.., N k =1,.., N

max ( S uy i ) max (SuYf ,P ) max ( S uy i ) max ( S uYm )

j =1,.., 6 f j ,P
j =1,..,6 j j =1,..., J k m j ,P j =1,..., J k j ,P

Fig. 2.14. Flowchart for process plan improvement

The different sensitivity indices applied in the proposed flowchart are

extensions of widely accepted indices [38]. Basically, a sensitivity index of a
SoV Based Quality Assurance for MMPs – Modeling and Planning 85

generic variable H w.r.t. a generic parameter ϕ can be described by the

differential equation
δH Δϕ
SϕH = , (2.66)
δϕ ΔH
where ΔH and Δϕ denote the admissible variation of variable H and the
expected variation of parameter ϕ , respectively. For a MMP design problem, the
admissible variation of a variable H is related to a KPC and it can be defined by
its tolerance from the design drawing, denoted by H tol . The expected variation of
a generic parameter ϕ is related to fixture-induced or machining-induced
variations, and it can be defined by its expected range of variation denoted as ϕ tol
and defined as 6σ ϕ (99.73% of confidence level), where σ ϕ is the standard
deviation of this source of variation. Thus, the generic sensitivity index is defined
as

δH ϕ tol
SϕH = . (2.67)
δϕ H tol
We consider a MMP composed of N stations, each of which is equipped with a
3-2-1 fixture device. For this MMP, the locator variations at station k are defined
as u kf = [u f , … , u f ]T , and the machining variations are defined as
1 6

um
k = [u m , … , u m
T
] , where J k is the number of machining sources of
1 Jk

variation. Consider that this MMP will be used to manufacture a product

composed of M KPCs, and each KPC is denoted by the index i ( i = 1,… , M ).
According to Eq. (2.67), a series of sensitivity indices can be defined as shown in
the following subsections. In addition to this definition, a summary of these
indices and their respective interpretations is shown in Table 2.1.

2.3.2.1 Sensitivity Indices Related to Fixtures

Locator sensitivity index: This type of indices evaluates the impact of the
variability of each fixture locator at station k . Two indices can be distinguished:
that w.r.t. an individual KPC, as defined in Eq. (2.68), and that w.r.t. a product, as
defined in Eq. (2.69).
⎛ tol ⎞ ⎛ tol ⎞
y ⎜ δyi u f j , k ⎟ ⎜ f u f j ,k ⎟
Su i = abs⎜ ⎟ = abs⎜ Γi , j tol ⎟, (2.68)
⎜ δu f j , k yi
tol
f j ,k ⎟ ⎜ yi ⎟
⎝ ⎠ ⎝ ⎠
M

∑S
1 yi
S uY = u f ,k , (2.69)
f j ,k M i =1 j
86 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

where abs (A) denotes the absolute value of A ; Γi,f j denotes the element at the
ith row and the jth column of the matrix Γ Nf ; yitol is the product dimensional
tolerance of the KPC yi ; and u tol
f , k is the tolerance of the jth fixture locator
j

component at station k . This sensitivity index is defined according to Eq. (2.67)

by substituting H , ϕ , H tol and ϕ tol with yi , u f , k , yitol , and u tol
f , k , respectively.
j j
y
Su i refers to the impact of the variability of the j th locator at station k on the
f j ,k

i th KPC, whereas S uY refers to the impact on the average of all KPCs.

f j ,k

Fixture sensitivity index: This type of indices evaluates the average impact of
locator variations on part quality at station k . The fixture sensitivity index w.r.t. a
y
KPC ( S f i,k ) and that w.r.t. the product ( S Yf ,k ) are defined respectively as

∑S
y 1 yi
S f i,k = u f ,k , (2.70)
6 j =1 j

∑S
1 yi
S Yf ,k = f ,k . (2.71)
M i =1

2.3.2.2 Sensitivity Indices Related to Machining Operations

Machining source of variation sensitivity index: This type of indices evaluate the
impact of the variability of a specific machining source of variation when
conducting a specific machining operation at station k . Two indices are defined
y
w.r.t. a KPC ( S u i ) and w.r.t. a product ( S uY ), respectively, as
m j ,k m j ,k

y δyi u mtol , k u mtol ,k

j j
Su i = abs( )= abs (Γim, j ), (2.72)
m j ,k δu m yitol yitol
j ,k

∑S
1 yi
S uY = um , k , (2.73)
m j ,k M i =1 j

where Γim, j denote the element in the i th row and j th column of the matrix Γ mN ;
and u mtol ,k defines the variability of the j th machining source of variation at the
j

k th station and j = 1,… , J k .

SoV Based Quality Assurance for MMPs – Modeling and Planning 87

Operation sensitivity index: This type of indices evaluates the average impact of
machining sources of variations at station k . Two indices are defined w.r.t. a KPC
y
( S mi,k ) and w.r.t. a product ( S mY ,k ), respectively, as

Jk

∑S
y 1 yi
S mi,k = um ,k , (2.74)
Jk j =1 j

∑S
1 yi
S mY ,k = m,k . (2.75)
M i =1

2.3.2.3 Sensitivity Index Related to Stations

Station sensitivity index: This type of indices evaluates the average impact of the
variability of both fixture and machining-induced variations at station k . Two
y
indices are defined w.r.t. a KPC ( S u i ) and w.r.t. a product ( S uY ), respectively, as
k k

y y
y S f i,k + S mi,k
Su i = , (2.76)
k 2

S Yf ,k + S mY ,k
S uY = . (2.77)
k 2

2.3.2.4 Sensitivity Indices Related to Manufacturing Process

Process sensitivity index considering fixture-induced variations: This type of

indices evaluates the average impact of fixture-induced variations along a MMP.
y
Two indices are defined w.r.t. a KPC ( S f i, process ) and w.r.t. a product ( S Yf , process ),
respectively, as
N

∑S
y 1 yi
S f i, process = f ,k , (2.78)
N k =1

∑S
1 yi
S Yf , process = f , process . (2.79)
M i =1

Process sensitivity index considering machining-induced variations: This type of

indices evaluates the impact of machining-induced variations along a MMP. Two
88 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

y
indices are defined w.r.t. a KPC ( S mi, process ) and w.r.t. a product ( S mY , process ),
respectively, as:
N

∑S
y 1 yi
S mi, process = m, k , (2.80)
N k =1

∑S
1 yi
S mY , process = m , process . (2.81)
M i =1

Note that in the flowchart for process planning improvement shown in Fig. 2.14,
y
the indices S u i and S uY are not evaluated since the purpose is to analyze
k k
separately the variation propagation in the MMP due to fixture- and machining-
induced variations in order to propose process plan adjustments.

Table 2.1. Summary of the sensitivity indices and physical meaning

Sensitivity index KPC Product Rationale for definition

Locator y
S uY Identifies the most influent locators on
Su i f j ,k KPC/product variability
index f j ,k

Fixture y
S uY Identifies the most influent fixtures on
Su i f ,k KPC/product variability
index f ,k

Machining source y
S uY Identifies the most influent machining
of variation index Su i m j ,k source of variation on KPC/product va-
m j ,k
riability
Machining opera- y
S mY ,k Identifies the most influent machining
tion index S mi,k operation on KPC/product variability
Station y
S uY Identifies the most influent station on
Su i k KPC/product variability
index k
Process index due y
S Yf , process Evaluates the influence on KPC/product
to fixture-induced S f i, process variability due to all fixture-induced
variations variations
Process index due y
S mY , process Evaluates the influence on KPC/product
to machining- S mi, process due to all machining-induced variations
induced variations

2.4 Case Study

In this section, the SoV model is derived in order to analyze the MMP shown in
Fig. 2.2 and its capability to manufacture the part geometry shown in Fig. 2.1. For
SoV Based Quality Assurance for MMPs – Modeling and Planning 89

this case study, the final part geometry is shown in Table 2.2. The positions of the
locators that compose the fixture at different stations are shown in Table 2.3, and
the characteristics of the MMP in terms of locator tolerances, expected thermal
variations of the machine-tool spindle at each station, and admissible cutting-tool
wear at each cutting-tool are shown in Table 2.4.
For the sake of simplicity and without loss of generality, the extended SoV
model applied in this case study only deals with cutting-tool wear-induced
variations, thermal-induced variations from the machine-tool spindle, fixture-
induced variations and datum-induced variations. The empirical coefficients that
relate the spindle thermal variations with the dimensional spindle expansion and
the cutting-tool wear effects with the cutting-tool tip loss of cut are obtained from
[18]. According to this research work and the results of a set of experiments, the
cutting-tool wear coefficients were adjusted to CfVk = 0.125 and CfVk = 0.135
B B
for frontal and peripheral machining operations, respectively, and the thermal
coefficients were adjusted to Cf xk = Cf yk = Cf αk = Cf βk = Cf γk ≈ 0 and
Cf zk = −0.0052 mm/ C .

Table 2.2. Product design information. Nominal position ( t S D ) and orientation ( ω S D ) of

i i
each feature w.r.t. DCS.

Feature
ω S D (rad) t S D (mm)
i i

S0 [0, π ,0] [150,125,0]

S1 [0, π ,0] [150,125,20]

S2 [0,0,0] [150,75,200]
S3 [0,0,0] [150,225,180]

S4 [π/2,−π/2,−π/2] [150,0,100]

S5 [0,−π/2,0] [0,75,100]
S6 [π/2, π/2, −π/2] [150,250,75]

S7 [0,π/2, π/2] [300,125,100]

S8 [π/2, π/2, −π/2] [150,200,190]

90 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

Table 2.3. Nominal position ( t F D ) and orientation ( ω F D ) of FCS w.r.t. DCS at each
k k
station and fixture layout

ω FD t FD
St. k k Locator position w.r.t. FCS
(rad) (mm) (mm)
1 [−π/2, π ,0] [0,0,0] L1x = 125, L1 y = 50, L2 x = 50, L2 y = 250, L3x = 200, L3 y = 250
p1 y = 50, p1z = −100, p 2 y = 250, p2 z = −100, p3 x = 125, p3 z = −100

2 [π/2,−π/2, π ] [0,0,200] L1x = 100, L1y = 50, L2 x = 50, L2 y = 150, L3 x = 150, L3 y = 150

p1y = 50, p1z = −125, p2 y = 250, p2 z = −125, p3x = 100, p3z = −125

3 [−π/2,0,0] [0,250,200] L1x = 125, L1 y = 50, L2 x = 50, L2 y = 250, L3x = 200, L3 y = 250
p1 y = 50, p1z = −100, p 2 y = 250, p2 z = −100, p3 x = 125, p3 z = −100

4 [ −π/2, π ,0] [0,0,20] L1x = 125, L1 y = 50, L2 x = 50, L2 y = 250, L3x = 200, L3 y = 250
p1 y = 50, p1z = −100, p 2 y = 250, p2 z = −100, p3 x = 125, p3 z = −100

5 [−π/2, π ,0] [0,0,0] Inspection station

Table 2.4. Process description of the case study

Datum Features Δl tol

j VB ΔTs
Station features machined
(mm) (mm) (ºC)
1 S0 − S 4 − S5 S2 0.035 [0,0.5] [ −5,+5]
2 S 4 − S 2 − S5 S6 0.035 [0,0.5] [ −5,+5]
3 S 2 − S 6 − S5 S1 0.035 [0,0.5] [ −5,+5]
4 S1 − S 4 − S 5 S 3 , S8 0.035 [0,0.5] [ −5,+5]
5 S0 − S 4 − S5 -CMM- - - -

Δl tol
j for j = 4,5,6 refer to tolerances at locators p1 , p 2 and p3 , respectively

2.4.1 Process Plan Evaluation

In order to estimate the capability of the MMP to manufacture the part geometry
defined in Fig. 2.1, 1,000 Monte Carlo simulations were conducted. Each
component of the fixture variations U kf was assumed independent and normally
distributed with a standard deviation of Δl tol
j /6 ; each component related to the

thermal variation of the machine-tool spindle in U m

k was assumed independent of
SoV Based Quality Assurance for MMPs – Modeling and Planning 91

other components and uniformly distributed within the expected range of thermal
variations; each component related to cutting-tool wear in U m k was assumed
independent of other components and uniformly distributed within the expected
range of the admissible cutting-tool wear; the measurement error and the term
related to non-linear errors were assumed to be negligible in comparison with
fixture- and machining-induced variations. Eq. (2.60) was evaluated for 1,000
simulations. The resulting distribution of the KPC variations was used to evaluate
the multivariate process capability ratio defined in Eq. (2.65).
After Monte Carlo simulations and considering the expected proportion of
conforming products as 1 − α = 0.99 , the capability index obtained was
MC p = 0.82 . Indeed, Monte Carlo simulations showed that 4.4% of the
simulated parts have nonconforming KPC1 ’s, whereas 0.5% of them have
nonconforming KPC 2 .

2.4.2 Process Plan Improvement

As shown above, the analyzed MMP is not able to produce 99% of conforming
parts ( MC p = 0.82 < 1 ). In order to investigate how to improve the process and
reduce part quality variability, a group of sensitivity indices are evaluated
following the flowchart shown in Fig. 2.14. According to the sensitivity indices
evaluated, the manufacturing variability due to fixture-induced variations is higher
than that due to machining-induced variations ( S mY , process = 0.044 and
S Yf , process = 0.094 ). Since the fixture-induced variations have a higher impact on
product quality than the machining-induced variations, at the second phase of the
flowchart it is analyzed the sources of fixture-induced variations. At this phase,
the sensitivity indices show that there is no special KPC with an important
variability w.r.t. other KPCs and thus, the sensitivity indices should be analyzed
w.r.t. product. In fact, the sensitivity indices for each KPC are similar
1 KPC 2
( S KPC
f , process = 0.101 and S f , process = 0.087 ), which means that both KPCs present
a similar manufacturing variability. By analyzing in more detail the sensitivity
indices related to fixture-induced variations, it can be observed that the fourth
station is the most critical station in the MMP. Indeed, the fourth station is
responsible for 35% of the part quality variability due to fixture-induced
variations. Following the third phase of the flowchart, a further investigation is
conducted at the fourth station. At this station, the locator sensitivity indices show
the particular impact of each fixture locator within the fixture-induced variations.
These indices are S uY ,4 = 0.14 , S uY ,4 = 0.25 , S uY ,4 = 0.32 , S uY ,4 = 0.03 ,
f1 f2 f3 f4

S uY ,4 = 0.03 , S uY ,4 = 0 . Thus, 41% of part fixture-induced quality variability

f5 f6

at this station is due to the low precision of locator number 3 .

92 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

In order to improve the MMP, it is reasonable to reduce the impact of fixture-

induced variations at station 4 . Additionally, the sensitivity analysis shows that
station 2 also contributes to the 23.4% of part variability due to fixture-induced
variations. Therefore, a process planning modification is proposed to increase the
locator precision at both stations. For this example, the locator tolerances at station
2 and 4 are reduced to 0.015 mm . By this modification, the multivariate
capability ratio is increased up to MC p = 0.91 . For this MMP with a first
improvement, the percentage of part quality variability due to fixture-induced
variations is 61% whereas due to machining-induced variations is 39%
( S mY , process = 0.047 and S Yf , process = 0.075 ).
For illustrative purposes, we propose a second improvement of the MMP but
now being focused on reducing the effect of machining-induced variations.
Analyzing the sensitivity indices it is observed that the variability of KPC1
depends more on machining-induced variations than that of the KPC 2
KPC 1
( S m, process > Cb ⋅ S mY , process , with Cb =1.3 ). Furthermore, it is observed that
33.3% of the variability of KPC1 due to machining-induced variations is
generated at station 1 , 3 and 4 in the same proportion. In any of these stations,
55% of the variability is due to the cutting-tool wear-induced variations whereas
45% of the variability is due to the thermal expansion of the machine-tool
spindle. According to this analysis, it seems reasonable to reduce the admissible
level of cutting-tool wear during machining at these three stations in order to
reduce the impact of machining-induced variations. For this example, it is
proposed to reduce the admissible flank wear at stations 1 , 3 and 4 to a value of
VB = 0.2 mm . By this modification, the multivariate capability ratio is increased
up to MC p = 1.02 and thus, the MMP is more capable to produce more than the
99% of parts within specifications.

SuYf 5 ,4 S uY
S fu ,P
S fu ,P
,4
f
f
f S fu ,P
S fu ,P
6 5
f
4
4
f f
S1 1

Sm
process
SuYf 1,4
S Yf ,1
f
S4

SmY , process S Yf ,4
S uY f ,4
S Yf , 2 S uY f 3
S Yf , process 2
,4 S fu ,P
3
f

Sf
process S fu ,P
f
f 2
S2

S Yf ,3 =0
f

S uY f
S3

6
,4

Fig. 2.15. Sensitivity analysis results of fixture-induced variations

SoV Based Quality Assurance for MMPs – Modeling and Planning 93

Sm,S3 Sm,S3
1 4

SmKPC S mKPC
Sm
process 1 1 KPC 1
S mY , process ,1 ,4 KPC 1 S m,S3

S um S um ,1
u
m
Ta
,P

,1 Ts
SmS3,1
S m,S3
u ,P VB
S Yf , process
m
VB

Sf
process
Sm,S3
2
SmKPC
,3
1

,2 = 0
S mKPC
Sm,S3
1 3

Fig. 2.16. Sensitivity analysis results of machining-induced variations

A summary of the results obtained after the sensitivity analysis is shown in

Fig. 2.15 and 2.16. The resulting probability density functions (p.d.f.) of the KPCs
for the initial MMP and the MMP with the proposed improvements are also shown
in Fig. 2.17.

12
16 MMP MMP
MMP with improv. I MMP with improv. I
14 10
MMP with improv. I & II MMP with improv. I & II
12
8
10
P.d.f.
P.d.f.

6
8

6 4
4
2 KPC Tolerance
1
2 KPC Tolerance
2

0 0
−0.1 −0.05 0 0.05 0.1 0.15 −0.1 −0.05 0 0.05 0.1 0.15
KPC deviations (mm) KPC deviations (mm)
2 1

(a) (b)
Fig. 2.17. P.d.f. obtained from Monte Carlo simulation results for a) KPC1 and b) KPC 2
variations according to the initial MMP and the MMP with improvement I and I and II

2.5 Conclusions
In this chapter, the derivation of the 3D stream of variation (SoV) model for
MMPs is presented in its extended version. With this model, engineers can
consider machining-induced variations such as those related to machine-tool
inaccuracy, cutting-tool deflections, thermal distortions or cutting-tool wear, as
well as fixture- and datum-induced variations in order to estimate dimensional and
geometrical part variations and their propagation in MMPs. The potential
94 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

applicability of the extended SoV model is extensive, especially in fields such as

manufacturing fault diagnosis, process planning, process oriented tolerancing and
dimensional quality control.
Besides the derivation of the extended SoV model, this chapter shows the
applicability of the model on process planning activities. Firstly, it was described
how to evaluate the robustness of a manufacturing process plan when fixture-
induced and machining-induced variations are present. Secondly, it was presented
a series of sensitivity indices and a flowchart to detect which are the sources of
variations that affect part quality and at which stations they can be shown up.
Through these indices, process plan improvements can be straight forward
conducted. The application of the SoV model on process planning was also
demonstrated through a 3D case study with a product design with two KPCs and a
MMP composed of 4 machining stations.

Acknowledgement. This work has been partially supported by Fundació Caixa-Castelló

Bancaixa (Research Promotion 2007, 2009 and 2011) and the Spanish National Project
number DPI2007-66871-C02-01 (PIA12007-83).

Appendix 2.1

A Homogeneous Transformation Matrix (HTM) in the 3 D space is a 4 × 4

matrix. It can be used to represent one CS w.r.t. another CS. Let us consider two
CS 1 and 2 . Given the position and orientation vector of CS 1 w.r.t. CS 2 , the
CS 1 can be expressed in 2 by a nominal HTM as [39]

H12 = T12 ⋅ R12 , (2.82)

where T12 and R12 are the translational and rotational matrix, respectively. For a
HTM, the superscript represents the CS we want the results to be represented in,
and the subscript represents the CS we are transferring from. In Eq. (2.82) the
translational matrix is defined as

⎛1 0 0 t12x ⎞⎟
⎜
2 ⎜0 1 0 t12y ⎟ ⎛⎜ I 3×3 t12 ⎞⎟
T1 = ⎜ ⎟= , (2.83)
⎜0 0 1 t12z ⎟ ⎜⎝ 01×3 1 ⎟⎠
⎜0 0 0 1 ⎟⎠
⎝

where t12 is the position vector of the origin of CS 1 expressed in 2 , defined as

t12 = [t12x , t12y , t12z ]T . The rotational matrix is defined according to the rotation
representation used. According to the Euler angles conversion, three Euler angles
( φ , θ , and ψ ) are used to define the rotational matrix. The orientation of CS 1
w.r.t. CS 2 can be obtained by three successive rotations as follows. First, CS 2
is rotated about the Z-axis by the angle φ . Then, the resulting CS is rotated
SoV Based Quality Assurance for MMPs – Modeling and Planning 95

around the new Y-axis by angle θ , and finally, the resulting CS is rotated around
the new Z-axis by ψ . According to this representation, the rotational matrix is
formulated as [40, Chapter 2]

⎛ R2 03×1 ⎞⎟
R12 = ⎜ 1 , (2.83)
⎜0 1 ⎟⎠
⎝ 1×3
where

R12 = R z ,φ ⋅ R y ,θ ⋅ R z ,ψ , (2.84)

and
⎛ cosφ − sinφ 0⎞
⎜ ⎟
R z ,φ = ⎜ sinφ cosφ 0 ⎟, (2.85)
⎜ ⎟
⎝ 0 0 1⎠

⎛ cosθ 0 sinθ ⎞
⎜ ⎟
R y ,θ =⎜ 0 1 0 ⎟, (2.86)
⎜ ⎟
⎝ − sinθ 0 cosθ ⎠

⎛ cosψ − sinψ 0⎞
⎜ ⎟
R z ,ψ = ⎜ sinψ cosψ 0 ⎟. (2.87)
⎜ 0 1 ⎟⎠
⎝ 0
For this rotation representation, the rotational matrix is defined as
⎛ cφ cθ cψ − sφ sψ − cφ cθ sψ − sφ cψ cφ sθ ⎞
⎜ ⎟
R12 = ⎜ sφ cθ cψ + cφ sψ − sφ cθ sψ + cφ cψ sφ sθ ⎟, (2.88)
⎜ ⎟
⎝ − sθ cψ sθ sψ cθ ⎠
where c and s refers to cos and sin respectively. As a result, Eq. (2.82) can be
rewritten as
⎛ R2 t12 ⎞⎟
H12 = ⎜ 1 . (2.89)
⎜0 1 ⎟⎠
⎝ 1×3
Using the HTM, an ith point in the CS 1 , defined as p1i = [ p1ix , p1iy , p1iz ] , is
related to the same point expressed in the CS 2 , defined as p i2 , by the following
equation
~ 2 = H2 ⋅ p
p ~1 , (2.90)
i 1 i

~ is equal to [p,1]T .
where p
96 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

Appendix 2.2
A Differential Transformation Matrix (DTM) in the 3D space is a 4 × 4 matrix
that is used to represent the small position and orientation deviation of one CS
w.r.t. another CS. For illustrative purposes, we consider two CSs, 1 and 2. If both
CSs are deviated from nominal values, the HTM between the actual CS 1 and CS
2 can be defined as

H12 = H 1
⋅ δH12 , (2.91)
2

where H 1
is the HTM between the nominal CSs 1 and 2 , and δH12 is the HTM
2
that defines the small position and orientation deviations of CS 2 w.r.t. 1 due to
the deviation from their nominal values, and it is defined as
⎛ 1 − θ 21 z θ 21 y d 21 x ⎞⎟
⎜
⎜ θ1 1 − θ 21 x d 21 y ⎟
δH 2 = ⎜ 21z
1
⎟. (2.92)
⎜ − θ2 y θ 21 x 1 d 21 z ⎟
⎜ 0 1 ⎟⎠
⎝ 0 0

Note that the rotational matrix in δH12 is defined as Eq. (2.88) using the
approximation of cos(θ ) ≈ 1 and sin(θ ) ≈ θ and neglecting second-order small
values since only small position and orientation variations are considered. From
Eq. (2.92), the small position and orientation deviation of CS 2 w.r.t. 1 is defined
as d12 = [d 21 x , d 21 y , d 21 z ]T and θ12 = [θ 21 x ,θ 21 y ,θ 21 z ]T , respectively.
The small position and orientation deviations defined by δH12 can be expressed
as

δH12 = I 4× 4 + Δ12 , (2.93)

where Δ12 is named the DTM, and is defined as

⎛ θˆ 1 d12 ⎞
Δ12 = ⎜ 2 ⎟, (2.94)
⎜0 ⎟
⎝ 1×3 0 ⎠

where θ̂12 is the skew matrix of θ12 and it is defined as

⎛ 0 − θ 21 z θ 21 y ⎞⎟
⎜
θˆ 2 = ⎜ θ 21 z
1
0 − θ 21x ⎟. (2.95)
⎜ 1 ⎟
⎜ − θ2 y θ 21 x 0 ⎟
⎝ ⎠
SoV Based Quality Assurance for MMPs – Modeling and Planning 97

It is important to remark that any DTM can also be expressed as a differential

motion vector (DMV) in a vector form. Indeed, given the DTM of CS 2 w.r.t. 1,
denoted as Δ12 , a DMV is then defined as

⎛ d1 ⎞
x12 = ⎜ 12 ⎟. (2.96)
⎜θ ⎟
⎝ 2⎠

References
[1] Zhou, S., Huang, Q., Shi, J.: State Space Modeling of Dimensional Variation Propa-
gation in Multistage Machining Process Using Differential Motion Vectors. IEEE
Transactions on Robotics and Automation 19, 296–309 (2003)
[2] Ogata, K.: Modern Control Engineering. Prentice Hall, New Jersey (2001)
[3] Ding, Y., Shi, J., Ceglarek, D.: Diagnosability analysis of multi-station manufacturing
processes. Journal of Dynamic Systems Measurement and Control-Transactions of
the Asme 124, 1–13 (2002)
[4] Zhou, S., Chen, Y., Ding, Y., Shi, J.: Diagnosability Study of Multistage Manufactur-
ing Processes Based on Linear Mixed-Effects Models. Technometrics 45, 312–325
(2003)
[5] Zhou, S., Chen, Y., Shi, J.: Statistical estimation and testing for variation root-cause
identification of multistage manufacturing processes. IEEE Transactions on Automa-
tion Science and Engineering 1, 73–83 (2004)
[6] Li, Z.G., Zhou, S., Ding, Y.: Pattern matching for variation-source identification in
manufacturing processes in the presence of unstructured noise. IIE Transactions 39,
251–263 (2007)
[7] Djurdjanovic, D., Ni, J.: Dimensional errors of fixtures, locating and measurement
datum features in the stream of variation modeling in machining. Journal of Manufac-
turing Science and Engineering-Transactions of the Asme 125, 716–730 (2003)
[8] Djurdjanovic, D., Zhu, J.: Stream of Variation Based Error Compensation Strategy in
Multi-Stage Manufacturing Processes. In: ASME Conference Proceedings,
vol. 42231, pp. 1223–1230 (2005)
[9] Djurdjanovic, D., Ni, J.: Online stochastic control of dimensional quality in multista-
tion manufacturing systems. Proceedings of the Institution of Mechanical Engineers
Part B-Journal of Engineering Manufacture 221, 865–880 (2007)
[10] Jiao, Y., Djurdjanovic, D.: Joint allocation of measurement points and controllable
tooling machines in multistage manufacturing processes. IIE Transactions 42, 703–
720 (2010)
[11] Zhong, J., Liu, J., Shi, J.: Predictive Control Considering Model Uncertainty for Var-
iation Reduction in Multistage Assembly Processes. IEEE Transactions on Automa-
tion Science and Engineering 7, 724–735 (2010)
[12] Liu, Q., Ding, Y., Chen, Y.: Optimal coordinate sensor placements for estimating
mean and variance components of variation sources. IIE Transactions 37, 877–889
(2005)
[13] Liu, J., Shi, J., Hu, J.S.: Quality-assured setup planning based on the stream-of-
variation model for multi-stage machining processes. IIE Transactions 41, 323–334
(2009)
98 J.V. Abellan-Nebot, J. Liu, and F. Romero Subiron

[14] Chen, Y., Ding, Y., Jin, J., Ceglarek, D.: Integration of Process-Oriented Tolerancing
and Maintenance Planning in Design of Multistation Manufacturing Processes. IEEE
Transactions on Automation Science and Engineering 3, 440–453 (2006)
[15] Ding, Y., Jin, J., Ceglarek, D., Shi, J.: Process-oriented tolerancing for multi-station
assembly systems. IIE Transactions 37, 493–508 (2005)
[16] Abellan-Nebot, J.V., Liu, J., Romero, F.: Design of multi-station manufacturing
processes by integrating the stream-of-variation model and shop-floor data. Journal of
Manufacturing Systems 30, 70–82 (2011)
[17] Shi, J.: Stream of Variation Modeling and Analysis for Multistage. CRC Press Taylor
and Francis Group (2007)
[18] Abellan-Nebot, J.V., Liu, J., Romero, F.: State Space Modeling of Variation Propaga-
tion in Multi-station Machining Processes Considering Machining-Induced Varia-
tions. Journal of Manufacturing Science and Engineering-Transactions of the Asme
(in press, 2012)
[19] Joshi, P.H.: Jigs and Fixtures Design Manual. McGraw-Hill, New York (2003)
[20] Wang, M.Y.: Characterizations of positioning accuracy of deterministic localization
of fixtures. In: IEEE International Conference on Robotics and Automation, pp.
2894–2899 (2002)
[21] Choi, J.P., Lee, S.J., Kwon, H.D.: Roundness Error Prediction with a Volumetric Er-
ror Model Including Spindle Error Motions of a Machine Tool. International Journal
of Advanced Manufacturing Technology 21, 923–928 (2003)
[22] Lei, W.T., Hsu, Y.Y.: Accuracy test of five-axis CNC machine tool with 3D probe-
ball. Part I: design and modeling, International Journal of Machine Tools and Manu-
facture 42, 1153–1162 (2002)
[23] Chen, J.S., Yuan, J., Ni, J.: Thermal Error Modelling for Real-Time Error Compensa-
tion. International Journal of Advanced Manufacturing Technology 12, 266–275
(1996)
[24] Chen, G., Yuan, J., Ni, J.: A displacement measurement approach for machine geo-
metric error assessment. International Journal of Machine Tools and Manufacture 41,
149–161 (2001)
[25] Yang, S.H., Kim, K.H., Park, Y.K., Lee, S.G.: Error analysis and compensation for
the volumetric errors of a vertical machining centre using a hemispherical helix ball
bar test. International Journal of Advanced Manufacturing Technology 23, 495–500
(2004)
[26] Tseng, P.C.: A real-time thermal inaccuracy compensation method on a machining
centre. International Journal of Advanced Manufacturing Technology 13, 182–190
(1997)
[27] Haitao, Z., Jianguo, Y., Jinhua, S.: Simulation of thermal behavior of a CNC machine
tool spindle. International Journal of Machine Tools and Manufacture 47, 1003–1010
(2007)
[28] Kim, G.M., Kim, B.H., Chu, C.N.: Estimation of cutter deflection and form error in
ball-end milling processes. International Journal of Machine Tools and Manufac-
ture 43, 917–924 (2003)
[29] López de Lacalle, L.N., Lamikiz, A., Sanchez, J.A., Salgado, M.A.: Effects of tool
deflection in the high-speed milling of inclined surfaces. International Journal of Ad-
vanced Manufacturing Technology 24, 621–631 (2004)
[30] Dow, T.A., Miller, E.L., Garrard, K.: Tool force and deflection compensation for
small milling tools. Precision Engineering 28, 31–45 (2004)
SoV Based Quality Assurance for MMPs – Modeling and Planning 99

[31] Gere, J.M., Goodno, B.J.: Mechanics of Materials, 7th edn. Nelson Engineering
(2008)
[32] ISO 8688-1:1989, Tool-life testing in milling – Part 1: face milling
[33] Taam, W., Subbaiah, P., Liddy, J.W.: A note on multivariate capability indices. Jour-
nal of Applied Statistics 20, 339–351 (1993)
[34] Chen, H.F.: A Multivariate Process Capability Index Over A Rectangular Solid To-
lerance Zone. Statistica Sinica 4, 749–758 (1994)
[35] Wang, F.K., Chen, J.C.: Capability indices using principle components analysis.
Quality Engineering 11, 21–27 (1998)
[36] Wang, F.K., Du, T.C.T.: Using principle component analysis in process performance
for multivariate data. Omega 28, 185–194 (2000)
[37] Pearn, W.L., Kotz, S.: Encyclopedia And Handbook of Process Capability Indices: A
Comprehensive Exposition of Quality Control Measures. World Scientific Publishing
Company (2006)
[38] Zhang, M., Djurdjanovic, D., Ni, J.: Diagnosibility and sensitivity analysis for multi-
station machining processes. International Journal of Machine Tools and Manufac-
ture 47, 646–657 (2007)
[39] Okafor, A.C., Ertekin, Y.M.: Derivation of machine tool error models and error com-
pensation procedure for three axes vertical machining center using rigid body kine-
matics. International Journal of Machine Tools and Manufacture 40, 1199–1213
(2000)
[40] Spong, M.W., Hutchinson, S., Vidyasagar, M.: Robot modeling and control. John Wi-
ley & Sons, Inc., New York (2006)
3

Finite Element Modeling of Chip Formation

in Orthogonal Machining

Amrita Priyadarshini, Surjya K. Pal, and Arun K. Samantaray

Department of Mechanical Engineering, Indian Institute of Technology,

Kharagpur, Kharagpur- 721 302, West Bengal, India
[email protected],
[email protected],
[email protected]

Finite element method has gained immense popularity in the area of metal cutting for
providing detailed insight in to the chip formation process. This chapter presents an
overview of the application of finite element method in the study of metal cutting
process. The basics of both metal cutting and finite element methods, being the fore-
most in understanding the applicability of finite element method in metal cutting,
have been discussed in brief. Few of the critical issues related to finite element mod-
eling of orthogonal machining have been cited through various case studies. This
would prove very helpful for the readers not simply because it provides basic steps
for formulating FE model for machining but also focuses on the issues that should be
taken care of in order to come up with accurate and reliable FE simulations.

3.1 Introduction
Metal cutting or machining is considered as one of the most important and versa-
tile processes for imparting final shape to the preformed blocks and various manu-
factured products obtained from either casting or forging. Major portion of the
components manufactured worldwide necessarily require machining to convert
them into finished product. This is the only process in which the final shape of the
product is achieved through removal of excess material in the form of chips from
the given work material with the help of a cutting tool. Basic chip formation
processes include turning, shaping, milling, drilling, etc., the phenomenon of chip
formation in all the cases being similar at the point where the cutting edge meets
the work material. During cutting, the chip is formed by deforming the work ma-
terial on the surface of the job using a cutting tool. The technique by which the
metal is cut or removed is complex not simply because it involves high straining
and heating but it is found that conditions of operation are most varied in the
102 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

machining process as compared to other manufacturing processes. Although nu-

merous researches are being carried out in the area of metal cutting both for its
obvious technical and economical importance, machining still is sometimes re-
ferred to as one of the least understood manufacturing processes because of the
complexities associated with the process. Efforts are being made continuously to
understand the complex mechanism of cutting in a simple and effective way.
The knowledge of basic mechanism of chip formation helps understanding the
physics governing the chip formation process. This enables one to solve the prob-
lems associated with metal cutting so that the metal removal process can be car-
ried out more efficiently and economically. A significant improvement in process
efficiency may be obtained by process parameter optimization that leads to desired
outputs with acceptable variations ensuring a lower cost of manufacturing. Re-
searchers have developed various techniques and the associated studies that have
contributed remarkably in the area of metal cutting. The age old technique of trial
and error experiment has contributed greatly and is still widely used in metal cut-
ting research. Various mathematical models have also been developed which form
the core of the metal cutting theory. These models range from simplified analyti-
cal models to complex, computer-based models using Finite Element Method
(FEM).
Metal cutting tests have been carried out for at least 150 years using a variety of
cutting conditions in tremendously increasingly volume [1]. Frederick W. Taylor
is regarded as the great historical figure in the field of metal cutting whose works
published at the end of the nineteenth century had a great impact on manufactur-
ing industry. In industries, the conventional way to optimize metal cutting process
is to conduct experimental tests with varying cutting conditions and then analyze
its effect on various outputs, e.g., cutting forces, temperatures, tool wear and sur-
face finish based on optimization techniques such as Design of Experiment
(DOE), Taguchi or Response Surface Methodology (RSM). The experimental
study is as important as the theoretical study. Unfortunately, the experimental op-
timization approach requires large number of experiments to achieve desired accu-
racy, is very time consuming and expensive.
For more than half a century the analysis of the machining problems has been
carried out by researchers using different analytical models such as shear plane so-
lution and slip line field method [2]. Pioneering works to determine cutting tem-
perature using analytical approach began in 1951 by Hahn [3] followed by Chao
and Trigger [4], Leone [5], Loewan and Shaw [6], Weiner [7] and Rapier [8]. In
analytical modeling, especially in earlier works, uniform plane heat source and ve-
locity discontinuity were often assumed by considering that the chip is formed in-
stantaneously at the shear plane. The secondary deformation zone was usually
neglected and the tool chip frictional force was taken as uniform. Such simplified
assumptions modify the original problem to some extent. Besides, various other
important machining features such as strain hardening, temperature dependence,
chip curling, etc., are neglected owing to the complexity associated with the ma-
chining process.
In case of metal cutting process, taking all the boundary conditions into consid-
eration renders the mathematical equations that make the process so complicated
Finite Element Modeling of Chip Formation in Orthogonal Machining 103

that solutions to such problem seldom exist. One prospect, as done in case of analyt-
ical approach, is to make simplifying assumptions in order to ignore the difficulties
and reduce the problem to one that can be solved. This, however, modifies the actual
problem to great extent and leads to serious inaccuracies. Now that more and more
powerful computers have emerged and are being widely used, a more viable alterna-
tive is to obtain approximate numerical solutions rather than exact form solutions.
This enables one to retain the complexity of the problem on one hand, and the de-
sired accuracy on the other. The most popular numerical technique that has evolved
in recent decades for the analysis of metal cutting is FEM. FEM allows the coupled
simulation of plastic and thermal process and is capable of considering numerous
dependencies of physical constants on each other.
The advantages of FEM over empirical and analytical methods in machining
process can be summed up as follows [9]:
• some of the difficult to measure variables, namely, stress, strain, strain
rate, temperature, etc., can be obtained quantitatively, in addition to cut-
ting forces and chip geometry.
• non-linear geometric boundaries such as the free surface of the chip can
be represented and used.
• material properties can be handled as functions of strain, strain rate and
temperature.
• the interaction between the chip and the tool can be modeled as sticking
and sliding conditions.
Thus, FEM-based analysis provides detailed qualitative and quantitative insight
in to the chip formation process that is very much required for profound under-
standing of the influence of machining parameters. While experimental tests and
analytical models serve as the foundation of metal cutting, FEM leads to the ad-
vancement and further refinement of knowledge in the area of metal cutting.

3.2 Basics of Machining

Metal cutting process involves various independent and dependent variables. In-
dependent variables are the input variables over which the machinist has direct
control [10]. These include type of workpiece material and tool material, shape
and size of workpiece material, cutting tool geometry, type of machining process,
cutting parameters (speed, feed and depth of cut) and cutting fluids. The type of
input parameters selected during machining process decides much about the de-
pendent variables. The important dependent or output variables are cutting force
and power, surface finish, tool wear and tool failure, size and properties of the fi-
nished product. A small change in input variables, say, cutting parameters, tool
geometry and workpiece or tool material may alter the forces to great extent. Sur-
face finish is again a function of tool geometry, tool material, workpiece material
and cutting parameters. Tool wear is also a crucial aspect of machining pertaining
to the economics of machining since longer tool life leads to higher productivity.
Moreover, as the tool wear takes place, it changes in both geometry and size. Such
a change can result in increased cutting forces which in turn increase deflection in
104 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

the workpiece and may create a chatter condition. Again due to increased power
consumption, there can be increased heat generation, thus accelerating the wear
rate. The enormous variety in input variables leads to infinite combinations and
understanding the interrelationship between these input variables and output va-
riables again becomes an arduous task.

3.2.1 Orthogonal Cutting Model

During cutting as the tool touches the work material, severe plastic deformation
occurs within the workpiece material and the excess of material starts flowing
over the rake face of the cutting tool in the form of chip. The zone of plastic de-
formation known as shear plane lies between the chip and the undeformed work
material. Based on the position of cutting edge, the cutting process can be classi-
fied as orthogonal cutting and oblique cutting. In orthogonal cutting, also known
as two-dimensional cutting, the cutting edge is perpendicular to the cutting veloci-
ty vector. On the other hand, in oblique cutting, the edge is inclined with the cut-
ting velocity by a certain angle called the inclination angle. Oblique cutting is a
common type of three-dimensional cutting used in machining process. Orthogonal
cutting is only a particular case of oblique cutting such that any analysis of ortho-
gonal cutting can be applied to oblique cutting [11]. As far as practical require-
ments of rake and other angles are concerned, the ideal conditions of orthogonal
cutting are rarely achieved. But analysis of oblique cutting is much more difficult,
so focus has been mainly given on the orthogonal cutting with very few excep-
tions dealing with the mechanics of oblique cutting. The important theories based
on orthogonal cutting model which have paved the way for the analysis of chip
formation process include Merchant’s model [12], Lee and Shaffer’s model [13],
Oxley’s model [14], just to name a few.
The significant geometrical parameters involved in chip formation (as shown in
Fig. 3.1) are described as follows:

Chip

tc
bc
bo
γo Contact length

to β

Workpiece Vc

Fig. 3.1. Orthogonal cutting model

Finite Element Modeling of Chip Formation in Orthogonal Machining 105

β = Shear plane angle which is defined as the angle between the shear plane
(plane of separation of chip from the undeformed work material) and the cutting
velocity vector

γ0 = Rake angle of the cutting tool

φp = Principal cutting edge angle (shown in Fig. 3.2)

λ = Inclination angle
d = Depth of cut (mm)
s = Feed (mm/rev)
Vc = Cutting velocity (m/min)
to , bo , lo are thickness, width and length (mm) of the uncut chip thickness, re-
spectively, such that

s sin φ p
t0 =
cos λ
(3.1)
d
b0 =
sin φ p

For pure orthogonal cutting case, λ = 0° and φ p = 90° giving t0 = s and

b0 = d .

tc , bc , lc are thickness, width and length (mm) of the deformed chip thickness, re-
spectively.
ζ , Chip reduction coefficient is defined as the index of the degree of deformation
involved in the chip formation such that
tc
ζ = (3.2)
t0
The degree of chip thickening can also be expressed in terms of cutting ratio
⎛ 1⎞
r⎜= ⎟ .
⎝ ζ ⎠
Since tc > t0 , ζ is generally greater than one. Larger value of ζ signifies high-
er value of cutting forces that are required to carry out the machining process. The
chip reduction coefficient is affected by the tool rake angle and chip-tool interfacial
106 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

friction coefficient ( μ ). This is clear from the relation obtained from velocity anal-
ysis based on D’Alemberts’ principle applied to chip motion, given as [15]:

ζ =e
(
μ Π 2 −γ 0 ) (3.3)

The ζ can be reduced considerably by using cutting tools with larger positive
rake angles and by using cutting fluids.
Shear angle can be expressed in terms of ζ :

cos γ 0
tan β = (3.4)
ζ − sin γ 0
It is evident from Eq. (3.4) that decrease in ζ and increase in rake angle tend to
increase the shear angle. This suggests that increasing values of shear angle re-
quires lesser forces for cutting resulting in favorable machining.

3.2.2 Cutting Forces

The first scientific studies of metal cutting were conducted in about 1850, which
were aimed at establishing the power requirements for various operations. A con-
siderable amount of investigations has been directed toward the prediction and
measurement of cutting forces. This is because cutting force is a result of the ex-
treme conditions at the tool-workpiece interface and this interaction can be direct-
ly related to many other output variables such as generation of heat and conse-
quently tool wear and quality of machined surface as well as chip formation
process and the chip morphology [16, 17]. Measurement of forces becomes man-
datory at certain cases say, adequate equations are not available, evaluation of ef-
fect of machining parameters cannot be done analytically and theoretical models
have to be verified. Several works are available in the literature that makes use of
different types of dynamometers to measure the forces. The dynamometers being
commonly used nowadays for measuring forces are either strain gauge or piezoe-
lectric type. Though piezoelectric dynamometer is highly expensive, this has al-
most become standard for recent experimental investigations in metal cutting due
to high accuracy, reliability and consistency.
Estimation of forces acting between tool and work material is one of the vital
aspects of mechanics of cutting process since it is essential for:
• Determination of the cutting power consumption
• Structural design of machine-fixture-tool system
• Study of the effect of various cutting parameters on cutting forces
• Condition monitoring of both the cutting tools and machine tools
Analysis of force during machining includes magnitude of cutting forces and
their components and location of action of those forces as well as the pattern of
the forces, say, static or dynamic.
Finite Element Modeling of Chip Formation in Orthogonal Machining 107

A single cutting force is known to act in case of a single point cutting tool be-
ing used in turning operation which can be resolved into three components along
three orthogonal directions i.e. X , Y and Z , as shown in Fig. 3.2.

Workpiece

FX
FY FT

Φp

Cutting tool

Feed

Fig. 3.2. Forces acting on a cutting tool in turning

The resultant cutting force R is resolved as:

R = Fc + FT (3.5)

FT = FX + FY (3.6)

such that FX = FT sin φ p and FY = FT cos φ p .

The force components are described as follows:

Fc – This is the main or tangential cutting force acting in the direction of cutting
velocity. This when multiplied with cutting velocity gives the value of cutting
power consumption.

FX – It is feed force acting in the direction of feed.

FY – This is radial force acting in the direction perpendicular to the feed force.
FT – This is known as thrust force.
The above-mentioned forces act on the cutting tool during machining process and
can be measured directly using tool force dynamometers. But there exist another
set of forces acting on the chip both from the tool side and the workpiece side that
108 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

cannot be measured directly. Merchant’s circle diagram, one of the oldest

techniques, is widely used for analytical estimation of these kind of forces for
orthogonal cutting process.
Merchant’s circle diagram is shown in Fig. 3.3 with schematic representation of
forces acting on a portion of continuous chip passing through the shear plane at a
constant speed which is in a state of equilibrium.

Chip

Shear plane
Fs γo Tool
Fc Machined surface
β

FT Ns R
F
η
N
Workpiece

Fig. 3.3. Merchant’s circle diagram for 2D orthogonal cutting

Fs and N s are called shear force and normal force, respectively, that act on
the chip from workpiece side i.e. in the shear plane. F and N are friction force
at chip-tool interface and force normal rake face, respectively, that act on the chip
from the tool side i.e. in the chip-tool interface. These forces can be determined as
follows [10]:

Fs = Fc cos β − FT sin β
(3.7)
N s = FT cos β + Fc sin β
F = FT cos γ 0 + Fc sin γ 0
(3.8)
N = Fc cos γ 0 − FT sin γ 0
The average coefficient of friction between chip and tool ( μ ) can be deduced
either in terms of friction angle ( η ) or F and N , as given:
F
μ = tan η = (3.9)
N
Finite Element Modeling of Chip Formation in Orthogonal Machining 109

3.2.3 Cutting Temperature

Heat has a critical influence on machining process. It is found that almost all of
the mechanical energy is converted into heat during chip formation process that
tends to increase the temperature to very high values in the cutting zone. The heat-
ing action is known to pose many of the economic and technical problems of ma-
chining. During machining heat is generated in the cutting zone from three distinct
zones: Primary shear deformation zone (PSDZ), secondary shear deformation
zone (SSDZ) and tertiary shear deformation zone (TSDZ) (see Fig. 3.4). Major
part of the heat is generated due to severe plastic deformation in a very narrow
zone called PSDZ. Further heating is generated in the SSDZ due to friction be-
tween tool and chip as well as shearing at the chip-tool interface. It is considered
that both these zones together account for almost 99% of the total energy con-
verted into heat in the cutting process [18]. A very small fraction of heat is gener-
ated at the TSDZ due to rubbing between flank of the tool and machined surface.
The effect of this zone comes into play only when the flank wear develops. The
TSDZ is therefore neglected when studying unworn tools [19]. The heat so gener-
ated is shared by the chip, cutting tool and the workpiece, the major portion of
heat being carried away by the flowing chip. As chip is meant for disposal only,
focus should be on the chip to take as much heat as possible with it so that small
amount of heat gets retained in the cutting tool.

Chip SSDZ
PSDZ
Tool

TSDZ
Workpiece

Fig. 3.4. Heat generation zones during metal cutting

Elevated temperature in the cutting zone adversely affects the strength, hard-
ness and wear resistance of the cutting tool, thereby inducing rapid tool wear and
reduced tool life. Temperature rise in workpiece material may cause dimensional
inaccuracy of the machined surface and can also damage the surface properties of
the machined component by oxidation, rapid corrosion, burning, etc. Thus, estima-
tion of cutting temperature is a crucial aspect in the study of metal cutting.
Cutting temperatures are more difficult to measure accurately than cutting
forces. No simple analog to the cutting force dynamometer exists for measuring
the cutting temperatures; rather numerous methods have been found in the litera-
ture to experimentally measure the machining temperature [20]. These methods
are thermocouples, radiation methods, metallographic techniques and application
of thermal paints, fine powders and Physical Vapor Deposition (PVD) coatings
[21]. A particular method generally gives only limited information on the com-
plete temperature distribution.
110 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

3.2.4 Chip Morphology

Chip formation and its morphology are the key areas in the study of machining
process that provide significant information on the cutting process itself. The
process variables such as cutting force, temperature, tool wear, machining power,
friction between tool-chip interface and surface finish are greatly affected by the
chip formation process and chip morphology. Chip is formed due to deformation
of the metal lying ahead of the cutting tool tip by shearing process. The extent of
deformation that the work piece material undergoes determines the nature or type
of chip produced. The extent of deformation of chips again depends upon cutting
tool geometry (positive or negative rake angle), workpiece material (brittle or duc-
tile), cutting conditions (speed, feed and depth cut), machining environment (dry
or wet machining) [10].
The main chip morphologies observed in cutting process are described briefly
as follows:
Discontinuous chip: These chips are small segments that adhere loosely to each
other. The phenomenon can be attributed to the repeated fracturing that limits the
amount of deformation the chip undergoes. Hard and brittle materials like gray
cast iron, bronze, brass when machined with larger feed and negative rake cutting
tool in the absence of cutting fluid produce discontinuous chips.
Continuous chip: Continuous chips are in the form of long coil. These are
formed by the continuous plastic deformation of material without fracture ahead
of the cutting edge of the tool resulting in a smooth flow of chip up the tool face.
Ductile material when machined under low feed and high cutting speed, generally,
in the presence of cutting fluid produces such kind of chips.
Continuous chip with built-up edge: This kind of chip is not as smooth as that of
continuous chip and affects the surface finish adversely. Built-up edge is mostly
found when ductile materials are machined under larger feed and lower cutting
speed with inadequate or no cutting fluid.
Cyclic or serrated chips: These chips are continuous, but, possess a saw-tooth
appearance that is produced by a cyclical chip formation of alternating high shear
strain followed by low shear strain. Machining of certain difficult to machine
metals such as titanium alloys, nickel-base super alloys and austenitic stainless
steels are generally characterized by formation of segmented chip formation [22].

3.3 Basics of FEM

The considerable rise in computer processing speeds makes it possible to incorpo-
rate the FEM for analyzing various aspects of metal cutting process, thus over-
coming most of the restrictive assumptions associated with analytical models.
Many investigators have adopted FEM to gain better understanding of machining
process. One of the first FE models developed for metal cutting process was by
Klamecki in the year 1973 [23]. Since then many works have been found to use
FEM for gaining better understanding of the machining process. Numerous FE
Finite Element Modeling of Chip Formation in Orthogonal Machining 111

codes such as DEFORM, FORGE2, ABAQUS/Standard, ABAQUS/Explicit, An-

sys/LS-DYNA, MSC Marc and Thirdwave AdvantEdge have come up that are be-
ing used by the researchers, thus giving results closer to the experimental ones.
However, it should be noted that the accuracy of FE modeling is determined by
how adequately the characterization of selected input parameters reflects the de-
formation behavior undergoing during the chip formation in the actual practice.

3.3.1 Generalized Steps in FEM

The finite element method is a numerical method seeking an approximated solu-
tion to a wide variety of engineering problems. The engineering problems are bas-
ically boundary value problems (or field problems) containing one or more depen-
dent variables that must satisfy a differential equation everywhere within a known
domain and satisfy specific conditions on the boundary of the domain. The field is
the domain of interest, often representing a physical structure while the dependent
variables of interest are called field variables. The specified values of the field va-
riables on the boundaries of the field are called boundary conditions. The field va-
riables may include displacement in solid mechanics problem, temperature in heat
flow problem, velocity in fluid flow problem and so on. In formulating an FE
model for a physical problem, few basic steps are required to be followed that are
common to all kinds of analysis whether structural, heat flow or fluid flow prob-
lems. The basic concepts, procedures and formulations of FEM can be found in
many available books [24–30].
Discretization: The idea of FEM is to approximate the problem domain by replac-
ing it with a finite number of distinct elements and solving them separately. This
is done by subdividing the domain into smaller units of simpler geometry called
finite elements. Such a process is called discretization of domain [25]. The number
of elements used to mesh the domain (mesh density) depends upon the type of
element chosen, accuracy deserved and computer resource available. The elements
selected must be small enough to give practical results and yet large enough to
reduce computational effort.
Interpolation functions: The interpolation function or shape function is in the
form of polynomial which is defined within an element using the values of field
variables at the nodes. For problems involving curved boundaries, the construction
of shape functions that satisfy consistency requirements for higher order elements
becomes increasingly complicated. Moreover, exact evaluation of the integrals
that appear in the expressions of the element stiffness matrix is not always feasible
because of the algebraic complexity. These two obstacles can be surmounted
through the concepts of isoparametric elements and numerical integration, respec-
tively. The numerical integration or numerical quadrature, the most popular of
which is Gaussian (Gauss–Legendre) quadrature involves approximation of the in-
tegrand by a polynomial of sufficient degree because the integral of a polynomial
can be evaluated exactly [30].
112 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

Derivation of element stiffness equation: Derivation of basic equations for a fi-

nite element is based on governing equations and constitutive laws. This prefera-
bly involves weak form of system equation which is often an integral form and re-
quires a weaker requirement on the field variables and the integral operation. A
formulation based on weak form usually leads to a set of algebraic system equa-
tions when the problem domain, no matter how complex, is discretized into small-
er elements [29]. There are three different methods for obtaining weak form,
namely, direct method (for simple problems), variational method and weighted re-
sidual method [30]. The variational method is generally based on either principle
of virtual work, principle of minimum potential energy or Castigliano’s theorem to
derive the element equation [25]. Whereas, method of weighted residual (MWR)
uses trial functions to minimize the error over the problem domain [28]. MWR is
particularly useful when functional such as potential energy is not readily availa-
ble. Galerkin method is one such technique which is widely used for its simplicity
and easy adaptability in FEM. The constitutive equations are employed which are
nothing but the stress-strain laws giving much of the information about the materi-
al behavior. Finally, the element stiffness matrix is derived by substituting the in-
terpolation functions into the weak form.
Assembly of elements: The individual element equations are combined together
using a method of superposition called direct stiffness method to obtain the global
stiffness matrix.
Boundary conditions: After the global equations are ready, necessary boundary
conditions are imposed to the global equations which lead to a set of simultaneous
equations
Solution of system equations: The system equations are often solved by using direct
method such as Guass elimination method and LU decomposition method for small
systems. While iterative methods namely, Gauss–Jacobi method, Gauss-Siedel me-
thod, generalized conjugate residual method, etc. are generally employed for large
systems. For transient problems (highly dynamic and time-dependent problems),
another loop is required meant for time stepping. The widely used method is called
direct integration method such as explicit and implicit methods [29].

3.3.2 Modeling Techniques

With the emergence of more and more potential computers as well as powerful
general purpose FE software packages, FEM is becoming a popular tool in the
analysis of various kinds of manufacturing processes. It is true that there have
been remarkable developments in both computer hardware and software but prop-
er background in FEM is a must otherwise using FEM packages for analysis
would simply appear as a black box. It is the task of the researcher to make use of
such resources tactfully and judiciously to come up with favorable and meaningful
results. Thus, in order to develop a good FE model, few critical issues which may
be referred as modeling techniques must be taken care of. This not only improves
the computational efficiency but also ensures the reliability and accuracy of the
obtained results [29].
Finite Element Modeling of Chip Formation in Orthogonal Machining 113

3.3.2.1 Type of Approach

The selection of an appropriate formulation or approach is of prime importance

especially for problems involving nonlinearity and large deformations. There are
mainly three ways to formulate the problems involving the motion of deformable
materials, given as:
• Lagrangian approach
• Eulerian approach
• Arbitrary Lagrangian Eulerian (ALE) approach
The formulations are distinguished from each other by mesh descriptions, kinetics
and kinematics description [31].
Lagrangian and Eulerian formulations are the two classical approaches for
description of motion in continuum mechanics that differ mostly in terms of the
behavior of nodes. In Lagrangian meshes, the nodes and elements move with
the material i.e. the nodes are coincident with material points. But if the mesh is
Eulerian, the Eulerian coordinates of nodes are fixed, i.e. the nodes are coincident
with spatial points.
Lagrangian approach: It can be seen that in Lagrangian approach, both the ma-
terial points and nodes change position as the body deforms such that position of
the material points relative to the nodes remains fixed. Boundary nodes tend to
remain along the element edges throughout the evolution of the problem which in
turn simplifies the imposition of boundary conditions and interface conditions.
Quadrature points also remain coincident with the material points which signify
constitutive equations are always evaluated at the same material points, thus lead-
ing to a straightforward treatment of constitutive equations. For these reasons, La-
grangian approach is widely used in solid mechanics. However, this approach may
encounter severe mesh distortion because the mesh deforms with the material.
This becomes highly critical when the problems involve large deformation as this
may affect the performance of FE simulations greatly.
Lagrangian formulation can be again of two types, namely, total Lagrangian
formulations and updated Lagrangian formulations [31]. In the total Lagrangian
formulation, the weak form involves integrals over the initial (reference) configu-
ration and derivatives are taken with respect to the material coordinates. In the up-
dated Lagrangian formulation, the weak form involves integrals over the current
i.e. deformed configuration and the derivatives are with respect to the spatial (Eu-
lerian) coordinates. It has been suggested that the updated and total Lagrangian
formulations are two representations of the same mechanical behavior and each
can be transformed to the other, thus, implying that the choice of formulation is
purely a matter of convenience.
Many of the researchers have used Lagrangian formulation in metal cutting
models since it is easy to implement and is computationally efficient because of
easy tracking of the material boundaries. But difficulties arise when elements get
highly distorted as the mesh deforms with the deformation of the material in front
of the tool tip. Since element distortion degrades its element accuracy, the degree
114 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

of deformation that can be simulated with a Lagrangian mesh is restricted. In order

to overcome this, either chip separation criteria or remeshing techniques have been
used in the literature to simulate the cutting action at the cutting zone [32].

Eulerian approach: In Eulerian approach, the reference frame is fixed in space

that allows for the material to flow through the grid [33]. As the mesh is fixed in
space, the numerical difficulties associated with the distortion of elements are
eliminated. This approach permits simulation of machining process without the
use of any mesh separation criterion. The main drawback of Eulerian formulation
is that it is unable to model free boundaries since boundary nodes may not be
coincident with the element edge. This may only be used when boundaries of the
deformed material are known a priori. Hence, dimension of the chip must be spe-
cified in advance to produce a predictive model for chip formation in an Eulerian
framework [34]. Furthermore, handling and updating of constitutive equations
become complex because of the fact that material flows through the mesh i.e.
element quadrature points may not be coincident with material points. These
drawbacks often limit the application of Eulerian formulation in modeling the
metal cutting process.

Arbitrary Lagrangian Eulerian approach: It is known that Eulerian and La-

grangian approaches have their own advantages and disadvantages. As the name
suggests, in ALE formulation the nodes can be programmed to move arbitrarily in
such a way that advantages of both Lagrangian and Eulerian formulations can be
exploited and combined into one while minimizing their disadvantages. The idea
is that the boundary nodes are moved to remain on the material boundaries, while
the interior nodes are moved to minimize mesh distortion. While simulating metal
cutting process in ALE framework, ALE reduces to a Lagrangian form on free
boundaries while maintains an Eulerian form at locations where significant defor-
mations occur, as found during the deformation of material in front of the tool tip.
This approach not only enables to reduce mesh distortion without the need of re-
meshing but also eliminates the need of a priori assumption about the shape of the
chip [35].

3.3.2.2 Geometry Modeling

Creating the geometry of the problem domain is the first and foremost step in any
analysis. The actual geometries are usually complex. The aim should not be simp-
ly to model the exact geometry as that of the actual one, instead focus should be
made on how and where to reduce the complexity of the geometry to manageable
one such that the problem can be solved and analyzed efficiently without affecting
the nature of problem and accuracy of results much. Hence, proper understanding
of the mechanics of the problem is certainly required to analyze the problem and
examine the geometry of the problem domain. It is generally aimed to make use of
2 D elements rather than 3 D elements since this can drastically reduce the number
of degrees of freedom (DOFs).
Finite Element Modeling of Chip Formation in Orthogonal Machining 115

3.3.2.3 Computation Time

Computation time which is nothing but the time that CPU takes for solving finite
element equation is affected markedly by the total number of DOFs in the FE equ-
ation as shown [29]:

tcomp ∝ DOF β (3.10)

Here, β is a constant which generally lies in the range of 2-3 depending on the
type of solver used and the structure or bandwidth of the stiffness matrix. A small-
er bandwidth yields smaller value of β resulting in faster computation. The Eq.
(3.10) suggests that a finer mesh with larger number of DOFs increases the com-
putation time exponentially. Thus, it is always preferred to create FE model with
elements possessing lower dimensions so that number of DOFs are reduced as far
as possible. In addition, meshing should be done in such a way that critical areas
possess finer meshing while others possess coarse meshing. This is one way of re-
ducing the computation time without hampering the accuracy of results.

3.3.2.4 Mesh Attributes

Choice of element type: Based on the shape, elements can be classified as one
dimensional (line or beam), two-dimensional or plane (triangular and quadrilater-
al) and three-dimensional (tetrahedral and hexahedral) elements. Each of these
elements can again be either in their simplest forms i.e. linear or higher order
forms such as quadratic or cubic depending upon the number of nodes an element
possess. As discussed earlier, 2D elements are generally preferred over 3D ele-
ments as far as cost of computation is concerned. The linear triangular element
was the first type of element developed for defining 2D geometries, the formula-
tion of which is simplest of all. However, quadrilateral elements are mostly pre-
ferred nowadays for 2D solids [36]. This is not simply because quadrilateral ele-
ment contains one node more than that of triangular elements but also gradients of
quadrilateral elements are linear functions of the coordinate directions compared
to the gradient being constant in triangular elements. Besides, the number of ele-
ments is reduced to half that of a mesh consisting of triangular elements for the
same number of nodes. In nearly all instances, a mesh consisting of quadrilateral
elements is sufficient and usually more accurate than that of triangular elements.
Besides for higher order elements, more complex representations are achieved that
are increasingly accurate from an approximation point of view. However, care
must be taken to evaluate the benefits of increased accuracy against the computa-
tional cost associated with more sophisticated elements.
Mesh density: A mesh of varying density is generally preferred. The critical areas
need to be finely meshed while rest of the areas may contain coarse mesh. The
control of mesh density can be performed by placing the nodes according to the
given geometry and required degree of accuracy in FEM packages.
116 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

Element distortion: It is not always possible to have regularly shaped elements

for irregular geometries. Irregular or distorted elements are acceptable in the FEM,
but there are limitations, and one needs to control the degree of element distortion
in the process of mesh generation. If these are overlooked then there are chances
of termination of simulation program because of element distortion. Few possible
forms of distortion error of elements are aspect ratio distortion, angular distortion,
curvature distortion and so on. Aspect ratio error arises when the elongation of
element occurs beyond a particular limit disturbing the aspect ratio of the element.
Angular distortion occurs when the included angle between edges of the element
approaches either 0° or 180°. Similarly, when the straight edges of the element are
distorted into curves, curvature distortion is most likely to occur [29].

3.3.2.5 Locking and Hourglassing

Approximating the problem domain using simpler elements is relatively easy, but,
very often formulation of these elements can give inaccurate results due to some
abnormalities namely, shear locking, volumetric locking and hourglassing. Lock-
ing, in general, is the phenomenon of exhibiting an undesirable stiff response to
deformation in finite elements. This kind of problem may arise when the element
interpolation functions are unable to approximate accurately the strain distribution
in the solid or when the interpolation functions for the displacement and their de-
rivatives are not consistent. Shear locking predominantly occurs in linear elements
with full integration and results in underestimated displacements due to undesira-
ble stiffness in the deformation behavior. This problem can be alleviated by using
reduced integration scheme. Volumetric locking is exhibited by incompressible
materials or materials showing nearly incompressible behavior resulting in an
overly stiff response [37].
It is known that the stress comprises of two components: deviatoric (distortion-
al) and volumetric (dilatational). The volumetric component is a function of bulk
modulus and volumetric strain. The bulk modulus is given by:

E
K= (3.11)
3 − 6ν
where E is Young’s modulus and ν is Poisson’s ratio. It is understood that
when ν = 0 , there is no volumetric locking at all. But in the incompressibility
limit, when ν → 0.5 then lim K = ∞ resulting in overly stiff response. In
ν →0.5
this limit, the finite element displacements tend to zero and thus, the volumetric
locking prevails [37]. Hybrid Elements are often used to overcome volumetric
locking [38]. The idea is to include the hydrostatic stress distribution as an addi-
tional unknown variable, which must be computed at the same time as the dis-
placement field. This allows the stiff terms to be removed from the system of fi-
nite element equations. Further details on hybrid elements can be found in existing
literatures [39, 40].
Finite Element Modeling of Chip Formation in Orthogonal Machining 117

It has been also observed that fully integrated elements have higher tendency of
volumetric locking. This is because at each of the integration points the volume
remains nearly constant and when the integration points are more, as in case of
full integration scheme, it results in overconstraining of the kinematically admiss-
ible displacement field. Reduced integration scheme is a possible way to avoid
locking wherein the element stiffness is integrated using fewer integration points
than that are required for full integration scheme [24]. Reduced integration is an
effective measure for resolving locking in elements such as quadratic quadrilateral
and brick effectively but not in elements such as 4 noded quadrilateral or 8 noded
brick elements. The error occurs because the stiffness matrix is nearly singu-
lar which implies that the system of equations includes a weakly constrained de-
formation mode. This phenomenon is known as hourglassing which results in
wildly varying displacement field but correcting stress and strain fields. This can
be cured either by employing selectively reduced integration or by adding an ar-
tificial stiffness to the element that acts to constrain the hourglass mode (Reduced
Integration with Hourglass Control) [37].

3.3.2.6 Solution Methods

Time integration methods are employed for time stepping to solve the transient
dynamic system of equations. There are two main types of time integration me-
thods [29, 31]:
• Implicit
• Explicit
Both the procedures have their own benefits and limitations from solution point of
view. Selection of an appropriate solution method should be done carefully ac-
cording to the nature of the problem. The implicit method is generally suitable for
linear transient problems. In an implicit dynamic analysis, a global stiffness matrix
is assembled and the integration operator matrix must be inverted and a set of non-
linear equilibrium equations must be solved at each time increment. Suitable im-
plicit integrator, which is unconditionally stable, is employed for non-linear for-
mulations and the time increment is adjusted as the solution progresses in order to
obtain a stable, yet time-efficient solution. However, there are situations where
implicit method encounters problems finding a converged solution. In analyses
such as machining which involves complex contact problems, this algorithm is
less efficient due to the simultaneous solving of the equation system in every in-
crement. In such situations, explicit time integration method proves to be robust.
The simulation program seldom aborts due to failure of the numerical algorithm
since the global mass and stiffness matrices need not be formed and inverted.
However, the method is conditionally stable i.e. if the time step exceeds critical
time step, the solution may grow unboundedly and give erroneous results. The
critical time step for a mesh is given by [31]:

Δtcrit ≤ min ( Le ce ) (3.12)

118 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

where Le is the characteristic length of the element and ce is the current wave
speed of an element.

3.3.2.7 Mesh Adaptivity

Generally speaking, mesh adaptivity can be described as the capability of the

mesh to adapt according to the nature of problem in order to maintain a high quali-
ty mesh throughout the analysis. The need of such technique becomes more evi-
dent while simulating problems that involve heavy mesh distortion such as in the
case of cutting process undergoing large deformations. This is well illustrated by
citing an example which considers cutting process as a purely large deformation
problem comprising a rectangular block as workpiece and a portion of cutting tool
(see Fig. 3.5). Figure 3.5 (a) shows the initial mesh configuration of the workpiece
while Fig. 3.5 (b) shows the deformed mesh configuration as tool comes in contact
with the workpiece. Since no mesh adaptivity technique is employed, elements get
highly distorted near the tool tip as highlighted in Fig. 3.6 (b) that leads to termi-
nation of the simulation program.

(a) (b)
Fig. 3.5. (a) Initial mesh configuration and (b) Deformed mesh configuration

Figures 3.6 (a) and 3.6 (b) show two possible mesh adaptivity techniques that
are implemented to control the mesh distortion near the tool tip.

(a) (b)
Fig. 3.6. (a) Mesh refinement and (b) Relocation of nodes near the tool tip
Finite Element Modeling of Chip Formation in Orthogonal Machining 119

In the first case (Fig. 3.6 (a)), it is noted that refinement of mesh takes place at
critical areas i.e. finer mesh in the region closer to the tool tip wherein the size of
elements is adjusted based on selected error indicators and loading history. Here, the
number of elements and connectivity of the nodes are changed. This type of tech-
nique is often known as h–adaptivity [41]. In Fig. 3.6 (b), mesh distortion is con-
trolled by relocation of nodes without altering the number of elements and connec-
tivity of nodes. This approach is often termed as r–adaptivity [41]. Since the number
of elements remains the same, this approach is computationally less expensive
as compared to h–adaptivity. However, both h-refinement and r–refinement are
widely implemented in simulating metal cutting process, the former being used in
Lagrangian framework while the latter being employed in ALE framework.

3.4 Brief History of FEM in Machining

Pioneering works in FEM include works by Klamecki [23], Tay et al. [42], Mura-
ka et al. [43]. Klamecki [23] was the first one to apply FEM in the analysis of
metal cutting process, whereas Tay et al. [42] were the first to calculate the tem-
perature field distribution in orthogonal machining by the application of FEM.
Muraka et al. [43] investigated the effect of process variables such as cutting
speeds, flank wear, coolant, etc., on the temperature distribution using finite ele-
ment method. Morikawa et al. [44] developed a rigid-plastic finite element model
to examine the effect of the tool edge radius to depth of cut ratio on the micro ma-
chining process. Kim and Sin [45] developed a thermo-visco-plastic cutting model
by using finite element method to analyze the mechanics of steady state orthogon-
al cutting process. Liu and Guo [46], on the other hand, proposed a thermoelastic-
viscoplastic FE model to investigate the friction at the tool chip interface. Cerreti
et al. [47] used DEFORM 2D to estimate the chip and tool temperatures for conti-
nuous and segmented chip. Li et al. [48] employed Johnson–Cook’s model as con-
stitutive equation of the workpiece material for qualitative understanding of crater-
wear from the calculated temperature distribution by using commercial FE code
ABAQUS. Arazzola et al. [49] incorporated ALE formulation along with Johnson
Cook model as the material model for the workpiece using ABAQUS/Explicit. In
many of the works, researchers have attempted to make comparative study of the
effect of various material models such as Johnson-Cook model, Zerilli–Armstrong
model, Power law rate-dependent model, Oxley model, Litonski–Batra and many
more on the predicted results [50–52]. Adibi–Sedeh et al. [51] found that John-
son–Cook model predicts chip thickness better than that of other models when
compared with the experimental ones. In few of the works, comparison of some
selected friction models has been presented showing their effect on FE simulations
for orthogonal cutting of various work materials [53–55].
Mabrouki et al. [56] compared the results obtained from ABAQUS/Explicit and
Thirdwave Systems’ AdvantEdge with those obtained experimentally. Coelho et
al. [57] developed an ALE–FEM model to help understand the wear results found
in case of coated PCBN tools for turning hardened AISI 4340. Lorentzon and
Jarvstrat [58] investigated the impact of different wear models and friction models
on the parameters affecting the wear process for cemented carbide tools while
120 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

machining alloy 718, using the commercial software MSC Marc. Davim et al. [59]
made a comparative study between the performances of PCD (polycrystalline di-
amond) and K10 (cemented carbide) tools while machining aluminum alloys using
AdvantEdge software. Attanasio et al. [60] presented 3D numerical predictions of
tool wear based on modified Takeyama and Murata wear model using SFTC De-
form 3D software.

3.5 Formulation of Two-Dimensional FE Model for Machining

This section attempts to make understand the basic methodology and the related
theory for developing the FE model of orthogonal cutting process. Furthermore, a
brief description of various modules contained in ABAQUS program has been
presented so that readers become conversant with ABAQUS. It is agreed that
ABAQUS software is provided with extensive and excellent manuals [61–63], but
our aim is to abridge those details in a very lucid manner that may serve as a quick
guide to a beginner. This would facilitate the readers to have a much clearer un-
derstanding of the finite element concepts being implemented in the finite element
software for simulating machining operation.

3.5.1 Formulation Steps

The present study focuses on developing a 2D FE model of chip formation process
based on orthogonal cutting conditions. Although two–dimensional analysis is a re-
strictive approach from practical point of view, it reduces the computational time
considerably and provides satisfactory results regarding the details of the chip for-
mation. A plane strain condition was assumed because the feed value is generally
very less as compared to the depth of cut. A fully coupled thermal-stress analysis
module of finite element software ABAQUS/Explicit version 6.7–4 has been em-
ployed to perform the study of chip formation process. This makes use of elements
possessing both temperature and displacement degrees of freedom. It is necessary
because metal cutting is considered as a coupled thermo-mechanical process, where-
in, the mechanical and thermal solutions affect each other strongly.

3.5.1.1 Geometric Model, Meshing and Boundary Conditions

The 2D model comprises a rectangular block representing the workpiece and only
a portion of cutting tool which participates in the cutting. The nose radius is neg-
lected for the simplicity of the problem. Moreover, there is hardly any effect of
nose radius once a steady state is reached in cutting. The cutting tool includes the
following geometrical angles: inclination angle χ = 90º , rake angle γ = −6º
and flank angle α = 5º . An intermediate layer of elements known as damage
zone (highlighted region in Fig. 3.7) has been considered in the workpiece block
such that width of the upper surface is equal to the undeformed chip thickness or
in other words feed in case of orthogonal cutting process. As mechanical boundary
conditions, bottom of work piece is fixed in Y direction and left vertical edge of
Finite Element Modeling of Chip Formation in Orthogonal Machining 121

workpiece is fixed in X direction. The former not only constrains the movement
of workpiece in the Y direction but also aids in calculating feed force during ma-
chining while the latter, not only constrains the movement of workpiece in the
X direction but also aids in calculating cutting force during machining. The reac-
tion force components when added at all the constrained nodes of the left vertical
edge of the workpiece in X direction give the cutting force values while at bot-
tom edge of the workpiece in Y direction give the feed force. Tool is given the
cutting velocity in negative X direction and top edge of the tool is constrained in
Y direction.
Similarly, as thermal boundary conditions the tool and the workpiece are in-
itially considered at the room temperature. Heat transfer from the chip surface to
cutting tool is allowed by defining the conductive heat transfer coefficient (h),
given as:

∂T
−k = h (To − T ) (3.13)
∂n
where k is thermal conductivity and To is the ambient temperature.
The geometric model taken into consideration is shown in Fig. 3.7.

Tool

Y
Damage zone
Z X

Workpiece

Fig. 3.7. 2D geometric model

Damage zone is a sacrificial layer of elements that defines the path of separation
between chip surface and the machined surface which is going to take place as the
tool progresses. In actual practice, these two surfaces should be same but such as-
sumption has been taken only for the modeling purpose as a chip separation crite-
rion. The choice of the height of the designated damage zone is purely based on
computational efficiency. Generally, a very small value (say, 10–30 µm) which is
computationally acceptable has been taken as the width of the damage zone.
Four-node plane strain bilinear quadrilateral (CPE4RT) elements with reduced
integration scheme and hourglass control are used for the discretization for both
122 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

the workpiece and the cutting tool. The workpiece is meshed with CPE4RT-type
elements by unstructured grid generation which utilizes advancing front algorithm
in ABAQUS/Explicit.

3.5.1.2 Governing Equations

The governing equations of a body undergoing deformation consists of two sets of

equations, namely, the conservations laws of physics and the constitutive equa-
tions. The conservation laws can be applied to body of any material. But to distin-
guish between different materials undergoing deformation of varied degrees, one
needs the constitutive equations.
Conservation laws: The mass, momentum and energy equations governing the
continuum are given as follows:
ρ + ρ div v = 0 (3.14)

ρ v = f + div σ (3.15)

ρ e = σ : D − div q + r (3.16)

where ρ is the mass density, v material velocity, f body forces, σ Cauchy

stress tensor, e specific internal energy, D strain rate tensor, q heat flux vector
and r is body heat generation rate. The superposed ‘ • ’ denotes material derivative
in a Lagrangian description and ‘:’ denotes contraction of a pair of repeated indic-
es which appear in the same order such as, A:B = Aij Bij .
The basic idea of using FEM is to discretize the above equations and then to
seek a solution to the momentum equation.
Workpiece material constitutive equations: Constitutive equations describe the
thermo-mechanical properties of a material undergoing deformation. Based on the
simplicity or complexity of the material behavior, there could be one constitutive
equation or a set of constitutive equations that relate the deformation in the body
to the stresses. The elastic response can be described by Hook’s law. Whereas, the
constitutive equations of plasticity deal with yield criterion, flow rule and strain
hardening rule. The yield criterion describes the stress state when yielding occurs,
the flow rule defines increment of plastic strain when yielding occurs, and the har-
dening rule describes how the material is strain hardened as the plastic strain in-
creases. For large deformation problems, as in case of machining, plasticity mod-
els based on von Mises yield criterion and Prandtl–Reuss flow rule are generally
used to define the isotropic yielding and hardening [64].
It is known that during cutting process, the workpiece material is generally
subjected to high levels of, strain, strain rate and temperature which significantly
influences flow stress. Thus, accurate and reliable rate-dependent constitutive
models are required that can reflect such phenomenon effectively. Johnson–Cook
constitutive equation is one such model that considers the flow stress behavior of
Finite Element Modeling of Chip Formation in Orthogonal Machining 123

the work materials as multiplicative effects of strain, strain rate and temperature,
given as follows [65]:

⎛ ⎛ T −T ⎞
( )
m
⎞
σ = A + B (ε )
p n
(1 + C lnε ) ⎜⎜1 − ⎜ T − Troom ⎟ ⎟⎟
p∗
(3.17)
⎝ ⎝ m room ⎠ ⎠
εp ⎛ T − Troom ⎞
ε = p for ε 0 p = 1 s-1 and T * = ⎜
∗
⎟ (3.18)
ε0 ⎝ Tm − Troom ⎠
where Troom is the room temperature taken as 25 ºC, Tmelt is the melting temper-
ature of the workpiece, A is the initial yield stress (MPa), B the hardening mod-
ulus, n the work-hardening exponent, C the strain rate dependency coefficient
(MPa), and m the thermal softening coefficient. A, B, C, n and m used in the mod-
el are actually the empirical material constants that can be found from different
mechanical tests. Johnson-Cook model has been found to be one of the most suit-
able one for representing the flow stress behavior of work material undergoing
cutting. Besides, it is also considered numerically robust as most of the variables
are readily acceptable to the computer codes. This has been widely used in model-
ing of machining process by various researchers. [52, 66–68].

3.5.1.3 Chip Separation Criterion

A damage model should be incorporated in the damage zone along with the ma-
terial as a chip separation criterion in order to simulate the movement of the cut-
ting tool into workpiece without any mesh distortion near the tool tip. Specifica-
tion of damage model includes a material response (undamaged), damage
initiation criterion, damage evolution and choice of element deletion.
Damage initiation criterion is referred to as the material state at the onset of
damage. In the present case, Johnson–Cook damage initiation criterion has been
employed. This model makes use of the damage parameter ωD defined as the
sum of the ratio of the increments in the equivalent plastic strain Δε to the frac-
p

ture strain ε , given as follows [61].

f

Δε p
ωD = ∑ (3.19)
εf
The fracture strain εf is of the form as follows [69]:

ε f = ⎡⎣ D1 + D2exp ( D3σ ∗ ) ⎤⎦ × ⎡⎣1 + D4 lnε p∗ ⎤⎦ × ⎡⎣1 + D5T * ⎤⎦ (3.20)

−P
σ∗ = (3.21)
σ
124 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

where P is the hydrostatic pressure and D1 to D5 are failure parameters deter-

mined experimentally. The damage constants can be determined by performing
several experiments such as tensile test, torsion test and Hopkinson bar test. On
the basis of results obtained, a graph is generally plotted between equivalent plas-
tic strain at fracture and pressure stress ratio such that the constants related to
pressure i.e. D1, D2, and D3 can be found out at T * = 0 and ε p = 0 . Similarly,
*
D4 and D5 can be obtained by plotting another graph between T and ratio of frac-
ture strain at different strain rates. The damage initiation criterion is met when
ωD (Eq. 3.20) reaches one [61].
Once the element satisfies the damage initiation criterion, it is assumed that
progressive degradation of the material stiffness occurs, leading to material failure
based on the damage evolution. At any given time during the analysis, the stress
tensor in the material is given by:

σ = (1 − D ) σ (3.22)

where σ is the effective (undamaged) stress tensor computed in the current in-
crement. When overall damage variable D reaches a value 1, it indicates that the
material has lost its load carrying capacity completely. At this point, failure occurs
and the concerned elements are removed from the computation.
p
The effective plastic displacement ( u ), after the damage initiation criterion is
met can be defined with the evolution law as follows:

u p = Leε p
(3.23)

where Le is the characteristic length of the element and ε p equivalent plastic

strain.
When a linear evolution of the damage variable with plastic displacement is as-
sumed, the variable increases as per the following [61]:

Leε p u p
D= = p
uf p uf
(3.24)
2G f
uf =p

σ y0

where u f p is the plastic displacement at failure, G f is the fracture energy per

unit area and σ y 0 is the value of the yield stress at the time when the failure crite-
rion is reached.
Finite Element Modeling of Chip Formation in Orthogonal Machining 125

The model ensures that the energy dissipated during the damage evolution
process is equal to G f only if the effective response of the material is perfectly
plastic (constant yield stress) beyond the onset of damage. In this study, G f is
provided as an input parameter which is a function of fracture toughness KC,
Young's modulus E and Poisson's ratio ν as given in the equation for the plane
strain condition [70]:

⎛ 1 −ν 2 ⎞ 2
Gf = ⎜ ⎟ KC . (3.25)
⎝ E ⎠
The ELEMENT DELETION = YES module along with the Johnson Cook damage
model of the software enables to delete the elements that fail. This produces the
chip separation and allows the cutting tool to penetrate further into the workpiece
through a predefined path (damage zone).

3.5.1.4 Chip-Tool Interface Model and Heat Generation

A contact is defined between the rake surface and nodes of the workpiece materi-
al. Coulomb’s law has been assumed in the present study to model the frictional
conditions as the chip flows over the rake surface.
During the machining process, heat is generated in the primary shear deforma-
tion zone due to severe plastic deformation and in the secondary deformation zone
due to both plastic deformation and the friction in the tool–chip interface. The
steady state two-dimensional form of the energy equation governing the orthogon-
al machining process is given as:

⎛ ∂ 2T ∂ 2T ⎞ ⎛ ∂T ∂T ⎞
k⎜ 2 + 2 ⎟ − ρ Cp ⎜ u x + vy ⎟+q = 0 (3.26)
⎝ ∂x ∂y ⎠ ⎝ ∂x ∂y ⎠

q = qp + qf (3.27)

qp = η pσε p (3.28)

qf = ηf Jτγ (3.29)

where q p is the heat generation rate due to plastic deformation, ηp the fraction of
the inelastic heat, qf is the volumetric heat flux due to frictional work, γ the slip
rate, ηf the frictional work conversion factor considered as 1.0, J the fraction of
the thermal energy conducted into the chip, and τ is the frictional shear stress.
The value of J may vary within a range, say, 0.35 to 1 for carbide cutting tool
[71]. In the present work, 0.5 (default value in ABAQUS) has been taken for all
126 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

the cases. The fraction of the heat generated due to plastic deformation remaining
in the chip, ηp , is taken to be 0.9 [56].

3.5.1.5 Solution Method

An ALE approach is incorporated to conduct the FEM simulation. This avoids se-
vere element distortion and entanglement in the cutting zone without the use of
remeshing criterion.
ALE formulation: In the ALE approach, the grid points are not constrained to re-
main fixed in space (as in Eulerian description) or to move with the material
points (as in Lagrangian description) and hence have their own motion governing
equations. ALE description is given as follows:
∼ •
( ) = ( )+ c ∇ ( ) (3.30)

c = v − vˆ (3.31)

where superposed ‘ ∼ ’ is defined in the ALE description, c is the convective ve-

locity, v is the material velocity, v̂ is the grid velocity and ∇ is the gradient op-
erator. The Eqs. (3.14) to (3.16) can now be rewritten in the ALE framework, as
given below:
ρ + c ∇ρ + ρ div v = 0 (3.32)

ρ v + ρ c ∇v = f + div σ (3.33)

ρ e + ρ c ∇e = σ : D − div q + r (3.34)

Explicit dynamic analysis: The explicit dynamic ALE formulation is primarily

used to solve highly non-linear problems involving large deformations and chang-
ing contact, as observed in the case of machining. The explicit dynamic procedure
in ABAQUS/Explicit is based upon the implementation of an explicit integration
scheme for time dicretization of Eqs. (3.32) to (3.34). This analysis calculates the
state of a system at a later time from the state of system at the current time.
The explicit formulation advances the solution in time with the central differ-
ence method [31]:

(
vi = M -1 f ext − f int
(i ) (i)

) (3.35)

Δt (i +1) − Δt i i
v (i +1/ 2) = v ( i −1/ 2) + v (3.36)
2
Δt (i +1) − Δt i (i +1/ 2)
x (i +1) = xi + v (3.37)
2
Finite Element Modeling of Chip Formation in Orthogonal Machining 127

where x is the displacement, v is the velocity and v is the acceleration, M is the

ext int
lumped mass matrix, f is the external force vector and f is the internal
force. The subscript (i) refers to the increment number (i–1/2) and (i+1/2) and
refers to mid-increment values. In this algorithm, the accelerations are first inte-
grated to obtain the velocities, integration of which is broken into two half-steps
so that the velocities are available at an integer step in the computation of the
energy balance. The displacements are then computed in each time step by inte-
grating the velocities.

3.5.2 ABAQUS Platform

The type of software package chosen for the FE analysis of metal cutting
process is equally important in determining the quality and scope of analysis
that can be performed. There are currently large number of commercial software
packages available for solving a wide range of engineering problems that might
be static, dynamic, linear or non-linear. Some of the dominant general purpose
FE software packages include ABAQUS, ANSYS, MSC/NASTRAN, SRDC-
IDEAS, etc. It is obvious that different packages would possess different
capabilities. This makes it critical to select the suitable software package with
appropriate features required for performing a given analysis successfully. The
present study selects ABAQUS as a platform to explore the capabilities of finite
element method in analyzing various aspects of metal cutting process. ABAQUS
is known to be powerful general purpose FE software that can solve problems
ranging from relatively simple linear analyses to the highly complex non-linear
simulations. This software does not have any separate module for machining as
in the case of Deform or AdvantEdge. As a result, the user has to explicitly de-
fine the tool and the workpiece, process parameters, boundary conditions, mesh
geometry and simulation controls. This may certainly require more skill, effort
and time to set up simulations as no preset controls and assumptions are availa-
ble. But this is the feature that not only ensures very high level of details from
modeling point of view but also a thorough analysis by allowing a precise
control on boundary conditions, mesh attribute, element type, solver type and
so on.
A complete ABAQUS program can be subdivided into three distinct modules,
namely, ABAQUS/CAE, ABAQUS/Standard or ABAQUS/Explicit and ABAQUS/
Viewer as shown in Fig. 3.8. These modules are linked with each other by input and
output files. ABAQUS/Standard and ABAQUS/Explicit are the two main types of
solvers that are available for performing analysis, ABAQUS/Explicit being mainly
used for explicit dynamic analysis. It is said that the strength of ABAQUS program
greatly lies in the capabilities of these two solvers.
128 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

ABAQUS/CAE
(Preprocessing)

Input file
(*.inp)

ABAQUS/Standard or ABAQUS/Explicit
(Analysis)

Output file
(*.odb)

ABAQUS/CAE or ABAQUS/Viewer
(Post-processing)

Fig. 3.8. Work flow of ABAQUS program

The model of the physical problem is created in the pre-processing stage, de-
tails of which such as discretized geometry, material data, boundary conditions,
element type, analysis type and output request are contained in the input file.
ABAQUS/CAE is divided into functional units called modules with the help of
which the FE model can be created, input file can be generated and results can be
extracted from the output file. Each module has been designed to serve a specific
portion of the modeling task. The subsequent subsections would discuss about
various modules of ABAQUS/CAE in brief.
Part module: Individual parts are created in the part module either by sketching
their geometry directly in Abaqus/CAE or by importing their geometry from other
geometric modeling programs. Depending upon the analysis the parts can be 2D
or 3D deformable, discrete rigid, analytical rigid or Eulerian parts. In the present
study, both the cutting tool and the workpiece are considered as 2D deformable
bodies. The part tools contained in this module allow editing and manipulating the
existing parts defined in the current model.
Property module: Property module allows assigning sections to a part instance or
region of a part instance to which various material properties are defined. A ma-
terial definition specifies all the property data relevant to a particular analysis. For
a coupled temperature displacement analysis, both the mechanical strength proper-
ties (elastic moduli, yield stress, etc.) and heat transfer properties (conductivity,
specific heat) must be given as inputs. Various plasticity models and damage
models are also contained in the property module. As an input the material model
Finite Element Modeling of Chip Formation in Orthogonal Machining 129

constants of the selected plastic model and damage model such as Johnson–Cook
material model and Johnson–Cook damage model, respectively, are defined in a
tabular form.
Assembly module: The individual parts that are created in part module exist in
their own coordinate system. It is in the assembly module that these parts are as-
sembled by relative positioning with respect to each other in a global coordinate
system.
Step module: The step module is used to create and configure analysis steps, as in
the present case is coupled temperature displacement explicit dynamic analysis.
The associated output requests can also be created. The sequence of steps provides
a convenient way to capture changes that may occur in the model during the
course of the analysis such as changes in the loading and boundary conditions of
the model or changes in the interaction, etc. In addition, steps allow you to change
the analysis procedure, the data output and various controls. An output request
contains information regarding which variables will be output during an analysis
step, from which region of the model they will be output, and at what rate they
will be output. The general solution controls and solver controls can also be cus-
tomized. Furthermore, adaptive mesh regions and the controls for adaptive mesh-
ing in selected regions can be specified in this module. It is noted that implemen-
tation of suitable mesh adaptivity (remeshing in Lagrangian framework or
repositioning of nodes in ALE framework) depends upon the type of analysis taken
into consideration i.e. implicit (ABAQUS/Standard) or explicit (ABAQUS/Explicit).
Adaptive remeshing is available in ABAQUS/Standard while ALE adaptive
meshing is available in ABAQUS/Explicit.
Interaction module: Interaction module allows to specify mechanical and thermal
interactions between regions of a model or between a region of a model and its
surroundings. Surface-to-surface contact interaction has been used to describe
contact between tool and workpiece in the present study. The interaction between
contacting bodies is defined by assigning a contact property model to a contact in-
teraction which defines tangential behavior (friction) and normal behavior. Tan-
gential behavior includes a friction model that defines the force resisting the rela-
tive tangential motion of the surfaces. While normal behavior includes a definition
for the contact pressure–overclosure relationship that governs the motion of the
surfaces. In addition, a contact property can contain information about thermal
conductance, thermal radiation and heat generation due to friction. ABAQUS/ Ex-
plicit uses two different methods to enforce contact constraints, namely, kinematic
contact algorithm and penalty contact algorithm. The former uses a kinematic pre-
dictor/corrector contact algorithm to strictly enforce contact constraints such that
no penetrations are allowed, while the latter has a weaker enforcement of contact
constraints. In this study, a kinematic contact algorithm has been used to enforce
contact constraints in master–slave contact pair, where rake surface of the cutting
tool is defined as the master surface and chip as the slave (node-based) surface.
Both the frictional conditions and the friction-generated heat are included in the
kinematic contact algorithm through TANGENTIAL BEHAVIOUR and GAP
HEAT GENERATION modules of the software.
130 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

Load module: The Load module is used to specify loads, boundary conditions and
predefined fields.
Mesh module: The Mesh module allows generating meshes on parts and assem-
blies created within ABAQUS/CAE as well as allows selecting the correct ele-
ment depending on the type of analysis performed (as in the present case is
CPE4RT) for discretization. Variety of mesh controls are available that help in se-
lecting the element shape (tri, quad or quad dominated), meshing technique (struc-
tured or unstructured) or meshing algorithm (medial axis or advancing front). The
structured meshing technique generates structured meshes using simple predefined
mesh topologies and is more efficient for meshing regular shapes. Free meshing,
however, allows more flexibility than structured meshing. Two commonly used
free mesh algorithms while meshing with quadrilateral elements are medial axis
and advancing front. The advancing front is generally preferable because it gene-
rates elements of more uniform size (area–wise) with more consistent aspect ratio.
Since in ABAQUS/Explicit small elements control the size of the time step,
avoidance of large differences in element size reduces solution stiffness, i.e.,
makes the numerical procedure more efficient.
Job: The Job module allows to create a job, to submit it to ABAQUS/Standard or
ABAQUS/Explicit for analysis and to monitor its progress.
Visualization: The Visualization module provides graphical display of finite ele-
ment models and results. It extracts the model and result information from the
output database.

3.6 Case Studies in ABAQUS Platform

This section demonstrates the efficiency of finite element models to replicate the
actual phenomena occurring during the cutting process under varied conditions as
well as to understand basic mechanism of chip formation process in terms of vari-
ous numerical results. Two different work materials are considered, namely AISI
1050 and Ti6Al4V, the former producing continuous chips and the latter produc-
ing segmented chips. In the first case study, not only the simulation results for ma-
chining AISI 1050 are presented but the predicted results are confirmed with the
experimental ones. In the second case study, Ti6Al4V is considered for the study
and the predicted results are compared with those of AISI 1050 under similar con-
ditions, thus showing the effect of work material, being one of the machining in-
puts, on various output variables. Apart from machining inputs, being a numerical
model, it is obvious that various FE inputs such as mesh size, material models,
friction models and so on also have a typical impact on the output results. Hence,
effect of FE inputs, namely, mesh size and Johnson-Cook material model con-
stants have also been studied while simulating segmented chip formation during
machining Ti6Al4V.
Finite Element Modeling of Chip Formation in Orthogonal Machining 131

3.6.1 Case Study I: FE Simulation of Continuous Chip

Formation While Machining AISI 1050 and Its
Experimental Validation
AISI 1050 is a standard grade carbon steel. This particular grade of steel is both
relatively easy to machine as well simulate. Therefore, AISI 1050 was chosen as a
reference material in the present study. The basic aim of the present section is to
develop the FE model for simulating the continuous chip formation process during
machining of AISI 1050 as well as to explain the basic mechanism governing the
continuous chip formation process in terms of numerical results viz. distributions
stress, strain and temperature. The simulated results are then validated with the
experimental ones at varied cutting conditions. Table 3.1 and Table 3.2 show the
thermo-mechanical properties of the workpiece material as well as the values of
Johnson-Cook material and damage constants used as input, respectively. The
tool-chip interaction is based on the commonly used Coulomb’s law with the
mean coefficient of friction 0.2.

Table 3.1. Thermo-mechanical properties of tungsten carbide tool and AISI 1050 [72, 73]

Properties Carbide tool AISI 1050

Density, ρ (Kg/m3) 11900 8030
Elastic modulus, E (GPa) 534 210
Poission’s ratio,μ 0.22 0.3
Specific heat, Cp (J/KgoC) 400 472
Thermal conductivity, λ (W/moC) 50 51.9
Expansion, (μm/moC) - 13.7
Tmelt (oC) - 1460
Troom (oC) - 25
Fracture toughness,Kc (MPa m½) - 60
Conductive heat transfer coefficient h, (kW m–2 K–1) - 10

Table 3.2. Johnson-Cook material and damage constants for AISI 1050 [72]

Material Constants

A (MPa) B (MPa) n C m

880 500 0.234 0.013 1.00

Damage Constants

D1 D2 D3 D4 D5

0.06 3.31 –1.96 0.018 0.58

132 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

3.6.1.1 Simulation Results

Figure 3.9 shows the distributions of stress, strain and temperature while machin-
ing of AISI 1050 using tungsten carbide tool for a cutting speed of 120 m/min and
feed of 0.2 mm/rev.

(a) (b) (c)

Fig. 3.9. Distribution of (a) stress (b) strain and (c) temperature while machining of AISI
1050 at Vc =120 m/min and f =0.2 mm

As the tool touches the workpiece, compression occurs within the workpiece.
With the further advancement of the tool into the workpiece, stresses start develop-
ing near the tool tip and attaining high localized values in a confined region called
primary shear deformation zone or shear plane as shown in Fig. 3.9 (a). The stresses
in these regions are as high as 1.3 GPa. Consequently, such high values of stresses
cause high strains to occur in the shear zone. This allows the workpiece material to
deform plastically and shear continuously in the form of chip that flows over tool
rake face. The type of chip, thus, formed depends largely on the distribution of strain
and temperature within the chip surface. As the cutting continues the effective
strains (especially, around the tool tip) increase and spread over a wider area of the
chip surface with a maximum value not exceeding 1.7 in the shear plane. Conse-
quently, this phenomenon tends to make the temperature distribution uniform in the
chip region, thereby resulting in a steady state continuous chip having unvarying
chip thickness. It is interesting to note that the maximum equivalent plastic strain
and temperature are found along the tool–chip interface.
This can be attributed to the fact that the chip which is entering into the second-
ary shear deformation zone already possesses accumulated plastic strain and heat.
The instant it begins to flow over the rake surface, further plastic straining and lo-
cal heating occur because of severe contact and friction in the contact zone [74],
thus, attaining higher temperature values at the tool-chip interface, specifically in
the sliding region.

3.6.1.2 Model Verification

In order to validate the developed model, both the cutting speed and feed or uncut
chip thickness values are varied and their effect on predicted cutting forces is
Finite Element Modeling of Chip Formation in Orthogonal Machining 133

studied and compared with the experimental ones. The cutting speed is varied in
the range of 72–164 m/min for feed values of 0.1 and 0.2 mm.
It is known that the cutting force and thrust force increase with increasing feed
rate almost linearly [75] where as decrease with the increasing cutting velocity.
The reason for this can be explained from the expressions of the cutting force and
thrust forces which are given as follows:

Cutting force, Fc = tsτ s (ζ − tan γ + 1) (3.38)

Thrust force, Ft = tsτ s (ζ − tan γ − 1) (3.39)

where t is the depth of cut (mm), s is the feed rate (mm/rev), τ s is the dynamic
shear strength of the workpiece, γ is the rake angle and ζ is the chip reduction
coefficient, i.e., the ratio of deformed chip thickness to undeformed chip thick-
ness. From the equations, correlation between feed rate and forces is straightfor-
ward. As far as the variation of cutting velocity is concerned, as the former in-
creases, temperature of the shear zone increases. This causes softening of the work
piece, which means the value of τ s decreases and thereby reduces the value of
cutting and thrust force.
Predicted cutting forces also show similar kind of trend with variation of speed
and feed as shown in Fig. 3.10. Cutting forces measured during experimental tests
under conditions similar to that of simulations have also been presented in Fig. 3.10
for comparison. It is noted that at lower feed, predicted results closely match with
experimental ones. At higher feed, a flatter curve is observed for predicted values of
forces implying that there is, no doubt, a decrease in cutting force values with the in-
creasing speed but not as pronounced as in case of the experimental one. However,
the maximum deviation between predicted and measured values of cutting forces
remains within an acceptable range of 15%.

1400
Fc(exp)
Fc(sim)
1200

Feed=0.2 mm
1000
Force (N)

800

600 Feed=0.1 mm

400

60 80 100 120 140 160 180

Cutting velocity (m/min)

Fig. 3.10. Predicted and measured forces while machining AISI 1050
134 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

3.6.2 Case Study II: FE Simulation of Segmented Chip

Formation While Machining Ti6Al4V
The present study deals with the simulation of chip formation while machining
Ti6Al4V. It is a known fact that machining of Ti6Al4V yields segmented chips.
Formation of segmented chips makes not only the machining of Ti6Al4V chal-
lenging but also simulating the chip formation process of the same very difficult.
In this study, simulation of segmented chip formation is performed based on adia-
batic shearing phenomenon. Thermal softening becomes highly pronounced when
the rate of heat generation exceeds the rate of heat dissipation which comes into
picture when thermal conductivity of the workpiece is very low and cutting speed
is high. During cutting process, high stresses and strains along the primary shear
zone accumulate and plastic deformation occurs, thereby increasing cutting tem-
perature. As the tool proceeds further, temperature increases in a highly localized
area in the shear zone in the form of bands called adiabatic shear bands that extend
from tool tip toward the chip free surface. It has been found that temperature at the
back of the chip reaches such a high value that thermal softening occurs in a high-
ly localized region on the free side of the chip surface. This induces wave like ir-
regularities on the back of the chip.
Firstly, an attempt is made to simulate the desired chip by fine tuning two of
the FE inputs that seem to be critical as far as reliability of the model is concerned.
Then a comparative study is done between the predicted results of AISI 1050 and
Ti6Al4V to explain the significance of simulating the right kind of chip formation
and its correlation with other output variables.

3.6.2.1 Role of FE Inputs

Meshing: Selection of a suitable mesh size is a critical factor from both accuracy
and computational time points of view [76]. As discussed earlier, finer mesh leads
to greater accuracy but at the cost of higher computational time. It is important to
mesh the model in such a way that the model gives results closer to the experi-
mental ones on one hand and consumes fairly less time on the other hand. Moreo-
ver, it is illogical to start with a very fine mesh, instead a mesh refinement study
should be performed wherein the degree of meshing is gradually changed from
coarser mesh to a finer mesh and the corresponding results are compared among
each other. There exists a limit beyond which if the mesh is refined further, the
CPU time would, no doubt, increase but there would not be any significant
changes in the numerical results. It is the duty of the analyst to consider an opti-
mum meshing that comes as a fair compromise between accuracy and computa-
tional time. This shows the need of mesh refinement study to prove the reliability
of the developed model.
Researchers have pointed that with the decrease in element size, the tempera-
ture at the integration point increases. Since this study primarily deals with the
mechanism of adiabatic shearing during the formation of segmented chip, it is
Finite Element Modeling of Chip Formation in Orthogonal Machining 135

logical to carry out a mesh refinement study. The meshing is mainly varied in the
chip region by considering 5, 10 and 15 elements (on an average) in the chip
thickness direction as shown in Fig. 3.11. The average element size for 5, 10 and
15 elements are: 50x50 µm, 20x20 µm and 12x12 µm, respectively. Since advanc-
ing front algorithm (free mesh algorithm) is employed for meshing the workpiece,
slight skewness has been observed at certain mesh regions of the workpiece with
finer meshing at the right edge of the workpiece.

(a) (b) (c)

Fig. 3.11. Mesh configurations for (a) 5 (b) 10 and (c) 15 elements

Figure 3.12 shows the predicted chip morphology and the distribution of tem-
perature within the chip surface for all the three cases for a cutting speed of 210
m/min and uncut chip thickness 0.2 mm.

(a) (b) (c)

Fig. 3.12. Temperature distributions for (a) 5 (b) 10 and (c) 15 elements

It is interesting to note that there is a gross difference in the predicted chip

morphologies for the first two cases. The difference can be attributed to the varia-
tion in stress, strain and temperature distributions within the chip region. Though
the maximum value of stress in both the cases is nearly same but the distribution
136 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

pattern varies. As the meshing gets finer, the stresses become more concentrated
along the shear plane, as a result of which the strain and temperature not only
get localized in a very narrow zone but also attain very high values. Since an in-
crease in the maximum temperature is found with finer meshing, the tendency to
invoke chip segmentation due to thermal softening by adiabatic shearing be-
comes higher, thus producing segmented chips in the second case. However,
with further refinement of mesh as in the case of 15 elements, not much varia-
tion in chip morphology is observed when compared with that of 10 elements.
The adiabatic shear band, no doubt, appears relatively more distinct (as very thin
strip) in the former case, but the average temperature values in the shear band
( Tadia ) increases not even by 1% (see Table 3.3). This implies that further mesh
refinement may not be required as far as consistency of the results is concerned.
Moreover, there is a limit to the reduction of element size from the software
point of view. Since the time step in ABAQUS/Explicit is controlled by the size
of the smallest element, reducing the element size beyond this may not allow the
simulation to run properly.

Table 3.3. Mesh refinement study

Mesh type Stress (GPa) Strain Tadia (°C)

5 elements 1.48 1.32 570

10 elements 1.50 2.0 956
15 elements 1.54 2.5 963

It is observed that finer meshing increases CPU time considerably. Therefore,

fixed mass scaling is used to expedite the analysis in the last two cases. Scaling
the material density of the small elements throughout the analysis proves to be
very useful as it increases the stable time increment allowing the analysis to take
less time, while retaining the accuracy to a great extent [77]. Table 3.4 presents a
comparative study of the actual CPU time with respect to simulation time for 5, 10
and 15 elements.

Table 3.4. CPU time for 5, 10 and 15 elements

Simulation time CPU time (h:min:s)

(ms) 5 elements 10 elements 15 elements
0.4 0:09:53 1:05:20 1:50:02
0.6 0:21:37 1:44:55 ---
1 0:35:07 2:55:52 ---
Finite Element Modeling of Chip Formation in Orthogonal Machining 137

From Table 3.4, it can be inferred that the model considering 10 elements along
the uncut chip thickness is the best compromise between accuracy and computa-
tion time. In case of 15 elements, though adiabatic shearing is very prominent, it
was very difficult to run the simulation beyond 0.4 ms.
Johnson–Cook material model constants: Several classical plasticity models have
been widely employed that represent, with varying degrees of accuracy, the material
flow stresses as a function of strain, strain rate and temperature. These models
include the Litonski-Batra model [78, 79], Usui model [80], Maekawa model [81],
the Johnson–Cook model [65], Zerilli–Armstrong model [82], Oxley model [83],
Mechanical Threshold Stress (MTS) model [84] etc. It is very important to carefully
select the appropriate material model that satisfactorily predicts the desired chip
morphology and other output variables. Johnson-Cook model, being most widely
used, is employed to describe the flow stress property of workpiece material
Ti6Al4V in this study (see Eq. 3.17). This material model defines the flow stress as
a function of strain, strain-rate and temperature such that, it not only considers the
strain rates over a large range but also temperature changes due to thermal softening
by large plastic deformation. The work material constants contained in this constitu-
tive equation have been found by various researchers by applications of several me-
thods, thus producing different values of data sets for a specific material. As a result,
selection of suitable data sets along with appropriate material model becomes equal-
ly important [85]. The present study selects two sets of Johnson-Cook material con-
stants from the available literature namely, M1 and M2 as listed in Table 3.4 [86,
87]. Lee and Lin [87] obtained the Johnson–Cook material constants from a high
strain rate mechanical testing using Split Hopkinson Pressure Bar (SHPB) method
under a constant strain rate of 2000 s-1 within the temperature range of 700–1000 ºC
and maximum true plastic strain of 0.3. While Meyer and Kleponis [86] obtained the
material constants by considering strain rate levels of 0.0001, 0.1 and 2150 s-1and a
maximum plastic strain of 0.57.

Table 3.5. Johnson-Cook material parameters for Ti6Al4V [86, 87]

Johnson-Cook constants A (MPa) B (MPa) C n m

M1 model 862.5 331.2 0.012 0.34 0.8
M2 model 782.7 498.4 0.028 0.28 1.0

Besides the material model constants, all other parameters are kept constant
so that the results can be compared on the same conditions. Meshing in both the
cases is similar to the one that was considered as optimum in the previous case
study.
Figures 3.13 and 3.14 show the distribution of stress, strain and temperature for
the models M1 and M2 at a cutting speed of 210 m/min and uncut chip thickness
of 0.2 mm.
138 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

(a) (b) (c)

Fig. 3.13. Distribution of (a) Stress (Pa) (b) Strain and (c) Temperature (ºC) for M1 model

(a) (b) (c)

Fig. 3.14. Distribution of (a) Stress (Pa) (b) Strain and (c) Temperature (ºC) for M2 model

It is observed that model M1 is unable to simulate the formation of segmented

chips. The distribution of stress, strain and temperature when compared with that
of model M2 explains much about the variation in chip morphology. This suggests
that material model constants used in the constitutive equations also affect the
predicted results. Thermal softening coefficient m and work hardening exponent n
are often considered as the critical parameters that influence the segmentation of
chip [88]. Baker [88] suggested increase in m leads to increase in chip segmenta-
tion while increase in n causes lesser tendency to undergo chip segmentation. In
the present study, similar kind of observation has been found. Greater value of m
and lower value of n can be attributed for the formation of well-defined segmented
chips in case of model M2.

3.6.2.2 Significance of FE Outputs

Various difficult to measure output variables such as stresses, strains, temperatures

and many more can be determined easily from the FE simulations of chip forma-
tion process. These predicted outputs greatly help in understanding the basic me-
chanism of chip formation under different cutting conditions. Chip morphology,
considered as an important index in study of machining, is dependent on various
input variables like workpiece material, cutting conditions, tool material and so
on. The type of chip produced affects the cutting force, the residual stresses and
the cutting temperatures to a great extent. This section, thus, demonstrates the
Finite Element Modeling of Chip Formation in Orthogonal Machining 139

same by a comparative study of predicted forces and cutting tool temperature

while machining AISI 1050 and Ti6Al4V.

Cutting forces: Figure 3.15 shows the variation of predicted cutting force with
time while machining AISI 1050 and Ti6Al4V.

500
Cutting force (N)

400

300

200
Ti6Al4V
AISI 1050
100

0.0000 0.0002 0.0004 0.0006 0.0008 0.0010 0.0012

Time (s)

Fig. 3.15. Time signature of cutting forces for Ti6Al4V and AISI 1050

The force signatures obtained for both the materials vary remarkably not only
in magnitude but also in nature from each other. As expected, the cutting force
profile closely resembles the predicted chip morphology i.e. continuous for AISI
1050 and wavy profile for Ti6Al4V. While machining Ti6Al4V, chip formation
process begins with the bulging of the workpiece material in front of the tool as a
result of which the forces increase gradually but drop suddenly when the shear
band begins to form due to thermal softening. Obviously, magnitude of the wavi-
ness (difference between the peak and lower values) affects the surface finish
quality and thermal load that the cutting tool undergoes during cutting process.

Cutting temperature: In cutting operation, cutting tool is subjected to highly adverse

conditions such as high stresses and high temperatures due to which tool wears out.
As the tool wear occurs, tool geometry changes affecting residual stresses, surface
finish and dimensional accuracy of the machined surface. Hence, cutting tempera-
ture is a critical parameter which is again a function of various machining inputs
such as cutting conditions, workpiece material, tool material, tool geometry, etc.
Tool condition monitoring and optimization of process parameters are of paramount
importance that can lead to higher productivity and accuracy.
Figures 3.16 (a) and (b) present the temperature distributions over the cutting tool
that has developed during machining of AISI 1050 and Ti6Al4V, respectively.
140 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

(a) (b)
Fig. 3.16. Temperature contours on cutting tool while machining (a) AISI 1050 and
(b) Ti6Al4V

While machining Ti6Al4V the maximum temperature attained by carbide tool

is as high as 1000ºC as compared to that of AISI 1050 with maximum value not
even exceeding 400ºC. It can be seen that not only the maximum value of temper-
ature attained by cutting tool is different but also the distribution pattern over the
cutting tool varies widely depending upon the type of worpiece material ma-
chined. This gives a basic idea of how the tool wear progress takes place. Fur-
thermore, the local variables such as temperature, stress, strain, etc., at different
nodes that are predicted prove helpful in developing a tool wear prediction method
by integrating FEM simulation of the cutting process with a suitable tool wear rate
model.
In general, it can be concluded that FEM can prove to be a very effective tool in
analyzing various aspects of metal cutting process, if used with caution. The
stresses, strains, temperatures and many other outputs that are obtained from FE
simulations greatly help in understanding the basic mechanism of chip formation
under different cutting conditions. This consequently helps in the cutting tool de-
sign, optimization of cutting parameters for higher tool life and good surface
finish, prediction of tool wear growth, etc., thus leading to higher productivity.

References
[1] Astakhov, V.P.: Tribology of Metal Cutting. Elsevier (2006) ISBN: 978-0-444-52881-0
[2] Shaw, M.C.: Metal Cutting principles, 2nd edn. Oxford University Press, Oxford
(2005)
[3] Hahn, R.S.: On the temperature developed at the shear plane in the metal cutting
process. In: Proceedings of First U.S. National Congress Appl. Mech. ASME 661
(1951)
[4] Chao, B.T., Trigger, K.J.: An analytical evaluation of metal cutting temperature.
Trans. ASME 73, 57–68 (1951)
[5] Leone, W.C.: Distribution of shear-zone heat in metal cutting. Trans. ASME 76, 121–
125 (1954)
Finite Element Modeling of Chip Formation in Orthogonal Machining 141

[6] Loewen, E.G., Shaw, M.C.: On the analysis of cutting tool temperatures. Transactions
of the ASME 71, 217–231 (1954)
[7] Weiner, J.H.: Shear plane temperature distribution in orthogonal machining. Trans.
ASME 77, 1331–1341 (1955)
[8] Rapier, A.C.: A theoretical investigation of the temperature distribution in the metal
cutting process. Br. J. Appl. Phys. 5, 400–405 (1954)
[9] Bagci, E.: 3-D numerical analysis of orthogonal cutting process via mesh-free me-
thod. Int. J. the Physical Sciences 6(6), 1267–1282 (2011)
[10] ASM Handbook, Volume 16-Machining, ASM International Handbook Committee,
ASM International, Electronic (1989) ISBN: 978-1-61503-145-0
[11] Juneja, B.L., Sekhon, G.S., Seth, N.: Fundamentals of metal cutting and machine
tools, 2nd edn. New Age International Publishers, New Delhi (2003)
[12] Merchant, M.E.: Mechanics of the metal cutting process. J. Appl. Phys. 16, 318–324
(1945)
[13] Lee, E.H., Shaffer, B.W.: The theory of plasticity applied to a problem of machining.
Trans. ASME, J. Appl. Mech. 18, 405–413 (1951)
[14] Oxley, P.L.B.: Shear angle solutions in orthogonal machining. Int. J. Mach. Tool.
Des. Res. 2, 219–229 (1962)
[15] Bhattacharya, A.: Metal cutting theory and practice. Central book publishers, Kolkata
(1984)
[16] Lacale, L.N., Guttierrez, A., Llorente, J.I., Sanchez, J.A., Aboniga, J.: Using high
pressure coolant in the drilling and turning of low machinability alloys. Int. J. of Adv.
Tech. 16, 85–91 (2000)
[17] Tobias, S.A.: Machine tool Vibration. Blackie and Sons Ltd, Scotland (1965)
[18] Ghosh, A., Mallik, A.K.: Manufacturing science (1985) ISBN: 81-85095-85-X
[19] Jacobson, S., Wallen, P.: A new classification system for dead zones in metal cutting.
Int. J. Mach. Tool. Manufact. 28, 529–538 (1988)
[20] Trent, E.M., Wright, P.K.: Metal cutting, 4th edn. Butterworth-Heinemann (2000)
[21] Komanduri, R., Hou, Z.B.: A review of the experimental techniques for the measure-
ment of heat and temperatures generated in some manufacturing processes and tribol-
ogy. Tribol. Int. 34, 653–682 (2001)
[22] Groover, M.P.: Fundamentals of modern manufacturing: materials processes, and sys-
tems, 2nd edn. Wiley, India (2002)
[23] Klamecki, B.E.: Incipient chip formation in metal cutting- A three dimensional finite
element analysis. Ph.D. Thesis. University of Illinois, Urbana (1973)
[24] Hughes, J.R.T.: The Finite Element Method. Prentice-Hall International, Inc. (1987)
[25] Reddy, J.N.: An introduction to the finite element method, 2nd edn. McGraw-Hill
Inc. (1993)
[26] Bathe, K.J.: Finite Element Procedures. Prentice Hall, Englewood Cliffs (1996)
[27] Rao, S.S.: The Finite Element in Engineering, 3rd edn. Butterworth-Heinemann
(1999)
[28] Zienkiewicz, O.C., Taylor, R.L.: The Finite Element Method, 5th edn. Butterworth-
Heinemann (2000)
[29] Liu, G.R., Quek, S.S.: The Finite Element Method: A Practical Course. Butterworth
Hienemann (2003)
[30] Hutton, D.V.: Fundamentals of finite element analysis, 1st edn. Mc Graw Hill (2004)
[31] Belytschko, T., Liu, W.K., Moran, B.: Nonlinear Finite Elements for Continua and
Structures. John Wiley and Sons, New York (2000)
142 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

[32] Strenkowski, J.S., Moon, K.-J.: Finite element prediction of chip geometry and
tool/workpiece temperature distributions in orthogonal metal cutting. ASME J. Eng.
Ind. 112, 313–318 (1990)
[33] Raczy, A., Elmadagli, M., Altenhof, W.J., Alpas, A.T.: An Eulerian finite-element
model for determination of deformation state of a copper subjected to orthogonal cut-
ting. Metall Mater. Trans. 35A, 2393–2400 (2004)
[34] Mackerle, J.: Finite element methods and material processing technology, an adden-
dum (1994–1996). Eng. Comp. 15, 616–690 (1962)
[35] Rakotomalala, R., Joyot, P., Touratier, M.: Arbitrary Lagrangian-Eulerian thermome-
chanical finite element model of material cutting. Comm. Numer. Meth. Eng. 9, 975–
987 (1993)
[36] Pepper, D.W., Heinrich, J.C.: The Finite Element Method: Basic Concepts and Appli-
cations. Hemisphere Publishing Corporation, United States of America (1992)
[37] Bower, A.F.: Applied mechanics of solid. CRC Press, Taylor and Francis Group,
New York (2010)
[38] Dhondt, G.: The finite element method for three-dimensional thermomechanical ap-
plications. John Wiley and Sons Inc., Germany (2004)
[39] Pian, T.H.H.: Derivation of element stiffness matrices by assumed stress distribu-
tions. AIAA J. 2, 1333–1336 (1964)
[40] Zienkiewicz, O.C., Taylor, R.L.: The Finite Element Method. Basic formulations and
linear problems, vol. 1. McGraw-Hill, London (1989)
[41] Liapis, S.: A review of error estimation and adaptivity in the boundary element me-
thod. Eng. Anal. Bound. Elem. 14, 315–323 (1994)
[42] Tay, A.O., Stevenson, M.G., de Vahl Davis, G.: Using the finite element method to
determine temperature distribution in orthogonal machining. Proc. Inst. Mech.
Eng. 188(55), 627–638 (1974)
[43] Muraka, P.D., Barrow, G., Hinduja, S.: Influence of the process variables on the tem-
perature distribution in orthogonal machining using the finite element method. Int. J.
Mech. Sci. 21(8), 445–456 (1979)
[44] Moriwaki, T., Sugimura, N., Luan, S.: Combined stress, material flow and heat analy-
sis of orthogonal micromachining of copper. CIRP Annals - Manufact. Tech. 42(1),
75–78 (1993)
[45] Kim, K.W., Sin, H.C.: Development of a thermo-viscoplastic cutting model using fi-
nite element method. Int. J. Mach. Tool Manufact. 36(3), 379–397 (1996)
[46] Liu, C.R., Guo, Y.B.: Finite element analysis of the effect of sequential cuts and tool-
chip friction on residual stresses in a machined layer. Int. J. Mech. Sci. 42(6), 1069–
1086 (2000)
[47] Ceretti, E., Falbohmer, P., Wu, W.T., Altan, T.: Application of 2D FEM to chip for-
mation in orthogonal cutting. J. Mater Process Tech. 59, 169–180 (1996)
[48] Li, K., Gao, X.-L., Sutherland, J.W.: Finite element simulation of the orthogonal met-
al cutting process for qualitative understanding of the effects of crater wear on the
chip formation. J. Mater Process Tech. 127, 309–324 (2002)
[49] Arrazola, P.J., Ugarte, D., Montoya, J., Villar, A., Marya, S.: Finite element modeling
of chip formation process with abaqus/explicit. VII Int. Conference Comp., Barcelona
(2005)
[50] Davies, M.A., Cao, Q., Cooke, A.L., Ivester, R.: On the measurement and prediction
of temperature fields in machining AISI 1045 steel. Annals of the CIRP 52, 77–80
(2003)
Finite Element Modeling of Chip Formation in Orthogonal Machining 143

[51] Adibi-Sedeh, A.H., Vaziri, M., Pednekar, V., Madhavan, V., Ivester, R.: Investigation
of the effect of using different material models on finite element simulations of metal
cutting. In: 8th CIRP Int Workshop Modeling Mach Operations, Chemnitz, Germany
(2005)
[52] Shi, J., Liu, C.R.: The influence of material models on finite element simulation of
machining. J. Manufact. Sci. Eng. 126, 849–857 (2004)
[53] Ozel, T.: Influence of Friction Models on Finite Element Simulations of Machining.
Int. J. Mach. Tool Manufact. 46(5), 518–530 (2006)
[54] Filice, L., Micari, F., Rizzuti, S., Umbrello, D.: A critical analysis on the friction
modeling in orthogonal machining. International Journal of Machine Tools and Man-
ufacture 47, 709–714 (2007)
[55] Haglund, A.J., Kishawy, H.A., Rogers, R.J.: An exploration of friction models for the
chip-tool interface using an Arbitrary Lagrangian-Eulerian finite element model.
Wear 265(3-4), 452–460 (2008)
[56] Mabrouki, T., Deshayes, L., Ivester, R., Regal, J.-F., Jurrens, K.: Material modeling
and experimental study of serrated chip morphology. In: Proceedings of 7th CIRP Int
Workshop Model. Machin, France, April 4-5 (2004)
[57] Coelho, R.T., Ng, E.-G., Elbestawi, M.A.: Tool wear when turning AISI 4340 with
coated PCBN tools using finishing cutting conditions. J. Mach. Tool Manufact. 47,
263–272 (2006)
[58] Lorentzon, J., Jarvstrat, N.: Modelling tool wear in cemented carbide machining alloy
718. J. Mach. Tool Manufact. 48, 1072–1080 (2008)
[59] Davim, J.P., Maranhao, C., Jackson, M.J., Cabral, G., Gracio, J.: FEM analysis in
high speed machining of aluminium alloy (Al7075-0) using polycrystalline diamond
(PCD) and cemented carbide (K10) cutting tools. Int. J. Adv. Manufact. Tech. 39,
1093–1100 (2008)
[60] Attanasio, A., Cerretti, E., Rizzuti, S., Umbrello, D., Micari, F.: 3D finite element
analysis of tool wear in machining. CIRP Annals – Manufact. Tech. 57, 61–64 (2008)
[61] ABAQUS Analysis User’s manual. Version 6.7-4 Hibbitt, Karlsson & Sorensen, Inc.
(2007)
[62] ABAQUS Theory manual, Version 6.7-4 Hibbitt, Karlsson & Sorenson, Inc. (2007)
[63] ABAQUS/CAE User’s manual. Version 6.7-4 Hibbitt, Karlsson & Sorensen, Inc.
(2007)
[64] Wu, H.-C.: Continuum Mechanics and Plasticity. Chapman and Hall/CRC (2004)
[65] Johnson, G.R., Cook, W.H.: A constitutive model and data for metals subjected to
large strains, high strain rates and high temperatures. In: Proceedings of 7th Int Symp
Ballistics, the Hague, The Netherlands, pp. 541–547 (1983)
[66] Umbrello, D., M’Saoubi, R., Outeiro, J.C.: The influence of Johnson–Cook material
constants on finite element simulation of machining of AISI 316L steel. Int. J. Mach.
Tool Manufact. 47, 462–470 (2007)
[67] Davim, J.P., Maranhao, C.: A study of plastic strain and plastic strain rate in machin-
ing of steel AISI 1045 using FEM analysis. Mater Des. 30, 160–165 (2009)
[68] Vaziri, M.R., Salimi, M., Mashayekhi, M.: A new calibration method for ductile frac-
ture models as chip separation criteria in machining. Simulat Model Pract. Theor. 18,
1286–1296 (2010)
[69] Johnson, G.R., Cook, W.H.: Fracture characteristics of three metals subjected to vari-
ous strains, strains rates, temperatures and pressures. Eng. Fract. Mech. 21(1), 31–48
(1985)
144 A. Priyadarshini, S.K. Pal, and A.K. Samantaray

[70] Mabrouki, T., Girardin, F., Asad, M., Regal, J.-F.: Numerical and experimental study
of dry cutting for an aeronautic aluminium alloy. Int. J. Mach. Tool Manufact. 48,
1187–1197 (2008)
[71] Mabrouki, T., Rigal, J.: -F A contribution to a qualitative understanding of thermo-
mechanical effects during chip formation in hard turning. J. Mater. Process
Tech. 176, 214–221 (2006)
[72] Duan, C.Z., Dou, T., Cai, Y.J., Li, Y.Y.: Finite element simulation & experiment of
chip formation process during high speed machining of AISI 1045 hardened steel. Int.
J. Recent Trend Eng. 1(5), 46–50 (2009)
[73] Priyadarshini, A., Pal, S.K., Samantaray, A.K.: A Finite Element Study of Chip For-
mation Process in Orthogonal Machining. Int . J. Manufact., Mater. Mech. Eng. IGI
Global( accepted, in Press, 2011)
[74] Shi, G., Deng, X., Shet, C.: A finite element study of the effect of friction in ortho-
gonal metal cutting. Finite Elem. Anal. Des. 38, 863–883 (2002)
[75] Lima, J.G., Avila, R.F., Abrao, A.M., Faustino, M., Davim, J.P.: Hard turning: AISI
4340 high strength alloy steel and AISI D2 cold work tool steel. J. Mater. Process.
Tech. 169, 388–395 (2005)
[76] Priyadarshini, A., Pal, S.K., Samantaray, A.K.: Finite element study of serrated chip
formation and temperature distribution in orthogonal machining. J. Mechatron Intell.
Manufact. 2(1-2), 53–72 (2010)
[77] Wang, M., Yang, H., Sun, Z.-C., Guo, L.-G.: Dynamic explicit FE modeling of hot
ring rolling process. Trans. Nonferrous Met. Soc. China 16(6), 1274–1280 (2006)
[78] Litonski, J.: Plastic flow of a tube under adiabatic torsion. Bulletin of Academy of
Pol. Science, Ser. Sci. Tech. XXV, 7 (1977)
[79] Batra, R.C.: Steady state penetration of thermo-visoplastic targets. Comput Mech. 3,
1–12 (1988)
[80] Usui, E., Shirakashi, T.: Mechanics of machining–from descriptive to predictive
theory: On the art of cutting metals-75 Years Later. ASME PED 7, 13–55 (1982)
[81] Maekawa, K., Shirakashi, T., Usui, E.: Flow stress of low carbon steel at high tem-
perature and strain rate (Part 2)–Flow stress under variable temperature and variable
strain rate. Bulletin Japan Soc Precision Eng 17, 167–172 (1983)
[82] Zerilli, F.J., Armstrong, R.W.: Dislocation-mechanics-based constitutive relations for
material dynamics calculations. J. Appl. Phys. 61, 1816–1825 (1987)
[83] Oxley, P.L.B.: The mechanics of machining: An analytical approach to assessing ma-
chinability. Ellis Horwood Limited, Chichester (1989)
[84] Banerjee, B.: The mechanical threshold stress model for various tempers of AISI
4340 steel. Int. J. Solid Struct. 44, 834–859 (2007)
[85] Priyadarshini, A., Pal, S.K., Samantaray, A.K.: On the Influence of the Material and
Friction Models on Simulation of Chip Formation Process. J. Mach. Forming Tech.
Nova Science (accepted, 2011)
[86] Meyer, H.W., Kleponis, D.S.: Modeling the high strain rate behavior of titanium un-
dergoing ballistic impact and penetration. Int. J. Impact Eng. 26(1-10), 509–521
(2001)
[87] Lee, W.S., Lin, C.F.: High temperature deformation behaviour of Ti6Al4V alloy ev-
luated by high strain rate compression tests. J. Mater. Process. Tech. 75, 127–136
(1998)
[88] Baker, M.: The influence of plastic properties on chip formation. Comp. Mater.
Sci. 28, 556–562 (2003)
4

GA-Fuzzy Approaches: Application to Modeling

of Manufacturing Process

Arup Kumar Nandi

Central Mechanical Engineering Research Institute (CSIR-CMERI)

Durgapur-713209, West Bengal, India
[email protected], [email protected]

This chapter presents various techniques using the combination of fuzzy logic and
genetic algorithm (GA) to construct model of a physical process including manu-
facturing process. First, an overview on the fundamentals of fuzzy logic and fuzzy
inferences systems toward formulating a rule-based model (called fuzzy rule based
model, FRBM) is presented. After that, the working principle of a GA is discussed
and later, how GA can be combined with fuzzy logic to design the optimal know-
ledge base of FRBM of a process is presented. Results of few case studies of
modeling various manufacturing processes using GA-fuzzy approaches conducted
by the author are presented.

4.1 Introduction
Optimal selection of machining parameters is an imperative issue to obtain a better
performance of machining, cost effectiveness as well as to achieve a desired accu-
racy of the attributes of size, shape and surface roughness of the finished product.
Selection of these parameters is traditionally carried out on the basis of the expe-
rience of process planners with the help of past data available in machining hand-
books and tool catalogs. Practitioners continue to experience great difficulties due
to the lack of sufficient data on the numerous new cutting tools with different ma-
terials. Specific data on relevant machining performance measures such as tool
life, surface roughness, chip form, etc. are very difficult to find due to the lack of
reliable information or predictive models for these measures. In automated manu-
facturing processes, it is required to control the machining process by determining
the optimum values of machining parameters online during machining. Therefore,
it is important to develop a technique to predict the attribute of a product before
machining to evaluate the robustness of machining parameters for keeping a de-
sired attribute and increasing product quality. Construction of suitable machining
process model and evaluation of the optimal values of machining parameters using
this model as predictor are essential and challenging tasks.
146 A.K. Nandi

The model of a machining process represents a mapping of input and output va-
riables in specific machining conditions. The input variables differ corresponding
to the type of machining process and the desired output. For example, in turning,
the surface roughness (output variable) is dependent on a number of variables that
can be broadly divided into four groups: major variables which include cutting
speed, feed rate, depth of cut and tool wear; flow of coolant, utilization of chip
breaker, work-holding devices and selection of tool type belong to the second
group. The third group includes machine repeatability, machine vibration and
damping, cutting temperature, chip formation and chip exit speed, thermal expan-
sion of machine tool and power consumption; room temperature, humidity, dust
content in the air and fluctuation in the power source are involved in the fourth
group. Among these four groups of input variable, the major variable can be
measured and controlled during machining process. Though the other variables are
not directly involved and uncontrolled during machining but their effect cannot be
neglected to obtain a desired surface roughness. In a specific machining condition,
these variables are assumed to be fixed at a particular state.
Various approaches have been proposed to model and simulate the machining
processes. Analytical methods, which are generally based on the established
science principles, are probably the first modeling approach. Experimental or em-
pirical approaches use experimental data as the basis for formulating the models.
Mechanistic and numerical methods integrate the analytical and empirical me-
thods, generally by the use of modern computer techniques.
Due to the complex and nonlinear relationship among the input–output variables,
influence of uncontrollable parameters and involvement of random aspect, predic-
tion of an output of machining processes using mathematical/analytical approaches
is not accurate. This leads to the development of empirical equations for a particular
machine tool, machining parameters and work piece-cutting tool material combina-
tion. Empirical models do not consider the underlying principles and mathematical
relationships. These are usually obtained by performing statistical analysis or
through the training of data-driven models to fit the experimental data [1].
The significant drawback of empirical models is their sensitivity to process var-
iation though they have the advantages of accuracy due to the use of experimental
data. The accuracy of a model degenerates rapidly due to the variation of experi-
mental data as the machining conditions deviate from the experimental settings. In
addition, quality characteristics of machined parts exhibit stochastic variations
over time due to changes in a machine tool structure and the environment. There-
fore, the modeling techniques should have enough capability to adapt the varia-
tions in machining process. Most of the statistical process control models do not
account for time-varying changes. Involvement of uncertainty and imprecision in
machining processes is another aspect affecting the variation of machining output.
In such cases, techniques of modeling using fuzzy logic are most useful because
fuzzy logic is a powerful tool for dealing with imprecision and uncertainty [2].
The basic concept of fuzzy logic is to categorize the variables into fuzzy sets with
a degree of certainty in the numerical interval (0 and 1), so that ambiguity and
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 147

vagueness in the data structure and human knowledge can be handled without
constructing complex mathematical models. Moreover, fuzzy logic-based control
system has the capability to adapt the variations of a process by learning and ad-
justing itself to the environmental changes by observing current system behavior.
Fuzzy logic is an application of fuzzy set theory, and was first proposed by
Prof. L.A. Zadeh [3]. Fuzzy logic rules, which are derived based on fuzzy set
theory, are used in fuzzy inference system toward formulating a rule-based model
(called fuzzy rule-based model, FRBM). The performance of a FRBM mainly de-
pends on two different aspects: structure of fuzzy logic rules and the type/shape of
associated fuzzy subsets (membership function distributions, MFDs) those consti-
tute the knowledge base (KB) of FRBM. Manually constructed KB of a FRBM
may not be optimal in many cases since it strongly demands a thorough know-
ledge of the process which is difficult to acquire, particularly in a short period of
time. Therefore, design of an optimal KB of a fuzzy model needs the help of other
optimization/learning techniques. Genetic algorithm (GA), a population-based
search and optimization technique is used by many researchers to design the op-
timal KB of FRBM for various processes. The systems of combining Fuzzy logic
and genetic algorithm are called genetic-fuzzy systems.

4.2 Fuzzy Logic

4.2.1 Crisp Set and Fuzzy Set
A set (A) is a collection of any objects (a1, a2, a3, ….., an), which according to
some law can be considered as a whole and it is usually written as
A={a1, a2,…..,an} or
A={x |P(x)}, means A is the set of all elements (x) of universal set X for which
the proposition P(x) is true (e.g. P(x) > 3).
In crisp set, the function XA(x) (so-called characteristic function) assigns a val-
ue of either 1 or 0 to each individual object, x in the universal set, thereby discri-
minating between members and non-members of the crisp set, A under considera-
tion. That means there exists no uncertainty or vagueness in the fact that the object
belongs to the set or does not belong to the set. Set A is defined by its characteris-
tic function, X A (x ) as follows
⎧1 for x ∈ A
X A (x) = ⎨
⎩0 for x ∉ A
In the year, 1965, Lotfi A. Zadeh [3] proposed a completely new and elegant
approach to vagueness and uncertainty in a seminal paper, called fuzzy set theory.
In his approach an element, x can belong to a set to a degree, k (0 ≤ k ≤ 1) in con-
trast to classical set theory where an element must definitely belong or not to a set.
A fuzzy set, Ã is usually written as
Ã={x, μÃ(x) |x ∈X}
148 A.K. Nandi

The function μà (x) is called membership function (MF) of the (fuzzy) set Ã
and defined as μÃ(x)→[0,1]. The value of μÃ(x) is called the degree of member-
ship of x in the fuzzy set Ã.

4.2.2 Fuzzy Membership Function

Graphically, a membership function is represented as a curve (as shown in Figure 4.1)
that defines how each element in the set is mapped to a membership value (or degree
of membership) between 0 and 1. There are many ways to assign membership values
or functions to fuzzy variables compared to that of assigning probability density func-
tion to random variables. The membership function assignment process can be intui-
tive or based on some algorithmic or logical operations.

Fig. 4.1. A graphical representation of fuzzy set

4.2.2.1 Some Key Properties of Fuzzy Set

i) Having two fuzzy sets à and B̃ based on X, then both are equal if
their membership functions are equal, i.e.
~ ~
( ) ( )
A = B ⇔ μ A x = μ B x for all x X
~ ~ ∈
ii) ∈
Given a fuzzy set à defined on X and any number α [0,1], the α-
cut, αÃ, and the strong α-cut, α+Ã, are the crisp sets:
α
{ }
A = x | A(x ) ≥ α and
~ ~ α+
{ ~
}
A = x | A(x ) > α
~
iii) The height of a fuzzy set is the largest membership grade obtained
~
()
by any element in that set i.e., height A = max μ A~ (x )
x∈X
iv) The crossover points of a membership function are defined as the
elements in the universe for which a particular fuzzy set A has values
equal to 0.5, i.e., for which μ A~ (x ) = 0.5
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 149

v) A fuzzy set à is called normal when ( )

~
height A = 1 and subnor-
~
mal when height A < 1 . ()
vi) The support of a fuzzy set à is the crisp set that contains all the ele-
ments of X that have non-zero membership grades, i.e.
( ) { }
support A = x ∈ X | μ ~ (x ) > 0 , refer to Figure 4.1.
~
A
vii) The core of a normal fuzzy set à is the crisp set that contains all the
elements of X that have the membership grades of one in Ã, i.e.
()
core A = {x ∈ X | μ A~ (x ) = 1}, refer to Figure 4.1.
~
viii) The boundary is the crisp set that contains all the elements of X that
have the membership grades of 0 < μ A~ (x ) < 1 in Ã, i.e.
()
boundary A = {x ∈ X | 0 < μ A~ (x ) < 1} , refer to Figure 1.
~
ix) Having two fuzzy sets à and B̃ based on X, then both are similar if
~
( ) ~
() ~ ~
core A = core B and support A = support B . ( ) ()
x) If the support of a normal fuzzy set consists of a single element xo of
(~ ) (~ )
X, which has the property support A = core A = {x0}, this set is
called a singleton.
xi) A fuzzy set à is said to be a convex fuzzy set if for any elements x, y
and z in fuzzy set Ã, the relation x<y<z implies that
[ ]
μ A~ (y ) ≥ min μ A~ (x ), μ A~ (z ) . The intersection of two convex
fuzzy sets is also a convex fuzzy set, i.e., if à and B̃ are convex
~ ~
fuzzy sets, then A ∩ B is also convex fuzzy set.
xii) If à is convex single-point normal fuzzy set defined on the real line,
then à is often termed as a fuzzy number.
xiii) Any fuzzy set à defined on a universe X is a subset of that universe.

4.2.2.2 Various Types of Fuzzy Membership Function and Its Mathematical

Representation

The common types of membership function (MF) used in FRBM are triangular,
(higher order) polynomial, trapezoidal, Gaussian, etc.
Trapezoidal MF: Mathematically a trapezoidal MF can be represented as shown in
Figure 4.2(a).
0 x < a, x > d
x−a
a≤x≤b
μ A~ (x, a, b, c, d ) = b − a for
1 b<x<c
d−x
c≤x≤d
d−c
150 A.K. Nandi

The controlling parameters toward the configuration of trapezoidal MF (as

shown in Figure 4.2(a)) are b1=b-a, b2=c-b and b3=d-c.

Polynomial MF: A polynomial MF can be expressed mathematically as shown in

Figure 4.2(b):

0 x < a, x > c
f 1 (x, b1) a≤x<b
μ A~ (x, a, b, c, d ) = for
1 x=b
f 2 (x, b 2) b<x≤c

where the functions f1 and f2 are the polynomial type. Polynomial MF is treated as
triangular type MF when the functions, f1 and f2 of the above empirical expression
are linear. f1 and f2 may be also exponential or any other kind of functions.
The controlling parameters toward the configuration of polynomial MF (as
shown in Figure 4.2(b)) are b1=b-a and b2=c-b. Mathematically, the second order
μ
polynomial function can be represented as Ã(x)= c0x + c1x2, where x is the
distance measured along the base-width of membership function distributions,
μÃ(x) is the fuzzy membership function value and, c0 and c1 are the coefficients
which can be determined based on some specified conditions, such as

⎧1 at x = b1 ∂ μ A~
μ A~ = ⎨ and = 0, at x = b1,
⎩ 0 at x = 0 ∂x

(a) (b)

Fig. 4.2. Membership function configuration (a) Trapezoidal (b) Polynomial

2
Finally the coefficients of the 2nd order polynomial function become c0 =
b1
1
and c1 = − 2
.
b1
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 151

From Figure 4.2, it has been seen that for a given value of support of the mem-
bership function of a fuzzy set, only one parameter, b1 is required to describe tri-
angular and polynomial type MFDs (membership function distributions) with 3
fuzzy sub-sets, whereas two parameters, b1 and b2 are required to explain the
(semi) trapezoidal MFDs with two fuzzy subsets. The number of controlling pa-
rameter increases with increasing the number of fuzzy sub-sets involved in the
MFDs.

4.2.3 Fuzzy Set Operators

Defining the fuzzy sets à and B̃ on the universe X, for a given element x of the
universe, the fuzzy set operations, intersection (t-norm), union (t-conorm) and
complement are expressed as follows and the corresponding Venn diagrams are
shown in Figure 4.3.

Intersection: μ A~ ∩ B~ (x ) = μ A~ (x ) ∧ μ B~ (x )
Union: μ A~ ∪ B~ (x ) = μ A~ (x ) ∨ μ B~ (x )
Complement: μ A~ (x ) = 1 − μ A~ (x )

Fig. 4.3. (a) Intersection of fuzzy sets à and B̃ (b) Union of fuzzy sets à and B̃ (c) Com-
plement of fuzzy set Ã

4.2.4 Classical Logical Operations and Fuzzy Logical Operations

Classical logic deals with classical proposition (P) which is a collection of elements,
that is, a set, where all the truth values, T(P) for all elements in the set are either all true
(1) or all false (0), and follows the two-valued logical operations and Boolean algebra.
Let the sets A and B are defined from universe X and the proposition, P measures
the truth of the statement that an element, x from the universe X is contained in set,
A and Q measures the truth of the statement that this element, x is contained in set,
152 A.K. Nandi

if x ∈ A, T(P ) = 1; otherwiase T(P ) = 0 and if x ∈ B, T(Q ) = 1;

B, i.e.,
otherwise T(Q ) = 0 . There are five logical connectives (defined as follows) used
to combine multiple simple propositions and to form new propositions:

Disjunction (OR):
P ∨ Q : x ∈ A or x ∈ B, Hence, T(P ∨ Q ) = max (T(P ), T (Q ))
Conjunction (AND):
P ∧ Q : x ∈ A or x ∈ B, Hence, T(P ∧ Q ) = min (T (P ), T(Q ))
Negation: If T (P ) = 1, then T (P ) = 0; if T(P ) = 0, then T (P ) = 1
Implication: P → Q : x ∉ A or x ∈ B, Hence, T (P → Q ) = T (P ∪ Q )
1 for T (P ) = T (Q )
(P ↔ Q ) : T(P ↔ Q ) = ⎧⎨
⎩0, for T (P ) ≠ T(Q )
Equivalence:

For two different universes of discourse where P is a proposition described by

set A, which is defined on universe X, and Q is a proposition described by set B,
which is defined on universe Y. Then the implication P→Q (which is also equiva-
lent to the linguistic rule form, IF A THEN B) can be represented in set-theoretic
terms by the relation, R which is defined by

( )
R = (A × B) ∪ A × Y ≡ IF A, THEN B
IF x ∈ A, where x ∈ X and A ⊂ X
THEN y ∈ B, where y ∈ Y and B ⊂ Y
The other connectives are applicable to two different universes of discourse as
usual. Classical logical compound propositions that are always true irrespective of
the truth values of the individual simple propositions are called tautologies.
Fuzzy propositional logic generalizes the classical propositional operations by us-
ing the truth set [0, 1] instead of either 1 or 0. The above logical connectives are also
defined for a fuzzy logic. Like classical logic, the implication connective in fuzzy
logic can be modeled in rule-based form: P̃→Q̃ is, IF x is à THEN y is B̃ (where IF

(~ )
part is called antecedent and THEN part is called consequent) and it is equivalent to
~
(~ ~ )
the fuzzy relation R = A × B ∪ A × Y where the fuzzy proposition P̃ is as-
signed to fuzzy set à which is defined on universe X, and the fuzzy proposition Q̃ is
described by fuzzy set B̃, which is defined on universe Y. The membership function
[( )( )]
of R̃ is expressed by μ R~ (x, y ) = max μ A~ (x ) ∧ μ B~ (y ) , 1 − μ A~ (x ) . The im-
plication connective can be defined in several distinct forms. While these forms of
implication are equivalent in classical logic, their extensions to fuzzy logic are not
equivalent and result in distinct classes of fuzzy implications.
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 153

4.2.5 Fuzzy Implication Methods

The fuzzy implication operation is used to find the fuzzy relation R̃ between
two fuzzy sets à and B̃ which are defined on the universes of discourse X and Y,
∈
respectively based on the rule IF x is Ã, THEN y is B̃, where x X and y Y. Ma- ∈
[ ]
thematically, the fuzzy relation, R̃ is defined as R (x, y ) = ℘ A (x ), B(y ) , where
~

℘ is called the implication operator. Besides the implication method as presented

in Section 4.2.4, there are other forms of implication operator, among them min
and product implication operators are mostly used in fuzzy inference system for
practical applications. The membership function values of fuzzy relation R̃ de-
fined on the Cartesian product space X × Y using min and product implication
operators are obtained by the Equation (4.1) and Equation (4.2), respectively and
graphically represented in Figure 4.4.

min: μ (x, y ) = min[μ (x ), μ (y )]

~
R
~
A
~
B
(4.1)

product: μ (x, y) = μ (x ) • μ (y )
~
R
~
A
~
B
(4.2)

In Figure 4.4, the MF value, 0.7 of μB(̃ x) corresponds to rule weight obtained after
decomposition of the IF part of rule.

(i) (ii)

Fig. 4.4. Graphical representation of fuzzy implication (i) min (ii) product

4.2.6 Decomposition of Compound Rules

The most common techniques for decomposition of compound linguistic rules into
simple canonical forms are described as follows:
154 A.K. Nandi

Multiple conjunctive antecedents: IF x is Ã1 AND x is Ã2 …. AND x is ÃL THEN

y is B̃S
This fuzzy rule can be written in canonical form as: IF ÃS THEN B̃S,
where the fuzzy subset ÃS is defined as A ~ = ~ ∩ ~ ∩ ...... ∩ ~ ) with the
A1 A2 AL
[ ]
S
membership function μ ~ (x ) = min μ ~ (x ), μ ~ (x ),..........., μ ~ (x ) obtained by
A S A1 A2 AL
the definition of the fuzzy intersection operation.
Multiple disjunctive antecedents: IF x is Ã1 OR x is Ã2 …. OR x is ÃL THEN y is B̃S
Assuming a new fuzzy subset ÃS (as A ~ = ~ ∪ ~ ∪ ...... ∪ ~ ) expressed by
A1 A 2 AL
[ ]
S
means of membership function μ ~ (x ) = max μ ~ (x ), μ ~ (x ),..........., μ ~ (x )
AS A1 A2 AL
based on the definition of the fuzzy union operation, the compound rule may be
written in canonical form as: IF ÃS THEN B̃S.

4.2.7 Aggregation of Rule

The technique of obtaining the overall rule consequent by combining the individu-
al consequents contributed by each rule in the rule base (which comprises multiple
rules) is known as aggregation of rules. The most popular aggregation techniques
of fuzzy rules are as follows:
Conjunctive system of rules (MIN): In the case of a system of rules that must be
jointly satisfied, the rules are connected by AND connectives. In this case, the ag-
gregated consequent, y is obtained by fuzzy intersection of all individual rule con-
sequents (y1, y2, …, yr), as y = y1 ∩ y 2 ∩ ...... ∩ y r which is defined by the
membership function: [ ] ∈
μ y (y ) = min μ y1 (y ), μ y2 (y ),..........., μ yr (y ) for y Y.
Disjunctive system of rules (MAX): For the case of a disjunctive system of rules
where the satisfaction of at least one rule is required, the rules are connected by the
OR connectives. In this case, the aggregated consequent, y is obtained by fuzzy un-
ion of the individual rule consequents as y = y1 ∪ y 2 ∪ ...... ∪ y r which is de-
fined by the membership function: [ ]
μ y (y ) = max μ y1 (y ), μ y2 (y ),..........., μ yr (y )
∈
for y Y.

4.2.8 Composition Technique of Fuzzy Relation

Let R and S be the relations that relate elements from universe X to universe Y, and
elements from universe Y to universe Z, respectively. Now, composition is an op-
eration to find another relation, T that relates the same elements in universe X
that R contains to the same elements in universe Z that S contains. The composition
operation of fuzzy relation reflects the inference of a fuzzy rule-based system and is
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 155

~ ~ ~
expressed by B = A R , where à is the input, or antecedent defined on the un-
iverse X, B̃ is the output or consequent defined on universe Y and R̃ is a fuzzy rela-
tion characterizing the relationship between specific input(s), x and specific out-
put(s), y. Among various methods of composition of fuzzy relation, max-min and
max-product are the most commonly used techniques and defined by membership
function-theoretic expressions as follows.
max-min: μ B~ (y ) = max {min [μ A~ (x ), μ R~ (x, y )]} (4.3)
x∈X

max-product: μ B~ (y ) = max μ A
x∈X
[
~ (x ) • μ ~ (x, y )
R ] (4.4)

The method of composition of fuzzy relation basically includes the implication

method and technique of aggregation of fuzzy rule. In the above methods of com-
position of fuzzy relation, max is the aggregation technique of rule, whereas min
and product are the implication methods used in Equation (4.3) and Equation (4.4),
respectively.

4.2.9 Fuzzy Inferences

Inference is a process of combining the measurement of input(s)/antecedent(s)
with one or more relevant fuzzy rules in a proper manner to infer the out-
put(s)/consequent(s). In order to demonstrate the inference method, consider a sys-
tem of one input (antecedent) and a single output (antecedent) described by two
IF-THEN rules as follows:
Rule 1: if input1 is Ã1, then output1 is B̃1
Rule 1: if input1 is Ã2, then output1 is B̃2
Now given the IF-THEN rules (Rule 1 and Rule 2) and a fact/measurement “In-
∈ ∈
put1 is Ô it is inferred that “output1 is B̃”, where Ã, Ã1 ₣( ) and B̃, B̃1 ₣( ),
where ₣( ) and ₣( ) denote the sets of all ordinary fuzzy sets that can be defined
within the X and Y, respectively. and are the sets of values (x and y) of va-
riables, input1 (condition variable) and output1 (action variable), respectively. In
order to determine B, method of interpolation is used which consists of two steps
explained as follows:
Step 1: Calculate the degree of consistence, rj(x) between the given
fact/measurement and the antecedent of each rule, j in terms of the height of inter-
section of the associated sets Ã1 and Ã. rj(x) is expressed using the standard fuzzy
intersection and the definition of height of a fuzzy set ( described in Section
4.2.2.1) as follows:

⎡ ⎤
r (x ) = max min ⎢⎣μ (x ), μ (x )⎥⎦ , j=1, 2
j x∈ X
~
A ~
Aj
(4.5)
156 A.K. Nandi

Step 2: Calculate the conclusion B̃ by truncating each set B̃j by the value of rj(x)
(i.e., min implication method) which expresses the degree to which the antecedent
Ãj is compatible with the given fact Ã1, and taking the union of the truncated sets
as the rules are satisfied independently (i.e., max aggregation method) .

[ ]
μ B~ (y ) = max min r j (x ), μ B~ j (y ) for all y Y
j =1,2
∈ (4.6)

The above steps are graphically presented in Figure 4.5.

Now, μ (y ) = max min ⎡⎢⎣r (x ), μ (y)⎤⎥⎦

~
B j =1,2 j ~
Bj

⎡
= max min max min ⎜
j =1,2 ⎢⎣ x∈X μ ~ μ A~ j (x )⎞⎟⎠, μ B~j (y )⎤⎥⎦
⎛ (x ),
⎝ A

⎡
= max max min ⎜
⎢
j =1,2 x∈X ⎣
μ ~ μ A~ j (x )⎞⎟⎠, μ B~j (y )⎤⎥⎦
⎛ (x ),
⎝ A

⎡ ⎛ ⎞⎤
= max max min ⎢ μ ~ (x ), min ⎜ μ ~ ( x ), μ ~ ( y ) ⎟⎥
x∈X y =1, 2 ⎣ A ⎝ Aj Bj ⎠⎦
⎡ ⎛ ⎞⎤
= max min ⎢ μ ~ (x ), max min⎜ μ ~ ( x ), μ ~ ( y ) ⎟⎥
x∈X ⎣ A j =1,2 ⎝ Aj Bj ⎠⎦

= max
x∈X
[
min μ ~ (x ), μ ~ (x, y ) ,
A R
] (4.7)

which is equivalent to the expression of max-min composition presented in Equa-

tion (4.3) and accordingly, this inference method is also called max-min inference
method. In the above step 2, if product implication technique and max aggrega-
tion method are used, we can obtain the expression (defined in Equation 4.7) simi-
lar to max-product composition (Equation (4.4)). Moreover, the above inference
method may also be applicable to the system with any number of fuzzy rule (j)
and inputs (Input1, input2). In case of multiple inputs/antecedents of fuzzy rule,
the (effective) degree of consistence, rj(x) between the given facts/measurements
and the antecedents of each rule is obtained by first finding the degree of consis-
tence between each fact/measurement and the related antecedent of the rule (using
Equation (4.5)), then adopting the technique of decomposition of compound rules
according to the type of logical connectives (AND or OR) as explained graphical-
ly in Figure 4.6 for the following two IF-THEN rules as follows:
Rule 1: if input1 is Ã11 AND input2 is Ã12 then output1 is B1
Rule 1: if input1 is Ã21 AND input2 is Ã22 then output1 is B2
Fact: input1 is Ã1 AND input2 is Ã2
Conclusion: output1 is B
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 157

Fig. 4.5. Illustration of the method of interpolation in fuzzy inferences

Fig. 4.6. Illustration of the method of interpolation in fuzzy inferences with multiple inputs
158 A.K. Nandi

In the above illustration of fuzzy inference, we have considered fuzzy value (Ã)
for the variable input1. Figure 4.7 demonstrates the above (max-min) inference
method for a two input and a single output system where the values of input va-
riables are considered as crisp type (for instance, Fact: input1 is x1 AND input2 is
x2) and the (max-product) inference method for the same system is demonstrated
in Figure 4.8.

Fig. 4.7. Graphical representation of max-min inference method with crisp type of input
values

Fig. 4.8. Graphical representation of max-product inference method with crisp type of input
values
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 159

4.2.10 Fuzzification and De-fuzzification

The fuzzification is a process of transforming the crisp value into a grade of mem-
bership using a membership function of the associated fuzzy set as shown in Fig-
ure 4.9. Figure 4.9 demonstrates that the given (crisp) value x0 of variable x be-
longs to a grade of μ ~ (x 0 ) = 0.8 to the fuzzy set Ã1 and with a grade of
A1
μ ~ (x 0 ) = 0.2 to the fuzzy set Ã2. Fuzzification is required in the fuzzy inference
A2
system when the values of input variables to system are considered as crisp type
(as described in Figure 4.7/Figure 4.8).

Fig. 4.9. Fuzzification method

Defuzzification is the conversion of a fuzzy quantity to a crisp quantity analog-

ous to fuzzification. Defuzzification is used in fuzzy inference system to convert
the fuzzy value of the output (i.e., B̃ of output1 in Section 4.2.9) to a crisp value
(*y). Among various defuzzification methods available in the literature, centroid
method (also called center of area (COA) or center of gravity) is most popular,
which is mathematically expressed by Equation (4.8) and graphically expressed in
Figure 4.10:

y =
∫ μ (y )⋅ ydy
~
B
(4.8)
COA
∫ μ (y )dy
~
B
160 A.K. Nandi

Fig. 4.10. Centroid defuzzification method

4.2.11 Fuzzy Rule-Based Model

4.2.11.1 Working Principle of a Fuzzy Rule-Based Model

A fuzzy rule base, a fuzzy inference engine, fuzzification and de-fuzzification:

these are the four modules involved in a FRBM (Fuzzy rule-based model). Figure
4.11 shows a schematic diagram explaining the working cycle of a FRBM. The
following steps are involved in the working cycle of a FRBM:
• The output(s) (action) and input(s) (condition) variables needed to control a
particular process are chosen and measurements are taken for all the condition
variables.
• The measurements taken in the previous steps are converted into appropriate
fuzzy sets to express measurement uncertainties (called fuzzification as de-
scribed in Section 4.2.10).
• The fuzzified measurements are then used by the interference engine to eva-
luate the control rules stored in the rule base and a fuzzified output is deter-
mined (as discussed in Section 4.2.9).
• The fuzzified output is then converted into a single (crisp) value (called a de-
fuzzification as illustrated in Section 4.2.10). The de-fuzzified value(s)
represent action(s)/prediction(s) to be made by the FRBM in controlling a
process.
The kernel of a FRBM that is the knowledge base (KB) is constituted by rule
base, RB (a set of fuzzy logic rules) and membership functions/membership func-
tion distributions, MFDs (fuzzy subsets). Two types of fuzzy logic rules (FLRs)
that are commonly used for constructing the RB of a FRBM are Mamdani-type
and TSK-type. In both types of FLRs, the input variables are expressed by linguis-
tic terms (fuzzy subsets) in the rule antecedent part. But the main difference be-
tween these two types of fuzzy rules lies in the rule consequent part. The output
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 161

variable in Mamdani-type FLR is defined by linguistic term also, whereas in TSK-

type FLR, it is not defined by linguistic term rather it is defined by a linear combi-
nation of the input variables. The shape of fuzzy subsets (MFDs of input-output
variables) is also an important factor that is to be decided appropriately to achieve
the best performance of a FRBM for a typical process.

Fig. 4.11. A schematic showing the working cycle of a FRBM

4.2.11.2 Various Types of Fuzzy Rule-Based Model

Mamdani-type [4]:
The structure of Mamdani-type fuzzy logic rule is expressed as follows:
IF x1 is A1 AND x2 is A2 AND……..AND xn is An THEN y is B
where xi (i=1, 2, ……, n) are input variables and y is the output variable. A1, A2,
…, An and B are the linguistic terms (say, Low, Medium, High, etc.) used for the
fuzzy subsets (membership function distributions) of the corresponding input and
output variables, respectively.
Sugeno-type [5]:
The Sugeno-type fuzzy rule is defined as follows:
IF x1 is A1 AND x2 is A2 AND……..AND xn is An THEN y =f(x1, x2, .., xn)
Unlike Mamdani-type, the rule consequent/output is expressed by a function of the
input variables.
Tsukamoto [6]:
The Tsukamoto -type fuzzy rule is defined as follows:
IF x1 is A1 AND x2 is A2 AND……..AND xn is An THEN y =z
where z is a monotonical membership function.
162 A.K. Nandi

4.2.11.2.1 Mamdani-Type Fuzzy Rule Based Model

The output of a Mamdani-type FRBM whose rule base (RB) is constructed using
Mamdani-type fuzzy logic rule is obtained as follows (when centroid method is
considered for defuzzification):
R lf
∑ A αr C A αr
r =1
Y= (4.9)
R lf
∑A αr
r =1

where A αr is the area of fuzzy subset of output variable, covered by α membership

value (equivalent to degree of consistence as obtained in Step 1 in Section 4.2.9) that
is obtained by rth rule after fuzzy inference method. C Aαr is the center distance of the
area, A αr . R lf ( R lf ⊆ R f ) is the number of rules fired out of a total of Rf rules
present in the rule base for a set of input values.
In order to determine the output in Equation 4.9, all the four modules (fuzzy
rule base, a fuzzy inference engine, fuzzification and de-fuzzification) as men-
tioned in Section 4.2.11.1 are involved in a Mamdani-type FRBM. The perfor-
mance of a Mamdani-type fuzzy model is relied on the appropriate fuzzy subsets
of rule consequents and antecedent, and the type of fuzzy subsets (membership
function distributions) considered for input-output variables. Therefore, the tasks
of designing FRBMs with Mamdani-type FLRs are:
• construction of an optimal set of rules (Rf) with its appropriate outputs (B),
• selection of shape of fuzzy subsets/MFDs for both the input and output va-
riables
• tuning of MFDs.

4.2.11.2.2 TSK-Type Fuzzy Rule-Based Model

The TSK-type fuzzy logic rule is defined as follows [5, 7]:
If x1 is A1 and x2 is A2 and………………….and xn is An, then
y = ∑ c j f j (x 1,..., n )
K

j =1

where A1, . . . , An are the fuzzy subsets of the respective input variables, x1, …, xn.
The output function of fuzzy rule is a linear function (say, polynomial) in the form
of

y = ∑ c j f j (x 1,..., n ) = c1 f 1 (x 1,.., n ) + c 2 f 2 (x 1,.., n ) + ..... + c k f K (x 1,..., n )

K
(4.10)
j =1

The overall output of the TSK-type fuzzy model can be obtained for a set of inputs
(x1, x2, …., xn) using the following empirical expression.
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 163

( )
R lf
⎛ n
⎞
∑ ⎜ ∏ μ v (x v )⎟ ∑ c j f j x 1,..., n
K
r r
r =1⎝ v =1 ⎠ j=1
Y= (4.11)
⎛ n r ( ⎞ )
R lf
∑ ⎜ ∏ μ v x 1,.., n ⎟
r =1⎝ v =1 ⎠
∏ is the product representing a conjunction decomposition method.

∑
K

j =1
r
cjf
r
j (x ) is the output function of r
1,.., n
th
rule and r
c j are the function coeffi-
cients of the corresponding rule consequent, where K is the number of coefficients
present in the consequent function of each rule.
Unlike Mamdani-type FRBM, TSK-type FRBM includes only the fuzzy rule
base, a fuzzy inference engine, and fuzzification module to determine the output in
Equation (4.11). The performance of a TSK-type fuzzy model is mainly depended
on the optimal values of the rule output (consequent) functions which are de-
pended on the coefficients (cj), the exponential parameters of the input variables
(not shown in the Equation (4.10)) and choice of the fuzzy subsets (membership
function distributions). Thus, the steps of developing FRBM with TSK-type FLRs
are:
• construction of an optimal set of rules (Rf) with the appropriate structures of
rule output/consequent functions
• selection of shapes of fuzzy subsets/MFDs of input variables
• determination of optimal values of coefficients and power terms of rule conse-
quent functions
• tuning of MFDs of the input variables.

4.3 Genetic Algorithm

Genetic algorithm is a search and optimization technique which mimics the prin-
ciple of natural selection and natural genetics [8] to find the best solution for a
specific problem. The genetic algorithm is an approach to solve problems which
are not yet fully characterized or too complex to allow full characterization, but
for which some analytical evaluation or physical interpretation to evolutes the per-
formance of a solution, is available. GA is a stochastic global search method that
mimics the metaphor of natural biological evolution. Genetic algorithms operate
on a population of feasible solutions by applying the principle of survival of the
fittest to produce better approximations to a solution. At each generation, a new
set of approximations is created by the process of selecting individuals according
to their level of fitness in the problem domain and breeding them together using
operators borrowed from natural genetics. This process leads to the evolution of
populations of individuals that are better suited to their environment than the indi-
viduals that they were created from, just as in natural adaptation.
164 A.K. Nandi

The basic concept of a genetic algorithm is to encode a potential solution to a

problem as a series of parameter strings, chromosomes, composed over some al-
phabet, so that chromosome values (genotypic) are uniquely mapped onto the de-
cision variable (phenotypic). A single set of parameter string or chromosome is
treated as the genetic material of an individual solution. Initially a large population
of candidate solutions is created with random parameter values. These solutions
are essentially bred with each other for several simulated generations under the
principle of survival of the fittest, meaning that the probability that an individual
solution will pass on some of its parameter values to subsequent children is direct-
ly related to the fitness of individual i.e. how good that solution is relative to the
others in the population.
Breeding takes place through use of recombination operators such as crossover,
which simulates basic biological cross-fertilization, and mutation, essentially the
introduction of noise. The simple application of these operators with a reasonable
selection mechanism has produced surprisingly good results over a wide range of
problems. After recombination and mutation, the individual strings are then, if ne-
cessary, decoded, the objective function evaluated, a fitness value assigned to each
individual and individuals selected for mating according to their fitness, and so the
process continues through subsequent generations. In this way, the average per-
formance of individuals in a population is expected to increase, as good individu-
als are preserved and bred with one another and the less fit individuals die out.
The GA is terminated when some criteria are satisfied, e.g. a certain number of
generations, a mean deviation in the population, or when a particular point in the
search space is encountered.

4.3.1 Genetic Algorithms and the Traditional Methods

The working principle of GA clearly indicates that the GA significantly differs in
some very fundamental way from other traditional search and optimization me-
thods. The four major significant differences are:

• GAs search a population of solutions instead of a single point solution as in tra-

ditional search methods.
• GAs do not use derivative-based algorithm. It does not use any derivative in-
formation or other auxiliary knowledge; only the objective function and corres-
ponding fitness levels influence the directions of search.
• GAs use probabilistic transition rules instead of deterministic principle.
It is important to note that the GA provides a number of potential solutions to a
given problem and the choice of final solution is left to the user. In cases where a
particular problem does not have one individual solution, for example a family of
Pareto-optimal solutions, as is the case in multi-objective optimization and sche-
duling problems, then the GA is potentially useful for identifying these alternative
solutions simultaneously.
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 165

4.3.2 Simple Genetic Algorithm

The schematic shown in Figure 4.12 illustrates the structure of a simple genetic
algorithm (SGA) as described by Goldberg [8]. GA starts with initial random pop-
ulation consisted of potential solution points called individuals. The decision is
made whether the individual is good or bad for the given problem based on the
fitness obtained from the evaluation of objective function. Once the fitness value
is evaluated and assigned to each individual, then initial population meets the first
genetic operator, selection process. This operator provides more chances of sur-
vival for the strong individuals and to decay the weakest ones according to their
fitness.

Fig. 4.12. A schematic representation of simple genetic algorithm outline population repre-
sentation and initialization

Next, crossover operator is performed on selected individuals to build the new

individuals by combining the existing ones. Crossover follows reproduction and
allows two individuals to swap their structures depending on the probability fac-
tor. This result in the creation of a pair of offspring solution containing characte-
ristics of their parents. Then the mutation operator is applied to supply diversity in
the population. As the fitness of a population may remain static for a number of
generations before a superior individual is found, the application of conventional
termination criteria becomes problematic. A common practice is to terminate the
GA after a pre-defined number of generations and then test the quality of the best
members of the population against the problem definition. If no acceptable solu-
tions are found, the GA may be restarted or a fresh search initiated with more
number of generations.
166 A.K. Nandi

4.3.2.1 Coding and Initialization of GA

4.3.2.1.1 Binary Coding

The most commonly used representation of chromosomes in the GA is that of the
single-level binary string. Here, each decision variable (say, d and h) in the para-
meter set is encoded as a binary string (sd and sh, respectively) and these are con-
catenated to form a chromosome as shown in Figure 4.13. Binary-coded GAs are
not restricted to use only integer and for the given lower bound (dmin) and upper
bound (dmax) of a variable (say, d), the value of the variable (d) is calculated from
the GA-string using the decoding scheme represented by Equation (4.12).
max
− d min
DV(sd )
d
d = d min + (4.12)
2ld − 1
where ld is the string length used to code the d variable and DV(sd) is the decoded
value of the string sd . This mapping function allows

Fig. 4.13. A schematic representation of a chromosome with 5 bits for Sd and 5 bits for Sh

4.3.2.1.2 Real Coding

For a continuous search space, binary-coded GA faces many problems such as
• Hamming cliffs associated with certain strings (e.g., 01111 and 10000) from
which a transition to a neighboring solution requires the alteration of many bits.
• Inability to achieve any arbitrary precision in the optimal solution. The more the
required precision, the larger is the string length, results in more computational
complexity.
In real coding, the variables are directly represented in real type as 25 8 .
h d

4.3.2.1.3 Initialization of GA
Initial population of a GA is normally determined at random. With a binary popu-
lation of Nind individuals whose chromosomes are Lind bits long, Nind × Lind ran-
dom numbers uniformly distributed from the set {0, 1} would be produced.

4.3.2.2 Objective Function and Fitness Function

The objective function is used to provide an analytical measure of how individuals

have performed in the problem domain. The objective function, f is a function of
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 167

decision variable(s). The fitness value of an individual/solution in the population

is determined based on the fitness function which consists of the objective func-
tion value and the penalty value for constraint violation which is determined by a
penalty function, fcostraint. Thus, the fitness value is calculated as follows:

Fitness value= f +fcostraint ,

Since GAs mimic the survival-of-the-fittest of the nature to make a search

process, GAs are suitable for maximization problem where objective function is
directly used as fitness function F(x). In the case of a minimization problem, the
fit individuals will have the lowest numerical value of the associated objective
function. This situation is handled using a conversion of the objective function in-
to an equivalent maximization problem and used as fitness function so that the op-
timum point remain unchanged.

4.3.2.3 Selection

Selection guides the tool to find the optimized solution by preferring individu-
als/members of the population with higher fitness over one with lower fitness. It is
the operator which generates the mating pool. This operator determines that the
number of times a particular individual will be used for reproduction and the
number of offspring that an individual will produce. Some of the popularly used
selection methods are as follows:

Roulette Wheel Selection Methods: Roulette wheel selection scheme chooses a

certain individual with a probability proportional to its fitness.

( )
[ ]
p I j, t =
f I j, t
(Ik,t )
(4.13)
∑ nk =1 f

[ ]
where p I j, t is the probability of getting selected of any jth individual at a genera-
( )
tion t, f I j, t and ∑ nk =1 f (Ik, t ) are corresponding individual fitness and the sum of
the fitness of the population with size n, respectively.
The property as represented by Equation (4.13) is satisfied by applying a ran-
dom experiment that has some similarity with a generalized roulette game. In the
roulette game, the slots are not equally wide that is, why different outcomes occur
with different probabilities. Figure 4.14 gives a graphical representation of how
this roulette wheel game works.
Linear rank selection: In this plan, a small group of individuals is taken from the
population and the individual with best fitness is chosen for reproduction. The size
of the group chosen is called the tournament size. A tournament size of two is
called binary tournament.
168 A.K. Nandi

Fig. 4.14. Graphical representation of the Roulette wheel selection mechanism

In addition another scheme for selection is applied along with all three selection
schemes discussed above which is called ‘elitism’. The idea of elitism is to avoid
the observed best-fitted individual dies out just by selecting it for the next genera-
tion without any random experiment. Elitism significantly influences the speed of
the convergence of a GA. But it can lead to premature convergence also.

4.3.2.4 Crossover (Recombination)

The basic operator for producing new chromosomes in the GA is that of crossov-
er. Like natural reproduction, crossover produces new individuals so that some
genes of a new child come from one individual while others come from the other
individual. In essence, crossover is the exchange of genes between the chromo-
somes of the two parents. The process may be described as cutting two strings at
a randomly chosen position and swapping the two tails. It is known as the single-
point crossover, and the mechanism is visualized in Figure 4.15. An integer posi-
tion, i is selected at random with a uniform probability between one and the
∈
string length, l, minus one (i.e., i [1, l-1]). When the genetic information is ex-
changed among the parent individuals (represented by the strings, P1 and P2)
about this point, two new offspring (represented by the strings, O1 and O2) are
produced. The two offspring in Figure 4.15 are produced when the crossover
point, i=4 is selected.

Fig. 4.15. A typical example of single point crossover

GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 169

For multi-point crossover, multiple crossover positions (m) are chosen at ran-
dom with no duplicates and sorted into ascending order. Then the bits between
two successive crossover points are exchanged between the two parents to pro-
duce two new offspring. The process of multi-point crossover is illustrated in Fig-
ure 4.16 with shaded color.

Fig. 4.16. Typical example of multi-point crossover (with m=5)

The idea behind multi-point crossover is that the parts of the chromosome re-
presentation that contributes to the most to the performance of a particular indi-
vidual may not necessarily be contained in adjacent substrings. Further, multi-
point crossover appears to encourage the exploration of the search space, thus
making the search more robust.

4.3.2.5 Mutation

Mutation is nothing but deformation of the genetic information of an individual

(solution) by means of some external influences. The bit-wise mutation operator
changes a bit, 1 to 0, and vice versa, with a prescribed probability (called, muta-
tion probability) as shown in Figure 4.17. In real reproduction, the probability that
a certain gene is mutated is almost equal for all genes. So, it is near at hand to use
the mutation technique for a given binary string, where there is a given probability
that a single gene is modified. The probability should be rather low in order to
avoid chaotic behavior of the GA.

Fig. 4.17. Mutation effect on offspring’s strings

4.3.2.6 Termination of the GA

Because the GA is a stochastic search method, it is difficult to formally specify

convergence criteria. As the fitness of a population may remain static for a number
of generations before a superior individual is found, the application of convention-
al termination criteria becomes problematic. A common practice is to terminate
the GA after a pre-specified number of generations and then test the quality of the
170 A.K. Nandi

best members of the population against the problem definition. If no acceptable

solutions are found, the GA may be restarted or a fresh search can be initiated. At-
taining a pre-specified fitness function value or when best fitness of the population
does not change appreciably over successive iterations may be also considered as
a termination criteria.

4.3.3 Description of Working Principle of GA

The working principle of a (binary-coded) GA is described here with an engineer-
ing optimization problem, namely designing a can [9] as shown in Figure 4.18.
The objective of the problem is to determine the optimum values of the diameter
(d, in cm) and height (h, in cm) of the can in order to minimize its cost, f(d, h)
subject to some constraints and the problem is defined as follows:
⎛ π d2 ⎞
Minimize f (d, h ) = c⎜⎜ + πdh ⎟⎟,
⎝ 2 ⎠

Subject to
π dh 2
g(d, h ) ≡ ≥ 300,
4
d min ≤ d ≤ d max
,
h min ≤ h ≤ h max
where c is the cost of the can material per square cm, which is taken as 0.005 and
the minimum and maximum values of d and h are taken as, dmin=hmin=0 and
dmax=hmax=31.
Therefore, in this problem the number of decision variables is two (d and h).
The population size of GA is considered as six and it is kept constant throughout
the GA-operation. GA operates in a number of iterations until a specified termina-
tion criterion is satisfied and in the Figure 4.18, the maximum number of itera-
tion/generation (max_gen) is treated as the termination criteria.
GA iteration starts with the creation of six random solutions which are treated as
the parent solutions. The chromosome structure of each solution is the same as pre-
sented in Figure 4.13. Then, the fitness values of all the parent solutions in the popu-
lation are calculated using following fitness function (discussed in Section 4.3.2.2).

⎧f (d, h ) + P p × [300 - g(d, h)] if g(d, h ) ≤ 300

Fitness value = ⎨ (4.14)
⎩f (d, h ) if g(d, h ) > 300
where Pp is the penalty parameter and the value of Pp is taken as 25. After that se-
lection/reproduction (discussed in Section 4.3.2.3) operation is performed on the
parent solutions based on the fitness values of solutions (as depicted in the binary
tournament selection table in Figure 4.18) to form the mating pool. The number
of solutions contained in the mating pool is to be equal to the number of parent
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 171

Fig. 4.18. Working principle of a genetic algorithm

solution in order to maintain a constant population size through out the GA-iteration.
After that, crossover operator (discussed in Section 4.3.2.4) is applied among the
two randomly chosen solutions from the mating pool based on a given probability
(crossover-probability, say 0.9). Then, bit-wise mutation (discussed in Section
4.3.2.5) is carried out on each of the six solutions obtained after employing the cros-
sover operator using a given mutation probability (say 0.02) and produces six new
(offspring) solutions. The objective function values corresponding to each solution
are depicted in the single-point crossover and bit-wise mutation tables in Figure
4.18. It completes one iteration/generation of GA. Now, the checking of GA-
termination criterion is performed and if it satisfies the termination criteria, stop the
GA-iteration process, otherwise start another generation/iteration with treating the
172 A.K. Nandi

six offspring solutions obtained in the previous iteration as the parent solutions and
then assigning fitness value to all the solutions in the population and continue the
same iteration procedure as described above.

4.4 Genetic Fuzzy Approaches

The performance of a FRBM depends on its knowledge base (KB) which consists of
database (DB) (that is, information regarding membership functions) and rule base
(RB). It is important to mention that determination of an appropriate knowledge
base for a FRBM is not an easy task. The genetic algorithms (GAs) have been used
by several investigators to design database and/or rule base of a FRBM. The fuzzy
systems making use of a GA in their design process are called genetic-fuzzy systems
(GFS). Figure 4.19 shows the schematic diagram of a genetic-fuzzy system, in
which a GA-based learning/tuning based on example data (training data) is adopted,
off-line to design the KB a FRBM. GA can also be used to tune the existing KB of a
FRBM which may be designed based on the some common knowledge or expert in-
formation, to improve the performance of the existing FRBM. As the GA is found to
be computationally expensive due to the nature of population-based optimization,
the GA-based tuning is normally carried out off-line.

Fig. 4.19. Schematic layout of genetic-fuzzy system

A Mamdani-type fuzzy logic rule for a particular process having say, two input
variables (x1 and x2) and one output variable (y) (each having triangular-type
MFDs with 3 fuzzy subsets) may be expressed as
IF x1 is A1 AND x2 is A2 THEN y is B,
where A1, A2 and B are the fuzzy subsets (those can be expressed by suitable lin-
guistic expression, such as LOW, MEDIUM, HIGH, etc.) of triangular-type mem-
bership function. In Section 4.2.2.2, it was discussed that in order to describe tri-
angular-type MFDS with 3 fuzzy subsets one controlling parameter is required. A
typical binary coded GA-string for optimizing the KB will look as shown in
Figure 4.20.
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 173

Fig. 4.20. A GA-string representing the rule base and the parameters related to membership
functions of input-output variables a FRBM

where b1, b2 and b3 are the (continuous) control (GA) variables related to MFDs
corresponding to the two inputs and a single output variables x1, x2 and y, respec-
tively. The number of bits used for optimizing the RB is equal to the number of
maximum possible rules present in the RB. In this case, the number of rules will
be 3 × 3 = 9 , since each of the two input variables comprise of 3 fuzzy subsets.
The information of b1, b2 and b3 are coded by the next bits in GA-string.
There are, in fact, three different approaches of designing genetic-fuzzy system
(GFS), according to the KB components including in the learning process. These
are as follows

Genetic Learning /Tuning of the Fuzzy Logic Controller Data Base

Here, the GA is used to optimize the appropriate value(s) of the controlling para-
meter(s) that define the typical type of MFDs. In other words, for examples in case
of triangular type MFDs, it is used to move and to expand or shrink the base
width(s) (b1/b2/b3) of each interior isosceles triangle. The extreme triangles will be
right triangles and the GA will make it either bigger or smaller. During GA-based
optimization the parameters (b1, b2 and b3) are allowed to vary in a range specified
by the designer. Thus, in this approach through evolution, the GA will find a good
database for the FRBM but the RB will be kept as same what was initially consi-
dered by the designer. In this case, the GA-string shown in Figure 4.21 will look
as follows:

Fig. 4.21. A GA-string representing the parameters related to membership functions of in-
put-output variables of a FRBM

Genetic Learning/Tuning of the Fuzzy Rule Base

All methods belonging to this family assume the existence of a pre-defined DB for
the FRBM. In one method, an initial user-defined rule base (assignment of fuzzy
subset to the output variable of each rule) constructed based on designer’s expe-
rience of the process to be controlled is tuned using GA. In GA-sting (as shown in
Figure 4.22), each rule is represented by a single bit (1 or 0), where 1 and 0 re-
spectively indicate the presence or absence of the rule in the optimum rule base.
174 A.K. Nandi

Fig. 4.22. A GA-string representing the rule base of a FRBM

In the method where the RB is designed using GA automatically, additional bits

will be included in the GA-string demonstrated in Figure 4.22 (as shown in Figure
4.23) [10]. For the case where output variable is considered to have 2 fuzzy subsets,
one bit will be required in order to determine the fuzzy subset to the output variable
for each rule (e.g., 0 for LOW and 1 for MEDIUM). Two bits required for the case
where output variable is considered to have 3/4 fuzzy subsets (e.g., 00 for LOW, 01
for MEDIUM, 10 for HIGH and 11 for VERY HIGH), and so on.

Fig. 4.23. A GA-string representing the rule base and automated rule development of a
FRBM

Genetic Learning/Tuning of the Fuzzy Knowledge Base

In this approach both the RB and DB are designed/optimized using GA as stated
above simultaneously and the corresponding GA-string will look as shown in
Figure 4.24.

Fig. 4.24. A GA-string representing the rule base with automated rule development and the
parameters related to membership functions of input-output variables of a FRBM

Besides the way how to construct/design the KB of FRBM, selection of the ap-
propriate shape of fuzzy subsets/membership function distributions (MFDs) for
both the input and output variables in case of Mamdani-type FRBM and selection
of shapes of fuzzy subsets of input variables as well as the appropriate structure(s)
of rule output/consequent function(s) and determination of optimal values of coef-
ficients and power terms of rule consequent functions are important issues. In
order to overcome these problems, a rigorous study with different choices is
required in order to obtain a good model for a manufacturing process.
GA is also used in the genetic-fuzzy system where the TSK-type fuzzy logic
rules (as defined in Section 4.2.11.2.1) are employed. Genetic Linear Regression
(GLR) approach [11] is one of the most popular approaches in designing the KB
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 175

of TSK-type FRBM. In genetic linear regression (GLR) approach, the GA is in-

troduced partly in the multiple regression method. The GLR method will take the
advantages of both the regression technique and GA, and have a capability to find
global optimum and good convergence properties. In this approach, the KB of the
FRBM is optimized using the combined method of linear regression (LR) ap-
proach and genetic algorithm. In this method the coefficients of output function of
each rule are determined using a linear regression approach whereas the input va-
riables’ exponential parameters are simultaneously optimized using a GA. In addi-
tion to that, the MFDs also tune using a GA. In order to accomplish this, besides
the exponential parameters of input variables, the controlling parameter(s) describ-
ing the MFDs (as discussed in Section 4.2.2.2) are considered as GA-variables.
The working principle of GLR approach (as illustrated in Figure 4.25) consists of
following five major steps:

• Step-I: Set an initial set (population) of values of power terms of a given re-
gression function at random
• Step-II: Evaluate the function coefficients based on least square method
• Step-III: Checking of fitness value (if satisfied terminate the iteration proce-
dure)
• Step-IV: Update the values of power terms of regression function using GA-
operators
• Step-V: Repeat Step-II

Fig. 4.25. Flow chart of genetic linear regression approach to construct KB of a FRBM
176 A.K. Nandi

The GA-string of GLR approach will look as presented in Figure 4.26.

Fig. 4.26. A GA-string representing the rule base, parameters related to membership func-
tions of input variables and exponential parameters of rule consequent functions of a TSK-
type FRBM

The proposed GLR has an added facility to carryout the task of tuning MFDs of
input variables simultaneously in the same framework of GA.
In order to determine the coefficients of the output functions of TSK-type fuzzy
rules, a general expression of multiple linear regression system with TSK-fuzzy
model is derived as follows. The Equation (4.11) may be rewritten by denoting

∏ μ v (x1,..., n ) = η r
n
for simplicity, in the following form:
v =1

Y = F (x 1 , .. x n )

∑ ηr (a 1r f 1r (x1,.., n ) + a r2 f r2 (x1,..., n ) + .......... .. + a rk f rk (x1,..., n ))

R lf

= r =1
R lf
∑ ηr
r =1

η1(a f (x1,.,n ) + .. + a f (x1,.,n )) + ..

1 1
1 1
1 1
k k

(
+ ηtr a1tr f 1tr (x1,..,n ) + .. + a ttrj f ttrj (x1,..,n ) + .. + a 2k f 2k (x1,.,n ) + . )
=
(
. + ηRl a1 f 1 (x1,.,n ) + .. + a k f k (x1,.,n )
Rlf Rlf Rlf Rlf
) (4.15)
η1 + η2 + ........+ ηRlf
Let us assume we have a set of input-output tuple (D) of S number of sample data
(i )
where the output y is assigned to the input (x 1(i ) , x (2i ) ,......... , x (ns ))

D= {( ( ) ,....., x ( ) , y ( )), (x ( ) ,....., x ( ) , y ( )).......... .(x ( ) ,....., x ( ) , y ( ))}

x 11 n
1 1
1
2
n
2 2
1
s
n
s s

Now, the total quadratic error that is caused by the TSK-type FRBM with respect
to the given data set:
s

l =1
(
E = ∑ f (x 1(l ) , x (2l ) ,.... x (nl )) − y l
()
)
2
(4.16)

In order to minimize E, we have to choose the parameters,

{(a 1
1 ,...., a ), (a ,..... a ),......,
1
k
2
1
2
k (a R
1
f R
)}
,...., a f appropriately.
k
r
where the parameter a indicates the jth coefficient of the output function of rth rule.
j
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 177

To determine the above parameters, we take the partial derivatives of E with re-
∂E
spect to each parameter and require them to be zero, i.e., = 0 , where
∂ a rj
j = {1,2,...., k} and r = {1,2,....., R f }.
Now, we obtain the partial derivation of E with respect to the parameter a tt rj ,

(l ) ∂f (x1 ,........, x n )
(l ) (l )
∂E
=
S
∑
∂ a ttrj l=1
2 ⋅ f ( (l )
x1 (,......., (l ))
xn y − ⋅ )
∂ a tt rj

⎛ R lf ⎞
⎜ ∑ ηr (a 1r f 1r (x1,.., n ) + ... + a k f k (x1,.., n ))
⎟
l ⎟ η t r f t j ( x1,.., n )
l r r l
S
⎜ ⋅ tr l
= 2⋅∑ r =1
− y ⋅
l =1 ⎜ R lf ⎟ R lf
⎜ ∑ ηr ⎟ ∑ ηr
⎝ r =1 ⎠ r =1

⎛⎛ ⎞ ⎛ ⎞⎞
⎜ ⎜ S R lf ⎟ ⎜ S R lf ⎟⎟
⎜ ⎜ ∑ ∑ η r (a 1r )f 1r (x 1,..,
l
)n η t r f tt rj (x1,..,l n ) ⎟ ⎜ ∑ ∑ η r (a rk )f rk (x 1,..,
l
) ( )
n η t r f t j x 1,.., n ⎟
tr l ⎟
= 2 ⎜ ⎜ l =1 r =1 ⎟ + .. + ⎜ l =1 r =1 ⎟ ⎟
⎜⎜ ⎛ R lf ⎞
2
⎟ ⎜ ⎛ R lf ⎞
2
⎟⎟
⎜⎜ ⎜ ∑ ηr ⎟ ⎟ ⎜ ⎜ ∑ ηr ⎟ ⎟⎟
⎜⎜ ⎜ r =1 ⎟ ⎟ ⎜ ⎜ r =1 ⎟ ⎟⎟
⎝⎝ ⎝ ⎠ ⎠ ⎝ ⎝ ⎠ ⎠⎠

⎛ S l ⎞
⎜ ∑ y ⋅ η f tt rj (x1,..,
l
) ⎟
⎜ ⎟=0,
tr n
−2 l =1 (4.17)
⎜ R lf ⎟
⎜ ∑ ηr ⎟
⎝ r =1 ⎠
Thus, the Equation (4.16) provides the following system of linear
equations from which we can compute the coefficients
{(a 1
1 ,...., a 1k ), (a 12 ,..... a 2k ) (
,......, a 1R f ,...., a Rk f : )}
∏ μ v (x1,.., n)
n
rl

f j (x1,..,n )⋅ f tt rj (x1,.., n )⋅ ∏ μ v (x1,..,n )

Rf K S n
v =1
∑ ∑ a rj ∑ 2
r l l tr l
r =1 j=1 l =1 ⎛ n r l ⎞ v =1
⎜ ∑ ∏ μ v (x1,.., n )⎟
R lf

⎜ r =1 v =1 ⎟
⎝ ⎠

y ∏ μ vt r (x 1,.., n)
n
l l

f tt rj (x 1,.., n )
S
=∑ v =1 l (4.18)
∑ ∏ μ v (x 1,.., n)
l
l =1 R f n
l r

r =1 v =1
178 A.K. Nandi

In matrix form, Equation (4.18) will be written as:

⎤⎡ r ⎤ ⎡β ⎤
r
⎡ r
⎢ α 11 α α 1Kr ⎥ ⎢ a 1r ⎥ ⎢ 1r ⎥
r r
12
.
⎢ α 21 α α 2K ⎥ ⎢ a 21 ⎥ = ⎢⎢ β 2 ⎥⎥
r r
22
.
(4.19)
⎢ . . . . ⎥⎢ . ⎥ ⎢ . ⎥
⎢ r ⎥⎢ r ⎥
⎢⎣ α K1 α α KK ⎥⎦ ⎢⎣ a K ⎥⎦ ⎢⎢⎣β K ⎥⎥⎦
r r r
K2
.
S

∑ f (x , x
S l
α tj =
r r l l l r l l l r r
where j 1 2
.., x n ) f t ( x 1 , x 2 .., x n ) ; βt = ∑ y f t ,
l =1 l =1

where s is the number of training (input-output) sample data. Thus, Equation

(4.18) provides a solution of the function coefficients ( a rj ) of the TSK-type fuzzy
rule consequents for a given value of the input variable’s exponential terms. To
solve the Equation (4.18) in order to find the values of function coefficients, the
Gauss Algorithm with Column Pivot Search Method is used here. More generally,
any conventional numerical method, which provides a representative solution of
Equation (4.18), may be adopted.

4.5 Application to Modeling of Machining Process

Numerous works on modeling of machining processes using soft computing tools
including different GA-fuzzy approaches can be found in the review paper [12]. In
this chapter, some of the previous research works of the author on modeling of
manufactuing processes using GA-fuzzy approaches are presented.

4.5.1 Modeling Power Requirement and Surface Roughness in

Plunge Grinding Process
In [13], a study was carried out to model power requirement and surface roughness in
plunge grinding process using a GA-fuzzy approach. The model considers the input
variables such as wheel speed, work speed and feed rate those mainly influence the
power requirement and surface roughness obtained on the grind surface. In this study
the main objective was to find the effect of various types of MFs considered for in-
put-output variables on the performances of Mamdani-type FRBMs. The approach of
tuned rule base and MFs simultaneously using GA was adopted. The performance of
the models was tested (as shown in Figure 4.27) with experimental results consider-
ing 52100 steel as work material and D126K5V as grinding wheel specification. A
digital clamp power meter is used to take the measurements of power requirement in
grinding and the surface roughness is measured with the help of a perthometer (S6R).
In Figure 4.27, Error 1, Error 2 and Error 3 are the percentage deviations of results
predicted by FRBMs with triangular, 2nd order polynomial and 3rd order polynomial-
type MFDs, respectively from that obtained in experimentation. It is observed that
higher order polynomial type MFs showed better results. It may happen because the
input-output relationship in grinding is highly nonlinear and the linear MFDs (trian-
gular type) may not be sufficient.
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 179

(a)

(b)
Fig. 4.27. Comparison of performances of FRBMs (with different types MFs) with those of
experimental results (a) surface roughness (b) power requirement
180 A.K. Nandi

In the above study, the GA was used to optimize the manually-defined KB of the
FRBM. The manually-defined KB is designed based on the expert’s knowledge of
the process that may not be complete. Sometimes, it becomes difficult to gather
knowledge of the process beforehand. To overcome this difficulty, the method for au-
tomatic design of fuzzy KB is adopted to model power requirement and surface
roughness in plunge grinding process [10]. Table 4.1 describes the comparative re-
sults of root mean square (RMS) percentage deviations exhibited by fuzzy rule-based
models (FRBMs) from those of real experimental values. It is found that the approach
of automatic design of RB and tuning of MFs simultaneously using GA provides bet-
ter result over the approach of tuned manually constructed RB and MFs simulta-
neously using GA. It happens because all the manually-designed fuzzy rules may not
be good, whereas the GA has the capability of finding the good fuzzy rules through
extensive search. Moreover, the main disadvantage of using later approach (the ap-
proach of tuned manually constructed RB and MFs) lies in the fact that the designer
is required to have a thorough knowledge of the process to be controlled. Thus, a
considerable amount of time is spent on manual construction of fuzzy RB. In the ap-
proach of automatic design of RB and tuning of MFs simultaneously using GA, no
effort is made for designing the fuzzy rule-base manually and a good KB of the
FRBM is designed automatically using a GA from a set of example (training) data.

Table 4.1. Comparison of RMS percentage deviations exhibited by fuzzy rule-based mod-
els from those of real experimental values

Power requirement
Mathematical model FRBM based on the FRBM based on the approach of
approach of tuned automatic design of rule base and
rule base and MFs tuning of MFs simultaneously
simultaneously using GA (approach 1)
using GA
(approach 2)
RMS percen- 31.51 8.13 5.34
tage error
Surface roughness
RMS percen- 16.44 10.22 6.32
tage error

4.5.2 Study of Drilling Performances with Minimum Quantity

of Lubricant
The main objective of this study is to investigate the performances of FRBMs
based on Mamdani-type and TSK-type of fuzzy logic rules, and different shapes
(namely, 2nd order polynomial and trapezoidal types) of MFs for prediction and
performance analysis of machining with minimum quantity of lubricant (MQL) in
drilling of Aluminium (AA1050). In this study, predictions of surface rough-
ness obtained in drilling and the corresponding cutting power and specific cutting
force requirements for different amounts of lubricant rate will be carried out
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 181

through a comparative analysis of the results of models with experimental results

as well as those published in the literature. The approach of tuned RB and MFs
simultaneously using GA and genetic linear regression method was adopted to
construct the KB of Mamdani-type and TSK-type FRBM, respectively. The struc-
ture of rule-consequent function for TSK-type fuzzy rules is used as follows

y = c1 V c p1 + c 2 Fr p2 + c3 L r p3 (4.20)
where c1, c2 and c3 are the function coefficients and p1, p2 and p3 are the exponen-
tial parameters of rule consequent function. Vc, Fr and Lr are the input variables,
cutting speed, feed rate and rate of lubricant, respectively.
A helical K10 drill (R415.5-0500-30) was manufactured according to DIN6537
by Sandvik(R). The drill has a point angle of 140º, 28 mm of flute length and is of
10% cobalt grade. The drills possess a diameter of 5 mm and are coated with
TiAlN. A Kistler® piezoelectric dynamometer 9272 with a load amplifier was
used to acquire the torque and the feed force. Data acquisitions were made through
piezoelectric dynamometer by interfacing RS-232 to load amplifier and PC using
the appropriate software, Dynoware Kistler(R). The surface roughness was eva-
luated (Ra according to ISO 4287/1) with a Hommeltester T1000 profilometer.
Here, 4 different models related to surface roughness and, other 4 different
models for cutting power/specific cutting force requirement) are developed. The
four different FRBMs are constructed using two different types of fuzzy logic
rules (Mamdani-type and TSK-type) and two different shapes of MFs.
The comparative results of surface roughness, cutting power and specific cut-
ting force with different lubricant flow rates for different cutting speed and feed
rate are described in Figure 4.28, Figure 4.29 and Figure 4.30, respectively. In this
study, nine different cases (of cutting speed and feed rate) are considered based on
which the effects of lubricant flow rate on machining performances are analyzed.
In these Figures, Model I indicates FRBM with Mamdani-type FLR and 2nd order
polynomial MFs, Model II represents FRBM with Mamdani-type FLR and trape-
zoidal MFs, Model III shows FRBM with TSK-type FLR and 2nd order polynomi-
al MFs and Model IV indicates FRBM with TSK-type FLR and trapezoidal MFs.
In the following subsections the evolutions of the prediction performances of these
models toward the effects of surface roughness, cutting power and specific cutting
force with lubrication rate are discussed.

Surface roughness
In Figure 4.28, the predicted values of surface roughness by FRBMs are compared
with the experimental values for 9 cases (Figure 4.28(i) to Figure 4.28(ix)). It has
been observed that the performance of Model II is better than Model I for first 5
cases and case number 6. For other cases Model I outperforms over model II. But,
the consistencies of deviations of results from that of the experimental values are
not good for both the Model I and Model II in all the cases. In contrast, the results
of Model III shows better than Model I and Model II for some cases (Figure 4.28
(iv), (v) and (vii)), but for other cases that are deteriorated compared to Model I
and Model II. On the other hand it is noticed that the results of Model IV (FRBM
with TSK-type FLR and trapezoidal MFs) yield less error (deviation from the
182 A.K. Nandi

Fig. 4.28. Comparative results of surface roughness with different lubricant flow rates for
different cutting speed and feed rate

experimental values) in majority than that of the other models for all (9) cases.
Furthermore, it is noticed that the results of Model IV are also consistent for dif-
ferent values of lubrication rate. The maximum value of percentage error exhibited
by Model IV is 2.2428, which is well accepted in industrial practice.
By analyzing the experimental values as well as results obtained by Model IV, it
has been observed that the surface roughness is improved by increasing the flow rate
for lower values of cutting speed (60 m/min) and constant feed rate. But, for higher
values of cutting speed, the surface roughness deteriorated with increasing flow rate
(75 and 90 m/min). In contrast, for a constant cutting speed, the rate of change of
surface quality with flow rate is minimized as feed rate increase.
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 183

Cutting power
By analyzing the results of various models and experimental values as depicted in
Figure 4.29, it has been observed that Model I as well as Model II provides poor
results than other two models (Model III and Model IV). In contrast, it is found
that both the models, Model III and Model IV obtain the best performance for
predicting cutting power with the quantity of lubricant for a given cutting speed
and feed rate. However, in cases Vc=90; f=0.15 and Vc=90; f=0.25, the Model IV
shows better results than Model III.

Fig. 4.29. Comparative results of cutting power with different lubricant flow rates for dif-
ferent cutting speed and feed rate
184 A.K. Nandi

By analyzing the experimental values as well as results obtained by Model III and
Model IV, it has been revealed that for a fixed value of cutting speed and feed rate,
the cutting power increase to a certain value of lubrication flow rate. After that the
value of cutting power decreases with increasing flow rate. From Figure 4.29, it is
found that for constant cutting speed, the cutting power requirement increase with
feed rate. It is also observed that the value of cutting power increases with cutting
speed when feed rate is kept as a constant value.
Specific cutting force
As like cutting power, here also Model III and Model IV show the best result in
predicting specific cutting force with the quantity of lubricant for a given cutting
speed and feed rate (Figure 4.30). This is because; both the cutting power and spe-
cific force are depended on the same parameter, torque and a linear relationship is
maintained among them. The variation of specific cutting force requirement with
lubricant flow rate and other input parameters, cutting power and feed rate exhibit
the same phenomena as found in cutting power.

Fig. 4.30. Comparative results of specific cutting force with different lubricant flow rates
for different cutting speed and feed rate
GA-Fuzzy Approaches: Application to Modeling of Manufacturing Process 185

From above discussions, it may be pointed out that FRBMs with TSK type fuzzy
logic rules provide best result in predicting surface roughness, cutting power and
specific cutting force. Specifically for surface roughness, trapezoidal MFs is well
suited, while trapezoidal as well as second order polynomial MFs give almost simi-
lar performances in predicting cutting power/specific cutting force requirements in
drilling of Aluminium AA1050 with emulsion with oil Microtrend 231L lubricant.
The above techniques may be adopted for developing FRBMs for other machining
(drilling) performance parameters. Once the model is developed, it may be used on-
line in drilling machine to control the MQL as per desired outputs.

References
[1] Groover, M.: Automation, Production System, and Computer Integrated Manufactur-
ing. Prentice-Hall Int’l, Upper Saddle River (2001)
[2] Kosko, B.: Neural Network and Fuzzy Systems. Prentice-Hall, New Delhi (1994)
[3] Zadeh, L.A.: Fuzzy sets. Information and Control 8(3), 338–353 (1965)
[4] Mamdani, E.H., Assilian, S.: An experiment in linguistic synthesis with a fuzzy logic
controller. International Journal of Man-Machine Studies 7(1), 1–13 (1975)
[5] Sugeno, M., Kang, G.T.: Structure identification of fuzzy model. Fuzzy Sets and Sys-
tems 28(1), 15–33 (1988)
[6] Tsukamoto, Y.: Fuzzy information theory. Daigaku Kyoiku Pub. (2004)
[7] Takagi, T., Sugeno, M.: Fuzzy identification of systems and its application to model-
ing and control. IEEE Transactions on Systems, Man, and Cybernetics 15(1), 116–
132 (1985)
[8] Goldberg, D.E.: Genetic Algorithms in Search, Optimization and Machine Learning.
Addison-Wesley, Reading (1989)
[9] Deb, K.: Multi-Objective Optimization using Evolutionary Algorithms. John Wiley &
Sons Ltd, England (2001)
[10] Nandi, A.K., Pratihar, D.K.: Automatic Design of Fuzzy Logic Controller Using a
Genetic Algorithm – to Predict Power Requirement and Surface finish in Grinding.
Journal of Material Processing Technology 148(3), 288–300 (2004)
[11] Nandi, A.K.: TSK-Type FRBM using a combined LR and GA: surface roughness
prediction in ultraprecision turning. Journal of Material Processing Technolo-
gy 178(1-3), 200–210 (2006)
[12] Chandrasekaran, M., Muralidhar, M., Krishna, C.M., Dixit, U.S.: Application of soft
computing techniques in machining performance prediction and optimization: a lite-
rature review. Int. J. Advance Manufacturing Technology 46, 445–464 (2010)
[13] Nandi, A.K., Pratihar, D.K.: Design of a Genetic-Fuzzy System to Predict Surface
finish and Power Requirement in Grinding. Fuzzy Sets and Systems 148(3), 87–504
(2004)
[14] Nandi, A.K., Davim, J.P.: A Study of drilling performances with Minimum Quantity
of Lubricant using Fuzzy Logic Rules. Mechatronics 19(2), 218–232 (2009)
5

Single and Multi-objective Optimization Methodologies

in CNC Machining

Nikolaos Fountas1, Agis Krimpenis1, Nikolaos M. Vaxevanidis1,

and J. Paulo Davim2
1
Department of Mechanical Engineering Technology Educators,
School of Pedagogical and Technological Education (ASPETE),
GR-14121, N. Heraklion Attikis, Greece
[email protected], [email protected],
[email protected]
2
Department of Mechanical Engineering, University of Aveiro, Campus Santiago,
3810-193, Aveiro, Portugal
[email protected]

Aiming to optimize both productivity and quality in modern manufacturing

industries, a wide range of optimization techniques and strategies has been
presented by numerous researchers. Optimization modules such as Genetic
Algorithms, Evolutionary Algorithms and Fuzzy systems are capable of exploiting
manufacturing data with high efficiency and reliability, in order to provide
optimal sets of solutions for machining processes. The main scope of this chapter
is to present the fundamentals in formulating and developing optimization
methodologies, which ultimately offer optimal cutting conditions for both
prismatic and sculptured surface part machining and actually improve industrial
practice.

5.1 Introduction

A large number of techniques have been implemented in the field of the manufacturing
parameters optimization and others are considered candidate optimization tools.
Optimization problems include product design, process engineering, quality control,
process planning and different machining operations numerically, or conventionally,
controlled. Moreover, optimization has also been investigated in other highly
sophisticated and/or non-conventional manufacturing processes, such as Electro-
discharge machining (EDM) or abrasive water jet machining (AWJM) [1, 2].
188 N. Fountas et al.

Initial approaches in the field of machining optimization involve classical

optimization, Geometric Programming [3, 4], as well as various Graphical
techniques [5]. Owing to the complexity of the machining problem – mixed real-
valued and integer-valued parameters with unknown interactions – the classical
optimization algorithms take too long to respond and have a high calculation cost,
since they mainly involve deterministic and gradient-based operators. This
practically makes them unsuitable for industrial usage. Thus, Artificial Intelligence
algorithms appear to be a more suitable tool provided that they are well calibrated
and carefully adjusted to the machining problem.
The main goal of this chapter is to provide an overview of the existing
optimization methodologies in CNC machining, i.e., Neural Networks, Genetic
and Evolutionary Algorithms, variations of optimization schemes and hybrid
approaches. In order to better demonstrate the functionality of these systems, basic
terms, definitions and standards about modeling of machining processes are
analyzed, along with quality targets, objective functions and distinction between
single- and multi-objective methodologies.

5.2 Modeling Machining Optimization

Depending on the number of optimization goals, optimization process can be

qualified as single-objective, if a single quality objective is addressed, or multi-
objective, if several quality objectives are undertaken. Since manufacturing problems
have a vast solution space, more than one quality objectives are usually included in
the optimization scheme. In case of multi-objective optimization, the problem should
be carefully designed and the quality attributes suitably determined, according to their
effect on the machining goals.

5.2.1 Definition of Optimization

Optimization of manufacturing processes is the effort to enhance performance of
manufacturing environments by providing an elite set of solutions that perform
best according to the target quality characteristics. Applying optimization
methodology to a given machining task leads to finding maximum or minimum
values of the corresponding fitness function (depending on whether it is
maximization or maximization problem, respectively). This optimization
philosophy can be applied to any scientific field were the aim is to enhance
efficiency, productivity and quality. When optimization refers to a maximization
problem with a quality goal that is measured by function f, the optimal solution is
the one with the highest f value, or equivalently, the one with the lowest value of
the opposite function, - f [6].
Single and Multi-objective Optimization Methodologies in CNC Machining 189

In its mathematical description, an optimization process, if minimizing its

quality goal, may be determined as follows:

For function f : R → R , find

n
xˆ ∈ R n , such that f ( xˆ ) ≤ f ( x ) , ∀x ∈ R n . (1)
Similarly, for a maximization problem, the expression is the following:
For function f : R → R , find xˆ ∈ R n , such that f ( x ) ≤ f ( xˆ ) , ∀x ∈ R . (2)
n n

n
The domain R of f is the search space. Each parameter vector of this domain
is a candidate solution in the search space, with x̂ being the optimal solution.
The value n represents the number of dimensions of the search domain and hence,
the number of parameters involved in the optimization problem. The function f is
the objective (or fitness) function that maps the search space to the function space.
If the objective function has one output, then the respective function space is one-
dimensional and thus provides a single fitness value for each set of parameters.
This single fitness value specifies the optimality of the parameter set for the
desired task. Usually, the function space can be directly mapped to the fitness
space. However, the distinction between function space and fitness space is
important in the case of multi-objective optimization problems, which include
several objective functions drawing input from the same independent variable
space [7, 8].
For a known differentiable function f, calculus may easily provide local and
global optima of f. However, in machining problems this objective function f is
not known. When addressing such problems, the objective function is treated as
“black-box”; input and output parameter values are obtained. The result of a
candidate solution evaluation is the solution’s fitness. Then, the final goal is to
point out parameter values that maximize or minimize this fitness function [7].
In order to suitably approach the optimization of a given problem, components
of candidate solutions should subject to certain constraints. In manufacturing
environments, such constraints may be the maximum available process values and
limitations of involved machine tools and equipment, product quality demand,
minimum manufacturing cost, etc. This statement implies that constraints can be
either economical or technological depending on the nature of the optimization
problem. Specifically, in machining processes optimization, constraints are both
economical and technological. Examples of technological constraints are
maximum available motor power of a CNC machine tool, its maximum capacity
in terms of cutting force load, the generated heat during the machining process,
the maximum available torque and the maximum range of feeds and speeds, etc;
see for details [9].

5.2.2 Objective Functions in Machining Problems

An Objective Function measures a candidate solution’s status or performance in
an optimization algorithm and in that sense it is used for its evaluation. Objective
Functions act as nodes between optimization modules and physical or digital
systems. The Objective Function of each candidate solution is individually
190 N. Fountas et al.

calculated, thus OF values vary in a solutions set. In order to sustain uniformity,

Objective Value V can be mapped into Fitness Value using mapping matrix M,
where f domain is usually greater than zero [10].
M:V→f (3)
When it comes to a minimization problem, individuals are considered better
when low values of the Objective Function are obtained. Further on, the Fitness
Function is used to transform the values obtained from the Objective Function to a
relative Fitness expression [11]. Hence,

F ( x ) = g ( f (x )) (4)
where
f : Objective Function
g: Transformation of the objective function values to nonnegative
F: Resulting Fitness Function.
In CNC machining applications, objective functions vary as far as process
characteristics and constraints are concerned. During different machining stages,
objective functions tend to optimize attributes referring either to productivity or
quality or even both. Posed process constraints are also chosen depending on the
studied machining stage, namely roughing, semi-finishing or finishing stage.
In most cases, objective functions of machining processes are discontinuous
and nondifferentiable. Stochastic optimization methodologies can be applied for
quality objectives optimization, since they are unresponsive to discontinuities and
they do not need derivative information to converge.
Quality objectives usually optimized by suitably developed objective functions
are determined in each machining stage. Considering that in most case a
machining process follows a two-stage scheme (roughing and finishing), the
following quality objectives are specified:

- Roughing Operations:
During rough machining operations, the primary target is to rapidly remove
material from the raw stock, until the roughed part geometry is close to its final
shape. Thus, quality characteristics are mainly related to productivity and time.
Targets, such as High Material Removal Rate, Minimum Roughing Time and
Minimum Remaining Volume for the finishing process to take up on, are the most
common optimization attributes.

- Finishing Operations:
During Finish machining Operation, the primary goal is to achieve final
specifications of the part geometry in terms of surface quality (low surface
roughness), dimensional accuracy and geometrical features within specified
tolerance, regarding the 3D target model, the blueprints and the engineering
drawings. As a matter of fact, quality targets like cutting forces, surface roughness
and machining time are the quality objectives to be minimized.
Single and Multi-objective Optimization Methodologies in CNC Machining 191

5.2.3 Mathematical Modeling of Process Parameters and Objectives

Machining processes values and significant level of quality characteristics are
usually different, according to the nature of the process. If a process strongly relies
on productivity, e.g. high Material Removal Rate in roughing processes, then quality
targets such as machining time play a vital role; much more than others such as
surface quality (see Section 2.2). On the contrary, if a process relies on product
quality, e.g. low Surface Roughness in finishing operations, then respective quality
targets may be of greater importance. Moreover, these function values are different
in order and magnitude. For instance, machining time expressed in seconds for a
rough machining stage may be 10,000 – 100,000 sec, while remaining volume
expressed in cubic millimeters may be of a much higher order, such as about
10,000,000 mm3. It is imperative to stress out that normalization procedure should
be carried out especially when it comes to contradictory quality characteristics put
together in the same problem. If, for example, the values of machining time and
remaining volume are summed up to formulate a single objective, then an inherent
bias in favor of machining time appears, owing to its values’ order. Obviously,
results can be quite disorienting as far as the problem overall response is concerned.
Significance level of quality characteristics is determined according to user’s
demands through weighing coefficients. Simple mathematical operations for data
normalization are applied to overcome this problem. Thus, linear combination of
quality targets, and their respective weight coefficients according to their
importance in the studied problem, can be used. This type of expressions offers no
obvious bias, provided that values are normalized beforehand. In mathematical
terms, it is described as follows:
Let X1, X2,......Xi be the given quality characteristics (as in multi-criterion
optimization) and w1, w2,......wj their respective weight coefficients. Practically, a
weight coefficient’s value – if it expressed as a percentage – represents the
respective quality characteristic’s significance in the problem. Significance, thus,
is applied through wi values in the Quality Characteristic Equation (5).
QC (X1, X2,......Xi, w1, w2,......wj) = X1.w1 + X2. w2 +...... Xi.wj (5)
Difference in magnitude of the quality attributes is yielded by their parameters
from which the attributes are controlled. The most common machining parameters
that are necessary for tool-path generation are listed here.
1. Machining Strategy / Toolpath Style: This option enables the user to
determine what strategy will be used to machine a work piece. Each
machining strategy offers additional parameters that define the way the
tool moves and the tool-path style in more detail.
2. Stepover (Radial cutting depth): This parameter is the Radial Distance
between two successive tool passes in X-Y plane (3-axis machining) or
in the plane that is vertical to tool axis (5-axis machining). It can be
determined either by an arithmetic value or as a ratio of cutting tool
diameter (also called engagement).
3. Stepdown (Axial cutting depth): This parameter adjusts the maximum depth
of cut in the vertical direction (or tool axis direction), between successive
192 N. Fountas et al.

machining planes. It can be determined either by setting the number of

cutting planes, or by specifying the distance between machining planes.
4. Machining Tolerance: The value of the maximum allowable distance
between the theoretical tool-path and the tool-path computed by the
CAM system.
5. Machining Direction: This option adjusts the cutting direction whether to
be of “Climbing” type (the front of the advancing tool cuts into the
material first) or of “Conventional” type (the back of the advancing tool
cuts into material first).

5.2.4 Single-Objective and Multi-objective Optimization

As mentioned previously, an optimization problem may be either single- or multi-
objective, depending on the number of its quality targets and the type –scalar or
vector– of the objective function. When dealing with machining optimization
problems, quality characteristics are usually more than one (and often conflicting),
due to high process complexity. The overall goal is to find solutions that are
acceptable by the decision maker. Although in single-objective optimization
problems, it is straightforward to find an optimal solution, multi-objective ones
provide users with sets of solutions, out of which one should be chosen so as to
apply in physical systems.
In multi-objective optimization, the objective function is a vector, thus belonging
to a multi-dimensional space. For each solution x in the decision variable space X,
there is a point in the objective space Z illustrated by f(x) = Z = (Z1, Z2,....Zn)k. Points
in the objective space are considered to reflect optimal solutions when their
coordinates lie in the nearest possible location to the coordinate system origin. The
set of solutions having the smallest vectorial distance to the coordinate system origin
is the optimal solution set; the so-called Pareto Front. When such a solution set is
found, then the optimization algorithm has converged.

5.2.4.1 Globally and Locally Pareto Optimal Sets

As in the case of single-objective optimization, there are also local and global
optimal solutions, or Global and Local Pareto-optimal sets. When the set SP is the
entire search space SS (SP=SS) then the resulting non-dominated set SP' is the
Pareto-Optimal Set. The non-dominated set of the entire feasible search space SS is
the globally Pareto-Optimal Set. If for every solution x in a the set SP no
solutions y are existed in the neighborhood of x, |y-x|≤ ε, dominating any member
of the set SP, then SP constitutes a locally Pareto-Optimal Set.

5.2.5 Choosing Optimization Philosophy

In general, manufacturing processes are dynamic procedures whose control
variables (range of values) vary depending on the specific case. Manufacturing
demands for products may be different, and so do manufacturing practices. Under
this assumption, suitability of an optimization methodology should be tested and
evaluated beforehand.
Single and Multi-objective Optimization Methodologies in CNC Machining 193

Differentiation in the structures of optimization algorithms establishes several

optimization philosophies. Therefore, an optimization approach is chosen based on
the quality objective (or objectives), the constraints, the process and the preferred
workflow of problem handling. More specifically, an optimization module should be
selected to be implemented by taking into account the following issues:

• The nature and the number of process parameters.

• The number of quality objectives
• Applying single- or multi-objective optimization philosophy
• The selection among stochastic or conventional optimization approaches
• The potential constraints on the solution space and on parameter values
• The computational cost
• The proper functionality per case study
• The iteration time (as short as possible).

5.3 Building Meta-Models for Machining Processes Using

Artificial Neural Networks

5.3.1 Basics of Artificial Neural Networks (ANNs)

Although there is not a single, or universally accepted, definition of ANNs, most
researchers agree that an ANN is a network of many simple processing “units”
(also referred to as “elements”, “nodes” or “neurons”) each with a small amount of
local memory and in some cases arranged in layers. Nodes are connected through
communication “channels” (usually referred to as “connections”, “connection
weights” or “weights”), see Figure. 5.1. Connections bear encoded arithmetic data.
Nodes process only local data and inputs brought to them through the connections.

Fig. 5.1. Feed-forward multi-layer perceptron (MLP) with two hidden layers.
194 N. Fountas et al.

Scientifically, Artificial Neural Networks (ANNs) are systems consisting of a

number of processing units or elements that calculate in parallel and whose
functioning is determined by the network structure, the connection weights and the
processing that is carried out on the elements or the nodes [4]. These systems tend
to store empirical knowledge and to make it available for use similarly to the
human brain. Knowledge is acquired through learning and it is stored in the
connections between neurons [12].
Formulation of ANNs is the mathematical modeling of the biological neuron
networks, so that they can perform complex, and possibly “intelligent”,
calculations similar to the ones performed by the human brain. Most ANN types
need a “training” stage, during which connection weights are adapted according to
a set of training data. Thus, ANNs “learn” or “are trained” by examples. When
trained carefully, ANNs can “generalize” (e.g. predict output values) beyond the
field of training data. Training algorithms commonly implemented are Gauss-
Newton [13-15], Levenberg-Marquardt [16, 17], conjugate-gradients algorithms
[18], etc. ANNs can operate on parallel computer systems, since calculations on
nodes are highly independent to each other.
Researchers have developed a large number of ANN types [19]. In the field of
machining applications, the most commonly used types are Multilayer Perceptrons
(MLP) and Radial Basis Function networks (RBF), that belong to the Supervised
learning - Feed forward ANN category, and Self-Organizing Maps (SOM) that
belongs to the Unsupervised learning - Competitive ANN category.
Figure 1 depicts a feed-forward MLP that is typically applied to function
approximation problems, such as machining time or surface roughness prediction.
It consists of four layers; one input layer, one output layer and two hidden layers.
Any node of a given layer is connected through connection weights (wijk) to all
nodes of the previous and the next layer. Apart from weights, biases (bij) are fed to
each node. A node’s output is an activation function (nonlinear, such as tanh or
1/(1+exp(-x))) value, whose argument is the node’s sum of weights, plus the
node’s bias. Training stage ultimately fixes wijk and bij values and it is concluded
when all inputs and outputs are matched with a predetermined error.
ANNs are mainly implemented in two major problem categories: (a)
Classification or sorting problems, and (b) Value Prediction or Estimation of
unknown functions. In classification problems, ANNs are tolerant to inaccurate
data and can cope with vast training sets at which strict and fast rules fail, such as
in Expert Systems. Moreover, almost any vector function defined at compact
spaces can be approximated with a given accuracy by a feed forward ANN, if
adequate data and resources are available [20-23].
“Case” or “example” is a vector whose values are presented to all input nodes
of an ANN. This vector can also include the objective or output values. An input
vector value is also called “input variable” or “independent variable”, while an
output vector value is called “output”, “objective” or “dependent variable”. “Data
sets” are matrices containing a number of “cases”.
The main objective of ANNs is to generalize, thus to accurately perform using
new data. A data set presented to an ANN at any time is called “sample” and is
Single and Multi-objective Optimization Methodologies in CNC Machining 195

divided in three subsets: training subset, which is used for parameter fitting
(learning), validation subset, which is used for network architecture tuning, and
test subset, which is used for accessing the generalization ability of a trained
network. In literature, the use of validation and test sets is often reversed [24, 25].
Finding an ANN that performs optimally in new cases, while it does not just
memorizes the already known cases with which it was trained, means that its
performance is measured by an error function (e.g. mean square error, total
absolute error, etc) when unknown –independent– data is presented to the network
[24]. The validation set consists of these new cases. However, ANN efficiency is
measured by a third –test– set, since validation procedure may lead to ANN over-
fitting (data memorizing).

5.3.2 Machining Data Sets for Training, Validation and Test

of ANNs
For specific parts, it is possible to immediately correlate machining parameters to
Quality Characteristics through ANN. Machining parameters involve tool
geometry, tool-path strategy, feed-rates, etc., while machining time and surface
quality are considered Quality Characteristics. Based on this concept, properly
trained ANNs are utilized as prediction functions. Machining parameters are the
independent variables and Quality Characteristics are the dependent variables.
Training back-propagation feed forward multi-layer perceptrons with
Levenberg-Marquardt algorithm leads to precisely predicting functions, under the
assumption that there are sufficient training, validation and test sets. Using Design
of Experiments [26], sets of data are systematically gathered. If ANOVA is carried
out on these sets and no interdependencies among machining parameters are
found, all parameters can be considered independent variables.
All necessary data can be accumulated through CAM software or by direct
experiments on a CNC machine tool. In the first case, an application which
interacts with the software must be developed for this purpose, so as to provide it
with machining parameter values and to extract machining time and remaining
volume values for further exploitation.
Accumulated machining data can be unequally divided into the three required
subsets, provided that the data selection for each is unbiased. Random choice of
data is a simple and adequate procedure, in order to achieve proper set division,
keeping in mind that the major part of existing data are needed as training subset
and the remaining data formulate validation and test sets. Although data coming
from actual experimentation on CNC machine tools could be mixed with data
from CAM software simulation for training, validation and test sets, users are
advised to avoid such practices; ANN training would be more complicated and
probably it will be not able to generalize. The measured results of actual
experiments on machine tools include “signal noise” owing to stochastic
phenomena such as machine tool vibrations and stiffness, chattering, tool and
material defects, etc which are usually not included in CAM software.
196 N. Fountas et al.

5.3.3 Quality Characteristics Predictions Using ANN

Meta-models
In machining processes, the main goal of ANN usage would be the creation of
prediction functions of quality targets such as the Machining Time, the Remaining
Volume (after a roughing operation) and/or Surface Roughness (in finishing), as
long interdependency among machining parameters and the aforementioned
Quality Characteristics exists. It has been seen that such interdependency exists.
Output level of a trained ANN contains either one or all the predicted Quality
Characteristics. It is noted here that Quality Characteristic’s values must have
been properly normalized prior to the ANN training process. Moreover, it is
possible that an ANN contains a weighed sum of more than one Quality
Characteristic in a single-node output level. Researchers use the approach of either
a single ANN that predicts all Quality Characteristics at the same time, which
means that a larger training is needed for training, or a number of separately
trained ANNs that predict a single Quality Characteristic.
An inherent drawback of using ANNs as prediction functions is that training
data are usually based on specific machine tools, tools, part materials and
geometries. In other words, ANN performance appears case-dependent. To
overcome this, researchers can use data sets stemming from experiments on
groups of similar (or parametrically designed) parts or ones that include the
philosophy of machining features. Thus, trained ANNs can predict Quality
Characteristics on any part of a certain group.
There is a plethora of publications concerning the application of ANNs for the
prediction and/or optimization of the resulted outputs (quality characteristics) in
machining. The discussion of such an extensive literature is beyond the scope of
this chapter. The readers interested in particular aspects of the topic may refer to
[27-30].

5.4 Genetic and Evolutionary Algorithms

Genetic Algorithms (GAs) are optimization modules that imitate the theory of
Natural Selection through evolution, as developed by the 19th century biologist,
Charles Darwin. In computers, GAs operate on a population –a set of potential
solutions– by applying the principle of “survival of the fittest” [31] to produce
better approximations to a solution of a given optimization task. Moreover, GAs
are characterized by their stochastic ability to seek for optimized solutions within
a search space, called exploration. GAs were first proposed and investigated by J.
Holland at the University of Michigan [32]. Populations evolved in iterations,
called “generations”. A new population, the offspring, is created in each
generation by applying mathematical operators on the old population, such as
crossover, mutation, inversion, etc.
Evolutionary Algorithms (EAs) are in fact generalized GAs. They include more
sophisticated operators than the ones in GAs and possibly implement other
Artificial Intelligence, deterministic or heuristic algorithms, in order to reduce
Single and Multi-objective Optimization Methodologies in CNC Machining 197

evaluation cost or increase the overall algorithm performance in terms of solution

quality and of high solution exploitation. In simple words, exploitation is a GA’s
or an EA’s ability to rapidly reach global optima after finding local optima inside
the solution space.
The criterion by which candidate solutions in EA populations are evaluated is
the Objective Function, as described earlier in this chapter. If the Objective
Function is scalar, then the method is called Single-objective optimization,
regardless of the number of physical objectives included. In the same sense, if the
Objective Function is a vector, EAs perform Multi-objective optimization and the
solutions are most commonly ranked based on Pareto optimality. For further
details on GAs and EAs the reader may refer to [9].

5.4.1 Basic Genetic Algorithm Structure

Basic Genetic Algorithms (GAs) have a simple architecture, but they have to be
suitably extended for a given optimization problem to be solved. A number of
genetic operators have been proposed to apply various optimization problems. The
fundamental operations a simple genetic algorithm may perform under its basic
architecture are the following:
1. GAs work with strings coded the candidate solution
2. GAs work on a population of strings
3. GAs need only fitness value, not derivative information.
A basic GA is composed of Roulette Wheel Selection (reproduction), one-point
Crossover (recombination) and a simple mutation (see Section 4.1.1). Each of
these basic mechanisms works on strings in a population only with simple bit
operations. The procedure of a GA may be as follows:

begin
Initialization
repeat
Roulette Wheel Selection
Crossover
Mutation
Evaluation
until Termination_condition = True
end.

5.4.1.1 Encoding

When an optimization problem is solved with GAs, the solution space should be
encoded into a string space. In other words, encoding means to map variables
from the solution space into a finite-length string space. Good encoding schemes
are thus required, so as to efficiently solve an optimization problem. Several
encoding methods have been proposed so far; see [33-35]. The most important
issue during encoding is to cover all the solution space with the mapped string
198 N. Fountas et al.

space without redundancy. Consequently, the phenotype of the string space should
be equal to the problem solution space, in order to make the problem simpler.
What is more, string space should be generated as a set of feasible candidate
solutions in order to avoid unnecessary search from the algorithm.
Genetic operators (see Section 4.1.2), however, directly work on the genotype
(often called “chromosome”) of GAs. Therefore, the performance of the genetic
search highly depends on symbolic operations for the genotype. GAs perform
better when substrings have consistent benefit throughout the string space. This is
based on the concept that an encoding method is deeply involved in crossover
operators. The neighborhood of a candidate solution in the solution space should
be similar to the neighborhood of the respective string in the string space. If the
offspring generated by a crossover operator is not similar to its parents, the
population would proceed towards a different evolutionary direction and genetic
search would probably result in failure. On the contrary, if good offspring is
generated, genetic search results in success. When relationships between encoding
methods and genetic operators are taken into account, two principles of the
encoding rule are applied [31]:

1. The smallest characters that represent a solution space should be selected.

2. An encoding method that generates consistent strings through genetic
operators should be designed.
There are three types of variable encoding: (a) real value, (b) binary and (c)
Gray binary. The binary types are usually preferred, since they are said to offer no
bias during the evolution being strings of binary digits. In some cases, encoding
with real values (between 0 and 1) tends to make the GA converge faster, but this
way the GA solutions are susceptible to entrapment in local optima.
GAs perform best when elite substrings tend to appear more frequently in a
population without having their genetic material altered by genetic operators.
Although the search depends on genetic operators, genetic information inherited
from parents to offspring is dependent on the encoding. Therefore, as it is already
been mentioned, a good encoding method must be adopted, so that genetic
operators can produce meaningful offspring.

5.4.1.2 Selection

A GA procedure starts with reproduction, recombination and mutation. Selection

simulates the process of natural selection and GAs need a similar mechanism to
make a population evolve toward a better direction of optimal solutions. The main
procedure of a selection module is to reproduce a population by pre-selecting an
individual with a selection probability proportional to its fitness value. In this kind
of selection, an individual with a higher fitness can reproduce more offspring.
Since selection is performed by stochastically, a number of identical (or the same
exactly) individuals may be chosen by chance. As a result, some types of strings
tend to occupy a population. This is called ''genetic drift'' [32] or ''random drift''
which often occurs in the case where the GA population size is relatively small
Single and Multi-objective Optimization Methodologies in CNC Machining 199

and has strong elitist behavior. Several selection schemes have been proposed in
order to prevent a population from genetic drift.
Selection schemes are classified into two main categories; the proportional
selection scheme and the competition selection scheme. In the first category,
selection is based on the fitness value of an individual compared to the total fitness
value of the overall population. In the second category, selection is based on the
fitness values of some other individuals. Some of the most commonly used
selection schemes are presented next [31]:

1. Roulette wheel selection scheme (or Monte Carlo method): This scheme
selects an individual with probability in proportion to its fitness values.
2. Elitist selection scheme: This scheme preserves the fittest individual
through all generation t, that is, the fittest individual is certainly selected
prior to others into the next generation.
3. Tournament selection scheme: the individual with the highest fitness
value between m randomly pre-selected individuals is selected. Note that
m is the number of competing members.
4. Ranking selection scheme: This scheme is based on the rank of the
individual's fitness value. Additionally, the individuals are selected by the
number of their reproduction into the next generation based on the
ranking table predefined earlier.
5. Expected value selection scheme: This scheme is based on the expected
value of the individual's fitness. According to a respective probability, the
expected value of an individual is calculated. Then, the probability of the
selected individual is decreased by 0.5. Thus, this selection scheme
relatively prevents an individual from being selected more than twice in
the population.

5.4.1.3 Genetic Operators

The genotype of an individual is changed by genetic operators, such as combination,

inversion and duplication, while evolution is controlled under natural selection. GAs
require genetic operators to make the evolution feasible to all successive
populations. Recombination generates new individuals with crossover operators.
Mutation generates new individuals with some perturbation operators. Each operator
contributes to evolution by performing specialized actions.

5.4.1.3.1 Crossover
The Crossover operator generates new individuals as solution candidates in GAs
[31]. GAs can search the solution space mainly by using one of the crossover
operators. With the absence of crossover operators, GAs would be random search
algorithms. The crossover operator exchanges each substring between two
individuals and replaces old individuals with others of a new genotype. The
recombination between two strings is performed according to the type of
crossover operator. Depending on the number of break-points among individuals,
the crossover mechanism recombines the strings of the genotype.
200 N. Fountas et al.

Some of the crossover operators are the following:

- One-point Crossover is one of the basic crossover operators. By
choosing a break-point as a cross site, the one-point crossover
recombines substrings between two individuals.

Example 100|11110 → 100|10010

101|10010 → 101|11110

- Multi-point Crossover is a crossover operator with more than one break-

points on a string. Multi-point crossover recombines some substrings
which cut by some break-points between individuals.

Example: 1|11|1|1 10010

0|00|0|0 01101

- Uniform Crossover. In the Uniform Crossover operator, a mask pattern is

randomly generated including “0” and “1”. Recombination between two
individuals is then achieved by exchanging the characters in the genotype
string according to the mask pattern.

Example: 11|11|1 11001

00|00|0 00110 (Mask Pattern: 00110)

- Cycle Crossover operator first selects a starting point (not a crossing

site). Then, it makes a closed round of substrings between individuals.
The following example describes the workflow of Cycle Crossover:

Example: Consider the Parents P1 and P2 as follows:

P1: 31245
P2: 24351

Firstly the starting point on the string is chosen and then the selection of
locus 2 of the P1 parent is done. A closed round of substring begins from
this starting point. The characters 1 and 4 in the locus 2 of the P1, P2 to
the same locus of offspring are copied, respectively:
O1: *1***
O2: *4***

Next, the character of the locus existed the character 4 in the locus 2 of
P2 is copied as follows:
O1: *1*4*
O2: *4***

The process is repeated until the first chosen character of P1 is reached

again. Consequently, the closed round of substring is achieved as
follows:
O1: *1*45
O2: *4*51
Single and Multi-objective Optimization Methodologies in CNC Machining 201

Finally, characters are filled from the former parents to the latter
offspring. The complete offspring becomes:
O1: 21345
O2: 34251

- Partially Matched Crossover (PMX). The PMX Crossover Operator is

performed by executing two procedures. Like the Cycle Crossover
operator, PMX Crossover is a specialized mechanism which may be
applied in order to overcome permutation problems [31]. The workflow
of the PMX Crossover is described below in the example:

Example: Consider the parents P1and P2 having genotypes as follows:

P1: 31|24|5
P2: 24|35|1

The exchange of substrings is performed between two break-points to

both parents without the overlapping of characters
O1: **35**
O2: **24**

The character 2 in the locus 3 of P1 is exchanged with the character 3 in

the same locus of P2, that is, the positions of the characters 2 and 3 in P1
and the characters 3 and 2 in P2. The characters on the substring rounded
by two break-points are also operated by this exchange, respectively.
Consequently, partially matching substrings can be obtained as follows:
O1: 2*354
O2: 3524*

The remained character is copied from P1 to O1 and from P2 to O2. As a

result, the complete offspring is generated as follows:
O1: 21354
O2: 35241

In this way, genetic algorithms can generate new individuals of feasible

candidate solutions satisfying the constraint of permutation problems [36]. It can
be clearly seen that an independent relation exists between coding and crossover.
When it comes to the design of crossover operation, one should take into account
the inheritance of genetic information from parents to offspring. The offspring
should inherit adequate genetic information from parents.

5.4.1.3.2 Mutation
Mutation occurs as the replicating error in nature. In GAs, the mutation operator
replaces a randomly selected character on the string with the other one [31].
Mutation is performed regardless of individual fitness values. A classic mutation
operation is the one-point changing per individual. Several mutation types are
occurred in nature such as inversion, translocation and duplication. These are also
the mutating mechanisms applied to GAs to simulate these phenomena.
202 N. Fountas et al.

- Inversion. Inversion partially changes a character sequence from one

direction to the opposite. At first, two points are chosen randomly and the
string is cut at these points. Then the substring is linked in the reverse
direction into the remained one.

Example:
Assume a string included 5 characters. Two points on the locus 2 and 4
are selected:
1|234|5

By linking the substring, inversion results in the following state:

14325

- Translocation (or Shift). This operator changes the character sequence

with moving to the different position. In translocation, a substring is
chosen randomly as occurs in Inversion. The substring is then
translocated to a randomly chosen break point (or locus).

Example:
Assume a string included 5 characters. The segment from the locus 1 to 2
is chosen as a substring.
|12|345

The substring is inserted from the locus 4 to 5 while the remained

substring of the string shifts to the leftmost point of the string as follows:
34512

- Duplication. This mechanism overlaps the same substring on a string. the

substring is randomly chosen as it occurs in inversion. The substring is
then overwritten to a randomly selected locus.

Example:
Assume a string included 5 characters. The segment from the locus 1 to 2
is chosen as a substring.
|12|345

The substring is overwritten over the locus 4 and 5 and the remained
substring in not operated as follows:
12312

However the above generated individual does not satisfy the constraint for
permutation problems. Therefore, duplication is not suitable to address this
problem.
Single and Multi-objective Optimization Methodologies in CNC Machining 203

Mutation operators have great influence on the GA performance because they

seriously affect populations. GAs are able to search a global solution space, since
the mutation mechanism randomly changes the strings of individuals. If the
probability of mutation is high, the mutation operator often breaks important
substrings on individuals. Therefore, it is essential that a very low mutation
probability should be used to overcome this drawback.

5.4.1.4 Advanced Reproduction Models

Reproduction operations are applied to individuals during their replacements by

others which have better fitness values. There are currently two reproductive
techniques in use. The first one is called the Generational reproduction model and
the second the Steady-State reproduction model.

5.4.1.4.1 Generational Reproduction Model

The Generational Reproduction model is probably the most widely used in the
applications of GAs. The generational reproduction mechanism replaces all the
population at once with a new population. This is achieved by selecting an
individual from the current population according to fitness value and adding this
individual to the next population repeatedly. This approach of elitist reproduction
has the drawback of low convergence speed.

5.4.1.4.2 Steady-State Reproduction Model

The Steady-State reproduction model replaces only a few individuals in a
generation. Further on, crossover and mutation is applied on these individuals. The
generated individuals are immediately available for generating and selecting next
individuals, hence the convergence speed increases. Generally, an individual is
replaced with one generated by genetic operators in each generation. Therefore,
methods that select individuals to be deleted should be introduced. Alternative
deletion methods were presented by [37]:

1. Delete Least Fit. Deletion of the least fit individual from the population.
2. Exponential Ranking: The worst individual has a probability p of being
deleted. If it is not selected, then the next to the last also has a
probability of p chance and so on.
3. Reverse Fitness: Each individual has a probability of being deleted
according to its fitness value.

5.4.2 Evolutionary Algorithms

5.4.2.1 Evolution Algorithm Structure

An Evolutionary Algorithm’s optimization philosophy is more or less the same to

a GA’s. In fact, there is not a discrete distinction between the definition of EAs
and GAs. One could say that GAs are a subset of the general EA family. EAs
204 N. Fountas et al.

utilize a population of solutions/individuals that successively evolves in

generations and constantly produces better solutions until the algorithm converges
and the global optima is obtained. Operators help EA population find potentially
better solutions by recombining, mutating and other operators that change the
original solutions, while they substitute bad/old individuals with better/younger
ones in the population.
EAs, however, can be different to GAs in a number of manners. One common
differentiation is that operators can be functions of time, in other words
probabilities changing with respect to the current generation number or have
completely different schemes that are chosen according to a specific probability or
to the current generation number.
A major trend that minimizes calculation cost is to divide EA population
appropriately in subpopulations. Each subpopulation optimizes a certain objective
function, thus the EA subpopulations are competitive to one another and they
exchange optimal values at the end of each generation. This notion comes from
Game Theory, at which players in a game behave competitively to each other and
try to make maximum profit, until a certain point is reached at which no player
makes more profit by changing his playing strategy (Nash Equilibrium point).
Alternatively, all subpopulations can optimize the same objective function, but
calculations are carried out on different computers or on different CPUs. This is
called parallelism of computers (or parallel systems, see Section 5.4.2.2.1) and
algorithms and can lead to serious cut-down of calculation cost. In either case,
subpopulation are parts of a single population, thus there is need for
intercommunication among them. This is performed introducing migration
operators that control the way that individuals move to other subpopulations
carrying all genetic information along with them.
In some optimization problems, such as machining or flow optimization, the
evaluators are very costly and take up too many resources, in order to calculate
objective function values. This means that the EA must “pay” high calculation
cost for all solutions, even for the bad ones. However, if there is a way to classify
solutions as “bad” or “good” prior to evaluation, then only the “good” would be
evaluated and cost associated with “bad” solutions/individuals would almost be
eliminated. During the last years, classification problems are commonly solved
using Artificial Neural Networks (ANNs). Both supervised and unsupervised
ANNs can perform decently in solution classification, but it has been shown that
the unsupervised ones work best at such problems in very small response times.
Including ANN operators in the EA optimizes its performance in terms of
calculation time. Nevertheless, solutions/individuals classified as “good” should
always be exactly evaluated by the EA. There is an inherent drawback in ANN
implementation; there is always the possibility that a “good” solution is classified
as “bad” and vice versa. This is overcome by the nature of EA, since “good”
genetic material is never lost or ignored through the use of the other EA operators,
such as mutation, crossover etc., that force it again back in the population either in
the current or a following generation.
Single and Multi-objective Optimization Methodologies in CNC Machining 205

Fig. 5.2. Indicative flowchart of an EA.

Figure 5.2 depicts the flowchart of an EA that works on the philosophy

described in this section as an example. As it is obvious, many variations and
differentiations can be met in the relative literature. Apart from Game Theory and
ANNs, heuristic methods are popular in the field of EAs. Heuristics are based on
experience; they help in minimizing calculation cost and facilitate the algorithm in
206 N. Fountas et al.

finding global optima rapidly, in other words augmenting both exploration and
exploitation abilities of the algorithm. In the same sense, other knowledge-based
systems can be implemented in EAs, such as Expert Systems.

5.4.2.2 Operators in Genetic Algorithms

In order to improve efficiency and speed of GAs, several advanced genetic

operators have been introduced and applied so far. Mechanisms of genetic
operators are based on biology, genetics and ecology. Moreover, advanced
operators in GAs are applied to address difficulties related to premature
convergence and minimal deceptive problems. The following sections represent
some of the most known and implemented schemes that integrate the architecture
of Genetic Algorithms.

5.4.2.2.1 Parallelism
Parallelism of GAs has been proposed by several researchers; see for example [38,
39]. Parallelism can be divided into two types. The first type works with a
population divided into several sub-populations. The genetic operators in a sub-
population prevent local minima from widely propagating to other sub-
populations. The second type is to facilitate the rapid computation by working
with parallel computer systems. Both types may be realized simultaneously. A
Parallel Genetic Algorithm where individuals in a population are placed on a
planar grid was proposed in [38]. Both Selection and Crossover are limited to
operate on individuals in neighbourhoods on that grid. During the next generation,
individuals from a specific location are selected. An individual of the old
population is replaced with the selected one. Crossover is performed by mating
individuals from the same neighborhood.
The main difference to the basic GA is the selection by replacement Operator.
Each individual is replaced with a selected individual with a higher fitness value.
The pseudocode of this parallel GA is:

begin
Initialization
repeat
Selection by Replacement
Crossover
Mutation
Evaluation
until Termination_condition = True
end.

Construction of the grid presented above, which constitutes an advanced

structural part of Parallel GAs, permits the local selection among individuals
found in sub-populations; this imitates the natural phenomenon of mating
individuals in local environments.
Single and Multi-objective Optimization Methodologies in CNC Machining 207

5.4.2.2.2 Migration
In Parallel GAs, genetic operations are performed locally and independently in the
divided subpopulations. Each subpopulation tries to locate good local minima.
New offspring is generated by genetic operators within each subpopulation.
Therefore, each subpopulation evolves toward different direction, much like a hill-
climbing algorithm. After some generations, the best solutions in a locus
(subpopulation) are propagated to neighboring subpopulations. This is called
Migration.
During Migration, the best individual in a generation is sent to its neighbors as
a migrating individual. The migration frequency is an essential parameter for the
efficient performance of this special operation. If the migration phenomenon
appears with high probability, the group of sub-populations works equivalently to
a single population. On the contrary, if no migration appears for a number of
generations, sub-populations only perform local hill climbing.
The division of a population into sub-populations prevents premature convergence
to local optima. GAs consisting of sub-populations are often run on parallel computer
systems with multi-processors, since each subpopulation can easily be assigned to a
different processor.

5.5 Variations of Evolutionary Algorithms

Except for the aforementioned modules of Genetic and Evolutionary algorithms, a
wide range of variations for process optimization has been presented in the
relative literature. Related evolution-based optimization algorithms are Particle
Swarm Optimization, Simulated Annealing, optimization with Tribes, Tabu
Search algorithms, etc. Some of the most known and extensively applied modules
for optimization are demonstrated and described in the following sections.

5.5.1 Particle Swarm Optimization

Particle Swarm Optimization (PSO) is an optimization strategy that explores the
search space in order to maximize a specific quality target. Kennedy and Eberhart
[40] were the first to introduce PSO. Their concept was based on the swarming
habits of certain kinds of animal species and the field of evolutionary computation.
PSO algorithm has the ability of the simultaneous maintenance of several
candidate solutions in a given search space. In each algorithm iteration, a
candidate solution is evaluated by the objective function being optimized, thus
determining the fitness of that solution. The candidate solutions in PSO algorithm
are treated as particles “flying” through the fitness range searching for the
minimum or maximum of the respective objective function. The initial selection of
candidate solutions within the search range is random. Every “'particle”, which
composes the candidate solution, sustains its position, its evaluated fitness and its
velocity. The best fitness value that may be achieved is referred to as the
individual’s best fitness and the solution, from which the best fitness is obtained,
is referred to as the individual best position (or individual’s best candidate
208 N. Fountas et al.

solution). Further on, PSO algorithm maintains the best solution in the group of
particles until the optimization procedure terminates. This best solution is known
as the global best position or global best candidate solution [41].
There are three major steps in the PSO algorithm development. These steps are
repeated until termination/convergence criteria are met:

- Evaluation of particles’ fitness values. This step is conducted by feeding

the objective function with candidate solutions.
- Updating individuals, global best fitness values and candidate solutions.
This is achieved via the comparison between the newly evaluated fitness
values against the previous individuals and global best fitness values,
replacing values when it is appropriate.
- Updating velocity and position of each particle. This step determines the
optimization ability of PSO algorithm. The velocity of the particles in the
swarm is updated using the following equation:

vi (t + 1) = wvi (t ) + c1r1[ xˆi (t ) − xi (t )] + c2 r2 [ g (t ) − xi (t )] (6)

In Equation (6), i represents the index number of each particle. Moreover,

vi (t ) is the velocity of the i-th particle at time t, whereas xi (t ) denotes the
position of i-th particle at time t. Parameters w, c1, c2, (0 ≤ w ≤ 1.2, 0 ≤ c1 ≤ 2 and
0 ≤ c2 ≤ 2) are coefficients adjusted by the user. The values r1 and r2 (0 ≤ r1 ≤ 1
and 0 ≤ r2 ≤ 1) are random values regenerated for each velocity update.
xˆi (t ) represents the individual best candidate solution for particle i at time t, and
g (t ) is the swarm’s global best candidate solution at time t.
As for the rest three terms existed in the velocity Equation (6), it should be
mentioned that they have different roles in the PSO algorithm. The term
wvi (t ) is the inertia component. The main role of this term is to keep the
particle’s movement in the same direction with the one that it was initially
heading. The term w typically ranges between 0.8 and 1.2 and can either
decelerate the particle’s inertia or accelerate it along its original direction [42].
The second term c1r1[ xˆi (t ) − xi (t )] is the cognitive component, acting as the
particle's memory, causing it to return to the regions of the search space in which
it experienced high individual fitness values. The cognitive coefficient c1 is
usually close to 2, affecting the step size of the particle when approaches to its
individual best candidate solution xˆi . The third term of Equation (6), i.e.,
c2 r2 [ g (t ) − xi (t )] , is called the social component. The social component moves
the particle to the best region that the algorithm has found so far. The social
coefficient c2 is close to 2, representing the step size of the particle when it moves
toward the global best candidate solution g(x) that the swarm has found up until
that point.
The r1 and r2 values in the cognitive and social component respectively enable
them to have a stochastic influence on the velocity update. The stochastic nature
Single and Multi-objective Optimization Methodologies in CNC Machining 209

of the two components causes each particle in the swarm to move in a semi-
random manner greatly influenced in the directions of the particle’s individual
best solution and hence, the swarm’s global best solution.
The velocity clamping technique offers the ability of keeping the particles from
moving too far beyond the search domain. This is achieved by limiting the
maximum velocity of each particle [7]. If [-xmax, xmax] is a specific search space,
then velocity clamping tends to limit the velocity to the range [-vmax, vmax], where
vmax = k * xmax. The k value is a user-defined parameter and represents the velocity
clamping vector taking values in the range 0.1≤ k ≤ 1.0. It has been noticed that
the search space in many optimization issues is not centered around 0 and, hence,
the range [-xmax, xmax] is not an adequate definition of the search domain. When it
comes to such problems, one may define vmax = k * (xmax-xmin)/2. After the
calculation of the particles velocities, the positions are updated by applying the
new velocities to the particles' previous positions. Finally,

xi (t + 1) = xi (t ) + vi (t + 1) (7)

This procedure is repeated until some stopping criteria are met. Some common
stopping conditions include a specific number of iterations of the PSO algorithm,
a number of iterations since the last update of the global best candidate solution,
or a predetermined fitness value of a quality target.

5.5.2 Simulated Annealing

Simulated Annealing (SA) methodology is a stochastic approach used in a wide
range of optimization tasks especially in Engineering. Simulated Annealing was
introduced in 1953, but actually first Kirkpatrick, Gelatt and Vecchi [43] used this
approach in computer applications and related tasks. The respective methodology
mimics the way thermodynamic systems go from one energy level to another [44].
In fact, the Simulated Annealing optimization algorithm is inspired by the
metallurgical annealing process where a controlled heating and cooling procedure
of a material through a temperature T is intended to produce a uniform distribution
of crystals with the lowest possible internal energy [45, 46].
In computational terms, SA tries to approach a domain’s global optima with
predetermined constraints, traversing its different sub-optimal solutions within a
neighbourhood, starting at the highest T. The parameter T enables the random
search over the largest possible range. If the initial T and “cooling” rate are
optimally selected, the number of sub-optimal solutions may be restricted, thus
stabilizing the system and shrinking the neighborhood barriers. As neighborhood
turns smaller (T parameter tends to 0), SA algorithm degenerates from stochastic
to deterministic for the reason that only improvements are accepted. However, if
parameter T is rapidly decreased, then the algorithm clusters to a local minimum.
The algorithm is capable of escaping local minimum if this may be imperative.
This also is the major advantage of this particular approach compared to other
stochastic optimization algorithms.
210 N. Fountas et al.

Davidson and Harell [47], summarize the parameters that are needed in the
implementation of SA algorithm. These parameters are listed below.

- Set of configurations or states of the system including an initial

configuration randomly chosen.
- A generation rule for new configurations, which is usually obtained by
determining the neighborhood of each configuration and choosing the
next configuration randomly out of the current neighborhood.
- The target or the fitness function to be minimized over the search space
(analogue of the energy).
- The “cooling” schedule of the target parameter, including initial values
and rules for when and how to change it (analog to the temperature and
its decreases).
- The termination criterion that is usually based on time, fitness function
values and/or target values.

The basic Simulated Annealing Algorithm follows the 18 steps described in [48].

1. Procedure Simulated Annealing

2. begin
3. t←0
4. initialize T
5. select a current point vcurrent randomly
6. evaluate vcurrent
7. repeat
8. repeat
9. select a new point vnew in the neighborhood of vcurrent
10. if eval (vcurrent) < eval (vnew)
11. then vcurrent ← vnew
eval ( vnew ) − eval ( vcurrent )
T
12. else if random [0,1] < e
13. then vcurrent ← vnew
14. until (termination-condition)
15. T←g(T,t)
16. t ← t+1
17. until (halting-criterion)
18. end

5.5.3 Tabu Search

Tabu Search (TS) is a heuristic approach for solving optimization tasks, designed
to prevent other methods or their components from being trapped in local optima
[49, 50]. The Tabu Search approach offers the ability of solving combinatorial
optimization problems whose applications vary on graph theory and matroid
settings to general pure and mixed integer programming tasks. It is an adaptive
procedure with the ability to utilize different methodologies, such as linear
Single and Multi-objective Optimization Methodologies in CNC Machining 211

programming algorithms and specialized heuristics, which it directs to overcome

the limitations of premature convergence to local optimality [49, 50].
The basic Tabu Search Algorithm follows the steps below:

1. Procedure Tabu Search

2. begin
3. select a current point vcurrent randomly
4. evaluate vcurrent
5. vtabu ← vcurrent
6. repeat
7. evaluate every point in the neighborhood of vcurrent
8. select a new point vnew in the neighborhood of vcurrent
9. if [eval (vcurrent) > eval (vnew)] ^ vnew ∉ vtabu
10. then vcurrent ← vnew
11. vtabu ← vtabu ∪ vcurrent
12. t ← t+1
13. until (halting-criterion)
14. end

5.5.4 Ant-Colony Optimization

Ant-Colony Optimization (ACO) is a metaheuristic algorithm that tries to provide
solutions to combinatorial optimization tasks. The function of ACO algorithm
mimics the collective foraging behavior of ants. The first ACO algorithm was
introduced in 1991 by Dorigo et al., [51], as a novel nature-inspired metaheuristic
optimization approach. The vital component of ACO algorithm is the pheromone
module. This particular module samples the search domain with a probabilistic
manner [52].
The main characteristic of ACO algorithm is the explicit use of elements of
previous solutions. In fact, ACO algorithm drives a constructive low-level solution
and incorporates it in a population framework, while randomizing the construction
in a “Monte Carlo” way. A “Monte Carlo” combination of different solution
elements is suggested also by Genetic Algorithms, but in the case of ACO the
probability distribution is explicitly defined by previously obtained solution
components.
Considering the suggestions reported in [53], a combinatorial problem to be
processed by implementing an ACO algorithm may be defined over a set C = C1,
C2,.......,Cn of basic elements. A subset S of elements represents a solution of the
problem; F ⊆ 2 is the subset of feasible solutions, hence a solution S is feasible
C

only if S ∈ F . A cost function f is determined over the solution domain, as f :

2C → R , the objective being to find a minimum cost feasible solution S*, i.e. to
find S*:S* ∈ F and f(S*) ≤ f(S), ∀S ∈ F . According to this model, a set of
computational concurrent and asynchronous agents (an ant colony) moves through
states of the problem corresponding to partial solutions of the solved problem. The
212 N. Fountas et al.

colony moves by applying a stochastic local decision policy, relying on two

parameters called trails and attractiveness. By moving each ant incrementally, a
solution to the problem is obtained. When an ant completes a solution or during
the construction phase, the ant evaluates the solution and modifies the trail value
on the elements used in its solution. This pheromone information directs the
search of future ants [51, 53].
Moreover, a typical ACO algorithm includes two more mechanisms: trail
evaporation and, optionally, daemon actions. Trail evaporation decreases all trail
values over time, in order to avoid unlimited accumulation of trails over some
components. Daemon actions can be used to implement centralized actions that
cannot be performed by single ants, such as the invocation of a local optimization
procedure or the update of global information to be used for deciding whether to
bias the search process from a non-local perspective [53].
As described in [52], a typical Ant Colony Optimization Algorithm follows the
steps below:

1. Procedure Ant Colony Optimization

2. Begin
3. Input P (S, f, Q)
4. Initialize Pheromone Values (T)
5. abs←NULL
6. while (termination criteria are not met) do
7. kinter ← Ø
8. for j = 1,..., ba do
9. a ← Construct Solution (T)
10. if a (is a valid solution,) then
11. a ← Local Search (a) {optimal}
12. if (f(a)<f(abs)) or (abs = NULL) then
13. abs ← a
14. kinter ← kinter ∪ {a}
15. end if
16. end for
17. Apply Pheromone Update (T, kinter, abs)
18. end While
19. output: The best-so-far solution, abs

5.5.5 Tribes
A tribe is a sub-swarm formed by particles which have the property that all
particles inform all others belonging to the tribe (a symmetrical clique in graph
theoretical language). The concept is therefore related to the “cultural vicinity”
(information neighborhood) and not on “spatial vicinity” (parameter-space
neighborhood). It should be noted that, due to this definition, the set of informers
of a particle (its so-called i-group) contains the whole tribe but is not limited to it
[54]. Note that Tribes mechanism can be also auto-parameterized.
Single and Multi-objective Optimization Methodologies in CNC Machining 213

The principles of Tribes are: 1) the swarm is divided in tribes; 2) at the

beginning, the swarm is composed of only one particle; 3) according to tribes’
behavior, particles are added or removed; and 4) according to the performances of
the particles, their displacement strategies are adapted [54-56]. The so-called
structural adaptation rules describe the time when a particle is created or removed
and when a particle becomes the informer of another, whereas the moving
strategies indicate how particles modify their positions. The Tribes algorithm is
executed by the following steps [54].

- Swarm Initialization. If the initial iteration is t =1 then a population of

one particle and one tribe can be initialized, with random values
generated regarding a uniform probability distribution to predefined
upper and lower bounds.
- Evaluation of Each Particle in the Swarm. The Fitness (Objective
Function) value of each particle is evaluated.
- Swarm Moves. Two moving strategies (“Simple Pivot” or “Noisy
Pivot”) can be applied to move the Swarm regarding the quality of the
Particles (“Excellent”, “Good” and “Bad”).
- Setting the Adaptation Scheme. Determining the number of links in the
population, after every L/2 iterations the structure of the swarm may be
adapted by applying one of the adaptation rules.
- Stopping Criteria. The number of generations is specified for t=t+1. The
procedure is continued with the evaluation of each particle in the swarm
until a stopping criterion is met. Usually the stopping criterion may be a
maximum number of iterations or a maximum number of evaluations of
the respective objective function.

5.5.6 Hybrids of Evolutionary Algorithms

In the pursuit of EAs with optimized performance both in terms of solution quality
and of calculation cost, researchers tend to blend various Artificial Intelligence
algorithms together or to embody operators of one method into another. This leads
to new evolutionary schemes, often referred to as “hybrids”. Since there is not a
specific definition of what exactly a hybrid is, even EAs described in Section 4.2
own properties of hybrids. A necessary step in the EA hybridization is to rearrange
or formulate “non-EA” operators properly, so that they can work on populations
of solutions/individuals instead of on single solutions, if a mixed optimization
scheme is implemented for the optimization problem.
Owing to their high response speed, ANNs are commonly coupled with EAs
either as surrogates for the Objective Function or as solution classification tools
(see Section 4.2), especially when expensive evaluators should be used. In the
field of machining processes, properly trained ANNs can predict values such as
machining time, mean surface roughness, total remaining volume, etc. [57], which
otherwise would be very time-consuming and costly to extract from experiments.
This is because iterative algorithms involve a very high number of solutions, not
all of them usable or exploitable.
214 N. Fountas et al.

Mixing and combining Evolutionary Algorithms have been investigated by

researchers with approaches such as Simulated Genetic Annealing; see [58, 59],
integrated SA, Tabu search and GA; see [60], coupled GA and PSO; see [61-63],
genetic and estimation of distribution algorithms; see [64], etc. It is obvious that
hybrids can be applied in many scientific fields and have already been
implemented in optimization of CNC machining processes; see [2].
Another powerful hybrid is created using Game Theory and Hierarchical
Games coupled with EAs. These have been proven very useful in optimization
problems with vast numbers of heterogeneous process parameters [65]. Apart
from controlling the subdivision of an EA population into subpopulations and the
way these communicate with each other, it is possible to create an optimization
scheme with various EAs behaving as leading and following players. “Leaders”
establish the overall optimization strategy, while “followers” take up specific tasks
or objectives to optimize. In this way, calculation load is distributed among the
parts that the overall optimization algorithm consists of and calculations are
carried out in parallel.

5.6 Conclusions
Machining processes, especially detailed CNC machining of high quality parts,
involve a large number of process parameters. Classical optimization schemes
often fail to produce global optima owing to the enormous calculation load. Thus,
stochastic optimization methodologies offer great advantages in this field. This is
mainly due to the probabilistic nature of their operators, which do not need any
derivative information of –in most cases- unknown functions, while they have
proven very efficient in searching the solution space. Artificial Neural Networks
are a relatively simple tool in predicting machining process values after proper
training or even classifying solutions. They can be utilized either as stand-alone
methods for high speed calculations or as internal functions in other optimization
algorithms.
Genetic and Evolutionary Algorithms are very powerful optimization tools that
perform only in a fraction of the calculation time of classical methods. Provided that
a machining problem is well defined, as far as independent and dependent variables
are concerned, and appropriately constrained, they yield optimal results with low
calculation cost. This chapter gives an overview of these algorithms, as well as their
variations, such as Simulated Annealing, Tabu Search, Particle Swarm
Optimization, Ant-Colony Optimization etc. Researchers have implemented these
methods in many scientific fields, including the CNC machining field, providing
users with optimal results in varying calculation times.
Owing to the large number and the different types of process parameters,
machining optimization problems are highly sophisticated; especially when 5-axis
CNC sculptured surface milling is involved. This calls for more complex
optimization schemes, which are named “hybrids”. Hybridization of stochastic
algorithms is the procedure of blending together methods of Artificial Intelligence
or lending operators from one algorithm to another. To that extend, Game Theory
helps in declaring interactions among operators and gives them priorities, as well
Single and Multi-objective Optimization Methodologies in CNC Machining 215

as defines the scope of each part of the algorithm. Overall, producing new hybrids,
or improving the existing ones, is an open field and can push optimization
algorithm domain to a whole new level.

References
[1] Vaxevanidis, N.M., Markopoulos, A., Petropoulos, G.: Artificial Intelligence in
Manufacturing Research. In: Paulo Davim, J. (ed.) Artificial neural network
modelling of surface quality characteristics in abrasive water jet machining of trip
steel sheet, ch. 5, pp. 79–99. Nova Publishers (2010)
[2] Rao, R.V.: Advanced Modeling and Optimization of Manufacturing Processes.
Springer, London (2011)
[3] Petropoulos, P.G.: Optimal selection of machining rate variable by geometric
programming. International Journal of Production Research 11, 305–314 (1973)
[4] Sönmez, A.İ., Baykasoğlu, A., Dereli, T., Filiz, İ.H.: Dynamic optimization of multi-
pass milling operations via geometric programming. International Journal of Machine
Tools and Manufacture 39, 297–320 (1999)
[5] Kiliç, S.E., Cogun, C., Şen, D.T.: Short Note: A computer-aided graphical technique
for the optimization of machining conditions. Computers in Industry 22, 319–326
(1993)
[6] Diwekar, U.: Introduction to Applied Optimization, 2nd edn. Springer (2008)
[7] Kennedy, J., Eberhart, R., Shi, Y.: Swarm Intelligence. Elsevier, Burlington (2001)
[8] Zitzler, E., Laumanns, M., Bleuler, S.: A tutorial on evolutionary multi-objective
optimization. In: Metaheuristics for Multiobjective Optimisation, pp. 3–37. Springer
(2004)
[9] Dixit, P.M., Dixit, U.S.: Modeling of Metal Forming and Machining Processes by
Finite Element and Soft Computing Methods. Springer, London (2008)
[10] Melin, P., Castillo, O.: Hybrid Intelligent Systems for Pattern Recognition Using Soft
Computing. Springer, Berlin (2005)
[11] De Jong, K.A., Spears, W.M.: A formal analysis of the role of multi-point crossover
in genetic algorithms. Annals of Mathematics and Artificial Intelligence 5(1), 1–26
(1992)
[12] Haykin, S.: Neural networks, a comprehensive foundation. Prentice-Hall, Englewood
Cliffs (1999)
[13] Bertsekas, D.P., Tsitsiklis, J.N.: Neuro-Dynamic Programming. Athena Scientific,
Belmont (1996)
[14] Fletcher, R.: Practical Methods of Optimization. Wiley, NY (1987)
[15] Gill, P.E., Murray, W., Wright, M.H.: Practical Optimization. Academic Press,
London (1981)
[16] Levenberg, K.: A method for the solution of certain problems in least squares. m
Quarterly of Applied Mathematics 2, 164–168 (1944)
[17] Marquardt, D.: An algorithm for least-squares estimation of nonlinear parameters.
SIAM Journal of Applied Mathematics 11, 431–441 (1963)
[18] Masters, T.: Advanced Algorithms for Neural Networks: A C++ Sourcebook. John
Wiley and Sons, NY (1995)
[19] Jain, A.K., Mao, J., Mohiuddin, K.M.: Artificial neural networks: a tutorial. IEEE
Computer 29(3), 31–44 (1996)
216 N. Fountas et al.

[20] Valiant, L.: Functionality in Neural Nets. In: Proceedings of the American
Association for Artificial Intelligence, St. Paul, Minnesota, August 21-26, vol. 2, pp.
629–634 (1988)
[21] Siegelmann, H.T., Sontag, E.D.: Turing Computability with Neural Networks.
Applied Mathematics Letters 4, 77–80 (1999)
[22] Orponen, P.: An overview of the computational power of recurrent neural networks.
In: Proceedings of the 9th Finnish AI Conference - STeP 2000 (2000),
http://www.math.jyu.fi/~orponen/papers/rnncomp.ps
[23] Sima, J., Orponen, P.: Computing with continuous-time Liapunov systems. In:
Proceedings of the 33rd Annual ACM Symposium on Theory of Computing - STOC
2001, Heraklion, Crete, Greece, July 06 - 08, pp. 722–731 (2001)
[24] Bishop, C.M.: Neural Networks for Pattern Recognition. Oxford University Press,
Oxford (1995)
[25] Ripley, B.D.: Pattern Recognition and Neural Networks. Cambridge University Press,
Cambridge (1996)
[26] Montgomery, D.C.: Design and analysis of experiments, 5th edn. John Wiley and
Sons, USA (2001)
[27] Chen, S.-L., Chang, C.-C., Chang, C.-H.: Application of a neural network for
improving the quality of five-axis machining. Proceedings of the Institution of
Mechanical Engineers, Part B: Journal of Engineering Manufacture 214(1), 47–59
(2000)
[28] Karpat, Y., Özel, T.: Multi-objective optimization for turning processes using neural
network modeling and dynamic-neighborhood particle swarm optimization.
International Journal of Advanced Manufacturing Technology 35(3-4), 234–247
(2007)
[29] Davim, J.P., Gaitonde, V.N., Karnik, S.R.: Investigations into the effect of cutting
conditions on surface roughness in turning of free machining steel by ANN models.
Journal of Materials Processing Technology 205, 16–23 (2008)
[30] Pontes, F.J., Ferreira, J.R., Silva, M.B., Paiva, A.P., Balestrassi, P.P.: Artificial neural
networks for machining processes surface roughness modeling. International Journal
of Advanced Manufacturing Technology 49(9-12), 879–902 (2010)
[31] Goldberg, D.E.: Genetic Algorithms in Search. Addison-Wesley, Reading (1989)
[32] Holland, J.: Adaptation in Natural and Artificial Systems, 2nd edn. The MIT Press,
Massachusetts (1992)
[33] Fogel, D.B.: Phenotype, Genotype and Operators in Evolutionary Computation. The,
IEEE International Conference on Evolutionary Computation, 193–198 (1995)
[34] Hinterding, R.: Mapping, Order-independent Genes and the Knapsack Problem. In:
Proceedings of the 1st IEEE Conference on Evolutionary Computing, vol. 1, pp. 13–
17 (1994)
[35] Tamaki, H., Kita, H., Shimizu, N., Maekawa, K., Nishikawa, Y.: A Comparison
Study of Genetic Codings for the Travelling Salesman Problem. In: The 1st IEEE
Conference on Evolutionary Computing, Florida, vol. 1, pp. 1–6 (1994)
[36] Bui, T.N., Moon, B.: A New Genetic Approach for the Travelling Salesman Problem.
In: Proceeding of the 1st IEEE Conference on Evolutionary Computing, vol. 1, pp. 7–
12 (1994)
[37] Syswerda, G.: A Study Reproduction in Generational and Steady-State Genetic
Algorithms. In: Rawlins, G.J.E. (ed.) Foundations of Genetic Algorithms, pp. 94–101.
Morgan Kaufmann Publishers, San Mateo (1991)
Single and Multi-objective Optimization Methodologies in CNC Machining 217

[38] Manderick, B., Spoessens, P.: Fine-grained parallel Genetic Algorithms. In: The 4th
International Conference on Genetic Algorithms, Virginia, pp. 428–433 (1991)
[39] Wang, Z.G., Rahman, M., Wong, Y.S., Sun, J.: Optimization of multi-pass milling
using parallel genetic algorithm and parallel genetic simulated annealing.
International Journal of Machine Tools and Manufacture 45(15), 1726–1734 (2005)
[40] Kennedy, J., Eberhart, R.: Particle swarm optimization. In: Proceedings of the IEEE
International Conference on Neural Networks, IV, pp. 1942–1948 (1995)
[41] Poli, R., Kennedy, J., Blackwell, T.: Particle swarm optimization - An overview.
Swarm Intelligence 1(1), 33–57 (2007)
[42] Shi, Y., Eberhart, R.: A modified particle swarm optimizer. In: Proceedings of the
IEEE International Conference on Evolutionary Computation, pp. 69–73 (1998)
[43] Kirkpatrick, S., Gelatt Jr., C., Vecchi, M.: Optimization by Simulated Annealing.
Science 220, 671–680 (1983)
[44] Fleischer, M.: Simulated Annealing: Past, Present, and Future. In: Alexopoulos, C.,
Kang, K., Lilegdon, W., Goldsman, G. (eds.) Proceedings of the, Winter Simulation
Conference, pp. 155–161. ACM (1995)
[45] Ryan, C.: Evolutionary Algorithms and Metaheuristics. In: Meyers, R.A. (ed.)
Encyclopedia of Physical Science and Technology, 3rd edn., pp. 673–685. Elsevier
(2001)
[46] Dréo, J., Pétrowski, A., Siarry, P., Taillard, E.: Metaheuristics for Hard Optimization:
Methods and Case Studies. Springer, Berlin (2006)
[47] Davidson, R., Harel, D.: Drawing Graphs Nicely Using Simulated Annealing. ACM
Transactions on Graphics 15(4), 301–331 (1996)
[48] Michalewicz, Z., Fogel, D.: How to Solve It: Modern Heuristics, 2nd edn. Springer,
Berlin (2004)
[49] Glover, F.: Tabu Search, Part I. ORSA Journal on Computing 1(3), 190–206 (1989)
[50] Glover, F.: Tabu Search, Part II. ORSA Journal on Computing 2(1), 4–32 (1990)
[51] Dorigo, M., Stützle, T.: The ant colony optimization metaheuristic: Algorithms,
applications and advances. In: Glover, F., Kochenberger, G. (eds.) Handbook of
Metaheuristics, Kluwer Academic Publishers (2002)
[52] Dorigo, M., Blum, C.: Ant Colony Optimization Theory: A survey. Journal of
Theoretical Computer Science 344, 243–278 (2005)
[53] Maniezzo, V.: Exact and approximate nondeterministic tree-search procedures for the
quadratic assignment problem. INFORMS Journal of Computing 11(4), 358–369
(1999)
[54] Dos Santos Coelho, L., Alotto, P.: Tribes Optimization Algorithm Applied to the
Loney’s Solenoid. IEEE Transactions on Magnetics 45(3), 1526–1529 (2009)
[55] Chen, K., Li, T., Cao, T.: Tribe-PSO: A novel global optimization algorithm and its
application in molecular docking. Chemometrics and Intelligent Laboratory
Systems 82(1-2), 248–259 (2006)
[56] Cooren, Y., Clerc, M., Siarry, P.: Tribes-A parameter-free particle swarm
optimization. In: The 7th EU/Meet, Adaptation, Self-Adaptation, Multi-Level
Metaheuristics, Paris, France (2006)
[57] Krimpenis, A., Vosniakos, G.C.: Optimisation of roughing strategy for sculptured
surface machining using genetic algorithms and neural networks. In: CD Proceedings
8th International Conference on Production Engineering, Design and Control,
Alexandria, Egypt, December 27-29 (2004), paper ID: MCH-05
[58] Mahfoud, S.W., Goldberg, D.E.: Parallel recombinative simulated annealing: A
genetic algorithm. Parallel Computing 21(1), 1–28 (1995)
218 N. Fountas et al.

[59] Wong, K.P., Wong, Y.W.: Genetic and genetic/simulated-annealing approaches to

economic dispatch, Generation, Transmission and Distribution. IEEE
Proceedings 141(5), 507–513 (1994)
[60] Mantawy, A.H., Abdel-Magid, Y.L., Selim, S.Z.: Integrating genetic algorithms, tabu
search, and simulated annealing for the unit commitment problem. IEEE Transactions
on Power Systems 14(3), 829–836 (1999)
[61] Esmin, A.A.A., Lambert-Torres, G., Alvarenga, G.B.: Hybrid Evolutionary
Algorithm Based on PSO and GA Mutation. In: Proceedings of 6th International
Conference on Hybrid Intelligent Systems, Brazil, 57I-62 (2006)
[62] Shi, X.H., Wan, L.M., Lee, H.P., Yang, X.W., Wang, L.M., Liang, Y.C.: An
improved genetic algorithm with variable population-size and a PSO-GA based
hybrid evolutionary algorithm. Machine Learning and Cybernetics 3, 1735–1740
(2003)
[63] Grimaccia, F., Mussetta, M., Zich, R.E.: Genetical Swarm Optimization: Self-
Adaptive Hybrid Evolutionary Algorithm for Electromagnetics. IEEE Transactions
on Antennas and Propagation 55(3), 781–785 (2007)
[64] Peña, J.M., Robles, V., Larrañaga, P., Herves, V., Rosales, F., Pérez, M.S.: GA-EDA:
Hybrid Evolutionary Algorithm Using Genetic and Estimation of Distribution
Algorithms. In: Orchard, B., Yang, C., Ali, M. (eds.) IEA/AIE 2004. LNCS (LNAI),
vol. 3029, pp. 361–371. Springer, Heidelberg (2004)
[65] Krimpenis, A., Vosniakos, G.C.: Rough milling optimisation for parts with sculptured
surfaces using Genetic Algorithms in a Stackelberg Game. Journal of Intelligent
Manufacturing 20(4), 447–461 (2009)
6

Numerical Simulation and Prediction of Wrinkling

Defects in Sheet Metal Forming

M.P. Henriques, T.J. Grilo, R.J. Alves de Sousa, and R.A.F. Valente

GRIDS Research Group

Centre for Automation and Mechanical Technology (TEMA)
Department of Mechanical Engineering
University of Aveiro
3810-193 Aveiro, Portugal
{marisaphenriques,a38296,rsousa,robertt}@ua.pt

The goal of the present work is to analyse distinct numerical simulation strategies,
based on the Finite Element Method (FEM), aiming at the description of wrinkling
initiation and propagation during sheet metal forming. From the FEM standpoint,
the study focuses on two particular aspects: a) the influence of a given finite ele-
ment formulation as well as the numerical integration choice on the correct predic-
tion of wrinkling in walls and flange zones of cup drawing formed parts; and b)
the influence of the chosen anisotropic constitutive model and corresponding pa-
rameters on the correct prediction and propagation of wrinkling deformation
modes during forming operations. In this sense, this work infers about the influ-
ence of accounting for distinct planar anisotropy behaviours within numerical
simulation procedures. Free and flange-forming examples will be taken into con-
sideration, with isotropic and anisotropic material models. Additionally, the influ-
ence on wrinkling onset and propagation as coming from different numerical
formulations will be accounted for shell and tridimensional continuum finite ele-
ments, along with implicit numerical solution procedures. Doing so, the present
work intends to provide some insights into how numerical simulation parameters
and modelling decisions can influence FEM results regarding wrinkling defects in
sheet metal formed parts.

6.1 Introduction
Over the last years investigators have paid special attention to wrinkling defects in
products coming from sheet metal-forming processes, particularly from the point
of view of the numerical simulation and prediction of such problems in metallic
parts. Nevertheless, when compared to other common defects in plastically formed
220 M.P. Henriques et al.

components (such as springback and thinning), wrinkling onset and propagation

during forming, in both single and multi-stage forming, is still an open problem
from the point of view of robust numerical modelling.
From the fundamental point-of-view, wrinkling in metallic sheets or tubes sub-
jected to plastic forming operations can be treated as localised buckling and, to
some extent, can be dealt with in a similar way as classical buckling structural
problems. This is, in fact, the basis of some analytical approaches, within the gen-
eral theory of plastic bifurcation and loss of uniqueness, as can be seen in the clas-
sical works of [1] and [2].
From the bifurcation theory, at the onset of wrinkling two possible, physically-
based, solutions can arise for equilibrium: the unwrinkled state and the wrinkled
state, characterizing a non-unique solution at this point. The basic idea is that, for
an unperturbed structure, wrinkling will start if the solution for the energy of the
structure is not unique (i.e., a bifurcation point arises). After this bifurcation point,
wrinkles may appear or the unwrinkled state may hold until the next bifurcation
point. This analytical approach can be applied, as mentioned in [3], to the onset of
wrinkling on unconstrained shell elements with double curvature, submitted to a
biaxial plane stress state, and including plasticity (J2-flow theory). In this sense, it
can deal to some extent with some types of wrinkling (wrinkling on unconstrained
structures or surfaces). Doing so, analytical approaches can possibly provide use-
ful estimates of elastic-plastic buckling for elementary shapes subjected to very
simple boundary conditions. Nevertheless, the more complex the boundary condi-
tions turn - as appearing, for instance, in flange-type wrinkling of blanks subjected
to blank-holder forces during forming - the more difficult it is to provide a general
analytical approach, although being still possible for very simple geometries. The
previous references are just representative of decades of research works, and a
comprehensive study can be found, for instance, in the work of [4].
Seeking for a higher generality in the analysis, numerical simulation (mainly by
the Finite Element Method - FEM) takes place in the study of wrinkling effects in
realistic sheet metal forming problems. For these cases, complexities in contact,
loading, geometry and process parameters can completely impair the use of ana-
lytical tools. Focusing on numerical simulation procedures based on the FEM, two
approaches are common: the so-called bifurcation and non-bifurcation types of
analysis.
Non-bifurcation analysis employs models including initial imperfections, artifi-
cially introduced in the formulation as geometry, material or loading eccentricities
and/or gradients [5] and [6]. This approach can lead to reasonable results - since
real structures indeed have imperfections - but the fact that these imperfections are
in some sense imposed into the formulation can lead to an inevitable sensitivity of
the results to the imperfection level and type chosen. A kind of non-bifurcation
procedure can be also retained, with good results, using explicit solution algo-
rithms in the FEM, and thus benefiting from intrinsic round-up errors associated
with this procedure [7]. These numerical errors can therefore act as the imperfec-
tions in the conventional non-bifurcation procedure. The main problem with
explicit calculations in FEM is related to its arguable robustness in multi-stage
sheet metal forming, particularly due to a poor evaluation of stress fields during
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 221

forming. Additionally, high mesh sensitivity, typical on non-bifurcation proce-

dures, is also observed in explicit analysis [8].
Bifurcation analysis, on the other hand, assumes no induced imperfections at
the model level [9] and [10], thus benefiting from no need for the introduction of
artificial empirical imperfection sets. Doing so, in order to catch wrinkling initia-
tion, energy or geometry-based wrinkling indicators are needed, in order to point
out the zones in the model that are prone to the appearance of wrinkling. This pro-
cedure has been used in conjunction with implicit solution techniques in FEM, in
order to analyse the wrinkling behaviour of sheet metal in a more exact and rigor-
ous way, either in free-form (side-wall) and constrained (flange) wrinkling, in
both single-and multi-step forming [11]-[15].
The use of wrinkling indicators in bifurcation analysis with the FEM is also
closely related to the attempts to establish a general Wrinkling Limit Diagram
(WLD), in a sense similar to what happens to the well-known Forming Limit Dia-
gram (FLD), in sheet metal forming. However, it turns out that seeking for a gen-
eral WLD represents a much more demanding task than its FLD counterpart, as
wrinkling limit diagrams, based solely on the final strain state of the formed part,
may not be robust enough for general applications.
These difficulties are related to the fact that wrinkling onset and propagation
depend, for instance, on the current local stress state during forming, the local cur-
vature of the sheet as forming evolves and, finally, thickness values and variations
during process, turning difficult to attain a general WLD. Attempts in determining
wrinkling limit diagrams and curves, for predefined conditions, are nevertheless
available in the literature, either for side-wall wrinkling [7,[16,[17] or flange
wrinkling [18], but there is still place for research in the search of general-purpose
robust WLD. An overview of the initial attempts in establishing limit diagram
procedures can be found, for instance, in references [18] and [19].
Since wrinkling sometimes represents a transitory behaviour, adaptive meshing
techniques based on wrinkling indicators are usually employed, as it would be
computationally expensive to simply use a very refined mesh for a global wrin-
kling consideration. This local character of search for wrinkling zones in a mesh,
together with the mentioned mesh sensitivity of non-bifurcation procedures [8],
favours the use of bifurcation approaches with adaptive meshing during forming.
However, successive mesh quality measurement and refinement, with proper state
variable transfer procedures, can affect computational efficiency. Recently, alter-
native ways to improve mesh quality without refinement were used, such as the
adoption of nodal enrichments [20] or mesh-free local sets [21] superposed to
conventional meshes within the FEM [15]. Nevertheless, it was observed that the
obtained solutions for wrinkling propagation were still dependent on the proper
choice of functions to be added to the finite element mesh, in the latter method, or
to the nodal enrichment level used, in the former approach. In this sense, none of
these procedures can be truly considered as natural approaches.
From the experimental standpoint, some studies on wrinkling limits of commer-
cially pure aluminium grades (subjected to different annealing treatments), when
drawing through conical dies, were published [22[23]. From these studies, re-
searchers concluded that the onset of wrinkling depends on the blank’s geometry,
222 M.P. Henriques et al.

the metal grading and the type of dies employed. Still relying on experimental data,
it turns out to be difficult to avoid wrinkles in metal formed parts, particularly for
some materials used in automotive industry, and therefore the use of blank-holder
is important.
A study on variable blank-holder force, and its influence on wrinkling onset,
was carried out, for instance, in reference [24]. Experimental results from two dif-
ferent shapes binder (a flat and cone-shaped ones) were compared with results
coming from the finite element method and verified. Even though more researches
were performed in the same study, only rigid shell modelling was used for the
numerical simulation models of the deformable sheet. In this study a box speci-
men was the chosen shape, which, contrasting to cylindrical and conical cups,
needed different amount of material to flow on different regions of the binder. To
simulate the binder in the variable blank-holder force forming process, three mod-
els were built: flat-shaped binder with elastic solid elements, cone-shaped binder
with also elastic solid elements and simplified segment binder with rigid shell
elements. The cone-shaped binder was shown to provide a better control over the
normal stress distribution, with a better formability of the blank in terms of the
thickness and major strain distribution [24].
Morovvati et al. [25] investigated the plastic wrinkling of a circular two-layer
blank to obtain the minimum blank-holder force required to avoid wrinkles, theo-
retically calculated by means of the energy method. The blank-holder force was de-
pendent on the material properties and geometry, and the study concluded that a
lower (a/b) ratio (where a is the punch plus die edge radius, while b is the blank ra-
dius) tends to increase the minimum blank-holder force required to avoid wrinkles.
As a consequence, for a certain blank diameter, an increase of the punch diameter
will decrease the blank-holder force. The study also concluded that the resultant
yield stress of the circular two-layer blank was a function of the yield stress of the
blank’s components. Additionally, the effect of anisotropy was investigated for three
different cases: a) the same material orientation for the two layers; b) a 45º differ-
ence in direction for the two materials’ orientation; and c) a 90º difference in that di-
rection. The results showed that, for the materials used, the minimum blank-holder
force required has reduced by about 20% from case a) to case c).
An investigation carried out by Stoughton and Yoon [26] tried to find an efficient
method for analysis of necking and fracture limits for sheet metals, combining a
model for the necking limit with fracture limits in the principal stress space by apply-
ing a stress-based forming limit curve and the maximum shear stress criterion. The
fracture model studied was applied on the opening process of a food can. Previous
studies in this area demonstrated that stress-based forming limit curves calculated di-
rectly from the strain-based forming limit curve are substantially less sensitive to
changes in strain path that normally happens in forming processes in industry. A new
failure model was presented in that reference, taking into consideration the stress dis-
tribution through thickness direction and localised necking prior to fracture, being
also capable of distinguishing forming processes where fracture could occur without
necking. Also recently, a new method to test formability, and evaluating various
modes of deformation, was investigated by Oh et al. [27]. The test consisted of three
steps: drawing a blank-holder force vs. punch stroke diagram, measuring the strain
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 223

level at the optimum condition on the stamped part and, finally, grading the test mate-
rials using a formability index. A new tool shape was designed and numerically simu-
lated to optimise the dimensional details. The blank-holder force vs. punch stroke
diagram had three failure loci to better evaluate the formability of the new tool and
also to help find the optimum process condition and formability index. The numerical
simulations were compared with experimental results and strain distribution in the
forming limit diagram.
In reference [28], a user-defined material (UMAT) – accounting for an anisot-
ropic material model based on non-associated flow rule and mixed isotropic-
nonlinear kinematic hardening – was studied and implemented into the commercial
finite element code ABAQUS. Two different forming processes were modelled: a
cylindrical drawing and a channel drawing processes, in order to estimate the capa-
bility of the constitutive model in predicting earing, springback and sheet metal ani-
sotropy effects. The achieved results demonstrated that applying the non-associated
mixed-hardening material model both anisotropy and hardening descriptions signifi-
cantly improved the prediction of earing in the cup drawing process and the predic-
tion of springback in the side wall of drawn channel sections, even though a simple
quadratic constitutive model and a single-backstress kinematic hardening model
were used [28]. Although focusing on important geometric dimensional defects, the
previous references lack in carrying out an in-depth study of wrinkling effects in the
formed parts.
Although the onset of wrinkling takes place when the ratio of strain increments
(dεr/dεθ) or the ratio of strain (εr/,εθ) reaches a critical value during forming, an at-
tempt was made to try to find a theory that predicted wrinkling based on results
obtained in the form of wrinkling limit diagram [29]. An aluminium alloy was
studied, for four different annealing treatments, and it was found that the annealed
sheet with higher n-values, R-value and UTS/σy ratios showed improved resistance
against wrinkling, along with a clear curve separating the safe and wrinkling re-
gions of service.
In reference [30], a study of different shapes for dies and blank-holders was
performed in order to observe the distribution of the blank-holder force, the punch
load at different drawing depths as well as the blank’s thickness reduction during
forming. The main goal of this investigation was to attempt to increase the deep
drawing ratio along with decreasing the blank-holder forces involved. On the other
hand, Port et al. [31] focused on surface defects of an industrial upper corner of a
front door panel, as well as in an initially planar L-shaped part designed on pur-
pose to reproduce at a small-scale surface defects after flanging. The simplified
geometry was measured using a tri-dimensional measuring machine. The re-
searchers achieved a good correlation between experiments and simulations, con-
cerning the spatial position of the defects, and a buckling analysis during spring-
back showed that the position of the defects effectively corresponded to a buckling
mode.
In the present contribution, and within the numerical and experimental frame-
work described, a sensitivity analysis of different numerical simulations, as well
as the robustness of the respective results in characterising wrinkling defects, is
carried out. The primary variables to be taken into account are related to numerical
224 M.P. Henriques et al.

simulation models only and correspond to (i) distinct mesh densities; (ii) different
finite element formulations; and (iii) basic and complex anisotropic constitutive
models. It is seen that completely different quantitative and qualitative solutions
(and, therefore, distinct wrinkling predictions) can be obtained for free and flange-
forming examples with small perturbations or variations in the chosen input pa-
rameters, in a somewhat more severe way than the one that occurs in springback
simulations. Also, and most noticeable, the correct prediction of wrinkling onset
and propagation is more related to a given finite element formulation than to com-
plex or more elaborate non-quadratic anisotropic constitutive models. All the
analyses performed in this work were carried out using the finite element com-
mercial package Abaqus/Standard [32], using fully implicit solution procedures,
for both shell and solid elements. The anisotropic constitutive models adopted
are not limited to those available in this commercial software, but also include
more recent ones, implemented by the authors by means of user subroutines
(UMAT).

6.2 Constitutive Isotropic and Anisotropic Models

Isotropic as well as anisotropic planar behaviours are considered in the following,
and in both cases an isotropic hardening evolution was taken into account, relating
( )
the effective stress (σ ) and equivalent plastic strain ε p .
Related to the numerical models for the characterisation of strain hardening ef-
fects, two isotropic hardening models were considered in the simulations. The first
model is described by a Swift's law, in the form

σ = K(ε Y + ε p )n , (6.1)

where K is a material constant, ε Y is the elastic strain at the yield state, n is the
strain-hardening exponent and ε p is given as

ε p = ε − ε e, (6.2)

where ε and ε e represent the logarithmic total and elastic strain terms, respec-
tively. On the other hand, the second model considered follows a Voce's law and
is written in the form

σ = σ Y + Rsat [1− exp(−Cr ε p )] (6.3)

where σ Y is the uniaxial yield stress, Cr is a material constant and Rsat is given
by

Rsat = σ sat − σ 0 . (6.4)

Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 225

In the present work, the anisotropic plastic behaviour was initially described by
means of the Hill's quadratic yield criterion in its original version of 1948 [33],
which represents a generalisation of the von Mises isotropic yield function, being
expressed by a yield function φ = φ (σ ) in the form

φ = F(σ 22 -σ 33 )2 + G(σ 33 -σ 11 )2 + H(σ 11 -σ 22 )2 + 2Lσ 232 + 2Mσ 312 + 2Nσ 122 . (6.5)

In this equation, the coefficients F, G, H, L, M and N are anisotropic constants

coming from experimental test data. These coefficients can be defined as

(σ 0 )2  1 1 1  (σ 0 )2  1 1 1 
F=  + 2 − 2 , G = + 2 − 2 
2  σ 22 σ 33 σ 11 
2
2  σ 332
σ 11 σ 22 
2
(σ 0 )2  1 1 1  3 τ0 
H= + − , L = (6.6)
2  σ 112 σ 22
2
σ 332  2  σ 23
2 

2 2
3 τ0  3 τ0 
M=  2 , N=  2
2  σ 13  2  σ 12 

For the determination of these parameters, σ Y is the normal yield stress stated
before, while τ Y is the corresponding yield stress on shear. The six anisotropic
parameters involved in the last equation can be determined by three uniaxial ten-
sion tests, performed at 0º, 45º and 90º directions, respecting the rolling direction
(RD). An alternative way of defining the anisotropic criterion is by means of the
so-called Lankford's r-values ( rθ ) for a specific (θ ) direction, in the form

H 2N − (F + G) H
r0 = , r45 = , r90 = . (6.7)
G 2(F + G) F

In order to perform the simulations in Abaqus/Standard software [32], it is also

necessary to transform these three coefficients into the equivalent anisotropic yield
stress ratios ( R11, R22 , R33, R12 , R13, R23 ) , required parameters for the input files of
that FEM package. In planar anisotropy (plane stress conditions), r0 and r90 coef-
ficients are generally different, and so are the ( R11, R22 , R33 ) . Assuming σ Y in the
metal plasticity model to be equal to σ 11 , then it is nevertheless valid to assume
that ( R11 = 1) , and therefore

r90 (r0 +1) r (r +1) 3(r0 +1)r90

R22 = , R33 = 90 0 , R12 = . (6.8)
r0 (r90 +1) (r0 + r90 ) (2r45 +1)(r0 + r90 )
226 M.P. Henriques et al.

Since the R13 and R23 coefficients refer to the thickness direction, and along
this direction an isotropic behaviour is assumed (normal isotropy), it comes that
R13 = R23 = 1.
The anisotropic yield criterion of Hill, on its version of 1948 as described be-
fore, is known to be well suited to the generality of steel alloys. Nevertheless, it
provides poor results when characterising the behaviour of aluminium alloys
sheets [34]. To fulfil this requirement in some of the following benchmark prob-
lems, non-quadratic anisotropic constitutive models were also implemented as
user subroutines in Abaqus commercial software. For the sake of completeness,
the implemented models will be summarised in the following.
The non-quadratic yield criteria described in the present work are suited for the
description of anisotropic effects in aluminium alloys, and although a large num-
ber of yield criteria for this purpose exist in the literature (see, for a comprehen-
sive survey on this topic, reference [34]), in the following only the criteria devel-
oped in the last decades by Barlat and co-workers (and particularly the 1991 [35]
and 2004 [36] versions) will be described. The reason for the specific choice of
those two yield criteria (Yld91 and Yld2004-18p, respectively) is related to their
ability to be numerically implemented in a three-dimensional framework, in oppo-
sition to other criteria that implicitly impose plane-stress conditions in the base
formulation. Both three-dimensional criteria were afterwards implemented in
Abaqus commercial finite element software, as UMAT subroutines [37].
The Yld91 criterion is a generalisation of the isotropic criterion of Hershey [38]
to anisotropic materials, with the anisotropic behaviour being considered in the
formulation by replacing the principal values of the stress tensor by the principal
values of an alternative tensor coming from linear transformations over the origi-
nal stress fields. The anisotropy effects are subsequently described by the coeffi-
cients present in the linear transformation operator.
The fourth order linear operator ( L) therefore appears in the formulation as

S = LT
 σ = Lσ , (6.9)

 groups the anisotropy coefficients, where T is the operator

where the operator L
that transforms the Cauchy stress tensor (σ ) into the deviatoric stress tensor
( S = Tσ ) . After this linear transformation, the components of the modified stress
 are introduced into the yield criterion and compared to the yield stress for
field S
uniaxial case, in the form

φ = ( S1 − S2 ) + ( S2 − S3 ) + ( S3 − S1 ) = 2σ Y2k ,

2k 2k 2k
(6.10)

where k is a parameter that affects the yield surface shape [35], whereas the ani-
sotropic effects may be reproduced by the knowledge of the coefficients affecting
the operator L , in the form
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 227

 c2 + c3 −c3 −c2 0 0 0 
 
 −c3 c1 + c3 −c1 0 0 0 
 −c −c1 c1 + c2 0 0 0 
1
L= 
2
, (6.11)
3 0 0 0 3c4 0 0 
 
 0 0 0 0 3c5 0 
 0 0 0 0 0 3c6 
 
as a function of the six anisotropic coefficients ci (i = 1,, 6 ) . Therefore, and ac-
counting for the exponent in the yield equation, the criterion is characterised by 7
coefficients, and the determination of the anisotropic coefficients can be carried
out by conventional tension tests (and respective yield stress values) at 0º, 45º and
90º, and also from the yield stress value coming from a biaxial stress state, ob-
tained, for instance, from a “bulge test”. Despite the easy obtaining of these ani-
sotropic coefficients, the ability to account for 3D stress states and the easy im-
plementation of the criterion into FEM codes, the major drawback of the
formulation would be the lack of reproducing distinct r0 and r90 coefficients
when uniaxial stresses in rolling and transverse directions are almost equal [34].
Seeking for a more general yield criterion, and after a succession of more
evolved plane stress models directly applicable to thin aluminium sheets, Barlat
and co-authors [36] have presented a more evolved formulation with a yield func-
tion proven to be convex and taking into account a large number of anisotropic
coefficients (in this case, 18 parameters). According to this Yld2004-18p, suited
for aluminium alloys and its anomalies (when compared to steel alloys), the yield
function can be given in the form
3,3

 S (i1) − S (j2) = 4σ Ya ,
a
φ= (6.12)
i=1, j=1

where index i, j = 1,, 3 , whereas tensor fields S (i1), S (j2) are defined by linear ( )
transformations of the type S  ( k)  (k)
= L S , for operators L  (k) in the form

 0 − L(12k ) − L(13k ) 0 0 0 
 
 − L( k ) 0 − L(23k ) 0 0 0 
 21

 − L31k )
(
− L(32k ) 0 0 0 0 
 ( k) = 
L . (6.13)
 0 0 0 ( k )
L44 0 0 
 
 0 0 0 0 L(55k ) 0 
 
 0 0 0 0 0 L(66k ) 

Therefore, the two combined linear transformations allow for the characterisation
of anisotropy based on a total of 18 parameters, which proves to lead to a very
228 M.P. Henriques et al.

general 3D constitutive modelling. For the special case of plane stress analysis, the
criterion degenerates in a simpler version involving 14 parameters. Furthermore,
where these coefficients are equal altogether to 1.0, the criterion turns out to be
equal to the isotropic criterion of Hershey [38], and for the particular case of
L (1) = L ( 2) , then the Yld91 anisotropic criterion is obtained.
The 18 coefficients involved in the Yld2004-18p yield criterion come from a series
of experimental analyses seeking for the determination of the uniaxial yield stress
values in tension (σ Y ) , as well as the Lankford’s coefficients ( rθ ) for seven distinct
directions, in the plane of the sheet, respective to the rolling direction (at 0º, 15º, 30º,
45º, 60º, 75º, 90º), the yield stress on biaxial loading solicitations (σ b ) and the ani-
sotropy coefficient ( rc ) for the compression loading experiment of a metallic disc.
The remaining factors are related to mechanical properties in the out-of-plane
directions of the metallic sheet. Since standard experimental essays for out-of-
plane properties are quite difficult to be obtained, a simplified version of the crite-
rion, involving less material parameters, is also available (Yld2004-13p, see refer-
ence [36]), pointing to an yield function in the form
a
{
φ = S 1(1) − S (21) + S (21) − S (31) + S (31) − S 1(1) − S 1(1) + S (21) + S (31) +
a a a a a
} (6.14)
+ S 1( 2) + S (22) + S (32) = 2σ Ya
a a a

( )
where S (1), S ( 2) are now obtained by means of the transformation S ( ) = L
k  ( k )S , for
the linear operators

 0 −1 − L(131) 0 0 0 
 
 − L(1) 0 − L(231) 0 0 0 
 21

 (1) =  −1 −1 0 0 0 0  (6.15)
L  
 0 0 0 L(441)
 0 0 
 0 0 0 0 L(551) 0 
 
 0 0 0 0 0 L(661) 
 
and

 0 − L(122) − L(132) 0 0 0 
 
 − L( 2) 0 − L(232) 0 0 0 
 21

 ( 2) =  −1 −1 0 0 0 0 
L  . (6.16)
 0 0 0 L(442)
 0 0 
 0 0 0 0 L(552) 0 
 
 0 0 0 0 0 L(662) 
 
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 229

It can be seen that for this simplified anisotropic criterion a lower number of anisot-
ropic coefficients (13) are needed for a full 3D model, whereas for the reduction to
plane stress problems this number turns out to be equal to 9. Although the approxima-
tion to experimental results is not completely perfect with these modified versions,
the Yld2004-13p can be a valid alternative to the original Yld2004-18p yield criterion.
More discussion and details on this can be found in references [36] and [37].

6.3 Benchmarks’ Numerical Analyses

Numerous factors can influence wrinkling in plastically formed metallic parts, such
as the material properties, friction, lubrication, the tool and blank geometries as well
as the holding conditions imposed to the sheet to be formed. Although some re-
searchers have experimentally investigated wrinkling effects and their relation to the
process parameters, the present work will focus on the influence of decisions in the
numerical modelling of the process and the consequent prediction of wrinkling ini-
tiation and propagation. The key idea is to infer about the robustness of numerically
obtained solutions respective to the numeric input to FEM codes, by the user.
To achieve this goal, several numerical simulations were performed with the FEM
code Abaqus/Standard ® [32], and attention was focused on the influence of mesh
densities, finite element formulations (shell, solid and “solid-shell” element types) as
well as constitutive material models (isotropic and anisotropic formulations) on the
overall quality of the numerical results, against experimental references.
The following sections describe the most important features employed in the
numerical simulations, for a set of examples involving unconstrained or free
(conical cup) as well constrained or flange (cylindrical cup) forming.

6.3.1 Free Forming of a Conical Cup

The main idea of this test case is that when a conical die is used, there is no need
to apply a blank-holder or clamping ring in the process of forming. Also, the coni-
cal form of the die, together with the absence of blank-holder or clamping devices,
leads easily to the formation of unconstrained wrinkling modes in the plastically
formed parts. Experimental results for this test case can be obtained, for instance,
in the original paper of Narayanasamy and Sowerby [39], being further described
in later references of the authors ([40[41]). On the other side, examples of numeri-
cal modelling and analysis for this specific problem can also be found in refer-
ences [16[17], for instance.
Following reference [39], the experimental onset is represented in Figure 6.1,
where can be seen the conical die as well as the cylindrical punch configurations
as proposed by Narayanasamy and Sowerby. The metallic blank sheet to be
formed has a circular drawing, and the plastic-forming process is defined as being
completed as soon as the entire sheet is drawn into the die.
Following the indications and data from the previous experimental references, the
diameter of the metallic sheet to be formed is taken equal to Ø = 104.94 mm, with a
thickness value equal to t = 1.90 mm. This specific combination of diameter and
thickness values is prone to the formation of a predefined number of wrinkling
230 M.P. Henriques et al.

zones along the perimeter of the circular blank, after forming, as experimentally
verified in references [39[41]. From the point of view of geometric and FEM model-
ling, the tools (the punch and conical die) were considered as rigid bodies. The tools'
dimensions follow those represented in Figure 6.1, that is: 49 mm for the punch di-
ameter, 5 mm on the punch's radius and 260 mm of height, while the die has a 19º
angle and an opening diameter of 54 mm. A detailed view of the tools involved can
be inferred from Figure 6.2.
The punch stroke, responsible to form the final metallic part, was taken as equal
to 60 mm. The metallic circular blank is therefore the only part to be considered as
deformable, and being meshed by distinct configurations of finite elements, as
seen in the following sections. Due to symmetry reasons, only one-quarter of the
total blank will be discretized by finite elements.

Fig. 6.1. Dimensions of the tools for the free-forming example, from references [39[41].

Fig. 6.2. Detailed description of the relevant dimensions for the tools in the conical cup
drawing problem (dimensions in millimetres).
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 231

Regarding the constitutive modelling for the application of the anisotropic yield
criterion of Hill (1948) [33], and following the references before, it was initially
considered an aluminium alloy with anisotropy coefficients ( rθ ) with values
r0 = 0.17, r45 = 0.58, r90 = 0.46 , and, from equation (6.8), leading to values of
R12 = 1.09, R22 = 1.47, R33 = 0.92 . The elastic part of the constitutive behaviour is
characterised by a Young's modulus and Poisson's ratio as E=69.0 GPa and υ=0.3,
respectively. The effective plastic stress-strain relationship used in the numerical
simulations is considered to be given by a Swift's law, with as main parameters
K = 127.83 MPa, ε Y = 0.0003 and n = 0.03 .
On the other hand, for the case of modelling this problem by means of the
Yld91 and Yld2004-18p constitutive criteria, parameters available in the literature
(experimentally obtained) for a 2090-T3 aluminium alloy were considered in the
numerical simulations. The strain-based hardening model is considered to follow
the parameters stated before for Swift’s law, while in both cases a friction coeffi-
cient of μ = 0.15 was assumed between all parts in contact. Furthermore, the
forming process to be simulated involved only one work stage for achieving the
final shape.
Since the tools are modelled as discrete rigid bodies, they were meshed by rigid
elements in Abaqus/Standard FEM program. The FE mesh of the punch was com-
posed of 5925 rigid finite elements of type R3D4 (4-node, 3-D bilinear quadrilat-
eral, rigid element), while the die was discretized by 1770 elements of the same
type. More information on this kind of FE formulation can be found in the pro-
gram manual [32].
Initially, the blank to be plastically formed was assumed to be a solid shape for
modelling purposes, since the adopted thickness values cannot be considered as
extremely small compared to the overall blank dimensions and also in order to
correctly describe the double-sided contact patterns involved in the process. In this
sense, different three-dimensional finite elements were adopted, as available from
the library of the FEM software: C3D8R (8-node, tri-linear 3-D solid element, re-
duced integration, that is, 1 integration point per element), C3D8 (8-node, tri-
linear 3-D solid element, full integration, that is, 8 integration points per element)
and C3D8I (8-node, tri-linear 3-D solid element, full integration and incompatible
deformation modes) [32].
Nevertheless, in a second phase shell elements were also considered, in order it
would be possible to infer about the influence of distinct finite element formula-
tions in the obtained results. In this sense, a second group of simulations were
considered, now including thin shell elements of type S4R (4-node, bilinear shell
element, reduced integration, one integration point in the element reference plane
and multiple integration points through thickness direction) as well as of type S4
(4-node, bilinear shell element, full integration within the element), for the sake of
comparisons [32].
The blank zone to be meshed with finite elements included three mesh parti-
tions in order a more refined mesh could be obtained in its central zone, as can be
seen in Figure 6.3. The partition divides the blank into a square area in the middle
232 M.P. Henriques et al.

of the blank, where the full contact with the punch will take place. A line from the
corner of the square to the perimeter of the blank (45º) divides the rest of the parti-
tion in two parts. The adopted mesh density for both solid and shell elements can
be shown in Table 6.1.

Fig. 6.3. Adopted mesh for the blank, with a refined centre zone.

Table 6.1. Mesh information for the deformable blank.

Number of elements along

Element type thickness (Gauss points for Total number of elements
shell elements)
1 5451
C3D8R/C3D8/C3D8I 2 12312
3 21312
S4R/S4 3, 5 , 7 7500

Besides the aluminium alloys mentioned before, the numerical simulations

were also performed for a blank of stainless steel grade 301, now with a blank
thickness of 1.60 mm and two different diameters: 110 mm and 130 mm. Doing so
it would be possible to infer about the influence of different materials and geome-
tries in the final results as predicted by numerical simulations, following the ex-
perimental analysis in [41]. For these setups and material, the adopted r-values are
given as: r0 = 1.159 , r45 = 1.147 , r90 = 0.759 ( R12 = 0.88 , R22 = 0.90 ,
R33 = 0.92 ). The elastic parameters are given by the Young's modulus and Pois-
son's ratio, as E=212.0 GPa and υ=0.3 being, respectively [41]. The friction coef-
ficient between the blank and tools was given as μ=0.10.
The simulations included shell (S4R) and solid elements (C3D8, C3D8R and
C3D8I) and in addiction the “solid-shell” element SC8R, a eight-node, quadrilat-
eral in-plane general-purpose continuum shell, reduced integration with hourglass
control [32], which turns to be a finite element topologically identical to a solid
one, but with a kinematics of a shell finite element. This last continuum shell
element was also tested with 3, 5 and 7 Gauss points through thickness.
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 233

6.3.2 Flange Forming of a Cylindrical Cup

The flange-forming simulations (constrained wrinkling) performed in the present
study considered two different materials (a mild steel DDQ and an aluminium al-
loy of the series 6111-T4). Three tools in each simulation were used: blank-holder,
die and punch, whose dimensions can be seen in Fig. 6.4, along with a circular
blank, as in the previous example. Once again, and due to symmetry reasons, only
one-quarter of the whole blank was considered. In Table 6.2 the constitutive
properties considered for both materials, following reference [42] are shown.

Fig. 6.4. Dimensions of the tools used in flange - forming (picture adapted from [42]).

Initially the tests start with the reference point of the die being fixed, with the
punch being above 6 mm from the top surface of the blank, while the blank-holder
is 5 mm from that surface. The simulations start with the blank-holder coming
down 1 mm (so the gap between the blank and blank-holder is now 4 mm) and af-
terwards a vertical downward movement of 45 mm is imposed to the punch's ref-
erence point.

Table 6.2. Materials properties (mm / MPa).

Properties Mild Steel DDQ AA 6111-T4

Initial dimensions (mm) Ø 180 x 1 (thickness) Ø 180 x 1 (thickness)
Friction coefficient 0.0426 0.0096
Young modulus (MPa) 221370 70500
Poisson ratio 0.30 0.342
Yield strength (MPa) 151.70 187.53
Hardening model Swift’s law Voce’s law
Hardening C=544.27 Rsat=420.86
parameters n=0.2701 Cr=8.448
Anisotropic F=0.25649 F=0.71518
parameters G=0.31646 G=0.52798
H=0.68354 H=0.47202
N=1.20949 N=1.38115
234 M.P. Henriques et al.

Different anisotropic coefficients were calculated for the different materials

used. For the mild steel (DDQ), replacing F, G, H and N (with L and M being con-
sidered as equal to 1.5) from Table 6.2 in equation (6.7), obtaining the Lankford's
r-values and using these values in equation (6.8) it is possible to calculate the co-
efficients R22, R33 and R12 to be equal to 1.0314, 1.3211 and 1.1136, respectively
(R13 and R23 are equal to 1 for the same reason as in free - forming).
The tools have been modelled as analytical bodies in this example, and there-
fore they have no mesh, while the blank is meshed with solid and shell elements as
in the previous examples. In this case, the solid element used was the C3D8 (8-
node trilinear 3-D brick, full integration), while the finite element S4 (4-node, bi-
linear shell element, full integration) was adopted when considering the blank to
be a shell part [32].

6.4 Results and Discussions – Free - Forming of a Conical Cup

The first study to be carried out was a mesh dependency study, including three dif-
ferent mesh refinements, a coarse, base and refined mesh, each with one and two
finite elements through the thickness direction. For all cases, the same finite ele-
ment formulation (solid element, reduced integration C3D8R [32]) was consid-
ered. The material parameters, for the anisotropic case following Hill 1948 model,
were specified in the last sections. The numerical results, for the evolution of the
punch force along its displacement, can be seen in Fig. 6.5 (one element through
thickness) and Fig. 6.6 (two finite elements through thickness). Due to the reduced
formulation character of the solid element employed, the meshes in Fig. 6.5 have
one single integration point along the thickness (constant stress field), while the
meshes in Fig. 6.6 have two integration points along thickness direction (linear
stress field). Meshes with more elements through the thickness direction were also
analysed, but the obtained results were the same as those shown in Fig. 6.6. For
each graph, isotropic (von Mises yield criterion) and anisotropic (Hill 1948 yield
criterion) results are represented.
What is interesting to note in this first set of simulation is that, although the
mesh refinement in plane induces (as expected) different results, comparing
Fig. 6.5 and Fig. 6.6 it can be seen that the description of wrinkling onset (and
propagation) is strongly related to the number of integration points through the
thickness direction, rather than to the in-plane refinement. It can be seen that
a “poor” representation of the stress field along thickness (obtained with the
single integration point in that direction for meshes in Fig. 6.5) has led to the ex-
perimentally verified wrinkling pattern in the numerical solution, while the more
refined modelling of Fig. 6.6, under the same conditions (in-plane mesh, material
model), leads to a numerical solution representative of a safe, wrinkling-free,
metallic part.
It is also worth noting that the consideration (or not) of anisotropic effects does
not improve the quality of results when compared to the simple von Mises
isotropic constitutive model.
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 235

Fig. 6.5. The behaviour of the blank with different refinements, constitutive model and one
element through thickness.

Fig. 6.6. The behaviour of the blank with different refinements, constitutive model and two
elements through thickness (the same happens when using three elements along thickness).

Concerns might be raised at this time due to the fact that the Hill 1948 anisotropic
yield criterion would not be the most appropriate constitutive modelling to correctly
infer the plastic deformation of an aluminium sheet. Due to this fact, further analyses
were carried out with the more advanced Yld91 and Yld2004-18p anisotropic non-
quadratic criteria detailed before.
236 M.P. Henriques et al.

To this end, and in order to have the appropriate anisotropic parameters avail-
able, it was assumed that the aluminium alloy was a representative of the series
2090-T3. From the literature, a representative hardening law for this alloy was se-
lected [43[44] in the form

σ Y = 646.0 ( 0.025 + ε )
0.227
(MPa), (6.17)

and, following the same references, the anisotropic coefficients to be included in

the Yld91 yield criterion are set as
c1 = 1.0674, c2 = 0.8559, c3 = 1.1296,
c4 = 1.000, c5 = 1.0000, c6 = 1.2970 (6.18)
( a = 8) .
For the Yld2004-18p anisotropic constitutive model, and following reference
[45], the coefficients for the 2090-T3 aluminium alloys are defined as
c1 = −0.0698, c2 = 0.9364, c3 = 0.0791, c4 = 1.0030, c5 = 0.5247, c6 = 1.3631,
c7 = 1.0237, c8 = 1.0690, c9 = 0.9543, c10 = 0.9811, c11 = 0.4767, c12 = 0.5753,
(6.19)
c13 = 0.8668, c14 = 1.1450, c15 = −0.0792, c16 = 1.0516, c17 = 1.1471, c18 = 1.4046
( a = 8) .
Focusing in this last yield criterion, and once again fixing the finite element for-
mulation involved (reduced integration solid element C3D8R, in Abaqus commer-
cial program), for different mesh densities in the plane of the blank, but keeping just
one element (one integration point) along the thickness direction, the profile of the
wrinkling appearance in the deformed blank is represented in Figure 6.7. In the pic-
ture, RD refers to “rolling direction”, while TD refers to “transverse direction”.

Fig. 6.7. Wrinkling profiles after forming for a one-quarter area of the initial circular blank.
Results obtained with Yld2004-18p criterion, for 2739 (Mesh 1), 5368 (Mesh 2) and 8109
(Mesh 3) elements in the plane of the blank (one element, i.e., one integration point through
thickness direction).
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 237

Taken the second refinement level (Mesh 2, in Figure 6.7), but now increasing
the mesh refinement through thickness, the results for the deformed configuration
after forming, and still accounting on the Yld2004-18p anisotropic criterion, are
represented in Figure 6.8. In this graph, wrinkling profiles are shown for one, two
and three elements along thickness direction, which directly corresponds to the
same number of integration points.

Fig. 6.8. Wrinkling profiles after forming for a one-quarter area of the initial circular blank.
Results obtained with Yld2004-18p criterion, for one, two and three elements (integration
points) through thickness direction.

From the last two graphs, it can be seen that even for a more sophisticated yield
criterion, the more dominant effect in this example is the number of integration
points through thickness direction. That is, the correct wrinkling profile after
forming can only be numerically attained with a proper low order integration rule
in the out-of-plane direction of the blank.
Since all the results shown before were valid for the same finite element formu-
lation, it would be interesting to infer about the influence of distinct methodolo-
gies into the quality of the numerical results obtained. To this end, keeping the
Yld2004-18p anisotropic criterion and once again the mesh density with one ele-
ment along the thickness direction, in Figure 6.9 the deformed profile for the
C3D8R (reduced integration) element, as well as the deformed configurations ob-
tained in Abaqus by means of the formulations C3D8 (full integration) and C3D8I
(full integration, enhanced strain modes), is again represented. Once again, the
strong influence of the wrinkling appearance (or not) on the choice of the numeri-
cal integration rule adopted is visible.
238 M.P. Henriques et al.

Fig. 6.9. Wrinkling profiles after forming for a one-quarter area of the initial circular blank.
Results obtained with Yld2004-18p criterion, for reduced integrated (C3D8R) as well as
fully integrated (C3D8 and C3D8I) formulations in Abaqus.

For the sake of confirmation of the conclusions being stated until now, a final
analysis must be carried out for this example when analysed by means of the
library of solid elements in Abaqus. In this case, it might be interesting to infer
about the influence of Yld91, Yld2004-18p and an isotropic (von Mises) on
the final configuration after forming, for the mesh system correctly inducing
the wrinkling patterns. To this end, and starting from the Mesh 2 defined before
(5368 elements in the plane of the blank), with one element along the thickness
and adopting the C3D8R formulation (to ensure a single integration point in
the out-of-plane direction), the obtained profiles after forming can be seen in
Figure 6.10.
Once again, and as expected from the previous results, the numerical integra-
tion rule showed to be the dominant factor in the prediction of the correct wrin-
kling pattern in the final conical part, rather than the in-plane refinement or the
constitutive model adopted. Increase on the order of integration along thickness,
for the different yield criteria shown in Figure 6.10, will promote the complete
disappearance of the wrinkling zones in the numerical solution.
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 239

Fig. 6.10. Wrinkling profiles after forming for a one-quarter area of the initial circular
blank. Results obtained with distinct yield criteria (isotropic and anisotropic), for a reduced
integrated formulation in Abaqus.

Considering once again the Hill 1948 yield criterion, for simplicity reasons,
but now focusing on alternative shell formulations in Abaqus, will lead to the
analysis that follows. Although the relative dimensions of the blank would not di-
rectly point to a thin-shell problem, it might be useful anyway to infer about the
performance of distinct shell element formulations and the corresponding ob-
tained results. This can be seen in Figures 6.11, 6.12 and 6.13, where for the
same in-plane mesh density, distinct number of integration points was adopted
through thickness. Since dealing with shell elements, only one element is as-
sumed along thickness, and it is possible to automatically define the integration
order along that direction by an increase in the number of its integration points.
Doing so, Figure 6.11 shows the deformed configuration and the evolution of the
punch force during forming for meshes with three integration points, for both iso-
tropic and anisotropic (Hill 1948) criteria, while Figures 6.12 and 6.13 do the
same for 5 and 7 integration points in thickness direction. In each graph the varia-
tion of results as coming from a full in-plane integration rule (as in S4 shell
element) against a reduced in-plane integration rule (as present in S4R shell
element) [32] is also shown.
It can be seen from the pictures that none of the shell models was able to
correctly predict the wrinkling pattern distribution, with some of the models even
inducing a single non-physical wrinkling mode in the final obtained part (see
Figures 6.11 and 6.12 for reduced integration S4R shell element).
240 M.P. Henriques et al.

Fig. 6.11. Reaction force evolution and sheet metal deformation for elements S4 and S4R
with three Gauss points and distinct constitutive behaviours (isotropic and anisotropic).

Fig. 6.12. Deformed shape of the blank for different elements formulation (S4 and S4R)
with five Gauss points through thickness and distinct constitutive behaviours (isotropic and
anisotropic).
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 241

Fig. 6.13. Reaction force evolution and blank deformation for element S4R and S4R with
seven Gauss points and both constitutive behaviour (isotropic and anisotropic).

For the sake of completeness, Figure 6.14 shows the deformed configuration of
the conical cup inside the die after the whole displacement of the punch, while in
Figure 6.15 the respective dimensions after forming are shown.

Fig. 6.14. Deformed configuration for a solid finite element formulation (C3D8R), one
element through thickness direction and an anisotropic constitutive behaviour.
242 M.P. Henriques et al.

Fig. 6.15. Wrinkling patterns (in mm) for the solid finite element C3D8R mesh, with one
element through thickness and anisotropic behaviour.

Following reference [41], in the following the forming of a circular blank of a

distinct material (stainless steel grade 301) is considered, and anisotropic effects
are taken into account by means of the Hill 1948 yield criterion. As specified in
the last section, two distinct diameters are now analysed
(φ1 = 110 mm, φ2 = 130 mm ) , for a thickness value t = 1.60 mm. The anisotropic
coefficients, as mentioned before, are equal to r0 = 1.159 , r45 = 1.147 , r90 = 0.759
( R12 = 0.88 , R22 = 0.90 , R33 = 0.92 ).
In Figure 6.16 the evolution of the punch force against its displacement during
forming, for both diameters and yield criteria, as a result of modelling the circular
blank with a reduced integrated solid formulation (C3D8R, in Abaqus), is shown.
As expected, that a geometric variation (i.e., the diameter) would induce distinct
wrinkling patterns, while (as can be seen in the previous examples) the constitu-
tive yield model does not induce noticeable differences in the results. The de-
formed final configurations can be seen in Figure 6.17, for the same punch dis-
placement. Regarding the stress levels predicted in the final parts, the maximum
equivalent stress level for the blank with a diameter of 110 mm is about 467 MPa,
while for the blank with a diameter of 130 mm it reaches 377 MPa, with the first
geometry attaining a higher number of wrinkles.
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 243

Once again, no noticeable wrinkling patterns were attained for those meshes in-
cluding fully integrated formulations. The same happened for shell formulations
employing three or more integration points along the thickness direction.

Fig. 6.16. Evolution of the punch force during forming, for both initial geometries (diame-
ters of 110 and 130 mm), and isotropic and anisotropic (C3DR formulation, one element
through thickness direction).

Fig. 6.17. The deformed shapes (360º) for the two diameters of the steel blank.

In this particular case, and when employing the “solid-shell” formulation avail-
able in Abaqus, some differences can be seen between the results coming from the
use of isotropic or anisotropic models, although in neither case the wrinkling ten-
dency is seen. Figure 6.18, for instance, shows the results for the circular blank with
244 M.P. Henriques et al.

an initial diameter of 110 mm, while in Figure 6.19 the corresponding results for the
larger diameter of 130 mm can be seen. In both cases, distinct number of integration
points through the thickness direction was used, with no distinguishable differences
between the results.

Fig. 6.18. Deformed shape of the 110 mm steel sheet with the element SC8R for 3, 5 and 7
Gauss points through thickness and isotropic and anisotropic constitutive behaviours.

Fig. 6.19. Deformed shape of the 130 mm steel sheet with the element SC8R for 3, 5 and 7
Gauss points through thickness and isotropic and anisotropic constitutive behaviours.
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 245

6.5 Results and Discussions – Flange-Forming of a Cylindrical

Cup
For this example, focusing on a constrained forming problem leading to flange
type wrinkles, and schematically previously shown in Fig. 6.4, different partitions
were considered altogether in the meshing of the blank model, as can be seen in
Fig. 6.20 and Fig. 6.21.
In the following, two mesh refinements were adopted, including both 10 and 15
elements in each edge of each mesh area in Fig. 6.20 (for the case of solid ele-
ments in Abaqus) and a more elaborated mesh refinement procedure for shell ele-
ments, as schematically represented in Fig. 6.21.

Fig. 6.20. Mesh partitions for solid elements, for the flange-forming example.

Fig. 6.21. Mesh partitions for shell elements, for the flange-forming example.

For the simulations considering solid finite element formulations, one element
layer was taken into account, and in Fig. 6.22 and 6.23 the evolution of the
punch's force throughout the analysis until the punch stroke is completed, for re-
spectively, the first and second mesh systems considered can be seen.
The graphics lines reproduce the results obtained for both aluminium (AA 6111-
T4) and mild steel (DDQ) alloys, also accounting for isotropic and anisotropic be-
haviours. It can be seen in these graphs that the force needed in the punch is larger
for the aluminium alloy (6111-T4) than for the mild steel (DDQ). It is also shown
that for the aluminium alloy with isotropic behaviour, the punch's force is bigger
when compared with the anisotropic constitutive model. The opposite happens with
the mild steel (DDQ) where the punch's force is lower for isotropic behaviour.
246 M.P. Henriques et al.

Here, and in opposition to the group of results in the last section, the influence
between the isotropic (von Mises) and anisotropic (Hill 1948) criteria, for the ani-
sotropic coefficients defined previously is more noticeable. The results were ob-
tained with the fully integrated solid element of Abaqus library (C3D8), while the
reduced formulation showed to suffer from severe hourglass effects, mainly in the
regions where double-sided contact situations were dominant.

Fig. 6.22. Punch reaction during forming, for 10 solid elements in each edge (total the 700
elements).

Fig. 6.23. Punch reaction during forming, for 15 solid elements in each edge (total the 1235
elements).
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 247

Fig. 6.24 and Fig. 6.25 show the deformed configuration for the same mesh, but
accounting for different constitutive models, and it can be seen that, despite the
differences in the evolution of the punch force against its displacement, the final
aspects of the predicted plastically formed parts are quite similar for both isotropic
and anisotropic models.

Fig. 6.24. Deformed shape for solid finite element formulation and an isotropic constitutive
model.

Fig. 6.25. Deformed shape for solid finite element formulation and an anisotropic constitutive
model.

Focusing on the particular case of the aluminium alloy (6111-T4), and now
taking into account numerical simulations considering shell elements (S4) in
Abaqus, with full in-plane integration but five integration points along the thick-
ness directions, the evolution of the punch force during forming (for an isotropic
behaviour) can be seen in Fig. 6.26, for the two mesh densities represented before
in Fig. 6.21.
248 M.P. Henriques et al.

Fig. 6.26. Punch reaction during forming, for distinct mesh refinements and shell elements.

Distinct wrinkling patterns are formed, for this case, as can be seen in Fig. 6.27
and Fig. 6.28, the first one (coarse mesh) being non-uniform and overlapped,
while the second mesh system (refined mesh) gives rise to a coherent set of wrinkling
patterns, qualitatively in accordance with published numerical and experimental
results.

Fig. 6.27. Deformed shape for shell elements and a coarse mesh.
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 249

Fig. 6.28. Deformed shape for shell elements and a refined mesh.

It seems that for the constrained forming into a cylindrical cup, the results are
quite dependent on the element type, and not only its numerical integration type (a
situation not seen in the free - forming examples) mostly due to the more severe
contact conditions involved in the flange area.

6.6 Conclusions
This work aimed to provide a preliminary insight into the influence of finite ele-
ment formulation, finite element discretization through the thickness direction and
constitutive material modelling in the onset and propagation of wrinkling patterns
in sheet metal formed parts, as reproduced by numerical simulations based on the
Finite Element Method.
It was seen that a correct prediction of wrinkling defects is very sensitive to the
initial decisions on the modelling phase of analysis, and a conclusion coming from
this work can be stated in the sense that - more than the correct constitutive mod-
elling to be adopted - distinct finite element formulations and discretization levels
show high influence on the quality of results obtained.
Nevertheless, and concentrating on the aspects related to the mesh systems and
formulations to be adopted in a given numerical simulation, it is not yet clear what
are the specific main driving effects when considering the correct prediction of
wrinkling effects. Contrary to the correct prediction of springback effects in sheet
metal formed products, where the dominant aspect to be taken into account is
known to be the numerical integration procedure and number of integration points
along the thickness direction, it is shown in the present work that for wrinkling ef-
fects a complex conjunction of (i) in-plane mesh refinement, (ii) out-of-plane mesh
refinement (or, alternatively, increase of integration points through the thickness di-
rection) and, finally, (iii) the finite element formulation itself (shell or solid ele-
ments) have a strong influence on the obtained simulation results, rather than the
constitutive model adopted. Also, and most importantly, these conclusions seem to
be extremely dependent on the examples chosen.
250 M.P. Henriques et al.

Based on that, proposals of future work are related to the research on alterna-
tive solid-shell finite element formulations in complex wrinkling prediction, where
the main advantages of solid and shell formulations alone are gathered in the same
formulation. Doing so, the sensitivity of the results to the mesh refinement levels
would also be inferred. In particular, and trying to avoid the sensitivity to distinct
numerical integration schemes, it would be useful for the development of a wrin-
kling criterion based on the use of enhanced-assumed strain finite solid-shell ele-
ments, following previous works of the authors in this field [46 - [48].

Acknowledgments. The authors would like to acknowledge the financial support

coming from the Fundação para a Ciência e a Tecnologia (Portuguese Science
and Technology Foundation), through the grants PTDC/EME-TME/66435/2006
and PTDC/EME-PME/113835/2009.

References
[1] Hill, R.: A general theory of uniqueness and stability in elastic-plastic solids. Journal
of the Mechanics and Physics of Solids 6, 236–249 (1958)
[2] Hutchinson, J.W.: Plastic buckling. Advances in Applied Mechanics 14, 67–144 (1974)
[3] Hutchinson, J.W., Neale, K.W.: Wrinkling of curved thin sheet metal. Plastic Instabil-
ity, pp. 71-78. Presses des Ponts et Chaussées, Paris (1985)
[4] Petryk, H.: Plastic instability: criteria and computational approaches. Archives of
Computational Methods in Engineering 4, 111–151 (1997)
[5] Cao, J., Boyce, M.C.: Wrinkling behaviour of rectangular plates under lateral con-
straint. International Journal of Solids and Structures 34, 153–176 (1997)
[6] Cao, J.: Prediction of plastic wrinkling using the energy method. Journal of Applied
Mechanics - Transactions of ASME 66, 646–652 (1999)
[7] Magalhães Correia, J.P., Ferron, G.: Wrinkling of anisotropic metal sheets under
deep-drawing: analytical and numerical study. Journal of Materials Processing Tech-
nology, 155–156, 1604–1610 (2004)
[8] Kawka, M., Olejnik, L., Rosochowski, A., Sunaga, H., Maknouchi, A.: Simulation of
wrinkling in sheet metal forming. Journal of Materials Processing Technology 109,
283–289 (2001)
[9] Wang, X., Lee, L.H.N.: Post-bifurcation behaviour of wrinkles in square metal sheet
under Yoshida test. International Journal of Plasticity 9, 1–19 (1993)
[10] Wang, C.T., Kinzel, Z., Altan, T.: Wrinkling criterion for an anisotropic shell with
compound curvatures in sheet forming. International Journal of Mechanical Scienc-
es 36, 945–960 (1994)
[11] Nordlund, P.: Adaptivity and wrinkle indication in sheet metal forming. Computer
Methods in Applied Mechanics and Engineering 161, 114–127 (1998)
[12] Wang, X., Cao, J.: On the prediction of side-wall wrinkling in sheet metal forming
processes. International Journal of Mechanical Sciences 42, 2369–2394 (2000)
[13] Kim, J.B., Yang, D.Y., Yoon, J.W., Barlat, F.: The effect of plastic anisotropy on
compressive instability in sheet metal forming. International Journal of Plasticity 16,
649–676 (2000)
[14] Kim, J.B., Yoon, J.W., Yang, D.Y.: Investigation into the wrinkling behaviour of thin
sheets in the cylindrical cup deep drawing process using bifurcation theory. Interna-
tional Journal for Numerical Methods in Engineering 56, 1673–1705 (2003)
Numerical Simulation and Prediction of Wrinkling Defects in Sheet Metal Forming 251

[15] Lu, H., Cheng, H.S., Cao, J., Liu, W.K.: Adaptive enrichment meshfree simulation
and experiment on buckling and post-buckling analysis in sheet metal forming. Com-
puter Methods in Applied Mechanics and Engineering 194, 2569–2590 (2005)
[16] Magalhães Correia, J.P., Ferron, G.: Wrinkling predictions in the deep-drawing pro-
cess of anisotropic metal sheets. Journal of Material Processing Technology 128,
199–211 (2002)
[17] Magalhães Correia, J.P., Ferron, G., Moreira, L.P.: Analytical and numerical investi-
gation of wrinkling for deep-drawing anisotropic metal sheets. International Journal
of Mechanical Sciences 45, 1167–1180 (2003)
[18] Kim, Y., Son, Y.: Study on wrinkling limit diagram of anisotropic sheet metal. Jour-
nal of Materials Processing Technology 97, 88–94 (2000)
[19] Obermeyer, E.J., Majlessi, S.A.: A review of recent advances in the application of
blank-holder force towards improving the forming limits of sheet metal parts. Journal
of Materials Processing Technology 75, 222–234 (1998)
[20] Belytschko, T., Moes, N., Usui, S., Parimi, C.: Arbitrary discontinuities in finite ele-
ments. International Journal for Numerical Methods in Engineering 50, 993–1013
(2001)
[21] Belytschko, T., Lu, Y.Y., Gu, L.: Element-free Galerkin methods. International Jour-
nal for Numerical Methods in Engineering 37, 229–256 (1994)
[22] Narayanasamy, R., Loganathan, C.: Some studies on wrinkling limit of commercially
pure aluminium sheet metals of different grades when drawn through conical and
tractrix dies. International Journal of Mechanics and Materials in Design 3, 129–144
(2006)
[23] Loganathan, C., Narayanasamy, R.: Effect of die profile on the wrinkling behaviour
of three different commercially pure aluminium grades when drawn through conical
and tractrix dies. Journal of Engineering & Materials Sciences 13, 45–54 (2006)
[24] Wu-rong, W., Guan-long, C., Zhong-qin, L.: The effect of binder layouts on the sheet
metal formability in the stamping with Variable Blank Holder Force. Journal of Mate-
rials Processing Technology 210, 1378–1385 (2010)
[25] Morovvati, M.R., Mollaei-Dariani, B., Asadian-Ardakani, M.H.: A theoretical, nu-
merical, and experimental investigation of plastic wrinkling of circular two-layer
sheet metal in the deep drawing. Journal of Materials Processing Technology 210,
1738–1747 (2010)
[26] Stoughton, T.B., Yoon, J.W.: A new approach for failure criterion for sheet metals.
International Journal of Plasticity 27, 440–459 (2011)
[27] Oh, K.S., Oh, K.H., Jang, J.H., Kim, D.J., Han, K.S.: Design and analysis of new test
method for evaluation of sheet metal formability. Journal of Materials Processing
Technology 211, 695–707 (2011)
[28] Taherizadeh, A., Green, D.E., Ghaei, A., Yoon, J.W.: A non-associated constitutive
model with mixed iso-kinematic hardening for finite element simulation of sheet met-
al forming. International Journal of Plasticity 26, 288–309 (2010)
[29] Ravindran, R., Manonmani, K., Narayanasmay, R.: An analysis of wrinkling limit di-
agrams of aluminium alloy 5005 annealed at different temperatures. International
Journal of Material Forming 3, 103–115 (2010)
[30] Savaş, V., Seçgin, Ö.: An experimental investigation of forming load and side-wall
thickness obtained by a new deep drawing die. International Journal of Material
Forming 3, 209–213 (2010)
252 M.P. Henriques et al.

[31] Port, A.L., Thuillier, S., Manach, P.Y.: Occurrence and numerical prediction of sur-
face defects during flanging of metallic sheets. International Journal of Material
Forming 3, 215–223 (2010)
[32] Hibbitt, Karlsson, Sorensen: ABAQUS/Standard v.6.5 User’s manual. Habbitt
Karlsson & Sorensen, Inc., USA (1998)
[33] Hill, R.: A theory of the yielding and plastic flow of anisotropic metals. Mathematical
and Physical Sciences 193, 281–297 (1948)
[34] Habraken, A.M.: Modelling the plastic anisotropy of metals. Archives of Computa-
tional Methods in Engineering 11, 3–96 (2004)
[35] Barlat, F., Lege, D.J., Brem, J.C.: A six-component yield function for anisotropic ma-
terials. International Journal of Plasticity 7, 693–712 (1991)
[36] Barlat, F., Aretz, H., Yoon, J.W., Karabin, M.E., Brem, J.C., Dick, R.E.: Linear trans-
formation-based anisotropic yield functions. International Journal of Plasticity 21,
1009–1039 (2005)
[37] Grilo, T.J.: Study of anisotropic constitutive models for metallic sheets. MSc Disser-
tation, University of Aveiro, Portugal (2011) (in Portuguese)
[38] Hershey, A.V.: The plasticity of an isotropic aggregate of anisotropic face centered
cubic crystals. Journal of Applied Mechanics 21, 241–249 (1976)
[39] Narayanasamy, R., Sowerby, R.: Wrinkling behaviour of cold-rolled sheet metals
when drawing through a tractrix die. Journal of Materials Processing Technology 49,
199–211 (1995)
[40] Loganathan, C., Narayanasamy, R.: Wrinkling of commercially pure aluminium sheet
metals of different grades when drawn through conical and tractrix dies. Materials
Science and Engineering A 419, 331–343 (2006)
[41] Narayanasamy, R., Loganathan, C.: The influence of friction on the prediction of
wrinkling of prestrained blanks when drawing through conical die. Materials and De-
sign 28, 904–912 (2007)
[42] Alves, J.L.C.M.: Numerical Simulation of the Sheet Metal Forming Process of Metal-
lic Sheets. PhD Thesis, University of Minho, Portugal (2003) (in Portuguese)
[43] Yoon, J.W., Barlat, F., Chung, K., Pourboghrat, F., Yang, D.Y.: Earing predictions
based on asymmetric nonquadratic yield function. International Journal of Plastici-
ty 16, 1075–1104 (2000)
[44] Yoon, J.W., Barlat, F., Dick, R.E., Chung, K., Kang, T.J.: Plane stress yield function
for aluminum alloy sheets - part II: FE formulation and its implementation. Interna-
tional Journal of Plasticity 20, 495–522 (2004)
[45] Yoon, J.W., Barlat, F., Dick, R.E., Karabin, M.E.: Prediction of six or eight ears in a
drawn cup based on a new anisotropic yield function. International Journal of Plas-
ticity 22, 174–193 (2006)
[46] Valente, R.A.F., Alves de Sousa, R.J., Natal Jorge, R.M.: An enhanced strain 3D el-
ement for a large deformation elastoplastic thin-shell applications. Computational
Mechanics 34(1), 38–52 (2004)
[47] Parente, M.P.L., Valente, R.A.F., Natal Jorge, R.M., Cardoso, R.P.R., Alves de Sou-
sa, R.J.: Sheet metal forming simulation using EAS solid-shell elements. Finite Ele-
ments in Analysis and Design 42, 1137–1149 (2006)
[48] Alves de Sousa, R.J., Yoon, J.W., Cardoso, R.P.R., Valente, R.A.F., Grácio, J.J.: On
the use of a reduced enhanced solid-shell (RESS) element for sheet forming simula-
tions. International Journal of Plasticity 23, 490–515 (2007)
7

Manufacturing Seamless Reservoirs by Tube Forming:

Finite Element Modelling and Experimentation

Luis M. Alves1, Pedro Santana2, Nuno Fernandes2, and Paulo A.F. Martins1
1
IDMEC, Instituto Superior Técnico, Universidade Técnica de Lisboa,
Av. Rovisco Pais s/n, 1049-001 Lisboa, Portugal
[email protected], [email protected]
2
OMNIDEA, Aerospace Technology and Energy Systems, Tv. António Gedeão,
9, 3510-017 Viseu, Portugal
[email protected], [email protected]

This chapter introduces an innovative manufacturing process that is capable of

shaping industrial tubes into seamless, reservoirs made from different materials
and available in multiple shapes. The process offers the potential for the manufac-
turing of medium to large batches of, for instance, high pressure vessels, for a
wide variety of industrial and commercial applications. The elimination of weld
seams allows for much less non-destructive test inspection requirements with the
consequent reduction in production time and avoidance of potential failures. Spe-
cial emphasis is given to practical aspects related to the tools and techniques used
to fabricate spherical and cylindrical reservoirs (with hemispherical and semi-
ellipsoidal domes). The presentation is supported by experimental data and
numerical modelling, the latter based on independently determined mechanical
properties aimed at understanding the deformation mechanics, identifying the
formability limits and demonstrating the overall performance and feasibility of the
proposed manufacturing process.

7.1 Introduction
Large-size reservoirs, like silos and tanks, are usually fabricated from curved steel
panels joined by circumferential and meridional welds. The presence of welds
adds defects, residual stresses and geometrical imperfections due to joint mis-
matching between panels that may lead to reduction in the overall performance,
namely the buckling strength [1].
Medium-size reservoirs (diameters up to 1 m) are fabricated by joining panels
or, alternatively, by multiple-stage fabrication processes. For instance, the central
cylindrical section of medium-size cylindrical reservoirs can be fabricated by
254 L.M. Alves et al.

rolling a sheet into a cylindrical surface and then joining the two ends by meri-
dional welding [2, 3], while medium-size spherical reservoirs can be fabricated in
two-half shells by deep drawing or spinning and then joined by circumferential
welding [4].
However, conventional fabrication processes running on panel joining or two-
stage manufacturing technologies are only suitable for producing single or small
numbers of reservoirs because they involve long production lead times and are
usually not appropriate to fabricate reservoirs in other materials than steel. This
prevents conventional fabrication processes from meeting the challenges imposed
by the increased demand of small reservoirs, for a wide variety of applications
such as anaesthetic and analgesic medical systems, supplemental and emergency
oxygen needs for patients, scuba diving tanks, high altitude ‘oxygen aid’ vessels
and compressed gas reservoirs for transportation systems, high pressure gas sto-
rage systems for aeronautical and space applications and compressed air tanks for
paintball and other leisure equipment, among others.
Despite recent efforts in fabricating small-size reservoirs from stainless steel and
aluminium by casting, conventional or hydromechanical deep drawing [5] as well
as explosive forming [6] the need for other undeveloped manufacturing technolo-
gies, able to produce medium to large batches of small-size reservoirs in a wide
range of materials exists, because casting is limited to simple shapes (e.g. cylinder
liners) and its operating costs demand very high production rates, explosive forming
suffers from industrialisation problems and conventional or hydromechanical deep
drawing, although being a flexible solution, is required subsequently for joining op-
erations for the produced half-shells into a reservoir by means of welding (tungsten
inert gas or, for space applications, more frequently electron beam welding).
This chapter focuses on the above-mentioned problems and presents an innova-
tive manufacturing process for producing seamless, low-cost, axisymmetric metal-
lic reservoirs by tube forming (Fig. 7.1).

Fig. 7.1. Spherical and cylindrical reservoirs made from aluminium AA7050 and AA6063
fabricated by the proposed manufacturing process.
Manufacturing Seamless Reservoirs by Tube Forming 255

The proposed manufacturing process avoids joint mismatching and geometrical

imperfections from half-shells welding, allowing utilisation of materials other than
steel, enabling mass production but still maintaining manufacturing flexibility to
allow possible customisation requirements from smaller batch production sizes,
such as the ones usually required by specialised markets such as the space market.
In other words, the process is flexible enough to fulfil the requirements of the
growing trend for the development of manufacturing processes which can demon-
strate very short life- cycles and production lead times as well as medium devel-
opment time requirements.
The chapter is organized into five parts; the first part describes the proposed
manufacturing process, the second part presents material characterisation, the
third part introduces the tools and techniques employed in the numerical and expe-
rimental modelling of the process, the fourth part characterises the mechanics of
deformation and the influence of the major operating parameters on the workabili-
ty window of the process and the final part is centred on demonstrating the feasi-
bility of the process for fabricating small-size cylinder reservoirs made from
aluminium and stainless steel.
The overall approach draws from accumulated knowledge in tube end forming
[7] and is based on flexible tool design, experimental trials performed under la-
boratory-controlled conditions, independently determined mechanical properties
of the raw materials and process modelling using the in-house finite element com-
puter program i-form [8].
The chapter is expected to effectively contribute to transferable knowledge on
an innovative manufacturing process to produce small-size, low-cost, seamless re-
servoirs.

7.2 Innovative Manufacturing Process

7.2.1 Tooling Concept

The new manufacturing process for producing small-size, seamless, metallic res-
ervoirs is schematically shown in Fig. 7.2a. As seen, the forming operation is car-
ried out by axial pressing the open ends of a tube against two profile shaped dies,
while providing internal support by means of a low melting point recyclable man-
drel, until the desired shape of the reservoir is achieved.
The basic components of the process, which is protected by an international
patent request [9], are; (i) the upper and lower profile shaped dies, (ii) the con-
tainer, (iii) the mandrel and (iv) the tubular preform.
The upper and lower dies are the active tool components and the sharp - edge of
the upper die is protected against collapse due to circumferential tensile stresses by
means of the container which acts as a shrink fit tool part. The dies are dedicated to
256 L.M. Alves et al.

(b)

(a) (c)

Fig. 7.2. Forming a tubular preform into a seamless cylindrical reservoir with profile
shaped ends. The photograph in (b) shows the preform and the reservoir with semi-
ellipsoidal ends and the photograph in (c) shows successful and non-successful modes
of deformation that were obtained when forming the reservoir with and without internal
mandrels.

a specific outside radius of the tube r0 and its profile defines the geometry of the
reservoir. The container constrains material from flowing outwardly in order to
avoid the occurrence of buckling and helps minimizing the errors due to misalign-
ment between the tubular preforms and the individual dies. The mandrel provides
internal support to the tubular preform during plastic deformation in order to avoid
collapse by wrinkling and local instability at the equatorial region. Figure 7.3
shows an exploded view drawing and a picture of the tool with its major active
components.
Manufacturing Seamless Reservoirs by Tube Forming 257

Fig. 7.3. Tool for fabricating seamless reservoirs with profile shaped ends from tubular
preforms.

7.2.2 Preforms and Mandrels

In case of bending thin-walled tubes over tight radii, internal support with man-
drels or loose fillers is commonly utilized to prevent the collapse by flattening or
wrinkling. Mandrels are preferentially chosen for large batch sizes (mass produc-
tion) and can be rigid, flexible or articulated. The last two types are employed
when support of the tube is needed far into the bending region. Loose fillers such
as sand and low-melting alloys are mainly chosen for small-batch production.
End forming of thin-walled tubes may also require internal support. Simple op-
erations such as one-side nosing and reduction can take advantage of conventional
mandrels to prevent collapse by local instability and wrinkling. However, other
operations might not benefit from the advantages of having internal support if, for
example, the flow of material creates difficulties (or even, renders it impossible) to
withdraw the mandrels. This often limits the mass production of sound thin-walled
tubular formed parts to components having geometrical features within a compact
range.
In case of small-batch applications, difficulties of extracting the mandrels from
the tubular formed parts can, whenever the geometry is favourable, be overcome
by the use of loose fillers, expanding mandrels or state-of-the-art shape memory
mandrels [10] that fully recover their original (undeformed) shape when heated
above the transition temperature. However, if the objective is, for instance, to
manufacture spherical or cylindrical seamless reservoirs from tubular preforms by
means of the manufacturing process described in the previous section, there is a
need for a new type of internal support that combines the effectiveness of conven-
tional mandrels with the easy extraction of loose fillers.
258 L.M. Alves et al.

A possible solution is to employ ‘sacrificial polymer mandrels’ that consists

of a polymer tube shrunk into the inner diameter of a metallic tubular preform
(Fig. 7.4a). The resulting bi-material assembly ensures that the sacrificial polymer
mandrel is tightly held in position and the inner surface of the metallic thin-walled
tubular preform is under the action of tensile stresses. Since plastic deformation of
the sacrificial mandrel during forming is connected to the plastic deformation of
the metallic preform, a firm internal support is guaranteed and, therefore, wrin-
kling and local instability at the equatorial region can be avoided. However, these
mandrels are difficult to extract because their removal by melting is time consum-
ing and likely to produce toxic gases. The alternative of using solvents for remov-
ing the mandrels at the end of the forming process should not be considered be-
cause it is not environmentally acceptable.
An alternative environmental friendly solution that was recently developed by
the authors consists of employing mandrels made from low melting point alloys
comprising bismuth, lead, tin and cadmium, among other materials [11]. These
mandrels are capable of continuously adapting its shape to that of the formed tube
and are easily removed (and recyclable) by heating slightly below or above 100ºC,
while leaving the reservoir intact, at the end of the process (Fig. 7.4b).

(a)

(b) (c)

Fig. 7.4. Tubular preforms with internal mandrels made from (a) polyvinyl chloride (PVC),
(b) low melting point alloy MCP70 and an (c) aluminium alloy.
Manufacturing Seamless Reservoirs by Tube Forming 259

Typical commercial alloys utilized in mandrels made from low melting point
alloys are MCP70 and MCP137 with melting temperatures equal to 70ºC and
137ºC, respectively. The mandrels are cast and their edges deburred with slotted
angles (around 15º, Fig.7.4b) in order to avoid premature flow of the low melting
point alloy into the polar openings during forming. This would greatly increase the
compression force required at the end of the process and give rise to undesirable
material flow at the poles.
In case of forming reservoirs made from stainless steel AISI 316 it is sometimes
needed to employ internal mandrels made from aluminium alloys (Fig. 7.4c).

7.2.3 Lubrication
Forming seamless metallic reservoirs by means of the proposed manufacturing
process is consistent with the three basic mechanisms governing material flow be-
haviour in tube processing; (i) bending, (ii) compression along the circumferential
direction and (iii) friction. Bending takes place where the tubular preform contacts
the dies while circumferential compression and friction develop gradually as the
preform deforms against the profile shaped dies.
In case of friction, previous research work on tube forming put into evidence that
operating parameters giving rise to successful modes of deformation can easily lead
to unsuccessful modes of deformation if lubrication is inexistent or simply inappro-
priate [7]. In case of the process development described in this chapter, lubrication
with zinc stearate proved efficient in a wide range of operative conditions.

7.3 Mechanical Testing of Materials

7.3.1 Flow Curve

The raw materials utilized in the investigation consisted of commercial tubes of
aluminium AA6063-T0 and AA7050-T0, stainless steel AISI 316 and ingots of
commercial low melting point alloysMCP70 and MCP137. The aluminium and
stainless steel tubes were split into preforms that were formed into seamless reser-
voirs, whereas the low melting point alloys were utilized to cast the mandrels us-
ing a permanent mould by taking advantage of the good recycling properties of the
material and mould.
The stress-strain curves of the materials were determined by means of conven-
tional and stack compression tests [12] that were carried out at room temperature.
The stack compression tests made use of multi-layer cylinder specimens that were
assembled by pilling up circular discs cut from the commercial tubes. The prepa-
ration of the discs was critical for ensuring that all the layers were concentric and
had identical cross-sectional area in order to ensure homogenous deformation un-
der frictionless conditions along the contact interface with the platens. The test
specimens were machined from the supplied material stock.
260 L.M. Alves et al.

Figure 7.5 presents the stress-strain curves for the aluminium and low melting
point alloys. Two different stress responses with increasing strain are observed;
(i) the aluminium alloys present strain hardening while (ii) the low melting point
alloys present evidence of strain softening for values of strain above 0.2.

Fig. 7.5. True stress-strain curves obtained from conventional and stack compression tests
of aluminium alloys AA6063-T0, AA7050-T0 and low melting point alloys MCP70 and
MCP137.

7.3.2 Critical Instability Load

Axial compression of tubes gives rise to either symmetrical or unsymmetrical in-
stability mode of deformation. The latter, usually referred to as buckling, happens
when a tube is long and has relatively thick walls, while the former, usually re-
ferred to as local buckling, takes place when a tube, either short or long, has thin
walls.
The understanding of the physics behind local buckling of thin-walled
tubes was first given by Alexander [13] on the basis of inward and outward
movements of the tube, later improved by Allan [14] and more recently en-
hanced by Rosa et al.[15] who illustrated the contact between successive insta-
bility waves as well as the progressive changes taking place at the contact region
between the tube and the dies. The critical instability load that originates local
buckling on a thin-walled tube can be determined analytically, numerically or
experimentally [15].
Manufacturing Seamless Reservoirs by Tube Forming 261

Fig. 7.6. Experimental evolution of the load–displacement curve for the axial compression
of thin-walled AA6063T0 short and long thin-walled tubes between flat dies.

The experimental value of the critical instability load for the occurrence of lo-
cal buckling in thin-walled tubes subjected to axial loading can be determined by
compressing tubular specimens with different initial lengths between flat dies.
Figure 7.6 shows the critical instability load as a function of the displacement
of the upper flat die. As can be seen, the load increases sharply from zero and lo-
cal buckling occurs upon reaching a critical experimental value equal to 27 kN for
aluminium tubes AA6063T0 with 60 mm diameter and 2 mm thickness. The pic-
ture placed inside Fig. 7.6 shows that a diamond shaped instability prevails over
conventional axisymmetric instability when the ratio of initial tube length to di-
ameter is small (say, close to 1).

7.4 Theoretical and Experimental Background

7.4.1 Finite Element Flow Formulation

Because the experiments in forming tubes into small size, seamless, reservoirs
were performed at room temperature under a quasi-static constant displacement
rate (100 mm/min) of the upper-table of the universal testing machine, no inertial
effects on forming mechanisms were likely to occur and, therefore, no dynamic ef-
fects in the deformation mechanics were needed to be taken into account. These
operative conditions allowed numerical modelling of the manufacturing process to
262 L.M. Alves et al.

be performed with the finite element flow formulation and enabled the authors to
utilize the in-house computer program i-form that has been extensively validated
against experimental measurements of metal-forming processes since the end of
the 1980s [8].
The finite element flow formulation giving support to i-form is built upon the
following weak variational form expressed entirely in terms of the arbitrary varia-
tion in the velocity,

δ Π = ∫ σ δε dV + K ∫ ε V δε V dV − ∫ t i δu i dS = 0 (7.1)
V V SF

where V is the control volume limited by the surfaces SU and ST where velocity
and traction are prescribed, respectively, and K is a large positive constant penaliz-
ing the volumetric strain rate component ε v in order to enforce incompressibility.
The utilisation of the flow formulation based on the penalty function method
offers the advantage of preserving the number of independent variables, because
the average stress σm can be computed after the solution is reached through,
σm = K εv (7.2)

The effective stress and the effective strain rate are defined, respectively, by,

3
σ = σ ij′ σ ij′ (7.3)
2

2
ε = ε ij′ ε ij′ (7.4)
3
where σ ij′ is the deviatoric stress tensor and ε ij′ is the deviatoric strain-rate tensor.
The spatial discretization of the weak variational form by means of M finite
elements with constant pressure interpolation, linked through N nodal points,
results in the following set of nonlinear equations [8, 16],

⎧ σ ⎫
M
⎪ ⎪
∑ ∫ ⎨
m =1 ⎪V ε ∫ ∫
δv T K v dV m + K m δv T C T BvCT B dV m − δv T NTdS m ⎬ = 0 (7.5)
⎩ m
V m m
ST ⎪⎭

which can be written in the following compact form,

∑ {[σ P + K Q] {v} = {F}}

M
m
(7.6)
m =1

where,
1
P= ∫ε n −1
K dV m (7.7)
Vm

K = BT D B (7.8)
Manufacturing Seamless Reservoirs by Tube Forming 263

Q= ∫C
T
BC T B dV m (7.9)
Vm

F= ∫ NTdS
m
(7.10)
STm

The symbol N denotes the matrix containing the shape functions of the element,
B is the velocity-strain rate matrix, C is the matrix form of the Kronecker symbol
and D is the matrix relating the deviatoric stresses with the strain rates according
to the rate - form of the Levy-Mises constitutive equations.
The nonlinear set of Eq. 7.6 derived from the flow formulation based on the
penalty function approach can be efficiently solved by a numerical technique re-
sulting from the combination between the direct iteration and the Newton–Raphson
methods.
The direct iteration method, which considers the Levy–Mises constitutive equa-
tions to be linear (and therefore constant) during each iteration, is to be preferen-
tially utilized for generating the initial guess of the velocity field required by the
Newton-Raphson method. The Newton-Raphson method is an iterative procedure
based on a Taylor linear expansion of the residual force vector R(v) of the nonlin-
ear set of Eq. 7.6,

∑ {[σ P + K Q] {v} − {F} }

M
n
Rn = m n n
(7.11)
m =1

near the velocity estimate at the previous iteration,

⎡∂ R ⎤
( )
R v n ≅ R n = R n −1 + ⎢
∂
⎥ Δv = 0
n
(7.12)
⎣ v ⎦ n −1
where Δv is the first-order correction of the velocity field, the symbol n denotes
the current iteration number,

{v}n = {v}n −1 + α {Δv}n α ∈ ] 0, 1 ] (7.13)

and α is a parameter that controls the magnitude of the velocity correction term
Δv . This procedure is only conditionally convergent, but converges quadratically
in the vicinity of the exact solution.
The aforementioned numerical techniques are designed in order to minimise the
residual force vector R(v) to within a specified tolerance and control and assess-
ment is performed by means of appropriate convergence criteria.
The numerical evaluation of the volume integrals included in Eq. (7.6) is per-
formed by means of a standard discretization procedure. Due to the rotational
symmetry and as no anisotropy effects were taken into account, the finite element
models set up to replicate the experimental test cases were accomplished by
discretizing only the cross-section of the tubular preform and mandrel by means of
axisymmetric quadrilateral elements (Fig. 7.7).
264 L.M. Alves et al.

Mandrel Tube

Fig. 7.7. Finite element model of the manufacturing process. Discretization of the preform
and mandrel by means of quadrilateral elements.
Manufacturing Seamless Reservoirs by Tube Forming 265

7.4.2 Friction and Contact

Friction at the contact interface Sf between tubular specimens and tooling is as-
sumed to be a traction boundary condition and the additional power consumption
term is modelled through the utilization of the law of constant friction τf =mk,
where k is the shear yield stress in pure shear.
Implementation of friction as a traction boundary condition is performed by ex-
tending the weak variational form in Eq. 7.5 as follows [16],

⎧ σ ⎫
∫
⎪ ε K v dV + K ∫ ∫
C T B v C T B dV m − N T dS m
m m
⎪
M
⎪V m Vm STm ⎪
∑ ⎨
m =1 ⎪ ⎡ ⎤
⎬=0
(7.14)
2 Nv m⎪
∫
−1
+ mk N tan ⎢ r
⎥ dS ⎪
⎪ π ⎣ v0 ⎦
⎩ m
S FR ⎭

where vr denotes the relative sliding velocity between the tube.

The approximation of the frictional stress τf =mk by an arctangent function of
the relative sliding velocity eliminates the sudden change of direction of the fric-
tional stress at the neutral point,

⎧⎪ 2 ⎛ | u | ⎞ ⎫⎪ u
τ f = mk ⎨ arctan ⎜⎜ r ⎟⎟ ⎬ r (7.15)
⎪⎩ π ⎝ u 0 ⎠ ⎪⎭ | u r |

where v0 is an arbitrary value within the range from 10-3 to 10-4 in order to avoid
numerical difficulties.
The contact algorithm implemented in the finite element computer program
solves the interaction between the tubular specimens and tooling by means of an
explicit direct method. The algorithm requires the discretization of the tool surface
into contact–friction linear elements and is based on two fundamental procedures;
(i) identification of the nodal points located on the boundary of the mesh and (ii)
determination of the minimum increment of time Δtmin for a free nodal point lo-
cated on the boundary of the tubular preform to go in contact with the surface of
the tool. The minimum increment of time Δtmin can be computed in accordance
with the procedure described elsewhere [8].
The contact interface between tubular preforms and recyclable mandrels was
modelled by means of a nonlinear procedure based on a penalty approach. The ap-
k
proach is built upon the normal gap velocity gn for a nodal point k, contacting an
element side ab of the adjacent element, (Fig. 7.8a),

gnk = vnk − β vna − (1 − β ) vnb (7.16)

266 L.M. Alves et al.

where subscript n indicates normal direction and β and 1-β are the fractions of the
element side lab defining the velocity projection of nodal point k on the element
side ab. The penalty contact approach adds the following extra term to Eq. (7.1),
Nk
δ Πc = γ ∑g
k =1
k
n δg nk (7.17)

where Nk is the total number of contacting points and γ is a large positive constant
enforcing the normal gap velocity gn ≥ 0 in order to avoid penetration.
k

(a)

(b)
Fig. 7.8. Contact between deformable tube and mandrels. (a) Modelling the contact be-
tween nodal point k of the tubular preform (or mandrel) and element side ab of the mandrel
(or tubular preform) and (b) schematic illustration of the modifications that are performed
on the global stiffness matrix of the finite element model due to the contact between nodal
point k and element side ab.

The extra term in Eq. (7.17) gives rise to additional contact stiffness terms Kc in the
original stiffness matrix σ P + K m Q resulting from the minimization of Eq. (7.1),

K ijmn
c = γ α mα n ni n j
(7.18)
(i, j ) = 1, 2 (m, n ) = k , a, b α k = 1, α a = −β , α b = −(1 − β )
Manufacturing Seamless Reservoirs by Tube Forming 267

The positions ijmn of the contact stiffness terms in the overall stiffness matrix
are schematically illustrated in Fig. 7.8b for typical skyline storage. It is worth
noting that skyline storage usually needs to be expanded during numerical simula-
tion in order to include new contacting pairs. The penalty contact method has the
advantage of being purely geometrically based and therefore no additional degrees
of freedom have to be considered as in case of alternative approaches based on
Lagrange multipliers.
The numerical simulation of the manufacturing process was accomplished
through a succession of displacement increments each of one modelling approxi-
mately 0.1% the initial height of the test specimens. No remeshing operations
were performed and the overall CPU time for a typical analysis containing around
2500 elements was below 5 min on a standard laptop computer.

7.4.3 Experimental Development

The experimental work plan was designed in order to meet two different objec-
tives; (i) to understand the influence of the major operating parameters on the
workability window of the process and (ii) to demonstrate the feasibility of the
new manufacturing process for producing seamless reservoirs with profile shaped
ends.
The first objective was accomplished by means of an investigation centred on
the modes of deformation and strain loading while the second objective was ac-
complished by fabricating small size reservoirs with hemispherical and semi-
ellipsoidal ends and ends made from aluminium and stainless steel by means of
the proposed manufacturing process.
The experiments were performed at room temperature in a universal testing
machine under a constant cross-head displacement rate of 100 mm/min (1.7 mm/s)
and the overall work plan is listed in Table 7.1.

Table 7.1. The experimental work plan

Material Geometry Mandrel End t0 l1 r0

Shape (mm) (mm) (mm)
AA6063T0 spherical - - 1.1-2.5 - 8-50
AA6063T0 spherical MCP70 - 0.9-2.2 - 20-50
AA7050 & spherical MCP70 - 1.0-1.5 25-35
7075T0
AA6063T0 cylindrical MCP137 and spherical 1-2 0-150 25-35
MCP70
AA6063T0 cylindrical MCP137 ellipsoidal 0.8-1.5 0-150 30
AISI 316 cylindrical AA6063, spherical 0.5-3 0-175 30
MCP137 and
MCP70
268 L.M. Alves et al.

7.5 Mechanics of the Process

7.5.1 Modes of Deformation

Forming a tubular preform into a spherical reservoir by means of the proposed
manufacturing process is the result of three basic mechanisms that compete with
each other: plastic work, local bucking and wrinkling. Plastic work is caused by
compression along the circumferential direction which gradually deforms the tube
against the dies. Local buckling and wrinkling are associated with compressive in-
stability in the axial and circumferential directions and limit the overall formabil-
ity of the process by giving rise to non-admissible modes of deformation.
Figure 7.9 shows that forming a tubular preform with a deformable mandrel
made from a low melting point alloy successfully avoids collapse by local buck-
ling and wrinkling in order to produce a sound spherical reservoir. The finite ele-
ment predicted geometry of the spherical shell and mandrel at the end of the form-
ing process (Fig. 7.9b) compares very well with the experimental part shown in
Fig. 7.9a. In fact, even the small depression at the poles that is caused by the rela-
tive sliding between the mandrel and the tubular preform is effectively predicted
by finite element modelling.

Small Depression

(a) (b)

Fig. 7.9. Spherical reservoir fabricated by means of the proposed manufacturing process.
(a) Spherical reservoir and internal mandrel after being formed and (b) finite element pre-
dicted geometry at the end of the process.

Subsequent removal of the mandrel by melting, while leaving the shell intact,
installation of the upper valve and the lower end cap, polishing and painting, re-
sults in the spherical reservoir depicted in Fig. 7.1.
Manufacturing Seamless Reservoirs by Tube Forming 269

(a) (b)

Fig. 7.10. (a) Experimental and (b) finite element predicted collapse by local bucking due
to compressive instability in the axial direction.

Figure 7.10 shows the specimen and the predicted finite element geometry resulting
from an attempt to shape a tubular preform into a spherical reservoir by means of the
proposed manufacturing process without using an internal mandrel. As seen, formabil-
ity is limited by local buckling due to compressive instability in the axial direction.

7.5.2 Formability
The technique utilised for obtaining the experimental strain loading paths in the
principal strain space involved electrochemical etching of a grid of circles with
1 mm initial radius on the surface of the preforms before forming and measuring
the major and minor axes of the ellipses that result from shaping the tubes into
spherical shells. The experimental values of the in-plane strains were determined
from (Fig. 7.11),
⎛ a ⎞ ⎛ b ⎞
ε 1 = ln ⎜ ⎟ ε 2 = ln ⎜ ⎟ (7.19)
⎝ 2R ⎠ ⎝ 2R ⎠
where the symbol R represents the original radius of the circle and the symbols a
and b denote the major and minor axes of the ellipse.
The in-plane components of strain resulting from the application of the above
mentioned procedure in different locations taken along the meridional direction of
the reservoir and plotted in the principal strain space allow us to determine the
strain loading path resulting from the gradual deformation of the tubular preform
against the dies (Fig. 7.12).
The principal strain space is of major importance in the analysis of forming proc-
esses because it allows foreseeing if a loading path resulting from a manufacturing
process is likely to produce admissible or inadmissible modes of deformation. In
case of the proposed manufacturing process, measurements and finite element
270 L.M. Alves et al.

R
2
a

(a) (b)
Fig. 7.11. (a) Grid of circles that were utilized for obtaining the local values of strain and
(b) schematic deformation of a circle into an ellipse during the forming process.

predicted values of strain allow us to conclude that the strain loading path is similar
to that of pure compression. So lying thus, close to the onset of wrinkling (refer to
the photograph included in Fig. 7.12 that shows a spherical reservoir with wrinkles
at the upper end).
Under these conditions, the internal mandrel plays a key role in the proposed
manufacturing process because it is capable of avoiding collapse by local buckling
due to compressive instability in the axial direction and impeding the strain loading
path to approach the onset of wrinkling. These conclusions apply to other shapes of
reservoirs than spherical as will be seen in the following section of this chapter.

0.6
FEM
Experimental
Compression Test
Tensile Test
Meridional Strain

0.4

0.2

0.0
-1.0 -0.8 -0.6 -0.4 -0.2 0.0 0.2

Circumferential Strain

Fig. 7.12. Experimental and finite element predicted strain loading path in the principal
strain space resulting from forming a tubular preform into a spherical reservoir.
Manufacturing Seamless Reservoirs by Tube Forming 271

7.5.3 Forming Load

Figure 7.13 shows the experimental and finite element predicted evolution of the
load-displacement curve for a test case in which fabrication was successfully ac-
complished by means of the proposed forming process using a mandrel.
As seen, there is a steady monotonic increase of the forming load with the dis-
placement of the upper die. Because the ultimate forming load for producing cy-
lindrical reservoirs with semi-ellipsoidal ends made from AA6063T0 with 60 mm
diameter is below 350 kN, the process can be industrialized in a low-cost, small
capacity, press. The overall ability of the finite element computer model to
predict, not only the ultimate forming load but also the general shape of the
load-displacement curve, is very good.

350

300

250
Load (kN)

200

150

100

50 FEM
Experimental
0
0 20 40 60 80 100

Displacement (mm)

Fig. 7.13. Experimental and finite element predicted evolution of the load-displacement
curve.

7.6 Applications

High pressure reservoirs are fundamental for several industries. For terrestrial ap-
plications their role is important in markets such as transportation where they are
employed for the storage of compressed natural gas (over 11 million vehicles
worldwide) [17] and hydrogen (hailed as the ‘fuel of the future’, which according
to the DoE should translate to nearly 150 million in-circulation vehicles by 2050)
[18]. Besides transportation, high pressure reservoirs are used for scuba diving ap-
plications, professional paint ball nitrogen high pressure bottles, etc.
272 L.M. Alves et al.

7.6.1 Performance and Feasibility of the Process

Figure 7.14 shows the initial finite element discretization of the tubular preforms
together with the computed predicted geometry of the reservoirs at the end of the
process. The reservoirs have 60 mm diameter and are made from AA6063T0 at
the end of the process.
As seen, forming with an internal mandrel allows fabricating sound reservoirs
whereas forming without a mandrel will inevitably lead to a disastrous result similar
to that previously obtained when attempting to fabricate spherical reservoirs without
a mandrel (refer to Fig. 7.10). This result further confirms the key role played by in-
ternal mandrels in the overall success of the proposed manufacturing process.

Fig. 7.14. Forming tubular preforms into cylindrical reservoirs with semi-ellipsoidal ends.
Discretization of the tubular preform and mandrel (if exists) by finite elements and com-
puted predicted geometry at the end of the process.
Manufacturing Seamless Reservoirs by Tube Forming 273

The finite element predicted distribution of effective stress at intermediate and

final stages of the forming process with an internal mandrel is shown in Fig. 7.15.

Fig. 7.15. Forming a tubular preform into a cylindrical reservoir with semi-ellipsoidal ends.
Finite element predicted distribution of effective stress (MPa) after 45 mm and 90 mm dis-
placement of the upper profile shaped die.
274 L.M. Alves et al.

Figure 7.16 shows the finite element predicted distribution of average stress σm
at the end of the forming process. As seen larger, compression values of the aver-
age stress are found at the semi-ellipsoidal tubular ends and justify the need to
employ internal mandrels for avoiding collapse by wrinkling along the circumfer-
ential direction.
On the contrary, the polar openings of the reservoir and the opposite regions lo-
cated at the internal mandrel show evidence of tensile average stresses.

Fig. 7.16. Forming a tubular preform into a cylindrical reservoir with semi-ellipsoidal ends.
Finite element predicted distribution of average stress (MPa) at the end of the forming
process.

The feasibility of the proposed manufacturing process was further investigated

by analysing the experimental and numerical predicted variation of thickness in
the cross-section of a cylindrical reservoir with semi-ellipsoidal ends (Fig. 7.17).
The meridional distance is measured from the mid-point of the finished reservoir
Manufacturing Seamless Reservoirs by Tube Forming 275

and, as seen in the figure, thickness variation along the cross-section of the reser-
voirs shows a significant growth as the circumferential perimeter decreases with
values above 150% at the open poles.
The initial flat region of the graphic corresponds to nearly unstrained material
placed in the cylindrical region of the reservoir. The final thickness in this region
of the reservoir remains practically identical to the initial thickness of the preform.
The subsequent slight decrease in the variation of thickness is related to the por-
tion of the tubular preform that starts to bend in order to match the contour of the
die. Measurements and numerical predictions can even yield negative values, re-
sulting in local thicknesses smaller than that of the original preform, as can be ob-
served at 40 mm distance from the equatorial region.
The last part of the graphic (say, above 45 mm distance from the equator)
shows a significant growth rate in thickness variation. This is due to compression
in the circumferential direction and the significant increase of thickness at the po-
lar openings of the reservoir being very advantageous for subsequent installation
of valves and end caps by mechanical fixing or welding.

200
FEM
Experimental
160
Thickness Variation (%)

120

80

40

0
0 10 20 30 40 50 60 70

-40 Meridional Distance (mm)

Fig. 7.17. Experimental and finite element predicted variation of thickness in the cross-
section of a cylindrical reservoir with semi-ellipsoidal ends.

The initial flat region of the graphic corresponds to nearly unstrained material
placed in the cylindrical region of the reservoir. The final thickness in this region
of the reservoir remains practically identical to the initial thickness of the preform.
The subsequent slight decrease in the variation of thickness is related to the por-
tion of the tubular preform that starts to bend in order to match the contour of the
die. Measurements and numerical predictions can even yield negative values, re-
sulting in local thicknesses smaller than that of the original preform, as can be
observed at 40 mm distance from the equatorial region.
276 L.M. Alves et al.

The last part of the graphic (say, above 45 mm distance from the equator)
shows a significant growth rate in thickness variation. This is due to compression
in the circumferential direction and the significant increase of thickness at the po-
lar openings of the reservoir being very advantageous for subsequent installation
of valves and end caps by mechanical fixing or welding.

7.6.2 Requirements for Aerospace Applications

In the aerospace industry, high pressure reservoirs are required for the storage of
gases, specifically noble gases (such as helium or argon) which, because they form
no compounds and are difficult to store in compound form and, because they are
light tend to have very low boiling points and severe boil-off losses.
Hence, these gases are usually stored in high pressure form to achieve signifi-
cant volumetric density. The relevant application, as required by Omnidea, adds
the fact that pressure might be required for something more than volumetric densi-
ty, as pressure is required so that the stored gas can act as pressurant (helium and
nitrogen) in chemical propulsion or directly as propellant (xenon) in electric-ionic
space propulsion (Fig. 7.18). While xenon requires pressures of the order of 150-
200 bar, helium and nitrogen require 280-350 bar.
To withstand these pressures while still delivering good performance, a high
pressure vessel performance index Ip must be achieved,

pV
Ip = (7.20)
m
where p is pressure, V is volume and m is the tank’s mass. The performance in-
dex is usually presented in J/kg-1 reflecting the fact that an energy density is
stored.
The typical construction type of storage vessels with high values of the per-
formance index is a hybrid solution in which a metallic liner is wrapped around
with epoxy-embedded composite fibre in an arrangement usually known as
‘composite overwrapped pressure vessels’ (COPVs). The metallic liner provides
mainly the shape, gas tightness and the tank’s toughness, while the composite
overwrapping provides the strength required to withstand the tank’s internal
pressure.
The application here under discussion, xenon storage for electric propulsion in
satellites, requires the production of a metallic liner for a COPV, matching the per-
formance of current state-of-the-art COPVs with significant (over 50%) cost and
manufacturing time reductions versus conventional manufacturing technologies.
For this particular application, aluminium alloys best fit the requirements of fa-
bricating seamless high pressure reservoirs with a high strength-to-weight ratio,
adequate toughness, low cost and considerable availability in seamless extruded
tube form, including a multiplicity of diameters and thicknesses. It is important to
mention that this manufacturing technique cannot be employed without access to
seamless extruded tube. Also important, from a manufacturing point of view, is
Manufacturing Seamless Reservoirs by Tube Forming 277

Thruster Cathode A
(Anode) Cathode B

Electrical Xenon Flow

Filter Unit Controller

Pressure
Pressure
Regulation
Power Regulation Unit
Electronics
Processing
Unit

Power input Monitoring Xenon Tank

(50V DC) &
Control

Fig. 7.18. Electric–ionic propulsion system layout in ESA’s Smart-1 probe. The xenon tank
is of US manufacture and a candidate for European replacement.

the very high heat treated-to-annealed strength ratio (see Fig. 7.19), allowing the
forming process to be performed in annealed conditions, which in turn provides
reduce polar apertures, good surface roughness characteristics (important for seal-
ing) while maintaining a relatively small press force. This is not always easy to
achieve with cold forming processes while the subsequent heat treatment process,
applied to the already formed spherical pressure vessels, can avoid the residual
stress remaining from the cold forming operation.
As it is easy to understand from Fig. 7.19 only aluminium alloys are capable
of multiplying the yield strength by a factor of 3 to 4 between annealed and
heat-treated conditions, especially when compared to alternative metals for this
application: Titanium alloys (mainly the α-β Ti-6Al-4V alloy, the current industry
benchmark), Inconel 718 and AISI 316 stainless steel.
Figure 7.20 shows a cylindrical reservoir with hemispherical ends made of
AISI 316 stainless steel that was successfully formed with an internal mandrel to-
gether with a cylindrical reservoir with hemispherical ends also made of AISI 316
that was not successfully formed with an internal mandrel made of MCP137.
278 L.M. Alves et al.

1200

1000
Yield Strength (MPa)

800

600

400

200

0
AA 6061 "O" AA 7050 "O" AA 2219 "O" AISI 316 "O" Ti-6-4 "O"
vs. vs. vs. vs. vs.
"T6" "T73X" "T851" 30% cold red. "STA"

Fig. 7.19. Comparison, for different metallic alloys, of yield strength in annealed vs. treated
conditions. Aluminium alloys exhibit the smaller annealed strengths and greatest response
to treatment.

Fig. 7.20. Successful and non-successful forming of cylindrical reservoirs with hemispherical
ends made from AISI 316 stainless steel.
Manufacturing Seamless Reservoirs by Tube Forming 279

7.7 Conclusions
The proposed manufacturing process was developed to fit the specific require-
ments of InnovGas project, performed between the Portuguese SME Omnidea and
the European Space Agency, which required substantial development of manufac-
turing technologies capable of producing seamless high pressure reservoirs. The
process extends the tools and techniques commonly utilized in tube forming in or-
der to include two innovative features related to the utilization of sharp edge dies
and internal, recyclable, mandrels made from low melting point alloys.
Sharp edge dies and internal mandrels proved crucial to fabricate spherical and
cylindrical reservoirs from both aluminium and stainless steel, in a single forming
operation without the risk of collapse by local buckling and/or wrinkling. Local
buckling and wrinkling are associated with compressive instability in the axial and
circumferential directions during the forming process. Wrinkling can also be attri-
buted to the strain loading paths being close to uniaxial compression as has been
experimentally observed and numerically predicted by means of finite element
analysis.
The increase in thickness at the poles is useful for installing devices, fixing the
outlet ports; also because overwrapping is much more complex near the polar re-
gions, the metallic liner alone might be required to withstand the internal pressure,
thus requiring higher thicknesses. In addition, the ultimate forming load for produc-
ing seamless reservoirs is small enough for enabling the process to be industrialized
in a low-cost, small capacity, press.

References
[1] Hübner, A., Teng, J.G., Saal, H.: Buckling behaviour of large steel cylinders with pat-
terned welds. International Journal of Pressure Vessels and Piping 83, 13–26 (2006)
[2] Kawahara, G., McCleskey, S.F.: Titanium lined, carbon composite overwrapped pres-
sure vessel. In: 32nd AIAA/ASME/SAE/ASEE Joint Propulsion Conference, Lake
Buena Vista, FL, USA (1996)
[3] Teng, J.G., Lin, X.: Fabrication of small models of large cylinders with extensive
welding for buckling experiments. Thin-Walled Structures 43, 1091–1114 (2005)
[4] Lee, H.S., Yoon, H.S., Yoon, J.H., Park, J.S., Yi, Y.M.: A study on failure characteristic
of spherical pressure vessel. Journal of Materials Processing Technology, 164–165,
882–888 (2005)
[5] Ostwald, P., Muñoz, J.: Manufacturing processes and systems. John Wiley & Sons,
New York (1997)
[6] Fengman, H., Zheng, T., Ning, W., Zhiyong, H.: Explosive forming of thin-wall
semi-spherical parts. Materials Letters 45, 133–137 (2000)
[7] Rosa, P.A.R., Alves, L.M., Martins, P.A.F.: Experimental and numerical modelling of
tube end forming processes. In: Davim, J.P. (ed.) Finite Element Methods in Manu-
facturing Processes, pp. 93–136. ISTE –Wiley (2010)
[8] Alves, M.L., Rodrigues, J.M.C., Martins, P.A.F.: Simulation of three-dimensional
bulk forming processes by the finite element flow formulation. Modelling and Simu-
lation in Materials Science and Engineering – Institute of Physics 11, 803–821 (2003)
280 L.M. Alves et al.

[9] Alves, L.M., Martins, P.A.F., Pardal, T.C., Almeida, P.J., Valverde, N.M.: Plastic de-
formation technological process for production of thin-wall revolution shells from tu-
bular billets. Patent request no. PCT/PT2009/000007, European Patent Office (2009)
[10] Everhart, M.C., Stahl, J.: Reusable shape memory polymer mandrels. In: Proceedings
of the SPIE - The International Society for Optics and Photonics, vol. 5762, pp. 27–
34 (2005)
[11] Alves, L.M., Pardal, T.C.D., Martins, P.A.F.: Nosing thin-walled tubes into axisym-
metric seamless reservoirs using recyclable mandrels. Journal of Cleaner Produc-
tion 18, 1740–1749 (2010)
[12] Alves, L.M., Nielsen, C.V., Martins, P.A.F.: Revisiting the Fundamentals and Capa-
bilities of the Stack Compression Test. Experimental Mechanics (in press, 2011)
[13] Alexander, J.M.: An approximate analysis of the collapse of thin cylindrical shells
under axial loading. Quarterly Journal of Mechanics and Applied Mathematics 13,
10–15 (1960)
[14] Allan, T.: Investigation of the behaviour of cylindrical tubes subject to axial compres-
sive forces. Journal of Mechanical Engineering Science 10, 182–197 (1968)
[15] Rosa, P.A.R., Rodrigues, J.M.C., Martins, P.A.F.: External inversion of thin-walled
tubes using a die: experimental and theoretical investigation. International Journal of
Machine Tools and Manufacture 43, 787–796 (2003)
[16] Kobayashi, S., Oh, S.I., Altan, T.: Metal forming and the finite element method. Ox-
ford University Press, New York (1989)
[17] IANGV International Association for Natural Gas Vehicles,
http://www.iangv.org
[18] Report to Congress, Effects of a Transition to a Hydrogen Economy on Employment
in the United States, Department of Energy, US (2008)
Author Index

Abellan-Nebot, J.V. 55 Liu, J. 57

Alves, Luis M. 253
Astakhov, Viktor P. 1 Martins, Paulo A.F. 253

Davim, J. Paulo 187 Nandi, Arup Kumar 145

de Sousa, R.J. Alves 219
Pal, Surjya K. 101
Fernandes, Nuno 253 Priyadarshini, Amrita 101
Fountas, Nikolaos 187
Samantaray, Arun K. 101
Grilo, T.J. 219 Santana, Pedro 253
Subiron, F. Romero 57
Henriques, M.P. 219
Valente, R.A.F. 219
Krimpenis, Agis 187 Vaxevanidis, Nikolaos M. 187
Subject Index

ABAQUS platform 127–130 temperature 19, 102, 109, 134,

Aerospace applications 276–278 138–139, 146
Agregation rule 154–155 tool wear 61, 65, 77, 78, 89,
Anisotropic models 224–229, 243, 247 91–93
Ant-colony optimization 5, 211–212, Cylindrical cup 229, 233, 245–249,
214
Artificial Neural Networks (ANNs) Datum 55, 57–62, 65–66, 69, 71, 79,
193–196, 204–205, 213–214 81, 89–90, 93
Design of experiments (DOE) 1–4, 36,
Benchmarks’ 229 52, 195
Binary coding 166 Derivation 58, 60–62, 64, 67, 75, 78,
Boundary conditions 2, 102, 111–113, 93–94, 112, 117
120–121, 127–130, 220 Drilling 13, 33, 101, 180, 185

Chip 8, 13, 26, 101–111, 114, Encoding 197–198

119–121, 123, 125–126, 129–140, Evolutionary algorithms 187–188,
145–146 196–197, 203–214
formation 102–106, 109–111, 114,
120, 130–131, 134, 138–140, 146 Factorial experiments 10, 23, 25
morphology 106, 110, 135–139 Finite element (FE) 101–102, 111–112,
separation criterion 121, 123–125, 115–116, 119–120, 127, 130,
tool interface model 125 219–224, 226, 229–232. 234,
CNC machining 5, 188, 190, 214 236–237, 241–242, 245, 247, 249,
Computational time 120, 134 250, 255, 261–266, 268–275, 279
Conical cup 222, 229–230, 234, 241 inputs 130, 134
Contact 68, 70–71, 117–118, 125–126, outputs 138–140
129, 132, 220, 231–232, 246, 249, simulation 101, 113, 119, 131, 138,
259–260, 265–267 140
Coordinate systems 62–63 Finite element flow formulation 261,
Critical instability load 260–261 262,
Crossover 148, 164–165, 168–169, Finite element modelling (FEM) 253,
171, 196–201, 203–204, 206 268
Cutting 8–9, 13, 19, 21, 24, 26, 50, 56, Fitness function 166, 167, 170, 188,
59–61, 63–67, 72, 76–78, 89, 91–93, 189, 190, 210
101–110, 114, 123, 125–135, Fixture 55–57, 59–62, 65–72, 79, 80,
137–140, 145–146, 168, 180–185, 82–94, 106
187, 189–192 Flange forming 219, 233, 245
force 19, 76–77, 102–103, Flow curve 259
105–107, 109–110, 121, Formability 222, 223, 253, 268, 269
132–133, 138–139, 180–181, Forming loads 271, 279
184–185, 189–190 Free forming 229
284 Subject Index

Friction 102, 106, 108–110, 119, 125, Geometry modeling 114

129–132, 229, 231, 232, 259, 265 Group method of data handling
Full factorial design 7, 10, 16, 24, 44 (GMDH) 1, 5, 44, 45
Fuzzy 5, 145–156, 158–163, 172–174,
176, 178, 180, 181, 187 Heat generation 104, 122, 125, 129,
approches 145 134
implication methods 153, 155 Hybrids of evolutionary algorithms 213
inferences 145, 155 Hourglassing 116, 117
logic 5, 10, 23, 44, 45, 134, 135,
145–148, 151, 152, 156, Induced variations 55, 57, 59, 60, 61,
160–164, 172–174, 180, 181, 64–67, 69, 72, 74–77, 79, 82–85,
185, 187, 189, 194, 232 87–89, 91–94
logic operations 148, 151 Isotropic models 224, 243, 247
relation 2, 5, 10 ,18, 19, 27–31, 34,
44–47, 50, 57, 58, 61, 62, 64, 74, Kinematic 60, 61, 65, 72–74, 113, 117,
79, 80, 104, 106, 129, 133, 134, 129, 223, 232
146, 149, 152–155, 178, 184,
198, 201, 223, 229, 231 Level of factors 9, 19
rule based model 145, 147, Logical operations 148, 151
160–162, 180 Lubrication 26, 181, 182, 184, 229,
set 2, 3, 6, 8–12, 16, 19, 20, 28, 44, 259
46, 47 ,50, 51, 83, 89, 107, 112,
117, 122, 127, 137, 146–149, Machining 8, 14, 55–57, 59, 61–66,
151–153, 156, 159, 160, 75–85, 87–89, 91–94, 101–107,
162–164, 166, 175, 176, 180, 109–111, 113–115, 117, 119–123,
187–190, 192, 194–196, 198, 125–127, 129–135, 137–139, 140,
210, 212, 221, 229, 234, 248, 145, 146, 178, 180, 181, 186,
262, 263 188–197, 199, 201, 203, 204, 205,
set operators 151 207, 209, 211, 213, 214
Fuzzification 259, 160, 162, 163 operations 5, 50, 61, 76, 79, 86,
89, 106, 148, 151, 152, 188,
Genetic algorithm (GA) 145, 147, 191, 197, 198, 203, 207, 219,
163–165, 172, 175, 187, 196, 197, 220, 254, 257, 267
201, 206, 211 optimization 4, 15, 24, 25, 45, 46,
Genetic operators 197–199, 203, 206, 48, 50, 102, 139, 140, 147, 163,
207 164, 170, 172, 173, 187–193,
Geometric 7, 55, 57, 60, 61, 63, 65, 195–197, 199, 201, 203–205,
72–76, 93, 103, 104, 120, 121, 128, 207–215
188, 190, 223, 230, 242, 253, 255, processes 24, 46, 51, 55, 82, 101,
257, 267 102, 145, 146, 178, 187–193,
model 2, 5, 8, 10, 13, 14, 19, 21, 22, 196, 213–215, 222, 223, 255,
25, 26, 40–48, 50, 51, 55–62, 64, 262, 269, 277
72, 73, 76–78, 81–83, 88, 89, 93, Mandrels 256–259, 265, 266, 272, 274,
94, 101–104, 106, 110–115, 279
119–123, 125, 127–132, Manufacturing 1, 3, 5, 8, 13, 14, 23–25,
134–138, 140, 145–147, 46, 48, 55–59, 61, 66, 78, 81–83, 87,
160–162, 176, 178, 180–185, 91, 94, 102, 112, 145, 174, 187–189,
188, 190, 203, 211, 219–226, 192, 253–255, 257, 261, 268–270,
229, 233–236, 238, 239, 242, 272, 274, 276, 277
243, 245, 247, 263, 264, 266, 271 Mechanical testing of materials 259
Subject Index 285

Mesh 111, 113–116, 118–120, 122, Response 1, 6, 7, 11–16, 18–21, 24–25,

123, 126, 127, 129, 130, 134–137, 28, 31, 36–40, 44, 50, 51, 102, 116,
221, 224, 229–232, 234, 236–239, 122, 123, 125, 191, 204, 213, 260,
243, 245, 247, 249, 250, 265, 267 278
adaptativity 118, 129 Resolution level 22
attributes 115, 127
Meshing 114, 115, 120, 126, 129, 130, Seamless reservoirs 253, 255, 257, 259,
134–136, 221, 245, 267 261, 267, 279
Migration 204, 207 Sellection 8, 9, 19, 21, 42, 44, 46–48,
Minimum quantity of lubricant 180 51, 57, 83, 113, 117, 134, 137, 145,
Modes of deformation 222, 259, 267–269 146, 162–165, 167, 168, 170, 174,
Multi-objective optimization 164, 193, 195–199, 206, 207
187–189, 192, 193, 197 Sensitivity indices 83–88, 91, 92,
Mutation 164, 165, 169, 171, 196–199, 94
201–204, 206 Sheet metal forming 219–222
Sieve DOE 1, 25, 27, 28, 31, 33, 34
Numerical simulation 219, 220, 222, Simulated annealing 187, 207, 209,
223, 229, 231, 232, 249, 267 210, 214
Single-objective optimization 192,
Objective function 45, 164–167, 171, 197
188–190, 197, 204, 207, 208, 213 Spindle 8, 63, 75, 76, 89, 90, 92
Optimization 4, 15, 24, 25, 45, 46, 48, Split-plot DOE 1, 34, 36, 48
50, 102, 139, 140, 147, 163, 164, Stations 55, 58, 6079, 81, 85, 87, 89,
172, 173, 187–189, 192, 193, 196, 92, 94
197, 203, 204, 207, 215 Statistical method 1, 2, 4, 24
Orthogonal array 22, 23 Stream-of-variation 55, 57
Orthogonal machining 101, 119, 125 Surface roughness 145, 146, 178–182,
185, 190, 191, 194, 196, 213, 277
Parallelism 204, 206
Pareto optimal sets 192
Tabu search 187, 207, 210, 211, 214
Particle swarm optimization 187, 207,
Taguchi method 4, 22–25
214
Thermal 56, 61, 65, 67, 72–75, 89–93,
Plunge grinding process 178, 180
103, 109, 120, 121, 123, 125, 129,
Preforms 256, 259, 265, 269, 272
131, 134, 136–139, 146
Process plan 56, 57, 81–84, 90, 91, 94
Tooling concept 255
Process planning 55–57, 61, 81, 82, 88,
Tribes 207, 212, 213
92, 94, 188
Tube forming 253, 254, 259
Quality assurance 55
Variation propagation modeling 58
Real coding 166
Reproduction model 203 Wrinkling defects 219, 223