Professional Documents
Culture Documents
Gyung-Jin Park
Analytic Methods
for Design Practice
123
Gyung-Jin Park, PhD
Department of Mechanical Engineering
College of Engineering Science
Hanyang University
Republic of Korea
This book is a compilation of what I have learned and studied so far. For the past
twenty years, I have been teaching mechanical design. Most of the contents of this
monograph was created during the past fifteen years of my professorship at
Hanyang University, Ansan, Korea. Mechanical design is a process for
incorporation of the results of theoretical calculation into design. Generally, many
resources are invested in mechanical analyses, however, the incorporation process
is carried out by experiences or engineering intuition. Although it may seem
inefficient, the process shows the characteristics of design well. Since the design
process is more complicated than the analysis process, the designer’s decision
made based on engineering intuition may be more acceptable.
Recently, engineering systems have tended to become quite large. When a
system is large, it is not easy to design it by subjective judgement. Instead, a
design method that is objectively validated is needed. Many design methods have
been developed to this end. A decision making process is required in design,
therefore, the designer inevitably contemplates the solution. One may not even be
able to make the final decision. In this case, a design method with definite logic is
needed. That is, a design tool is exploited to facilitate the decision making process.
Design is classified into conceptual design, which determines the overall
characteristics of a product, preliminary design where selection and distribution of
parts are determined, and detailed design, which fixes the final dimensions. A
design generally proceeds in the order of conceptual, preliminary and detailed
designs. Do we have a universal method that can be applied to all the processes?
It seems that we do not have such a method yet. Instead, different methods are
employed in different steps. From the viewpoint of application of the design
methods, the design process is classified into two processes of conceptual design
and detailed design.
This book treats the design methods for each step of the entire design process.
It explains some analytic design methods and discusses their application. An
analytic design method indicates such a method with a definite theoretical logic.
Each chapter handles a method, and the methods are derived and explained.
Readers can practice the methods with examples. Appendices, which are included
at the end of the chapters, demonstrate case studies. Thus, each chapter is
somewhat independent.
Classes of higher levels of undergraduate or graduate courses can use this book
as a textbook. The students should have basic knowledge of engineering and
viii Preface
1 Introduction......................................................................................................1
1.1 What is Engineering Design? ...................................................................1
1.2 Role and Classification of Design ............................................................4
1.3 Analytic Design Methodology .................................................................6
1.4 Axiomatic Design.....................................................................................7
1.5 Design Optimization.................................................................................8
1.6 Structural Optimization ............................................................................9
1.7 Structural Optimization Under Dynamic Loads .....................................11
1.8 Design with DOE ...................................................................................12
1.9 Robust Design ........................................................................................13
1.10 Multidisciplinary Design Optimization ..................................................13
1.11 Summary.................................................................................................15
7 Robust Design...............................................................................................393
7.1 Introduction ..........................................................................................393
7.2 Mean and Variance...............................................................................394
7.3 Taguchi Method ...................................................................................399
7.3.1 Introduction ..............................................................................399
7.3.2 The Loss Function and S/N Ratio.............................................401
7.3.3 Parameter Design......................................................................404
7.3.4 Various Methods Using the Taguchi Method...........................412
7.4 Robust Optimization.............................................................................416
7.4.1 Introduction ..............................................................................416
7.4.2 Robustness of the Objective Function ......................................418
7.4.3 Robustness of Constraints ........................................................421
7.4.4 Optimization Methods ..............................................................425
Contents xiii
10.4.1
Multidisciplinary Feasible (MDF)............................................579
10.4.2
Individual Discipline Feasible (IDF) ........................................580
10.4.3
All-at-once (AAO)....................................................................582
10.4.4
Concurrent Subspace Optimization (CSSO) ............................585
10.4.5
Bilevel Integrated System Synthesis (BLISS) ..........................593
10.4.6
Collaborative Optimization (CO) .............................................597
10.4.7
Multidisciplinary Design Optimization Based
on Independent Subspaces (MDOIS) .......................................602
10.4.8 Comparison of MDO Methods .................................................605
10.5 Discussion ............................................................................................608
10.A Application of an MDO Algorithm to the Design of a Belt Integrated
Seat Taking Crashworthiness into Consideration.................................609
10.A.1 Problem Description .................................................................609
10.A.2 The Interdisciplinary Relationship Between the Disciplines
Involved....................................................................................610
10.A.3 Formulation for the Design of a BIS Taking Crashworthiness
into Consideration ....................................................................610
10.A.4 Application of MDOIS to the Design of a BIS.........................611
Index ....................................................................................................................621
1 Introduction
Production planning
Product design
Feedback
Manufacturing
Feedback
Marketing
engineering parts, and various aspects are defined for manufacturing. Therefore,
the subjects in the planning process are thoroughly examined and details of parts,
materials, dimensions, etc. are determined. In general, the output from the design
process is expressed by drawings. An example of a drawing is illustrated in Figure
1.2. The role of the design process is a bridge between the stages of the planning
and the manufacture of the final products.
A designer typically goes through decision-making processes. What would be
rational and logical decisions? Who is a good designer and who is not? Can the
characteristics of good decisions be generalized? Can a good design of a specific
product also be good for other products? Design activities can be viewed from
different aspects depending on the function of each part of a product. According
to the manner of the decision making process, there are two viewpoints for the
definition of design. In some parts aesthetic beauty is emphasized. An example is
the design of the external shape of an automobile. In other parts, the engineering
functionality is more important. The engine of an automobile is a prime example.
That is, the design process should be based on objective logic and a general design
principle. While the definition of design may include many views, Figure 1.3 and
Figure 1.4 contrast the different design requirements and objectives in automobile
design. External appearance is mostly considered in Figure 1.3. On the other hand,
strength and manufacturability are important in the design of Figure 1.4. If we
find a general principle for good designs, general design methodologies can be
developed. Recently, design methods have begun to emerge by the finding of
common aspects of good designs. An analytic design method is a design
methodology that has a certain process for design practice.
There are many excellent designs around us. For example, there are numerous
bicycles, automobiles, electrical devices, etc. Some of them have been created by
inspiration from unknown sources while others may have been improved gradually.
Excellent patents are being created in these ways as well. Can we derive specific
Introduction 3
Conceptual design
Overall selection of functions and parts
Detailed design
Determination of sizes of parts
Definition of problem
Determination of design
Satisfied?
No
Yes
Final design
Why? G Why?
Constraints
Why?
G
Constraints
G
Customer needs Functional requirements Design parameters Process variables
Find b Rn (1.1a)
to minimize f (b) (1.1b)
g j (b ) d 0 , j 1, ,m (1.1d)
bL d b d bU (1.1e)
where b is the design variable vector with n elements, f is the objective function,
hi is the ith equality constraint, g j is the jth inequality constraint, b L and b U are
Introduction 9
K (b)z f (1.2)
where K is the stiffness matrix, z is the nodal displacement vector, b is the design
variable vector for sizes and shapes, and f is the external load vector.
10 Analytic Methods for Design Practice
Structural
Design variable
optimization
Find b R n (1.3a)
to minimize f (b, z ) (1.3b)
bL d b d bU (1.3e)
The objective function in Equation 1.3b can be the weight of the structure or a
specific response. The inequality constraints in Equation 1.3d are generally
defined by limit values on displacements, stresses, natural frequencies, etc. In the
optimization process, we need sensitivity information for the function in Equation
1.3 with respect to design variables. It is known that the sensitivity evaluation is
quite expensive. Therefore, efficient calculation of sensitivity information is
significant.
Structural optimization is classified according to the characteristics of the
design variables. The classification is shown in Table 1.1. In structural design,
various analysis methods are employed to examine the status of the structure.
However, research has been concentrated mostly on linear static problems.
Therefore, the application of optimization is still limited. More research is needed
for problems requiring other analysis methods (i.e. nonlinear FEM).
Introduction 11
M (b )z K (b ) z f (t ) (1.4)
where M is the mass matrix and z is the acceleration vector. Equation 1.5 is the
general formulation for dynamic response optimization.
Find b Rn (1.5a)
to minimize f (b, z, t ) (1.5b)
bL d b d bU (1.5e)
other discipline and vice versa. Complicated relations occur due to the coupled
relations and entire system analysis is extremely difficult. A typical coupled
Aerodynamics
Pressure
Pressure
Deformation Deformation
Bending stress
Start End
Yes
Analysis 2 Analysis 3
Optimum sensitivity analysis
Analysis 1 Discipline 1
Global
sensitivity
Analysis 2 equation Discipline 2
Analysis 3 Discipline 3
1.11 Summary
Design methodologies are described briefly. As mentioned earlier, axiomatic
design is introduced as a method for conceptual design. There exist some other
methods to enhance efficiency of the design process. They are design for
assembly (DFA), design for manufacturing (DFM), etc. Also, TRIZ can be
exploited to create a design idea. These are the methods for conceptual design,
which however are not handled in this book.
For detailed design, optimization, DOE and robust design are briefly discussed.
Recently many other methods have been developed such as genetic algorithms
(GA), simulated annealing (SA), design with fuzzy set theory, design with neural
networks, particle swarm optimization algorithms, etc. Although they belong to
the area of optimization, they are not included herein.
References
Antonsson EK, Cagan J (2001) Formal Engineering Design Synthesis. Cambridge
University Press, New York
Altshuller G (1996) And Suddenly the Inventor Appeared. Technical Innovation
Center, Massachusetts
Arora JS (1988) Introduction to Optimum Design. McGraw–Hill, New York
Arora JS (1997) Guide to Structural Optimization. The American Society of Civil
Engineers, New York
Braun R (1996) Collaborative Optimization: An Architecture for Large-Scale
Distributed Design. Ph.D. Thesis, Stanford University, CA
Choi WS, Park GJ (2002) Structural Optimization Using Equivalent Static Loads
at All Time Intervals. Computer Methods in Applied Mechanics and
Engineering 191:2077í2094
Dieter GE (1999) Engineering Design: A Materials and Processing Approach. 3rd
ed., McGraw–Hill, New York
Haftka RT, Gürdal Z (1992) Elements of Structural Optimization. 3rd revised and
expanded ed., Kluwer Academic Publishers, Dordrecht
Haugen EB (1980) Probabilistic Mechanical Design. Wiley, New York
Hibbeler RC (2000) Mechanics of Materials. 4th ed., Prentice Hall, Englewood
Cliffs, NJ
16 Analytic Methods for Design Practice
2.1 Introduction
Axiomatic design is a design theory that was created and popularized by Professor
Suh of the Massachusetts Institute of Technology (Suh 1990, 2000). Actually, it is
a general design framework, rather than a design theory. As the word
“framework” indicates, it can be applied to all design activities. It consists of two
axioms. One is the Independence Axiom and the other is the Information Axiom.
A good design should satisfy the two axioms while a bad design does not. It is
well known that the word “axiom” originates from geometry. An axiom cannot be
proved and becomes obsolete when a counterexample is validated. So far, a
counterexample has not been found in axiomatic design. Instead, many useful
design examples with axioms are validated.
Design is the interplay between “what we want to achieve” and “how we
achieve it.” A designer tries to obtain what he/she wants to achieve through
appropriate interplay between both sides. The engineering sequence can be
classified into four domains as illustrated in Figure 2.1. Customer attributes (CAs)
are delineated in the customer domain. In other words, CAs are the customer
needs. CAs are transformed into functional requirements (FRs) in the functional
domain. FRs are defined by engineering words. This is equivalent to “what we
want to achieve.” FRs are satisfied by defining or selecting design parameters
(DPs) in the physical domain. Mostly, this procedure is referred to as the design
process. Production variables (PVs) are determined from DPs in the same manner.
The aspects for the next domain are determined from the relationship between the
two domains, and this process is called mapping. A good design process means an
efficient mapping process.
Design axioms are defined from common principles for engineering activities
as follows:
What? What?
Constraints Constraints
The axioms may look simple. However, they have significant meanings in
engineering. Details of the axioms will be explained later. Axiom 1 is an
expression that design engineers know consciously or subconsciously. When we
design a complex system, the axiom tells us that a DP should be defined to
independently satisfy its corresponding FR. In other words, the FRs of the
functional domain in Figure 2.1 should be independently satisfied by DPs of the
physical domain. Otherwise, the design is not suitable. When multiple designs are
found from Axiom 1, the best one can be chosen based on Axiom 2. That is, the
best design has minimum information content that is usually quantified by the
probability of success. It also corresponds to the engineering intuition that design
engineers usually have in mind. Axiom 2 is related to robust design and it will be
explained later. Although the axioms are expressed simply, real application can be
very difficult.
As explained earlier, axioms are defined in geometry. As in geometry,
theorems and corollaries are derived from axioms (see Appendix 2.A).
Axiomatic Design 19
The Independence Axiom indicates that the aspects in the proceeding domain
should be independently satisfied by the choices carried out in the next domain.
The domains are illustrated in Figure 2.1. The relationship of FR–DP is defined to
be independent. When plural FRs are defined, each DP should satisfy each
corresponding FR. The relationship can be expressed by a design matrix. Using
vector notations for FRs and DPs, the relationship is expressed as the following
design equation:
FR A DP (2.1)
Solution
Design parameters for the vertically hung door in Figure 2.2a are as follows:
DP1 : Vertically hung door
DP2 : Thermal insulation material in the door
Axiomatic Design 21
ª FR1 º ªX 0 º ª DP1 º
« FR » «X « » (2.3)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
ª FR1 º ªX 0 º ª DP1 º
« FR » «0 « » (2.4)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
When we open the horizontally hung door, cold air remains in the refrigerator
and energy loss can be minimized. Therefore, the horizontally hung door has an
uncoupled design and is a better design than the vertically hung door. Is the
horizontally hung door always better? As far as the functional requirements
defined here are kept, it is correct. Suppose that constraints are proposed for the
amount of stored food or convenience to access items. Then the problem will be
22 Analytic Methods for Design Practice
Solution
Analyzing the product in Figure 2.3a, DPs and the design equation are defined as
follows:
DP1 : Angle I1
DP2 : Angle I 2
I2
Y
I1
As shown in Equation 2.5, the design is coupled. Thus, the design is not
acceptable.
Another example is presented in Figure 2.3b. The design is analyzed as
follows:
DP1 : Angle I1
DP2 : Angle I 2
FR ཛ DP
ཝ
FR1 FR2 ˎ DP1 DP2 ˎ
ཞ
Lathe
2.2.2 Independence
An uncoupled design is presented in Figure 2.6a. The point A has ( FR1 ) A and
( FR2 ) A for ( DP1 ) A and ( DP2 ) A , respectively. The points B, C and D have the
same characteristics. If the design is to be changed from A to C, the path A–D–C
or the path A–B–C can be selected. That is, the uncoupled design is independent
of the design path.
The decoupled design in Figure 2.6b is different. Suppose we want to change
the design from A to C. First, DP2 should be changed from ( DP2 ) A to ( DP2 ) E .
In this process, DP1 is fixed and both FR1 and FR2 are changed. Second, DP2
is fixed and DP1 is changed from ( DP1 ) E to ( DP1 ) C . In this process, FR2 is
fixed and FR1 is changed. Thus, the decoupled design relies upon the design path.
That is, DP2 should be determined first and DP1 should be determined later.
Now, look at the coupled design in Figure 2.6c. When the design is changed
from A to C, the effect is the same no matter what design parameter is changed.
Suppose DP1 is changed first. To satisfy FR1 , DP1 can be changed from A to C' ,
and then FR2 is also changed. Thus the design should be changed from C' to C"
to satisfy FR2 . Then FR1 is changed again and DP1 should be changed again.
Therefore, the design process is repetitively performed until the design converges.
This can be quite a complicated process. In particular, convergence may be
impossible when the design is highly nonlinear.
Figure 2.7 briefly presents the above relationships. The characteristics of the
design equations can be expressed by D1 , D 2 and T in Figure 2.7. The ideal
uncoupled design is obtained when D1 D 2 0 and T 90 o. As an index for
coupling, the following index R called “reangularity” is defined:
FR2
DP2 B C
( DP2 ) A, D A D
( FR2 ) A, D
( DP1 ) A, B DP1
FR2 DP2
B E C F
( DP2 ) B ,C
( DP2 ) A, D A D
( FR 2 ) A, D
DP2
FR2
( DP2 ) B , C B
( FR2 ) C C
C cc
( DP2 ) A, D A
Cc
( FR2 ) D D
( DP1 ) A, B DP1
( DP1 ) C , D
( FR1 ) A ( FR1 ) C ,C c FR1
Figure 2.6. Mapping process from the FR domain to the DP domain for each design
Axiomatic Design 27
FR2 DP2
DP2 DP2
DP2
D2
T DP1
D1
(d) Coupled design ( D 1 z 0, D 2 z 0 )
DP1
T
T (c) Decoupled design ( D 1 z 0, D 2 0)
FR1
Figure 2.7. Schematic view of each design according to the coupling characteristics
If the numbers of FRs and DPs are n and each element of the design equation is
Aij , R is as follows:
1/ 2
ª n º
« ( ¦ Aki Akj )2 »
k 1
R «1 n n » (2.12)
i 1, n 1 «
j 1 i , n ( ¦ Aki2 )( ¦ Akj2 ) »
«¬ k 1 k 1 »¼
n | A jj |
S n
(2.13)
j 1
( ¦ Akj2 )1 / 2
k 1
28 Analytic Methods for Design Practice
Reangularity 1 R S 1 R z S 1
Angle between
90q T T
column vectors ( T )
Semangularity 1 S R 1 S z R 1
1/ 2
ª ( A11 A12 A21 A22 ) 2 º
R «1 2 2 2 2
» (2.14a)
«¬ ( A11 A21 )( A12 A22 ) »¼
ª ºª º
S « | A11 | » « | A22 | » (2.14b)
« A2 A2 »« A2 A2 »
¬ 11 21 ¼ ¬ 12 22 ¼
1/ 2 1/ 2
ª 2 2 º ª 2 2 º
A11 A12 A11 A22 | A22 |
R «1 2 2 2 » « 2 2 2 »
(2.15a)
¬« A11 ( A12 A22 ) ¼» ¬« A11 ( A12 A22 ) ¼» 2
A12 2
A22
ª| A | ºª | A | º | A22 |
S « 11 » « 22 » (2.15b)
« A2 » « A2 A2 » 2
A12 2
A22
¬ 11 ¼ ¬ 12 22 ¼
FR2
DP2c2
FR
D2
T
DP1c1 FR1
Figure 2.8. Vector representation of Example 2.3
Solution 2
Solve the problem geometrically. Define c1 [ A11 0]T and c 2 [ A12 A22 ]T . If
the two functional requirements are expressed by a vector FR, then
FR DP1c1 DP2 c 2 . This is geometrically represented in Figure 2.8. From
Figure 2.8, R and S are as follows:
c 1T c 2 A11 A12
cos T (2.16a)
| c 1 || c 2 | 2
A11 2
A12 2
A22
1/ 2
ª 2 2 º
2 1/ 2 A11 A12
R sin T (1 cos T ) «1 2 2 2 »
¬« A11 ( A12 A22 ) ¼»
1/ 2
(2.16b)
ª 2 2 º
A11 A22 | A22 |
« 2 2 2 »
¬« A11 ( A12 A22 ) ¼» 2
A12 2
A22
ª ºª º
S cos D 1 cos D 2 « | A11 | » « | A22 | » | A22 |
(2.16c)
« A2 0 2 »« A2 A2 » 2
A12 2
A22
¬ 11 ¼ ¬ 12 22 ¼
When the design equation is nonlinear with respect to design parameters, Aij
of Equation 2.2 may not be constant. Thus, although the uncoupled relationship is
satisfied at a design point, it may not be satisfied at other points. In this case, an
approximation by Taylor expansion can be employed. FRi in Equation 2.2 can be
approximated as follows:
30 Analytic Methods for Design Practice
3 wFRi
FRi ( FRi ) 0 GFRi ( FRi ) 0 ¦ GDPj (2.17)
j w
1 DPj
where ( FRi ) 0 is the current functional requirement. Using Equation 2.17, the
design equation at ( DP1 , DP2 , DP3 ) is defined as follows:
As shown in Equation 2.18, the design matrix is a matrix with partial derivatives,
which defines the relationship between increments of FRs and DPs. The effort to
find an uncoupled design is to find a design window where the design matrix is
diagonal. Therefore, although the Independence Axiom is satisfied at a design
point, it is not guaranteed if the design is changed.
Can we consider a design to be uncoupled when the off-diagonal terms are
quite small compared to the diagonal terms? The following equation is an
example:
ª FR1 º ªX x x º ª DP1 º
« FR » «x X x »» «« DP2 »»
« 2» « (2.19)
«¬ FR3 »¼ «¬ x x X »¼ «¬ DP3 »¼
where X !! x . The decision can be made based on the tolerance ranges and the
sizes of X and x. Theorem 2.A.8 in Appendix 2.A provides the reason for this.
Theorem 2.A.8 is as follows:
n § wFR ·
¦ ¨¨ i ¸ 'DPj
¸
(2.20)
j z i © wDPj ¹
j 1
When the magnitude of Equation 2.20 is very small, in other words, when x is
considerably small compared to X in Equation 2.19, Equation 2.19 can be regarded
as an uncoupled design.
Suppose that the current design is FR 0 A DP0 . The change of the design
parameters is 'DP . The change of the functional requirements is 'FR and it is
wFRi
obtained by replacing G with ' in Equation 2.18. ('FRi ) diag { 'DPi and
wDPi
('FRi ) diag is the change of FRi by the ith diagonal term with respect to the
change of DPi . Generally, the diagonal term is the largest. Therefore, ('FRi ) diag
has the largest impact on the FRi change. If we exclude ('FRi ) diag from 'FR ,
the remainder is the off-diagonal terms. When the influence from the off-diagonal
terms is very small, we do not need to consider them. This is expressed as
n § wFR ·
¦ ¨¨ i ¸'DPj d ('FRi ) allowable ,
¸
i 1,..., n (2.21)
j z i© wDPj ¹
j 1
ª FR1 º ªX x º ª DP1 º
« FR » «y « » (2.22)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
(1) Obtain the range of design parameters with which we can consider the
design as an uncoupled design.
(2) Obtain the range of design parameters with which we can consider the
design as a decoupled design.
Solution
(1) Ranges of 'DP1 and 'DP2 that satisfy x'DP2 d ('FR1 ) allowable and
y'DP1 d ('FR2 ) allowable .
32 Analytic Methods for Design Practice
(2) Ranges of 'DP1 and 'DP2 that satisfy x'DP2 d ('FR1 ) allowable or
y'DP1 d ('FR2 ) allowable .
There is a saying that a simple design is a good one. From this statement, we may
guess that a good design makes one DP satisfy multiple FRs. In other words, a
coupled design is better. This aspect is very confusing in axiomatic design.
However, from an axiomatic design viewpoint, this is the case where multiple DPs
make a physical entity. That is, multiple DPs satisfy FRs of the same number.
This is called “physical integration.” Physical integration is desirable because the
information quantity can be reduced. The following example is a typical example
of physical integration.
Solution
The device in Figure 2.9 has one physical entity for the bottle opener and can
opener. However, two DPs at both ends independently satisfy the two functional
requirements. Therefore, the design in Figure 2.9 satisfies the Independence
Axiom. If the constraint set includes “both functions should be simultaneously
used,” then a different design should be investigated.
Ii log 2 1 / p (2.23)
In Equation 2.23, the reciprocal of p is used to make the larger probability have
less information. Also, the logarithm function is utilized to enhance additivity.
The base of the logarithm is 2 to express the information content with the bit unit.
Suppose we have the following uncoupled design:
34 Analytic Methods for Design Practice
Suppose p1 , p2 and p3 are the probabilities of satisfying FR1 , FR2 and FR3
with DP1 , DP2 and DP3 , respectively. The total information I total is
3 3 § 1 ·
I total ¦ Ii ¦ log 2 ¨¨ ¸¸ (2.25)
i 1 i 1 © pi ¹
It is noted that the information content should only be defined based on the
corresponding functional requirement.
§ 1 · § 1 ·
I total I1 I 2 log 2 ¨ ¸ log 2 ¨ ¸ 0.1520 0.2345 0.3865 (bits)
© 0. 9 ¹ © 0.85 ¹
(2.26)
Solution
If we use the same machine for both cases, the probability of success is smaller
when the tolerance is small. Also, if the given length (nominal length) is longer,
the ratio of the tolerance to the total length is smaller. Thus, the probability of
success is as follows:
Axiomatic Design 35
§ tolerance ·
p f ¨¨ ¸¸ (2.27)
© nominal length ¹
tolerance
p c (2.28)
nominal length
where c is a constant.
ª FR1 º ªX 0 º ª DP1 º
« FR » «X « » (2.29)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
If p1 is the probability that DP1 satisfies FR1 , then the probability that
DP2 satisfies FR2 under satisfaction of FR1 by DP1 is a conditional probability.
Suppose it is p 21 . Then the probability of success p that both FR1 and FR2 are
satisfied is
p p1 p 21 (2.30)
Target
Probability Bias
density Probability density
function function of the system
Design range
Common
range
FR
Variation from the peak value
Figure 2.11. Calculation of the information content using the probability density function
the design range and the system range. The design range is defined by lower and
upper bounds and the system range is defined by a distribution function of the
system performance. A uniform distribution of a system range is illustrated in
Figure 2.12. The design should be directed to increase the common range. The
information content is defined as follows:
where Asr is the system range and Acr is the common range.
The person considers a house in city A or city B. Table 2.3 shows the conditions
of both cities. Where should the person buy a house to minimize the information
content?
Solution
The system range is defined from Table 2.3 and the design range is determined
from the functional requirements. It is assumed that all the probability densities
are uniform. Figure 2.12 presents the probability density for the price of the house
in city A. Other items can be illustrated in the same manner. The information
content for city A is as follows:
Axiomatic Design 37
city A city B
Price $45,000–$60,000 $70,000–$90,000
Commuting time 35–50 min 20–30 min
§ 1.5 · § 15 ·
I A1 log 2 ¨ ¸ 0.59, I A2 log 2 ¨ ¸ 1.59 (2.33)
© 1 ¹ ©5¹
IA I A1 I A2 2.18 (bits) (2.34)
§2· § 10 ·
I B1 log 2 ¨ ¸ 1.0 , I B 2 log 2 ¨ ¸ 0.0 (2.35)
©1¹ © 10 ¹
IB I B1 I B 2 1.0 (bits) (2.36)
The information content I A for city A is 2.18 and that for city B I B is 1.0.
Therefore, city B has the optimum house from an axiomatic design viewpoint.
Probability
density Probability density Design range
function function
of the system
Common range
1 3 5 7 9 FR
Figure 2.12. Probability density function of a uniform distribution
38 Analytic Methods for Design Practice
ª FR1 º ªX 0 º ª DP1 º
« FR » «X « » (2.37)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
ª wFR1 wFR1 º
ª GFR1 º « wDP wDP2 » ªGDP1 º ª A11 0 º ªGDP1 º
«GFR » « 1 » (2.38)
¬ 2¼ « wFR2 wFR2 » «¬GDP2 »¼ «A
¬ 21 A22 »¼ «¬GDP2 »¼
«¬ wDP1 wDP2 »¼
p s { p( 'FR1 d GFR1 d 'FR1 ) p ('FR2 d GFR2 d 'FR2 | 'FR1 d GFR1 d 'FR1 )
(2.39)
Axiomatic Design 39
Let us assume that Aij of Equation 2.38 is a positive constant and GDPi is
statistically independent. GFR2 is a statistically dependent random variable with
respect to GFR1. In the DP domain, the condition in Equation 2.39 can be
expressed as
A21
A22 'DP2 d GFR 2 GFR1 d A22 'DP2 (2.41d)
A11
ps Af / Adp (2.42a)
I log 2 p s (2.42b)
where Af is the feasible region, which is the shadowed area in Figure 2.13 and
Adp is the tolerance for design parameters, which is 4'DP1 'DP2 in Figure 2.13.
40 Analytic Methods for Design Practice
'FR1
GDP1
A11 ('DP1 , 'DP2 )
'FR2 A21GDP1
GDP2
§ 'FR2 · A22
¨¨ 0, ¸¸
© A22 ¹
§ 'FR2 ·
¨¨ 0, ¸¸
© A22 ¹ GDP1
( DP1* , DP2* )
'FR1
GDP1
A11
'FR1
GDP1
A11 § 'FR2 ·
¨¨ 0, ¸¸
© A22 ¹
'FR2 A21GDP1
GDP2
A22
Figure 2.13. The probability of success of the decoupled design in the DP range
Axiomatic Design 41
I log 2 p s (2.43b)
where Asr is the area of the system range, which is the area of the parallelogram in
Figure 2.14 and the design range is the shadowed area of Figure 2.14. The
common range Acr is the common area of the system range and the design range.
This is the same as the feasible region in Equation 2.41. It is noted that the
probability of success for Figure 2.13a is 1, but that for Figure 2.14a is not 1. It is
somewhat complicated to calculate the shadowed area in Figure 2.13b or the
common area of Figure 2.14b. The probability of success for Figure 2.14b is as
follows:
2
A11 § A21 ·
¨ 'FR1 A22 'DP2 'FR2 ¸¸
'FR1 A21 ¨© A11 ¹
ps (2.44)
A11'DP1 4 A11 'DP1 A22 'DP2
When the feasible region is such as the one in Figure 2.13a, the integration is
easy. However, if it is such as the one in Figure 2.13b, the integration is somewhat
more difficult. In that case, we employ the unit step function u (x) as follows:
u ( x x * ) 1 : when x t x *
0 : when x x * (2.46)
Figure 2.15 represents the unit step function. Using the unit step function, the
probability distribution can be defined not only in the feasible region but also in
the entire region as follows:
42 Analytic Methods for Design Practice
GFR2
A21 ( A11'DP1 , A21'DP1 A22 'DP2 )
GFR2 GFR1 A22 'DP2
A11
('FR1 , 'FR2 )
GFR2
(0, A22 'DP2 )
A21 ( A11'DP1 , A21'DP1 A22 'DP2 )
GFR2 GFR1 A22 'DP2
A11
('FR1 , 'FR2 )
Common range
System range
u( x x * )
x* x
To integrate in the feasible region, the parallelograms in Figure 2.13 are used as
the integration interval. The interval is
In some cases, we may not satisfy the target value FR * exactly with the design
parameters. In this case, the following equations hold:
FR c ADPc (2.50a)
FR c z FR * (2.50b)
44 Analytic Methods for Design Practice
where DPc [ DPc1 , DPc 2 ]T and FR c is the functional requirement vector made by
the current design parameters. The probability of success by the graphical method
is evaluated by transition of the rectangulars and parallelograms in Figures 2.13
and 2.14, so that DPc [ DPc1 , DPc 2 ]T becomes the origin. Then Equation 2.40
yields
FR2* 'FR2 d A21 ( DPc1 GDP1 ) A22 ( DPc 2 GDP2 ) d FR2* 'FR2
(2.51b)
'DP1 d GDP1 d 'DP1 (2.51c)
The advantage of the integration method is that the probability of success can
be calculated for many design parameters. If the number of design parameters is n,
the following multiple integration is utilized:
n n 1
FR n* ¦ Ani DPci 'FRn ¦ Ani GDPi
FR1* A11 DPc1 'FR1 i 1 i 1
A11 Ann
³ ³ p1... pn dGDPn dGDP1 (2.53)
n 1
FR1* A11 DPc1 'FR1 n
FR n* ¦ Ani DPci 'FR n ¦ Ani GDPi
A11 i 1 i 1
Ann
The integration method can be defined by the distribution function of FR for the
feasible region in Equation 2.41. Calculation of the information content in the FR
domain is more complicated than calculation in the DP domain, because the
distribution of DP is usually given.
The above methods can be applied to designs with FR–DP hierarchy of many
levels. When we have multilevel hierarchy, we can make an entire design matrix
for the FRs and DPs in the lowest level. We can apply the above methods to the
entire design matrix. The information content can be evaluated for a coupled
design. The method is defined by modification of the above methods. However, it
Axiomatic Design 45
is very complex and the coupled design is not considered in general design.
Therefore, the information content for the coupled design is not explained here.
ª FR1 º ª3 0º ª DP1 º
« FR » «2 5» « DP » (2.54)
¬ 2¼ ¬ ¼¬ 2 ¼
Solution
(1) From Equation 2.40, the feasible region in the DP domain is as follows:
1
I log 2 0.152 (bits) (2.56b)
ps
(2) From Equation 2.41, the feasible region in the FR domain is as follows:
46 Analytic Methods for Design Practice
GDP2
2
GDP2 GDP1 0.3
5
( 0 , 0.3 )
(0.3,0.18)
( 0 .3 , 0 ) ( 0 .3 , 0 ) GDP1
( 0.3 , 0.18 )
( 0 , 0 .3 )
2
GDP2 GDP1 0.3
5
2
5 u 0.3 d GFR2 GFR1 d 5 u 0.3 (2.57d)
3
1
I log 2 0.152 (bits) (2.58b)
ps
1 1
pGDP1 pGDP2 (2.59a)
2'DP1 0.6
Axiomatic Design 47
2 GFR2
GFR2 GFR1 1.5
3 (0, 1.5)
GFR1 0.9
GFR1 0.9
(0, í1.5) 2
GFR2 GFR1 1.5
3
From Equation 2.49, the probability of success and the information content
are
1
I log 2 0.152 (bits) (2.60b)
ps
(1) Calculate the information content in the DP range by using the graphical
method.
(2) Calculate the information content in the FR range by using the graphical
method.
(3) Calculate the information content in the DP domain by using the
integration method.
Solution
FR2* 'FR2 d A21 ( DPc1 GDP1 ) A22 ( DPc 2 GDP2 ) d FR2* 'FR2
(2.61b)
'DP1 d GDP1 d 'DP1 (2.61c)
GDP2
1.2 2GDP1 5GDP2
(0.15,0.3)
(0.3,0.3) (0.3,0.3)
(0.3,0.12)
GDP1
(0.3,0.24)
(0.3,0.3) (0.3,0.3)
(0.15,0.3)
1.8 2GDP1 5GDP2
A21
A22 'DP2 d GFR2 GFR1 d A22 'DP2 (2.64e)
A11
2
1.5 d GFR2 GFR1 d 1.5 (2.65d)
3
GFR2
(0, 1.5)
(0.9, 2.1)
(í2.25, 0) (1.8, 0)
(T1.2, 0) (í0.9, 0) (0.9, 0) GFR1
(1.8, 0.3)
(2.25, 0)
(0.9,Gí0.9)
(0,Gí1.5)
(í0.9,Gí2.1) (0,Gí1.8)
(í0.45,Gí1.8)
1
I log 2 0.193 (bits) (2.66b)
ps
(3) The integration range is within the parallelogram in Figure 2.18. Therefore,
the information content is calculated as
1.2 2GDP1
0.6 5
ps ³ ³ p1 p2 dGDP2 dGDP1 0.875 (2.67)
0.4 1.8 2GDP1
5
3 3
pGDPi (GDPi ) 2 , i 1, 2 (2.68)
4('DPi ) 3 4('DPi )
Axiomatic Design 51
(1) In Example 2.10, replace the distribution function with Equation 2.68. The
tolerances for the design parameters are the same. Calculate the
probability of success by the integration method.
(2) In Example 2.11, replace the distribution function with Equation 2.68. The
tolerances for the design parameters are the same. Calculate the
probability of success by the integration method.
Solution
(1) When the distribution has Equation 2.68 in Example 2.10, the probability
distribution pi for the ith design parameter is
ª 3 3 º
pi « (GDPi ) 2 u [u (GDPi ( 'DPi )) u (GDPi ('DPi ))]
¬ 4 u 0.3 3 4 u 0.3 »¼
(2.69)
1
I log 2 0.032 (bits) (2.70b)
ps
1.2 2GDP1
0.6 5
ps ³ ³ p1 p2 dGDP2 dGDP1 0.953 (2.71a)
0.4 1.8 2GDP1
5
The example shows various cases for calculating the information content
using the probability density function. A practical example is introduced
in Appendix 2.B.
52 Analytic Methods for Design Practice
Analysis of design
No
Is the no. of designs sufficient?
Yes
Yes
Multiple designs? Find the best design with
the Information Axiom
No
To satisfy the two FRs, a refrigerator with two compartments can be designed.
The design parameters are as follows:
DP1 : The freezer section
DP2 : The chiller section
FR11 : Maintain the temperature of the freezer section in the range of 18q C
r 2q C.
FR12 : Maintain a uniform temperature in the freezer section.
FR13 : Control the relative humidity to 50% in the freezer section.
ª FR12 º ªX 0 0 º ª DP12 º
« FR » «X X 0 »» «« DP11 »» (2.72)
« 11 » «
¬« FR13 ¼» ¬« X 0 X ¼» ¬« DP13 ¼»
For food storage in the chiller section, the temperature should be maintained in
the range of 2q C 3q C . The chiller section also activates the compressor and
circulates the air. Design parameters for the chiller section are
DP21 : Sensor/compressor system that activates the compressor when the
temperature of the chiller section is different from the preset one
DP22 : Air circulation system that blows the air into the chiller section and
circulates it uniformly
The design equation is a decoupled one as follows:
Axiomatic Design 55
ª FR22 º ªX 0 º ª DP22 º
« FR » «X « » (2.73)
¬ 21 ¼ ¬ X »¼ ¬ DP21 ¼
ª FR12 º ªX 0 0 0 0 º ª DP12 º
« FR » «X « »
« 11 » « X 0 0 0 »» « DP11 »
« FR13 » «X 0 X 0 0 » « DP13 » (2.74)
« » « »« »
« FR22 » «0 0 0 X 0 » « DP22 »
« FR21 » «¬ 0 0 0 X X »¼ «¬ DP21 »¼
¬ ¼
It is noted that the FRs of the lower level still keep the independence of the upper
level in Equation 2.74.
From the design equation in Equation 2.74, one compressor and two fans can
satisfy the FRs. DP11 and DP21 are sensor/compressor systems so that the
compressor is activated by the sensors. However, the fans of DP12 and DP22 will
not be activated unless the temperature is out of the range of the preset one.
Therefore, the design with one compressor and two fans satisfies the Independence
Axiom. An example is illustrated in Figure 2.21. Other designs can be proposed.
If multiple designs are proposed, we can select one that satisfies the Independence
Axiom and controls the temperature and humidity in a wide range. The new
design and the conventional refrigerator are compared.
F-fan Freezing
Capillary
tube room
Cold
air
Condenser Evaporator
R-fan Refrigerating
room
Compressor
Figure 2.21. A new design of a refrigerator that satisfies the Independence Axiom
56 Analytic Methods for Design Practice
Conventional refrigerator
Condenser
Damper
Refrigerating
room
Compressor
freedom (translation along the X and Y axes and rotation with respect to the Z axis).
Two products manufactured by Sanyo and Sharp will be compared.
Based on the Independence Axiom, we will select the better one.
Solution
The FR and DP of the highest level are stated as follows:
FR : Align the pixels of the LCD panels.
DP : The LCD holder that can align the pixels of the LCD panels
To align the pixels of all the LCD panels, the functional requirement is
decomposed as follows:
FR1 : Translate along the X axis = T(X).
FR2 : Translate along the Y axis = T(Y).
FR3 : Rotate with respect to the Z axis = R(Z).
The LCD projector uses three panels, and one is used as a reference one.
Therefore, holders with three degrees of freedom are needed for the two panels.
Sanyo Holder
Figure 2.24 shows the Sanyo holder. The holder is composed of three mechanisms
and is attached to each panel of Figure 2.23. All of them are lead screw structures.
If a screw moves, the attached plane moves accordingly. plate A is fixed to the
side frame. The LCD panel is attached to plate C. The three lead screws are
design parameters.
DP1 : The lead screw for conjunction of plate B and screw 1
DP2 : The lead screw for conjunction of plate C and screw 2
DP3 : The lead screw for conjunction of plate B and screw 3
58 Analytic Methods for Design Practice
Z X
If plate B is rotated, the LCD panel moves along the T(X) and R(Z) axes. Plate C
and the LCD panel move with plate B. If screw 1 is rotated, the LCD panel moves
in T(X) and R(Z). Therefore, DP1 affects FR1 and FR3 . DP2 is composed of
screw 2 and plate C. If screw 2 is rotated, plate C and the LCD panel move in the
Y direction. Since the rotation of screw 2 changes the position of the LCD panel in
the T(X) axis, DP2 only affects FR2 . DP3 has the same function as DP1 . The
rotation of screw 3 moves the LCD panel in the T(X) and R(Z) axes and DP3
affects FR1 and FR3 . The design equation is
ªT ( X )º ªX 0 X º ª DP1 º
« T (Y ) » «0 X 0 »» «« DP2 »» (2.75)
« » «
«¬ R( Z ) »¼ «¬ X 0 X »¼ «¬ DP3 »¼
Sharp Holder
The Sharp holder also has three mechanisms as illustrated in Figure 2.25. One is a
simple lead screw and the other two have a guideway and a guide boss. Plate A is
fixed to the side frame. The LCD panel is attached to plate C. The following three
DPs are defined:
DP1 : The lead screw for conjunction of plate B and screw 1
DP2 : The lead screw for conjunction of plate C, plate D, boss 1, boss 2 and
screw 2
DP3 : The lead screw for conjunction of plate B, plate E, boss 2 and screw 3
If screw 1 is rotated, plate B moves in the X direction. Since the LCD panel is
attached to plate C, it moves with plate B in the X direction. Thus, DP1 only
affects FR1 . Rotating screw 2 moves plate D in the X direction and the wall of the
guideway in plate D pushes boss 1. As a result, plate C moves along the Y axis
because the vertical groove in plate B guides the movement of boss 1 in the Y
direction. Therefore, rotating screw 2 moves the LCD panel in the Y direction and
DP2 only affects FR2 .
Screw 3 moves plate E in the X direction and the guideway of plate E pushes
boss 2. Since boss 1 does not have directional constraints, plate C rotates with
Boss 2
Boss 1
Guideway
Shift
linkage
Connected to
transmission
To valve Sprocket
Detent Tension
spring spring
Cam
Spring A Pawl
Figure 2.27 is the free body diagram of the forces acting on the pawl. FR is
the reaction force between the pawl and the sprocket. FC is the reaction force
between the pawl and the cam, FS is the spring force, FP is the force acting on the
pawl by the pin and P is the friction coefficient between the pawl and the cam.
As the slope of the tooth profile in the pawl increases, FR , FC and PFC increase
in order. FS is constant while the cam is engaged.
The functional requirements of the system are as follows:
FR1 : Engage the pawl in the locked position.
FR2 : Disengage the pawl from the locked position.
FR3 : Prevent accidental engagement.
Force exerted by
sprocket = FR
Pawl
FP
ȝFC
FC FS
ª FR1 º ªX 0 0 X X º ª DP1 º
« FR » «X « »
« 2» « X 0 0 X »» « DP2 »
« FR3 » «0 X X 0 0 » « DP3 » (2.77)
« » « »« »
« FR 4 » «0 X 0 X X » « DP4 »
« FR5 » «¬ X X 0 X X »¼ «¬ DP5 »¼
¬ ¼
Therefore, the current design is coupled. The reason is that the vehicle weight
transmitted by the automatic transmission is sustained by the pawl and the cam.
As the slope of the hill increases, the normal and friction forces on the cam
increase and disengagement of the parking mode becomes more difficult.
Tooth
profile =DP32
DP51
Pin = DP52
FR FP
Spring A/ Pawl
linkage =DP31 DP4 DP1 FS
Cam Tension
FC spring = DP2
ª FR 4 º ªX 0 0 0 0 º ª DP4 º
« FR » «X « »
« 2» « X 0 0 0 »» « DP2 »
« FR3 » «0 X X 0 0 » « DP3 » (2.78)
« » « »« »
« FR1 » «X 0 0 X 0 » « DP1 »
« FR5 » «¬ 0 0 0 0 X »¼ «¬ DP5 »¼
¬ ¼
FR31 : Control the force that pushes the pawl into sprockets.
FR32 : Generate the reaction force if the sprocket is turning.
DP51 : Nearly vertical surface of the pawl and the sprocket tooth profile
DP52 : Pin located collinearly with the force vector acting on the vertical
surface the pawl
To minimize the reaction force between the cam and the pawl, the reaction
force between the sprocket and the pawl should be close to the horizontal line. For
this, DP51 is nearly vertical. The small slope between the pawl and the sprocket is
used to minimize the reaction of the pawl. The design equation is diagonal. An
improved design is found by using the Independence Axiom. A better idea may be
created with application of the Independence Axiom.
The FRs and DPs satisfy the Independence Axiom as a decoupled design
because the design matrix is triangular as in Equation 2.79.
FR1 M 1 DP1
(2.80)
FR 2 M 2 DP2
M1 A11
DP1 (2.81)
M2 A21 A22
DP2
FR DP
C
M1 M2
C C
M121 M122
M121
M11 C S M21 C M22
M122
modules are utilized to control the module of the right hand side. This
represents a decoupled design.
Feedback junction (̺ F ): This is for a coupled design. In Figure 2.30, the
results of the right hand side return to the left hand side as feedback.
Thus, many repetitions are needed. When there are many feedback
junctions, the program is not manageable.
With the junctions, the hierarchical structure of FRs and DPs can be
represented by a tree structure. This is called a module junction structure
diagram. The example in Figure 2.29 is modified to that in Figure 2.31.
The module junction structure diagram can be modified to the flow of the
network type. Figure 2.32 shows the flow induced from Figure 2.31.
In this section, the application of axiomatic design is explained for the
development of software using conventional languages.
Customer 7 Software
needs 1 Code with system product
Define FRs architecture
2 6
Establish interfaces
el
Bu (Top
Map to DPs
roa mod
ild
5
the down
ch)
3
app ted
Identify classes
sof ap
-up ien
-
Decompose
tw pro
om ct or
4
are ac
Define
(B e obj
hie h)
rar
modules
th
ott
ch
ild
y
Bu
Identify leaves
(Full design matrix)
Figure 2.33. The axiomatic approach for objected oriented programming (V-model)
Axiomatic Design 69
Since the 1980s, the object oriented paradigm has received much attention in
software engineering. It is a new approach compared to process oriented
languages such as C, Pascal, Fortran, etc. The object oriented language, which is
popular these days, is appropriate for graphic user interface (GUI). The object
oriented technology provides methods to use existing programs. Common libraries
are prepared and specialized by customization. Also, a large program is divided
into independent objects and objects have relations by well defined interfaces.
Due to the above advantages, object oriented programming (OOP) is frequently
utilized in software development. The V-model has been proposed to exploit the
axiomatic approach in object oriented programming (Do 2000). In the V-model, a
designer defines the functional requirements of the software and establishes
independent modules from zigzagging decomposition. Each module is modified to
a class of the object oriented programming and coded. The process consists of two
steps: construction of the full design matrix with the top-down approach and
a b c d
Fd
F*
B
Gd
C G* Operation list
Hd
H*
D
Behavior list
(methods)
Figure 2.34. The full design matrix using object oriented programming
70 Analytic Methods for Design Practice
Figure 2.35. The class diagram for the full design matrix in Figure 2.34
coding the program with the bottom-up approach. The process is illustrated in
Figure 2.33. The top-down approach up to Step 4 is the same as the steps
explained in the previous section. Thus, this section describes the steps after Step
4.
Step 5. Identification of objects, attributes and operations
The full design matrix is constructed after the decomposition process. It
shows all FRs and DPs. An example of the full design matrix is presented
in Figure 2.34. Rows of Figure 2.34 are FRs and the columns are
DPs. ”X” means an algorithm of a logical relation between an FR and a
DP. The logical relation includes not only operators such as ”+” and ”
”
but also control statements such as ”if,” ”for,” etc. In Figure 2.34, a
rectangle with thick lines is an object. The object is composed of attributes
in columns (DPs) and the methods of the operations list within the
rectangle. The method of the object is the module. Therefore, an object
executes a method with attributes and satisfies the functional requirement.
Step 6. Establishment of interfaces between objects
An object is expressed by a class and a class is a template that defines the
format of the object. Classes share attributes that are the data structures
and behaviors. In this step, the relationships between classes are set up.
They are generalization, aggregation and association. Figure 2.35 presents
a class diagram according to the design matrix of Figure 2.34. We can see
the data and their functions in Figure 2.35 and the class diagram shows the
relations of classes. The design process is shown by the aforementioned
flow. Thus, software development easily proceeds with these.
Step 7. Coding with system architecture
Coding is the programming process based on the classes and their
relationships. The flow chart of the design matrix helps with the coding.
Axiomatic Design 71
2.6 Discussion
As explained earlier, the two axioms are independent of each other. Thus, we have
to apply them separately. Generally, the Independence Axiom should be satisfied
first. In many cases, the design is terminated only with the application of the
Independence Axiom. When both of the axioms are utilized, the flow in Figure
2.20 is recommended.
When we apply the Independence Axiom, the ideal design should be kept in
mind. The numbers of FRs and DPs are the same in an ideal design. The design
matrix should be a square diagonal or triangular one. If the numbers are different,
the design is coupled. When the number of DPs is smaller, new DPs should be
added. In a redundant design where the number of DPs is larger, the number
should be reduced or some specific DPs should be fixed.
Suppose we have the following redundant design:
ª DP1 º
ª FR1 º ªX X 0 º«
« FR » «0 DP2 »» (2.82)
¬ 2¼ ¬ X X »¼ «
«¬ DP3 »¼
First, we can fix DP2 . Then the design becomes an uncoupled one. That is,
redundant parameters are fixed to make the design uncoupled or decoupled with
the rest of the parameters.
The Information Axiom is utilized to quantitatively evaluate a design that
satisfies the Independence Axiom. It is especially useful when multiple designs
are compared. When multiple designs, which satisfy the Independence Axiom, are
found, the one with the minimum information content is selected as the final
design.
Basically, the axiomatic design can be exploited in creating a new design or
evaluating existing designs. It is quite useful in the conceptual design of new
products. Although the history of the method is relatively short, the usefulness has
been verified through many examples. There are some common responses from
application designers. First, they tend to easily agree with the axioms and think
that they can use them right away. However, they have difficulties in testing the
axioms with their existing products. In most cases, they tend to look at the designs
with previous concepts, not from an axiomatic viewpoint. Many designers tend to
stop applying axiomatic design at this stage. However, if the designers overcome
this stage, they realize the usefulness of axiomatic design. It is important not to
consider the existing products when the functional requirements are defined.
Instead, designers should think about the functional requirements in a solution
neutral environment. In recent research, axiomatic design is utilized in detailed
designs. Later examples will demonstrate how axiomatic design is applied to the
detailed design process of structures.
72 Analytic Methods for Design Practice
2.7 Exercises
2.2 Analyze the design of a cellular phone with the Independence Axiom.
2.3 Find a product that uses the idea of physical integration and analyze it with
the Independence Axiom.
Design a fuel tank that satisfies above four FRs. The new tank should cost
less than the current one.
ª f1 º ªX 0 0 º ª x1 º
«f » «X X 0 »» «« x 2 »»
« 2» «
¬« f 3 ¼» ¬« X X X ¼» ¬« x 3 ¼»
f1 x12
f2 x12 x1 x 2 x 22
f3 x12 x 22 x 2 x3 x32
2.6 Calculate the reangularity and semangularity of the design matrices and
discuss the characteristics.
Axiomatic Design 73
ª FR º ª3 1º ª DP1 º
Design 1 « 1 » «2 4» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
ª FR º ª3 0º ª DP1 º
Design 2 « 1 » «8 5» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
(1) Draw an FR–DP graph for each design and explain the order of the
design process.
(2) Calculate the reangularity and semangularity of each design and
compare the results.
ª FR º ª3 0º ª DP1 º
Design 1 « 1 » «0 4» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
ª FR º ª3 0º ª DP1 º
Design 2 « 1 » «2 4» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
74 Analytic Methods for Design Practice
ª FR º ª3 1º ª DP1 º
Design 3 « 1 » «2 4» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
(a) When DPs have uniform distribution in the DP range, calculate the
information content by the graphical method and compare them.
(b) Calculate the information content in the same manner as (a) in the
FR range.
(c) Distributions of the design parameters are as follows:
3 3
Distribution of DP1 pGDP1 3
(GDP1 ) 2
4('DP1 ) 4('DP1 )
1 1
Distribution of DP2 pGDP2 2
GDP2 when
('DP2 ) 'DP2
1 1
GDP2 0 , and pGDP2 2
GDP2 when GDP2 t 0
('DP2 ) 'DP2
2.10 We have a decoupled design and one element of the design matrix is
expressed by an unknown x. The value of x is 0.5, 1 or 2 in the following
design equation:
ª 0 FR1 2 º ª1 0 º ª 0 DP1 1º
«0 FR 2» «1 x » «0 DP 1»
¬ 2 ¼ ¬ ¼¬ 2 ¼
2.11 Make up a problem for buying a laptop computer in the same way as in the
house buying problem and solve it.
Axiomatic Design 75
ª FR º ª100 0 º ª DP1 º
Design 1 « 1 » « 0 100» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
ª FR º ª1 0º ª DP1 º
Design 2 « 1 » «1 1» « DP »
¬ FR 2 ¼ ¬ ¼¬ 2 ¼
The targets of FRs are 10 and 20, respectively. The allowable tolerances
are 9.5 FR1 10.5 and 19.5 FR2 20.5 , and the tolerance of each
DP is r 2.0.
2.14 Expand the graphical method for a decoupled design to a coupled design.
76 Analytic Methods for Design Practice
n § wFR ·
¦ ¨¨ i ¸'DPj
¸
j z i© wDP j ¹
j 1
Theorem 2.A.21 [Robust Design when the System Has a Nonuniform pdf]
If the probability distribution function (pdf) of the FR in the design range is
nonuniform, the probability of success is equal to one when the system
range is inside the design range.
Beam generator
Scanning head
Diode laser
Scanning head
Marking position
Figure 2.B.2. Schematic view of the inside of a laser marker
plates. The points are starting points of the marking process and illustrated
as hollow points as shown in Figure 2.B.3.
(2) The YAG laser is turned off. The diode laser is turned on. The solid points
in Figure 2.B.3a are the final destinations of the diode laser.
(3) The adjuster of the diode laser is utilized to make two identical points as
illustrated in Figure 2.B.3b. If the two points match, the angles and the
final destinations from the YAG and diode lasers are considered identical.
Now, we are sure that the two lasers have the same routes.
(4) The marking is simulated with the diode laser. That is, the motors in the
scanning head are simulated by a computer program. The program is the
one specifically developed for the marking process. As mentioned earlier,
the marking result is visible.
(5) If the results are validated, the test is terminated.
Many problems occur in the adjuster of the diode laser. Currently, screws are
used for the adjustment. Precise adjustment is difficult to obtain, since tolerances
and human errors are involved. Thus, the adjustment is a long process.
Since the laser marking machine has already been commercialized, there is an
existing design for the diode beam adjuster. Therefore, it is necessary to define the
functional requirements and corresponding design parameters to evaluate the
existing device. The relationship between FRs and DPs can be expressed by a
design matrix. The FRs of the existing device is defined as follows:
FR1 : Align the vertical position of the diode laser beam.
FR2 : Align the vertical angle of the diode laser beam.
FR3 : Align the horizontal position of the diode laser beam.
FR4 : Align the horizontal angle of the diode laser beam.
FR5 : Fix the beam alignment.
Figure 2.B.4 illustrates each functional requirement. The two beams from the
YAG and diode lasers should be properly matched. First, the horizontal and
vertical destinations of the diode laser should be the same as those of the YAG
laser ( FR1 , FR3 ). Second, the angles of the beams must be the same ( FR2 , FR4 ).
Figure 2.B.5 illustrates the existing product. DPs corresponding to FRs are
DP1
DP2
DP3
defined as follows:
DP1 : Vertically moving component
DP2 : Supporting block
DP3 : Fixing screw
ª FR1 º ªX 0 0º
« FR » «X
« 2» « 0 0 »» ª DP1 º
« FR3 » «0 « »
X 0 » « DP2 » (2.B.1)
« » « »
« FR 4 » «0 X 0 » «¬ DP3 »¼
« FR5 » «¬ 0 0 X »¼
¬ ¼
The design in Equation 2.B.1 is a coupled design because the number of DPs is
less than the number of FRs. When we move the solid points in Figure 2.B.3a
( DP1 ), the vertical angle also varies because FR1 and FR2 are coupled by DP1 .
In a similar manner, when we move the horizontal position ( DP2 ) the aligned
angle can vary. If a design is coupled in the way of Equation 2.B.1, it can be
decoupled by adding new DPs to make the numbers of FRs and DPs equal.
86 Analytic Methods for Design Practice
ª FR1 º ªX º
« FR » « X »>DP1 @ (2.B.2)
¬ 2¼ ¬ ¼
We can think of a design that has independent design parameters for the
vertical position and angle. As a result, two designs are made. The first one is
illustrated in Figure 2.B.6. The fastener at the back ( DP1 ) controls the vertical
position and the front one ( DP2 ) controls the vertical angle. After the position is
fixed, the fastener is tightened by a screw. The design matrix for Figure 2.B.6 is
decoupled as follows:
ª FR1 º ªX 0 º ª DP1 º
« FR » «X « » (2.B.3)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
ª FR1 º ªX 0 º ª DP1 º
« FR » «X « » (2.B.4)
¬ 2¼ ¬ X »¼ ¬ DP2 ¼
DP2 DP1
DP2 DP1
The above two designs show a method to make a decoupled design by defining
new design parameters. The method can be expanded for other DPs.
Figure 2.B.8 is the expansion of Figure 2.B.7. Five FRs are the same as before
and DPs for Figure 2.B.8 are as follows:
DP1 : Upper rear screw
DP2 : Upper front screw
DP3 : Side rear screw
DP4 : Side front screw
DP5 : Fixing screw
DP3 and DP4 are similar to the aforementioned DP1 and DP2 . We can think of
two designs in Figure 2.B.6 and Figure 2.B.7. DP3 and DP4 of Figure 2.B.8 are
selected in the same manner as in Figure 2.B.7. The expanded design is also a
decoupled design as
ª FR1 º ªX 0 0 0 0 º ª DP1 º
« FR » «X « »
« 2» « X 0 0 0 »» « DP2 »
« FR3 » «0 0 X 0 0 » « DP3 » (2.B.5)
« » « »« »
« FR 4 » «0 0 X X 0 » « DP4 »
« FR5 » «¬ 0 0 0 0 X »¼ «¬ DP5 »¼
¬ ¼
Therefore, the new designs satisfy the Independence Axiom. Using various new
ideas, we can create other designs that satisfy the Independence Axiom.
DP1
DP2
DP3
DP4 DP5
where r is the radius of the screw, T is the angle of the screw and p is the pitch.
When the radius is 1.5 mm and the pitch is 1 mm, the slope is 0.106 as shown
in Figure 2.B.10. Considering the environmental and geometrical aspects of a
certain existing design, the design matrices of the two designs are as follows:
ª FR º ª 1 0º ª DP1 º
Model #1: « 1 » « 4 4» « DP » (2.B.7)
¬ FR2 ¼ ¬ ¼¬ 2 ¼
ª FR º 1 ª 1 0º ª DP1 º
Model #2: « 1 » « 4 4» « DP » (2.B.8)
¬ FR2 ¼ 3S ¬ ¼¬ 2 ¼
DP
DP
Model #1 Model #2
FR Model #1
(distance)
mm
m=1
Model #2
m = 1/3ʌ
DP (distance) mm
1
pi [u (GDPi ('DPi )) u (GDPi ('DPi ))], i 1, 2 (2.B.9)
2 u1
Therefore, model #2 is better than model #1. It is the same for the design of DP3
and DP4 . In conclusion, the design in Figure 2.B.8 is determined as the final
design.
2.B.4 Summary
Strength analysis
Drawing generation
The FRs are mapped into design parameters in the physical domain.
DP1 : A set of basic data
DP2 : The three-dimensional shape structure for the panel and the funnel
DP3 : Loading conditions for the panel and the funnel
DP4 : A set of drawing data
ª FR1 º ªX 0 0 X º ª DP1 º
« FR » «X « »
« 2» « X 0 X »» « DP2 »
(2.C.1)
« FR3 » «X X X X » « DP3 »
« » « »« »
¬ FR4 ¼ ¬X X 0 X ¼ ¬ DP4 ¼
FR1 is enabled by the basic data and drawings given by a customer. In the
same manner, FR 4 is enabled by the basic data, the three-dimensional shape data
and drawings. Therefore, FR1 and FR 4 are coupled in the conventional design
process.
92 Analytic Methods for Design Practice
ª FR1 º ªX 0 0 0 º ª DP1 º
« FR » «X « »
« 2» « X 0 0 »» « DP2 »
(2.C.2)
« FR3 » «X X X 0 » « DP3 »
« » « »« »
¬ FR4 ¼ ¬X X 0 X ¼ ¬ DP4 ¼
The selected design parameters and the design matrix are as follows:
Axiomatic Design 93
ª FR11 º ªX 0 º ª DP11 º
« FR » «X « » (2.C.3)
¬ 12 ¼ ¬ X »¼ ¬ DP12 ¼
Since the ID number of a new product is used before making the specific
product data, the design matrix in Equation 2.C.3 is a decoupled one. The design
parameters for FR21 , FR22 and FR23 and the design matrix are as follows:
ª FR21 º ªX 0 0 º ª DP21 º
« FR » «X X 0 »» «« DP22 »» (2.C.4)
« 22 » «
«¬ FR23 »¼ «¬ X X X »¼ «¬ DP23 »¼
ª FR41 º ªX 0 º ª DP41 º
« FR » «0 « » (2.C.5)
¬ 42 ¼ ¬ X »¼ ¬ DP42 ¼
FR Top
M1 M2 M3 M4
C C S
Figure 2.C.3. Module junction structure diagram of the design system for the TV glass
bulb
ª FR231 º ªX 0 0 º ª DP231 º
« FR » «0 X 0 »» «« DP232 »» (2.C.6)
« 232 » «
«¬ FR233 »¼ «¬ 0 0 X »¼ «¬ DP233 »¼
M231
M41
M11 C M12 C M21 C M22 C M232 S M3 C S
M42
M233
M1 M2 M3 M4
User Interface
The user interface is developed by using X-window (MOTIF), which is a standard
graphics tool of UNIX (Heller 1994). The user interface provides various menu
systems and graphic displays. An example is illustrated in Figure 2.C.5.
Database
A commercial database management system is utilized to handle enormous data
(Oracle Co. 1990). The transaction commands such as insert, delete, update and
inquire, which are offered by the database management system, are utilized. Input
data or design variables are given by the designer and stored in the database. The
design variables can be viewed from the database contents such as tables and
records. Various graphs can help the designer in decision making. The designer
can find the mismatch of data from the graphs. The program can automate the
manual process of the conventional design. In a conventional design, a rough final
drawing is needed for the decision making process.
graphics library is used for this purpose (Hewlett–Packard Co. 1993). An example
is illustrated in Figure 2.C.6.
Strength Analysis
After the three-dimensional shape is constructed, strength analysis is performed. A
special mesh generation routine is used to generate an input file for strength
analysis. The failure criterion uses the maximum normal stress theory. If the
result of the strength analysis is not acceptable, an iterative process with the three-
dimensional shape generation module is carried out. This process is shown as the
dotted line in Figure 2.C.4. A commercial software system called ANSYS is
employed for the strength analysis (ANSYS Inc. 1993). Results of the strength
analysis are shown in Figure 2.C.7.
Drawing Generation
When all the activities are finished, the results are drawn. A commercial CAD
(computer-aided design) system called Unigraphics is used (Electronic Data
Systems Co. 1993). An example of the final drawing is illustrated in Figure 2.C.8.
2.C.5 Summary
and improved by the axiomatic approach. The approach uses the general methods
defined in the axiomatic design, such as the zigzagging process, module junctions
and system architecture for the flow. The software is designed based on the
improved design process at the early stage of software development. It is noted
that the flow of the software execution is the same as the design process.
Axiomatic Design 99
The V-model and the steps introduced in Section 2.5.3 are utilized. An automatic
design system is developed for conceptual and detailed designs. First, the
conventional design method is investigated. Customer attributes (CAs) are defined
by interviewing practical designers.
Table 2.D.1. Top level FRs of the design system for the EPS cushioning package of a
monitor
FRx DM DPx
1 Set up the options X O O O O Option data
Construct the data set for modelling Data for modelling and
2 X X O O O
and simulation drop test
Generate an FEM model of the Design variables of
3 X X X O O
cushioning material cushioning material
Recommend a good design value
4 X X X X O DYNA3D input deck
through simulation analysis
5 Manage the design data X X X X X Data manager
FR2.x DP2. x
Construct the data set for modelling and DM Data for modelling and
simulation drop test
1 Construct modelling data for monitor X O O Modelling data for monitor
Construct modelling data for cushioning Modelling data for
2 O X O
material cushioning material
FR21223. x DP21223. x
Translate the files for nodes and DM21223
Files for nodes and elements
elements into the DYNA format files
1 Read the files X O O O File names
X X O O Numbers of nodes and
2 Calculate the adding quantity
elements
3 Save the files in DYNA format X X X O DYNA format
As shown in Table 2.D.1, FR2 is “construct the data set for modelling and
simulation.” DP2 is input data such as modelling of the monitor and input data for
analysis. From DP2 , the detailed operations of FR2 are defined. That is, various
data constructions should be made for the modelling data, analysis data and
Axiomatic Design 101
B A
management module, the module for modeling, the module for the drop test
analysis and the module for automatic analysis and design. The main module
controls the graphic user interface for input and the overall design process. The
data management module has the function of managing the data for the system. It
handles the interface with external systems, the files for material properties and the
database for standard orthogonal arrays. The module for modelling generates the
input data for analysis, which are shape sizes and the input file for LS/INGRID
that automatically generates meshes for the finite element analysis. The module
for the drop test generates an input file for the analysis system LS/DYNA3D and
executes the system. The module for analysis and design performs the analysis,
analyzes the results and proposes a new design.
Each module is defined as an object. An object consists of the functions in the
row of the design matrix and the attributes of the column. For example, FR3 in
Figure 2.D.1 is defined by object A, which includes object B for inner shapes and
object C for external shapes. Object B is composed of four objects for four
positions from the user input and one object for common data. In the same manner,
object C has five objects.
CPackDesign
int designStep;
……
FileSave();
…….
and the class “CushionModel” handles the shapes of the cushioning materials.
They receive data from a user, perform coordinate transformation and generate the
input file for the finite element analysis. Finally, the class “IterationAnal” selects
an orthogonal array, performs the drop tests according to the orthogonal array and
analyzes the results. If the aforementioned basic model is varied, derived classes
can be made by inheritance from the classes “BasicModel” and “CushionModel.”
Using the above process, a software system is coded. The overall menus are
illustrated in Figure 2.D.3. The left hand side is for input and the right hand side is
for displaying results. The execution of the system is classified into two modes.
One is analysis with given parameters and the other is a design process for
multiple analyses with changing variables.
2.D.3 Summary
References
Albano LD, Suh NP (1992) Axiomatic Approach to Structural Design. Research in
Engineering Design l(4):171í183
Albano LD, Suh NP (1993) The Information Axiom and Its Implication. DE–Vol
66, Intelligent Concurrent Design: Fundamentals, Methodology, Modeling and
Practice, ASME
ANSYS Inc. (2004) ANSYS User’s Manual Version 8.1. Canonsburg, PA
104 Analytic Methods for Design Practice
Do SH, Park GJ (2001) Application of Design Axioms for Glass Bulb Design and
Software Development for Design Automation. Journal of Mechanical Design,
Transactions of the ASME 123(3):322í329
Do SH, Suh NP (2000) Axiomatic Design of Software Systems. CIRP Annals
49(1):95í100
Electronic Data Systems Co., (1993) User Function Programming Manual. Livonia,
MI
El-Haik B, Yang K (1999) The Components of Complexity in Engineering Design.
IIE Transactions 31:925í934
Frey DD, Jahangir E, Engelhardt F (2000) Computing the Information Content of
Decoupled Designs. Research in Engineering Design 12:90í102
Gebala DA, Suh NP (1992) An Application of Axiomatic Design. Research in
Engineering Design 3:149í162
Heller D, Ferguson PM (1994) Motif Programming Manual. O’Reilly &
Associates, Setastopol, CA
Hewlett–Packard Co. (1991) Starbase Graphics Techniques. TX
Hwang KH, Lee KW, Park GJ (2001) Robust Optimization of an Automobile
Rearview Mirror for Vibration Reduction. Structural and Multidisciplinary
Optimization 21(4):300í308
Kim SJ, Suh NP, Kim SK (1991) Design of Software Systems Based on Axiomatic
Design. Annals of CIRP, 40(1): 165í170
Lee J, Cho K, Lee K (1994) A New Control System of a Household Refrigerator-
Freezer. International Refrigeration Conference, Purdue University, IN
Livermore Software Technology Co. (1998) LS/INGRID Manual Ver. 3.5.
Livermore, CA
Livermore Software Technology Co. (1998) LS/DYNA Theory Manual.
Livermore, CA
Livermore Software Technology Co. (1999) LS/DYNA User’s Manual Ver. 950.
Livermore, CA
Mood AM, Graybill FA, Boes DC (1963) Introduction to the Theory of Statistics.
3rd ed. McGraw–Hill, New York
NSF (1998) Axiomatic Design Workshop for Professors. MIT, Cambridge, MA
Oracle Co., (1990) ORACLE 3GL Programmers Guide. Redwood Shore, CA
Park GJ, Do SH, Lee JW, (1995) Construction of Automatic Design Systems.
Final Report, Samsung Corning Co. (in Korean)
Phadke MS (1989) Quality Engineering Using Robust Design. Prentice Hall,
Englewood Cliffs, NJ
Rinderle JR, Suh NP (1982) Measures of Functional Coupling in Design.
Transactions of ASME, Journal of Engineering for Industry 104:383í388
Shin GS, Park GJ (2004) Supplementary Beam Adjuster for a Laser Device. Patent
No. 10-045398, Korea
Shin MK, Hong SW, Park GJ (2001) Axiomatic Design of the Motor-Driven
Tilt/Telescopic Steering System for Safety and Vibration. Proceedings of the
Institution of Mechanical Engineers, Part D, Journal of Automobile Engineering
215(2):179í187
Suh NP (1990) The Principles of Design. Oxford University Press, New York
Axiomatic Design 105
3.1 Introduction
Optimization was originally created by mathematicians out of curiosity. Basically,
a given mathematical function is minimized or maximized. Mathematical
optimization methods have been developed to solve a variety of different
optimization problem formulations. For example, finding the firing angle of a
cannon can be viewed as an optimization problem to obtain the maximum distance.
The flying distance can be expressed by the function of the angle. In optimization,
optimum values of the variables are found to minimize or maximize a function that
is expressed by the variables.
Optimization originates from the variational method, which finds an optimum
function for a given objective functional. In the variational method, a functional is
defined by the total potential energy of a system, and minimization of the
functional results in the governing equation of the system. When there are
constraints, they are multiplied by Lagrange multipliers and added to the objective
functional as penalties. This concept is utilized by modern optimization as well.
The Brachistochrone problem is the well known problem of minimizing a
functional. It calculates the shortest path where we move under gravity. Also, it
can be proved that the shortest path between two points is a straight line. Algebra
and variational calculus are adopted to solve such minimization or maximization
problems.
The variational methods in the early stage are mainly utilized to solve
mechanics problems using energy methods. Modern optimization methods have
been established as engineers apply them to design. In optimization, algebraic
functions are utilized instead of functionals and constraints are emphasized due to
the characteristics of design. Optimization theory is used to solve problems
formulated as follows:
Find b Rn (3.1a)
to minimize f (b) (3.1b)
bL d b d bU (3.1e)
Initial design
Performance analysis
Yes
Satisfied? End
No
Design change based on
experience and intuition
Initial design
Yes
Are convergence criteria satisfied? End
No
decision making processes, where the decisions are made by either the designer or
the computer.
Optimization gives an excellent solution when the formulation is accurate. In
other words, a mathematical optimum solution is obtained. However, when the
formulation is not appropriate, the optimum solution can be obsolete in real design
terms because the solution is obtained only from the formulated problem.
Meanwhile, a mathematically excellent solution may satisfy constraints at the
utmost limits. In this case, if tolerances or slight changes exist in design variables,
the constraints can be easily violated. Although many methods are proposed to
overcome these problems, optimization still has numerical difficulties in solving
mathematically sensitive problems.
Optimal design is the terminology for a technique to solve the problems
formulated as Equation 3.1. As mentioned earlier, it can be efficiently utilized in
detailed design. Also, it should be remembered that the impact of detailed design
is generally smaller than that of conceptual design. Therefore, optimization should
be carefully applied to appropriate cases.
Step 1. Design variables in Equation 3.1a are defined. Design variables are
parameters that are to be determined in the design process. In structural
design, they can be section properties, thickness and shapes, etc.
Step 2. The objective function in Equation 3.1b is defined. The objective
function is a performance measure that is to be improved (e.g., min or
max) in design. In structural design, the objective function can be the
weight of a structure, a specific response, etc.
Step 3. The constraints in Equations 3.1c and 3.1d are defined. Generally,
design is performed based on some design specifications and conditions.
They should be transformed to constraint functions.
The problem formulation in Equation 3.1 is referred to as the standard form for
constrained optimization. Problems with different forms can be modified to the
standard one. For example, a maximizing problem is transformed to a minimizing
problem as follows:
where
F (b) { f (b) (3.3)
When a constraint in Equation 3.1d has the greater than type inequality, it can be
transformed as follows:
l l
ཛ ཛྷ ཝ
l 1 2 3
x
T
y
P
Step 2. The objective function is defined by the weight of the structure. The
objective function is
gi | V i | V all d 0 , i 1, 2, 3 (3.6)
g4 G x G x, all d 0 (3.7)
g5 G y G y , all d 0 (3.8)
2lPx
Gx (3.9)
b1 E
2lPy
Gy (3.10)
(b1 2b2 ) E
where E is Young’s modulus, and Px ( P cos T ) and Py ( P sin T ) are the lateral
and vertical components of the external load P, respectively. The stresses at the
members are as follows:
1 §¨ Px Py ·
¸
V1 (3.11)
¨
2 © b1 b1 2b2 ¸
¹
2 Py
V2 (3.12)
b1 2b2
1 §¨ Py P ·
V3 x¸ (3.13)
2 ¨© b1 2b2 b1 ¸¹
2R
P P
d V all (3.15)
A 2SRt
Also, the following buckling constraint is defined from the condition that
the compressive stress should be within the allowable range:
S 2 EI S 3 ER 3 t
Pd (3.16)
4l 2 4l 2
D
P P
d G
d 4G
Spring constant: K (3.20)
8D 3 N
8kPD
Shear stress: W (3.21)
Sd 3
4 D d 0.615d
Wahl stress concentration factor: k (3.22)
4 D 4d D
d G
Frequency of surge waves: Z 2
(3.23)
2SD N 2U
where G is the shear modulus, U is the density of the material and k is the Wahl
stress concentration factor, which is determined from the experiments. Now
formulate an optimization problem.
Solution
An optimization problem is formulated as follows:
Step 1. Design variables are defined as d wire diameter, D mean coil
diameter and N number of active coils.
Step 2. The objective function is the mass as
1
f (b) NS 2 Dd 2 U (3.24)
4
Step 3. Constraints are defined by using the above design equations. The
displacement G should be larger than a certain value ' to store enough
energy. From Equations 3.19 and 3.20, the following constraint is defined:
Design Optimization 115
P 8 PD 3 N
' G ' ' d0 (3.25)
K d 4G
The shear stress W should be less than the allowable value W all . From
Equations 3.21 and 3.22,
8PD § 4 D d 0.615d ·
3 ¨
¸ W all d 0 (3.26)
Sd © 4 D 4d D ¹
d G
Z0 Z Z0 2
d0 (3.27)
2SD N 2U
D d Dlim d 0 (3.28)
Step 4. The lower and upper bounds on design variables are defined as
As shown from the examples, problem formulation is the first and the most
important step in engineering optimization. The sequence of the formulation is the
definition of design variables, the objective function and constraints. The
objective function and constraints should be expressed by functions of design
variables. Although small-scale problems are demonstrated in the examples, large-
scale problems also follow the same procedure.
Graphical optimization finds an optimum solution by drawing the objective and the
constraint functions on a plane. Since the graphics can be drawn on a two-
dimensional plane, problems with two design variables can only be solved.
Therefore, it is used for the explanation of an optimization method. The optimum
solution is obtained by the following sequence:
Step 1. Draw constraints on a two-dimensional plane and find the feasible
region.
Step 2. Draw the iso-objective contours.
Step 3. Find a design point that makes the objective function a minimum in
the feasible region.
Solution
The graph of each function is illustrated in Figure 3.6. The feasible region is
represented by three constraints. The concentric circles represent iso-objective
contours. In the feasible region, the optimum point is b [1 1]T . It is noted that
the constraint g1 is the most important for the optimum solution and it is a tangent
line for an iso-objective contour. At the optimum point, g1 is active.
b2
f 0.75 Iso-objective contour
2 Optimum
f 0.5
(1.5, 1.5) b* (1, 1)
*
f (b ) 0.5
1
Feasible g1 b1 b2 2 0
region
b1
1 2
The increment of the function f can be obtained from the linear expansion of
the Taylor series
Gf T f (b * )(b b * ) (3.34)
f (b * ) 0 (3.35)
That is, the gradient of f is the zero vector at b*. In the partial differentiation form,
Equation 3.35 becomes
wf (b * )
0, i 1, , n (3.36)
wbi
(b b * ) T H (b b * ) ! 0 (3.37)
f (b * ) 0 (3.38)
f cc(b * ) ! 0 (3.39)
To obtain a local minimum for a function f(b): (1) We obtain a solution of the
simultaneous equation of Equation 3.35 or 3.36. (2) We check whether the
Hessian matrix is positive definite at the solution of the first step. When the
Hessian matrix is positive semidefinite, we need to check for a higher order.
Details are given in the references (Arora 2004, Rao 1996).
Solution
Necessary condition:
wf
2b1 4b2 0 (3.41)
wb1
wf
10b2 4b1 2 0 (3.42)
wb2
Sufficient condition:
Design Optimization 119
ª w2 f w2 f º
« 2 »
« wb1 wb1 wb2 » ª 2 4º
H « 4 10 » (3.44)
« w2 f w2 f » ¬ ¼
« »
«¬ wb2 wb1 wb22 »¼
The Hessian matrix in Equation 3.44 is positive definite. Thus, the solution in
Equation 3.43 satisfies the necessary and sufficient conditions for the local
minimum.
l m
L(b, v, u, s) f (b) ¦ vi hi (b) ¦ u j ( g j (b) s 2j ) (3.46)
i 1 j 1
wL wf l whi m wg j
{ ¦ v*i ¦ u *j 0, k 1, , n (3.47a)
wbk wbk i 1 wbk j 1 wbk
u *j s j 0, j 1, , m (3.47d)
u *j t 0, j 1, , m (3.47e)
subject to b1 b2 0 (3.48b)
b2 3b1 t 0 (3.48c)
Solution
Since the problem has constraints, the Lagrangian function is defined as
wL
2b1 4b2 v 3u 0 (3.50a)
wb1
wL
10b2 4b1 2 v u 0 (3.50b)
wb2
b1 b2 0 (3.50c)
Design Optimization 121
b2 3b1 s 2 0 (3.50d)
us 0 (3.50e)
ut0 (3.50f)
b1 b2 0 (3.51c)
b2 3b1 s 2 0 (3.51d)
b1 b2 0 (3.52c)
ª b1 º ª 0.1º
«b » « 0.1 »
« 2» « »
«v» « 0.6 » (3.53)
« 2» « »
«s » « 0.4 »
«¬ u »¼ «¬ 0 »¼
122 Analytic Methods for Design Practice
where k is the iteration number. At the kth iteration, the change of the design
variable vector 'b ( k ) is evaluated based on the characteristics of the algorithm and
Design Optimization 123
added to the current design b ( k ) , and a new design b ( k 1) is obtained. The change
'b ( k ) consists of two parts as follows:
'b ( k ) D k d (k ) (3.55)
d ( k 1)
D k d (k ) d (k )
D k 1d ( k 1) (k )
b ( k 1)
b
d ( k 1)
b ( k 1)
f (b ( k 1) ) f (b ( k ) D k d ( k ) ) f (b ( k ) ) (3.56)
where constraints are not considered. When there are constraints, active
constraints are modified to the penalty function and added to the objective function.
Then the augmented objective function is treated as an unconstrained objective
function. This aspect will be explained later. The condition in Equation 3.56 is
called a descent condition. If we expand it with a Taylor series expansion
T
f (b ( k 1) ) # f (b ( k ) ) D k c ( k ) d ( k ) f (b ( k ) ) (3.57)
|| b ( k 1) b* ||
d K , K t 0, p t 1 (3.58)
|| b ( k ) b* || p
n
subject to a11b1 a12 b2 a1n bn ¦ a1i bi e1 t 0
i 1
126 Analytic Methods for Design Practice
n
a 21b1 a 22 b2 a 2 n bn ¦ a 2i bi e2 t 0
i 1
(3.60c)
n
a m1b1 a m 2 b2 a mn bn ¦ a mi bi em t 0
i 1
bi t 0, i 1, , n (3.60d)
Equation 3.61 shows that the right hand side can always be zero or positive.
(2) When Equation 3.60c is an inequality constraint
When Equation 3.60c is a less than type (<) function, a slack variable si is
added as follows:
a i1b1 a i 2 b2 a in bn s i ei (3.62a)
si t 0 (3.62b)
a i1b1 a i 2 b2 a in bn s i ei (3.62c)
si t 0 (3.62d)
where
bi t 0, bi t 0 (3.64)
Find b (3.65a)
to minimize f cT b (3.65b)
subject to Ab e (3.65c)
bt0 (3.65d)
b2 t 0 (3.66d)
128 Analytic Methods for Design Practice
Solution
Step 3. If slack and surplus variables are added, Equation 3.69 is converted
to the standard form as follows:
Minimize f (z ) z1 z 2 2 z 3 (3.71a)
subject to 2 z1 2 z 2 3 z3 z 4 8 (3.71b)
5 z1 5 z 2 2 z3 z5 1 (3.71c)
z i t 0, i 1, , 5 (3.71d)
Design Optimization 129
Since all the functions are linear in a linear programming problem, the feasible set
is convex from Theorem 3.A.3. Therefore, a local minimum is the global
minimum from Theorem 3.A.4. In Equation 3.60c, the number of linearly
independent constraints m must be less than the number of design variables n. If
m ! n , there is no solution satisfying constraints, and if m n only one solution
exists for the constraints and this is the optimum solution. When there are
inequality constraints, the optimum solution resides on the boundary of the feasible
set. If the optimum solution is inside the boundaries, the problem is the same as an
unconstrained problem. If the objective function is linear without constraints, the
problem is unbounded. Therefore, the optimum solution should be on a boundary.
From the condition m n, the number of solutions satisfying Equation 3.60c is
infinite and the one with the minimum objective function is the optimum out of the
infinite solutions. The above concept can be assured from the following example:
subject to b1 b2 d 16 (3.72b)
1 1
b1 b2 d 1 (3.72c)
28 14
b2
25 G
b1 b2
20 1
14 24
F
b1 b2 16
15
E
f= D
–7 I
10 36
0
b1 b2
f= 1
–4 28 14
5 800 C
f=
–2 f=
400 –8
800
B
J H
b1
A
5 10 15 20 25 30
Figure 3.8. Graphical presentation of Example 3.8
130 Analytic Methods for Design Practice
1 1
b1 b2 d 1 (3.72d)
14 24
b1 , b2 t 0 (3.72e)
Solution
Figure 3.8 illustrates each function in a plane. The feasible region is presented and
the dotted lines are the iso-objective contours. Since the objective function has the
maximum at point D, it is an optimum point. As shown in the graphical method,
the optimum is on a boundary of the feasible set. Therefore, the optimum can be
obtained from comparison of the vertex points. Now Equation 3.72 is transformed
to the standard problem. The objective function is to be minimized by multiplying
–1 and a slack variable is added to each inequality constraint.
The problem is transformed as follows:
subject to b1 b2 b3 16 (3.73b)
1 1
b1 b2 b4 1 (3.73c)
28 14
1 1
b1 b2 b5 1 (3.73d)
14 24
bi t 0, i 1, ,5 (3.73e)
Position in
No. b1 b2 b3 b4 b5 f
Figure 3.1
1 0 0 16 1 1 0 A
2 0 14 2 0 5/12 í8400 E
3 0 16 0 í1/7 1/3 í F (infeasible)
4 0 24 í8 í5/7 0 í G (infeasible)
5 16 0 0 3/7 í1/7 í J (infeasible)
6 14 0 2 1/2 0 í5600 B
7 28 0 í12 0 í1 í H (infeasible)
8 4 12 0 0 3/14 í8800 D
9 11.2 4.8 0 9/35 0 í7360 C
10 140/17 168/17 í36/17 0 0 í I (infeasible)
Design Optimization 131
where b3 ,b4 and b5 in Equations 3.73bí3.73d are slack variables. The equations
in Equations 3.73bí3.73d are linearly independent. Thus, there is an infinite
number of solutions for those equations. Let us define parameter p as
p n m (n ! m). Then p 2 in this problem. At the vertices of the feasible set,
p variables have zero and they are called basic solutions. If we set p variables as
zero in Equations 3.73bí3.73d, a unique solution exists. We have ten basic
solutions and they are analyzed in Table 3.1. From the basic solutions, the
optimum solution is the one with minimum objective while the constraints are
satisfied.
3.5.3 Terminologies of LP
subject to b1 b2 d 4 (3.74b)
2b1 b2 d 2 (3.74c)
b1 t 0, b2 t 0 (3.74d)
Solution
First, the given problem is converted to a standard form as follows:
subject to b1 b2 b3 4 (3.75b)
2b1 b2 b4 2 (3.75c)
bi t 0, i 1, , 4 (3.75d)
b1 b2 b3 b4
1 1 1 0 = 4
2 í1 0 1 = 2
í3 1 0 0 = f í3
Table 3.3. The simplex tableau after one iteration in Example 3.9
b1 b2 b3 b4
0 3/2 1 í1 = 3
1 í1/2 0 1/2 = 1
0 í1/2 0 3/2 = f
Design Optimization 133
Table 3.4. The simplex tableau after two iterations in Example 3.9
b1 b2 b3 b4
0 1 2/3 í1/3 = 2
1 0 1/3 1/3 = 2
0 0 1/3 4/3 = f+1
The third row of the tableau is for the objective function. The least negative
coefficient is found from the third row. It is í3 in Table 3.2. The negative value
means that the objective function can be reduced. Thus, a new basic feasible
solution having nonzero b1 is selected. Since we can have two nonzero variables,
one of the nonzero variables should be changed to zero. If we investigate the
column with í3 in the third row, there are positive numbers 1 and 2 in the other
rows. We evaluate the ratios between these numbers and the numbers in the right
most column and compare them. In this case, 4/1 > 2/2. The nonzero variable in
the row with the smaller rate is changed to zero. That is, b4 is changed to zero
and b1 is changed to nonzero.
A simplex tableau is again made with the new basic feasible solution. A new
tableau with the canonical form is shown in Table 3.3. The pivoting process is
already applied to Table 3.3. The solution from the tableau is b1 1, b3 3,
b2 0, b4 0 and f 0. Since there is a negative value in the third row, the
objective function can be further reduced. The previous process is applied to
Table 3.3. If we investigate the column with í1/2 in the third row, the positive
coefficient is 3/2. Therefore, b3 is changed to zero and b2 is changed to nonzero.
The simplex tableau after two iterations is shown in Table 3.4. The basic
feasible solution is b1 2, b2 2, b3 0 and b4 0 , and the objective function
is í1. Since there is no negative value in the third row of Table 3.4, the objective
function cannot be reduced and the solution is the optimum solution.
f (b ( k 1) ) f (b ( k ) Dd ( k ) ) f (D ) (3.76)
134 Analytic Methods for Design Practice
The stepsize D , which makes the function f minimum, is the best value. In
df (D )
mathematics, D should satisfy the condition 0. Generally, the process
dD
for the stepsize calculation is in the middle of an optimization algorithm. Thus, D
df (D )
is calculated in an iterative manner instead of calculating 0. A few
dD
algorithms are introduced below.
f (D ) f (D )
rG
G
D D
G 2G qG D * (q 1)G D Dl D* Du
(a) Initial bracketing of the minimum (b) Reduction of the interval of uncertainty
The equal interval search has weak points in that: (1) It is difficult to estimate the
size of G appropriately and (2) the number of calculations of f (D ) can be
extremely large. Function calculation is quite costly in engineering, so it should be
reduced. Various algorithms have been developed to overcome these difficulties.
The golden section search method is introduced here. In the golden section search
method, G is increased by multiplying the golden section ratio 1.618 as follows:
q
j
Dq ¦ G (1.618) , q 0, 1, 2, (3.77)
j 0
F0 1, F1 1, Fn Fn 1 Fn 2 , n 2, 3, (3.78)
Fn
lim 1.618 (3.79)
nof Fn 1
ƒ(Į)
q =0 1 2 3 ···
Į
0 į 2.618į 5.236į Į* 9.472į
Figure 3.10. Initial bracketing of the minimum point in the golden section method
136 Analytic Methods for Design Practice
IJI (1íIJ)I Įu
(a) Įl Įa
Įb
(1íIJ)I IJI
I’
Į'b
Į'l Į'u
(b)
IJI’ (1íIJ)I’
q q2
j j
I Du Dl D q D q2 ¦ G (1.618) ¦ G (1.618) (3.80)
j 0 j 0
f ( x) x2 5x 4 (3.81)
G = 1.
Solution
x0 1
x1 1 1.618 2.618
Iteration 1:
(1) f ( x0 ) 0, f ( x1 ) 2.2361, f ( x2 ) 5.2351
0
I 4.2359, xu 5.2359, xl 1
(2) xa 2.618
xb 1 0.618, I 0 3.6178
(3) f ( x a ) 2.2361, f ( x b ) 1.0005
f ( xa ) f ( xb ), thus, the range of the minimum point x * is 1 d x*
d 3.6178.
(4) xl' xl 1, xu' xb 3.6178, xb' xa 2.618
xa' xl' 0.382( xu' xl' ) 2
(5) I xu' xl' 2.6178
138 Analytic Methods for Design Practice
Iteration 2:
step is to find the direction vector d in the steepest descent direction. The descent
condition is explained in Equations 3.56 and 3.57. At the kth iteration, the change
T
of the objective function in the direction of d ( k ) is D k (c ( k ) d ( k ) ) in the linear
expansion of the Taylor series. Because stepsize D k is a positive number, the
vector d ( k ) to reduce the objective function the most is c ( k ) , as seen in the linear
expansion. The vector c( k ) is the gradient vector and d ( k ) f (b ( k ) ) yields the
steepest descent vector.
The steps of the method are as follows:
T
c ( k ) c ( k 1) 0 (3.83)
T
df (b ( k 1) ) § wf (b ( k 1) ) · db ( k 1)
¨ ¸
¨ ¸ dD 0 (3.84)
dD k © w b ¹ k
140 Analytic Methods for Design Practice
wf (b ( k 1) )
c ( k 1) (3.85)
wb
and
db ( k 1) d
(b ( k ) D k d ( k ) ) d(k ) c ( k ) (3.86)
dD k dD k
T T
c ( k 1) d ( k ) c ( k 1) c ( k ) 0 (3.87)
Solution
Iteration 1:
(1) b ( 0) [0 0]T
(2) f [2b1 4b2 10b2 4b1 4]T [0 4]T
(3) direction vector d f [0 4]T
(4) b (1) b ( 0 ) Dd [0 0]T D [0 4]T [0 4D ]T
f (b ( 0) Dd) 80D 2 16D 4
D* 0.1
Design Optimization 141
b2
b1
Figure 3.12. Orthogonal steepest descent paths when the difference of partial derivatives is
large
Iteration 2:
The conjugate gradient method is modified from the steepest descent method and
shows good performance. It was developed by Fletcher and Reeves and the
convergence rate was improved considerably through modification of the
orthogonal condition in Equation 3.83. It is exactly the same as the steepest
descent method at the initial iteration (k 0) . From the second iteration (k 1),
all the processes are the same except for the evaluation of direction vector d ( k ) in
Step 3. The direction vector is
d (k ) c ( k ) E k d ( k 1) (3.88)
If matrix H is positive definite, Equation 3.90 is convex and 'b for the minimum
of Equation 3.90 is unique. The condition is
df
c H'b 0 (3.91)
d('b)
All the processes of the algorithm are the same as the steepest descent method,
except for the evaluation of the direction vector in Step 3.
Newton’s method is a theoretically excellent method in that the order of
convergence is two. However, the method needs the Hessian matrix, which is very
expensive to evaluate. Moreover, when the Hessian matrix is not positive definite,
convergence is not guaranteed. If a design iteration comes to a point where the
Hessian is indefinite in the iterative process, the next design can be extremely
deteriorated. Therefore, it is not utilized in practice, although provides ideas for
quasi-Newton methods.
Quasi-Newton methods use an approximation of the Hessian matrix H or its
inverse H 1. They do not require the condition of positive definiteness of the
Hessian matrix or calculation of the second derivatives. It is known that the
convergence rate of the methods is inferior to Newton’s method and superior to the
conjugate gradient method. This is evidenced from numerical research and not
from rigorous mathematics. The Hessian matrix is assumed to be a unit matrix at
the initial iteration and is updated as the iteration proceeds. Two methods exist
and their characteristics are as follows:
method. The algorithm is the same as in the steepest descent method, except for
two aspects.
(1) In Step 3 of the steepest descent method, the direction vector is evaluated
as follows:
d (k ) A (k ) c (k ) (3.93)
A ( k 1) A (k ) B (k ) C (k ) (3.94)
T
(k ) s (k ) s (k )
B T
(3.95a)
s (k ) y (k )
T
z (k ) z (k )
C (k ) T
(3.95b)
y (k ) z (k )
z (k ) A (k ) y (k ) (3.95f)
Fletcher and Powell proved that matrix A (k ) is always positive definite and it
converges to the inverse Hessian when the objective function is a positive definite
quadratic function.
H (k ) d (k ) c ( k ) (3.96)
H ( k 1) H (k ) D (k ) E (k ) (3.97)
T
y (k ) y (k )
D (k ) T
(3.98a)
y (k ) s (k )
T
c (k ) c (k )
E (k ) T
(3.98b)
c (k ) d (k )
Find Gb ( k ) (3.99a)
g j (b ( k ) ) T g j (b ( k ) )Gb ( k ) d 0, j 1, , m (3.99d)
Gb (Lk ) d Gb ( k ) d Gb U( k ) (3.99e)
(1) Vector Gb ( k ) is the design variable vector of the subproblem and is the
design change from the current design point.
(2) Equation 3.99b is the linear term of the Taylor expansion and the change of
the objective function when the design variable vector b ( k ) is changed by
Gb ( k ) . It is minimized to reduce the objective function.
(3) Equation 3.99c requires the equality constraints to be satisfied. Suppose
the ith constraint is satisfied at the current point by hi (b ( k ) ) 0. Then
Equation 3.99c changes the constraint to satisfy T hi (b ( k ) )Gb ( k ) 0.
When the ith constraint is violated by hi (b ( k ) ) z 0, Equation 3.99c tries to
satisfy the ith constraint at the next iteration.
(4) An inequality constraint can be corrected by Equation 3.99d. In any cases,
Equation 3.99d tries to satisfy the jth inequality constraint by
g j (b ( k ) ) T g j (b ( k ) )Gb ( k ) d 0 at the next iteration.
(5) Equation 3.99e restricts the size of Gb ( k ) .
The subproblem is a linear programming problem, thus the optimum solution
of Gb ( k ) can be obtained by a linear programming method. In general, algorithms
of the primal method proceed as follows:
b1 3 t 0 (3.100c)
1
t0 (3.100d)
b2 1
Solution
g 2 [1 0]T , g 2 5
T T
ª 1 º 1 ª 1º
g 3 «0 » , g3 , g 3 «0
«¬ b2 1 »¼
2 2 ¬ 4 »¼
If we substitute the above information into Equation 3.99, then the following
linearized subproblem is obtained:
1 ª 1º
0 [Gb1 Gb2 ]T t 0 (3.101e)
2 «¬ 4 »¼
In engineering, the number of inequality constraints m in Equation 3.99d is
often extremely large. Then the size of the subproblem becomes very large.
Frequently some inequality constraints satisfied in an iteration are satisfied again
in the next iteration. Then those constraints have no roles in the iterative process.
In other words, those constraints do not need to be considered in the subproblem.
Therefore, if a constraint is sufficiently satisfied at the kth iteration, the
subproblem of the kth iteration does not include the constraint. Then the size of
the subproblem is reduced and the cost for the sensitivity analysis for the constraint
is saved. This is called the active set strategy.
When a satisfied constraint is close to zero, satisfaction and violation can
alternate between feasibility and infeasibility. The İ-active set strategy is utilized
148 Analytic Methods for Design Practice
to prevent it. Although the constraints are satisfied, the ones with the following
condition are included in the active set:
A feasible region is presented in Figure 3.13. If we use the İ-active set strategy,
the feasible region is reduced. With the active set strategy (not the İ-active set
strategy), when the design points such as B and C in Figure 3.13 are close to the
boundary, the constraint g j (b ( k ) ) is not included in the subproblem and the next
design can have a large probability of moving into the infeasible region. However,
if we use the İ-active set strategy, the feasible region is reduced, because the inside
region of the dotted lines yields the feasible region.
When the active set strategy is employed and the design point is at B or C, the
constraint g j (b) is excluded in the active constraint set and the design may move
to an infeasible point such as D. However, when the İ-active set strategy is
utilized, the constraints are active and the design point remains in the feasible
region. Because the design point frequently exists around the boundary of a
constraint, the İ-active set strategy works quite well. When the number of
constraints is reduced, the subproblem size is reduced and sensitivity analysis of
the excluded constraints is not needed. Of course the equality constraints are
included in the active set all of the time.
In the İ-active set strategy, active constraint set I k is defined as
D•
•
C
B•
İ
g j (b) 0
•A
Feasible region g j (b) H 0
b1 b22
g1 1 d 0 (3.104)
3 3
g2 b1 1 d 0 (3.105)
b1b2
g3 1 d 0 (3.106)
4
3
g4 1 d 0 (3.107)
2b1 2
3
g5 1 d 0 (3.108)
2b2
Solution
Primal methods have been developed based on the steps in Section 3.8.1. Each
algorithm includes slight modifications of the baseline steps. In this section,
representative algorithms are introduced by virtue of popularity and applicability.
whereas the move limit on the direction vector Gb ( k ) in Equation 3.99e is defined
as
|| Gb ( k ) || d [ 2 (3.115)
n
0.5 ¦ (Gbi( k ) ) 2 d [ 2 (3.116)
i 1
Find Gb ( k ) (3.118a)
g j (b ( k ) ) T g j (b ( k ) )Gb ( k ) d 0, j 1, , m (3.118d)
T
0.5Gb ( k ) WGb ( k ) d [ 2 (3.118e)
Since Equation 3.118e is nonlinear, the algorithm for linear programming cannot
be used. The KKT condition of Equation 3.118 is equivalent to that of Equation
3.119:
Find Gb ( k ) (3.119a)
T
to minimize T f( b ( k ) )Gb ( k ) 0.5Gb ( k ) WGb (3.119b)
g j (b ( k ) ) T g j (b ( k ) )Gb ( k ) d 0, j 1, , m (3.119d)
l
I f (b( k ) D k d(k ) ) ¦ vi | hi (b(k ) D k d( k ) ) |
i 1
(3.120)
m
¦ u j (max[0, g j (b(k ) D k d( k ) )])
j 1
the original SQP methods. It uses a specific active set strategy with W I and a
specific method for determination of D k (Arora 2004).
There is an SQP method that exploits the idea of Newton’s method for
unconstrained problems. The matrix W is replaced by the Hessian matrix of the
Lagrangian function in Equation 3.46 as
W 2L (3.121)
s (k ) D k d (k ) (3.122a)
z (k ) W (k ) s (k ) (3.122b)
0.8[ 2
if [1 t 0.2[ 2 T 1 , otherwise T (3.122f)
[ 2 [1
p( k ) șy ( k ) (1 T )z ( k ) (3.122g)
T
[3 s (k ) p (k ) (3.122h)
1 T
D (k ) p (k ) p (k ) (3.122i)
[3
1 T
E (k ) z (k ) z (k ) (3.122j)
[2
W ( k 1) W (k ) D(k ) E (k ) (3.122k)
It is well known that many excellent methods have already been developed for
unconstrained problems. If constrained problems can be transformed to
unconstrained problems, the existing unconstrained optimization methods can be
exploited. Numerous research investigations have been carried out to explore this
idea. Generally, the constraint functions are transformed to a penalty function and
added to the objective function. The transformed functions are usually expressed
as follows:
where < is the transformed objective function, P is the penalty function and r is a
vector for the penalty parameters. P is a real function and the magnitude of the
penalty is controlled by r. Based on the definition of <, an unconstrained
optimization method finds an optimum for < with initial values of b (0) and r.
During the optimization process, b is the design variable vector and r is considered
as a constant vector. The optimum values yield the initial design of the next
iteration and the new iteration finds an optimum of < with an adjusted r. The
iterative process terminates if no improvement is observed.
The sequential unconstrained minimization technique (SUMT) is the
representative method for the transformation methods. It is generally called
SUMT. The method is varied by virtue of the definitions of the penalty function
and the penalty parameter r. One well known penalty function is the quadratic loss
function below
l m
P(h(b), g(b), r ) r{ ¦ [ hi (b)] 2 ¦ [ max (0, g j (b))] 2 } (3.124)
i 1 j 1
m
P (g(b), r ) (1 / r ) ¦ [1 / g j (b)] : Inverse barrier function (3.125)
j 1
m
P (g(b), r ) (1 / r ) ¦ log( g j (b)) : Log barrier function (3.126)
j 1
Because these functions set a large barrier around the feasible region, a feasible
solution cannot jump to the infeasible region. When a constraint is active, P is
theoretically infinite. Also, the algorithm needs the condition that the initial design
154 Analytic Methods for Design Practice
l m
P (h(b), g(b), r ) 0.5 ¦ ri' (hi T i' ) 2 0.5 ¦ r j [( g j T j ) ] 2 (3.127)
i 1 j 1
where ( g j T j ) max(0, g j T j ) . T i' and ri' are parameters related to the ith
equality constraint. T j ! 0 and r j ! 0 are related to the jth inequality constraint.
If T j T i' 0 and ri ri' r , Equation 3.127 is the same as the quadratic loss
function as in Equation 3.124. The parameter r goes to infinity in the quadratic
loss function, whereas it has finite values here. The optimum of Equation 3.123
should be found by substituting Equation 3.127 into Equation 3.123 in an iterative
manner. The parameters are updated using various rules.
3.9 Exercises
3.2 A cantilever beam has a vertical load P at the tip and the length is L. The
vertical displacement at the tip (d) should be less than 10 mm and the
maximum stress at the clamped end should be less than 100 MPa.
Formulate the optimization problem to minimize the weight of the beam.
Determine the cross section. (Young’s modulus is E, the moment of inertia
is I, the density is U , and the width and the height of the cross section are
b and h, respectively.)
Design Optimization 155
s wG
h
d b
k1
l1
Qx P2
P2 Qy P1
Q
Q
P1
l2 k2
p(t)
r(t)
(1) The difference between the momentum by p(t) and that of r(t) should
be minimized.
(2) The differences between p(t1) and r(t1) and p(t2) and r(t2) should be
less than 0.5.
(3) p(t) is a polynomial equation of the 6th degree.
Formulate an optimization problem to obtain p(t) while the given
conditions are satisfied.
156 Analytic Methods for Design Practice
subject to g 1 (b) b1 1 d 0
(2)
Minimize f (b) b1 b2
g 3 (b) b1 t 0
3.7 Find a local optimum point of the following problems by using the KKT
necessary condition:
(1)
Minimize f (b) 10b12 2b22 5b1b2 3b1
subject to g 1 (b) b1 b2 4
(2)
Minimize f (b) 3b12 16b1 16b22 7
(3)
Minimize f (b) 2b1 3b 2 b12 b22
b1 , b2 t 0
f (b) 3b 2 6b 7 , b (0) 2
Repeat the five steps of Section 3.4 twice. The direction vector
d f and the stepsize D 0.2 .
3.9 Calculate the sensitivity of the following function by the finite difference
method with the following finite differences. Compare the values with
those from the analytic method. The current design point is b ( 0) 1 .
subject to b1 b2 2b3 t 2
b1 , b2 t 0
(2)
Minimize f (b) b1 3b2
b1 4b2 d 5
b1 2b2 t 6
b1 , b2 t 0
(3)
Maximize f (b) 3b1 4b2 5b3 5b4
b1 , b2 , b3 , b4 t 0
(4)
Minimize f (b) 10b1 b2
4b1 b2 t 10
b1 d 0
3.11 If the following problems have solutions, obtain them by the graphical
method. Otherwise, explain why there are no solutions.
(1)
Minimize f (b) b1 b2
5b1 7b2 1
b1 3b2 7
b1 , b2 t 0
(2)
Minimize f (b) b1 b2
b1 , b2 t 0
(3)
Minimize f (b) b1 b2
b1 3b2 7
b1 5b2 d 10
3.12 Find the basic solutions to the following problems and show them by the
graphical method:
Design Optimization 159
(1)
Maximize f (b) b1 2b2
subject to 2b1 b2 t 4
b1 2b2 d 3
2b1 3b2 5
b1 , b2 t 0
(2)
Maximize f (b) 3b1 2b2
subject to 7b1 b2 t 14
b1 4b2 t 5
b1 t 0
(3)
Minimize f (b) 9b1 8b2 2b3
b1 2b2 3b3 t 7
b1 , b2 , b3 t 0
3.13 Find the minimum point of f (b) by the equal interval search. The interval
of uncertainty for convergence is 0.001 and G 0.02 . b 1 is the initial
design. Use programming languages such as C or FORTRAN if needed.
f (b) 4b 2 5b 9
3.16 Solve the problems in Exercise 3.15 by using the golden section method.
The solution process can be programmed. G 0.02 and the interval of
uncertainty for convergence is 0.001.
3.17 Repeat two iterations of the steepest descent method to minimize the
following functions:
3.19 Solve Exercise 3.17(3) by using the steepest descent method and the
conjugate gradient method. Continue the iteration until the interval of
uncertainty is less than 0.001. Compare the histories of the two processes.
3.22 Normalize the following constraints and obtain the İ-active constraints.
Use H 0.1 and the current design point is b [3.3 3.3]T .
g 1 (b) b1 3.3 d 0
g 2 (b) b2 3.5 d 0
g 4 (b) b1b2 12 d 0
g 5 (b) b2 d 0
g 6 (b) b1 b22 10 d 0
g 7 (b) b2 b1 2 d 0
Design Optimization 161
b [3 1]T
(2)
Minimize f (b) b12 2b22 4b1 b2 1
3b1 7b2 d 10
b1 2b2 t 0
b [1 1]T
(3)
Minimize f (b) (b1 1) 2 (b2 1) 2
subject to b1 b22 2 d 0
b1 2b2 d 0
b [4 5]T
3.24 Make subproblems for Exercise 3.23 to use the sequential linear
programming (SLP) algorithm. The move limit is r 0.5.
subject to b1 b22 d 6
b1 , b2 d 0
b ( 0) [0 1]T
162 Analytic Methods for Design Practice
(2)
Minimize f (b) b12 2b22 4b1b2 3b2 1
b1 2b22 d 0
b (0) [1 3]T
3.26 Define the inverse barrier function to solve the following problem by the
transformation method. Draw the contour of the inverse barrier function
according to the parameter r. Explain the phenomena around the
boundaries of the design variable.
Minimize f (b) b 2 2b 5
subject to 3dbd7
Design Optimization 163
A design solution must be in the feasible region and thus the optimum solution
should be found in the feasible region. That is, the optimum solution makes the
objective function have the minimum within the feasible region. Suppose we have
the following constraint set:
Figure 3.A.1 illustrates the above constraint functions. The region satisfying each
constraint is the opposite side of the oblique lines. Therefore, the feasible region is
the shadowed area. In Figure 3.A.1, the design points A, B and C are in the
feasible region and the points D, E and F are outside the feasible region.
Global (absolute) minimum: At a certain design point b* , the objective function
has the global (absolute) minimum if the following condition is satisfied for any
design point b in the feasible region:
b2
2 F
1 E
g 2 { b12 b22 4 0
C
1 2 b1
B A
g1 { b1 b2 1 0
where, if the inequality (<) is always satisfied, the minimum is called the strict
global minimum.
N {b | b S , | b b * | G } (3.A.4)
T
ª wf wf wf º
f « » (3.A.5)
¬ wb1 wb2 wbn ¼
f(b) C
A
B
D
F
Feasible region
T
ª wf wf º
f « » [2b1 2b2 ]T (3.A.7)
¬ wb1 wb2 ¼
ª w2 f w2 f w2 f º
« 2
»
« wb1 wb1 wb2 wb1 wbn »
« w f
2
w2 f w2 f »
w2 f « »
H « wb2 wb1 wb22 wb2 wbn » (3.A.8)
wbwb
« »
« 2 2 2 »
« w f w f
w f »
« wbn wb1 wbn wb2 wbn2 »¼
¬
b2
f [ 2 2] T
1
b1
g2 b12 b22 2 0
1
f (b) f (b * ) T f (b * )(b b * ) (b b * ) T H(b b * ) R (3.A.9)
2
x2 x2
1
x1
x
x1 x2
x1 x2
Regular point: Assume that we have only equality constraints in the optimization
problem in Equation 3.1. If all the equality constraints are satisfied with
hi (b * ) 0, i 1, ..., l at point b * and the gradient vectors of the equality
constraints are linearly independent, then the point b * is said to be a regular point.
When inequality constraints exist, the active constraints are considered as equality
constraints and the gradient vectors of active constraints should be linearly
independent of each other at a regular point.
References
Arora JS (2004) Introduction to Optimum Design. 2nd ed., Elsevier, San Diego,
CA
Bazaraa MS, Sherali HD, Shetty CM (1993) Nonlinear Programming. Wiley, New
York
Belegundu AD, Chandrupatla TR (1999) Optimization Concepts and Applications
in Engineering. Prentice Hall, New Jersey
Dantzig G (1963) Linear Programming and Extensions. Princeton University Press,
Princeton, New Jersey
Fiacco AV, McCormic GP (1968) Nonlinear Programming: Sequential
Unconstrained Minimization Techniques. Wiley, New York
Fletcher R, Reeves CM (1964) Function Minimization by Conjugate Gradients.
Computer Journal 7:149í154
Fletcher R (1970) A New Approach to Variable Metric Algorithms. Computer
Journal 13:317í322
Goldberg DE, Samtani MP (1986) Engineering Optimization via Genetic
Algorithm. Proceedings of the 9th Conference on Electronic Computation,
ASCE, pp 471í482
Haftka RT, GĂrdal Z (1992) Elements of Structural Optimization. Kluwer,
Dordrecht
Haug EJ, Arora JS (1979) Applied Optimal Design. Wiley, New York
Papalambros PY, Wilde DJ (2000) Principles of Optimal Design: Modeling and
Computation. 2nd ed., Cambridge University Press, Cambridge
Parmee IC, Hajela P (eds) (2002) Optimization in Industry. Springer, Berlin
Heidelberg New York
Polak E (1971) Computational Methods in Optimization. Academic Press, New
York
Powell MJD (1964) An Efficient Method for Finding the Minimum of a Function
of Several Variables without Calculating Derivatives. Computer Journal
7:303í307
Rao SS (1996) Engineering Optimization. 3rd ed., Wiley, New York
Rosen JB (1960) The Gradient Projection Method for Nonlinear Programming,
Part I, Linear Constraints. SIAM Journal Appl Math 8:181í217
Rozvany GIN (1997) Topology Optimization in Structural Mechanics. Springer,
Berlin Heidelberg New York
Design Optimization 169
4.1 Introduction
In structural optimization, optimization theory is applied to the design of structures.
In Chapter 3, it was emphasized that a design problem should be formulated as in
Equation 3.1 in order to apply optimization theory to practical design. It may not
be easy to precisely define a design problem as in Equation 3.1. Especially, the
equality constraints in Equation 3.1c are difficult to define and manipulate.
Nevertheless, the optimization method is frequently utilized in structural design
because the formulation in Equation 3.1 can incorporate the objective and the
constraint conditions of design.
Structural optimization has been employed to reduce the weight of airplane
structures. Aerospace engineering is the field where structural optimization is
applied most extensively. Recently, application has expanded to other fields.
There are several reasons why optimization is actively applied to structural design.
First, the design process for structures is quite similar to the optimization
process. Reduction of weight is important in structural design and it is equivalent
to the objective function of optimization in that the reduction is similar to
minimization of the objective function. When maximization or minimization of a
specific response is pursued, it can be treated as an objective function in
optimization. Also, various design conditions for structural design can be easily
transformed into constraints in optimization.
Second, the finite element method (FEM) can be exploited in structural
optimization. In optimization, the objective function and constraints should be
defined in definite forms and they must be evaluated precisely. FEM is an analysis
method that facilitates this aspect and it is one of the computational methods that
are well developed for analysis of a system. Since many commercial systems are
available, it is easy to interface an optimization system with the systems of FEM.
The optimization theories in Chapter 3 are directly used in structural
optimization. The process to find optimum sizes of structures is explained by the
primal method in Section 3.8. With the current sizes, we evaluate the total weight
of a structure (objective function) and the stresses (constraints). A linearized
subproblem is defined based on the evaluation. A direction vector is calculated
172 Analytic Methods for Design Practice
Vy
W yz
W yx
W zy W xy
W xz
Vx
Vz W zx
z x
Stress is the reaction force per unit area. Each stress component in Figure 4.1
yields the following stress vector:
ı [V x V y V z W xy W yz W zx ]T (4.1)
At a point in the space, the stresses in Equation 4.1 should satisfy the equilibrium
equation in Equation 4.2
wV x wW xy wW xz
Fx 0 (4.2a)
wx wy wz
wV y wW xy wW yz
Fy 0 (4.2b)
wy wx wz
wV z wW xz wW yz
Fz 0 (4.2c)
wz wx wy
where Fx , F y , and Fz are components of the body force such as the gravitational
force.
Suppose u [u v w]T is a displacement vector at a point. Then the strains
are
174 Analytic Methods for Design Practice
wu
Hx (4.3a)
wx
wv
Hy (4.3b)
wy
ww
Hz (4.3c)
wz
wu wv
J xy J yx (4.3d)
wy wx
wv ww
J yz J zy (4.3e)
wz wy
ww wu
J zx J xz (4.3f)
wx wz
In vector form
İ [H x H y H z J xy J yz J zx ]T (4.4)
ı Eİ (4.5)
where E is a matrix defined by Young’s modulus and Poisson’s ratio.
The stiffness matrix is made in Step 2. There are a few methods to make a
stiffness matrix as follows:
(1) The matrix method: A simultaneous equation is defined from the
equilibrium equation.
(2) The principle of minimum potential energy: The principle that the
condition of minimum potential energy satisfies the equilibrium equation is
utilized.
(3) The principle of virtual work: The principle of virtual work is utilized. It
indicates that the sum of the work by infinitesimal virtual displacements
satisfying the boundary conditions and the internal strain energy is zero.
Table 4.1 shows the processes for making the element stiffness matrix by the
above methods. The matrix method can only make stiffness matrices for simple
elements such as truss, beam and triangular elements. The principle of minimum
potential energy or the principle of virtual work should be utilized to attain the
stiffness matrices for modern complicated elements. It is noted that the stiffness
matrices for truss, beam and triangular elements have explicit expressions.
However, the stiffness matrix is generally expressed by functionals as follows:
T
kL ³ B EB dV (4.6)
V
where V is the domain of each element. As shown in Table 4.1, the equilibrium
equation of each element is
k LzL fL (4.7)
176 Analytic Methods for Design Practice
Table 4.1. Formulation process of the finite element method (Choi 2002) (continued)
5. From the relationship 6. The functional of the total potential 6. The strain energy by the virtual strain is evaluated as
between ı and the nodal energy is obtained from the sum of the U ³ ı T İ dV ³ İ T Eİ dV
displacement vector z L , deformation energy and the potential of V V
wz L
T
z L k L z L z TL f L 0
dz TL k L z L dz TL f L 0
If z TL is eliminated
If dz TL is eliminated
k LzL fL
k Lz L fL
Structural Optimization 179
where z L R l ' is the nodal displacement vector at each element and l ' is the
degrees of freedom for each element. The vector f L is a vector of the external
loads imposed on the nodes.
The equilibrium equation of the entire structure is obtained by assembling the
element equilibrium equation in Equation 4.7. Equation 4.7 is defined in the
global coordinate. If the boundary condition is applied to the assembled equation,
the equation yields
Kz f (4.8)
where z R l is a displacement vector for all the degrees of freedom and l is the
total degrees of freedom. K is an (l u l ) matrix and f is an (l u 1) column vector
and they are the stiffness matrix and the external load vector for the entire structure,
respectively.
The governing equation for vibration of structures can be formulated in the
finite element method as follows:
Ky [My 0 (4.9)
Solution
The stiffness matrix k iL of the ith element in the local coordinate system is
bi E ª 1 1º
k iL « 1 1 » (4.10)
Li ¬ ¼
180 Analytic Methods for Design Practice
where Li is the length of the ith member. Equation 4.10 should be transformed to
the stiffness matrix in the global coordinate system. For the transformation, the
following transformation matrix T is utilized:
ªcos D sin D 0 0 º
T « 0 (4.11)
¬ 0 cos D sin D »¼
where D is the angle between a truss element and the x-axis of the global
coordinate system.
Using the transformation matrix T, the matrix in Equation 4.10 is transformed
as follows:
ki TT k iL T (4.12)
where k i is the stiffness matrix in the global coordinate system for the ith element.
Using Equation 4.12, the product of the stiffness matrix and the displacement
vector in the global coordinate system for each member is as follows:
ª § 3S · º
«cos ¨ 4 ¸ 0 »
« © ¹ »
« sin § 3S · »
¨ ¸ 0
b1 E « © 4 ¹ » ª 1 1º
k 1z 1 « »
L1 « § 3S ·» «¬ 1 1 »¼
0 cos ¨ ¸
« © 4 ¹»
« § 3S · »
« 0 sin ¨ ¸»
¬« © 4 ¹ ¼»
ª § 3S · § 3S · º ª u1 º
«cos ¨ 4 ¸ sin ¨ 4 ¸ 0 0 »« v »
« © ¹ © ¹ »« 1 »
« § 3S · § 3S ·» «u 2 »
« 0 0 cos ¨ ¸ sin ¨ ¸» « »
¬ © 4 ¹ © 4 ¹ ¼ ¬ v2 ¼
ª 1 1 1 1º ª u1 º
« »« »
b1 E « 1 1 1 1» « v1 »
(4.13)
2 L1 « 1 1 1 1 » «u 2 »
« »« »
¬ 1 1 1 1 ¼ ¬ v2 ¼
Structural Optimization 181
ª § S· º
«cos ¨ 2 ¸ 0 »
« © ¹ »
« sin § S · 0 »
¨ ¸
b2 E « © 2 ¹ » ª 1 1º
k 2z 2 « »
L2 « § S · « 1 1 »¼
0 cos ¨ ¸» ¬
« © 2 ¹»
« § S ·»
« 0 sin ¨ ¸ »
¬« © 2 ¹ »¼
ª § S· § S· º ª u1 º
«cos ¨ 2 ¸ sin ¨ 2 ¸ 0 0 »«v »
« © ¹ © ¹ »« 1 »
« § S· § S ·» «u3 »
« 0 0 cos ¨ ¸ sin ¨ ¸» « »
¬ © 2¹ © 2 ¹¼ ¬ v3 ¼
ª0 0 0 0 º ª u1 º
« »« »
b2 E «0 1 0 1» « v1 »
(4.14)
L 2 «0 »
0 0 0 » «u 3 »
« « »
«¬0 1 0 1 »¼ ¬ v3 ¼
ª § S· º
«cos ¨ 4 ¸ 0 »
« © ¹ »
« sin § S · 0 »
¨ ¸
b3 E « © 4 ¹ » ª 1 1º
k 3z 3 « »
L3 « § S · « 1 1 »¼
0 cos ¨ ¸» ¬
« © 4 ¹»
« § S ·»
« 0 sin ¨ ¸ »
«¬ © 4 ¹ »¼
ª § S· § S· º ª u1 º
«cos ¨ 4 ¸ sin ¨ 4 ¸ 0 0 »« v »
« © ¹ © ¹ »« 1 »
« § S · § S · «u »
« 0 0 cos ¨ ¸ sin ¨ ¸»» « 4 »
¬ © 4¹ © 4 ¹ ¼ ¬ v4 ¼
ª 1 1 1 1 º ª u1 º
« »« »
b3 E « 1 1 1 1» « v1 »
(4.15)
2 L3 « 1 1 1 1» «u 4 »
« »« »
¬ 1 1 1 1 ¼ ¬v4 ¼
182 Analytic Methods for Design Practice
where z i is the displacement vector of the ith element in the global coordinate
system. The product of the global stiffness matrix K and the displacement z is
obtained by assembling the element stiffness matrices as follows:
ª b1 b3 b1 b b1 b1 b3 b3 º ª u 1 º
« 2L 0 2L 0 3 00 00
« 1 3 2 L1 2 L3 2 L1 2 L1 2 L3 2 L3 »» « »
« »
« b1 0 b3 b1 b
2 3
b
1
b
1
b
00 0
b2 b3 b
3 »» « v1 »
« 2L 2 L3 2 L1 L 2 2 L3 2 L1 2 L1 L2 2 L3 2 L3 « »
« 1 »« »
« b b b1 b1
1
1 0 0 0 0 » «u 2 »
« 2 L1 2 L1 2 L1 2 L1 »« »
« b b b1 b1 »« »
« 1 1 0 0 0 0 »« »
Kz E« 2 L1 2 L1 2 L1 2 L1 » «v 2 »
« 00 00 0 0 0 0 0 0 » «u 3 »
« b b2 »« »
« 00 0 2 0 0 0 0 0 » « v3 »
« L2 L2 »« »
« b b3 b3 b3 » « »
« 3 0 0 0 0
2 L3 » « 4 »
u
2 L3 2 L3 2 L3
« »
« b3 b b3 b3 » «« »»
3 0 0 0 0
« 2 L3 2 L3 2 L3 2 L3 »¼ ¬« v 4 ¼»
¬
(4.16)
ª2 0 º ªu1 º
Kz 2 u 10 4 u 207 u 10 9 « »« » (4.17)
¬0 2 2 ¼ ¬ v1 ¼
ª 40000 u 1 / 2 º
f «40000 u 3 / 2» (4.18)
¬ ¼
Substituting Equations 4.17 and 4.18 into Equation 4.8, the following equilibrium
equation is obtained:
ª2 0 º ªu1 º ª 40000 u 1 / 2 º
2 u 10 4 u 207 u 10 9 « « » (4.19)
¬ 0 2 2 »¼ ¬ v1 ¼ «40000 u 3 / 2»
¬ ¼
ªu1 º ª3.4160 u 10 4 º
z «v » « 4 »
(4.20)
¬ 1¼ «¬3.4659 u 10 »¼
ı EBz L (4.21)
V p/b (4.22)
where b is the area. Equation 4.22 gives the same solution as Equation 4.21.
Equation 4.22 is utilized here for stress calculation.
Equation 4.23 is the equilibrium equation in the local coordinate for the truss
element in Figure 4.4
ª p1 º bE ª 1 1º ª z L1 º
«p » « »« » (4.23)
¬ 2¼ L ¬ 1 1 ¼ ¬ z L 2 ¼
bE
p2 p1 [ 1 1]z L (4.24)
L
E
V [ 1 1]z L (4.25)
L
zL Tz (4.26)
where the transformation matrix T represents the relationship between the two
coordinate systems as shown in Equation 4.11. By substituting Equation 4.26 into
Equation 4.25, the stress can be obtained by the displacement vector in the global
zL2 z L1
p2 p1
E
Vi [ 1 1]Tz i (4.27)
L
4
ª § 3S · § 3S · º ª3.4160 u 10 º
E «cos ¨ 4 ¸ sin ¨ 4 ¸ 0 0 » «3.4659 u 10 4 »
V1 [ 1 1]« © ¹ © ¹ »« »
L1 « § 3S · § 3S ·» « 0 »
« 0 0 cos ¨ ¸ sin ¨ ¸» « »
¬ © 4 ¹ © 4 ¹¼ ¬« 0 ¼»
71.227 MPa (4.28)
4
ª § S· § S· º ª3.4160 u 10 º
E
cos
« ¨ 2¸ sin ¨
2
¸ 0 0 «
» 3.4659 u 10 4 »
V2 [ 1 1]« © ¹ © ¹ »« »
L2 « § S· § S ·» « 0 »
« 0 0 cos ¨ ¸ sin ¨ ¸» « »
¬ © 2¹ © 2 ¹¼ ¬« 0 ¼»
71.744 MPa (4.29)
4
ª § S· § S· º ª3.4160 u 10 º
E « cos ¨ ¸
4¹
sin ¨ ¸ 0 0 «
» 3.4659 u 10 4 »
V3 [ 1 1]« © © 4¹ »« »
L3 « § S· § S ·» « 0 »
« 0 0 cos ¨ ¸ sin ¨ ¸» « »
¬ © 4 ¹ © 4 ¹¼ «¬ 0 »¼
516.5 kPa (4.30)
The aforementioned analysis details the process for obtaining the stresses for a
truss structure. Generally, the first process in the finite element method is the
calculation of the displacements. Other responses such as strains and stresses are
obtained from the displacement information.
Solution
The stiffness matrix of the four node element with thickness h is
kL h ³ B T EB dxdy (4.31)
Structural Optimization 185
100kN
4 3 100kN
y 2m
x
1 2
2m
ª º
«1 v 0 » ª1.06667 0.26667 0 º
E «
E v 1 0 » E ««0.26667 1.06667 0 »» (4.32)
1 v2 « 1 v» «¬ 0
«0 0 » 0 0.4»¼
¬ 2 ¼
Using the s–t natural coordinate system within an element, Equation 4.31 is
11
kL h ³ ³ BT ( s, t )EB( s, t ) | J ( s, t , x, y ) | dsdt (4.33)
11
where the natural coordinate system is an artificial coordinate system to define the
shape function in an isoparametric formulation. The matrix J is the Jacobian
matrix between the s–t coordinate and the x–y coordinate systems. Details are
explained in the references for the finite element method. Using four point
Gaussian quadrature, the integration of Equation 4.33 is approximated as
4
k L # h ¦ B T ( s i , t i )EB( s i , t i ) J ( s i , t i , x, y ) (4.34)
i 1
From Equation 4.34, it is noted that thickness h is the only property that can be a
design variable in size optimization. It is separated from the other terms.
Therefore, partial differentiation with respect to h is easy and the sensitivity
analysis is relatively simple.
186 Analytic Methods for Design Practice
The stiffness matrix in the global coordinate system can be obtained by using
Equation 4.12. Since the transformation matrix T is the unit matrix and we have
one element in this problem, the element stiffness matrix in Equation 4.35 is
identical to the assembled matrix in the global coordinate system. If we apply the
boundary condition, the final stiffness matrix is
Kz f (4.37)
z [u2 v2 u3 v3 ]T (4.38)
ı EBz (4.41)
Structural Optimization 187
If we substitute B(0,0) at the center, Equation 4.32 and Equation 4.40 into
Equation 4.41, the following stresses are obtained:
ª1.06667 0.26667 0 º
ı E ««0.26667 1.06667 0 »»
«¬ 0 0 0.4»¼
ª 0 º
« 0 »
ª 1 1 1 1 º« »
« 4 0 0 0 0 » « 4.0419 »
4 4 4 « »
« 1 1 1 1 » « 9.9649 »
« 0 0 0 0 » u 10 4 (4.42)
« 2 2 2 2 » « 4.0759 »
« 1 1 1 1 1 1 1 1 »« »
«¬ 2 » «15.0012»
4 2 4 2 4 2 4 ¼«
0 »
« »
«¬ 0 »¼
Find b R n , z R l , [ R1 (4.44a)
g j (b, z, [ ) d 0, j 1, , m (4.44e)
bL d b d bU (4.44f)
where Equation 4.44c is the equilibrium equation in the finite element method and
it is assumed to be linear. Equation 4.44d is the eigenvalue equation. In Equation
188 Analytic Methods for Design Practice
y T My 1 (4.45)
Find Gb ( k ) , Gz ( k ) , G[ ( k ) (4.46a)
subject to
w w
(Kz f ) (Kz f )Gb ( k ) (Kz f )Gz ( k ) 0 (4.46c)
wb wz
w w
(Ky [My ) (Ky [My )Gb ( k ) (Ky [My )G[ ( k ) 0 (4.46d)
wb w[
g j Tb g j Gb ( k ) Tz g j Gz ( k ) [T g j G[ ( k ) d 0, j 1, , m (4.46e)
constraints are eliminated by this process and the number of design variables is
reduced from (n + l + 1) to n.
We will investigate the elimination process of the equality constraints. At the
kth iteration, the solution z (k ) of the finite element method satisfies
Kz (k ) f 0 (4.47)
Since K and f are functions of the design variable vector b, the second term of
Equation 4.46c becomes
w § w (K~z ) wf · ( k )
(Kz f )Gb ( k ) ¨ ¸Gb (4.48)
wb © wb wb ¹
where ~z is a constant vector and the same as z (k ) obtained from Equation 4.47.
The third term of Equation 4.46c becomes
w
(Kz f )Gz ( k ) KGz ( k ) (4.49)
wz
dz
Gz Gb (4.50)
db
§ w (K~z ) wf · ( k ) dz ( k )
¨ ¸Gb K Gb 0 (4.51)
© wb wb ¹ db
w (K~z ) wf dz
K 0 (4.52)
wb wb db
wK ( k ) wf dz
z K 0, i 1, , n (4.53)
wbi wbi dbi
As shown in Equation 4.53, sensitivity analysis requires wK / wbi . When there are
no body forces and wf / wbi 0 , Equation 4.53 yields
190 Analytic Methods for Design Practice
dz wK ( k )
K z , i 1, , n (4.54)
dbi wbi
Ky [ ( k ) My 0 (4.55)
The second term of Equation 4.46d can be expressed with respect to the ith design
variable bi as follows:
w ~ dy wK ~ wM
(Ky [My ) (K [ M ) y [ y, i 1, , n (4.56)
wbi dbi wbi wbi
~
where [ is a constant and the same as [ ( k ) . The third term of Equation 4.46d is
w
(Ky [My ) My (4.57)
w[
d[
G[ Gb (4.58)
db
~ dy wK ~ wM d[
(K [ M ) y [ y My 0, i 1, , n (4.59)
dbi wbi wbi dbi
§ wK ~ wM ·
y T ¨¨ [ ¸y
d[ © wbi wbi ¸¹
, i 1, , n (4.60)
dbi y T My
Equation 4.46d is the same as Equation 4.60. The above process is derived based
on the assumptions that the eigenvalues are not multiple roots and the eigenvectors
are not included in the objective function and constraints. The eigenvalue [
Structural Optimization 191
physically the natural frequency or the buckling load. When the eigenvalue has a
multiple root, it is known that Equation 4.60 is defined by directional derivatives.
Details are not introduced here.
Now the inequality constraints in Equation 4.46e are considered. If we
substitute Equation 4.50 and Equation 4.58 into Equation 4.46e,
wg j § wg j dz · ( k ) § wg j d[ · ( k )
Gb ( k ) ¨¨ ¸Gb ¨
¸
¸
¨ w[ db ¸Gb d g j , j 1, , m (4.61)
wb © wz db ¹ © ¹
If we substitute dz/db and dȟ/db from Equations 4.54 and 4.60 into Equation
4.61, then the equality constraints in Equations 4.46c and 4.46d are eliminated and
the inequality constraints in Equation 4.46e are functions of Gb ( k ) only. It is
essential to obtain the roots (dz/db and dȟ/db) of Equations 4.54 and 4.60 in the
optimization process. This is called sensitivity analysis and is a fairly expensive
process. The detailed process will be introduced later.
The sensitivity of eigenvectors can be derived as well. If we differentiate
Equation 4.45,
dy 1 wM
yT M yT y (4.62)
dbi 2 wbi
ª dy º ª § wK wM · º
ªK [M My º « dbi » « ¨¨ [ ¸y »
« » « © wbi wbi ¸¹ »
« yT M (4.63)
¬ 0 »¼ « d[ » « 1 T wM »
«¬ dbi »¼ « 2 y wb y »
¬ i ¼
Find b, z (4.64a)
192 Analytic Methods for Design Practice
to minimize f b2 z (4.64b)
h1 bz 0 (4.64d)
Solution
Using Equation 4.50, a subproblem to find Gb is established in the manner of
Equation 4.46.
Find Gb (4.65a)
§ dz ·
to minimize ¨ Tb f Tz f ¸Gb (4.65b)
© db ¹
§ dz ·
subject to g1 ¨ Tb g1 Tz g1 ¸Gb t 0 (4.65c)
© db ¹
§ dz ·
h1 ¨ Tb h1 Tz h1 ¸Gb 0 (4.65d)
© db ¹
§ T dz ·
¨ b h1 Tz h1 ¸ 0 (4.66)
© db ¹
dz
is calculated from Equation 4.66 as follows:
db
If we substitute 4.67 into Equations 4.65b and 4.65c, the following subproblem is
obtained:
Find Gb (4.68a)
g2 V 2 d V all (4.69d)
g3 V 3 d V all (4.69e)
Kz f (4.69f)
where
ª b1 b3 b1 b º
« 2L 2L 3
2 L1 2 L3 »»
K E« 1 3 (4.69g)
« b1 b3 b1 b b
2 3 »
« 2L 2 L1 L2 2 L3 »¼
¬ 1 2 L3
z [u1 v1 ]T (4.69h)
When the linearized subproblem is defined, Equation 4.54 is used for Equation
4.69f. Equation 4.54 is recalled as
dz wK ( k )
K z , i 1, , n (4.70)
dbi wbi
194 Analytic Methods for Design Practice
To calculate the right hand side of Equation 4.70, Equation 4.69g is differentiated
as
wK ª1 / 2 L1 1 / 2 L1 º
E« » (4.71a)
wb1 ¬1 / 2 L1 1 / 2 L1 ¼
wK ª0 0 º
E« (4.71b)
wb2 ¬ 0 1 / L2 »¼
wK ª 1 / 2 L3 1 / 2 L3 º
E« (4.71c)
wb3 ¬ 1 / 2 L3 1 / 2 L3 »¼
Equations 4.69g, 4.71 and 4.20 are substituted into Equation 4.70, and the values
of parameters are used. Then Equation 4.70 is solved to obtain dz/dbi as follows:
dz
[ 0.86024 0.50392]T (4.72a)
db1
dz
[ 0.0 0.71781]T (4.72b)
db2
dz
[0.00624 0.00365]T (4.72c)
db3
Since the solution of Equation 4.72 is obtained from Equation 4.70, the
equality constraints can be eliminated. Now the subproblem is made by Equations
4.69aí4.69e by substituting Equation 4.72 into the subproblem. The subproblem
is as follows:
§ wV 1 wV 1 dz1 ·
subject to V 1 ¨¨ ¸Gb d 0 (4.73c)
© wb wz1 db ¸¹
§ wV 2 wV 2 dz 2 ·
V 2 ¨¨ ¸Gb d 0 (4.73d)
© wb wz 2 db ¸¹
§ wV 3 wV 3 dz 3 ·
V 3 ¨¨ ¸Gb d 0 (4.73e)
© wb wz 3 db ¸¹
Structural Optimization 195
where z i is the displacement vector of the ith truss element in the global
coordinate system.
The equation for the stress in a truss element Equation 4.27 is recalled as
E
Vi [ 1 1]Tz i (4.74)
L
wV i
0, i 1, ,3 (4.75)
wb
wV i E
[ 1 1]T, i 1, , 3 (4.76)
wz i L
If we substitute Equations 4.75 and 4.76 into Equation 4.73, we obtain the
following subproblem:
§ E ª dz dz1 dz1 º ·¸
subject to V 1 ¨¨ 0 [ 1 1]T « 1 » Gb d 0 (4.77c)
© L1 ¬ db1 db2 db3 ¼ ¸¹
§ E ª dz dz 2 dz 2 º ·¸
V 2 ¨¨ 0 [ 1 1]T « 2 » Gb d 0 (4.77d)
© L2 ¬ db1 db2 db3 ¼ ¸¹
§ E ª dz dz 3 dz 3 º ·¸
V 3 ¨¨ 0 [ 1 1]T « 3 » Gb d 0 (4.77e)
© L3 ¬ db1 db2 db3 ¼ ¸¹
Equation 4.72 and the transformation matrix are substituted into Equation 4.77
and the values of parameters are used. Then the following subproblem is obtained:
subject to
196 Analytic Methods for Design Practice
§ ª § 3S · § 3S · º
¨ 9 «cos ¨ ¸ sin ¨ ¸ 0 0 »
¨ 207 u 10 4 ¹ © 4 ¹
71.227 u 10 6 ¨ 0 [ 1 1]« © »
2 « § 3S · § 3S ·»
¨¨ « 0 0 cos ¨ ¸ sin ¨ ¸»
© ¬ © 4 ¹ © 4 ¹¼
ª 0.86024 0.0 0.00624 º ·
« 0.50392 0.71781 0.00365» ¸
« » ¸Gb d 0
« 0 0 0 »¸
« » ¸¸
¬ 0 0 0 ¼¹
(4.78c)
§ ª § S· § S· º
¨ «cos ¨ ¸ sin ¨ ¸ 0 0 »
¨ 207 u 109 2 2
6
71.744 u 10 ¨ 0 [ 1 1]« © ¹ © ¹ »
1 « § S· § S ·»
¨¨ « 0 0 cos ¨ ¸ sin ¨ ¸»
© ¬ © 2¹ © 2 ¹¼
ª 0.86024 0. 0 0.00624 º ·
« 0.50392 0.71781 0.00365» ¸
« » ¸Gb d 0
« 0 0 0 »¸
« » ¸¸
¬ 0 0 0 ¼¹
(4.78d)
§ ª § S· § S· º
¨ «cos¨ 4 ¸ sin¨ 4 ¸ 0 0 »
3 ¨ 207 u 109 © ¹ © ¹
516.5 u 10 ¨ 0 [ 1 1]« »
2 « § S· § S ·»
¨¨ « 0 0 cos¨ ¸ sin ¨ ¸»
© ¬ © 4¹ © 4 ¹¼
ª 0.86024 0.0 0.00624 º ·
« 0.50392 0.71781 0.00365» ¸
« » ¸Gb d 0
« 0 0 0 »¸
« » ¸¸
¬ 0 0 0 ¼¹
(4.78e)
subject to
ª AL2 AL2 º
« 0 0 0 0 » ª u1 º
« 2I 2I »« »
« 0 6 3L 0 6 3L » « v1 »
k i z iL
2 EI « 0 3L 2 L2 0 3L L2 » «T 1 » (4.80)
« 2
AL2 » «u »
L3 « AL 0 0 0 0 »« 2 »
« 2I 2I » «v 2 »
« 0 6 3L 0 6 3L » « »
« 0 ¬«T 2 ¼»
¬ 3L L2 0 3L 2 L2 »¼
where I is the moment of inertia, A is the area of the cross section, L is the length
of the element and the transformation matrix is
ª cos D sin D 0 0 0 0º
« sin D cos D 0 0 0 0»»
«
« 0 0 1 0 0 0»
T « » (4.81)
« 0 0 0 cos D sin D 0»
« 0 0 0 sin D cos D 0»
« »
¬« 0 0 0 0 0 1¼»
y P
L1 L2
x h
3
1 ཛ 2 ཛྷ
b
Solution
Since the global and the local coordinate systems are identical, the transformation
matrix is T I. Thus the entire stiffness matrix K in the global coordinate system
is
ª A1 E1 A1 E1 º
« L 0 0 0 0 0 0 0 »
« 1 L1 »
« 0 12 E1 I 1 6 E1 I 1 12 E1 I 1 6 E1 I 1 »
0 0 0 0
« L13 L12 L13 L12 »
« »
« 0 6 E1 I 1 4 E1 I 1 6 E1 I 1 2 E1 I 1 »
0 0 0 0
« L12 L1 L12 L1 »
« »
« A1 E1 0 0
A1 E1 A2 E 2
0 0
A2 E 2
0 0 »
« L1 L1 L2 L2 »
« 12 E1 I 1 6 E1 I 1 12 E1 I 1 12 E 2 I 2 6 E1 I 1 6E 2 I 2 12 E 2 I 2 6E 2 I 2 »
K « 0 0 0 »
« L13 L12 L13 L32 L12 L22 L32 L22 »
« 6 E1 I 1 2 E1 I 1 6 E1 I 1 6E 2 I 2 4 E1 I 1 4 E 2 I 2 6E 2 I 2 2E 2 I 2 »
« 0 0 0 »
« L12 L1 L12 L22 L1 L2 L22 L2 »
« A E A2 E 2 »
« 0 0 0 2 2 0 0 0 0 »
« L2 L2 »
« 12 E 2 I 2 6E 2 I 2 12 E 2 I 2 6E 2 I 2 »
« 0 0 0 0 0 »
« L32 L22 L32 2
L2 »
« 6E 2 I 2 2E 2 I 2 6E 2 I 2 4E 2 I 2 »
« 0 0 0 0 0
¬ L22 L2 L22 L 2 »¼
(4.82)
ª A1 E1 A2 E 2 A2 E 2 º
« L L 0 0
L2
0 »
« 1 2 »
« 12 E1 I 1 12 E 2 I 2 6 E1 I 1 6E 2 I 2 6E 2 I 2 »
0 0
« L13 L32 L12 L22 L22 »
« »
« 6 E1 I 1 6E 2 I 2 4 E1 I 1 4 E 2 I 2 2E 2 I 2 »
K 0 0
« L12 L22 L1 L2 L2 »
« »
« A2 E 2 0 0
A2 E 2
0 »
« L2 L2 »
« 6E 2 I 2 2E 2 I 2 4E 2 I 2 »
« 0 0 »
¬« L22 L2 L 2 ¼»
(4.83)
Find b1 , h1 , b2 , h2 (4.84a)
g2 V 1 d V all
Structural Optimization 199
g3 V 2 d V all
Kz f (4.84d)
where
ª b1 h1 E b2 h2 E b2 h2 E º
« L L 0 0
L2
0 »
« 1 2
3
»
« 12 E b1 h1 12 E b2 h23 3
6 E b1 h1 6 E b2 h2
3
6E b2 h23 »
« 0 3 2 0
«
3
L1 12 L2 12 L12 12
L2 12 L22 12 »»
« 6 E b1 h13 6 E b2 h23 4 E b1 h13 4 E b2 h23 2E b2 h23 »
K « 0 2 2 0 »
« L1 12 L2 12 L1 12 L2 12 L2 12 »
« b2 h2 E b2 h2 E »
« 0 0 0 »
« L2 L2 »
3 3 3
« 6 E b2 h2 2 E b2 h2 4E b2 h2 »
« 0 0 »
¬« L22 12 L2 12 L2 12 ¼»
(4.84e)
z [u 2 v2 T2 u 3 T 3 ]T (4.84f)
f [0 P 0 0 0]T (4.84g)
Solution
Find h (4.85a)
to minimize Uh (4.85b)
Kz f (4.85g)
where
200 Analytic Methods for Design Practice
z [u 2 v2 u3 v 3 ]T (4.85i)
The derivatives of the objective function and constraints can be obtained by the
finite difference method. It can be applied to Equation 4.44b or 4.44e. The
method is explained in Equation 4.44e. Using the forward difference method, the
sensitivity information of Equation 4.44e with respect to bi is
The accuracy of the finite difference method depends on the size of 'bi . The
forward and backward difference methods have the same accuracy. However, the
central difference method has better accuracy, even though it requires more
calculation. For accuracy, it is important to have an appropriate size for 'bi .
Although much research has been conducted, a general rule does not exist for
determination of 'bi . The size of 'bi tends to be reduced due to the enhanced
power of the computer. It is noted that small 'bi can increase the round-off error.
The finite difference method is numerically simple to implement. Moreover, it
has the advantage that a commercial software system can be treated as a black box
in the calculation process. It has the disadvantage that the numerical cost is quite
high. When a number of function calculations are repeatedly required, this method
may not be practical.
Example 4.7 [Sensitivity Analysis for the Three Bar Truss: Finite
Difference Method]
We have the symmetric three bar truss of Example 4.1. Evaluate the sensitivity of
the stress of member 1 with respect to the area by using the forward difference
method and the central difference method. Set the perturbation as 'b 10 8 .
Solution
Using Equation 4.69g, which is the stiffness matrix of the three bar truss,
ª b1 b3 b1 b º
« 2L 2L 3
2 L1 2 L3 »
K E« 1 3 » (4.91)
« b1 b3 b1
b2
b3 »
«¬ 2 L1 2 L3 2 L1 L2 2 L3 »¼
Structural Optimization 203
ª 40000 u 1 / 2 º
f «40000 u 3 / 2» (4.92)
¬ ¼
The displacement vector for the current design is as in Equation 4.20. The stress
of member 1 is as in Equation 4.28 by substituting the displacements in Equation
4.20 into Equation 4.27. The stress for the area increased by ('b) can be obtained
in the same way.
(1) Sensitivity using the forward difference method
Using Equation 4.86
dV 1
[ 1.4119 u 1011 7.4292 u 1010 2.6748 u 108 ] (4.93d)
db
dV 1
[V 1 (b1 'b / 2, b2 , b3 , z (b1 'b / 2, b2 , b3 ))
db1 (4.94a)
V 1 (b1 'b / 2, b2 , b3 , z (b1 'b / 2, b2 , b3 ))] / 'b
dV 1
[V 1 (b1 , b2 'b / 2, b3 , z (b1 , b2 'b / 2, b3 ))
db2 (4.94b)
V 1 (b1 , b2 'b / 2, b3 , z (b1 , b2 'b / 2, b3 ))] / 'b
204 Analytic Methods for Design Practice
dV 1
[V 1 (b1 , b2 , b3 'b / 2, z (b1 , b2 , b3 'b / 2))
db3 (4.94c)
V 1 (b1 , b2 , b3 'b / 2, z (b1 , b2 , b3 'b / 2))] / 'b
dV 1
[ 1.4119 u1011 7.4293 u1010 2.6748 u108 ] (4.94d)
db
dV x
5.6720 u 1010 (4.96a)
dh
dV y
6.4538 u1010 (4.96b)
dh
dW xy
5.0063 u 1010 (4.96c)
dh
Substituting the stresses from Equations 4.37 and 4.41 into Equation 4.97,
the sensitivity information is
Structural Optimization 205
dV x
5.6725 u 1010 (4.98a)
dh
dV y
6.4545 u1010 (4.98b)
dh
dW xy
5.0068 u 1010 (4.98c)
dh
In the analytic method, the differentiation process in Equations 4.53 and 4.60 is
directly programmed. It is the best method when the programming is possible,
since the calculation cost is the lowest. However, the programming is generally
extremely complicated and we have to use the source code of the finite element
method. Therefore, it is almost impossible to use it with commercial software
systems. The analytic method with continuum equations can be utilized to
overcome these difficulties. This will be introduced later. In this section, the
analytic method with discrete equations is introduced.
The governing equation in a discrete form comes from the algebraic equations
in Equation 4.44c or 4.44d. Equation 4.46e becomes Equation 4.61 by using
Equations 4.53 and 4.60. Generally, z and [ of Equation 4.46e do not appear
simultaneously in an equation. Therefore, the differentiation in Equation 4.61
results in either of the following two equations:
dg j wg j wg j dz
(4.99)
dbi wbi wz dbi
dg j wg j wg j d[
(4.100)
dbi wbi w[ dbi
The calculation method for Equation 4.99 is explained here. Equation 4.100
can be calculated in a similar way by using Equation 4.60. The complex
wg j dz
calculation in Equation 4.99 is the second term . Especially the
wz dbi
dz
calculation of is very expensive. There are two methods for the calculation
dbi
dz
of . These are the direct differentiation method and the adjoint variable
dbi
method.
In the direct differentiation method, the following simultaneous equation,
which is derived from Equation 4.53, is solved:
206 Analytic Methods for Design Practice
dz
K v i , i 1, , n (4.101a)
dbi
where
wK ( k ) wf
vi z , i 1, , n (4.101b)
wbi wbi
wg j dz wg j § wK ( k ) wf ·
K 1 ¨¨ z ¸
¸ (4.102)
wz dbi wz © wbi wbi ¹
wg j
Ȝ Tj K 1 (4.103)
wz
T
§ wg j ·
KȜ j ¨ ¸ , j 1,, m (4.104)
¨ wz ¸
© ¹
wg j dz § wK ( k ) wf ·
Ȝ Tj ¨¨ z ¸
¸ (4.105)
wz dbi © wbi wbi ¹
Structural Optimization 207
Example 4.9 [Sensitivity Analysis for the Three Bar Truss: Analytic
Method]
In Example 4.1, obtain the sensitivity information of the stress in member 1 with
respect to the area by using the analytic method.
Solution
Equation 4.99 can be modified as
dV 1 wV 1 wV 1 dz
(4.106)
db wb wz db
Thus
wV 1
0 (4.108)
wb
wV 1 E
[ 1 1]T (4.109)
wz L
dz
is obtained from Equation 4.101. In Equation 4.101, the
db
differentiation of the stiffness matrix should be obtained analytically and it
dz
has been derived in Equation 4.71. The solution of Equation 4.101 is
db
obtained in Equation 4.72. If we substitute Equations 4.72 and 4.108í
4.109 into Equation 4.106, then the sensitivity with respect to the area is
208 Analytic Methods for Design Practice
ª § 3S · § 3S · º
dV 1 E «cos ¨ 4 ¸ sin ¨ 4 ¸ 0 0 »
0 [ 1 1]« © ¹ © ¹ »
db L « § 3S · § 3S ·»
« 0 0 cos ¨ ¸ sin ¨ ¸»
¬ © 4 ¹ © 4 ¹¼
dV 1 wV 1 wV 1 1 § wK wf ·
K ¨ z ¸ (4.111)
db wb wz © wb wb ¹
wV 1 1
ȜT K (4.112)
wz
T
§ wV 1 ·
KȜ ¨ ¸ (4.113)
© wz ¹
If we substitute Equation 4.91 and Equation 4.109 with current values into
Equation 4.113, the following equation is obtained:
T
§ ª § 3S · § 3S ·º ·¸ (4.114)
9 ª2 0 º ¨ 207 u 109 cos ¨ ¸ sin ¨ ¸»
2 u 10 4
u 207 u 10 « »Ȝ ¨ [ 1 1] « © 4 ¹ © 4 ¹» ¸
¬0 2 2 ¼ ¨ 2 « ¸
© ¬« 0 0 ¼» ¹
dV 1 wV1 § wK wf ·
ȜT ¨ z ¸ (4.116)
db wb © wb wb ¹
wK
in Equation 4.116 is in Equation 4.71. Equation 4.71 with current
wb
values is
wK 207 u 10 9 ª1 1º
« » (4.117a)
wb1 2 2 ¬1 1¼
wK ª0 0º
207 u 10 9 « » (4.117b)
wb2 ¬0 1 ¼
wK 207 u 10 9 ª 1 1º
« » (4.117c)
wb3 2 2 ¬ 1 1 ¼
If we substitute Equations 4.115, 4.117 and 4.20 into Equation 4.116, then
the following solutions are obtained:
dV 1 § ª0 0º ª3.4160 u 10 4 º ·
0 [1767.78 1035.53]¨ 207 u 109 « » « » 0 ¸¸
db2 ¨ 4
¬0 1¼ «¬3.4659 u 10 »¼
© ¹
2.6741u 10 8 (4.118c)
Solution
dı wı wı dz
(4.119)
dh wh wz dh
Since the plane stress is calculated from Equation 4.41, each term in
Equation 4.119 is
wı
0 (4.120a)
wh
ª 1 1 1 1 º
ª º« 4 0 0 0 0 »
1 v 0 » 4 4 4
wı E « « 1 1 1 1 »
«v 1 0 »« 0 0 0 0 »
wz 1 v2 « 1 v »« 2 2 2 2 »
«0 0 » 1 1 1 1 1 1 1 1
¬ 2 ¼ «« »»
¬ 2 4 2 4 2 4 2 4¼
(4.120b)
dz
Using the stiffness matrix in Equation 4.36, of Equation 4.119 is
dh
obtained by solving Equation 4.101 as follows:
wK
where it is noted that is analytically calculated. In Example 4.12 it
wh
will be calculated by the semianalytic method.
The solution of Equation 4.121 is
dz z
[ 4.0419 9.9649 4.0759 15.0012]T u 10 1 (4.122)
dh h
Structural Optimization 211
Substituting Equations 4.120 and 4.122 into Equation 4.119, the sensitivity
with respect to the thickness is
dı wı wı dz
dh wh wz dh
ª 0 º
« 0 »
ª 1 1 1 1 º« »
ª º« 4 0 0 0 0 » «0.40419»
1 v 0 »« 4 4 4 « »
E «« 1 1 1 1 » «0.99649»
0 v 1 0 »« 0 0 0 0 »
1 v2 « 1 v »« 2 2 2 2 » «0.40759»
«0 0 »« 1 1 1 1 1 1 1 1 »« »
¬ 2 ¼ « » «1.50012 »
¬ 2 4 2 4 2 4 2 4 ¼«
0 »
« »
«¬ 0 »¼
(4.123)
dV x
5.6725 u 1010 (4.124a)
dh
dV y
6.4545 u 1010 (4.124b)
dh
dW xy
5.0068 u 1010 (4.124c)
dh
dı wı wı 1 § wK wf ·
K ¨ z ¸ (4.125)
dh wh wz © wh wh ¹
wı 1
ȜT K (4.126)
wz
T
§ wı ·
KȜ ¨ ¸ (4.127)
© wz ¹
212 Analytic Methods for Design Practice
wı
EB (4.128)
wz
KȜ (EB ) T (4.129)
ª 5.3333 1.3333 4º
« 2.6667 10.667 2 »
(EB) T « » u 1010 (4.130)
« 5.3333 1.3333 4»
« »
¬ 2.6667 10.667 2¼
where Equation 4.130 is the result after the boundary condition is applied.
If we substitute Equations 4.36 and 4.130 into Equation 4.129, the adjoint
variable vector (matrix) Ȝ is obtained as
dı wı § wK wf ·
Ȝ¨ z ¸ (4.132)
dh wh © wh wh¹
wı
0 (4.133a)
wh
wf
0 (4.133c)
wh
If we substitute Equations 4.40, 4.131 and 4.133 into Equation 4.132, then
the sensitivity is
T
ª 477.91 49.103 408.84º
« 89.348 694.55 91.837 »
dı
0« »
dh « 477.91 49.103 408.84 »
« »
¬ 89.348 694.55 91.837 ¼
ª 4.0419 º ·
« 9.9649 » ¸
« » u 10 4 0 ¸ (4.134)
« 4.0759 » ¸
« » ¸
¸
¬15.0012¼ ¹
dV x
5.6725 u 1010 (4.135a)
dh
dV y
6.4545 u1010 (4.135b)
dh
dW xy
5.0068 u 1010 (4.135c)
dh
214 Analytic Methods for Design Practice
In the analytic method, the differentiation of the stiffness or mass matrix is the
most difficult task in numerical implementation (programming). The semianalytic
method simplifies the process. The differentiation of the matrices is carried out by
the finite difference method. In other words, wK / wbi and wM / wbi in Equations
4.53 and 4.60 are calculated as follows:
Since the amount of calculation in Equations 4.136 and 4.137 is not great, the
central difference method is frequently employed for accuracy. The results of
Equations 4.136 and 4.137 are substituted into Equations 4.53 and 4.60. The
remaining procedures are exactly the same as the one used in the analytic method.
Since the numerical implementation of the semianalytic method is handy, it is
frequently used in commercial software systems. It is noted that the accuracy
depends on the size of 'bi and the error can be enlarged in the later processes.
There are reports that the error is large for shape design variables.
The stiffness matrix in Equation 4.91 is utilized. Using Equation 4.138 the
solution of Equation 4.101 is
dz
[ 0.86024 0.50392]T (4.139a)
db1
dz
[ 0.0 0.71781]T (4.139b)
db2
dz
[0.00624 0.00365]T (4.139c)
db3
The results in Equation 4.139 are the same as those in Equation 4.72.
Therefore, the remaining process is the same as that of Example 4.9.
(2) The adjoint variable method
wK
from the semianalytic method is shown in Equation 4.138. Equation
wb
4.138 is used instead of Equation 4.117 in Example 4.9. The remaining
process is the same as that of Example 4.9.
dz
of Equation 4.141 is
db
dz
[ 4.0419 9.9649 4.0759 15.0012]T u101 (4.142)
dh
So far, sensitivity analysis has been derived by using the discrete equations for the
finite element method. In this section, it is derived based on the continuum
equations. The objective or a constraint function using a continuum equation can
be written as
wg wg dz
where g z , gb and Gz Gb . Calculating Gz is the most important
wz wb db
task in Equation 4.144.
Using the principle of virtual work, the discrete equation in Equation 4.8 yields
z T Kz zT f fTz (4.145)
ab (z, z ) lb ( z ) (4.146)
Ab z f (4.147)
T
³ z ( Ab z f )dȍ 0 (4.148)
ȍ
Equation 4.146 and Equation 4.148 are identical from the following relationships:
lb ( z ) { ³ z T fdȍ (4.149b)
ȍ
wa wl
where aGc b Gb and lGcb Gb . ~z is the current value and is considered
wb wb
constant. If we compare Equation 4.150 with Equation 4.53, which has a
discretized form, then
218 Analytic Methods for Design Practice
§ dz wK ~z ·¸Gb § wf ·
zT ¨ K z T ¨ ¸Gb (4.151)
© db wb ¹ © wb ¹
ab ( Ȝ , Ȝ ) ³ g z Ȝdȍ (4.152)
ȍ
ab (Ȝ , Gz ) ³ g zGzdȍ (4.153)
ȍ
~
G\ ³ g bGbdȍ lGcb (Ȝ ) aGc b ( z , Ȝ ) (4.156)
ȍ
It is noted that the adjoint variable vector Ȝ in Equation 4.156 is the one obtained
from Equation 4.152.
The method with the continuum equations is mathematically well verified. It is
explained in detail in the references (Choi and Kim 2004a, 2004b). In the direct
differentiation method, the sensitivity information is obtained by substituting Gz
of Equation 4.150 into Equation 4.144. On the other hand, the sensitivity
information is obtained from Equation 4.156 in the adjoint variable method. The
difference from the method with discrete equations is that wK / wbi does not need
to be obtained in the method with continuum equations. The reason is that the
continuum equation is differentiated first and then discretized. A simple example
will be presented later.
Structural Optimization 219
The method with continuum equations is classified into two methods as shown
in Table 4.2. They are the continuum–continuum method and the continuum–
discrete method. The continuum–continuum method is used when Gz of Equation
4.150 can be obtained exactly. It can only be applied to simple problems. In
practical problems, the continuum–discrete method is generally employed. That is,
Equation 4.150 or Equation 4.156 is discretized by approximation and the
sensitivity information is evaluated.
l l
\ ³ fzdz ³ ( F JA) zdx (4.157)
0 0
Obtain the sensitivity of Equation 4.157 with respect to the design variables by
using the direct differentiation method and the adjoint variable method. Use the
continuum–continuum method.
Solution
Using the variational method, the governing equation for a beam is
l l l
ab ( z , z ) { ³ EDA2 z xx z xx dx ³ fz dx ³ ( F JA) z dx { lb ( z ) (4.158)
0 0 0
w2z
where z xx . Differentiation of Equation 4.157 yields
wx 2
x h
z l w
l
G\ ³ [( F JA)Gz JzGA]dx (4.159)
0
l
2
ab (Gz , z ) ³ EDA Gz xx z xx dx (4.160)
0
l
lGcb ( z ) ³ Jz dxGA (4.161)
0
l l
aGc b (~z , z ) 2
³ DA z xx z xx dxGE ³ 2 EDAz xx z xx dxGA
0 0
1l 2 2l 2
³ EDA z xx z xx dxGE ³ EDA z xx z xx dxGA (4.162)
E0 A0
1l 2l
aGc b ( z, z ) ³ ( F JA) z dxGE ³ ( F JA) z dxGA (4.163)
E0 A0
l l 1l 2l
2
³ EDA Gz xx z xx dx ³ Jz dxGA ³ ( F JA) z dxGE ³ ( F JA) z dxGA
0 0 E0 A0
(4.164)
x 2 (1 x 2 ) ª 1 2 º
Gz 2 «
JGA ( F JA)GE ( F JA)GA» (4.165)
24 EDA ¬ E A ¼
Substituting Equation 4.165 into Equation 4.159 with real values, the
sensitivity is
Structural Optimization 221
ab (O , O ) ³ ( F JA)O dx (4.167)
~
G\ ³ g bGbdȍ lGcb ( z ) aGc b ( z , z )
ȍ
l l
2 2 2
³ [2Jz 2 EDA( z xx ) ]GAdx [ ³ DA ( z xx ) dx]GE
0 0
ªl ª 25,000 º º
2 2
« ³ «385.6 x (1 x) (6 x 2 6 x 1) 2 » dx »GA 2.08 u 10 11GE
¬0 ¬ 3 ¼ ¼
Given problem
Linearization
Subproblem
LP or QP algorithm
(LP, QP)
No
Check convergence criteria
Yes
New design
End
The approximation method has been proposed to overcome the drawbacks of the
direct method. Approximated functions are generated by considering the
characteristics of structural design. The direct method is also an approximation
method in that the original functions are reduced to linear or quadratic functions.
However, the functions can be approximated to any nonlinear functions in the
approximation method. The functions in structural optimization are frequently
implicit functions with respect to design variables. The implicit functions are
approximated to explicit nonlinear functions.
An optimization problem is defined with the approximated functions. A
nonlinear programming algorithm finds an optimum of the approximated problem.
A typical flow of the approximation method is illustrated in Figure 4.9. The
optimization process of the approximated problem does not find the direction
vector; it finds a new design. When the process does not converge, the
approximated problem is redefined with the new design. The process is not called
Structural Optimization 223
Given problem
Approximation
New design
No
Check convergence criteria
Yes
End
an iteration, which is the term used in the direct method, but is called a cycle.
Therefore, a nonlinear programming algorithm is used once at each cycle.
A new design is found at each cycle in the approximation method. A nonlinear
programming algorithm finds the new design with explicit functions. Therefore,
the performance of the algorithm does not have a large impact on the overall
efficiency. Instead, the quality of the approximation determines the efficiency.
Generally, the convergence criterion is that the objective function is not changed
while all the constraints are satisfied. It is not guaranteed that the final solution
satisfies the KKT condition. It is noted that the İ-active set strategy is also used in
the approximation method. That is, only violated constraints are approximated.
Generally, commercial software systems are coded by using the approximation
method. The functions can be approximated in various manners according to the
application area. Therefore, there are many approximation methods. In this
section, the basic concept is introduced with the results of the early research.
Although Equation 4.46e is used in the direct method, it is a linear
approximation in some sense. From an approximation viewpoint, Equation 4.46e
is
n § wg ·
gL g (b 0 ) ¦ (bi bi 0 )¨¨ ¸
¸ (4.169)
i 1 © wbi ¹ b0
c c(b) (4.170)
n wg wbi
g A (b) g (b 0 ) ¦ (c i (bi ) c i (b0i )) (4.172)
i 1 wbi wc i b0
n b0i wg
g A (b) g (b 0 ) ¦ (bi b0i ) (4.173)
i 1 bi wbi b0
Many other approximation methods are proposed. They are usually defined as in
the above concept (Haftka and GĂrdal 1992).
There is a particular approximation method for a stress constraint function.
Suppose the stress in a beam element of the finite element method is as follows:
where A, I, J, F, M, T, c and r are the area, the moment of inertia, the polar
moment of inertia, the normal force on the area, the moment, the torsion, the
distance from the neutral axis and the distance from the center, respectively.
Equation 4.174 can be approximated to a nonlinear function of b as follows:
wF § wM · § wT ·
F Gb ¨ M Gb ¸c(b) ¨ T ¸r (b)
wb © w b ¹ © wb ¹
g V all 0 (4.175)
A(b) I (b) J (b)
pi
ci bi (4.176)
is used. Using Equation 4.176, the function g is linearly expanded by the Taylor
series as follows:
p
n ª§ b · i º§ b ·§ wg ·
gA g (b 0 ) ¦ «¨¨ i ¸¸ 1»¨¨ 0i ¸¸¨¨ ¸¸ (4.177)
i 1«© b0i ¹ »© pi ¹© wbi ¹b 0
¬ ¼
If we use the condition that the derivatives of g and g A are the same at an another
design point b1 , then p i is
ª§ wg · § wg · º
log «¨¨ ¸
¸
¨
¨ wb
¸ »
¸ »
«© wbi ¹ b1 © i ¹ b0 ¼
¬
pi 1 (4.178)
§b ·
log¨¨ 1i ¸
¸
© b0 i ¹
where p i has a value from –1 to 1. When the value in the logarithm is negative,
then p i 1 . If the result of Equation 4.178 is that p i 0 , the following equation
is used by using the L’Hospital theorem:
ª§ b pi º
·
«¨ i ¸ 1»
«¨© b0i ¸
¹ » §b ·
¬ ¼
lim log¨¨ i ¸
¸ (4.179)
pi o 0 pi © b0 i ¹
Two techniques are often used in structural optimization. They are the application
of multiple loading conditions and design variable linking. In some cases, we have
many loading conditions. When there are multiple loading conditions, the same
number of finite element analyses is performed and the results are incorporated in
the optimization process.
Suppose that the objective function is a function of only design variables and
that constraints are imposed on the displacements and the stresses. Then the
optimization formulation with multiple loading conditions is
Find b R n , z Rl (4.180a)
where NLC is the number of multiple loading conditions. Since the natural
frequency is not dependent on the external loads, it is excluded in Equation 4.180.
In Equation 4.180c, it seems that a finite element analysis is carried out for each of
the multiple loading conditions. However, it should be noted that the external load
Structural Optimization 227
vectors appear only on the right hand side. Therefore, the computational effort is
almost the same as that with the single loading condition because all the loading
conditions are simultaneously considered in the process of solving the
simultaneous equations. Since the worst case is taken into account in Equation
4.180d, the number of constraints does not increase compared to the single loading
case.
A problem can include the displacement vector z in the objective function. In
this case, an artificial variable E is adopted and Equation 4.180b is modified to
Minimize E (4.181a)
That is, the objective function is changed to an objective function and constraints.
In structural design, some design variables must have the same value. The
following equation is employed for design variable linking:
b Ab el (4.182)
where b el is the vector for all the variables and A is the matrix for linking them.
Design variable linking not only enables a meaningful design in an engineering
sense but also reduces the number of design variables in the optimization process.
Because the size of optimization depends on the number of design variables,
design variable linking is fairly useful.
The İ-active set strategy in Chapter 3 excludes constraints that are sufficiently
satisfied. Generally, the number of displacement constraints is not very large.
However, because the stress constraints are defined at all the elements of the finite
element analysis, the number of stress constraints can be extremely large. The İ-
active set strategy can be exploited in this case. With the İ-active set strategy, the
number of constraints in the subproblem is reduced and the effort for sensitivity
analysis is significantly reduced.
l w
Figure 4.10. A cantilever beam of a rectangular cross section with a load at the tip
228 Analytic Methods for Design Practice
h h
Pl Pl
2 2 Pl
V 2 2
6.67 u 1014
I DA 2Dw h
Each term of Equation 4.169 is
V 6.67 u 1014
g (b 0 ) = 1 1 (4.183)
V all V all
wg Pl 4.44 u 1017
(4.184)
ww Dw3hV all V all
wg Pl 1.33 u 1017
(4.185)
wh 2Dw 2 h 2V all V all
wM
0 (4.187)
wb
h
M
2 Plh
gA d1 (4.188)
IV all 2Dw 2 h 2V all
Structural Optimization 229
The sizes of structures are the design variables in size optimization. The
optimization process starts from design parameterization. Design variables can be
selected from any variables that do not change the domain of the finite element
analysis. For example, they are the cross sectional area of the truss element, cross
sectional properties of the beam element and the thickness of the shell element.
Examples of design variables are shown in Table 4.5. The theories in Sections
4.3í4.5 can be directly applied to size optimization. Thus, a detailed explanation
is omitted and instead, some examples are given.
Table 4.6. History of the optimization process for the three bar truss
Table 4.7. History of the optimization process for the member beam
Active
Itera- Objective
b1 h1 b2 h2 constr-
tion function
aint
1 5.0000Eí2 5.0000Eí2 5.0000Eí2 5.0000Eí2 3.9450E+1 2, 3
2 5.3312Eí2 1.4999Eí1 5.4378Eí2 1.5000Eí1 1.2745E+2 2, 3
3 2.5302Eí2 2.4990Eí1 1.6513Eí2 2.5000Eí1 8.2461E+1 2, 3
4 5.0000Eí3 4.1650Eí1 9.2194Eí3 4.1667Eí1 4.6740E+1 2
5 5.0000Eí3 5.4602Eí1 5.0003Eí3 4.6454Eí1 3.9867E+1 2, 3
6 5.0000Eí3 5.5591Eí1 5.0000Eí3 4.6879Eí1 4.0425E+1 2, 3
Based on the results of Example 4.4, the history of the optimization process is
shown in Table 4.6. In Table 4.6, it is noted that the design variable b3 converges
to the lower limit. Thus, we do not need the member of the right hand side.
Solution
The history of the optimization process is shown in Table 4.7. The height of the
cross section is increased and the width is reduced to the lower limit. Overall, the
mass is reduced and the moment of inertia is increased.
Structural Optimization 231
r1
L d e
r2
(a) Circle (b) Domains with some definite shapes (c) Arbitrary shape
The domain of the finite element method is the design variable in shape
optimization. As in size optimization, shape design parameterization is the first
step in shape optimization. The optimization problem is formulated as Equation
4.44 and the subproblem is the same as in Equation 4.46. The obvious difference
from size optimization is the process of sensitivity analysis.
Sensitivity Analysis
In shape optimization, sensitivity analysis is performed as well to establish the
subproblem in Equation 4.46. In general, the shape optimization problem is larger
than the size optimization problem. Therefore, the finite difference method is
rarely used in sensitivity analysis due to cost. The semianalytic method in
Equations 4.136 and 4.137 is usually adopted in commercial software systems. In
shape optimization, numerical error of the semianalytic method can be very large
according to the perturbation of the design variables. In this section, the method
232 Analytic Methods for Design Practice
WV (x )
īIJ
x xW
ȍ ȍW
where
ª z ( x ) z ( x) º
z ' ( x) lim « W » (4.190)
W o0 ¬ W ¼
Using the principle of virtual work and Equation 4.146, the governing equation
in the domain is
a ȍW ( z, z ) l ȍIJ ( z ) (4.191)
T
\ ³³ f ' ( x)dȍ ³ f ( x)(V n)dī ³³ [ f ' (x) div ( f (x)V (x))]dȍ (4.193)
ȍ ī ȍ
where n is the normal vector at the boundary and div means the operator of
wf wf wf
divergence and div f .
wx1 wx2 wx3
Equation 4.191 is
where
T
\ ³³ ( g z z ' g bGb)dȍ ³ g (V n)dī
ȍ ī
T T
³³ g z z g z ( zV ) g b Gb)dȍ ³ g (V n)dī (4.200)
: ī
234 Analytic Methods for Design Practice
aȍ ( Ȝ , Ȝ ) ³³ g z Ȝdȍ (4.201)
ȍ
aȍ (Ȝ , z ) ³³ g z z dȍ (4.202)
ȍ
T T T T
³³ [c(z, ȜV ) f ( ȜV) c( zV , Ȝ ) g z ( zV )]dȍ
ȍ
After the adjoint variable vector Ȝ is obtained from Equation 4.201, Equation
4.204 is calculated. It can be analytically derived in simple problems. In most
problems, it is obtained from a process similar to the assembling process of the
finite element method. The sensitivity in Equation 4.204 can be calculated in the
domain or the boundary. Green’s theorem is utilized in this process.
The results of Equation 4.204 depend on the velocity field V; therefore, the
definition of the velocity field is important. It is related to the design variables.
The boundary of the domain can be defined by certain vectors or functions. A
change of the mesh for the finite element follows and the velocity field is
determined. Details are explained in the references (Choi and Kim 2004a, 2004b).
l
\ ³ fzdx (4.205)
0
Using the continuum–continuum method, obtain \ by (1) the direct method and
(2) the adjoint variable method.
Structural Optimization 235
Solution
f
z ( x) ( x 2 (1 x) 2 ) (4.206)
2 EI
f 2l 5
\ (4.207)
720 EI
f 2l 4
\ Gl (4.208)
144 EI
The solution of the direct method in Example 4.13 can be obtained in the
same way.
(2) The adjoint variable method
As Example 4.13, O z . Equation 4.204 yields
l
\ ³ [2 EI ( z xV ) xx z xx 2 f ( z xV )]dx (4.209)
0
l l l
\ 2 EIz xx ( z xV ) x 0 2( EIz xx ) x ( z xV ) 0 (2 fz EI ( z xx ) 2 )V (4.210)
0
l
\ EI ( z xx ) 2 V (4.211)
0
Gl Gl
In Equations 4.211 and Equation 4.205, V (l ) and V (l ) are
2 2
applied, and then
236 Analytic Methods for Design Practice
f 2l 4
\ Gl (4.212)
144 EI
1 b
a
1
(a) Porous material (b) A hole in a finite element
in it. It is assumed that the holes are evenly distributed as illustrated in Figure
4.15a. In the average sense, the porous material can be modelled as Figure 4.15b
where a large hole exists in a finite element. If the hole is larger than a certain
value, the element is considered as an empty one. In the homogenization method,
the perturbation theory with the multiscale method is exploited. The method is
mathematically well established and it is reliable. The density method regards the
density of an element as a design variable. The density is used to calculate the
modulus of each element. If the density is close to zero, the corresponding
element is considered empty. The density method is slightly weaker in
mathematics than the homogenization method; however, it has been popularized
due to simplicity of the theory and numerical implementation. The density method
is introduced here. Details are given in the references (BednsIe 1998, 2003,
Hassani 1999, Mlejnek 1990).
In general, the stiffness of a structure is maximized in topology optimization.
Maximization of the stiffness is equivalent to minimization of the compliance. The
goal of topology optimization is generally minimization of the compliance. If we
substitute z into z in Equation 4.191, then
Minimize l (z ) (4.214)
E U p E0 (4.215)
where U is the density, which is a design variable and E0 is the real Young’s
modulus. The coefficient p is defined as p ! 1 and Young’s modulus depends on
the density.
Using Equation 4.125, the optimization formulation of topology optimization is
as follows:
Find U (4.216a)
to minimize l (z ) (4.216b)
³ U dȍ d V , 0 U min d U 1 (4.216d)
ȍ
Structural Optimization 239
where V is the given limit of the volume. The problem of Equation 4.216 has
many design variables and a small number of constraints. Thus, the optimality
criterion method is often employed to find the optimum solution. It numerically
solves the KKT condition.
Details are not explained here. The process to update design variables is
where k is the iteration number, K is a tuning parameter and 9 is the move limit.
Bk is calculated from the Lagrange multiplier / k as follows:
pU kp 1 EH (z k )H (z k ) /k (4.218)
Recently, the technique of size optimization has been utilized to solve the
topology optimization problem. Using the finite element equation and Equation
4.215, topology optimization is formulated as
to minimize f T z (4.220b)
NE
p
subject to ¦ Ui k i f (4.220c)
i 1
NE
¦ vi U i d V , 0 U min d U i d 1, i 1, , NE (4.220d)
i 1
where vi is the volume of the ith element and NE is the number of finite elements.
Sensitivity analysis requires the use of the technique of size optimization. As
shown in Equation 4.220c, the design variables (density) are separated from the
finite element equation. Therefore, sensitivity analysis is fairly simple. Since
topology optimization has many design variables, the transformation method in
Section 3.8.3 is often used rather than the primal method in Section 3.8.2.
Solution
Figure 4.16b presents the results of topology optimization by using the density
method. Materials should be kept in the dark areas.
4.7 Exercises
4.1 A vertical load is imposed at the tip of a structure with two truss elements.
Evaluate the displacement of the tip. A 200 m 2 , E 200 GPa ,
L 1.0 m and P 100 kN.
L
$
60
60$ P
Structural Optimization 241
4.2 Solve Exercise 4.1 when each member has two truss elements as
illustrated in the following figure. Compare the results with those of
Exercise 4.1 and analyze the results.
L
$
60
60$ P
4.3 Eliminate the equality constraints and make subproblems for the following
problems. The design variable is b and the initial value is b 0 1.
(1)
Find b, z
to minimize f b2 z 2
subject to g1 b2 z 1 t 0
h1 5z b 2 4 0
(2)
Find b, z
to minimize f b2
subject to g1 b2 z 2 1 t 0
g2 4b z 3 10 d 0
h1 z 2b 2 2 0
4.4 Make a subproblem for Exercise 4.1. The cost function is the volume of
the structure, the design variables are the areas of the members and a
displacement constraint is applied at the tip.
4.5 A clamped beam has a concentrated load at the center. The volume of the
structure is to be minimized while a displacement constraint is defined at
the center. Design variables are the width and the height of the cross
242 Analytic Methods for Design Practice
l/2
P
l w
4.6 In Exercise 4.1, evaluate the sensitivity of the displacement at the tip with
respect to the area by using the following methods when the perturbation
of the area is 10 3 m 2 :
(1) Forward and central difference methods of the finite difference
method.
(2) Direct and adjoint variable methods of the analytic method.
(3) Semianalytic method.
4.7 Evaluate the sensitivity of the displacement with respect to the width and
the height of the cross section for the following structures. Use the direct
and the adjoint variable methods with the continuum–continuum method.
Refer to the data of Example 4.13.
(1)
x h
z l w
(2)
z x h
l w
Structural Optimization 243
z x h
l w
(2)
l/2
P
z x h
l w
4.9 Perform optimization using Exercise 4.1. The objective function is the
total volume of the structure, the design variables are the cross sectional
areas of the members and the tip displacement should be less than 0.05 m.
The upper and lower limits of the design variables are 10 1 m 2 and
10 8 m 2 , respectively.
4.10 In Exercise 4.7, evaluate the sensitivity of the displacement with respect to
the shape variable l by using the direct and adjoint variable methods. Use
the data of Example 4.13 and refer to Example 4.17 for the derivation
process.
244 Analytic Methods for Design Practice
The panel typically consists of two blank pieces and the pieces have different
thicknesses (Part 1 and Part 2). The door in Figure 4.A.2b is manufactured by a
stamping process.
A design process is proposed in Figure 4.A.3. First, the design domain is
selected. The domain may be the entire area or in part of the structure. In
conceptual design, topology optimization is performed for several mass constraints
to determine the tendency of the stiffest structure. The objective is to minimize the
strain energy for the given loading conditions. According to the topology
optimization results, the design domain is divided into a few parts with different
thicknesses. The divided lines are equivalent to the welding lines.
In detailed design, size and shape optimizations are performed. Size
optimization begins with the results from the above part selection process. The
thickness of each part is determined. The final welding lines are rendered by
shape optimization. Sometimes, it is extremely difficult to use shape optimization
for a curved boundary such as the inner panel. Also, the design boundaries of
shape optimization exist within the domain of the finite element method. A
method for the problem is developed for usage with a commercial code.
In many applications, size and shape optimizations can be used simultaneously.
In this problem, there are large differences between size and shape sensitivities.
The order of the size sensitivity is considerably larger. The welding lines do not
move in simultaneous application. Therefore, they are utilized separately as
illustrated in Figure 4.A.3.
246 Analytic Methods for Design Practice
End design
Figure 4.A.3. Design process of the door using various optimization methods
Structural Optimization 247
(1) When the welding lines fluctuate, the lines of the current cycle are
considered as the final design.
(2) When there are no constraint violations at the ith cycle and some constraint
violations at the (i+1)th cycle, the final design is determined by the ith
cycle.
The design time increases with this method because the design change is small
in a cycle. Moreover, a better design may exist, which can be exploited when the
shape sensitivity is small and the shape boundary is within the domain.
x
Load case 1 Load case 2 Load case 3
Figure 4.A.5. Finite element model and loading conditions for the front door
The finite element model is illustrated in Figure 4.A.5 and represents boundary
conditions for analysis, the frame rigidity conditions (load cases 1 and 2) and the
sag condition (load case 3). The number 12345 means that only the rotation with
respect to the z-axis is free and 2356 means that the translation and rotation in the
x-axis are free. The number 2 means that only the translation in the y-axis is fixed.
The finite element model includes the inner and outer panels and the model is
verified via experiments.
The design domain is defined on the inner panel of the front door. The reasons
are that the TWB technology can be easily applied, and the additional effect of
reduction of both manufacturing processes and weight can be obtained by
elimination of reinforcements and some parts. Topology optimization is
performed to determine the number of parts and rough welding lines in conceptual
design. As mentioned earlier, the inner reinforcements are removed to apply the
TWB technology. This is illustrated in Figure 4.A.6. The stiffest structure is
found through topology optimization. A mass fraction constraint is defined for
three cases: 50%, 60% and 70% of the inner panel mass. The objective function is
the linear sum of strain energies under three loading conditions.
The results are illustrated in Figure 4.A.7. The inner panel is divided into three
parts based on the results of the topology optimization. The divided three parts are
Inner reinforcement
(c) 70%
where b is the design variable vector for the thicknesses of the plates for the inner
panel, G BASE
j
is the displacement of the existing model with reinforcements where
the jth load is imposed, G TWB
j
is the displacement of the model with TWB at the
same place of G BASE
j
.
The starting values of b are set to 0.7 mm. It is noted that the thickness
difference between adjacent parts should not be very large in the TWB application.
Therefore, Equation 4.A.1e is defined so as to make one thickness not larger than
twice the size of the other. The constraint is considered as the formability of TWB.
The initial and optimum designs are shown in Table 4.A.1. The results in
Table 4.A.1 are obtained in a continuous space. However, they are not usable in
practical design and therefore, they should be modified to have discrete values.
An available rounded-up value is given to each thickness: b1 = b3 = 0.65 mm and
b2 = 1.3 mm. The rounded-up values are checked for the constraints in Equation
4.A.1. Since all the constraints are satisfied, they are taken as the final design.
Shape optimization is performed to determine the location of the welding lines.
It is noted that shape optimization can reduce the increased weight of the inner
panel due to the rounded-ups of the thickness. The formulation is as follows:
11
{ XYZ} { XYZ}0 ¦ bi { XYZP}i (4.A.2d)
i 1
where { XYZP}i are 11 perturbation vectors to change the domain of X 2 plate and
bi are 11 design variables that determine the amount of perturbations. The
welding lines in shape optimization should maintain the initial shape of the
structure. Thus, the element-based shape optimization technique is utilized and the
final welding line is identical to the boundary of X 2 .
Table 4.A.2 shows the optimization cycles. The lower and upper limits are
defined as Figure 4.A.4. The process terminates in three cycles. The optimization
results are illustrated in Figure 4.A.9. The shape sensitivities are smaller than the
size sensitivities. Therefore, drastic change is not made in the shape optimization
process. From the optimization process, the weight of the inner panel is reduced
by 8.72% compared to the existing one with reinforcements. Since the outer panel
is not included in the optimization process the weight of the entire door is reduced
by 3.23% compared to the original structure. The reduction may not seem to be
very large. However, it is quite a large amount because a car has four doors. Also,
it is noted that a better design exists with the TWB technology, which can reduce
manufacturing cost considerably.
Conclusions
The optimization method for a lightweight TWB door has been investigated. The
design flow is defined with topology, size and shape optimizations. The sequence
of using optimization methods is appropriately defined for the door design. It is
found that the door with TWB is a typical industrial example where various
optimization methods can be utilized. In conceptual design, topology optimization
is performed to determine the number of TWB parts and rough welding lines. Size
and shape optimizations are conducted to find the optimum thickness of each part
and the optimum welding line locations, respectively. The inner panel of the door
is optimized, and the weight of the TWB inner panel is reduced by 8.72%
compared to the original inner panel.
References
Arora JS (1988) Introduction to Optimum Design. McGraw–Hill, New York
Arora JS (1997) Guide to Structural Optimization. The American Society of Civil
Engineers, New York
Arora JS, Haug EJ (1979) Methods of Design Sensitivity Analysis in Structural
Optimization. AIAA Journal 17:970í974
Bathe KJ (1996) Finite Element Procedure. Prentice Hall, Englewood Cliffs, NJ
BednsIe MP, Diaz A, Kikuchi N (1998) Topology and Generalized Layout
Optimization of Elastic Structures. Computer Methods in Applied Mechanics
and Engineering 71:197í224
BednsIe MP, Sigmund O (2003) Topology Optimization. Springer, Berlin
Heidelberg New York
Choi CK (2002) Finite Element Method. Techno Press, Daejon, Korea (in Korean)
Choi KK, Kim NH (2004a) Structural Sensitivity Analysis and Optimization I:
Linear Systems. Springer, Berlin Heidelberg New York
Choi KK, Kim NH (2004b) Structural Sensitivity Analysis and Optimization II:
Nonlinear Systems. Springer, Berlin Heidelberg New York
Cook RD, Malkus DS, Plesha ME (1989) Concepts and Applications of Finite
Element Analysis. 3rd ed., Wiley, New York
Haftka RT, GĂrdal Z (1992) Elements of Structural Optimization. Kluwer,
Dordrecht
Haftka RT, Adelman HM (1989) Recent Developments in Structural Sensitivity
Analysis. Structural and Multidisciplinary Optimization 1:137í151
Haftka RT, Grandhi RV (1985) Structural Shape Optimization–A Survey.
Computer Methods in Applied Mechanics and Engineering 57:99í106
Haug EJ, Arora JS (1979) Applied Optimal Design: Mechanical and Structural
Systems. Wiley Interscience, New York
Haug EJ, Choi KK, Komkov V (1986) Design Sensitivity Analysis of Structural
Systems. Academic Press, Orlando, FL
Hassani B, Hinton E (1999) Homogenization and Structural Topology
Optimization. Springer, Berlin Heidelberg New York
Structural Optimization 253
5.1 Introduction
In Chapter 4, structural optimization was introduced and the external loads were
imposed in a static manner. In this chapter, structural optimization when the
imposed loads have dynamic characteristics is explained. A dynamic load is a load
that has varying magnitude and direction with respect to time. Dynamic response
optimization is structural optimization based on the dynamic loads.
It is fairly important to precisely identify the external loads in structural design.
However, identifying the loads is difficult in most cases. Generally, the loads are
assumed to be static with some specific values and the design process is performed
with the static loads. In reality, loads are dynamically imposed on structures. A
dynamic load is transformed to a static load with a dynamic factor and a structure
is designed based on the transformed static load. However, the structure can be
over-designed or under-designed. Therefore, dynamic loads should be considered
directly in the design process. Dynamic response optimization directly handles
dynamic loads.
Structural optimization is classified into methods using sensitivity information
and methods without sensitivity information. In this chapter, the optimization
methods are assumed to use sensitivity information and the utilized analysis is
linear. Dynamic response optimization is classified into optimization in the time
domain and optimization in the frequency domain. The optimization methods
utilized are basically the same as those used in static response optimization. In
particular, the optimization process in the frequency domain is the same as that of
static response optimization. Since optimization in the time domain solves a
second-order differential equation, it is complicated. This chapter mostly deals
with optimization in the time domain. Optimization in the frequency domain will
be briefly explained at the end of this chapter.
256 Analytic Methods for Design Practice
Find b (5.1a)
to minimize f (b) (5.1b)
where t is time, M is the mass matrix, K is the stiffness matrix, f is the external
force vector, z is the displacement vector and z and z are velocity and
acceleration vectors, respectively. Other notations are the same as those of
Equation 3.1. Optimization in the time domain is attributed to structural
optimization; therefore, the contents in Chapter 4 are valid here. The difference is
that the equality constraints in Equation 5.1c yield a second-order differential
equation.
A typical objective function in the time domain is a min-max problem as
follows:
Minimize f E (5.3a)
Find b (5.4a)
to minimize f (b) (5.4b)
g j (b, z (t u ), z (t u ), z(t u ), t u ) d 0, j 1, , m, u 1, , q
(5.4d)
T
³ g j (b, z (t ), z (t ), z(t ), t ) ! dt d 0 (5.5)
0
where
0 if g j 0
g j (b, z (t ), z (t ), z(t ), t ) ! ® (5.6)
¯g j if g j t 0
258 Analytic Methods for Design Practice
where T is the duration of the time. The basic concept of this transformation is to
reduce many active or violated constraints to one constraint. Satisfaction of
Equation 5.5 is equivalent to satisfaction of Equation 5.1d and vice versa. The
transformation is reasonable in that the number of constraints is reduced
significantly.
When a constraint is violated in multiple sections as shown in Figure 5.1, the
number of the violated sections can be changed. Then Equation 5.5 becomes
nondifferentiable and it may be difficult to find an optimum. Moreover, when an
equivalent functional is used, we need a premise that the Lagrange multiplier is
time invariant during [0, T]. However, it is obvious that the Lagrange multiplier is
time variant. Therefore, the KKT necessary conditions are not generally the same
when Equation 5.1d is directly considered and an equivalent functional is used.
Each violated section can be separated, and a corresponding constraint is
defined. Equation 5.7 shows this method
tk 2
³ g j (b, z (t ), z (t ), z(t ), t ) ! dt d 0, j 1, , m (5.7)
tk 1
where tk1 is the starting point of the kth violated section and tk2 is the ending
point. This is illustrated in Figure 5.1.
0 Time
discrete time domain. The simplest method with this is the worst case approach,
which is illustrated in Figure 5.2.
This approach only takes the maximum violated response as a constraint. The
number of constraints is reduced from the number of time grid points to one,
unless the same valued multiple peaks exist along the time axis. However, the
time when the maximum violated response occurs generally varies as the
optimization process proceeds. A response in the time domain, which is not the
worst case in a previous iteration, can become the worst case in the next iteration.
Thus, the convergence can be slow or the problem can diverge.
g0
0 Time
0 Time
0 Time
0 Time
Figure 5.6. Pointwise constraints at all time grid points with an active set strategy
other words, the local maxima of the current iteration may not be local maxima in
the next iteration. Thus, a few points near the maximum can be included in the
constraint set. The method handles relatively many constraints and is illustrated in
Figure 5.4. The method in Figure 5.5 can be used. The points near the maximum
are considered in an equivalent functional.
Pointwise Constraints at All Time Grid Points with an Active Set Strategy
The simplest treatment of time-dependent constraints would be to consider
responses at all time grid points with an active constraint set strategy. Figure 5.6
represents this method and moreover, deals with the most constraints of all the
methods. This one is simple to implement. If a constraint is not violated at most
time grid points, it can be efficient, since the modern computing environment can
cover more constraints than ever.
g (t ) sin t sin 5t , 0 d t d 2
Solution
The time step D 0.1 , t 0 0, tf 2 , n min 1 and nmax 3. Table 5.1 shows
the calculation results.
The procedure is solved as follows:
(1) Since g (t 0 0) t g 0 g(t) is evaluated at t 0 0.1 0.1. Since g(t) =
0.57926, g(t) is evaluated until g (t ) g (t 0.1) at t = 0.2, t = 0.3,
(2) When t = 0.5, g (t ) g (t 0.1) . Thus, g (t 0.1) is a local maximum and
g(0.4) = 1.29872.
(3) Set t = t + 0.1 = 0.5. g(t) is evaluated at t = 0.4, t = 0.6, which are the
neighboring points of t = 0.5. The function is the largest at t = 0.4. From t
= 0.7, the function is searched with the time step until
262 Analytic Methods for Design Practice
t g(t) t g(t)
0 0.0 1.1 0.18567
0.1 0.57926 1.2 0.65262
0.2 1.04014 1.3 1.17868
0.3 1.29302 1.4 1.64244
0.4 1.29872 1.5 1.93549
0.5 1.07790 1.6 1.98893
0.6 0.70576 1.7 1.79015
0.7 0.29343 1.8 1.38597
0.8 T0.03945 1.9 0.87115
0.9 T0.19420 2.0 0.36528
1.0 T0.11745
2.1 g (t )
1.6
1.1
0.6
0.1
-0.4 t
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
1Tª m 2º
< (b, ș, r, T ) f (b) ³ « ¦ r j {max(0, g j (b) T j )} »dt (5.8)
2 0¬j 1 ¼
If V H 1 and || < || H 2 , then stop. Otherwise, make the following set:
I { j : max | max( g j , T j ) | ! K / D , j 1, , m}
t
264 Analytic Methods for Design Practice
I is the set with constraints whose violations are larger than 1 / D times the
maximum violation of the previous iteration.
Step 5. If V t V prev , which means that the violation is not reduced, update
the parameters as follows:
rj E r j , T (j k 1) T (j k ) / E (5.10)
It is noted that this step is carried out when the constraint violation is
improved. If V d V prev / D , set V prev V and go to Step 2. Otherwise, go
to Step 7.
Step 7. For j I , the parameters are updated as follows:
Go to Step 2. It is noted that this step is carried out when the constraint
violation is larger than 1 / D times the maximum violation of the previous
iteration.
The initial value of T j is usually set to 0. When D has a value from 2 to 3 and E
has a value from 6 to 10, the optimization process is stable.
In Equation 5.8, max(0, g T ) only accounts for the violated constraints at all
the time steps. This process has the effect that the active set strategy is employed.
The termination criteria in Step 4 guarantee that all the constraints are satisfied in
the time domain. Meanwhile, the adjoint variable method is effective for
sensitivity analysis of the augmented Lagrange function. This will be explained
later. In this method, the Lagrange multipliers should be updated carefully. There
are a few studies that investigate the update rules (Chahande and Arora 1993, Kim
and Choi 2000).
Dynamic Response Optimization 265
wf
T f (5.13a)
wb
wg j wg j dz wg j dz wg j dz
T g j (5.13b)
wb wz db wz db wz db
dz dz dz
The difficult terms to calculate in Equation 5.13 are , and . They are
db db db
obtained from the equilibrium (governing) equation in Equation 5.1c. Sensitivity
analysis with Equation 5.1c is discussed.
The direct differentiation method is classified into two methods: the direct
integration method and the modal approximation method. In the direct integration
method, Equation 5.1c is differentiated with respect to design variables and then
integrated. In the second method, the size of Equation 5.1c is reduced by modal
approximation, the reduced equation is differentiated with respect to design
variables and then integrated. The integration method is the same as that for
second-order differential equations.
d 2 § dz · § dz · df (t ) dM dK
M ¨ ¸ K¨ ¸ z z (5.14)
dt 2 ¨© dbi ¸
¹
¨ db
© i
¸
¹ dbi dbi dbi
266 Analytic Methods for Design Practice
dz
Equation 5.14 is a second-order differential equation of . A numerical method
dbi
such as the Newmark method can be employed to integrate Equation 5.14. Using
the Newmark method, the acceleration, velocity and displacement vectors can be
calculated with the time step 't as follows:
dz(t 't )
[M E ('t ) 2 K ] 1[ A1z (t ) A 2 z (t ) A 3z(t )
dbi
dz (t ) dz (t ) dz(t )
K A4 A5 A 6z(t 't )] (5.15a)
dbi dbi dbi
dK
A1 (5.15b)
dbi
dK
A2 't (5.15c)
dbi
§1 · 2 dK
A3 ¨ E ¸('t ) (5.15d)
©2 ¹ dbi
A4 't K (5.15e)
§1 · 2
A5 ¨ E ¸('t ) K (5.15f)
©2 ¹
dM dK
A6 E ('t ) 2 (5.15g)
dbi dbi
where
d 2 § dz · dz d § dz · dz
¨ ¸ , ¨ ¸ (5.18)
dt 2 ¨© dbi ¸
¹ dbi dt ¨© dbi ¸
¹ dbi
Numerical stability is assured when E has a value between 0 and 1/6. Equations
5.15í5.17 are utilized in a subproblem and a direction vector is evaluated from the
Dynamic Response Optimization 267
dM dK
subproblem. If and are analytically evaluated, it is the analytic method.
dbi dbi
When they are evaluated by the finite difference method, it is the semianalytic
method.
When sensitivity analysis is performed by using the direct integration method,
the number of differential equations to be solved is the same as that of the design
variables. Thus, the method may not be efficient for problems with many design
variables.
z (t ) ĭq(t ) (5.19)
where ĭ is the modal matrix and q is the generalized coordinate vector. Equation
5.19 is differentiated by the ith design variable as follows:
dz dĭ dq
qĭ (5.20a)
dbi dbi dbi
dz dĭ dq
q ĭ (5.20b)
dbi dbi dbi
dz dĭ
dq
ĭ
q (5.20c)
dbi dbi dbi
dĭ
in Equation 5.20 is the sensitivity of the modal matrix with respect to the ith
dbi
design variable and it is evaluated from Equation 4.63. q, q and q in Equation
5.20 are obtained from the following equation:
ĭT Mĭq
ĭT Kĭĭ ĭT f (5.21)
dq
Equation 5.21 is derived by substituting Equation 5.19 into Equation 5.1c. ,
dbi
dq
dq
and in Equation 5.20 can be obtained from the following differential
dbi dbi
equation:
268 Analytic Methods for Design Practice
dq dq wf
ĭ T Mĭ ĭ T Kĭ ĭT B 2 q
B1q (5.22a)
dbi dbi wbi
ª wM dĭ º
B1 «ĭ T ĭ ĭT M » (5.22b)
¬ wbi dbi ¼
ª wK dĭ º
B2 «ĭ T ĭ ĭT K » (5.22c)
¬ wbi dbi ¼
Equation 5.22 is made by substituting Equation 5.20 into Equation 5.14. Generally,
the size of Equation 5.22 is very small compared to Equation 5.14. That is, the
dimension of q is a lot smaller than that of z. Therefore, the CPU time for
Equation 5.22 is less than that for Equation 5.14 when a numerical method is used
to solve them.
Generally, the sensitivity of eigenvectors with respect to design variables is
quite small. Therefore, it can be assumed to be zero as follows:
dĭ
#0 (5.23)
dbi
dz dq
ĭ (5.24a)
dbi dbi
dz dq
ĭ (5.24b)
dbi dbi
dz
dq
ĭ (5.24c)
dbi dbi
dq dq wf wM wK
ĭT Mĭ ĭ T Kĭ ĭT ĭT ĭ T
ĭq ĭq (5.25)
dbi dbi wbi wbi wbi
Some commercial structural optimization software systems use the above method
(Moore 1994).
z
k = 4b
f(t) = sin t
m b2
Solution
The governing equation of the system is
1
zp sin t (5.28)
3
1
z zh z p A sin 2t B cos 2t sin t (5.29)
3
1 1
z sin 2t sin t (5.30)
6 3
If we use Equation 5.14, the sensitivity dz/db should satisfy the following
differential equation:
d 2 dz dz
2 db
4 0 2z 4 z (5.31)
dt db
0
dz/db
-0.1
-0.2
-0.3
-0.4
-0.5
-0.6 t
0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2
1 2
y A sin 2t B cos 2t t cos 2t sin t (5.35)
6 9
dz 1 1 2
(t ) sin 2t t cos 2t sin t (5.36)
db 36 6 9
T
g ³ h(z, z , z, b, t )dt (5.37)
0
T T
T
A ³ h(z, z , z, b, t )dt ³ Ȝ (Mz Kz f )dt (5.38)
0 0
where
Mz Kz f 0 (5.39)
dg dA
(5.40)
db db
Therefore, the sensitivity of the constraint is the same as the sensitivity of the
augmented response function.
Differentiation of Equation 5.38 with respect to the ith design variable is
dA T wh T § wM wK wf ·
³ dt ³ Ȝ T ¨¨ z z ¸¸dt
dbi 0 wbi 0 w
© ib wbi wbi ¹
T § wh dz wh dz wh dz · T § dz dz ·
³ ¨¨ ¸¸dt ³ Ȝ T ¨¨ M K ¸dt (5.41)
0 © wz dbi wz dbi wz dbi ¹ 0 © dbi dbi ¸¹
X
T § wh d wh d 2 wh T · dz
³ ¨¨ 2 Ȝ M Ȝ T K ¸¸ dt
0 © wz dt wz dt wz ¹ dbi
T T
§ wh · dz § wh d wh · dz
¨ ȜT M ¸ ¨ Ȝ T M ¸ 0 (5.43)
© wz ¹ dbi 0
© wz dt wz ¹ dbi 0
wh d wh d 2 wh T
Ȝ M ȜT K 0 (5.44a)
wz dt wz dt 2 wz
T
§ T wh · dz
¨Ȝ M ¸ 0 (5.44b)
© wz ¹ dbi 0
T
§ wh T d wh · dz
¨ Ȝ M ¸ 0 (5.44c)
© wz dt wz ¹ dbi 0
dz dz
(0) 0, (0) 0 (5.45)
dbi dbi
T
§ wh d wh d 2 wh ·
MȜ KȜ ¨ ¸
¨ wz dt wz dt 2 wz ¸ (5.46a)
© ¹
wh wh d wh
Ȝ T (T )M (T ) 0 , (T ) Ȝ T (T )M (T ) 0 (5.46b)
wz wz dt wz
dA T wh T § wM wK wf ·
³ dt ³ Ȝ *T ¨¨ z z ¸¸dt (5.47)
dbi 0 wbi 0 w
© ib wbi wbi ¹
dg T wh T § wM wK wf ·
³ dt ³ Ȝ *T ¨¨ z z ¸¸dt (5.48)
dbi 0 wbi 0 © wbi wbi wbi ¹
Equation 5.48 is the sensitivity of the constraints, which uses the adjoint variable
method.
The adjoint variable method can be applied to the constraint of the integral
form as shown in Equation 5.37. The constraint in Equation 5.5, which is an
equivalent functional, has an integral form. Therefore, the adjoint variable method
can be utilized in calculating sensitivity for the constraints of Equation 5.5 (Feng et
al. 1977). Moreover, the adjoint variable method is more efficient than the direct
differentiation method for problems with many design variables and a small
number of active constraints.
However, the adjoint variable method is not appropriate for constraints with a
nonintegral form. A representative case of the nonintegral form is Equation 5.4d
where the constraints are defined at particular times. The Dirac delta function can
be adopted in order to use the adjoint variable method for such constraints. A
constraint defined at a particular t i is
T
g ³ h(b, z, z , z, t )G (t t i )dt (5.50)
0
Since Equation 5.50 has an integral form, the adjoint variable method can be used.
If many constraints such as those in Equation 5.49 are considered at many time
grid points, the above method can be inefficient. If we have q constraints such as
in Equation 5.50, we will have q augmented response functions. Then we have to
solve the differential equation in Equation 5.50 q times. If q is large, the adjoint
variable method does not have any advantages compared to the direct
differentiation method. Therefore, the adjoint variable method is not usually used
for constraints such as those in Equation 5.49.
1Tª m 2º
h ³ « ¦ r j {max(0, g j (b, t ) T j (t ))} »dt (5.51)
2 0¬j 1 ¼
Thus, the sensitivity of Equation 5.8 with respect to the ith design variable is
d< wf T wh T § wM wK wf ·
³ dt ³ Ȝ T ¨¨ z z ¸¸dt (5.52)
dbi wbi 0 wbi 0 w
© ib wbi wbi ¹
Equation 5.52 consists of the sensitivity of the objective function f(b) and that of
the penalty term. The sensitivity of the objective function can be easily obtained.
We need an adjoint variable vector Ȝ for the sensitivity of the penalty term. The
adjoint variable vector is obtained from the following adjoint variable equation:
ª m wg j d m wg j ½
MȜ KȜ « ¦ r j max(0, g j T j ) ® ¦ r j max(0, g j T j ) ¾
¬« j 1 wz dt ¯ j 1 wz ¿
T
d2 m wg j ½º
2 ®
¦ r j max(0, g j T j ) ¾» (5.53a)
dt ¯i 1 wz ¿¼»
m wgi
Ȝ T (T )M ¦ r j max(0, g j (b, z, z , z, T ) T j (T )) (T ) 0 (5.53b)
j 1 wz
m wg j
Ȝ T (T )M ¦ r j max(0, g j (b, z, z , z, T ) T j (T )) (T )
j 1 wz
d m wg j
¦ r j max(0, g j (b, z, z , z, T ) T j (T )) (T ) 0 (5.53c)
dt j 1 wz
The time finite element method is a method to solve the differential equation of
Equation 5.1 by transforming it to an algebraic equation. When the time finite
element method is utilized, sensitivity is evaluated by direct differentiation of the
Dynamic Response Optimization 275
algebraic equation. Suppose Equation 5.1c has one degree of freedom. The
solution of Equation 5.1c is assumed to be as follows:
N
z (t ) ¦ q j I j (t ) (5.54)
j 1
T
³Q ( Mz Kz f )dt 0 (5.55)
0
ª B M ȥ (T )º ªq j º ª a* º
«ij(T ) T « » (5.56a)
¬ 0 »¼ «¬ P »¼ «¬ z (T )»¼
where
T
³ ( K\ iI j M\ iI j )dt , ȥ
T
B [\ 1 \ 2 \ n ] (5.56b)
0
T
a* M\ i (0) z (0) ³ f\ i dt , ij T [I1 I 2 I n ] (5.56c)
0
P z (T ) (5.56d)
ª wq j º ª wa * º
ª B M ȥ (T )º « wb » ª wB wM º ªq º « »
ȥ (T )» j wb
«ij(T ) T « » « wb wb « P » « wz (T ) » (5.57)
¬ 0 »¼ « wP » « 0 0 »¬ ¼ « »
¬ ¼
¬« wb ¼» ¬« wb ¼»
where the elements of the right hand side can be easily obtained by partial
differentiation. This is well explained in a reference (Kapania and Park 1996).
276 Analytic Methods for Design Practice
5.5 Approximation
Transient response analysis is required for optimization in the time domain. It is
well known that the analysis is costly. Therefore, many researchers have tried to
approximate the dynamic response. The approximation is classified into global
approximation and local approximation. The two methods are briefly explained
below.
Start
End
Figure 5.10. Flow chart of optimization using RSM in the time domain
Find b (5.59a)
~
to minimize f ( k ) (b) (5.59b)
278 Analytic Methods for Design Practice
s L d b b ( k 1) d s U (5.59d)
where ~ means approximation and s U and s L are the upper and the lower bounds
of the move limit, respectively. The approximated ~r is a vector of the explicit
functions of intermediate variables (N) or design variables. The intermediate
variables are used as follows:
~ wr
r ( N, t ) r0 (N, t ) 'N (5.60)
wN
N N(b) (5.61)
ª 2
T m
« § P j (t ) · 1 ° § P j (t ) ·½° º»
< f (b) ³ ¦ P j (t ) max ¨ ¸
, g j r j ®max ¨ ¸
, g j ¾ dt
0j 1
« ¨ rj ¸ 2 °̄ ¨ rj ¸° »
¬« © ¹ © ¹¿ ¼»
(5.62)
ª 2
~ ~ T m
« § P j (t ) ~ · 1 ° § P j (t ) ~ ·½° º»
< (D ) f (D ) ³ ¦ P j (t ) max¨ , g j ¸ r j ®max¨ , g j ¸¾ dt
0j 1
« ¨ rj ¸ 2 °̄ ¨ rj ¸° »
¬« © ¹ © ¹¿ ¼»
(5.63a)
where
~
f (D ) f (b ( k ) ) T f (b ( k ) )(Ds ( k ) ) (5.63b)
Dynamic Response Optimization 279
g~ j (D , t ) g j (b ( k ) , t ) T g j (b ( k ) , t )(Ds ( k ) ) (5.63c)
f eq { Kz f Mz (5.64)
When a dynamic load is applied to a structure, the equivalent static load is defined
as the static load that makes the same displacement field as a dynamic load at an
arbitrary time. After z is calculated from Equation 5.1c, f eq is evaluated by
multiplying the stiffness matrix and z as shown in Equation 5.64. The equivalent
static loads are calculated at all the time grid points.
In engineering practice, dynamic factors are usually used to transform dynamic
loads into static loads. This may generate a dangerous design. Figure 5.11 shows
the behavior of a structure under a dynamic load. If a dynamic load is imposed as
280 Analytic Methods for Design Practice
(a) (b)
Figure 5.12. A dynamic load (a) and the corresponding equivalent static loads (b)
illustrated in Figure 5.11a, the structure vibrates and has the modes shown in
Figures 5.11b and c. The design with a dynamic factor only accounts for the mode
in Figure 5.11b. However, it is obvious that the mode in Figure 5.11c should be
considered. Since the equivalent static loads are calculated at all the time grid
points, all the modes of the behavior are considered.
It is noted that although a dynamic load is imposed on a node as illustrated in
Figure 5.12a, the equivalent loads are generated for all nodes as shown in Figure
5.12b. The reason is that the equivalent static loads include the external force as
well as the inertia forces at all nodes. In Equation 5.64, it is noted that the
elements of the inertia force are not usually zeros and the corresponding elements
of the equivalent static load f eq are not zeros.
Equation 5.64 defines the equivalent static loads with respect to displacements.
In a similar way, the equivalent inertia load vector with respect to accelerations is
defined as follows:
p eq { Mz f Kz (5.65)
The equivalent inertia load is defined as the loads that make the same acceleration
field as a dynamic load at an arbitrary time. In real implementation, z is
calculated from Equation 5.1c and p eq is evaluated by multiplying the mass
matrix and z as shown in Equation 5.65. The equivalent inertia loads are utilized
when a constraint for acceleration is included in the optimization problem.
Dynamic Response Optimization 281
u
f eq K (b p )z u , u 1, , q (5.66)
Find b p 1 (5.68a)
to minimize f (b p 1 ) (5.68b)
u
subject to K (b p 1 )y u f eq , u 1, , q (5.68c)
M (b p 1 )a u p ueq , u 1, , q (5.68d)
g ju (b p 1 , z u , a u ) d 0, u 1, , q, j 1, , m (5.68e)
|| b p 1 b p ||
dH (5.69)
|| b p ||
equivalent static and inertia loads are updated by the optimum solution of Equation
5.68.
If the design variables are not changed in the process of Equation 5.68, then the
equivalent static and inertia loads are not changed. Therefore, when the design
variables are not changed, the optimization process terminates and the design
variables are considered as the optimum. It has been proved that the KKT
necessary condition of Equation 5.68 in the last cycle is the same as that of
Equation 5.1. The proof is shown in a reference (Park and Kang 2003).
In Step 3, the equivalent static and inertia loads are obtained at q time grid
points. It is noted that the total number of time grid points is q. They are treated
as q multiple loading conditions in the static response optimization process as
shown in Equations 5.68c and 5.68d. It is well known that the multiple loading
conditions are easily handled with low cost in structural optimization because only
the worst cases are considered in the optimization process.
The advantage of the proposed method is that the well developed static
response optimization is exploited. Modern structural optimization software
systems can easily handle large-scale structures. Therefore, the method using
equivalent static loads can easily optimize large structures in the time domain
(Choi and Park 2002, Kang et al. 2003).
f(t)
11.5 kN
f(t)
0.01 s
(a) (b)
Optimization is carried out with equivalent static loads. The results are shown
in Figure 5.15. When transient analysis is performed, the maximum stress occurs
at point A of Figure 5.15 and the magnitude is about 250 MPa. It is noted that the
constraints are satisfied.
5×6.35 m
y
f(t) f(t)
(10) 16 (8)12 (6) 8 (4) 4 (2) 1 (1) 59.3 kN
17 15 13 11 9 7 5 3 2 6.35 m
18 14 10 6
2s
(11) (9) (7) (5) (3) x
Cycle
Design Initial value 1 2 3 4 5
variable
Upper members
64.52 113.34 129.03 129.03 129.03 129.03
(cm2)
Lower members
139.68 121.22 149.16 152.60 168.74 143.08
(cm2)
Vertical
80.65 29.21 40.22 34.91 27.72 44.35
members (cm2)
Diagonal
45.61 53.11 64.51 64.01 62.97 64.52
members (cm2)
x3 (cm) 0.00 í508.00 í494.03 í495.30 í450.85 í500.38
y3 (cm) 0.00 204.47 185.42 187.96 250.19 138.43
x5 (cm) 0.00 í374.65 í294.64 í285.75 í232.41 í303.53
y5 (cm) 0.00 101.60 74.93 102.87 175.26 53.34
x7 (cm) 0.00 í222.25 í218.44 í210.82 8.89 í254.00
y7 (cm) 0.00 24.13 2.54 20.32 85.09 í12.70
x9 (cm) 0.00 í194.31 í220.98 í189.23 í106.68 í226.06
y9 (cm) 0.00 í30.48 í22.86 í33.02 í22.86 í27.94
Mass (kg) 2916.78 2688.72 3260.07 3241.71 3313.62 3260.90
Tip
displacement í27.4 í24.7 í20.3 í20.4 í20.3 í20.3
(cm)
í156.5 í133.1 í111.7 í112.7 í113.2 í106.5
Stress (MPa)
(16) (15) (15) (15) (15) (8)
*The number of the stress row indicates the member where the maximum stress occurs
Dynamic Response Optimization 285
y
x
Minimize the volume of the beam by using the equivalent inertia loads while the
tip acceleration is less than 40m/s 2 . The acceleration from 0.02 seconds to 0.1
seconds is considered and the number of time grid points is less than 4000. Since
the acceleration is drastically changed in the time domain, there should be many
time grid points. Design variables are the width (b) and the height (h) of the cross
section. The initial values are 0.01 m and 0.02 m, respectively. Young’s modulus
and the density are 300 MPa and 1700 kg/m 3 , respectively.
f(t)
b
1m
Initial Optimum
Width (m) 0.01 0.01195
Height (m) 0.02 0.02300
Volume (m3) 0.2 0.27483
No. of cycles 6
60
40
20
Acceleration (m/s2)
–20
–40
Time (s)
–60
0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
Solution
The volume is multiplied by 1000 and used as the objective function as follows:
The first natural frequency of the initial design is 1.356 Hz. The results of the
transient analysis for the initial design are illustrated in Figure 5.19. The
acceleration constraint is violated.
Table 5.3 shows the optimization results by using the equivalent inertia loads.
The tip acceleration is illustrated in Figure 5.20. It is noted that the acceleration
constraint is satisfied.
60
40
Acceleration (m/s2)
20
–20
–40
Time (s)
–60
0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09 0.1
and rigid joints are not yet clearly defined, flexible body dynamics is exploited for
systems with high speeds.
The loads in flexible multibody dynamics are external loads, reaction forces of
the joints, centrifugal forces of the elements, etc. They are dynamic and thus,
optimization of flexible multibody dynamic systems is attributed to dynamic
response optimization. The method using equivalent static loads is utilized for
optimization of flexible multibody dynamic systems (Kang et al. 2005).
Mz Kz fc fe fv (5.72a)
ªm RR m RT m Rf º
« »
M « mTT mTf » (5.72b)
« sym m ff »
¬ ¼
ª0 0 0 º
« »
K «0 0 0 » (5.72c)
«0 0 K ff »
¬ ¼
ªR º
« »
z «ș » (5.72d)
«q f »
¬ ¼
288 Analytic Methods for Design Practice
ª C TR º
« »
fc « CTT »Ȝ (5.72e)
«C T »
¬ qf ¼
ª(Q e ) R º
« »
fe « (Q e ) T » (5.72f)
«(Q e ) f »
¬ ¼
ª(Q v ) R º
« »
fv « (Q v ) T » (5.72g)
«(Q v ) f »
¬ ¼
i mi și mi qi
miRRR CTRi Ȝ (Qie ) R (Qiv ) R (5.73a)
RT Rf f
i mi și mi qi
miRT R CTT i Ȝ (Qie )T (Qiv )T (5.73b)
TT Tf f
i mi și mi q
miRf R i i i
CTqi Ȝ (Qie ) f (Qiv ) f
fT ff f K ff q f f
(5.73c)
i mi și mi q
Kiff qif miRf R i T i i
fT ff f Cqi Ȝ (Qe ) f (Qv ) f
f
(5.74)
Dynamic Response Optimization 289
f eq { K iff q if i m i ș i m i q
m iRf R i T i i
fT ff f C q i Ȝ (Q e ) f (Q v ) f
f
(5.75)
The first two terms on the right hand side in Equation 5.75 represent the influence
from the interrelation between the rigid body motion and the elastic deformation.
The third term is the inertia force from the deformation of the flexible bodies.
These three terms are developed by the flexible bodies and other terms are from
the rigid bodies.
u
f eq K ff (b p )q uff , u 1, , q (5.76)
Step 5. Using the equivalent loads in Step 4, the static response optimization
problem in Equation 5.77 is to be solved. Structural optimization software
can be employed.
Find b p 1 (5.77a)
to minimize f (b p 1 ) (5.77b)
u
subject to f eq K ff (b p 1 )z uff , u 1, , q (5.77c)
g ju (b p 1 , z uff ) d 0, u 1, , q, j 1, , m (5.77d)
290 Analytic Methods for Design Practice
|| b p 1 b p ||
dH (5.78)
|| b p ||
terminate the process and the solution of Equation 5.77 is the optimum. H
is a sufficiently small number. Otherwise, go to Step 7.
Step 7. Set p = p + 1 and go to Step 2.
0.5 T 2St
'X E 'YE (t sin ), t [0, 0.66] (5.79)
T 2S T
Equation 5.80 represents stress constraints and Gy a and Gy b in Equation 5.81 are
vertical displacements at nodes a and b, respectively. The initial value of each
design variable is 50 mm. Optimize the system.
a
D12 D21
D22
D11
b
E
Table 5.4. Optimization results of the robot arm using equivalent static loads (mm, N)
D11 54.500
D12 38.525
D21 30.400
D22 22.700
Weight 13.229
Solution
The problem converges in six cycles and the optimum solution is shown in Table
5.4. The number of modes for transient analysis is three. The first two modes are
for bending and the third one is for the deformation in the axial direction. Since
the number of time grid points is 300, 300 sets of equivalent loads are utilized in
Equation 5.80 as multiple loading conditions. The active set strategy and the
approximation by intermediate variables are adopted in static response
optimization in Equation 5.77.
2
d2
1 d1
3 d3
Table 5.5. Optimization results of the four bar linkage (mm, kg)
4 S
V ik 3
M ik (5.82)
Ai 2
where Ai is the area of the ith link and M ik is the bending moment at the kth node
of the ith link. The objective function is the total mass of the system. Determine
the design variables while the bending stresses in Equation 5.82 are less than 27.58
MPa in a steady state. The initial value of each design variable is 356.8 mm.
Solution
Table 5.5 shows the optimization results. This problem has many local optimum
values (Etman and Van 1998). Therefore, the optimum in Table 5.5 is one of
many local optima. Transient analysis is performed using the first three modes.
The problem converges in 16 cycles. The number of time grid points is 500.
p [ p1 p n ]T (5.83b)
z x sinZt (5.84)
(Z 2 M K )x p (5.85)
Find b R n (5.86a)
to minimize f (b) (5.86b)
g j (b, x i , Z i ) d 0, i 1, , r , j 1, , m (5.86d)
bL d b d bU (5.86e)
ªm 0 º ª z1 º ª 2k k º ª z1 º ª pº
« 0 m» « z » « k « »
2k »¼ ¬ z 2 ¼ « 0 » sin Zt (5.87)
¬ ¼¬ 2 ¼ ¬ ¬ ¼
ª z1 º ª x1 º
«z » « x » sin Zt (5.88)
¬ 2¼ ¬ 2¼
§ ªm 0 º ª 2k k º · ª x1 º ª pº
¨Z 2 « »« ¸« » (5.89)
¨
© ¬ 0 m¼ ¬ k 2k »¼ ¸¹ ¬ x 2 ¼ «0»
¬ ¼
(2k mZ 2 ) p
x1 (5.90a)
§ k ·2
2 ¨§
·§ § 3k · 2 ·
m ¨ ¸ Z 2 ¸¨ ¨ ¸ Z 2 ¸
¨© m ¹ ¸¨ © m ¹ ¸
© ¹© ¹
kp
x2 (5.90b)
§ k· 2
·§ § 3k · 2 ·
2 ¨§
m ¨ ¸ Z 2 ¸¨ ¨ ¸ Z 2 ¸
¨© m ¹ ¸¨ © m ¹ ¸
© ¹© ¹
10
x1 (5.91a)
36 Z 2
20
x2 (5.91b)
(4 Z 2 )(36 Z 2 )
Dynamic Response Optimization 295
X
90
70 xx1
1
50 xx2
2
30
10
Z
Ṗ㰚㭒䕢㑮
-10 0 1 2 3 4 5 6 7 8 9 10
-30
-50
-70
-90
The graph of Equation 5.91 with respect to the range of Z is illustrated in Figure
5.24.
Direct Differentiation
It is assumed that the external forces are not functions of design variables.
Differentiation of Equation 5.86c with respect to the ith design variable yields
dx § 2 dM dK ·
(Z 2 M K ) ¨¨ Z ¸x (5.92)
dbi © dbi dbi ¸¹
x ĭq (5.93)
where ĭ is the modal matrix. It is noted that the dimension of q is less than that
of x. Differentiation of Equation 5.93 with respect to the ith design variable yields
dx dĭ dq
qĭ (5.94)
dbi dbi dbi
dĭ dq
where is obtained from the sensitivity analysis of eigenvectors. is
dbi dbi
calculated from the following equation:
dq dĭ § wM wK ·
(Z 2 ĭT Mĭ ĭT Kĭ) ĭT (Z 2 M K ) q ĭT ¨¨ Z 2 ¸ĭq
dbi dbi © wbi wbi ¸¹
(5.95)
dx dq
#ĭ (5.96)
dbi dbi
dq § wM wK ·
(Z 2 ĭ T Mĭ ĭ T Kĭ) # ĭ T ¨¨ Z 2 ¸ĭq (5.97)
dbi © wbi wbi ¸¹
dq
Equation 5.96 represents sensitivity information and is calculated from
dbi
Equation 5.97.
dg j wg j wg j dx
(5.98)
dbi wbi wx dbi
Dynamic Response Optimization 297
dx § wM wK ·
(Z 2 M K ) 1 ¨¨ Z 2 ¸x (5.99)
dbi © wbi wbi ¸¹
dg j wg j wg j § wM wK ·
(Z 2 M K ) 1 ¨¨ Z 2 ¸x (5.100)
dbi wbi wx © wbi wbi ¸¹
wg j
ȜT (Z 2 M K ) 1 (5.101)
wx
T
§ wg j ·
(Z 2 M K )Ȝ ¨
¨ wx
¸
¸ (5.102)
© ¹
where the mass matrix M and the stiffness matrix K are assumed to be symmetric.
Using Equation 5.101, Equation 5.100 is
dg j wg j § wM wK ·
Ȝ T ¨¨ Z 2 ¸x (5.103)
dbi wbi © wbi wbi ¸¹
The adjoint variable vector Ȝ is obtained from Equation 5.102 and substituted into
dg j
Equation 5.103. Then is obtained.
dbi
Solution
The mass matrix and the stiffness matrix in the global coordinate system are as
follows:
298 Analytic Methods for Design Practice
ª b1 b3 b1 b º
« 2L 2L 3
2 L1 2 L3 »»
K E« 1 3 (5.104)
« b1 b3 b1 b b
2 3 »
« 2L 2L 2 L1 L 2 2 L3 »¼
¬ 1 3
ª b1 L1 b3 L3 b1 L1 b3 L3 º
« »
M U« 6 6 6 6 (5.105)
« b1 L1 b3 L3 b1 L1 b2 L 2 b3 L3 »»
¬« 6 6 6 3 6 ¼»
ª2 0 º
K 2 u 10 4 u 207 u 10 9 « » (5.106)
¬0 2 2 ¼
2 u 7830 u 10 4 ª4 2 0 º
M « » (5.107)
3 «¬ 0 1 4 2 »¼
ª 40000 u 1 / 2 º
p «40000 u 3 / 2» (5.108)
¬ ¼
wK E ª1 1º 207 u 10 9 ª1 1º
«1 1» « » (5.110)
wb1 2 L1 ¬ ¼ 2 2 ¬1 1¼
wM UL1 ª1 1º 7830 u 2 ª1 1º
« » «1 1» (5.111)
wb1 6 ¬1 1¼ 6 ¬ ¼
§ 0 º 2 u 7830 u 10 4 ª4 2 0 º ·¸ dx
¨ 20 2 u 2 u 10 4 u 207 u 10 9 ª2
¨ « 0 2 2 » « »¸
© ¬ ¼ 3 ¬« 0 1 4 2 ¼» ¹ db1
§ 9 ·
= ¨ 20 2 u 7830 u 2 ª«1 1º» 207 u 10 ª«1 1º» ¸ ª«0.00034160º» (5.112)
¨ 6 ¬1 1 ¼ 2 2 ¬1 1 ¼ ¸¹ ¬0.00034659¼
©
dx
[ 0.86026 0.50392]T (5.113)
db1
v1 [1 0]T (5.114b)
v2 [0 1]T (5.114c)
In this problem, the first eigenvector (horizontal direction) and the second
one (vertical direction) are equally important. Therefore, the modal matrix
for all the DOF’s should be used. However, only the first eigenvector is
utilized to show approximation. Therefore, x is approximated by q as
follows:
x v1q1 (5.115)
At the given design point, Equation 5.115 is substituted into Equation 5.85
as follows:
§ 0 º 2 u 7830 u 10 4 ª4 2 0 º ·¸ ª1º
¨ 20 2 u 2 u 10 4 u 207 u 10 9 ª2
¨ « » « » « » q1
© ¬0 2 2 ¼ 3 «¬ 0 1 4 2 »¼ ¸¹ ¬0¼
ª 40000 u1 / 2 º (5.116)
«40000 u 3 / 2»
¬ ¼
dx dq
# v1 1 (5.118)
db1 db1
dq1
After using Equation 5.97, is calculated from the following equation:
db1
dq1 § wM wK ·
v1T (Z 2 M K ) v1 v1T ¨¨ Z 2 ¸ v1q1 (5.119)
db1 © wb1 wb1 ¸¹
If we use the given data for Equation 5.119, the following is obtained:
dq1
0.4270 (5.120)
db1
dx
# [ 0.4270 0]T (5.121)
db1
dx j § wM wK ·
Ȝ Tj ¨¨ Z 2 ¸x, j 1, 2 (5.122)
db1 © wb1 wb1 ¸¹
§ 0 º 2 u 7830 u 10 4 ª4 2 0 º ·¸ T
¨ 20 2 u 2 u 10 4 u 207 u 10 9 ª2 « » ¸Ȝ1
¨ « »
© ¬0 2 2 ¼ 3 ¬« 0 1 4 2 »¼ ¹
ª1º (5.123)
«0 »
¬ ¼
§ 0 º 2 u 7830 u 10 4 ª4 2 0 º ·¸ T
¨ 20 2 u 2 u 10 4 u 207 u 10 9 ª2
¨ « 0 2 2 » « » ¸Ȝ 2
© ¬ ¼ 3 ¬« 0 1 4 2 ¼» ¹
Dynamic Response Optimization 301
ª0 º (5.124)
«1»
¬ ¼
From Equations 5.123 and 5.124, the following adjoint variable vectors are
calculated:
Ȝ 1T [1.7080 0] u 10 8 (5.125)
Ȝ T2 [0 1.0005] u 10 8 (5.126)
Substituting Equations 5.125 and 5.126 into Equation 5.122, the following
sensitivities are obtained:
g j (b, x i , Z i ) d 0, i 1, , r , j 1, , m (5.129)
Find b (5.130a)
to minimize the maximum amplitude at a DOF (5.130b)
0 Frequency
Z1Z 2 Z 3 … … Zr
g j (b, x i , Z i ) d 0, i 1, , r , j 1, , m (5.130d)
bL d b d bU (5.130e)
The methods to solve Equation 5.130 are classified into two types. One method is
to use an artificial variable and the other is a method to minimize the area under
the amplitude versus the frequency curve.
Find b, E (5.131a)
to minimize E (5.131b)
g j (b, xi , Z i ) d 0, i 1, , r , j 1, , m (5.131d)
Find b, x (5.132a)
Dynamic Response Optimization 303
x Maximum amplitude
0 Frequency
x Z2
³ xdZ
Z1
0 Frequency
Z2
to minimize ³ xdZ (5.132b)
Z1
g j (b, x i , Z i ) d 0, i 1, , r , j 1, , m (5.132d)
bL d b d bU (5.132e)
In Figure 5.27, the area under the amplitude versus the frequency curve as
illustrated. Equation 5.132 minimizes the area. Equation 5.132 is not the same as
Equation 5.130. However, when the area is reduced, the maximum amplitude
tends to be reduced. Thus, the formulation in Equation 5.132 is utilized to
minimize the maximum amplitude in an engineering sense.
h
1 2 3 4 5 b
Solution
Find b, x, E (5.133a)
to minimize E (5.133b)
volume d 15 u 10 6 (5.133d)
all amplitudes at x5 y in a given range of frequency d E
(5.133e)
20 Amplitude
18
16
14
12
10
8
6
4
2
0
9 14 19 24 29 34 39 44 49 54 f (Hz)
1.6 Amplitude
1.4
1.2
1
0.8
0.6
0.4
0.2 f (Hz)
0
9 15 21 27 33 39 45 51 57
Figure 5.30. Harmonic response of the optimum results–the method with an artificial
variable
bL d b d bU (5.133f)
The optimization results are shown in Table 5.6. The harmonic response
from the optimum in Table 5.6 is illustrated in Figure 5.30. The maximum
amplitude (1.48 mm) is significantly reduced and the volume is also
reduced by half of the initial design.
(2) Minimization of the area
The maximum amplitude is indirectly reduced by minimizing the area
between the amplitude curve and the frequency axis. The following
optimization problem is solved:
Find b, x (5.134a)
( x5 y1 x5 y n ) n 1
to minimize 'f ¦ | x5 y i | 'f (5.134b)
2 i 2
volume d 15 u 10 6 (5.134d)
bL d b d bU (5.134e)
306 Analytic Methods for Design Practice
3 Amplitude
2.5
1.5
0.5 f (Hz)
0
9 15 21 27 33 39 45 51 57
Figure 5.31. Harmonic response of the optimum results: minimization of the area
The area is calculated by the trapezoid rule for numerical integration. The
area of the initial design is 17.9062. The frequency domain is discretized
by 1 Hz. The optimization results are shown in Table 5.7 and the harmonic
response of the optimum is illustrated in Figure 5.31.
References
Afimiwala KA, Mayne RW (1974) Optimal Design of an Impact Absorber.
Journal of Engineering for Industry, Transactions of the ASME 96:124í130
Barthelemy JFM, Haftka RT (1993) Approximation Concepts for Optimum
Structural Design–A Review. Structural Optimization 5:129í144
Bennett JA, Park GJ (1995) Automotive Occupant Dynamics Optimization. Shock
and Vibration 2:471í479
Chahande AI, Arora JS (1993) Development of a Multiplier Method for Dynamic
Response Optimization Problems. Structural Optimization 6:69í78
Choi WS, Park GJ (2002) Structural Optimization Using Equivalent Static Loads
at All the Time Intervals. Computer Methods in Applied Mechanics and
Engineering 191:2077í2094
Craig KJ, Stander N, Balasubramanyam S (2003) Worst-Case Design in Head
Impact Crashworthiness Optimization. International Journal for Numerical
Methods in Engineering 57:795í817
Etman LFP, Van CDH (1998) Design Optimization of Multibody Systems by
Sequential Approximation. Multibody System Dynamics 2:393í415
Dynamic Response Optimization 307
Feng TT, Arora JS, Haug JR (1977) Optimal Structural Design under Dynamic
Loads. International Journal for Numerical Methods in Engineering 11:39í52
Fox RL, Kapoor MP (1970) Structural Optimization in the Dynamic Regime: A
Computational Approach. AIAA Journal 8:1798í1804
Grandhi RV, Haftka RT, Watson LT (1986) Design-oriented Identification of
Critical Times in Transient Response. AIAA Journal 24:649í656
Greene WH, Haftka RT (1991) Computational Aspects of Sensitivity Calculations
in Linear Transient Structural Analysis. Structural Optimization 3:176í201
Gu YX, Chen BS, Zhang HW, Grandhi RV (2002) A Sensitivity Analysis Method
for Linear and Nonlinear Transient Heat Conduction with Precise Time
Integration. Structural and Multidisciplinary Optimization 24:23í37
Haftka RT, GĂrdal Z (1992) Elements of Structural Optimization. Kluwer,
Dordrecht
Hong UP, Park GJ (2003) Determination of Crash and Optimization of Crash
Components Using Response Surface Approximate Approximation. Journal of
Automobile Engineering 217:203í214
Kang BS, Choi WS, Park GJ (2001) Structural Optimization under Equivalent
Static Loads Transformed from Dynamic Loads Based on Displacement.
Computers and Structures 79:145í154
Kang BS, Park GJ, Arora JS (2005) Optimization of Flexible Multibody Dynamic
Systems Using the Equivalent Static Load. AIAA Journal 43:846í852
Kang BS, Park GJ, Arora JS (2006) A Review of Optimization of Structures
Subjected to Transient Loads. Structural and Multidisciplinary Optimization 31:
81í95
Kapania RK, Park S (1996) Nonlinear Transient Response and Its Sensitivity
Using Finite Elements in Time. Computational Mechanics 17:306í317
Kim MS, Choi DH (1997) Multibody Dynamic Response Optimization with ALM
and Approximation Line Search. Multibody System Dynamics 1:47í64
Kim MS, Choi DH (2000) A New Penalty Parameter Update Rule in the
Augmented Lagrange Multiplier Method for Dynamic Response Optimization.
KSME International Journal 14:1122í1130
Kurtaran H, Eskandarian A (2001) Design Optimization of Multibody Systems
under Impact Loading by Response Surface Methodology. Proceedings of the
Institution of Mechanical Engineers, Part K: Journal of Multibody Dynamics
215:173í185
Kurtaran H, Eskandarian A, Marzougui D, Bedewi NE (2002) Crashworthiness
Design Optimization Using Successive Response Surface Approximations.
Computational Mechanics 29:409í421
Logan DL (2002) A First Course in the Finite Element Method. Brooks/Cole,
Pacific Grove, CA
Marklund PO, Nilsson L (2001) Optimization of a Car Body Component Subjected
to Side Impact. Structural and Multidisciplinary Optimization 21:383í392
Moore GJ (1994) MSC/Nastran Design Sensitivity and Optimization. The
MacNeal–Schwendler Corporation, Los Angeles, CA
Myers RH, Montgomery DC (1995) Response Surface Methodology. Wiley, New
York
308 Analytic Methods for Design Practice
6.1 Introduction
Experiments are performed to investigate characteristics of a system. Responses or
characteristics are the results evaluated from experiments. The responses should
be analyzed appropriately according to the objectives of the experiments. The
factors are the sources that affect the experiments. It may not be easy to identify
all the factors and give right values because there may be many factors and they
may have distributions and be affected by the variations of the environment.
Generally, the objectives of experiments are as follows:
(1) To measure or evaluate characteristic values without analyzing them
statistically.
(2) Identification through experiment of factors that are significant to the
responses and determination of how large their impact is. The results of
experiments are analyzed statistically.
(3) Statistical identification of factors with little influence.
(4) Determination of the values of significant factors. Then, the optimum
condition can be found.
Design of Experiments (DOE) determines the allocation and method of
experiments to satisfy the objectives. Various factors are determined by analysis
of the experimental results. Generally, DOE proceeds as illustrated in Figure 6.1.
As mentioned earlier, DOE has been utilized for efficient experiments and analysis
of the results. Recently, this method has been applied to design. In design, the
fourth method is adopted to determine optimum values. The characteristic value
and factors of DOE are equivalent to the objective function and design variables of
the design, respectively. Many errors are involved in the experiments. When the
DOE method is utilized in design, the functions are usually evaluated numerically.
The numerical method does not have such errors and is a deterministic method.
Design using DOE can be regarded as a method using metamodels. The
metamodel is the model of a model. A function can be approximated by some
specific function values. The approximated function is a metamodel. Suppose we
have a function y with design variable vector b as follows:
310 Analytic Methods for Design Practice
Allocation of experiments
Randomization of the order of the experiments
Operation of experiments
Analysis of results
y f (b) (6.1)
yi f (b i ), i 1,..., k (6.2)
yˆ g (b) (6.3)
In the real design process, the metamodel ŷ is used. When the error in the
approximation process is H , the relation between y and ŷ is
Design of Experiments 311
y yˆ H (6.4)
In a design using DOE, well developed DOE methods are employed to define
ŷ , and the design variable vector b is obtained. The design problem in this case
can also be expressed as the optimization in Equation 3.1 Generally, a design
using DOE is not mathematically complicated; however, somewhat inferior
solutions are obtained compared to optimization. Two methods are usually used:
(1) the orthogonal arrays method and (2) the response surface method, which uses
the metamodel of explicit functions for approximation.
Terminologies Terminologies
Description
in DOE in design
One of the simplest DOE theories is the one-way factorial design. It is a method
for allocation of experiments when the number of factors is one and the number of
levels is more than one. For each level, experiments are repeatedly performed.
The number of repetitions (observations) does not need to be the same for each
level. In this section, the case when the number of repetitions is the same is
explained. The data for the one-way factorial design are shown in Table 6.2 for
the case when the number of levels for factor A is l ( A1 , A2 , ..., Al ) and the number
of repetitions is m. In Table 6.2
m
Ti. ¦ x ij , x i. Ti. / m , i 1, 2 , ... , l (6.5a)
j 1
l
T ¦ Ti. , x T / lm (6.5b)
i 1
where xij is the characteristic value (result of experiment) of the jth experiment
when A is Ai . The population exists for each level, the mean is P i and the
variance of the error is V E 2 . Then the mean of the entire population P is
l
P ¦ Pi / l (6.6)
i 1
where the error eij has a normal distribution N (0, V E 2 ) and is independent.
l
¦ ai 0 should be satisfied because it is an additive model. a i is the main effect
i 1
of A at the ith level. In Equation 6.7, the effect of each level is independently
added. Assumption of the additive model is exploited in DOE.
The analysis of variance (ANOVA) is explained for the one-way factorial
design. ANOVA is a frequently used for statistical analysis. The distribution of a
characteristic function is expressed by the sum of squares (This is called
“variance.”). The sum of squares is decomposed into the sum of squares for each
source and the significant sources are identified. The source means the source of
variation and the variation is the sum of squares of a characteristic function. For
instance, factors, errors and interactions can be sources.
From the data in Table 6.2, the deviation between xij and x is decomposed as
Design of Experiments 313
x11 x 21 x31 … x l1
x12 x 22 x32 … xl 2
Repetition of x13 x 23 x33 xl 3
…
experiments
x1m x 2m x3 m … xlm
( x ij x ) ( x ij x i. ) ( x i. x ) (6.8)
l m
2 2 2
¦ ¦ ( x ij x ) ¦ ¦ ( x ij x i. ) ¦ ¦ ( x i. x ) 2¦ ¦ ( x ij x i. )( x i. x )
i 1j 1 i j i j i j
(6.9)
The third term of the right hand side of Equation 6.9 is zero. Thus, Equation 6.9
becomes
2 2 2
¦ ¦ ( xij x ) ¦ ¦ ( x ij x i. ) ¦ ¦ ( x i. x ) (6.10)
i j i j i j
Therefore, the total sum of squares S T is decomposed into the sum of squares S A
due to A and the sum of squares S E due to errors as follows:
ST S A SE (6.11a)
l m 2
2
ST ¦ ¦ ( xij x ) ¦ ¦ xij CT (6.11b)
i 1j 1 i j
314 Analytic Methods for Design Practice
Ti. 2
SA ¦ ¦ ( xi . x )
2
m ¦ ( xi . x ) 2 m¦ xi. 2 CT ¦ CT (6.11c)
i j i i i m
2 2 2
SE ¦ ¦ ( xij xi. ) ¦ ¦ xij m ¦ xi.
i j i j i
§ · § T 2 ·
¨¨ ¦ ¦ xij 2 CT ¸¸ ¨ ¦ i. CT ¸ ST S A (6.11d)
©i j ¨
¹ ©i m ¸
¹
T2
where CT is the correction term.
lm
Next, the degree of freedom (DOF) is derived. Subtraction of the number of
constraints from the number of deviations for squares yields the DOF of the sum of
squares. Thus, each DOF is
IT lm 1 (6.12a)
IA l 1 (6.12b)
IE lm l l (m 1) (6.12c)
and
IT I A IE (6.13)
The ANOVA table shows the above results in a compact form. The mean
square V is defined as follows:
l SA VA
A SA m ¦ ( x i. x ) 2 IA l 1 VA F0
i 1 IA VE
SE
E SE ST S A IE l (m 1) VE
IE
l m
2
T ST ¦ ¦ ( x ij x ) IT lm 1
i 1j 1
Design of Experiments 315
S
V (6.14)
I
where S is the sum of squares and I is the degree of freedom. The ANOVA table
for the one-way factorial design is shown in Table 6.3.
From the ANOVA table, the population mean at each level P i P ai is
estimated. The population mean is the characteristic value at each level. m data of
the ith level are regarded as random samples with a normal distribution
N ( Pi , V E 2 ). The unbiased estimate for V E 2 is VE . Using Equations
6.A.20í6.A.22 of Appendix 6.A, the estimate of P i with 100(1 D )% confidence
interval is as follows:
§ D· V
xi. r t ¨ I E ; ¸ E (6.15)
© 2¹ m
SE
where V E and the degree of freedom I E for V E is l (m 1) . The t-
IE
distribution in Equation 6.15 is briefly introduced in Appendix 6.A (Walpole
SA
1982). If V A , F0 is defined to compare the mean squares of the variations,
IA
VA
F0 (6.16)
VE
The value in Equation 6.16 is used in an F test and it is the relative ratio of the
sum of squares due to a factor with respect to the sum of squares due to error.
When we have multiple factors, the factors with low F values can be ignored.
E S E =3.3214 IE = 6 V E = 0.5536
T S T =7.4078 IT = 8
Also, the F test is used for statistical hypothesis to accept or reject a hypothesis.
However, since it is rarely used in design, it is not explained here.
3 3
2
ST ¦ ¦ ( xij x )
i 1j 1
7.4078 (6.17a)
3
SA m ¦ ( x i. x ) 2
i 1
4.0992 (6.17b)
3 3
2
SE ¦ ¦ ( xij xi. )
i 1j 1
3.3214 (6.17c)
The above results can be obtained by using CT in Equation 6.11. The ANOVA
table is shown in Table 6.5.
Design of Experiments 317
Solution
Equation 6.15 is utilized for the calculation of the confidence interval. From the
95% confidence interval, D = 0.05. From Tables 6.4 and 6.5, m 3, x1. 7.41,
x2. 8.41, x3. 9.05, V E 0.5536 and I E 6. Substituting these values into
Equation 6.15, the confidence interval is
VE 0.5536
t (I E ; 0.025) t (6; 0.025) (6.18)
m 3
where eij has a normal distribution N (0, V E 2 ). xij is the measured value for Ai
and B j , P is the population mean, ai is the main effect of Ai , b j is the main
effect of B j and eij is the measuring error for Ai and B j . The interaction cannot
be measured from the two-way factorial design without repetitions. In other words,
the interaction effects are confounded with the error eij . The interaction will be
explained later.
318 Analytic Methods for Design Practice
Factor A
A1 A2 … Al Total Mean
Factor B
B1 x11 x21 … xl1 T.1 x.1
Mean x1. x 2. … xl . x
m
Ti. ¦ xij , xi. Ti. / m, i 1,..., l (6.21a)
j 1
l
T. j ¦ xij , x. j T. j / l , j 1,..., m (6.21b)
i 1
l m
2 2 2
¦ ¦ ( x ij x ) ¦ ¦ ( x i . x ) ¦ ¦ ( x. j x )
i 1j 1 i j i j
¦ ¦ ( x ij x i. x. j x ) 2 (6.22b)
i j
l m 2
2
ST ¦ ¦ ( x ij x ) ¦ ¦ x ij CT (6.22c)
i 1j 1 i j
2 Ti. 2
SA ¦ ¦ ( x i. x ) m ¦ ( x i. x ) 2 ¦ CT (6.22d)
i j i i m
2
T. j 2
SB ¦ ¦ ( x. j x ) l ¦ ( x. j x ) 2 ¦ CT (6.22e)
i j j j l
Design of Experiments 319
Table 6.7. ANOVA table for a two-way factorial design without repetitions
Factor S I V E (V ) F0 F (D )
VA
A SA l 1 VA V E 2 mV A 2 F (I A , I E ;D )
VE
VB
B SB m 1 VB V E 2 lV B 2 F (I B , I E ;D )
VE
E SE (l 1)(m 1) VE V E2
T ST lm 1
2 2
SE ¦ ¦ ( xij xi. x. j x ) ¦ ¦ {( xij x ) ( xi. x ) ( x. j x )}
i j i j
2 2 2
¦ ¦ ( xij x ) ¦ ¦ ( xi. x ) ¦ ¦ ( x. j x ) S T S A S B (6.22f)
i j i j i j
§ D· V
xi . r t ¨ I E ; ¸ E (6.23a)
© 2¹ m
§ D· V
x. j r t ¨ I E ; ¸ E (6.23b)
© 2¹ l
§ D· V
( xi. x. j x ) r t ¨ I E ; ¸ E (6.23c)
© 2 ¹ ne
lm
respectively. ne is the effective number of repetitions and ne . The
l m 1
effective number of repetitions is the number of repetitions to which a specific
factor contributes at a specific level. For example, m is the number of repetitions
in Equation 6.15. In a two-way factorial design, when a factor is repeated, the
levels of other factors are changed. Therefore, the number of repetitions of a
factor does not directly contribute to the errors. The effective number of
repetitions includes this phenomenon. The optimum levels for the population
320 Analytic Methods for Design Practice
K K B3 K
B3
B3 B2 B2
B2 B1 B1
B1
A1 A2 A3 A1 A2 A3 A1 A2 A3
Figure 6.2. Interaction: (a) No interaction, (b) synergistic interaction, (c) antisynergistic
interaction
mean are the levels for the optimum ( xi. x. j x ) . This is further explained in
(Montgomery 1991).
When experiments are performed with two factors, experimental results can be
measured repeatedly for specific levels. This is called a “two-way factorial design
with repetitions” and the cost increases due to the increments of the number of
experiments. However, the interactions can be measured. When factor A has l
levels, factor B has m levels and the number of repetitions is r, the additive
statistical model is
where eijk has a normal distribution N (0, V E 2 ) . (ab) ij represents the interaction
effect of factors A and B in the ith and the jth levels, respectively. The interaction
is the effect due to a combination of factors. Figure 6.2a does not have
interactions. Figures 6.2b and 6.2c show interaction effects. Since Figure 6.2b
shows the synergistic interaction, it can be ignored. However, the effect of the
antisynergistic interaction as illustrated in Figure 6.2c is significant and the
interactions should be considered in the design process.
Data from a two-way factorial design with repetitions are shown in Table 6.8.
The following equations are obtained:
m r
Ti.. ¦ ¦ xijk , xi.. Ti.. / mr , i 1,..., l (6.25a)
j 1k 1
l r
T. j. ¦ ¦ xijk , x. j. T. j. / lr , j 1,..., m (6.25b)
i 1k 1
l m
T ¦ Ti.. ¦ T. j. , x T / lmr (6.25c)
i 1 j 1
Design of Experiments 321
Factor A
A1 A2 … Al Total Mean
Factor B
x111 x 211 xl11
x112 x 212 xl12
B1 … T.1. x.1.
x11r x 21r xl1r
x121 x 221 xl 21
x122 x 222 xl 22
B2 … T.2. x.2.
x12 r x 22 r xl 2 r
x1m1 x 2 m1 xlm1
x1m 2 x 2m 2 xlm 2
Bm … T.m. x.m.
x1mr x 2 mr xlmr
Total T1.. T2.. … Tl .. T
Mean x1.. x 2.. … xl .. x
r
Tij. ¦ xijk (6.25d)
k 1
Tij.
xij. (6.25e)
r
r
Tij. ¦ xijk (6.26)
k 1
l m r
2 2 2
¦ ¦ ¦ ( xijk x ) ¦ ¦ ¦ ( xi.. x ) ¦ ¦ ¦ ( x. j. x )
i 1 j 1k 1 i j k i j k
Each term of Equation 6.28 is the total sum of squares ( ST ) , the sum of squares
due to A ( S A ) , the sum of squares due to B ( S B ) , the sum of squares due to the
interaction between A and B ( S AuB ) and the sum of squares due to error ( S E ).
Actually, it is convenient to use Equation 6.29 instead of Equation 6.28,
2
ST ¦ ¦ ¦ xijk CT (6.29a)
i j k
Ti..2
SA ¦ CT (6.29b)
i mr
T. j.2
SB ¦ CT (6.29c)
j lr
S AuB S AB S A S B (6.29d)
SE ST S AB (6.29e)
Tij .2
S AB ¦¦ CT (6.29f)
i j r
T2
where CT . The degrees of freedom for above sums of squares are IT
lmr
lmr 1, I AB lm 1, I A l 1, I B m 1, I AuB (l 1)(m 1) and I E lm
(r 1), respectively. The ANOVA table is shown in Table 6.9.
The (1 D ) confidence intervals for the combination of Ai B j are
§ D· V
xi.. r t ¨ I E ; ¸ E (6.30a)
© 2 ¹ mr
§ D· V
x. j. r t ¨ I E ; ¸ E (6.30b)
© 2 ¹ lr
§ D· V
xij . r t ¨ I E ; ¸ E (6.30c)
© 2¹ r
Design of Experiments 323
Table 6.9. ANOVA table for a two-way factorial design with repetitions
Factor S I V E (V ) F0 F (D )
VA
A SA l 1 VA V E 2 mrV A 2 F (I A , I E ; D )
VE
VB
B SB m 1 VB V E 2 lrV B 2 F (I B , I E ; D )
VE
V Au B
Au B S Au B (l 1)(m 1) V Au B V E 2 rV AuB 2 F (I Au B , I E ; D )
VE
E SE lm(r 1) VE V E2
T ST lmr 1
respectively. The optimum values are the levels that make the optimum of xij . .
The design of experiments (DOE) is explained briefly. As mentioned earlier,
experiments are appropriately allocated in DOE. Through statistical analysis,
values are estimated or optimum values are derived. When there are more than
two factors, a multiway factorial design must be utilized. Since the number of
experiments is considerably increased, it is rarely used. Moreover, there are many
other statistical theories for estimation, testing, error analysis, etc. However,
because those theories are not generally used for design, they are not explained
here.
T2
CT 630.38 (6.31a)
lmr
2
ST ¦ ¦ ¦ xijk CT 787 630.38 156.62 (6.31b)
i j k
Ti..2
SA ¦ CT 33.25 (6.31c)
i mr
324 Analytic Methods for Design Practice
Factor A
A1 A2 A3 T. j. x. j .
Factor B
2 3 4 18 3
B1
1 5 3
3 7 6 28 4.67
B2
4 3 5
6 10 7 48 8
B3
8 12 5
4 8 6 29 4.83
B4
2 5 4
Ti.. 30 53 40 123
T. j.2
SB ¦ CT 78.45 (6.31d)
j lr
Tij.2
S AB ¦¦ CT 130.12 (6.31e)
i j r
SE ST S AB 26.5 (6.31g)
Square
Factor DOF I Mean square V F
sum S
A 33.25 IA l 1 2 16.63 V A VE 7.52
Table 6.12. The characteristic function for each factor of Example 6.3
Factor A A1 A2 A3
Factor B Total Mean Total Mean Total Mean
B1 3 1.5 8 4 7 3.5
B2 7 3.5 10 5 11 5.5
B3 14 7 22 11 12 6
B4 6 3 13 6.5 10 5
§ D· V 2.21
x23. r t ¨ I E ; ¸ E 11 r t (12; 0.025)
© 2¹ r 2
11 r 2.179 u 1.051 11 r 2.29 (6.32)
t (12; 0.025) can be found in the table for t-distribution (Walpole 1982).
The aforementioned DOE is the factorial design. Although the factorial design
provides a lot of precise data, it is an expensive method due to the large number of
experiments required. The fractional design is utilized to save cost. In fractional
design, the metamodel for the factorial design is derived and subsequently the
metamodel is analyzed. Therefore, allocation of factors and the analysis of the
DOE results are quite important.
326 Analytic Methods for Design Practice
There are many methods in fractional design. In this section, the method using
orthogonal arrays is explained. Orthogonal arrays are generally utilized when
fractional design is directly applied to design. Other than this method, fractional
design is usually utilized to generate the design surfaces in the response surface
method. Fractional design is indirectly used in the response surface method. An
orthogonal array is an array that easily enables the design of experiments. When
there are many factors, each factor and the plausible interactions between factors
are allocated to columns of an orthogonal array. Since their introduction by Rao in
1946, orthogonal arrays have been widely used in many fields. They are mostly
used in the statistics area (Hedayat et al. 1999) where there are many theories
concerning them. However, aspects related to design are briefly discussed in this
chapter.
An orthogonal array is represented by a matrix. There are a few methods to
express them and we will use Taguchi’s method as follows (Taguchi 1987):
n
LN ( sik i ) (6.33)
i 1
where N is the number of rows for an orthogonal array, n is the number of factors
with a specific level, si is the number of levels, ki is the number of factors with
si levels. For example, L18 (2137 ) is an orthogonal array having 18 rows, one
factor with 2 levels and 7 factors with 3 levels.
The orthogonal array L9 (3 4 ) is shown in Table 6.13. It enables an experiment
of up to four factors with three levels. In a factorial design, 34 81 times the
experiments are required. Only nine times the experiments are required with the
orthogonal array. At each column, each level appears at the same time. Each
column is a vector with nine components. Substitute the components 1, 2 and 3 by
í1, 0 and 1. If we make inner products with two columns, they are zeros. In other
words, the vectors are orthogonal to each other. This orthogonality is the basis of
its name. Various orthogonal arrays are shown in Appendix 6.B.
Solution
Let each column of Table 6.13 n i and substitute 1, 2 and 3 in Table 6.13 by í1, 0
and 1 as follows:
n1 [ 1 1 1 0 0 0 1 1 1]T (6.34a)
n2 [ 1 0 1 1 0 1 1 0 1]T (6.34b)
n3 [ 1 0 1 0 1 1 1 1 0]T (6.34c)
Design of Experiments 327
n4 [ 1 0 1 1 1 0 0 1 1]T (6.34d)
Level
Factor
1 2 3
A A1 A2 A3
B B1 B2 B3
C C1 C2 C3
D D1 D2 D3
Experiment
K1 K2 K3 K4 K5 K6 K7 K8 K9
number
Characteristic
20 50 30 25 45 30 45 65 70
function
can be made from the results of the orthogonal array. This will be shown in
Example 6.8.
From Table 6.15, the mean of the characteristic function is
1 9
m ¦ Ki (6.37)
9i 1
The effect by the levels of each factor is investigated. m A3 is the mean when
factor A has A3 as follows:
1
m A3 (K 7 K8 K9 ) (6.38)
3
1
m A1 (K1 K2 K3 ) (6.39a)
3
1
m A2 (K 4 K5 K 6 ) (6.39b)
3
Design of Experiments 329
Ș
80
70
60
50
40
30
20
10
A1 A2 A3 B1 B2 B3 C1 C2 C3 D1 D2 D3
Figure 6.3. Graphs for the results of each factor in Table 6.15
The means for factors B, C and D are obtained by the mean with specific levels
as well. Results derived from Table 6.15 are illustrated in Figure 6.3. The levels
minimizing K are considered optimum. Factor A has the same results for the first
and second levels. Therefore, the optimum levels are A1 B1C1 D3 or A2 B1C1 D3 .
This process is called the “analysis of means (ANOM).” Instead of Figure 6.3, the
one-way table in Table 6.16 can be utilized. This will be explained later.
The design process from ANOM is performed based on the assumption that the
statistical model for K satisfies additivity as follows:
K ( Ai , B j , Ck , Dl ) P ai b j c k d l e (6.40)
where P is the population mean. ai , b j , ck and d l are the deviations from the
population mean when factors A, B, C and D have the i, j, k and lth level,
respectively. Error e includes the errors from the experiments and the additive
model. From Equation 6.40, the following equations are derived:
a1 a 2 a3 0 (6.41a)
b1 b2 b3 0 (6.41b)
c1 c2 c3 0 (6.41c)
d1 d 2 d 3 0 (6.41d)
Level
Factor
1 2 3
K1 K 2 K 3 K 4 K5 K 6 K 7 K8 K 9
A m A1 m A2 m A3
3 3 3
K1 K 4 K 7 K 2 K 5 K8 K3 K6 K9
B m B1 m B2 m B3
3 3 3
K1 K 6 K8 K 2 K 4 K9 K3 K5 K7
C mC1 mC2 mC3
3 3 3
K1 K 5 K 9 K2 K6 K7 K 3 K 4 K8
D m D1 m D2 m D3
3 3 3
1
m A3 (K 7 K8 K 9 )
3
1
[( P a3 b1 c3 d 2 e7 ) ( P a3 b2 c1 d 3 e8 )
3
( P a 3 b3 c 2 d1 e9 )]
1 1 1
(3P 3a3 ) (b1 b2 b3 ) (c1 c 2 c3 )
3 3 3
1 1
(d1 d 2 d 3 ) (e7 e8 e9 )
3 3
1
( P a3 ) (e7 e8 e9 ) (6.42)
3
pred
K opt m (m A1 m) (mB1 m) (mC 2 m) (mD2 m) (6.43)
When a term in Equation 6.43 is not significant, it can be pooled to the error
term. For example, if the effects of C2 and D2 are negligible, the prediction can
be made by Equation 6.44,
pred
K opt m (m A1 m) (mB1 m) (6.44)
The decision can be made from the ANOVA table introduced in Example 6.8.
When the optimum level is different from any rows of the orthogonal array and
the characteristic function is estimated by Equation 6.43, a confirmation
experiment with the optimum levels should be performed. If the results of the
confirmation experiment are similar to the estimated value, the process is correct.
On the other hand, if the difference is large the additive model is not appropriate.
That is, the interaction effect compounded to the error is considerably large. It is
not easy to consider the interactions; this will be explained later. Selection of an
orthogonal array is quite important. Example 6.6 illustrates the method.
Mean 1
A 2 (levels) – 1 = 1
Au B (2 – 1) u (3 – 1) = 2
Total 14
332 Analytic Methods for Design Practice
the smallest one is selected. When the second smallest one is utilized, we may
obtain better results. On the other hand, then the number of experiments is
increased.
Solution
(1) The function consists of perfect square terms. Thus, the optimum solution
is A = 0, B = 1.5 and C = 2.5.
(2) The factorial design is shown in Table 6.18 where function f is the smallest
when the design point is A1B3C2 . The values of the design point are (í0.2,
2.0, 2.8) and the function value is f = 0.38.
(3) The function value of Equation 6.45 is calculated according to the rows of
the orthogonal array in Table 6.13. They are
In Table 6.16, K is replaced by f and the one-way table in Table 6.19 is set up.
Level
Factor
1 2 3
A –0.2 0.6 1.4
B –2.0 0 2.0
Exp.
A B C f
no.
1 –0.2 –2.0 2.0 12.54
2 –0.2 –2.0 2.8 12.38
3 –0.2 –2.0 3.6 13.5
4 –0.2 0 2.0 2.54
5 –0.2 0 2.8 2.38
6 –0.2 0 3.6 3.5
7 –0.2 2.0 2.0 0.54
8 –0.2 2.0 2.8 0.38
9 –0.2 2.0 3.6 1.5
10 0.6 –2.0 2.0 12.86
11 0.6 –2.0 2.8 12.7
12 0.6 –2.0 3.6 13.82
13 0.6 0 2.0 2.86
14 0.6 0 2.8 2.7
15 0.6 0 3.6 3.82
16 0.6 2.0 2.0 0.86
17 0.6 2.0 2.8 0.7
18 0.6 2.0 3.6 1.82
19 1.4 –2.0 2.0 14.46
20 1.4 –2.0 2.8 14.3
21 1.4 –2.0 3.6 15.42
22 1.4 0 2.0 4.46
23 1.4 0 2.8 4.3
24 1.4 0 3.6 5.42
25 1.4 2.0 2.0 2.46
26 1.4 2.0 2.8 2.3
27 1.4 2.0 3.6 3.42
In each row of Table 6.19, the level for each factor that makes the objective
function the smallest is chosen. Therefore, the optimum levels are A1 B3C 2 and
their values are (í0.2, 2.0, 2.8). This is the same as the case of the factorial design
because there are no interactions.
Using Equation 6.43, the function is estimated at the optimum point. The mean
m is
334 Analytic Methods for Design Practice
Design Level
variable 1 2 3
9
¦ fi
i 1
m 6.22 (6.46)
9
Substituting the values in Table 6.19 into Equation 6.43, the estimated value is
pred
f opt 6.22 (5.47 6.22) (1.55 6.22) (5.79 6.22) 0.37 (6.47)
The sum of squares due to factor A is made from the first column of the
orthogonal array in Table 6.13.
3Ti..2
SA ¦ CT
i 1 3
1
(T1.. 2 T2.. 2 T3.. 2 ) CT
3
1
[( y111 y122 y133 ) 2 ( y 212 y 223 y 231 ) 2
3
1
[(12.54 2.38 1.5) 2 (12.7 3.82 0.86) 2
3
where Ti.. is the sum of f for the ith level ( Ai ) of factor A. In the same manner,
3T.i.2
SB ¦ CT
i 1 3
1
(T.1. 2 T.2. 2 T.3. 2 ) CT
3
1
[( y111 y 212 y313 ) 2 ( y122 y 223 y321 ) 2
3
248 (6.49b)
3 T..i 2
SC ¦ CT
i 1 3
1
(T..12 T..2 2 T..3 2 ) CT
3
1
[( y111 y 231 y321 ) 2 ( y122 y 212 y332 ) 2
3
2.20 (6.49c)
The sum of squares due to error is calculated from column D of Table 6.13,
336 Analytic Methods for Design Practice
Factor S I V
A 6.34 2 3.17
B 248 2 124
C 2.20 2 1.10
Error (E ) 0 2
T 256.54 8
1
SE [( y111 y 223 y332 ) 2 ( y122 y 231 y313 ) 2
3
ST S A S B SC S E 256.54 (6.50)
In another way,
2
ST ¦ ¦ ¦ y ijk CT 604.74 348.20 256.54 (6.51)
i j k
Find b Rn (6.52a)
where the number of design variables n = 4 and the number of levels for each
design variable is three. The DOE notation is somewhat different from the design
notation. For instance, the objective function f in design is equivalent to the
Design of Experiments 337
Level
Factor
1 2
A – 0.5 0.3
B – 0.6 0.4
C – 0.9 1.0
D – 0.5 0.2
Solution
(1) The levels of each factor in Table 6.21 are substituted into the L8 (2 7 )
Factor
Exp. no.
1 2 3 4 5 6 7
1 1 1 1 1 1 1 1
2 1 1 1 2 2 2 2
3 1 2 2 1 1 2 2
4 1 2 2 2 2 1 1
5 2 1 2 1 2 1 2
6 2 1 2 2 1 2 1
7 2 2 1 1 2 2 1
8 2 2 1 2 1 1 2
Interaction A B Au B C Au C D e
Design of Experiments 339
f1 f2 f3 f4 f5 f6 f7 f8
f1 f 2 f 5 f 6 f3 f 4 f 7 f8
A 2.24 1.76
4 4
f1 f 2 f 7 f 8 f3 f 4 f5 f6
B 2.2 1.8
4 4
f1 f 3 f 5 f 7 f 2 f 4 f6 f8
C 2.29 1.72
4 4
f1 f 3 f 6 f 8 f 2 f 4 f5 f7
D 2.32 1.69
4 4
Exp. no. A B C D f
1 – 0.5 (1) – 0.6 (1) – 0.9 (1) – 0.5 (1) 4.96
2 – 0.5 (1) – 0.6 (1) – 0.9 (1) 0.2 (2) 4.33
3 – 0.5 (1) – 0.6 (1) 1.0 (2) – 0.5 (1) 1.35
4 – 0.5 (1) – 0.6 (1) 1.0 (2) 0.2 (2) 0.72
5 – 0.5 (1) 0.4 (2) – 0.9 (1) – 0.5 (1) 3.76
6 – 0.5 (1) 0.4 (2) – 0.9 (1) 0.2 (2) 3.13
7 – 0.5 (1) 0.4 (2) 1.0 (2) – 0.5 (1) 0.15
8 – 0.5 (1) 0.4 (2) 1.0 (2) 0.2 (2) -0.48
9 0.3 (2) – 0.6 (1) – 0.9 (1) – 0.5 (1) 0.64
10 0.3 (2) – 0.6 (1) – 0.9 (1) 0.2 (2) 0.01
11 0.3 (2) – 0.6 (1) 1.0 (2) – 0.5 (1) 3.11
12 0.3 (2) – 0.6 (1) 1.0 (2) 0.2 (2) 2.48
13 0.3 (2) 0.4 (2) – 0.9 (1) – 0.5 (1) 1.04
14 0.3 (2) 0.4 (2) – 0.9 (1) 0.2 (2) 0.41
15 0.3 (2) 0.4 (2) 1.0 (2) – 0.5 (1) 3.51
16 0.3 (2) 0.4 (2) 1.0 (2) 0.2 (2) 2.88
If one of the above three conditions is satisfied, the effect of interaction is large.
However, interactions can still exist even though all the conditions are not satisfied.
Therefore, the conditions are sufficient, but not necessary. Although the existence
of interactions can be verified, the factors interacting with each other should be
identified through the process shown in Figure 6.2. The identification of the
interactions is a complicated and expensive process. Therefore, if we use the
algorithm of the previous section, a good solution can be found in an engineering
sense, although it is not the best one.
As explained above, it is not easy to identify the interactions but they should be
identified before DOE is carried out. What if the interactions are identified
beforehand? In this case, they should be allocated to a column of the orthogonal
array. For example, we have four factors of two levels (A, B, C and D) and the
interaction between A and B. If we use the L8 (2 7 ) orthogonal array in Table 6.22,
A, B are allocated to the first two columns, respectively. The interaction A u B is
allocated to the third column. Moreover, C and D are allocated to the fourth and
fifth columns, respectively. No factors are allocated to the other columns. Details
about allocation are given in a reference (Phadke 1989).
Column no. 1 2 3 4 5 6 7
Factor e A B C D e e
Level 1 2 1 2 1 2 1 2 1 2 1 2 1 2
4.96 3.13 4.96 1.04 4.96 1.04 4.96 0.72 4.96 0.72 4.96 0.72 4.96 0.72
0.72 0.15 0.72 2.88 0.72 2.88 1.04 2.88 1.04 2.88 2.88 1.04 2.88 1.04
1.04 0.01 3.13 0.01 0.01 3.13 3.13 0.15 0.15 3.13 3.13 0.15 0.15 3.13
2.88 3.11 0.15 3.11 3.11 0.15 0.01 3.11 3.11 0.01 3.11 0.01 0.01 3.11
(1) Total 9.6 6.4 8.96 7.04 8.8 7.2 9.14 6.86 9.26 6.74 14.08 1.92 8 8
(2) Sum of
16 16 16 16 16 16 16
(1)
(3)
Difference – 3.2 – 1.92 – 1.6 – 2.28 – 2.52 – 12.16 0
of (1)
(4) (3)2 10.24 3.69 2.56 5.20 6.35 147.87 0
(5) (4)/8 1.28 0.46 0.32 0.65 0.79 18.48 0
pred
f opt m (m A2 m) (mB2 m) (mC 2 m) (mD2 m)
0.96 (6.54a)
exp
f opt 2.88 (6.54b)
Factor S I V F
A 0.46 1 0.46 0.07
B 0.32 1 0.32 0.05
C 0.65 1 0.65 0.1
D 0.79 1 0.79 0.12
Error 19.76 3 6.59
T 21.98 7
Design of Experiments 343
Figure 6.4 shows the graphs for this problem for the interactions. From Figure
6.4, we can identify that the interactions A u B and A u C are large.
f
C = – 0.9
5 D = – 0.5
B = 0.4
B = – 0.6
– 0.5 0.3 A
f
B = – 0.6
5 D = – 0.5
C = 1.0
C = –0.9
– 0.5 0.3 A
f
A = – 0.5
5 D = – 0.5
C = – 0.9
C = 1.0
– 0.6 0.4 B
Table 6.28. The objective function of Example 6.11 for each row of the L8 (2 7 ) orthogonal
array
f1 f2 f3 f4 f5 f6 f7 f8
4.96 0.72 3.13 0.15 0.64 2.48 0.41 3.51
Column no. 1 2 3 4 5 6 7
Factor A B A uB C A uC D e
Level 1 2 1 2 1 2 1 2 1 2 1 2 1 2
4.96 0.64 4.96 3.13 4.96 3.13 4.96 0.72 4.96 0.72 4.96 0.72 4.96 0.72
0.72 2.48 0.72 0.15 0.72 0.15 3.13 0.15 3.13 0.15 0.15 3.13 0.15 3.13
3.13 0.41 0.64 0.41 0.41 0.64 0.64 2.48 2.48 0.64 0.64 2.48 2.48 0.64
0.15 3.51 2.48 3.51 3.51 2.48 0.41 3.51 3.51 0.41 3.51 0.41 0.41 3.51
(1) Total 8.96 7.04 8.8 7.2 9.6 6.4 9.14 6.86 14.08 1.92 9.26 6.74 8 8
(2) Sum of
16 16 16 16 16 16 16
(1)
(3)
Difference – 1.92 – 1.6 – 3.2 – 2.28 – 12.16 – 2.52 0
of (1)
(4) (3)2 3.69 2.56 10.24 5.20 147.87 6.35 0
Factor S I V
A 0.46 1 0.46
B 0.32 1 0.32
Au B 1.28 1 1.28
C 0.65 1 0.65
Au C 18.48 1 18.48
D 0.79 1 0.79
Error 0 1
T 21.98 7
Table 6.31a. In this case, A2 B1 is the solution. The selection of A and C is shown
in Table 6.31b. The solution is A1C2 . The solution for D is the second level from
the one-way table in Table 6.31c, since there are no interactions due to D.
When a factor has interactions with the other two factors, two two-way tables
may give different solutions. Then a one-way table is made for A. From the one-
way table of Example 6.9, the solution of A is A2 . Therefore, the design solution
is A2 B1C2 D2 . In this problem, the interaction of A u C is much greater than the
interaction of A u B from Table 6.30. From Table 6.31b, A1C2 is determined as
the design solution. When A A1 , the good solution from Table 6.31a is A1B2 .
Thus, the final solution is A1B2C2 D2 .
The objective function value is estimated by using the solution determined
above. Using Equation 6.43, the estimated value is
pred
f opt m (m A1B2 m) (m A1C2 m) (m D2 m) 0.24 (6.55)
By using the additive model in Equation 6.43, the components of Equation 6.55
can be expressed as
mD2 m d2 (6.56c)
346 Analytic Methods for Design Practice
A1 A2
B1 f1 f 2 5.68 f5 f6 3.12
B2 f3 f4 3.28 f 7 f8 3.92
A1 A2
C1 f1 f 3 8.09 f5 f7 1.05
C2 f2 f4 0.87 f 6 f8 5.99
D1 D2
f1 f 4 f 5 f 8 9.26 f2 f3 f6 f7 6.74
pred
f opt m (m A1 B2 m) (m A1C 2 m) (mD2 m) a1 0.48 (6.57)
a1 in Equation 6.57 is calculated from (1) of the first column in Table 6.29 as
follows:
(8.96 7.04) / 4
a1 0.24 (6.58)
2
exp
The value of the confirmation experiment is f opt 0.48 . Since this problem
does not have errors, Equation 6.57 predicts the solution precisely.
9.144m
A
D
C 9.144m
4
Factor ( u 10 m2 )
Level
A B C D E
1 3.00 8.00 10.10 10.00 6.00
2 17.00 17.00 11.00 25.00 18.00
3 30.00 26.00 20.00 40.00 30.00
Maximize [ (6.59)
where [ is the natural frequency of the first mode. Factors (design variables) are
the areas of the cross sections. The discrete values of the factors are shown in
Table 6.32. Identify the interactions between the factors.
Solution
Since we have five factors with three levels, the orthogonal array L18 in Table
6.33 is chosen. The interactions are evenly distributed in L18 . Thus, it is
frequently chosen when DOE is performed regardless of the interactions. All the
factors are allocated to columns of L18 . The results are shown in Tables 6.33 and
6.34. The ANOVA table is shown in Table 6.34. The mean for the levels of each
factor is shown in Table 6.34 and it is the same as the solution from the one-way
exp
table. The solution is A3 B3C 2 D2 E3 from Table 6.34 and [ opt 42.806. The
methods mentioned before can be used to identify interactions. The mean square
is employed for this problem and VE 6.5563 30.8026 14.6699 52.0288.
The mean square is somewhat large. However, since the interactions are evenly
exp
distributed in L18 , the solution is considered as the final one. It is noted that [ opt
348 Analytic Methods for Design Practice
Table 6.33. Experiments of Example 6.12 using the L18 ( 21 u 3 7 ) orthogonal array
1 1 1 1 1 1 1 1 1 34.40398
2 1 1 2 2 2 2 2 2 34.80376
3 1 1 3 3 3 3 3 3 31.75859
4 1 2 1 1 2 2 3 3 38.07272
5 1 2 2 2 3 3 1 1 40.39151
6 1 2 3 3 1 1 2 2 29.83928
7 1 3 1 2 1 3 2 3 41.13371
8 1 3 2 3 2 1 3 1 32.97368
9 1 3 3 1 3 2 1 2 42.08462
10 2 1 1 3 3 2 2 1 27.82737
11 2 1 2 1 1 3 3 2 28.18135
12 2 1 3 2 2 1 1 3 39.06658
13 2 2 1 2 3 1 3 2 31.56198
14 2 2 2 3 1 2 1 3 34.98063
15 2 2 3 1 2 3 2 1 42.04833
16 2 3 1 3 2 3 1 2 36.60479
17 2 3 2 1 3 1 2 3 36.44593
18 2 3 3 2 1 2 3 1 37.88146
is better than any row in Table 6.33. Therefore, a good solution is obtained
without considering interactions, although they exist.
Design of Experiments 349
Table 6.35. Experiments with the L27(313) orthogonal array in Example 6.13
Table 6.36. The mean and the analysis of variance for Example 6.13
Level of factor A
A1 A2 A3
E1 36.4251 36.4902 34.7501
Level of
factorٻEٻ
E2 31.2400 37.2814 37.1633
E3 28.8474 38.2535 39.4374
Factor A B C D E [ (Hz)
Factorial 3 3 1 3 3 43.315
Solu-
L18 3 3 2 2 3 42.806
tion
L27 3 3 1 2 3 43.140
exp
is A3 B3C1 D2 E3 and [ opt 43.140. In Table 6.38, the results from L18 and L27
are compared with the results from the factorial experiments. In an engineering
sense, the two results from the orthogonal arrays are sufficiently excellent as
compared to the factorial solution. Although the interactions are ignored, a good
solution can be obtained by using the orthogonal L18 where the interactions are
evenly distributed.
Design of Experiments 351
where b is the design variable vector and f is the objective function. Using the
notation of Equation 3.1, P is the penalty function defined by the maximum
violation of the constraints,
Equation 6.61 is obtained by multiplication of the scale factor s and the maximum
violation. The influence of the constraints depends on s. It is recommended that
the constraints be normalized.
The characteristic function < in Equation 6.60 is equivalent to K in Section
6.3.3. Therefore, the process of Section 6.3.3 can be directly utilized with <.
Some constraints may represent the overall characteristics of a system. They are
usually functions of most design variables. It is noted that the interactions become
important in these cases.
The above method is used to find an optimum from ANOM of <. The
solution from an orthogonal array (1) may not be a good one when the interactions
are ignored or (2) may violate constraints. To overcome (1), after a confirmation
experiment is performed, the best one is picked from the rows of the orthogonal
array and the confirmation experiment. This method does not provide the best
solution. However, a good solution can be obtained with low cost. Therefore, the
method in Section 6.3.3 is utilized for a constrained problem. However, we still
have the possibility of a constraint violation. It should be noted that each row of
the orthogonal array can be evaluated with a different constraint because only the
maximum constraint violation is considered. Therefore, there is a strong
possibility that the solution of Section 6.3.3 violates the constraints.
To overcome this difficulty, two methods are proposed as follows:
(1) A method that directly uses an orthogonal array
The method in Section 6.3.3 is directly used and the constraint violation is
checked. First, the cases satisfying the constraints are picked from the
confirmation experiment and the rows of the orthogonal array. Second, the
one with the best < is picked as the final solution. In other words, K is
352 Analytic Methods for Design Practice
replaced by < in Table 6.13, <i is obtained for each row, the
confirmation experiment and the rows of the orthogonal array are
compared, and the best one is picked from those satisfying the constraints.
(2) A method to arrange the estimated values in ascending order
In Equation 6.43, the estimated value is obtained when the interactions are
ignored. The estimation is performed for the full combination. That is,
< pred is estimated for a full combination. These are approximated values
calculated from the matrix experiment with an orthogonal array. The
objective function and the penalty term for constraints are estimated,
respectively. < pred is derived by addition of the estimated objective
function and the penalty term. The full combination of < pred is arranged
in ascending order for the magnitude of < pred . An example with the
L9 (3 4 ) orthogonal array is shown in Table 6.39. The experiment number
in Table 6.39 is the ranking for the magnitude of < pred . That is,
<ipred <ipred
1 , i 1, 2, ... , 80. Confirmation experiments are performed
for the cases of Table 6.39. The first experiment that satisfies the
constraints is picked as a candidate design. In Table 6.39, the fifth one is
the candidate design. The candidate design is compared with all the rows
of the orthogonal array. The final design is the best one out of the
1 <1 Violated
2 <2 Violated
3 <3 Violated
4 <4 Violated
5 <5 Satisfied
The ith row for each factor is the level of
6 the factor <6 Violated
. . .
. . .
. . .
81=34 <81 .
Design of Experiments 353
Find b1 , b2 , b3 , b4 (6.62a)
b1 2 b2 2 b3 2 b4 2 b1 b2 b3 b4
subject to 1 d 0 (6.62c)
8
b1 2 2b2 2 b3 2 2b4 2 b1 b4
1 d 0 (6.62d)
10
2b1 2 b2 2 b3 2 2b1 b2 b4
1 d 0 (6.62e)
5
Solution
(1) Using Equation 6.60, < is defined. The scale factor of Equation 6.61 is
arbitrarily defined as 100. Since the problem has four design variables
with three levels, the L9 (34 ) orthogonal array is chosen. The experiments
are performed as shown in Table 6.40. The seventh row violates the
constraints and the other rows satisfy all the constraints. Based on the
results in Table 6.40 the one-way table for < is made as shown in Table
6.41. The design solution from Table 6.41 is (b1 ) 2 (b2 )3 (b3 )3 (b4 )1 . If we
substitute the design solution into Equation 6.62 for the confirmation
354 Analytic Methods for Design Practice
Table 6.40. Experiments with the L9 (3 4 ) orthogonal array for Example 6.14
1 1 1 1 1 29 ( <1 )
2 1 2 2 2 6 ( <2 )
3 1 3 3 3 –9 ( <3 )
4 2 1 2 3 14 ( <4 )
5 2 2 3 1 –25 ( <5 )
6 2 3 1 2 19 ( <6 )
7 3 1 3 2 23 ( <7 )
8 3 2 1 3 27 ( <8 )
9 3 3 2 1 –14 ( <9 )
Level
Design
variable
1 2 3
b1 8.67 2.67 12
b2 22 2.67 –1.33
b3 25 2 –3.67
b4 –3.33 16 10.67
Table 6.42. Arrangement of estimated values in ascending order for Example 6.14
1 2 3 3 1 – 29 Satisfied
2 3 3 3 2 – 27
3 2 2 3 1 – 25
4 1 3 3 1 – 23
. . . . . .
. . . . . .
. . . . . .
78 3 1 1 3 43
79 3 1 1 3 53
80 3 1 1 2 65
81=34 3 1 1 1 79
the constraints is (b1 ) 2 (b2 )3 (b3 ) 3 (b4 )1 and the objective function is – 29. It
is compared with the fifth row of Table 6.40. The final solution is the
same as that of (1).
Find b1 , b2 , b3 (6.63a)
where V i (i = 1,2,3) is the stress of the ith member. The design variables can have
the levels in Table 6.43.
(1) Find a solution by using an orthogonal array.
(2) Find a solution by arranging the estimated values in ascending order.
356 Analytic Methods for Design Practice
Table 6.43. Levels of design variables for Example 6.15 (unit: m2)
Design variable
Level
b1 b2 b3
1 0.00018 0.0001 0.00002
2 0.0002 0.00012 0.00004
3 0.00022 0.00014 0.00006
Solution
(1) The number of design variables is three and the number of levels for each
design variable is three. Thus, the L9 (3 4 ) orthogonal array is selected.
The results of the experiments are shown in Table 6.44. The scale factor is
set as s = 100 and the process of Example 6.14 is performed. The results
from the one-way table for < in Equation 6.60 are shown in Table 6.45.
From Table 6.45, the design solution is (b1 )3 (b2 )3 (b3 )3 . The objective
function is 1.485 from the confirmation experiment and the constraints are
satisfied. Comparing the results in Table 6.44 and the solution from the
one-way table, the final solution is (b1 ) 2 (b2 ) 2 (b3 )1 , which is the fifth row
of the orthogonal array. Therefore, the design variables are changed from
b [0.0002 0.00012 0.00004]T to b [0.0002 0.00012 0.00002]T
and the mass is reduced from 1.273 kg to 1.194 kg.
Design of Experiments 357
Level
Design
variable
1 2 3
1 3 1 1 1.2172 Violated
2 2 3 1 1.2496 Satisfied
. . . . . .
. . . . . .
(2) The estimated values for < are arranged in ascending order as shown in
Table 6.46. In Table 6.46, b [0.0002 0.00014 0.00002]T is the first
one that satisfies the constraints. Therefore, the design solution is (b1 ) 2
(b2 ) 3 (b3 )1 and the objective function is 1.2496. However, the solution of
(1) is better. Therefore, the design solution of the fifth row of the
orthogonal array is selected as the final solution.
satisfied. The overall flow is illustrated in Figure 6.6 and the steps are as follows
(Lee et al. 2003):
Step 1. (Definition of the problem) The objective function and design
variables are defined. Candidate values are defined in a discrete space.
The candidate values are the values that the design variables can have.
When we define the initial design b (0) , the existing design is usually used
for the initial design to improve the existing design. The iteration number
is set to k 0 and an appropriate orthogonal array is selected. The
number of levels is usually defined as three because three levels easily
include a larger value, a smaller value and the curvature. A larger number
Start
Matrix experiment
k=k+1
ANOM for the characteristic function
No
Convergence criteria satisfied?
Yes
End
Table 6.47. Levels of design variables at the first iteration of Example 6.17 (unit: m2)
Design variable
Level
b1 b2 b3
1 0.00026 0.0001 0.00014
2 0.00024 0.00008 0.00012
3 0.00028 0.00012 0.00016
6.7.1 Introduction
f f (b) (6.64)
f(x)
f(x)
f1
f2
f3
x1 x2 x3 x
Yes
Optimization with the response surfaces
[1 [2 … [n [ n 1 [ n2 … [l
f
b1 b2 … bn b12 b22 … bn2 b1b2 … bn 1bn
2 2
f1 b11 b12 b1n b11 b12 b12n b11b12 b1, n 1b1n
2 2
f2 b21 b22 b2 n b21 b22 b22n b21b22 b2,n 1b2 n
fk bk1 bk 2 bkn bk21 bk22 2
bkn b k 1b k 2 bk ,n 1bkn
where H is the error in the approximation process. In Equation 6.65, the linear
and quadratic terms are treated in the same manner. l n(n 1) / 2 n is set and
the coefficients are substituted by a0 o c0 , a1 o c1 , ..., a11 o cn 1 ,..., an 1, n o cl .
As shown in Table 6.48, the variables are substituted by b1 o [1 , ..., b12 o [ n 1 , ...,
bn 1bn o [ l . Then Equation 6.65 yields
f x1 x2 x3 … xl
f1 x11 x12 x13 … x1l
f2 x 21 x 22 x 23 … x 2l
…
fk x k1 xk 2 xk 3 … x kl
364 Analytic Methods for Design Practice
[ ij (max ([ mj ) min ([ mj )) / 2
m m
xij (6.68)
(max ([ mj ) min ([ mj )) / 2
m m
The data in Table 6.48 are shown in Table 6.49 with xij and Equation 6.67 is
modified to
fi yˆ i H i (6.70)
To determine ŷ , L is defined by the data in Table 6.49 and the error at each
candidate point ( H i ) as follows:
k k k l
2 2 2
L ¦Hi ¦ ( f i yˆ i ) ¦ ( f i (c0 ¦ c j xij )) (6.71)
i 1 i 1 i 1 j 1
wL
0, i 0,1, 2, ... , l (6.72)
wci
ª k k k º ª k º
« k ¦ xi1 ¦ xi 2 ¦ xil » « i¦1 f i »
i 1 i 1 i 1 ªc º
«k k k k »« 0 » «k »
« ¦ xi1 2
¦ xi1 ¦ xi1 xi 2 ¦ xi1 xil » « c1 » « ¦ xi1 f i »
(6.73)
«i 1 i 1 i 1 i 1 »« » «i 1 »
« »« » « »
«k k k k
2 » c «k »
« ¦ xil ¦ xil xi1 ¦ xil xi 2 ¦ xil » ¬ l ¼ « ¦ xil f i »
¬i 1 i 1 i 1 i 1 ¼ ¬i 1 ¼
are included in the approximation. When l is large, more terms are generally
considered and the cost is higher. Therefore, the user should appropriately
determine l according to the environment.
Solution
The values for b1 and b2 are allocated to the orthogonal array in Table 6.50. x in
Table 6.50 is the normalized value and the function f at the candidate points is
shown in Table 6.50. Because we have two design variables with three levels, the
experiment with the L9 (34 ) orthogonal array is the same as the factorial
experiment. The approximation function is defined as
f c 0 c1 x1 c 2 x 2 c3 x12 c 4 x 22 c5 x1 x 2 H (6.75a)
f # yˆ c 0 c1 x1 c 2 x 2 c3 x12 c 4 x 22 c5 x1 x 2 (6.75b)
Observation b1 ( x1 ) b2 ( x 2 ) f
1 –1.5 (–1) –3 (–1) –1.022
2 –1.5 (–1) 0 (0) 4.503
3 –1.5 (–1) 3 (1) 31.997
4 1.25 (0) –3 (–1) 8.704
5 1.25 (0) 0 (0) 1.636
6 1.25 (0) 3 (1) 8.793
7 4 (1) –3 (–1) 37.341
8 4 (1) 0 (0) 10.243
9 4 (1) 3 (1) 4.157
366 Analytic Methods for Design Practice
ª9 0 0 6 6 0 º ªc0 º ª 106.35169 º
«0 « »
« 6 0 0 0 0»» « c1 » « 16.26412 »
« »
«0 0 6 0 0 0 » «c 2 » « 0.07552 »
« »« » « » (6.76)
«6 0 0 6 4 0 » « c3 » « 87.21919 »
«6 0 0 4 6 0 » «c 4 » « 89.96951 »
« »« » « »
«¬0 0 0 0 0 4»¼ «¬c5 »¼ «¬ 66.20445»¼
Therefore,
1
V (2 2b2 2(1 3 )b3 ) u 10 4 (6.78)
b1b2 2b1b3 b 2 b3
Approximate the function in Equation 6.78 by using RSM. The candidate points
are defined by the central composite design for 0.1 u 10 4 d b1 , b2 , b3 d 5.0 u 10 4
(Myers and Montgomery 1995).
Solution
Table 6.51 shows the allocation of the candidate points made by the central
composite design. x in Table 6.51 is the normalized value and the function f at the
candidate points is shown in Table 6.51. The function in Equation 6.78 is
approximated as
V c 0 c1 x1 c 2 x 2 c 3 x 3 c 4 x1 x 2 c 5 x1 x 3
c 6 x 2 x 3 c 7 x12 c 8 x 22 c 9 x 32 H (6.79a)
V # yˆ c 0 c1 x1 c 2 x 2 c 3 x 3 c 4 x1 x 2 c 5 x1 x 3
c 6 x 2 x 3 c 7 x12 c 8 x 22 c 9 x 32 (6.79b)
Table 6.51. Allocation of candidates and function calculation of Example 6.19 (unit of bi: 10 4 )
Observation b1 ( x1 ) b2 ( x 2 ) b3 ( x3 ) V
1 5 (1) 5 (1) 5 (1) 4.858E+07
2 5 (1) 5 (1) 0.1 (–1) 5.605E+07
3 5 (1) 0.1 (–1) 5 (1) 7.593E+07
4 5 (1) 0.1 (–1) 0.1 (–1) 6.813E+07
5 0.1 (–1) 5 (1) 5 (1) 1.582E+08
6 0.1 (–1) 5 (1) 0.1 (–1) 1.448E+09
7 0.1 (–1) 0.1 (–1) 5 (1) 2.268E+09
8 0.1 (–1) 0.1 (–1) 0.1 (–1) 2.429E+09
9 5 (1) 2.55 (0) 2.55 (0) 5.672E+07
10 0.1 (–1) 2.55 (0) 2.55 (0) 2.971E+08
11 2.55 (0) 5 (1) 2.55 (0) 8.092E+07
12 2.55 (0) 0.1 (–1) 2.55 (0) 1.465E+08
13 2.55 (0) 2.55 (0) 5 (1) 9.262E+07
14 2.55 (0) 2.55 (0) 0.1 (–1) 1.090E+08
15 2.55 (0) 2.55 (0) 2.55 (0) 9.525E+07
1.814 x1x3 1.431x2 x3 3.564 x12 2.932 x22 2.803 x32 ) (6.81)
368 Analytic Methods for Design Practice
Find x Rn (6.82a)
xL d x d xU (6.82d)
where x is the design variable vector transformed from the original design
variable vector b, fˆ and ĝ are the objective function and the jth constraint
j
approximated by RSM, respectively. When an original function has an explicit
form, it can be directly used without approximation. Equation 6.82 is a metamodel
of Equation 3.1 and an optimum point is obtained from it. Since only explicit
functions are included in Equation 6.82, it is easy to find an optimum point.
When a highly nonlinear function is approximated in a wide range, the
accuracy can be deteriorated. There are methods to evaluate the accuracy. The
methods are introduced in a reference (Myers and Montgomery 1995).
Find b
g2 b12 b2 d 0
Use the same allocation method as Example 6.18 at b1 0.5, 0, 0.5 and b2
1.0, 0, 1.0.
Design of Experiments 369
Solution
Find x
to minimize fˆ 72 x1 33.33 x 2
gˆ 2 0.167 x 2 d 0
1 d x1 , x2 d 1
Find b (6.85a)
For the candidate points and allocation of Example 6.19, approximate the
functions in quadratic forms and obtain the optimum point.
Solution
If the response surface method is performed by normalizing the design variables,
the problem in Equation 6.85 is approximated to the following:
Find x (6.86a)
0.6482 x 32 d 0 (6.86h)
When the initial design is (0.102, í 0.878, í 0.388), the objective function is
f 6.62 u 10 4 and the constraints g1 , g 3 and g 5 are violated. The optimum
point is (0.159, í 0.171, í 0.886). The objective function is f 6.824 u 10 4 and
the constraints g1 and g 3 are active.
Solution
The objective function is a linear function and the constraints are rational functions.
The optimum point of the original problem is b* 10 4 u [2.0230 1.0460 0.1]T
and the objective function is f = 4.048 u 10 4. The solution in Example 6.21 is the
normalized one. Two solutions are similar. Moreover, the original constraints are
satisfied at the optimum of the approximated problem. Thus, the performance of
the approximation is fairly good for this problem.
6.8 Exercises
3.5 2.3 3.6 1.9 4.2 2.1 3.5 2.9 4.2 3.8
2.0 3.3 4.3 1.8 2.7 3.2 3.7 1.6 2.6 3.9
6.2 Regard the grades in Exercise 6.1 as 20 samples of all the alumni.
Estimate the mean and the variance of the population.
6.3 For Exercise 6.2, find the 95% confidence interval for the mean of the
entire population. Use the table of t-distribution.
A1 A2 A3
Factor A
A1 A2 A3
Factor B
12.35 17.89 26.89
14.78 12.78 32.24
B1
20.89 25.23 23.24
15.98 18.99 12.38
18.67 23.35 37.27
26.78 23.45 35.89
B2
31.26 35.47 25.67
23.87 42.89 47.64
25.92 23.78 34.78
12.98 32.87 25.98
B3
23.89 45.84 42.34
35.62 59.82 56.73
23.82 48.93 35.89
34.89 34.89 56.83
B4
63.78 63.21 54.23
39.83 33.89 63.67
6.6 Show that the columns of the L8 (2 7 ) orthogonal array in Table 6.22 are
orthogonal to each other.
6.7 There are two factors A and B with two levels, four factors C, D, E and F
with three levels. The interactions A u C , A u D and E u F are significant.
Select the minimum orthogonal array for a matrix experiment.
Minimize f 2 A 2 ( B A) 2 2C 2
Level
Factor
1 2 3
A –1.5 0 1
B –1 0 1
C –1 0 1
Level
Factor
1 2
A 2 5
B 5 8
C 4 7
D 1 4
Level b1 b2 b3
(1) Find a design solution by using the L9 (34 ) orthogonal array. Use the
method in Section 6.3.3.
(2) Find a design solution by using the factorial combination and compare
the solution with the one from (1).
(3) Make an ANOVA table and discuss the effects of the factors and the
error.
(4) From the effect of the error, discuss the effect of the interactions.
(5) If interactions exist, make a design using an orthogonal array by
considering the interactions and compare the results with the one from
(2).
6
2
f ¦ (bi 2)
i 1
Level b1 b2 b3 b4 b5 b6
(1) Find a design solution by using the L18 (21 u 3 7 ) orthogonal array.
Use the method in Section 6.3.3.
(2) Find a design solution by using the factorial combination and compare
the solution with the one from (1).
(3) Make an ANOVA table and discuss the effects of the factors and the
error.
(4) Using the effect of the error, discuss the effect of the interactions.
(5) If interactions exist, make a design using an orthogonal array by
considering the interactions and compare the results with the one from
(2).
6.14 A design problem for the three bar truss in Figure 3.3 is as follows:
Maximize [1
where [1 is the first natural frequency. Young’s modulus (E) is 207 GPa,
and the density ( U ) is 2770 kg/m3. The levels of the design variables are
as follows:
Level b1 b2 b3
Maximize [1
9.144 m 9.144 m
a b
g
i
e
f 9.144 m
h
j
c d
where [1 is the first natural frequency. Young’s modulus (E) is 207 GPa,
and the density ( U ) is 2770 kg/m3. The levels of design variables are as
follows:
(1) The L18 (21 u 37 ) orthogonal array is selected and the design variables
are allocated as follows:
378 Analytic Methods for Design Practice
6.16 In Example 3.8, the design variables can have the following discrete
values:
Design of Experiments 379
Level
1 2
b1 12 16
b2 0 2
Find the solutions by using Equation (6.60) when the scale factor s is 0.1,
1.0, 10, and 100. Use the method in Section 6.5 with an orthogonal array.
Find b1 ,b2
subject to (1 b1 ) 3 b2 t 0
0 d bi , i 1, 2
Level
1 2 3
b1 –1 1 3
b2 –2 0 2
Find solutions by using Equation (6.60) when the scale factor s is 10 and
50. Use the method in Section 6.5 with an orthogonal array.
Find b1 , b2 , b3
b1 2b2 2b3 t 0
0 d bi d 42, i 1, 2, 3
Level
1 2 3
b1 14 24 34
b2 2 12 22
b3 2 12 22
Find solutions by using Equation (6.60) when the scale factor s is 5 and 50.
Use the method in Section 6.5 with an orthogonal array.
Find b1 , b2 , b3 , b4
subject to 8 b1 2b2 t 0
12 4b1 b2 t 0
12 3b1 4b2 t 0
5 b3 b4 t 0
8 2b3 b4 t 0
8 b3 2b4 t 0
Level
1 2 3
b1 –2 0 2
b2 1 3 5
b3 –2 0 2
b4 2 4 6
Find solutions by using Equation (6.60) when the scale factor s is 5 and 10.
Use the method in Section 6.5 with an orthogonal array.
Design of Experiments 381
6.20 Using SOA, perform two iterations for Exercise 6.16. The candidate
values of the design variables are as follows:
s = 10 and the initial design is the solution of Exercise 6.16 when s = 10.
6.21 Using SOA, perform two iterations for Exercise 6.17. The candidate
values of the design variables are as follows:
Candidates b1 , b2 { 5, 4, 3, 2, 1, 0, 1, 2, 3, 4, 5 }
s = 10 and the initial design is the solution of Exercise 6.17 when s = 10.
6.22 Using SOA, perform two iterations for Exercise 6.18. The candidate
values of the design variables are as follows:
s = 50 and the initial design is the solution of Exercise 6.18 when s = 50.
6.23 Using SOA, perform two iterations for Exercise 6.19. The candidate
values of the design variables are as follows:
Candidates b1,b2 ,b3 ,b4 { 2.0, 1.5, 1.0, ,8, 8.5, 9 , 9.5, 10 }
s = 10 and the initial design is the solution of Exercise 6.19 when s = 10.
6.24 Solve Exercise 6.17 by using RSM. The candidates are those in Exercise
6.17. Use the central composite design. Compare the solution with the
mathematical optimum and the solution from Exercise 6.17.
6.25 Solve Exercise 6.18 by using RSM. The candidates are those in Exercise
6.18. Use the L9 (3 4 ) orthogonal array. Compare the solution with the
mathematical optimum and the solution from Exercise 6.18.
6.26 Solve Exercise 6.19 by using RSM. The candidates are those in Exercise
6.19. Use the central composite design. Compare the solution with the
mathematical optimum and the solution from Exercise 6.19.
382 Analytic Methods for Design Practice
6.A Statistics
Statistical theories are briefly explained.
1 n
P ( x1 x 2 ... x n ) / n ( ¦ xi ) (6.A.1)
n i1
Variance and standard deviation: From Equation 6.A.1, the variance V 2 and the
standard deviation V are defined as:
n
2
¦ ( xi P )
V2 i 1
(6.A.2)
n
b
P{a d X d b} ³ p( x)dx (6.A.5)
a
F ( x) P{ X d x} (6.A.6)
Design of Experiments 383
2
1 § xP ·
¨ ¸
1
f ( x) e 2© V ¹ (6.A.7)
2SV 2
where e is the base of the natural logarithm. The normal distribution is denoted as
N ( P , V 2 ) and the random variable is denoted by Z. The normal distribution
N(0,1) is called the standard normal distribution and the pdf is
1
1 z2
I ( z) e 2 (6.A.8)
2S
In most textbooks on statistics, a table for the following zD , which has 1 D for
the cumulative probability, is included.
zD
1D P{Z d zD } ³ I ( z )dz (6.A.9)
f
1–D
D
0 ZD Z
Z
t ( n) (6.A.12)
F 2 ( n)
n
F-distribution: Suppose F12 (n1 ) and F 22 (n2 ) are independent and their degrees
of freedom are n1 and n 2 . When they have F 2 -distributions, a random variable
is defined as follows:
F12 (n1 ) / n1
F (6.A.14)
F 22 (n2 ) / n2
The random variable F has n1 DOF for the numerator, n2 DOF for the
denominator and has F-distribution. In most textbooks on statistics, a table for the
following cumulative probability 1 D of F (n1 , n2 ;D ) is included:
n
¦ Xi
i 1
P̂ X (6.A.16)
n
Design of Experiments 385
n
2
¦(Xi X )
Vˆ 2 S2 i 1
(6.A.17)
n 1
The estimated values P̂ and Vˆ 2 are called “unbiased estimators.” The reason
is that their expected values are as follows:
E (X ) P (6.A.18)
E (S 2 ) V2 (6.A.19)
X P
T (6.A.20)
S/ n
§ ·
P¨ t D T t D ¸ 1D (6.A.21)
¨ ¸
© 2 2 ¹
§ S S ·¸
P¨ X t D P X tD 1D (6.A.22)
¨ n n ¸
© 2 2 ¹
Equation 6.A.22 represents the probability for the distribution of the population.
The degrees of freedom of the t-distribution in Equation 6.A.22 is n í 1.
386 Analytic Methods for Design Practice
Column
Exp. no.
1 2 3
1 1 1 1
2 1 2 2
3 2 1 2
4 2 2 1
Column
Exp. no.
1 2 3 4 5 6 7
1 1 1 1 1 1 1 1
2 1 1 1 2 2 2 2
3 1 2 2 1 1 2 2
4 1 2 2 2 2 1 1
5 2 1 2 1 2 1 2
6 2 1 2 2 1 2 1
7 2 2 1 1 2 2 1
8 2 2 1 2 1 1 2
Design of Experiments 387
11
Table 6.B.3. The L12 ( 2 ) orthogonal array
Exp. Column
no. 1 2 3 4 5 6 7 8 9 10 11
1 1 1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 1 2 2 2 2 2 2
3 1 1 2 2 2 1 1 1 2 2 2
4 1 2 1 2 2 1 2 2 1 1 2
5 1 2 2 1 2 2 1 2 1 2 1
6 1 2 2 2 1 2 2 1 2 1 1
7 2 1 2 2 1 1 2 2 1 2 1
8 2 1 2 1 2 2 2 1 1 1 2
9 2 1 1 2 2 2 1 2 2 1 1
10 2 2 2 1 1 1 1 2 2 1 2
11 2 2 1 2 1 2 1 1 1 2 2
12 2 2 1 1 2 1 2 1 2 2 1
Exp. Column
no. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2
3 1 1 1 2 2 2 2 1 1 1 1 2 2 2 2
4 1 1 1 2 2 2 2 2 2 2 2 1 1 1 1
5 1 2 2 1 1 2 2 1 1 2 2 1 1 2 2
6 1 2 2 1 1 2 2 2 2 1 1 2 2 1 1
7 1 2 2 2 2 1 1 1 1 2 2 2 2 1 1
8 1 2 2 2 2 1 1 2 2 1 1 1 1 2 2
9 2 1 2 1 2 1 2 1 2 1 2 1 2 1 2
10 2 1 2 1 2 1 2 2 1 2 1 2 1 2 1
11 2 1 2 2 1 2 1 1 2 1 2 2 1 2 1
12 2 1 2 2 1 2 1 2 1 2 1 1 2 1 2
13 2 2 1 1 2 2 1 1 2 2 1 1 2 2 1
14 2 2 1 1 2 2 1 2 1 1 2 2 1 1 2
15 2 2 1 2 1 1 2 1 2 2 1 2 1 1 2
16 2 2 1 2 1 1 2 2 1 1 2 1 2 2 1
388 Analytic Methods for Design Practice
4
Table 6.B.5. The L9 (3 ) orthogonal array
Column
Exp. no.
1 2 3 4
1 1 1 1 1
2 1 2 2 2
3 1 3 3 3
4 2 1 2 3
5 2 2 3 1
6 2 3 1 2
7 3 1 3 2
8 3 2 1 3
9 3 3 2 1
Design of Experiments 389
Exp. Column
no. 1 2 3 4 5 6 7 8 9 10 11 12 13
1 1 1 1 1 1 1 1 1 1 1 1 1 1
2 1 1 1 1 2 2 2 2 2 2 2 2 2
3 1 1 1 1 3 3 3 3 3 3 3 3 3
4 1 2 2 2 1 1 1 2 2 2 3 3 3
5 1 2 2 2 2 2 2 3 3 3 1 1 1
6 1 2 2 2 3 3 3 1 1 1 2 2 2
7 1 3 3 3 1 1 1 3 3 3 2 2 2
8 1 3 3 3 2 2 2 1 1 1 3 3 3
9 1 3 3 3 3 3 3 2 2 2 1 1 1
10 2 1 2 3 1 2 3 1 2 3 1 2 3
11 2 1 2 3 2 3 1 2 3 1 2 3 1
12 2 1 2 3 3 1 2 3 1 2 3 1 2
13 2 2 3 1 1 2 3 2 3 1 3 1 2
14 2 2 3 1 2 3 1 3 1 2 1 2 3
15 2 2 3 1 3 1 2 1 2 3 2 3 1
16 2 3 1 2 1 2 3 3 1 2 2 3 1
17 2 3 1 2 2 3 1 1 2 3 3 1 2
18 2 3 1 2 3 1 2 2 3 1 1 2 3
19 3 1 3 2 1 3 2 1 3 2 1 3 2
20 3 1 3 2 2 1 3 2 1 3 2 1 3
21 3 1 3 2 3 2 1 3 2 1 3 2 1
22 3 2 1 3 1 3 2 2 1 3 3 2 1
23 3 2 1 3 2 1 3 3 2 1 1 3 2
24 3 2 1 3 3 2 1 1 3 2 2 1 3
25 3 3 2 1 1 3 2 3 2 1 2 1 3
26 3 3 2 1 2 1 3 1 3 2 3 2 1
27 3 3 2 1 3 2 1 2 1 3 1 3 2
390 Analytic Methods for Design Practice
Column
Exp. no.
1 2 3 4 5 6 7 8
1 1 1 1 1 1 1 1 1
2 1 1 2 2 2 2 2 2
3 1 1 3 3 3 3 3 3
4 1 2 1 1 2 2 3 3
5 1 2 2 2 3 3 1 1
6 1 2 3 3 1 1 2 2
7 1 3 1 2 1 3 2 3
8 1 3 2 3 2 1 3 1
9 1 3 3 1 3 2 1 2
10 2 1 1 3 3 2 2 1
11 2 1 2 1 1 3 3 2
12 2 1 3 2 2 1 1 3
13 2 2 1 2 3 1 3 2
14 2 2 2 3 1 2 1 3
15 2 2 3 1 2 3 2 1
16 2 3 1 3 2 3 1 2
17 2 3 2 1 3 1 2 3
18 2 3 3 2 1 2 3 1
References
Alberto GD, Phillips DT (1994) Principles of Experimental Design and Analysis.
Chapman and Hall, Boca Raton, FL
Box GEP, Draper NR (1987) Empirical Model-Building and Response Surfaces.
Wiley, New York
Box GEP, Hunter WG (1978) Statistics for Experimenters. Wiley, New York
Freund JE, Walpole RE (1987) Mathematical Statistics. Prentice Hall, Englewood
Cliffs, NJ
Goldberg DE (1989) Genetic Algorithms in Search Optimization and Machine
Learning. Addison Wesley, New York
Design of Experiments 391
Hedayat AS, Sloane NJA, Stufken J (1999) Orthogonal Array: Theory and
Applications. Springer, Berlin Heidelberg New York
Korte BH, Vygen J (2003) Combinatorial Optimization: Theory and Algorithms.
Springer, Berlin Heidelberg New York
Ku KJ, Rao SS, Chen L (1998) Taguchi-aided Search Method for Design
Optimization of Engineering Systems. Engineering Optimization 30:1í23
Lee KH, Eom IS, Park GJ, Lee WI (1996) Robust Design for Unconstrained
Optimization Problems Using Taguchi Method. AIAA J. 34(4):1059í1063
Lee KH, Park GJ (1996) Discrete Post Process of the Constrained Size
Optimization Using Orthogonal Arrays. Proceedings of 6th AIAA/USAF/
NASA/ISSMO Symposium on Multidisciplinary Analysis and Optimization,
Part 1, AIAA, Sept. 4í6, Bellevue, pp. 56í61
Lee KH, Park GJ (2002) Robust Optimization in Discrete Design Space for
Constrained Problems. AIAA J. 40(4):774í780
Lee KH, Yi JW, Park JS, Park GJ (2003) An Optimization Algorithm Using
Orthogonal Arrays in Discrete Design Space for Structures. Finite Elements in
Analysis and Design 40(1):121í135
Lim JM, Wu, DH, Park GJ (2000) Analysis and Design Consideration of an
Energy Absorbing Steering System using Orthogonal Arrays. International
Journal of Crashworthiness 5(3):271í278
Montgomery DC (1991) Design and Analysis of Experiments. Wiley, New York
Mood AM, Grabill FA, Boes DC (1974) Introduction to the Theory of Statistics.
McGraw Hill, New York
Myers RH, Montgomery DC (1995) Response Surface Methodology. Wiley, New
York
Osman, IH, Kelly JP (1996) Meta-heuristics: Theory & Application. Kluwer,
Dordrecht
Park GJ, Hwang WJ, Lee WI (1994) A Structural Optimization Post-process Using
Taguchi Method. International Journal of Japan Society of Mechanical
Engineers, Series A 37(2):166í172
Phadke MS (1989) Quality Engineering Using Robust Design. Prentice Hall,
Englewood Cliffs, NJ
Vanderplaats GN (1984) Numerical Optimization Techniques for Engineering
Design with Applications. McGraw Hill, New York
Taguchi G (1987) Systems of Experimental Design (Vols. I, II). Kraus
International Publications, New York
Walpole MS (1982) Introduction to Statistics. Macmillan, New York
7 Robust Design
7.1 Introduction
As the engineering environment becomes more and more competitive, quality
products are required in industry. Unexpected deviations from the function that a
designer initially intended are caused by variations. Robust design seeks to
prevent such phenomena and has been developed to improve the quality and
reliability of the product in engineering. Recently, this technology has expanded
to various design areas.
Many researchers have defined robust design. Taguchi, who is the pioneer of
robust design, stated that “robustness is the state where the technology, product, or
process performance is minimally sensitive to factors causing variability (either in
the manufacturing or the user’s environment) and aging at the lowest unit
manufacturing cost” (Taguchi et al. 2000). Suh stated that “robust design is
defined as the design that satisfies the functional requirements even though the
design parameters and the process variables have large tolerances for ease of
manufacturing and assembly. This definition of robust design states that the
information content is minimized” (Suh 2001). Box defined that “robustifying a
product is the process of defining its specifications to minimize the product’s
sensitivity to variation” (Box and Fung 1993). Although different expressions are
used, their meanings are similar. A common aspect of the definitions is that robust
design is insensitive to variations and this concept is accepted in the engineering
community.
Conventional design methodologies have been developed to improve the
performances of the products. If consideration of robustness is added, a new
procedure should be made in addition to the conventional concept. Theories of
robust design have been developed by using existing design theories. The first one
is the Taguchi method and it is the foundation of robust design. Taguchi proposed
methods of determining variables to make a performance insensitive to noises in
the manufacturing process. The concept of robustness was established by the
Taguchi method. In a product design, robustness is obtained by appropriate
modification of the Taguchi method. The second one is robust optimization where
the robustness concept is added to conventional optimization. The objective
394 Analytic Methods for Design Practice
function and constraints are redefined with robustness indices. The last one is
axiomatic design, which can be exploited in conceptual design. In axiomatic
design, the Independence Axiom is utilized to find excellent designs. When
multiple designs are found by using the Independence Axiom, the best one is
selected by the Information Axiom and the robustness concept is employed in that
process.
Various robust design methodologies are overviewed based on the above three
methods. The trend of the research is surveyed and the applications are discussed.
However, theoretical aspects are emphasized, rather than the application aspects.
The advantages and disadvantages of the methods are also discussed. In the robust
design area, some terminologies are used in different ways. Thus, those
terminologies are precisely defined and a unified viewpoint of robust design is
proposed.
There are some methods that are similar to robust design such as design
methods considering reliability and uncertainty. Uncertainties or noises are
included in these methods as well. Optimization terminologies can be utilized to
distinguish the methods from robust design. In robust design, low sensitivity of
the objective function is emphasized. In reliability design, reliability of constraints
is important. The reliability of constraints is similar to the robustness of the
constraints in robust design. Design with uncertainties is similar to robust
optimization. However, robustness of the objective function is not considered.
The distinctions are made based on the methods of the early stages. These days,
the methods are often fused; therefore, it may not be possible to distinguish them
in some applications. Methods considering reliability or uncertainties are not
included (Special Supplement on Robust and Reliability-based Design, 2006).
f (b) o f (b z b , p z p ) (7.1a)
g j (b ) o g j ( b z b , p z p ) (7.1b)
Find b Rn (7.2a)
to minimize F (b, p, z ) (7.2b)
bL d b d bU (7.2d)
l
1
Sample mean f or (m f ) f
l ¦f
i 1
i
2 1 l 2
Sample variance V or ( s f ) V ¦ ( fi f )
l 1 i 1
P 1 l
Population mean P ¦ fi
li 1
1 l
Population variance V2 V2 ¦ ( fi P )
2
li 1
Generally, F and G j are defined as the functions of the mean and the variance of f
and g j , respectively.
Robust design should not be developed by the deterministic concept. It should
be developed by the stochastic concept, since the deterministic approach cannot
include the noises. The performance measures are defined as the data to be
measured experimentally or mathematically, which represent the performances of a
product or a process. The performance measures are usually modified to form a
characteristic function. The performance measures could vary under the same
circumstances due to the distribution of noise factors. Generally, the distribution is
measured by the mean and the variance. The formulae of mean and variance are
summarized in Table 7.1. f i is the ith characteristic function in the experiment or
the ith response function in the design.
The mean and the variance for the design are explained. The characteristic
function f in Table 7.1 is a specific response. A design variable (factor) has a
distribution around a nominal value for many reasons. Thus, the response f also
has a distribution. The distribution is quantified by the variance and the sources
for the variance are shown in Table 7.2. The mean of f is expressed by the mean
P f and the variance is expressed by V 2f . V f is the standard deviation. The
mean P f and the variance V 2f of the characteristic function f are represented as
Factor Example
b p
Pf E[ f (b)] ³ ³ ³ f (b z , p z )
(7.3a)
u1 ( z1b ) u n ( z nb )v1 ( z1p ) vo ( z op ) dz1b dz nb dz1p dz op
V 2f E[( f (b) P f ) 2 ] b p
³ ³ ³ [ f (b z , p z ) P f ]
2
(7.3b)
u1 ( z1b ) u n ( z nb ) v1 ( z1p ) v o ( z op ) dz1b dz nb dz1p dz op
1 ª (bi P b ) 2 º
u i ( z ib ) exp « i
» (7.4)
V bi 2S «¬ 2V b2i »¼
where P bi and V bi are the mean and the standard deviation of the ith design
variable bi . The robust design techniques have been developed by replacing the
functions of the mean and variance into the functions of Equation 7.2. In the next
sections, the robust design methodologies are examined.
b1 and b2 have uniform distributions around (b1 , b2 ) (0.0, 0.0) and the ranges
are 0.5 d b1 d 0.5 and 0.5 d b2 d 0.5 . Obtain the mean and the variance of f at
(b1 , b2 ) (0.0, 0.0) .
Solution
From Equation 7.3a, the mean is
0.5 0.5
Pf ³ ³ (b1 b2 2 2b1b2 ) [u (b1 0.5) u (b1 0.5)]
0.5 0.5
[u (b2 0.5) u (b2 0.5)]db1db2 (7.6)
2
where u is the unit step function enabling the uniform distribution within the
design ranges.
By using Equation 7.3b, the variance is
398 Analytic Methods for Design Practice
0.5 0.5
V 2f ³
2
³ [(b1 b2 2 2b1b2 ) (2)] [u (b1 0.5) u (b1 0.5)]
0.5 0.5
[u (b2 0.5) u (b2 0.5)]db1db2 (7.7)
0.1944
Solution
G x and G y are calculated as
2lPx 1
Gx 1.414 u10 8 (7.8a)
b1 E E
2lPy 1
Gy 1.325 u 10 8 (7.8b)
(b1 2b2 ) E E
209 E 9 1 1
8
PG x ³ 1.414 u 10
203E 9 E 209 u 109 203 u 109
[u ( E 203 u 109 ) u ( E 209 u 109 )]dE (7.9a)
0.000686456
209 E 9 1 1
8
PG y ³ 1.325 u 10
203E 9 E 209 u 10 203 u 109
9
209 E 9 1 1
V G2x ³ ( 1.414 u 10
8
0.000686456) 2
203E 9 E 209 u 10 203 u 109
9
7.3.1 Introduction
By the end of the 1940s the Taguchi method had been developed by Dr. G.
Taguchi for quality improvement. His technique for quality engineering is referred
to as the Taguchi method or robust design (Taguchi 1987). While the Taguchi
method was successfully applied at the Electrical Communications Laboratories of
the Nippon Telephone and Telegraph Company, the AT&T Bell Laboratories
became interested in it. In 1980, Phadke invited Taguchi to the AT&T Bell
Laboratories. After being impressed by the Taguchi method, Phadke published a
textbook on the Taguchi method in 1989 (Phadke 1989).
The Taguchi method has greatly contributed to quality improvement of various
designs. In early case studies, the Taguchi method was applied to the process
design rather than the product design. That is, it was regarded as a method for
design of experiments rather than a design methodology.
In Equation 7.2 of the design formulation, the original Taguchi method does
not deal with the constraints in Equation 7.2. This leads to unconstrained
optimization problems. Although the design parameters p and the noise factors z
are not explicitly defined in the design, the Taguchi method can be used to
determine the settings of control factors for robust design. To explain this in detail,
let us use an example with three distributions of the objective function f as
illustrated in Figure 7.1.
In Figure 7.1, P A , P B and P C are the means of distributions A, B and C,
respectively. Then, robustness can be measured by the magnitude of variation.
When the design is focused only on the robustness of the performance, the
candidate designs B and C are better than design A. However, the target value of
the performance generally exists in the design. The target value can be zero,
infinite or a specific nominal value. Thus, the Taguchi method suggests a design
400 Analytic Methods for Design Practice
Probability
density C B
function
A
PA PC PB
f
m Target value
with minimum variation around the target value (m). From the Taguchi method
viewpoint, design C is the best out of the three candidate designs.
The procedure to obtain a robust design is now explained. The design
candidates D and E with two arbitrary distributions are represented in Figure 7.2.
The design range is equivalent to the allowable range and the means of the two
distributions coincide with the target value m of the design. A small part of
candidate design D is outside the design range while no part of candidate design E
is outside the design range. However, if any unexpected noise factors become
active, the distribution of E has a larger probability to be outside the design range
Design range
Probability D
density
function
E
m Target value f
Figure 7.2. Probability density functions of the two designs with the design range
Robust Design 401
than that of D. For instance, the unexpected noise is the tolerances of product
sizes, variances of the environment such as temperature, humidity, external forces,
manufacturing processes, etc. Thus, it is clear that candidate design D is more
robust than candidate design E.
L( f ) k ( f m) 2 (7.11)
where k is the constant to define the loss and m is the target value. The expected
value of the loss function is defined as
Q E[ L( f )] k[V 2 ( P m) 2 ] (7.12)
z Noise factors
m f
Product / Process
Target value Response
(Signal factor)
b Design variables
(Control factors)
design increases the cost. The parameter design is dealt with here.
When the target value of a response is given as illustrated in Figure 7.3, the
Taguchi method determines the optimum setting of control factors, so that the
variation of a response is minimized, although uncontrollable factors (noise
factors) exist. Equation 7.12 can be regarded as the index to find a robust design.
Suppose that we have a scale factor s to adjust the current mean to the target value.
The scale factor s is described as
m
s (7.13)
P
When the current mean is adjusted to the target value, the average loss function of
Equation 7.12 is changed to
§§ m 2 2
· §m · · V2
Qa k ¨ ¨¨ P m ¸¸ ¨¨ V ¸¸ ¸ km2 (7.14)
¨© P ¹ © P ¹ ¸¹ P2
©
P2
K 10 log10 (7.15)
V2
Equation 7.15 is the ratio of the power of the signal factor P and the power of
noise factors V . Thus, it is called the S/N ratio. Maximizing Equation 7.15 is
equivalent to minimizing Equation 7.14. That is, a robust design is obtained by
maximizing Equation 7.15.
mf 2
Nominal-the-best with scale factor: K 10 log or K 10 log s f 2
sf 2
ª1 n º
Smaller-the-better K 10 log «
¬« n
¦f
i 1
i
2
»
¼»
ª1 n
1 º
Larger-the-better K 10 log «
«¬ n
¦f
i 1
2
i »
»
¼
Robust Design 403
As in the previous example of Figure 7.2, the response with target value m is
referred as the nominal-the-best type characteristic. On the contrary, the responses
with target values of zero and infinity are referred as the smaller-the-better type
and the larger-the-better type, respectively. The S/N ratios of the smaller-the-
better type and the larger-the-better type can be derived by a similar procedure as
for the nominal-the-best type. As the index of robustness, the S/N ratio of
Equation 7.15 is preferred to the average loss function of Equation 7.14. The
examples of S/N ratios are summarized in Table 7.3.
Example 7.4 [The Loss Function and S/N Ratio of the Three Bar Truss]
In Example 7.2, E has the values 203, 208, 209 and 211 GPa. Calculate the loss
functions and the S/N ratios with Equation 7.15. The target values are the means
in Example 7.2.
Solution
The loss functions have the nominal-the-best type and the target values are mG x
0.000686, mG y 0.000643 from Example 7.2. Table 7.5 shows the loss
functions. The constant k for the loss function is 1.
404 Analytic Methods for Design Practice
The means of the loss functions are P G x = 1.23 u 10 10 and P G y = 1.09 u 10 10.
The variances are V G2x = 8.32 u 10 21 and V G2y = 6.69 u 10 21 . If we substitute
these into Equation 7.15, the S/N ratios are 2.57 and 2.53.
The Taguchi method is classified into two submethods. One is the parameter
design, where the loss function is reduced by determining the values of parameters
when noises or tolerances exist. The other is the tolerance design, where the loss
function is reduced by determining the tolerances. Since the tolerance design is
usually expensive, the parameter design is generally utilized. In this section, the
parameter design is explained.
As mentioned earlier, the Taguchi method is generally used in experiments;
however, it is explained here from the viewpoint of application of the method to
the design. The method evaluates the characteristic function, which is equivalent
to the objective function of optimization. The characteristic function is not
evaluated by experiments but is numerically calculated. The Taguchi method has
been applied to various areas. In this section, only the basic aspects are dealt with.
References present many other cases (Taguchi 1987, Phadke 1989).
There are two goals in performing robust design. One is to minimize the
variability produced by the noise factors. The other is to make the mean value
close to the target value. The design that attains one goal is not usually consistent
with the one that attains the other goal. To meet the two goals, Taguchi developed
a two-step optimization strategy. The first step is to reduce the variation, and the
second step is to adjust the mean on the target. The procedure is illustrated in
Figure 7.4. The first step in Figure 7.4a concentrates on minimization of the
variation, while the mean is ignored. In the second step of Figure 7.4b, the mean is
moved to the target value while somewhat sacrificing the improved variation.
The first step is to find the optimum setting of control factors (design variables)
to maximize the S/N ratio or to reduce the variation. Orthogonal arrays are
employed to assign the control factors to an experimental matrix. The best
Robust Design 405
Target f
Probability G
densityG Low variability
performance is
function
shifted onto the
target Step 2
Target f
condition can be selected from the full combinations of control factors. However,
it is efficient to select the smallest orthogonal array because full combinations are
costly in most cases. The orthogonal arrays are explained in detail in Chapter 6.
Experiments are performed repeatedly for each row of an orthogonal array and the
mean and the variance are calculated from the repeated experiments. The noises
are automatically included in the repetition.
The S/N ratio is calculated for each row of an orthogonal array. Then, the
optimum setting of control factors is determined by performing ANOM of the S/N
ratios. A one-way table is utilized in the ANOM process. If the average loss of
Equation 7.12 is defined as the characteristic function to find a robust design, the
effect on the variation is confounded by the effect on the mean. Then the factors
are determined to have the mean close to the target although the variance is large.
This may be opposite to the goal of robust design, which tries to obtain a small
variance. Therefore, it is assumed that there is a scale factor that can move the
mean. Then the S/N ratio, which is the variance when the target is met, is utilized.
It is noted that the interactions should be allocated to the columns of an orthogonal
array. However, it is difficult to identify the existence of the interactions before
experiments. This is an obstacle in using orthogonal arrays; therefore, a
406 Analytic Methods for Design Practice
confirmation experiment should be performed with the results from the one-way
table.
A computer simulation model provides the same output with the same input
since it is a deterministic method. While this is a disadvantage, the computer
simulation model has the advantage that it can define the design parameters p and
the noise factors z exactly. In this case, Taylor series expansion, Monte-Carlo
simulation or an outer array can be utilized to obtain the variations. An outer array
is an orthogonal array that is used for a row of the orthogonal array (inner array).
The outer array has multiple cases of experiments for a row; therefore, it
accommodates the repeated experiments of an experiment case.
The second step is to set the mean to the target value by selecting a control
factor called the scale factor. The scale factor should only have an effect on the
mean, while the S/N ratio is maintained. In general experiments, time can be one
of the scale factors. However, it is difficult to find a scale factor in a general
design. Furthermore, it is possible that no scale factor exists and this prevents the
designer from applying the Taguchi method directly to the design. Therefore, each
row can use a function such as the loss function in Equation 7.12, which considers
the mean and the variance.
When it is impossible to find a scale factor, only the first step can be performed
out of the two steps. In this case, the choice of the characteristic function is crucial.
Moreover, the design problem generally has constraints as shown in Equation 3.1.
Thus, the characteristic function should include the constraints. When the
Outer array
Exp. Noise factor
no. z1 z 2 z 3 z 4
2
s f , v, <1
1 1 1 1 1
2 1 2 2 2
Inner array
. . . . .
Exp. Control factor
. . . . .
no. b1 b2 b3 b4
Nout . . . .
1 1 1 1 1
2 1 2 2 2 Exp. Noise factor
. . . . . no. z1 z 2 z 3 z 4
. . . . . 1 1 1 1 1
s 2f , v, < N in
Nin . . . . 2 1 2 2 2
. . . . .
. . . . .
Nout . . . .
Equation 7.16 has a similar concept to that of Equations 6.60 and 6.61. P f and
V f are the mean and the standard deviation of the objective function, respectively.
The designer can determine the weighting factors w1 and w2 with values from 0
to 1.
< (b) of Equation 7.16 is considered as the characteristic function and the
design method using orthogonal arrays in Chapter 6 is directly used. When we use
Equation 7.16, it is not easy to obtain a robust design when the constraint violation
is large. Thus, it is recommended that the Taguchi method be used in the feasible
region. It is noted that the characteristics of the method, which are explained in
Chapter 6, are also valid here.
The one step method using calculations or computer simulations is explained.
It is assumed that the number of design variables (control factors) is n = 4 and the
number of levels for each variable is three.
Step 1. An appropriate orthogonal array (inner array) is chosen according to
the number of design variables and levels. It is assumed that the
orthogonal array L9 (3 4 ) in Table 6.13 is chosen.
Step 2. An outer array is chosen by considering the tolerances and noises.
The process is presented in Figure 7.5. The number of total experiments is
N in u N out .
Step 3. By using the outer array, < (b) of Equation 7.16 is evaluated for
each row of the inner array. Define K { < (b) .
Step 4. The one-way table as shown in Table 6.16 is made. For mbij , j = 1, 2,
3 in Table 6.16, the smallest level oi , i=1,…,4 is found for the design
variable bi .
Step 5. The design solution is bi , i = 1, …,4, which has the oi th level. With
It is noted that the estimated value can be utilized as in Section 6.5. That is, the
solution is arranged in ascending order and the best one is included in the set in
Step 5 for comparison.
Solution
Design variable
Exp. no. < (b)
b1 b2 b3 b4
1 í1 í1 í1 í1 17.32
2 í1 0 0 0 5.13
3 í1 1 1 1 í2.88
4 0 í1 0 1 9.25
5 0 0 1 í1 í11.12
6 0 1 í1 0 12.17
7 1 í1 1 0 33.00
8 1 0 í1 1 16.29
9 1 1 0 í1 í5.17
Robust Design 409
Table 7.7. The outer array for the first row of Table 7.6
Level
Design variable
1 2 3
b1 6.53 3.44 14.71
1 0 1 1 í1 í13.18 Satisfied
2 1 1 1 0 í12.15 .
. . . . . . .
. . . . . . .
. . . . . . .
81 1 í1 í1 í1 72.17 .
Example 7.6 [Robust Design of the Three Bar Truss Using the Taguchi
Method]
In Example 6.15, the levels of the design variables are shown in Table 7.10 and the
tolerances are 0.00002 d z1 , z 2 d 0.00002, 0.000002 d z3 d 0.0000015. Solve
the following problems:
(1) Use the above method with five steps for robust design.
(2) Find a design solution from the estimated values in ascending order,
compare it with that from (1), and find the final solution for robust design.
Solution
Level
Design variable
1 2 3
b1 0.0001 0.0002 0.0004
b2 0.0001 0.0002 0.0003
b3 0.000015 0.00002 0.00003
Robust Design 411
Level
Design variable
1 2 3
b1 79.93 1.92 1.11
b2 32.28 27.35 23.33
compared with the best one in the orthogonal array. The fifth row is the
best in Table 7.11 and it is better than that from the one-way table.
Therefore, the fifth row of Table 7.11 is selected as the final solution,
which is b = [0.0002 0.0002 0.000015]T . The mean and variance of the
objective function are 1.396 kg and 0.00694, respectively.
(2) Design by the estimated values in ascending order
The characteristic function is estimated in the same way as in Example 7.5.
The estimated values are arranged in ascending order and the constraint
violation is examined from the first row. Since the first row satisfies the
constraints as shown in Table 7.13, it is selected as the design solution.
The mean and variance of the objective function are 1.415 kg and
0.006941, respectively. Since the design is obtained from the first row, the
confirmation experiment is not performed from the second row. The
characteristic function is 0.711. Therefore, the solution of (1) is chosen as
the final solution since it is better.
Equation 7.17 is useful when the factors can be divided into the scale factor that
affects the mean, and the control factors that affect the variance. When this
division is not possible, the solution might not be very good because the effects on
the mean and the variance are compounded in the S/N ratio. That is, the two-step
optimization method proposed by Taguchi is valid when the factors can be divided.
It is noted that the method of the previous section, which uses computer simulation,
uses the one-step method. Phadke defined qualification of the characteristic
function (Phadke 1989). However, it is quite difficult to use in practical design.
The mean and variance are compounded in the S/N ratio in the cases of the
smaller-the-better and the larger-the-better types as well. Therefore, the direct use
of the S/N ratios in Table 7.3 may lack pertinence. It is the same for the
characteristic functions for the smaller-the-better and the larger-the-better types.
Therefore, when the factors are not obviously divided, appropriate solutions may
not be obtained from the S/N ratio. Box proposed a general transformation
approach (Box 1988). However, this solution may not be precise for highly
nonlinear cases. Montgomery (Montgomery 1991) and Leon (Leon et al. 1987)
showed that minimizing the variance is more effective than using the S/N ratio.
Lee (Lee et al. 1996) recommended the use of a multiobjective function with
weighting factors. Wu evaluated robustness for the nominal-the-best type problem
with the indices made by the S/N ratio, mean standard deviation (MSD) and the
variance (Wu and Chen 2002).
Second, the experiment of each row of an orthogonal array is carried out
repeatedly to consider noise. In a design using computer simulation, the outer
array is employed for this. Since many experiments are needed, they can be
inefficient. Welch (Welch et al. 1990) and Wu (Wu and Chu 2002) proposed the
use of a combined array that unifies the inner and outer arrays.
Third, the number of characteristic functions is one in the original Taguchi
method. Phadke proposed a method by using trade-off for the problem with
multiple characteristic functions (Phadke 1989). However, this method can
increase the uncertainty for robustness. Shiau solved the problem with multiple
characteristic functions by defining the weighting factors. It is noted that the
weighting factors are subjectively determined by the user (Shiau 1990). There are
many methods to solve the problem (Tong and Su 1997, Chen 1997, Vining 1998,
Tsui 1999).
Fourth, the Taguchi method cannot consider many constraints that are
generally considered in design. To overcome this problem, Lee proposed the
following steps (Lee and Park 2001): (1) An optimum solution is obtained by a
classical optimization method regardless of the tolerances. It is considered that the
target is obtained. (2) Candidate values are defined around the optimum solution.
An orthogonal array is defined with the candidate values and the method of the
previous section is utilized. Recent research on the Taguchi method has been
carried out to overcome the difficulties of the classical Taguchi method.
414 Analytic Methods for Design Practice
Level
Design variable
1 2 3
b1 0.0001 0.0002 0.0003
Example 7.7 [Robust Design of the Three Bar Truss by Using the
Taguchi Method after Optimization]
The optimum of Example 7.6 is b* = [0.0002034 0.0001064 0.00000001]T .
The candidate values around the optimum are defined as shown in Table 7.14. The
tolerances are 0.00002 d z1 , z 2 d 0.00002 , 0.000000002 d z3 d 0.000000002 .
Solve the following problems by using the inner and outer arrays.
(1) Carry out robust design by using the steps in Section 7.3.3.
(2) Arrange the solutions in ascending order by using the estimated values.
Evaluate the best value and obtain the final solution by comparing it with
the solution of (1).
Solution
Design variable
Exp. no. Ȍ (b)
b1 b2 ignored b3
1 í1 í1 í1 í1 109.425
2 í1 0 0 0 102.359
3 í1 1 1 1 102.418
4 0 í1 0 1 109.576
5 0 0 1 í1 5.112
6 0 1 í1 0 1.617
7 1 í1 1 0 109.778
8 1 0 í1 1 5.300
9 1 1 0 í1 0.799
Robust Design 415
Level
Design variable
1 2 3
Three levels for each design variable are considered. The L9 (34 )
orthogonal array is selected for the inner and outer arrays. The outer arrays
for each row of the inner array are evaluated. The characteristic function
K { < (b) is calculated from the mean P f and the variance V f . The
results are shown in Table 7.15. The weighting factors and the scale factor
are the same as those of Example 7.6.
The one-way table is made as shown in Table 7.16. The solution from
Table 7.16 is (3 3 1) levels. The design variables are b =
[3.0E 4 1.5E 4 5.0E 9]T .
The confirmation experiment is carried out with the determined levels.
The mean is 1.591 kg, the variance is 0.00691 and the characteristic
function is 0.799. The constraints are satisfied. The solution is compared
with the best one from the orthogonal array. The ninth row of the
orthogonal array is the best and it is the same as the one previously
obtained. Thus, it is selected as the solution.
(2) Arranging the solutions in ascending order using the estimated values
The estimated values for the objective function and the maximum
constraint violation are calculated for the full combination. They are
arranged in ascending order as shown in Table 7.17 and the confirmation
experiment is carried out from the first one. The second row satisfies the
constraints and it is selected as the solution. The mean is 1.591, the
variance is 0.00691 and the characteristic function is 0.799. Since this is
the solution of (1), it is selected as the final solution. The confirmation
experiment is not carried out from the third row of Table 7.17.
7.4.1 Introduction
The aforementioned robust design problems have been solved based on well
developed mathematical optimization techniques during the last decade. Since the
Taguchi method uses an orthogonal array and design variables are defined in a
discrete space, it is difficult to treat a wide design range. Design in a discrete
space may be useful; however, design in a continuous space is often required.
Generally, a design may have many design constraints that the Taguchi method
may not resolve. Because optimization techniques can easily handle the
constraints, robust optimization based on optimization techniques has been studied.
In a general design problem, we can hardly separate the mean and variance of
design variables. Optimization techniques can obtain design variables by
considering the mean and the variance simultaneously. The goal of robust
optimization is to solve these problems.
Robust optimization is formulated as Equation 7.2. The functions F and G in
Equation 7.2 are defined to enhance robustness and the robustness of functions is
defined as follows: (1) Robustness of the objective function can be achieved by
C
B
A
b
reducing the change of the objective function with respect to the changes of
tolerances for the design variables. Figure 7.6 presents an objective function that
has a minimum at the peak point A, where the objective function can drastically
change by small variations of the design variables. Thus, the local minimum point
B or C is better for an insensitive (robust) design. (2) Robustness of constraints
means that all the constraints are satisfied within the range of tolerances for the
design variables, i.e., the feasible region is reduced as illustrated in Figure 7.7.
Probabilistic consideration of the objective function and constraints is known as
reliability optimization or stochastic optimization. They will be briefly mentioned
later.
Now let us consider the formulation of robust optimization as follows:
Find b Rn (7.18a)
to minimize [ P f (b, p) V f (b, p)] (7.18b)
bL d b zb d bU (7.18d)
It is noted that the objective function in Equation 7.18b is defined in terms of the
mean and the standard deviation, i.e., the variance. Equation 7.18c means that all
the constraints should be satisfied despite the noise in design variables and
parameters. Therefore, robust optimization methods have been efficiently
developed to solve the problem defined as Equation 7.18.
g 1, new
g1 Robust
b2 g2,new
feasible region
'b2 g2
b*
'b1
b1
Robustness of the objective function has been emphasized more than that of
constraints because an insensitive design of the objective function is pursued in
robust design. The mean P f and the variance V f of the objective function are
approximated as follows:
P f # f (ȝ b , ȝ p ) (7.19a)
2 2
n § wf · 2 o § wf · 2
V 2f # ¦ ¨¨ ¸¸ V b ¦ ¨¨ ¸¸ V p (7.19b)
i 1© wbi i 1© wpi
i i
¹ ¹
where ȝ b is the mean vector of design variable vector b, ȝ p is the mean vector of
the design parameter vector p, V bi is the standard deviation of bi and V pi is the
standard deviation of pi . If b and p are assumed to have normal distribution, the
mean P f of Equation 7.19a can use the nominal value of the objective function
during the optimization process. Note that Equation 7.19 is meaningful when the
interactions of the design variables and the design parameters are negligible, and
their deviations are small.
If we assume that the design variable vector b and the parameter vector p are
symmetrically distributed around the nominal values, then f(b) in Equation 3.1b
can be directly used for the mean P f in Equation 7.19a. Thus, the mean can be
easily included in the optimization process. It has been reported that the variance
is the same as the square mean of the forward difference, backward difference and
central difference, which mean the design sensitivity. It is noted that Equation
7.19 can be used based on the assumption that the design variables and parameters
are statistically independent and their distribution ranges are small.
The variance is made of the sensitivity, i.e. first derivative, as shown in
Equation 7.19b. If the objective function includes the sensitivity, second-order
derivatives of the objective are required in the optimization process. However,
calculation of the second-order derivatives of the objective function is so
expensive that it is not recommended. Since the constraints of Equation 7.18c are
represented in terms of the variance, second-order derivatives of the constraints are
also needed. Thus, robust optimization without the second-order derivatives is
recommended.
In the simplest method, the objective function is defined by the sensitivity and
the second-order derivatives are directly evaluated. This method is quite
expensive but reliable. It has some limits in that the solution may include local
information. In the early research on robust optimization, robustness was obtained
by minimizing the sensitivity information of the objective function. In this case,
we also need the second-order derivatives of the objective function. When the
weight is the objective function in structural optimization, the objective function is
Robust Design 419
an explicit function of the design variables and the second-order derivatives are
easy to evaluate. However, robustness of the weight may not be meaningful in
engineering.
In some research, the second-order derivatives are used for robustness of the
objective function and only first-order derivatives are used for robustness of
constraints. This will be explained later. Calculation of the second-order
derivatives can be viable as computer capacity increases. However, we still have
difficulties with large-scale problems.
Various methods have been proposed in order to avoid calculation of the
second-order derivatives. The sensitivity index (SI), which represents the variance
of the worst case of the current design, has been proposed. The worst case is
found by orthogonal arrays to save computational time (Sundaresan et al. 1995).
The objective function can be approximated by the regression model, which is an
explicit function. In other words, the function is approximated to a quadratic
equation by using RSM. Then evaluation of the second-order derivatives is quite
easy. Many researchers have utilized RSM to approximate the objective function.
When a function is approximated by RSM, the sensitive part (the part where
the first-order derivative is large) can be ignored. In this case, it is noted that the
sensitive part can be determined as the final solution. Robust optimization can be
obtained by the direct use of the Taguchi method, as explained in the previous
section. An outer array is utilized for each row of the inner array to calculate the
mean and the variance. In this method, sensitivity calculation is not required.
1.1 x
E (b1 ) ³ dx 1 (7.20)
0.9 .2
0
The means of other design variables are obtained in the same manner, ȝ b =
[1 1 1 1]T . Using Equation 7.19a, the mean of the objective function is
P f # f ( P b ) = í19 (7.21)
1.1 (bi 1) 2
V b2i ³ dbi 0.00333, i 1, ..., 4 (7.22)
0.9 0.2
420 Analytic Methods for Design Practice
wf wf
2b1 5 = í3, 2b2 5 = í3, (7.23a)
wb1 wb2
wf wf
4b3 21 = í17, 2b4 7 = 9 (7.23b)
wb3 wb4
2
n § wf · 2
V 2f # ¦ ¨¨ ¸¸ V b = (9 + 9 + 289 + 81)Ý0.00333 = 1.292 (7.24)
i 1© wbi
i
¹
Example 7.9 [Robustness of the Objective Function for the Three Bar
Truss]
Suppose each variable of Example 6.15 has the value of the second level. The
tolerances are 0.00002 d z1 , z 2 d 0.00002 , 0.000002 d z3 d 0.0000015 and
they have uniform distributions. Calculate the mean and the variance of the
objective function in Equation 7.19.
Solution
The design variables are uniformly distributed. If we use the equations in Example
7.8, the mean is ȝ b = [0.0002 0.0002 0.00001975]T . The mean of the
objective function by using Equation 7.19a is P f = 1.415 kg.
The variance of each design variable is
2.2 E 4 ( x 2.0 u 10 4 ) 2
V b21 , V b22 ³ dx 1.333 u 10 10 (7.25a)
1.8 E 4 4.0 u 10 5
2.15 E 5 ( x 1.975 u 10 5 ) 2
V b23 ³ dx 1.021u 10 12 (7.25b)
1.8 E 5 3.5 u 10 6
wf wf wf
2 U = 3917.372, U = 2770, 2 U = 3917.372 (7.26)
wb1 wb2 wb3
2
n § wf ·
V 2f # ¦ ¨¨ ¸¸ V b2 (7.27)
i 1 wb ©
i
i ¹
As mentioned earlier, the constraints can be transformed into Equation 7.18c in the
case where noise exists. Treatment of constraints varies according to the detailed
definition of Equation 7.18c. First, we can reduce the feasible region. The
feasible region is reduced by the noise as illustrated in Figure 7.7. In other words,
the constraints in Equation 7.18c must be satisfied even though we have z b and
z p . Generally, z b and z p are considered as tolerances. These can be defined by
ranges or distributions.
When noise exists as distribution, Equation 7.18c can be expressed as
g j , new { g j kV g2 j d 0 (7.28)
2 2
n§ wg j · 2 o § wg j · 2
V g2 j # ¦ ¨¨ ¸ V b ¦¨
¸ i i 1¨ wp
¸ Vp
¸ (7.29)
1© wbi ¹
i
i © i ¹
If the noise is given by a range, the worst case of g j , new and Equation 7.18d
become the following:
n wg j o wg j
g j ,new g j k bj ¦ | zib | k jp ¦ | zip | (7.30a)
i 1 wbi i 1 wpi
b L d b z bmax d b U (7.30b)
where | z ib | and | z ip | denote the maximum values of the tolerance ranges, and
k bj and k jp are user defined constants depending on the design purpose.
Equations 7.29 and 7.30a have the first-order derivatives of constraints. Thus, the
second-order derivatives of constraints are needed if mathematical programming is
422 Analytic Methods for Design Practice
In addition to the methods mentioned above, g j , new can be defined based on the
theory of tolerance design, but the concept is very similar.
Probabilistic analysis is adopted from the theories of reliability optimization. It
is also being used in design that considers uncertainties. Satisfaction of constraints
can be written as probabilistic forms for normal distribution of noise as follows:
P[ g j (b z b , p z p ) d 0] t Pj ,0 j 1,..., m (7.32)
P[ g j (b z b , p z p ) d 0] b p b p
³ d ( z , z ) dz dz , j 1,..., m (7.33)
g j ( b z , p z p )d 0
b
g j (b , p )
Vgj Ij(pj)
1 2 1 2
³ e 2T 2
dT t ³ e 2t 2
dt (7.34)
f 2S f 2S
where I j ( p j ) is the variable for the probability p j of the normal distribution and
is given as
Robust Design 423
g j (b z b , p z p ) g j (b, z )
Ij(pj) (7.35)
V gj
From the inequality of Equation 7.34, Equation 7.32 can be modified as follows:
g j , new { g j I j ( p j )V g2 j d 0 (7.36)
Equation 7.36 provides the basis for Equation 7.28. Equation 7.36 can be
employed in robust optimization; however, Equation 7.28 is used more often.
b1 2 b2 2 b3 2 b4 2 b1 b2 b3 b4
g1 = 1 d 0 (7.37a)
8
b1 2 2b2 2 b3 2 2b4 2 b1 b4
g2 = 1 d 0 (7.37b)
10
2b1 2 b2 2 b3 2 2b1 b2 b4
g3 = 1 d 0 (7.37c)
5
The sensitivities of the constraints are shown in Table 7.18. The variance of
each design variable has been obtained in Example 7.8. Using Equation 7.29, the
variance of each constraint can be obtained as follows:
2
n § wg ·
V g21 # ¦ ¨¨ 1 ¸¸ V b2i = 0.001041 (7.38a)
i 1© wbi ¹
w w w w
Constraint
wb1 wb2 wb3 wb4
When the noise has ranges, Equation 7.30a is utilized. The constraints are
V1
g1 d 0 , g 2 = 1 d 0 (7.41b)
100 MPa
V2 V 2
g3 = 1 d 0 , g4 = 1 d 0 (7.41c)
140 MPa 100 MPa
V3 V 3
g5 = 1 d 0 , g6 = 1 d 0 (7.41d)
140 MPa 100 MPa
w w w
Constraint
wb1 wb2 wb3
The variance of each constraint has been obtained in Example 7.9. Using
Equation 7.29, the variance of each constraint is
2
n§ wg1 · 2
V g21 # ¦ ¨¨ ¸¸ V bi = 0.00267, V g2 = 0.00522 (7.42a)
1© wbi ¹
2
i
When the noise has the range, Equation 7.30a is utilized. The constraints are
P f (b, p) V f (b, p)
f new D (1 D ) (7.45)
P *f V *f
where P *f and V *f are the base values for the mean P and standard deviation V
of the objective function, respectively. They are used for normalization and
usually serve as the starting values of the optimization process. The weighted sum
method is known to be inefficient, but it is popular due to its simple application.
One type of multiobjective optimization method is goal programming.
Designers determine the design goal, transform the goal into goal constraints, and
then optimize them with real constraints by means of nonlinear programming. A
goal constraint is another expression of “an objective function must achieve the
design goal.” The compromise decision support problem is a modified version of
goal programming for robust optimization (Mistree et al. 1993). A compromise
decision support problem is a hybrid multiobjective optimization technique that
incorporates both traditional mathematical programming and goal programming.
The system constraints must be retained without compromise while objectives are
properly compromised.
The objective function in a compromised decision support problem is a
deviation function x(d i , d i ) in terms of deviation variables. Therefore, a robust
optimum can be found by minimizing the deviation variables while the constraints
are satisfied. The compromise decision support problem with t multiobjective
functions is formulated as follows:
f j (b, p) d j d j g jf (7.46d)
Robust Design 427
d j d j 0 (7.46e)
d j t 0, d j t 0 , j 1,..., t (7.46f)
where w j is the weight and g jf is the goal that the jth objective function achieves.
d j and d j are the deviation variables that represent the distance between the
target level and the actual attainment of the goal. d j and d j indicate the degrees
of underachievement and overachievement of the goals, respectively. As a result,
we can achieve the goal as much as possible by minimizing these deviations. In
contrast to traditional goal programming, the compromise decision support
program uses real design constraints as shown in Equation 7.46.
Various deviation functions are defined in a compromise decision support
program. The most general and simplest form is the weighted sum of deviations
according to the level of requirement to achieve each goal. The distance function
considering the weighted deviation vector has been used as a deviation function.
When a compromise decision support program is used in robust optimization,
both the robustness and feasibility of constraints are satisfied and various design
goals are achieved by compromising goals. Moreover, deviation variables provide
the degrees of underachievement and overachievement of the goals. This method
can obtain a Pareto solution for a nonconvex problem. There are some other
methods for dealing with multiobjective functions (Osyczka 1984, Messac 1996).
Various optimization methods have been applied to robust optimization, some
of which are successful. However, a unified method has not yet emerged and
large-scale problems have rarely been solved. More sophisticated case studies are
needed for robust optimization.
FR A DP (7.47)
Suppose that there are two candidate DPs in Equation 7.47, and the
relationships between the FR and the DP are depicted in Figure 7.8. The gradient
of the design is called the stiffness of the design. The specified tolerance of the FR
can be more easily achieved by lowering the stiffness of the system (design A).
Allowing 'DP to be larger leads to a robust design. Therefore, the process of
robust design is stated as follows:
Step 1. Select the design that has low stiffness in order to minimize the
variation of the system range.
Step 2. Minimize the bias between the target and the mean from the selected
design.
As mentioned earlier, common aspects exist between the Taguchi method and
axiomatic design in that a robust design is induced. However, different
characteristics are discovered as well. The Taguchi method emphasizes how each
parameter is determined after the conceptual design. That is, the value of each
parameter is determined from the S/N ratio of each row in an orthogonal array.
Axiomatic design emphasizes which parameters are chosen in the conceptual stage,
Robust Design 429
FR Design B
Design A
¨DP ¨DP DP
for design B for design A
because the choice of the DP is important as shown in Figure 7.8. In other words,
selection of a part out of multiple candidates is determined by axiomatic design.
Robust design in axiomatic design is obtained by minimizing the information
content. The methods to calculate the information content are explained in
Chapter 2. Appendix 2.B shows an example of how to obtain a robust design with
less information content out of the two designs satisfying the Independence Axiom.
Design range
Probability
density
function A
B
FR FR
Figure 7.9. A normal distribution and a uniform distribution in the system range
mean resides at the outside of the design range as illustrated in Figure 7.10. The
mean of A is the same as that of B. Since the standard deviation of A is smaller,
the loss function of A is smaller. However, B has more probability to reside within
the design range. Thus, the loss function ignores the design range. The
probability of success and the loss function may not reach the same conclusion
when two designs are compared.
A robustness index is proposed to overcome the above difficulties: the
weighted robustness index. A weighting function is employed to emphasize the
closeness to the target. The weighted robustness index is composed using the
design range, the weighting function and the probability density function of the
system as follows:
Probability
density Design range
function A
FR FR0 FR
Figure 7.10. The distribution of the system range when the mean is outside of the design
range
Robust Design 431
FR * '
Pr ³ W ( FR)I ( FR)dFR (7.48)
FR * '
where Pr is the weighted robustness index, ' is half of the entire design range,
FR * is the target, I (FR) is the probability density function of the system, and
W(FR) is the weighting function.
The weighting function W(FR) is defined between the upper and lower bounds
and has the maximum at the target. It is assumed to be a linear or a quadratic
function. The linear weighting function is
A1 * * *
° ' ( FR ( FR ')) ( FR ' d FR d FR )
W1 ( FR) ®A (7.49)
° 1 (( FR * ') FR ) ( FR * d FR d FR * ')
¯'
For the ideal case, the probability density function of the system range is
expressed by the Dirac delta function as follows:
I ( FR ) G ( FR FR* ) (7.50)
The Dirac delta function is infinite at the target. If we make the weighted
robustness index have a unit value (1) with the Dirac delta function, A1 in
Equation 7.49 is determined by the following equation:
FR * '
*
Pr ³ W1 ( FR)G ( FR FR )dFR 1 (7.51)
*
FR '
1 *
° ' ( FR ( FR ')) ( FR * ' d FR d FR*)
W1 ( FR) ®1 (7.52)
° (( FR * ') FR) ( FR* FR d FR * ')
¯'
A quadratic equation can be used for the weighting function. The quadratic
weighting function W2 ( FR) is calculated in a similar fashion to the case of the
linear weighting function. In Equation 7.53 A2 1 and the quadratic weighting
function is
432 Analytic Methods for Design Practice
Design range
Probability A1 or A2
density System pdf
function (b)
(a)
FR í¨ FR FR +¨ FR
Figure 7.11. Weighting functions and the distribution of the system ranges: (a) linear
function (b) quadratic function
A2
W2 ( FR) 2
(( FR FR * ) 2 '2 ) (7.53)
'
Unless the distribution is the Dirac delta function, the integration of Equation 7.51
is not guaranteed to be 1 when A1 1. However, multiple designs can be
compared by Equation 7.51 regardless of the value of A1. If we use the logarithm,
the weighted information index can be defined as follows:
1
I log 2 (7.54)
Pr
The above two weighting functions are illustrated in Figure 7.11. As shown in
Design range
Probability
Design A Design B
density
function
PA m PB FR
Figure 7.12. Distribution of two system ranges when the mean is inside the design range
Robust Design 433
Table 7.20. The mean and the standard deviation of Example 7.14
Uniform
Mean / Standard Normal distribution
distribution
deviation
a b c d
Mean (P) 15 15 30 30
Standard
5.78 4 4 6
deviation (V)
Figure 7.12, suppose the mean is inside the design range and the distributions are
the same (e.g. normal distribution). In this case, the comparison of the two
distributions reaches the same conclusion regardless of the selection of the
probability of success, the loss function and the weighted robustness index.
Design range
Probability
density c
b
function
a d
5 15 25 30 FR
Table 7.21. The probability of success, the loss function and the weighted robustness index
of Example 7.14
Uniform
Normal distribution
Index distribution
a b c d
Probability
1 0.99 0.11 0.20
of success
Robustness
0.5 0.68 0.02 0.07
index (1st order)
Robustness
0.67 0.84 0.03 0.10
index (2nd order)
weighted robustness index is calculated by using the linear and quadratic equations.
The results are shown in Table 7.21. Table 7.22 shows the probability of success
and the robustness index by using Equation 2.23 and Equation 7.54, respectively.
The probability of success is to be large and the information content and the loss
function are to be small.
In Table 7.22, the information content and the loss function show the
differences in comparisons of systems a, b and c. System a is a perfect system
from the viewpoint of the information content; however, it has twice the loss
compared to b. That is, the information content calculated from the probability of
Table 7.22. The information content, loss function and the weighted information index of
Example 7.14
Uniform
Normal distribution
Index distribution
a b c d
Information
0.00 0.02 3.24 2.31
content
Information
1.00 0.55 5.63 3.93
(1st order)
Information
0.58 0.25 4.89 3.36
(2nd order)
Robust Design 435
success only considers the design range. However, it hardly considers robustness
from the variance. The weighted information index, when the quadratic weighting
function is used, shows that the information content of a (0.58) is 2.3 times of that
of b (0.25).
The weighted information content of c is larger than that of d. This means that
the responses from d reside more in the design range. However, the loss function
of c is less than that of d. In this case, the weighted information index of c (4.89)
is larger than that of d (3.36), since only the distribution within the design range is
considered. The weighted information index is more influenced by the probability
of success than by the deviation. The means of c and d are out of the design range.
Therefore, the loss function is not appropriate in this case and the weighted
information index can prevent this weakness.
7.7 Summary
Methodologies for robust design are classified into three groups. They are
summarized as follows:
(1) The Taguchi method is appropriately defined for the design of
experiments. Many design engineers have made efforts to adopt the
robustness concept in design. However, it is difficult to find a scale
factor and the S/N ratio in design. The method should be developed
further to find a mathematically robust solution.
(2) Robust optimization employs optimization for robust design. However,
expensive processes are required for the multiobjective function and the
second-order differentiation. Therefore, a compromise solution is
obtained or some approximation methods are utilized.
(3) The second axiom of axiomatic design is used for robust design. When
the information content is minimized, robust design is obtained and the
loss function of the Taguchi method is also minimized. However, the
method is mainly applied to conceptual design and needs to be applied to
detailed design.
Each method is excellent and has many case studies. It is also true that these
methods are difficult to use in practical problems. Therefore, we need some type
of a unified method for robust design.
7.8 Exercises
0.5 d b2 d 1.5. Calculate the expectation value for the mean and the
variance of f (1, 1).
7.3 In Example 7.2, l has the mean of 1 m and the standard deviation of 0.05
m. Calculate the expectation value of the means and the variances of the
horizontal displacement G x and the vertical displacement G y .
7.4 In Exercise 7.1, b1 0.9, 1.1 and b2 0.8, 1.2 . Calculate f (b1 , b2 ) for
all combinations and evaluate the loss function and the S/N ratio in
Equation 7.15. The target is the mean of Exercise 7.1.
7.5 In Exercise 7.2, z1 0.1, 0.1 and z 2 0.1, 0.1 . Calculate f (b1 , b2 )
for all combinations and evaluate the loss function and the S/N ratio in
Equation 7.15. The target is the mean of Exercise 7.2.
7.6 In Example 7.2, l = 0.9, 0.95, 1.0, 1.05 m. Calculate the loss function and
the S/N ratio in Equation 7.15 for the horizontal displacement G x . The
target is the mean of Exercise 7.3.
Level b1 b2 b3
1 1.5 2.0 í0.2
2 3.0 4.0 0.0
3 4.5 6.0 2.0
Use the L9 (3 4 ) orthogonal array as the inner array. The tolerances for
design variables are
Robust Design 437
Level z1 z2 z3
1 í0.1 í0.1 í0.1
2 0 0 0
3 0.1 0.1 0.1
7.8 The three bar truss in Figure 3.3 has the following design problem:
Maximize [1
where [ 1 is the first natural frequency. Young’s modulus (E) is 207 GPa,
and the density ( U ) is 2770 kg/m3. l = 1 m and T 60 $ in Figure 3.3.
The levels of the design variables are
( u 10 4 m 2 )
Level
b1 b2 b3
1 0.1 0.1 0.1
2 2.5 2.5 2.5
3 5.0 5.0 5.0
Use the L9 (3 4 ) orthogonal array as the inner array. The tolerances for
design variables are
( u 10 4 m 2 )
Level
z1 z2 z3
1 í0.001 í0.002 í0.001
2 0 0 0
3 0.001 0.002 0.001
438 Analytic Methods for Design Practice
7.9 Use Exercise 6.15. The tolerances for the design variables are
( u 10 4 m 2 )
Level
za zb ze zg zi
1 í1 í1 í2 í1 í1
2 0 0 0 0 0
3 1 1 2 1 1
Find b1 , b2
subject to (1 b1 ) 3 b2 t 0
0 d bi , i 1, 2
Level b1 b2
1 í1.0 í2.0
2 1.0 0.0
3 3.0 2.0
Robust Design 439
Find b1 , b2 , b3
b1 2b2 2b3 t 0
0 d bi d 42, i 1, 2, 3
Level b1 b2 b3
1 14.0 2.0 2.0
2 24.0 12.0 12.0
3 34.0 22.0 22.0
Find b1 , b2 , b3 , b4
12 3b1 4b2 t 0 5 b3 b4 t 0
8 2b3 b4 t 0 8 b3 2b4 t 0
Level b1 b2 b3 b4
1 í2 1 í2 í2
2 0 3 0 4
3 2 2 2 6
7.13 In Exercise 7.10, suppose each variable has the value of the second level.
The tolerances are 0 . 2 d z i d 0 . 2 , i = 1, 2 and have uniform
distributions.
(1) Calculate the mean and variance of the objective function in Equation
7.19.
(2) Obtain the variance of each constraint in Equation 7.29 and make
Equations 7.28 and 7.30 when k = 10.
7.14 In Exercise 7.11, suppose each variable has the value of the second level.
The tolerances are 0 . 1 d z i d 0 . 1 , i = 1, 2, 3 and have uniform
distributions.
(1) Calculate the mean and variance of the objective function in Equation
7.19.
(2) Obtain the variance of each constraint in Equation 7.29 and make
Equations 7.28 and 7.30 when k = 5.
7.15 In Exercise 7.12, suppose each variable has the value of the second level.
The tolerances are 0 . 1 d z i d 0 . 1 , i = 1,…,4 and have uniform
distributions.
(1) Calculate the mean and variance of the objective function in Equation
7.19.
Robust Design 441
(2) Obtain the variance of each constraint in Equation (7.29) and make
Equations 7.28 and 7.30 when k = 10.
7.16 Perform robust optimization by using the results of Exercise 7.13. Use the
objective function in Equation 7.45 and the constraints in Equation 7.30.
D 0.5 in Equation 7.45.
7.17 Perform robust optimization by using the results of Exercise 7.14. Use the
objective function in Equation 7.45 and the constraints in Equation 7.30.
D 0.5 in Equation 7.45.
7.18 Perform robust optimization by using the results of Exercise 7.15. Use the
objective function in Equation 7.45 and the constraints in Equation 7.30.
D 0.5 in Equation 7.45.
References
Belegundu AD, Zhang S (1992) Robustness of Design Through Minimum
Sensitivity. Journal of Mechanical Design, Transactions of the ASME
114:213í217
Box GEP (1988) Signal-to-Noise Ratios, Performance Criteria, and
Transformation (with Discussion). Technometrics 30(1):1í17
Box GEP, Fung C (1993) Is Your Robust Design Procedure Robust? Quality
Engineering 6(3):503í514
Chen LH (1997) Designing Robust Products with Multiple Quality Characteristics.
Computers Operations Research 24:937í944
Han JS, Kwak BM (2001) Robust Optimal Design of a Vibratory Microgyroscope
Considering Fabrication Errors. Journal of Micromechanics and Micro-
engineering 11:662í671
Lee KH, Eom IS, Park GJ, Lee WI (1996) Robust Design for Unconstrained
Optimization Problems Using the Taguchi Method. AIAA Journal
34(5):1059í1063
Lee KH, Park GJ (2001) Robust Optimization Considering Tolerances of Design
Variables. Computers & Structures 79(1):77í86
Lee KH, Park GJ (2002) Robust Optimization in Discrete Design Space for
Constrained Problems. AIAA Journal 40(4):774í780
Leon RV, Shoemaker AC, Kacker RN (1987) Performance Measures Independent
of Adjustment. Technometrics 29(3):253í285
Mistree F, Hughes OF, Bras BA (1993) The Compromise Decision Support
Problem and the Adaptive Linear Programming Algorithm. In: Kamat MP (ed)
Structural Optimization: Status and Promise. Washington, D.C., pp 247í286
Messac A (1996) Physical Programming Effective Optimization for Computational
Design. AIAA Journal 34:149í158
Montgomery DC (1991) Design and Analysis of Experiments. Wiley, New York
442 Analytic Methods for Design Practice
8.1 Introduction
Various design methodologies have been discussed and each method shown to
have specific procedures. A single method can be employed for an engineering
product. However, multiple methods should be used for some products. In this
case, it is quite important to appropriately define the design scenario according to
the characteristics of the problem.
Application of design methodologies varies according to whether the design
phase is the conceptual design or the detailed design. The design phases can be
divided into conceptual design, preliminary design and detailed design. However,
division into two phases is sufficient from the viewpoint of application. The
design process can be defined differently based on the product and given
conditions. Therefore, a general definition of the design scenario is almost
impossible. In spite of this, examples are solved by using design case studies in
this chapter. The scenarios can be changed if the product is changed. Only the
design methods introduced in the previous chapters are utilized. Therefore,
different scenarios can be defined with different methods.
Axiomatic design can be exploited for conceptual design. When a problem is
identified, the Independence Axiom is applied by defining the functional
requirements and design parameters. If multiple designs, which satisfy the
Independence Axiom, are found, the Information Axiom is applied to find the best
one with the least information content. As explained in Chapter 7, the design with
the least information content is robust. A brief scenario is illustrated in Figure
2.20 for the application of axiomatic design. This scenario can be modified
according to the characteristics of the problem.
When the conceptual design process is finished, a detailed design is carried out
to determine the specific details of design. The overall process is illustrated in
Figure 1.6. In this procedure, various methods can be utilized according to the
environment. The environment can be defined according to whether the design
apace is continuous or discrete, whether the designer pursues optimization or
robust design and the analysis methods utilized. Figure 8.1 shows a scenario for
444 Analytic Methods for Design Practice
Characteristics
of the design
Design Design
space variable
Continuous Discrete
Optimization DOE
Multiple DPs
Analysis
Axiomatic
design
Static Dynamic
response response Robust Taguchi
optimization optimization optimization method
structural design using the analytic methods of this book. Examples are introduced
in this chapter and use the framework in Figure 8.1 directly or with modification.
The twin-roll strip casting process is a casting technology under which hot-rolled
coils are directly manufactured bypassing the conventional continuous casting
process and the hot-rolling process. The conventional continuous casting process
casts slabs of 200–250 mm in thickness, and produces coils through the rolling
process. The thin slab casting process has a route similar to the continuous casting
process. However, it is a process technology to cast thin slabs so that they may be
50–75 mm in thickness to decrease the load of the hot-rolling process.
Case Studies with Analytic Design Methods 445
Shell
Solidification end
Solidification End
Hot deformation
Hot Deformation
Roll nip
Roll Nip Zone
zone
Roll
Cast Strip
Cast strip
In the hot rolling process, flatness, bending and shearing of the strip cast are
generated from large external forces imposed on the rolling mill. On the other
hand, the external forces are not large in the twin-roll process because
solidification is made around the roll nip. Therefore, the physical external forces
caused by the rolling step do not have much contribution in the casting roll. Thus,
the external forces are neglected and the research is focused on the thermal stress
on the roll surface.
Figure 8.3 illustrates the axial cross section of the finite element model used to
calculate stresses. In order to reduce calculation time, the entire cross section is
not considered as illustrated in Figure 8.4. Only several parts of the axial cross
section of the roll are selected for analysis. The analysis condition of a selected
model is the plane strain condition.
It is assumed that the roll is deformed radially and is not deformed in the
direction of the circumference. The rotation condition of the roll is changed to the
rotation of the heat flux. When the molten steel comes into contact with the roll, a
thermal loading is imposed on the roll; otherwise, it is cooled down. Then this
condition is equivalent to the rotation of the roll.
When the roll temperature is calculated, initial boundary conditions are twofold.
The adiabatic condition is utilized at the faces except for the roll surface, and heat
x
z
is absorbed into the roll through the contact area with molten steel while it is
radiated into the atmosphere from the noncontact area. Examination is conducted
based on the case where the copper roll, the POSCO pilot caster of 1300 mm in
width and 1250 mm in diameter, is used to cast Type 304 austenitic stainless steel
at a speed of 45 m/min with 6.2 MW/m2 of heat flux based on a total of 20
revolutions. The time interval applied during calculation is to be one step when
the roll turns by 2q.
In order to view the trends of deformation and thermal stress, the results of
analysis are obtained by using the elastic finite element method. Since a high
temperature gradient is expected near the roll nip of Figure 8.2, the effective (von
Mises) stress may exceed the yield stress. Thus, it may be desirable to use elasto-
plastic finite element analysis. The results of analysis by the elasto-plastic finite
element method are compared with those of the elastic finite element method. In
addition, since the rotating roll has different boundary conditions depending on the
positions, transient analysis is performed.
As the molten steel makes direct contact with the casting rolls, cooling of the
rolls is extremely important. Therefore, the important variables of cooling include
the position, arrangement and the diameter of the cooling holes. In conjunction
with this, the diameter and the position of the channel of the rolls are defined as
the design variables to optimize the rolls. In addition, the internal stress of the
casting rolls exposed to a high thermal load is expected to be very high, and
accordingly, the degree of deformation will also be high. Since such stress and
deformation are caused by heat, cooling through the cooling channel reduces the
temperature of the rolls, which in turn reduces the stress and deformation to the
same level. The position and the diameter of the cooling holes are, therefore,
optimized using the stress and displacement obtained from the temperature, and
thermal deformation analysis as the constraints. As the volume of the rolls varies
according to the diameter of the cooling hole, design optimization is carried out by
designating the volume of the rolls as the objective function to make the rolls as
light as possible and to facilitate machining.
Copper alloys and nickels are used as materials. Design optimization is carried
out by applying the variation of the heat flux and the casting speed as load
conditions. The formulation of the optimization procedure is made as follows:
H d H max 3 u 10 2 (8.1d)
G d G max 3 u 10 4 (8.1e)
where D is the diameter of the cooling hole, PCD is the position of the cooling
hole, V is the stress, H is the strain and G is the displacement. They are
calculated from the finite element method. Optimization is performed with the
448 Analytic Methods for Design Practice
In the strip casting process, hot molten steel should be solidified and shaped
uniformly over the width of the roll in the rapid cooling process. The roll for strip
casting generally has a large contact area with molten steel in order to ensure high
productivity. Thus, it has a much greater diameter than that of an ordinary rolling
mill. Two aspects should be considered in the roll design of the strip casting
process. First, cooling channels within the roll should be designed with good
efficiency because the roll plays the role of the mold in this process. Second, the
roll acting as a rolling mill should be considered from the viewpoint of
deformation and thermal expansion during the casting process. However, the
external forces are not large in the twin-roll process because solidification occurs
around the roll nip. Therefore, the physical external forces caused by the rolling
step do not have much contribution to the stress distribution in the casting roll.
Thus, the external forces can be considered negligible and the research is focused
on roll cooling because the internal stress has an extremely close relationship with
the roll cooling process.
Elastic stress and deformation are examined prior to applying design
optimization, and the stress and deformation are also examined after the design
optimization is applied. Moreover, changes in the design variables are also
observed when applying design optimization. Figure 8.5 shows the temperature
distribution on the roll nip after the roll has revolved 20 turns. The copper alloy
hardened by zirconium oxide (ZrO2) precipitate is used for strip casting rolls.
Since this alloy may lose its own properties when exposed to a temperature higher
than 400 o C, the temperature should not exceed 400 o C. The results of calculation
350
315
280
Temperature 245
210
175
140
105
70
35
0
0 8 16 24 32 40 48 56 64 72 80 88 96 104 110
Time
Figure 8.6. Temperature variation of the roll surface with respect to time
show that the temperatures on the roll nip surface are less than 350q C . In
particular, the temperature of the copper alloy, which is below the nickel plating
layer of the casting rolls, is much lower than the temperature of the surface. Figure
8.6 shows the temperature variation on the surface of the casting roll when the roll
makes 20 turns.
As illustrated in Figure 8.6, the roll nip temperature is approximately 330q C
and the temperature of the location prior to the meniscus is approximately 70q C.
From the results of the temperature variation, we can see that the temperature
sharply changes while the casting rolls are revolving. Figure 8.7 indicates the
temperature distribution near the cooling channel on the roll nip.
Figure 8.7. Temperature distribution around the cooling channel at the roll nip
450 Analytic Methods for Design Practice
Table 8.1. Radial displacement due to thermal expansion at the roll nip (20 rev.)
After revolving
Nip (after 20 rev.) Before meniscus
180q
Temperature (o C) 330 104 70
Table 8.1 shows the tabulation of the temperature, deformation and stress of
the casting rolls at the roll nip. The maximum thermal expansion is approximately
0.282 mm in the radial direction, and such thermal expansion directly affects the
strip thickness. During casting, strips, excluding those of initial casting, should
almost have a flat state in the direction of the width. To achieve this, the thermal
expansion of the casting rolls should be estimated in advance, and the thermal
crown is given to the casting rolls. The amount of the crown is equivalent to the
thermal expansion of the casting rolls. The thermal crown means the shape of a
crown, which compensates the difference of expansions in the direction of the
width. The rolls continue to expand. When they reach a steady state, thermal
expansion no longer increases and periodic contraction and expansion are repeated.
Figure 8.8 shows that the position of the cooling hole is moved by the thermal
deformation of the rolls. Such an expansion has maximum deformation at the nip.
It is contracted in a region that is not exposed to the thermal load. Expansion and
contraction occur repeatedly; this phenomenon is depicted in Figure 8.9.
Figures 8.10 and 11 show the effective stress distribution at the roll nip. They
are the results from elastic analysis and elasto-plastic analysis, respectively. The
Figure 8.8. Change of the cooling channel position due to thermal expansion
Case Studies with Analytic Design Methods 451
10-4
2.823
2.701
2.578
2.456
Displacement
2.333
2.211
2.088
1.966
1.843
1.721
1.598
0 8 16 24 32 40 48 56 64 72 80 88 96 104 110
Time
Figure 8.10. von Mises stress distribution near the roll surface at the roll nip (elastic
deformation)
stress, which is used to determine whether plastic deformation has occurred, is the
effective stress. If the effective stress is greater than the yield stress, this means
that plastic deformation has occurred. In Figure 8.10, the maximum effective
stress is 1.33E09 N/m2. This is an extremely high value and does not occur in
reality. Such a high effective stress occurred because the elastic model was used.
Since the results of elastic analysis are not actual values, elasto-plastic analysis
must be carried out. Figure 8.11 illustrates the results of calculation using the
elasto-plastic model, and the value does not exceed that of the yield stress.
However, even though deformation will continue to change, there will be no more
452 Analytic Methods for Design Practice
Figure 8.11. von Mises stress distribution near the roll surface at the roll nip (elasto-
plastic deformation)
Figure 8.12. von Mises stress distribution around the cooling channel
Case Studies with Analytic Design Methods 453
cooling channel is below the yield stress. As shown in Table 8.1, the stress
distribution of the casting rolls is the highest at the roll nip and the lowest at the
place in front of the meniscus.
The stress is extremely high at the roll nip due to the thermal load and plastic
deformation that may occur. The stress and deformation can be reduced according
to the cooling process. The diameter and the position of the cooling hole are
important variables for the cooling capacity of the casting rolls.
Since the position and the diameter of the channel change the stress and
deformation, they should be appropriately selected. If the position of the channel
is closer to the surface, the surface temperature decreases and this leads to a
decrease in the stress in the rolls. On the other hand, stress increases as the
channel goes away from the roll surface. If the diameter of the channel is large,
the interval between channels is also large. An uneven temperature and stress
distribution occur. However, a smaller diameter used for the channel leads to
uniform temperature and stress distributions. If the diameter of the channel is too
small, it is difficult to make such a channel and therefore, an appropriate diameter
of the channel should be selected.
In Equation 8.1, constraints are imposed on strains and displacements in the
radial direction. The constraints in Equation 8.1c are not utilized because elasto-
plastic analysis is employed. In elasto-plastic analysis, when the stress reaches the
yield stress the stress is fixed and the strain becomes larger. Therefore, the
constraints in Equation 8.1d are utilized to limit the strains. The stress constraints
are indirectly considered. When the roll is expanded excessively, the size of the
crown should be larger to flatten the strip. Therefore, Equation 8.1e is imposed to
prevent excessive expansion of the roll. The objective function is defined by the
volume of the roll.
Table 8.2 shows the thermal load conditions. Design optimization is carried
out based on these conditions, and the results show that the diameter of the cooling
hole is approximately 0.64% of the roll diameter. The position from the origin of
the roll center is 96% of the radius. These results are presented for Case 1 in Table
8.2. Optimization is carried out based on the conditions in Table 8.2 and the
results are shown in Table 8.3.
2
Speed (mpm) Heat flux (MW/m )
Case 1 45 6.2
Case 2 60 8.4
Case 3 80 9.3
Hole
Deformation Channel
Stress (Pa) Strain diameter
(m) position (%)
(%)
Case 1 3.71E8 1.43Eí2 1.03Eí4 0.643 96.003
The velocity of the casting roll and heat flux are loading conditions, and the
stresses and displacements are constrained. The maximum effective stress is the
yield stress. The displacements increase as the loading increases. However, they
are within the ranges of the constraints. It is noted that the optimum variables are
similar in all cases.
In all cases, the optimum values are similar. The diameter and the position of
the cooling hole change at a minor level despite the changes in the heat flux and
the casting speed. As illustrated in Figure 8.13, the objective function also shows
almost the same results.
8.2.4 Summary
Stress analysis is carried out considering heat transfer and plastic deformation.
The results of the design optimization are as follows:
6.2300E-03
(m3)3)
6.2250E-03
Volume (m
Volume
6.2200E-03
6.2150E-03
6.2100E-03
0 1 2 3 4 5
Case
Case
Figure 8.13. The value of the objective function for each case
Case Studies with Analytic Design Methods 455
(1) From the viewpoint of effective stress, the roll surface may undergo plastic
deformation due to the large temperature gradient imposed near the roll nip.
However, since the compressive stresses in the direction of the
circumference are predominant, the amount of plastic deformation is
negligible.
(2) The stress around the cooling channel is lower than the yield stress of the
copper alloy. The region having contact with molten steel from the
meniscus to the roll nip shows plastic deformation and the other region of
the casting roll is within the elastic region.
(3) The diameter and the position of the cooling holes are selected as the
design variables for optimization of the casting rolls.
(4) Each design variable shows only minimal changes with the increase of
thermal loads.
In pressurized light water reactor (PWR) fuel assemblies, spacer grids are
important components to laterally and vertically support nuclear fuel rods. It is
required that nuclear fuel maintain mechanical soundness for safety and efficiency
during operation. The loss of mechanical soundness of nuclear fuel is generated
by excessive deformation and damage of a spacer grid, fuel rod, etc.
The structural grids support the fuel rods both laterally and axially with a
friction grip. Each cell in the spacer grid employs a fuel rod support system
consisting of two orthogonal sets of two dimples and a spring. The support system
Fuel rod
Grid dimple
Grid spring
not only allows the fuel rod to move through the cell axially, but also follows fuel
rod diametral changes in operation while maintaining alignment.
A schematic view of a grid set is illustrated in Figure 8.14. A spacer grid for a
nuclear fuel assembly has grid springs and opposing dimples, which contact a fuel
rod and soften any vibration impact. The impact is generated during reactor
operation and shipping. Moreover, external forces are imposed on the springs
during insertion of the fuel rod into the spacer grid. Because the strength of the
grid dimples is relatively large compared to that of the grid spring, the deformation
of the dimples is negligible.
The functions of the spacer grid assembly are as follows: First, the grid
assemblies provide both lateral and axial support for the fuel rods. The support
system allows for thermal expansion and growth of the fuel rods. It also retains
the rods with sufficient spring contact force to prevent coolant flow-reduced
vibration damage. However, it should not overstress locally at the fuel rod. In a
recent design, the spring system mitigates bending of the fuel rods. Second, the
grid maintains its spacing under both accidental and operational loading conditions
to maintain the cooling capacity of the fuel rod lattice. The pitch of the fuel rods
in the core is a parameter that has a major effect on the nuclear and
thermal/hydraulic performance of the core. Third, the grids must support the guide
thimbles to keep them sufficiently straight so as not to impede control rod insertion
under any normal or accidental conditions. Fourth, the grid makes a water channel
between the nuclear fuel rods to improve coolant mixing through the core.
During operation in a nuclear reactor, the grid springs and dimples are exposed to
intensive radiation, which causes the grid springs to lose initial stiffness. Thus, the
grid spring must permit the fuel rods to vibrate and chatter against the dimples.
Some conventional spacer grids have an additional problem in that the fuel rod
springs and dimples may not accurately locate the fuel rod at the center of a
standard cell.
The conventional design of the grid spring has been conducted for the axial and
lateral forces. Thus, the grid spring was designed to be stiff and large
displacement was not considered. However, the grid spring should satisfy not only
force but also displacement conditions. Hence, the grid design requirements are
defined in two categories: displacement and force conditions.
Spring
0.4 mm
Spacer grids
Figure 8.15. The manufacturing process and schematic of the manufacturing tolerance
Force Requirement
Throughout the lifetime of the reactor, the grid spring loses 92% of its strength due
to the radiation-induced relaxation. The changes from the radiation induced
property are strength and ductility change, stress relaxation, and growth. If they
are not considered in the design stages, the integrity of the grid can be jeopardized.
During reactor operation, the unit grid spring endures flow-induced vibration of 2
N. The spring should normally perform up to 25 N. The design of all grid
assembly parts should limit flow-induced vibration; therefore, it will not lead to
vibration-induced fatigue, fretting, and out of position problems. During shipping,
the grid spring should withstand vertical and lateral accelerations without allowing
the fuel rods to shift or the grid spring to deform permanently. Therefore, the
requirement should be larger than 25 N to include the unknown disturbances.
Considering the above design requirements, the force–displacement curve for
the spring has to be linear in the range of 0.4 mm for displacement. Moreover, the
maximum spring force is defined by 40 N.
Modification of the
Satisfied? design by engineering
No intuition
Yes
Satisfied?
No
Yes
Stop
the previous section are defined in a linear range. The spring just reaches the
plastic (nonlinear) range. Therefore, the conceptual design is carried out with
nonlinear analysis. The design is directed to have maximum flexibility and lower
maximum stress by trial and error. A spacer grid cell with imposed displacement
boundary conditions is modelled for analysis of spacer grid spring characteristics.
The geometric model is created by using I-DEAS (IDEAS 2002). Because of
symmetry, only a quarter of the structure modelled by thin shell elements is
analyzed by using ABAQUS (ABAQUS 2003).
For the spring characteristics satisfying the foregoing proposed design
requirements, the stiffness of the spring is selected as a performance measure. The
stiffness is computed by the ratio of the reaction force developed at the final load
step and the displacement. The linear range of the stiffness is also selected as a
performance measure. The stiffness is related to the maximum stress in a model
with nonlinear material property. The linearity of the stiffness is assured when the
Case Studies with Analytic Design Methods 459
Section point 1
Mises Value
3.06E-02
9.96E+01
1.67E+02
2.38E+02
3.12E+02
Figure 8.17. The von Mises stress contour after conceptual design
maximum stress does not exceed the yield stress of the spring at the load step. The
final model for the conceptual design is illustrated in Figure 8.17.
Nonlinear analysis is carried out in nine load steps and the maximum
displacement at the final step is 0.9 mm. Figure 8.17 shows the plot of the von
Mises stresses. Figure 8.18 presents the force–displacement curve for the grid
spring of the initial design.
The area A in Figure 8.17 is further investigated. The fuel rod vibrates against
area A due to the coolant flow. The relative infinitesimal motion between the two
bodies causes fretting wear on the contact surfaces of the fuel rod and the grid
60
50
Force (N)
40
Force(N)
30
20
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Deflection(mm)
Displacement (mm)
Figure 8.18. The force–displacement characteristic curve of the conceptually designed spring
460 Analytic Methods for Design Practice
Value
+0.00E+00
+5.36E+00
+1.07E+01
+1.79E+01
+2.32E+01
Value
+0.00E+00
+9.72E+00
+2.19E+01
+3.16E+01
Figure 8.19. Contact pressure configuration of the grid spring with 0.4 mm
spring. Design should be performed to minimize the wear and this can be achieved
through minimization and uniform distribution of the local contact pressure. This
design process cannot be included in the shape optimization (detailed design) since
nonlinear analysis is involved. Therefore, the design for area A in Figure 8.17 is
carried out in the conceptual design process. The contact pressure is evaluated by
ABAQUS. Figure 8.19a shows the contact pressure for the original shape of area
A. Improved shapes are discovered by performing trial and error analysis. The
final result is illustrated in Figure 8.19b.
The overall shape is determined in the conceptual design process. The detailed
design is carried out by shape optimization. A slight modification of the shape is
made. As mentioned earlier, the optimization is performed in a linear range.
Generally, the objective function in structural optimization is defined by the
weight of the structure. However, it has been proved that the structure does not
have a feasible solution for allowable stress with the given configuration in Figure
Case Studies with Analytic Design Methods 461
8.17. Therefore, the problem is formulated with the maximum stress as the
objective function.
The optimization problem is defined as follows:
where V is the stress, G is the displacement at the center of the spring and Fspring
is an external force on the spring. The displacement condition in Equation 8.2c is
from the first requirement and the force condition in Equation 8.2d is from the
second requirement. The optimization problem in Equation 8.2 is converted as
follows:
subject to V E d0 (8.3c)
G 0.4 0 (8.3d)
Fspring 40 0 (8.3e)
D
B
C
E
G F
x
Y X H
y
I
Z
z
Initial Optimum
Design variable
value value
Scale factor of the perturbation vector for
b1 0.0 0.01
y-coordinates of nodes A, B, C in Figure 8.20
Scale factor of the perturbation vector for
b2 0.0 0.14
y-coordinates of nodes D, E, F in Figure 8.20
Scale factor of the perturbation vector for
b3 0.0 í0.11
y-coordinate of node G in Figure 8.20
Scale factor of the perturbation vector for
b4 0.0 í0.11
y-coordinate of node H in Figure 8.20
Scale factor of the perturbation vector for
b5 0.0 0.54
z-coordinate of node I in Figure 8.20
Scale factor of the perturbation vector for
b6 0.0 1.61
x-coordinate of node G in Figure 8.20
Scale factor of the perturbation vector for
b7 0.0 0.94
x-coordinate of node H in Figure 8.20
Scale factor of the perturbation vector for
b8 0.0 í2.0
z-coordinate of node H in Figure 8.20
E Bound on stress 2.0 0.371
Objective
2.0 0.371
function
A quarter model for the finite element analysis is illustrated in Figure 8.20.
Eight design variables are selected. The design variables are indicated in Figure
8.20 and Table 8.4. Optimization is carried out by GENESIS (GENESIS 2004).
The initial design for the optimization process is the design result of the conceptual
design. The results of the optimization are shown in Figure 8.21 and Table 8.4.
The maximum stress exceeds the yield stress of 0.382 GPa. Because a little plastic
deformation is allowed, the design is considered acceptable. Nonlinear analysis is
performed with the optimum design and the result is illustrated in Figure 8.22. It is
noted that the optimum design is more flexible than the initial design. The
equivalent plastic stress and the strains of the initial and the optimum designs are
compared in Figure 8.23. The new design is compared to the current existing
designs by nonlinear analysis. The performance of the spring is significantly
improved.
8.3.4 Summary
(1) Design requirements for a spacer grid spring are defined for manufacturing
and working conditions.
Case Studies with Analytic Design Methods 463
x
Y y X
Z
z
in it ial design
Initial design oOptimal
p t im al design
design
80
70
60
(N)
Force (N)
50
Reaction force
40
Reaction
30
20
10
0
0 0 .1 0 .2 0 .3 0 .4 0 .5 0 .6 0 .7 0 .8
Displacement
Deflect io n (m m(mm)
)
(2) A design process is proposed for the spacer grid design in a practical way.
Conceptual and detailed designs are defined appropriately with linear and
nonlinear finite element analyses and optimization. Basically, the design is
directed to enhance the flexibility of the spring.
464 Analytic Methods for Design Practice
80
70
60
force (N)
new design
New design
50
Reaction Force
inconel AA
Inconel
40 inconel BB
Inconel
Reaction
30 zircalloy A
Zircalloy A
zircalloy B
Zircalloy B
20
10
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
Deflection (mm)
Displacement (mm)
(3) In the conceptual design, the overall shape of the grid spring is determined.
Nonlinear analysis is exploited to help the decision making process, which
is made based on engineering intuition.
(4) Shape optimization is employed for the detailed design. The maximum
stress of the spring is minimized, while the design requirements are
satisfied.
(5) Because the shape optimization is conducted in a linear range, the
nonlinear performance of the optimum design is evaluated. It is found that
the new design is much more excellent than the current designs.
Housing
Shaft Frame Actuator
z
y
finite element method. Various finite elements should be utilized for the
assembled structure as shown in Figure 8.24. The finite element model is
complicated and the model is validated by real experiments.
The finite element model of a rearview mirror is presented in Figure 8.25. It
mainly consists of shell elements. Parts such as the actuator in Figure 8.24 have
only a mass effect, thus, they are modelled by mass elements. The connections of
the parts are modelled by beam elements and the section properties of the beam
elements are determined by comparison with the experimental results. The
boundary conditions are determined by fixing the base plate in Figure 8.24. In
experiments, the displacement on the surface of the mirror is measured.
The natural frequency of each part is measured and compared with the finite
element model for validation. As illustrated in Figure 8.26, the natural frequency
is measured by an impact hammer and the fast Fourier transform (FFT) analyzer.
466 Analytic Methods for Design Practice
Impact hammer
Housing
Base plate
Frame subassembly
Charge amplifier Conditioning amplifier
FFT
Figure 8.26. Schematic view of the experiment for measuring the natural frequency of the
parts
The measured values are incorporated into the finite element model to tune the
values of the material properties, the thickness, etc.
The finite element models for the parts are modelled and the assembled model
is tuned to be close to the experiments. The schematic view of the experiment is
illustrated in Figure 8.27. An experiment of harmonic vibration is carried out. L
in Figure 8.27 is the distance from the mirror to the measuring plate and G is the
displacement on the surface of the mirror. Vibrations of 10í200 Hz are imposed.
A laser beam is projected onto the mirror and the displacement angle and the
Small laser
Laser beam
G
Rearview mirror
Jig
Vibration generator
Direction of vibration
1.5
Displacement (mm)
0.5
0
0 50 100 150 200
Frequency (Hz)
where ξ1 is the first natural frequency and ξ 2 is the second natural frequency.
Since ξ1 has more influence on stiffness, the weighting factor is larger. In order to
maximize the natural frequency, the reciprocal is minimized.
The results of topology optimization are presented in Figure 8.29. The dark
areas are recommended for stiffeners. The initial design has ξ1 = 67.20 Hz and
ξ 2 = 82.07 Hz. After topology optimization, ξ1 = 73.03 Hz and ξ 2 = 86.71 Hz.
The displacement on the surface of the mirror changes from 0.396 mm to 0.222
mm. Thus, stiffness is increased. The results in Figure 8.29 cannot be used
directly. The final stiffeners are determined by the designer.
Size optimization can be adopted for design of the rearview mirror. The design
variables are defined by the thicknesses of parts. Six of them are selected as
Figure 8.30. Finite element model of the rearview mirror for size optimization
Case Studies with Analytic Design Methods 469
4
Thickness
0
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21
Iteration
shown in Figure 8.30, which also presents the finite element model utilized in the
size optimization. The formulation is as follows:
The rearview mirror is designed by using the method given in Section 6.3. The
design variables are shown in Table 8.5. The displacement on the surface of the
mirror is minimized to maximize stiffness. The formulation is as follows:
Find A, B, C , D, E , F , G, H , I , J , K , L, M (8.6a)
to minimize G mirror (8.6b)
470 Analytic Methods for Design Practice
As shown in Table 8.5, the number of design variables is 13; their levels are
shown in Table 8.6. If we use the full factorial design we need 313 = 1,594,323
experiments (finite element analysis). The number can be considerably reduced by
using an orthogonal array. The interactions are ignored. A minimum orthogonal
array is selected by using the method in Section 6.3. Since this orthogonal array is
L27 (313 ), only 27 experiments are required. Equation 8.6b is calculated for each
row of the orthogonal array and the results are shown in Table 8.7. The one-way
Level
Level 1 Level 2 Level 3
Factor
A E = 200.0, ȡ = 1.07Ý10í10 E = 1662.40, ȡ = 1.74Ý10í10 E = 7241.73, ȡ = 2.78Ý10í10
B 1 mm 3 mm 5 mm
C 1 mm 5 mm 9 mm
D 5 mm (horizontal) No change 5 mm (vertical)
E 1mm 3 mm 5 mm
F E = 1.05Ý104 E = 2.1Ý104 E = 3.15Ý104
G 11.25 mm 14.5 mm 17.45 mm
H E = 1662.40, ȡ = 1.74Ý10í10 E = 4698.15, ȡ = 4.238Ý10í10 E = 7733.90, ȡ = 6.735Ý10í10
I 0.4Ý10í10 1.07Ý10í10 1.74Ý10í10
J 5 mm (up and down) No change 5 mm (up, down and outside)
K 6.664Ý10í11 8.33Ý10í11 9.996Ý10í11
L 1 mm 3 mm 5 mm
M 1 (right side) No change 2 (both sides)
Case Studies with Analytic Design Methods 471
table is made as shown in Table 8.8. From the table, the solution is
( A1 , B 2 , C 2 , D2 , E 3 , F3 , G3 , H 3 , I 3 , J 3 , K 2 , L3 , M 3 ) where the subscript denotes
pred
the level. From this solution, the value of G opt in Equation 6.43 is estimated as
0.2634. Because the interactions are ignored, the confirmation experiment must be
exp
performed; its result is G opt 0.2880 . Since this is better than any row of the
orthogonal array in Table 8.7, it is selected as the final solution.
Interactions are checked. Using the method in Section 6.4, the following value
is calculated:
472 Analytic Methods for Design Practice
A* B C D E* F G H* I J* K L M
1 0.3395 0.3538 0.3522 0.3512 0.3628 0.3579 0.3571 0.3576 0.3520 0.3555 0.3514 0.3523 0.3531
2 0.3551 0.3474 0.3445 0.3453 0.3451 0.3477 0.3483 0.3512 0.3500 0.3564 0.3430 0.3521 0.3530
3 0.3544 0.3476 0.3522 0.3553 0.3411 0.3433 0.3436 0.3401 0.3469 0.3371 0.3545 0.3444 0.3428
pred exp
| G opt G opt |
exp
0.085 (8.7)
| G opt |
The value in Equation 8.7 is small. Although we do not have a column for error,
the solution from the one-way table is better than any row of the orthogonal array.
Therefore, the interactions can be ignored in this problem.
Robust design is carried out with respect to the displacement on the surface of the
3
Sum of squares (10í3)
2.5
1.5
0.5
0
A B C D E F G H I J K L M
Design variables
rearview mirror. The outer array is adopted. If we assume that each variable has
three levels, the L27 (313 ) orthogonal array can be utilized for the inner and outer
arrays. Then we need 27 u 27 729 analysis. Since the number of analyses is
very large, only sensitive variables are selected. The ANOVA table is utilized for
the selection and the results are given in Table 8.9. The sensitivity is illustrated in
Figure 8.32. Sensitive variables A,E,H and J are to be used in robust design.
Other variables have the values given in the previous section.
The number of design variables is four and the tolerances are shown in Table
8.10. The following S/N ratio is used:
where ĭ i is the standard deviation V i of the ith row of the orthogonal array.
Since we have four design variables with three levels, the L9 (3 4 ) orthogonal array
is used for the inner and outer arrays. When the outer array is defined, the lower
Table 8.11. Characteristic value of the inner array for the L9 (3 4 ) orthogonal array
Level / Factor A E H J
value has level 1 and the upper one has level 3, and the current value has level 2.
Table 8.11 shows the standard deviation of the inner array. The one-way table
for robust design is shown in Table 8.12. This procedure is explained in Section
7.3.3. From Table 8.12, the solution is A1 E 2 H 2 J 2 , which is the same as the
second row of Table 8.11. Thus, A1 E 2 H 2 J 2 is the final solution.
8.5.2 The Aluminum Space Frame and the Low Speed Vehicle
z
y
x
elements for the thin space frame parts and solid elements for thick ones. Some
parts are connected to other ones through welding. For these joint parts, adjacent
nodes are shared or constrained. Material properties of the joints may be varied
due to the welding condition and can be obtained from experiments. However,
this variant is not considered here because the LSV model is under development.
Aluminum T6-6061 is used and the details of the material properties are
summarized in Table 8.13.
For structural analysis of the automobile, stiffness analysis, strength analysis and
crashworthiness analysis are usually performed. The purpose of stiffness analysis
is to improve comfort when a car is running, and the stiffness analysis is performed
by evaluating the bending and torsional characteristics. Strength analysis pursues
automobile durability. Crashworthiness is evaluated to see whether or not the
regulations are satisfied.
For a general passenger car, the loading conditions are well defined, but the
LSV does not have appropriate metrics. Stiffness and strength analyses are
performed for the LSV based on the loading conditions of a general passenger car.
Crashworthiness aspects are not considered here. The analyses are performed with
the current model. The current model is the initial design of the structure. Only
478 Analytic Methods for Design Practice
body in white (BIW) is used for structural analysis. Analysis is performed for the
initial design model in Figure 8.35.
Stiffness Analysis
The automobile body is deformed due to the energy from the engine and the
external forces from the surface of the road. The engine makes the body vibrate.
These deformation patterns are mainly bending or torsional types and cause the
passenger to feel uncomfortable. In the design stage, these problems should be
considered carefully. Static displacement responses are analyzed to find the
rigidity under a given load that causes the body to be bent or twisted. Similarly,
modal analysis finds the natural frequency of the automobile body.
In static stiffness analysis, test loads are imposed on the BIW to check bending
or torsional rigidity. Bending stiffness provides structural rigidity against a force
loaded to the side sill, which is located at the lower center parts between the front
and the rear mounts. Bending stiffness K B is obtained as follows:
y x
|
z kG
y
x
z z
y
y x x
KB P /Gz (8.9)
where P is a force to bend the BIW globally and G is the displacement in the
direction of the force ( z direction in this LSV model).
Torsional stiffness K T is obtained when torsion is applied to the front mounts
while the rear mounts are fixed. It is
KT T /Tx (8.10)
where T is the torsion to twist the BIW globally and T is a twisting angle at the
front mount in the x direction. Figure 8.36 presents the deformed shape of the
current LSV. The initial bending stiffness is 3420 N/mm and the initial torsional
stiffness is 3455 Nm/deg.
Modal analysis finds the vibration of the BIW with the corresponding natural
frequency. Global patterns such as bending and torsional modes are examined. In
general, the first bending and torsional modes should occur above at least 25 Hz
and 40 Hz, respectively. In our example, the first bending and torsional modes
occur at 25.6 Hz and 41.4 Hz, respectively. Therefore, the current LSV satisfies
the design requirements. Each mode shape is illustrated in Figure 8.37.
Strength Analysis
There are no rigorous loading conditions for the low speed vehicle yet. Therefore,
the loading conditions for general passenger cars can be utilized with some
modification. There are several load cases for strength analysis of general
passenger cars. These simulate various cases when an automobile is running on
the road. Actually, an automobile receives loads from the road through the
suspension system. This process is omitted and the loads are directly imposed on
the front and rear mounts. The BIW can be considered simply as a cantilevered
480 Analytic Methods for Design Practice
beam. This assumption makes the analysis very easy. However, since it does not
have real boundary conditions, the analysis results can be incorrect. Therefore, the
inertia relief technique is employed to evaluate the stresses in the BIW.
Basic loadings are defined to make various loading conditions. The loading
conditions are made by multiplying the basic loadings by a constant. Five of the
basic loadings are defined and they are the reaction forces for the stopping and
turning situations. It is assumed that the center of mass is located at 782.5 mm
from the front axle and 520 mm from the ground. The acceleration of gravity g is
9.81 m/s 2 . The other sizes are shown in Table 8.14. Basic loadings are calculated
from the static equilibrium and called Ri (i 1,...,5) .
First, consider the situation when a vehicle moves forward or backward and
stops at the acceleration of 1.0 g and 0.8 g, respectively. For these cases, a free
body diagram of the left hand side of the LSV is illustrated in Figure 8.38.
Therefore, 2 R1 and 2 R2 are reaction forces of the two wheels at the front and the
730 kg
1.0 g 0.8 g
520 mm
x
782.5 mm 782.5 mm
Figure 8.38. Free body diagram of the left hand side of the LSV in a forward/reverse brake
situation
Case Studies with Analytic Design Methods 481
Table 8.15. Reaction forces at wheel centers for the forward/reverse brake
R1 R2
Forward braking at 1.0 g 600 N 2980 N
Reverse braking at 0.8 g 2742 N 838 N
rear axle, respectively. At the center of mass, 730 kg is loaded in the íz direction.
Using moment equilibrium at the front and the rear mount, respectively, reaction
forces are obtained as shown in Table 8.15. Two pairs of R1 and R2 are
calculated. One pair is from the forward braking at 1.0 g and the other is from the
reverse braking at 0.8 g. The reaction forces R1 and R2 are directly applied to the
mounts and are illustrated by dotted arrows in Figure 8.38.
Second, consider the situation when a vehicle turns left with the acceleration of
1.0 g. To find the reaction forces acting at the left/right wheel, a free body
diagram is drawn as illustrated in Figure 8.39. Because there are two wheels on
the left and right sides of the vehicle, 2 R3 and 2 R4 are used instead of R3 and
R4 , respectively. Figure 8.39 represents the front side. The reaction forces for
this case are obtained by using equilibrium of the moments. When the vehicle
turns right with 1.0 g, the reaction forces are similarly obtained. Therefore, two
pairs of R3 and R4 are calculated and they are summarized in Table 8.16. Finally,
when a vehicle does not move, the reaction force at each wheel R5 is one fourth of
the total weight of 730 kg.
1.0 g 730 kg
520 mm
602.5 mm 602.5 mm
z
y
Figure 8.39. Free body diagram of the front side of the LSV for the left cornering at 1.0g
482 Analytic Methods for Design Practice
Table 8.16. Reaction forces at wheel centers for left/right cornering at 1.0 g
R3 R4
Left cornering at 1.0 g 3335 N 245 N
Right cornering at 1.0 g 245 N 3335 N
Strength analysis is a durability test for various loading conditions when a car
is running on the road. Pothole brake simulates the situation when the front
wheels fall into a pothole after a sudden stop. Pothole corner (left hand/right hand)
is the case when the wheels on the left or right hand side of the vehicle fall into it.
The ultimate vertical load case is when an automobile passes an elevated place,
and the reverse brake is for a sudden backward stop. When an automobile hits a
curb, it might climb up the curb. This is called a lateral curb strike (left hand/right
hand).ٻ
Using the basic loadings, the loading conditions are defined to simulate the
running situation and are summarized in Table 8.17. In the pothole brake
condition, loadings are applied one time in the x direction and three times in the z
direction with R1 and R2 of the forward brake situation. For both the pothole
corner and lateral curb strike conditions, basic loadings R3 and R4 are used. Two
or three times the basic loadings are applied to the specific wheels, which may
mainly resist the given loading conditions.
In the case of the left hand side pothole corner side situation, basic loadings R3
and R4 for right cornering at 1.0 g are used. The reason is as follows: when a
driver finds a pothole in the left front of the LSV, she/he actually tries to avoid it
with right cornering. However, the vehicle lands in the pothole anyway and the
left wheels fall into it eventually. Therefore, three times the reaction force R3 is
applied in the z direction to the front and the rear left wheels. Three times the
reaction force R5 is applied as the ultimate vertical load case. Finally, for the
reverse brake condition, R1 and R2 in the reverse brake situation are applied in
the z direction, and 0.8 times these values in the x direction. It is noted that the
negative sign is purposely used for the applied loading to be directed toward the
center of mass in the vehicle.
Strength analysis is performed and the results are summarized in Table 8.18.
Evaluation criteria should be used to evaluate the stress results. The yield strength
of the material is used for the pothole brake, ultimate vertical and reverse brake
loading conditions. On the other hand, the ultimate strength is compared with the
results of the pothole corner and lateral curb strike analysis. It is noted the current
LSV is safe in a running situation.
Although the vehicle is safe at all loading conditions, it is important to find
parts that have relatively high stresses. Relatively high stress occurs in the pothole
brake. In particular, the highest stress occurs in the place where the A-pillar in
Figure 8.35 meets the mount. The second highest stress occurs in the pothole
corner loading condition. The same area is weak. Additionally, the lower area of
the B-pillar should be carefully designed. In Figures 8.40 and 8.41, the
z
y
Figure 8.40. Deformation and stress concentration for the pothole brake condition
Figure 8.41. Deformation and stress concentration for the left corner loading conditions
deformation and stress concentration for these loading conditions are illustrated
and the areas with high stress are circled.
In Figure 8.42, the maximum stress for each loading condition is drawn by a
bar chart, and the allowable stresses for each safety factor are represented by lines.
At present, there are no standard methods of stress evaluation for the LSV. The
stresses are evaluated based on the safety factors. If the safety factor is 1.0, all the
maximum stresses of the current design are within the allowable ranges. However,
if the safety factor is set to 1.5, the current design does not satisfy the stress
constraint in the case of the pothole brake. When the safety factor is 2.0, the
Case Studies with Analytic Design Methods 485
ͤͦ͡
ͤ͡͡
;ΒΩΚΞΦΞ͑ΧΠΟ͑;ΚΤΖΤ͙͑;Β͚
ͣͦ͡
ͣ͡͡
ͦ͢͡
͢͡͡ ͲΟΒΝΪΤΚΤ
΄ΒΗΖΥΪ͑ΗΒΔΥΠΣ͑ͮ͑͢͟͡
ͦ͡ ΄ΒΗΖΥΪ͑ΗΒΔΥΠΣ͑ͮ͑ͦ͢͟
΄ΒΗΖΥΪ͑ΗΒΔΥΠΣ͑ͮ͑ͣ͟͡
͡
ΠΥΙΠΝΖ ΠΥΙΠΝΖ ΠΥΙΠΝΖ ΆΝΥΚΞΒΥΖ ΖΧΖΣΤ Ζ ͽΒΥΖΣΒΝ ͽΒΥΖΣΒΝ
ΓΣΒΜΖ ΔΠΣΟΖΣ ΔΠΣΟΖΣ ΧΖΣΥΚΔΒΝ ΓΣΒΜΖ ΔΦΣΓ͑Τ ΥΣΚΜΖ ΔΦΣΓ͑Τ ΥΣΚΜΖ
͙ͽ΄͚ ͙΄͚ ͙ͽ΄͚ ͙΄͚
ͽΠΒΕΚΟΘ͑ΔΠΟΕΚΥΚΠΟΤ
current design does not satisfy four conditions. Figure 8.42 shows the safety
margin for each case.
PL(3L2 4 x 2 )
EI (8.11)
48G
486 Analytic Methods for Design Practice
where P is a concentrated load (9800 N), L is the wheelbase length (1565 mm), x is
the distance (778.5 mm) from the rear axle to the position of P, and G is the
displacement of the point where P is imposed.
For torsional rigidity GJ, the vehicle is assumed to be a cantilevered beam with
torsion. The value of GJ is defined as
TL
GJ (8.12)
T
where T is a torque (980 Nm) and T is the twisted angle where T is imposed.
The stiffness data for general passenger cars are surveyed and they are
expressed by Equations 8.11 and 8.12. The survey is conducted for Korean
passenger cars whose engine sizes are less than 2000 cc. Using Equations 8.11
and 8.12, bending and torsional stiffnesses with respect to vehicle weight and
wheelbase for general passenger cars are illustrated in Figures 8.43 and 8.44.
Stiffnesses for several models are marked by diamonds. Based on these marks,
solid lines are approximated to represent the trend of stiffness with respect to the
weight and the wheelbase.
The rigidity expressions for bending ( K BS ) and torsion ( KTS ) can be
individually established with respect to the weight and the wheelbase. However,
the average values are used to define the target values for the current LSV model
as follows:
KW B
BS K BS 417263e 0.0016 X 13351e 0.002Y
K BS (8.13)
2 2
W B
KTS KTS 580134e 0.0004 X 230820e 0.005Y
KTS (8.14)
2 2
where superscripts W and B are the vehicle weight and the wheelbase,
respectively. Subscripts BS and TS are the bending stiffness and the torsional
stiffness, respectively. X and Y are the vehicle weight (kg) and the wheelbase
length (mm). Equations 8.13 and 8.14 are derived from the solid lines in Figures
8.43 and 8.44.
The rigidity for the current LSV is marked by a rectangle in Figures 8.43 and
8.44. According to Equations 8.13 and 8.14, the target stiffness for the current
LSV is determined and marked as “x” in Figures 8.43 and 8.44. It is found that the
Table 8.19. Target rigidity, stiffness and displacement for the LSV
flexural rigidity with respect to the wheelbase for the LSV is higher than the target
value, and the other rigidities are lower than the target values. Therefore, the
current LSV should be improved to some extent. The LSV model has 600 kg for
the vehicle weight, 1565 mm for the wheelbase length, 9800 N for a concentrated
load, and 980 Nm for a torque. Target values are calculated and shown in Table
1.E+07
Bending stiffness (Nm2)
1.E+06
General passenger
General passengercars
cars
LSV
LSV
Target value
Target valuefor
for LSV
LSV
Approximation considered
Approximation considering general passenger cars
general passenger cars
1.E+05
500 600 700 800 900 1000 1100 1200 1300 1400
Net vehicle weight (kg)
1.E+06
General passenger
General passengercars
cars
LSV
LSV
Targetvalue
Target valuefor
for LSV
LSV
Approximation considering
Approximation considered only 4-wheeled
general cars
passenger cars
1.E+05
1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500 2600 2700
Wheelbase (mm)
Figure 8.43. Comparison of the flexural rigidity of the LSV and general passenger cars
488 Analytic Methods for Design Practice
8.19.
The target stiffness is obtained from the assumption that this low speed vehicle
is in the general passenger car group. The bending and torsional stiffness levels
for the LSV model are 79% and 50% of the target values, respectively. Structural
optimization is employed to achieve the targets.
1.E+07
LSV
LSV
Target value
value for
forLSV
LSV
Approximation consideringonly
Approximation considered general passenger
general cars
passenger cars
1.E+06 1.E+05
500 600 700 800 900 1000 1100 1200 1300 1400
Net vehicle weight (kg)
LSV
LSV
Target value
value for
forLSV
LSV
Approximationconsidered
Approximation consideringonly
general passenger
general cars cars
passenger
1.E+06
1.E+05
1500 1600 1700 1800 1900 2000 2100 2200 2300 2400 2500 2600 2700
Wheelbase (mm)
Figure 8.44. Comparison of the torsional rigidity of the LSV and general passenger cars
Case Studies with Analytic Design Methods 489
where Ebending and Etorsion are the strain energies for bending and torsional loading
conditions, respectively. mass and massinitial are the vehicle mass and the initial
z
y
x
Nondesign
domain
Design domain
Shell element
with thickness t
(a) The original A-pillar section and the modified A-pillar section for topology optimization
(b) Nondesign domain and some design domains of the A-pillar for topology optimization
Figure 8.46. Finite element model for topology optimization of the A-pillar section
Case Studies with Analytic Design Methods 491
hT
A-pillar
where k means 130 design domains in both A-pillars. mass k and massinitial k
are
the mass of the kth design domain and the corresponding initial mass, respectively.
The mass constraint is defined for each design domain. Mass distributions are
output from the topology optimization, and some of them are selected as the
candidates for initial rib design. In Figure 8.47, some results for the 65 design
domains are illustrated. Among these results, six cases of the initial rib design are
selected and they are presented in Figure 8.48. The selection is carried out by the
designer’s judgment. Each rib design is assigned a name such as Section 2, 3, 4, 5,
6 or 7. The original A-pillar section without ribs is named Section 1.ٻ
As mentioned earlier, structural analysis is performed for every LSV model
with each rib candidate design. The stiffness increase of the six models relative to
the original model without ribs is illustrated in Figure 8.49. All sections have
better stiffness properties than Section 1. Section 5 has the best stiffness. The
bending and torsional stiffnesses of Section 5 are increased by 4.6% and 5.0%,
respectively. Therefore, Section 5 is selected for the A-pillar.
Size optimization is performed for the entire LSV model with Section 5 to
satisfy the stiffness target values. Size optimization is also performed for the LSV
with Section 1. The formulation of size optimization is as follows:
(a) Topology optimization result (left) for determination of Sections 2 (center) and 3 (right)
(b) Topology optimization result (left) for determination of Sections 4 (center) and 5 (right)
Figure 8.48. Rib design candidates from some topology optimization results
G T d 0.1415q (8.17d)
V j dV j
allowable (8.17e)
Case Studies with Analytic Design Methods 493
where G and T are the displacements of the bending and torsional loading
conditions, respectively. The limit values of G and T are taken from Table 8.19.
V j and V j allowable are the maximum and the allowable stresses for strength
analysis. Superscript j refers to each loading condition in the strength analysis.
Design variables are the thicknesses of the entire LSV members including the ribs.
The members in the front and the rear parts, which play a role in crashworthiness
performance, are not included in this optimization process.
In Figure 8.50, the history of the objective function is illustrated. Because the
current LSV model has a much lower stiffness than the target stiffness, the
ࣲ࣪࣬࣬࣡
ࣱ࣪࣬࣬࣡
एरथढढपडययࣜथऩबमफलडऩडपरय
ࣰ࣪࣬࣬࣡
࣯࣪࣬࣬࣡
࣮࣪࣬࣬࣡
࣭࣪࣬࣬࣡ ࣾडपठथपणࣜयरथढढपडयय
ऐफमयथफपࣜयरथढढपडयय
࣬࣪࣬࣬࣡
एडटरथफप࣭ࣜ एडटरथफप࣮ࣜ एडटरथफप࣯ࣜ एडटरथफपࣰࣜ एडटरथफपࣱࣜ एडटरथफपࣲࣜ एडटरथफपࣜࣳ
एडटरथफपࣜरवबडय
ͥ͢͟͡͡
ͤͦ͢͟͡ ΄ΖΔΥΚΠΟ͑ͦ
ͤ͢͟͡͡ ΄ΖΔΥΚΠΟ͑͢
;ΒΤΤ͙͑ΜΘ͚
ͣͦ͢͟͡
ͣ͢͟͡͡
ͦ͢͢͟͡
͢͢͟͡͡
ͦ͢͟͡͡
͢ ͣ ͤ ͥ ͦ ͧ ͨ ͩ ͪ ͢͡
͵Ζ Τ ΚΘΟ ͑Δ Ϊ Δ ΝΖ Τ
constraint violation is very large. Therefore, the initial values of thickness are
increased by 1 mm compared to the current LSV. In the first three design cycles,
the mass of the LSV is increased to satisfy the stiffness constraints. Because the
rib is included in the A-pillars, the initial weight with Section 5 is a little heavier
by 0.9 kg than the original LSV without the rib. However, the situation is reversed
at the optimum. When the LSV with Section 5 is optimized, weight reduction is
improved by 1.9 kg compared to the case when the LSV with Section 1 is
optimized. Therefore, the rib design obtained from topology optimization
significantly reduces the weight.ٻ
8.5.5 Summary
(1) Structural analysis is performed for the low speed vehicle, which is
currently under development. Loading conditions of the general passenger
car are used for stiffness and strength analyses. The stiffness of the LSV is
compared with that of a general passenger car group. From the results of
stress, the LSV model is found to be safe when the safety factor is 1.0. A
high stress concentration occurs where the A-pillar and the mount meet.
(2) An approximation formula is proposed to find the stiffness with respect to
the vehicle weight and the wheelbase. Using this formula, the current LSV
with 600 kg for the weight and 1565 mm for the wheelbase should have
4345 N/mm and 6927 Nm/deg for bending and torsional stiffnesses,
respectively.
(3) The stiffness levels of the current LSV model are lower than the target
values defined by the formula. Structural optimization is carried out to
satisfy the stiffness target values. First, topology optimization is used to
find parts whose stiffness should be increased. After the A-pillar is
selected to be reinforced by the rib, topology optimization is performed
again to find six candidates for the rib shape. A rib design is selected from
the structural analysis. Finally, size optimization is performed to find the
optimum thickness of the entire LSV with the rib. Although the mass of
the LSV with the rib is initially heavier than that of the LSV without the rib,
the one with the rib shows better performance in optimization. Therefore,
the rib design process defined in this study is found to be efficient.
(4) To satisfy the stiffness target values, the weight of the vehicle should be
increased by 24% of the current model. This is reasonable because the
current model is designed by engineering intuition. Nevertheless, the rib
design process can be used for other parts such as the side sill and the rear
mount to further reduce the weight. Loading conditions for stiffness and
strength analyses were derived from requirements for general passenger
cars. Actually, the size and the operation range of the low speed vehicle
are different from general passenger cars and appropriate loading
conditions for the low speed vehicle should be investigated.
Case Studies with Analytic Design Methods 495
References
ABAQUS/Standard User’s Manual Version 6.4 (2003) Hibbitt, Karlsson, and
Sorensen Inc., Providence, RI
ANSYS User’s Manual Version 8.1 (2004) ANSYS Inc., Canonsburg, PA
GENESIS User's Manual Version 7.5 (2004) Vanderplaats Research &
Development, Colorado Springs, CO
Hwang KH, Lee KW, Park GJ (2001) Robust Optimization of an Automobile
Rearview Mirror for Vibration Reduction. Structural and Multidisciplinary
Optimization 21(4):300í308
I-DEAS nx10 Help Library Bookshelf (2002) Unigraphics Solutions Inc., Cypress,
CA
National Highway Traffic Safety Administration (1998) U.S. Code of Federal
Regulations. Title 49, Part 571
Park KJ, Kang BS, Song KN, Park GJ (2003) Design of a Spacer Grid Using
Axiomatic Design. Journal of Nuclear Science and Technology 40(12):989í997
Park CM, Kim WS, Park GJ (2003) Thermal Analysis of the Roll in the Strip
Casting Process. Mechanics Research Communications 30(4):297í310
Shin MK, Shin JK, Hwang SJ, Park GJ (2006) Structural Analysis and
Optimization of a Low Speed Vehicle Body. Proceedings of the Institution of
Mechanical Engineers, Part D, Journal of Automobile Engineering, accepted
9 Application of Design Methods to Automobile
Safety Problems
9.1 Introduction
Various design methodologies are applied to problems involving design for
automobile safety. Generally, the safety of an automobile is measured by
experiments. Figure 9.1 shows a crash test and the test measures forces or
accelerations on the head, chest, legs, etc. The safety is evaluated by these
measured values. The crash test requires fairly expensive facilities and the test
itself is also expensive. Thus, sled tests are performed to simulate the crash test.
A sled test is presented in Figure 9.2. The sled is moving based on the acceleration
curve measured in the crash test and the behavior of parts is measured.
Sometimes, even the sled test can be quite costly. Therefore, computer
simulation is often utilized as a surrogate for real tests. Since computer simulation
is relatively inexpensive, it is frequently utilized when many experiments are
Figure 9.1. Crash test (courtesy of the Korea Automobile Test and Research Institute)
498 Analytic Methods for Design Practice
needed. The simulation for the crash test in Figure 9.1 generally uses nonlinear
finite element analysis and an example of a model is illustrated in Figure 9.3. The
simulation for the sled test in Figure 9.2 uses multibody dynamics and a model is
illustrated in Figure 9.4.
When a design is performed to improve safety, computational simulation is
usually employed. If nonlinear finite element analysis is used for computer
simulation, the computational cost is also high. Therefore, when a design
methodology such as optimization, which requires many function calculations, is
utilized, simulation using multibody dynamics is usually employed.
Figure 9.2. Sled test (courtesy of the Korea Automobile Test and Research Institute)
Figure 9.3. Finite element model of an automobile body for nonlinear analysis
(http://www.ncac.gwu.edu)
Application of Design Methods to Automobile Safety Problems 499
Rim
Upper column
Shear capsule Plate bend
Lower column
Upper shaft
Hub Lower shaft
Upper shaft
Upper bracket
Lower bracket
Universal joint
SD2 SD4
Force
Force
Displacement Displacement
SD6 SD7
Force
Force
Displacement Displacement
Figure 9.6. Force–displacement curves for the springs of the steering system
Displacement
⎛⎛ ⎞
2.5 ⎞
⎜ ⎜ 1 t2 ⎟
HIC = sup⎜ ∫ a (t )dt ⎟ (t 2 − t1 ) ⎟ (9.1)
⎜
t1 , t2 ⎜ ⎝ t 2 − t1 t1
⎟ ⎟
⎝ ⎠ ⎠
where a(t) is the acceleration of the center of the head, t1 and t 2 are the times
which maximize the HIC value during the impact. There is a chest injury index,
which is called CSI (chest severity index) and is defined as follows:
tend
CSI = ∫ a(t ) 2.5 dt (9.2)
tbegin
where a(t) is the acceleration of the center of the chest and t begin and t end are the
beginning and end times of the impact, respectively. There are injury indices for
the femur load and the maximum acceleration of the chest. Moreover, the
following index is used:
where C is the damping matrix. The damping matrix is frequently replaced by the
linear combination of the mass matrix M and the stiffness matrix K.
There are geometric nonlinearity and material nonlinearity. M, C and K are
functions of z. The methods to solve the differential equation in Equation 9.4 are
classified into the explicit method and the implicit method. The explicit method
can solve the equation quickly; however, it does not obtain accurate solutions.
Since large-scale problems, as shown in Figure 9.3, require a large amount of
calculation, the explicit method is exploited. Details are explained in the reference
(Cook et al. 1989).
If we solve the nonlinear finite element equation for Figure 9.3, the crash test
in Figure 9.1 is reconstructed. Nonlinear finite element analysis can be applied to
parts as well and then the behaviors due to external forces can be calculated.
where HIC and CSI are shown in Equations 9.1 and 9.2, respectively.
Design variables should be selected to solve Equation 9.5. Variables that
influence HIC or CSI are chosen as design variables. The design variables of the
airbag are the mass flow rate, the vent hole area, the firing time of the inflator, the
stiffness of the seat belt and the stiffness of the knee bolster. When a design
variable is selected for the tendency of a force–displacement curve, the scale factor
of the curve yields a design variable.
Table 9.1 shows the design variables and their levels for an airbag. Each
design variable has five levels and the initial designs are allocated to the third
levels. The initial design is adopted from the current existing design. The mass
Level
1 2 3 4 5
Design variable
Scale factor of the mass flow rate 0.7 0.85 1.0 1.15 1.3
2
Vent hole area (cm ) 26 31 36 41 46
Firing time (ms) 23 26 29 32 35
Scale factor for the seat belt 0.7 0.85 1.0 1.15 1.3
Scale factor for the knee bolster 0.7 0.85 1.0 1.15 1.3
Application of Design Methods to Automobile Safety Problems 503
flow rate and the stiffnesses of the belt and knee bolster are defined by given
curves and the scale factors are increased or decreased from 15% to 30%. In Table
9.1, the value 1 for a scale factor means the initial design, 1.3 means an increase by
30% and 0.7 means a decrease by 30%. The vent hole area and the firing time are
changed in size. The orthogonal array L25 (5 6 ) is selected and is shown in Table
9.2. The simulation of each row is carried out by using the SAFE program
Design Scale
variable Scale Objec-
Mass Vent hole Firing factor for
factor for HIC CSI tive
flow rate area time knee
seat belt function
Exp. no. bolster
Scale factor
Mass flow Vent hole Scale factor
Level Firing time for knee
rate area for seat belt
bolster
1 1859.1 2414.9 1445.1 1749.3 2393.4
2 1960.6 2218.4 1931.4 2240.6 1944.1
3 2115.0 1903.8 2126.3 2218.5 1996.7
4 2160.1 1877.9 2240.0 2105.6 1800.7
5 2020.7 1700.6 2372.7 1977.9 1980.7
Figure 9.8. Facility of the body block test (courtesy of the Korea Automobile Test and
Research Institute)
The energy absorbing steering system consists of the steering wheel and the
column as illustrated in Figure 9.5. The rim and the hub are essential parts of the
wheel. The steering column consists of four components. The first is the upper
shaft, which is attached to the wheel. The second is the lower shaft, which
translates the torque from the upper shaft to the steering gear box through the
universal joint. The third is the upper column, which is connected to the upper
shaft through ball bearings and is attached to the mounting frame of the vehicle. It
supports the upper column. The fourth is the lower column, which fastens the
lower shaft on the mounting frame.
Six rigid segments are utilized to model the steering wheel and column, as
illustrated in Figure 9.5. Nine spring–damper elements are utilized to model the
energy absorbing parts and brackets. Using the general force–displacement
506 Analytic Methods for Design Practice
function to consider the unloading and the reloading processes, the spring–damper
can realistically estimate the behavior of the energy absorbing parts and brackets.
The force–displacement characteristics of the energy absorbing parts are obtained
from experiments with the components. The body block in Figure 9.8 is modelled
by one rigid segment with the mass and the moment of inertia.
The spring–damper SD1 in Figure 9.5b represents ball bearings which connect
the upper shaft to the upper column and transmit the rotation between them. Thus,
the force–displacement characteristics of the SD1 are modelled in order to prevent
the relative displacement between the upper shaft and the upper column. The
spring–damper SD2 represents the polyacetal molding that connects the upper
shaft to the lower shaft. SD2 absorbs the crash energy through compression. The
spring–damper SD4 represents the plate bend that absorbs the crash energy
through tension. The spring–damper SD6 represents the column fitting part that
connects the upper column to the lower column and absorbs the crash energy
through compression. The spring–damper SD7 represents the rubber flange that
connects the lower shaft to the steering gear box. The force–displacement
characteristics of the energy absorbing parts are represented in Figure 9.6. The
spring–dampers SD3, SD5, SD8 and SD9 represent brackets that connect the
steering system to the vehicle structure.
The body block test is conducted in order to validate the simulation model of the
steering system. To evaluate the performance of a steering system, the test utilizes
the standard FMVSS 203 (FMVSS 2003). The standard indicates that the impact
force developed on the chest of a test dummy may not exceed 11,125 N (1,134
kgf). On impact, the dummy hits the steering system at a relative velocity of 24.14
km/h (15 mph). To correlate the analysis result to the test result, the scale factor of
the force–displacement curve is adjusted. The scale factor is the multiplier of the
1 .2
1 Test
Analysis
Normalized force
0 .8
0 .6
0 .4
0 .2
0
0 20 40 60 80
Tim e (m s)
Time (ms)
1.5
Normalizeddisplacement
displacement SD2
1
SD4
SD6
0.5
SD7
0
Normalized
0 10 20 30 40 50 60 70 80
-0.5
-1
-1.5
Time (ms)
force term in the force–displacement curve. The shape (trend) of the force–
displacement curve for the energy absorbing parts is not modified from the
variation of the scale factor. As illustrated in Figure 9.9, the analysis results agree
very well with the test results.
The first peak, the second peak and the lowest concave value of the force curve
on the body block are 1.000, 0.826 and 0.315, respectively. The values for the
force curve are normalized and the forces are below 11125 N. Displacements and
forces of the energy absorbing parts are represented in Figures 9.10 and 9.11.
After approximately 22 ms, the maximum force on the body block occurs before
breakdown of the polyacetal molding (SD2) and the column fitting part (SD6).
After the breakdown, the upper column and the upper shaft move down to the
1.2
1
SD2
0.8 SD4
Normalized force
0.6 SD6
0.4 SD7
0.2
0
-0.2 0 10 20 30 40 50 60 70 80
-0.4
-0.6
-0.8
Time (ms)
0 ms 40 ms 80 ms
Figure 9.12. Behavior of the body block and the steering system
direction of the lower part of the steering system. The force on the body block
decreases to a lower value because of the breakdown of the molding and the
column fitting part. The force on the body block increases again according to the
metal-to-metal contact of both shafts. At a time between 22 ms and 45 ms, the
spring–damper SD4 absorbs the impact energy through the tensile deformation and
prevents the upper part of the steering system from excessive movement toward
the lower part. The SD4 controls the behavior of the upper part of the steering
system; therefore, the force on the body block does not rise too much at the time of
the metal-to-metal contact.
The behaviors of the body block and the steering system are represented in
Figure 9.12. Contact between the body block and the steering system occurs at
0 ms. The body block is separated from the steering system after 80 ms. In Figure
9.12, the steering wheel penetrates into the body block. The deformation is
evaluated by the penetration and the contact force is calculated from the
deformation. Thus, the penetration length is assumed to be the displacement
(deformation) and the impact force on the body block is evaluated from the force–
displacement curve. The same size of reaction force in the opposite direction acts
on the steering wheel. This force breaks the mold so that the upper column and the
upper shaft can have rigid body motions.
Figure 9.13. The increase of the SD4 stiffness results in the increase of the first
peak and the lowest concave value, as illustrated in Figure 9.14. The second peak
is the increase after the initial decrease. This is the result due to the increase of the
lowest concave value. The trend of SD6 is similar to SD4, as illustrated in Figure
9.15. However, the increase or the decrease of the SD6 stiffness rarely alters the
trend of the force curve. The increase or decrease of the SD7 stiffness influences
the second peak of the force curve as illustrated in Figure 9.16. The trend of SD2
is opposite to that of SD7 and the softness of SD7 delays the destruction of SD2
(molding) according to the deformation increase of SD7. As represented in
Figures 9.13í9.16, the stiffness for SD2, SD4 and SD6 may be decreased and the
SD7 stiffness may be increased to reduce the first peak on the body block. In
1.2 1.2
1 1
Normalized force
Normalized force
0.2 0.2
0 0
0 10 20 30 40 50 60 70 80 DN30 DN15 ORG UP15 UP30
1.2 1.2
1 1
Normalized force
Normalized force
0.2 0.2
0 0
0 10 20 30 40 50 60 70 80 DN30 DN15 ORG UP15 UP30
1.2 1.2
1 1
Normalized force
Normalized force
0.2 0.2
0 0
0 10 20 30 40 50 60 70 80 DN30 DN15 ORG UP15 UP30
Time (ms) Stiffness
1.2 1.2
1 1
Normalized force
Normalized force
0.2 0.2
0 0
0 10 20 30 40 50 60 70 80 DN30 DN15 ORG UP15 UP30
order to reduce the second peak, the SD2 stiffness may be increased and the SD7
stiffness may be decreased. According to the above characteristics for the spring–
dampers, it is not easy to reduce both peaks of the force curve.
CB
f2
Force
f1 CA
Time
t1 t2 tm tf
Curve C A Curve C B
§ f1 · § f2 ·
¨¨ ¸¸ u t1 0 u (t m t1 ) f1 ¨¨ ¸¸ u t 2 0 u (t m t 2 ) f2
From 0 to t m
© t1 ¹ © t2 ¹
§ f1 · § f2 ·
From t m to t f
¨¨ ¸¸ u t1 0 u (t m t1 ) f1 ¨¨ ¸¸ u t 2 0 u (t m t 2 ) f2
© t1 ¹ © t2 ¹
Total 2 f1 2 f2
2
§ df ·
Fobjective (b) ³ ¨ ¸ dt (9.6)
© dt ¹
5
Table 9.6. Results using the L25 (5 ) orthogonal array
2
Exp. Design variable Normalized force § df ·
no.
³ ¨ ¸ dx
SD2 SD4 SD6 SD7 1st peak 2nd peak Valley © dx ¹
1 1 1 1 1 0.951 0.878 0.153 1.134
2 1 2 2 2 0.964 0.855 0.233 1.055
3 1 3 3 3 0.967 0.852 0.310 0.983
4 1 4 4 4 0.975 0.864 0.379 0.943
5 1 5 5 5 0.982 0.868 0.440 0.878
6 2 1 2 3 0.963 0.870 0.185 1.188
7 2 2 3 4 0.968 0.858 0.261 1.066
8 2 3 4 5 0.974 0.841 0.334 0.985
9 2 4 5 1 0.991 0.827 0.374 0.915
10 2 5 1 2 0.998 0.830 0.388 0.918
11 3 1 3 5 0.979 0.926 0.216 1.234
12 3 2 4 1 0.994 0.812 0.251 1.025
13 3 3 5 2 0.999 0.830 0.329 0.989
14 3 4 1 3 1.001 0.826 0.345 0.964
15 3 5 2 4 1.010 0.826 0.405 0.923
16 4 1 4 2 0.990 0.827 0.202 1.139
17 4 2 5 3 0.989 0.825 0.284 1.034
18 4 3 1 4 0.993 0.824 0.300 1.019
19 4 4 2 5 1.005 0.814 0.365 0.966
20 4 5 3 1 1.029 0.777 0.410 0.883
21 5 1 5 4 0.981 0.899 0.238 1.174
22 5 2 1 5 0.986 0.888 0.258 1.132
23 5 3 2 1 1.006 0.784 0.290 0.971
24 5 4 3 2 1.024 0.799 0.359 0.946
25 5 5 4 3 1.043 0.779 0.419 0.916
and the lowest concave value are 0.864, 0.987, 0.839 and 0.425, respectively. The
summation of the derivatives and the first peak are decreased by 13.6% and 1.3%,
respectively. The second peak and the lowest concave value are increased by
1.58% and 34.9%, respectively.
Generally, it is easy to obtain an improved solution for a problem with one
peak. However, a problem with two peaks is difficult to handle due to the valley.
In this case, the engineer can select an appropriate solution from the orthogonal
Application of Design Methods to Automobile Safety Problems 513
array. Suppose we select the third row in Table 9.6. The summation value is
reduced by 1.7% and the first peak is reduced by 3.3%. The lowest concave value
is reduced by 1.6% and the second peak is increased by 3.2%. This is the case
where the solution is picked from the rows of the orthogonal array by using
engineering sense.
9.4.7 Summary
Using the occupant analysis program SAFE, the characteristics of the energy
absorbing steering system are examined and improved. The analysis model is
established by comparing the analysis model and the experiments based on the
FMVSS 203 regulation. To reduce the impact load on the body block, an objective
function is proposed as the integrated value of the squared sum of the derivatives
of the load curve. An orthogonal array is employed for an improved design. A
design solution is obtained from the one-way table. Because there are two peaks, a
better solution is suggested by picking one from the rows of the orthogonal array.
In practical engineering, designers may generate multiple cases and determine the
final design from these cases. The solution here is the same as in the method
where the performances are evaluated using only orthogonal arrays.
Uniform velocity
800 mm/s
In side impact, the distance between an occupant and the impact beam is short.
Thus, enhancement of the stiffness of the impact beam is more important than
energy dissipation. If the size of the impact beam is increased, then the stiffness is
increased. However, the weight, which should be minimized, is also increased.
Therefore, an optimization problem is formulated as follows:
Find a, b, t (9.7a)
to maximize f(a, b, t) (9.7b)
subject to w < w0 (9.7c)
The impact beam has an ellipsoidal shape. a is the major axis, b is the minor axis,
t is the thickness of the impact beam, f(a, b, t) is the contact force between the
impact beam and the impactor (cylinder), w0 is the initial weight of the impact
beam, and w is the weight of the impact beam.
Generally, the impact beam has a circular shape. However, here it is defined
by an ellipsoid. The impact beam is designed using the method in Section 6.5 and
the response surface method in Section 6.7. The contact force is the performance
index for the resistance to penetration.
516 Analytic Methods for Design Practice
Using the penalty function P(b) in Equation 6.60, a characteristic function <(b)
to be maximized is defined as follows:
The response surface method is explained in Section 6.7. The objective function is
approximated by a quadratic function. The candidate points for the approximation
Design Level
variable 1 2 3
a 17.0 15.5 14.0
b 14.0 12.5 11.0
t 2.6 2.4 2.2
Application of Design Methods to Automobile Safety Problems 517
are the points of L18 (21 u 37 ) in Table 9.9 and the results in Table 9.9 are utilized.
The approximated function is
9.5.5 Summary
The cross section of the impact beam is designed to maximize the contact force
and to reduce the weight. The design process is carried out by using an orthogonal
array and the response surface method.
(1) The design results from the two methods are compared and are found to be
similar.
(2) When an orthogonal array is used for the design, the contact force is
increased by 19.6% and the weight is reduced by 10.4%. The response
surface method increases the contact force by 19.44% and reduces the
weight by 10.39%.
Application of Design Methods to Automobile Safety Problems 519
The axiomatic approach is utilized to analyze the overall design process. The
functional requirements(FRs) for the product are as follows:
FR1 : Let the system have the capability of tilt/telescopic operation.
FR2 : Let the transmission of the torque be easy.
FR3 : Let the strength of the steering column be sufficient.
DP1 for FR1 is defined because the tilt/telescopic function is enabled by the
rack and pinion gears, which are operated by a motor. The function becomes easy
to operate if we use the universal joints and splines. The positions and angles of
the two universal joints determine the steering function; therefore, they are
520 Analytic Methods for Design Practice
selected as DP2 . Since the strength of the system depends on the positions and
sizes of the parts, they are defined as DP3 .
The main interests are the safety and vibration performances. Thus, only FR3
and DP3 are explained. The lower level function requirements ( FR3 x ) for FR3
are defined as follows:
FR31 : Maintain the bending strength.
FR32 : Absorb impact energy from vehicle collision.
FR33 : Avoid the resonance caused by the combustion of the engine, the
irregularity of the road and the variation of the brake torque.
The bending strength should be maintained to keep the configuration of the
structure. If there is no airbag, the steering system directly contacts the driver.
Thus, the steering system should absorb the impact energy as much as possible.
The steering system should not vibrate excessively from the external forces during
operation. When the system is resonant, comfort is deteriorated. Therefore, DP3 x
are defined as follows:
DP31 : The positions of the support and bolt (tilt point)
DP32 : The mechanism for the energy absorption such as the bolt ejection,
tube-in/tube-out insertion, the main energy absorber and three-point
support
DP33 : The shapes, sizes and gaps of components
Because the positions of the support and the bolt for the tilt point are crucial for
FR31 , DP31 is defined as above. DP32 is selected with respect to FR32 as
follows: During impact, the bolt of the main bracket is ejected. The main bracket
is a part that is connected to the car body. The upper and lower columns are
compressed, forces are imposed on the energy absorbing parts and impact energy
is absorbed. Figure 9.21 shows the hierarchies for the functional requirements and
the design parameters.
Equation 9.10 is the overall design equation and Equation 9.11 is the design
equation for FR3 .
FR DP
Figure 9.21. Hierarchies of the FR–DP mapping for the steering system
Application of Design Methods to Automobile Safety Problems 521
ª FR1 º ªX 0 0 º ª DP1 º
« FR » «X X 0 »» «« DP2 »» (9.10)
« 2» «
«¬ FR3 »¼ «¬ X X X »¼ «¬ DP3 »¼
ª FR31 º ªX 0 0 º ª DP31 º
« FR » «0 X 0 »» «« DP32 »» (9.11)
« 32 » «
«¬ FR33 »¼ «¬ X X X »¼ «¬ DP33 »¼
where 0 means there is no relation and X means there is a relation. Equations 9.10
and 9.11 represent decoupled designs.
Design for safety and vibration reduction should be performed according to the
sequence indicated by Equation 9.11. Then the Independence Axiom is satisfied.
The detailed design is carried out to find the detailed sizes of DP3 x . Figure 9.22
presents a prototype of the steering system.
In this section, detailed design of DP32 is carried out to satisfy FR32 . The force
displacement curves for energy absorbing parts are determined. The steering
system consists of the steering wheel and the steering column. The rim and hub of
the steering wheel are modelled by a contact ellipsoid that does not have mass, as
explained in Section 9.4. Therefore, the corresponding mass and moment of
inertia are included in the steering shaft. The other parts are modelled by segments
that have the mass and the moment of inertia. The connections between parts are
modelled by slip joints or universal joints. Since the slip joint prevents the motion
in the axial direction, a spring–damper, which has a large stiffness coefficient, is
used. Figure 9.23 illustrates the model.
522 Analytic Methods for Design Practice
The force displacement curves for the spring–dampers are obtained from static
experiments and they are tuned in the modelling process. The parts in Figure 9.23
are as follows:
G S1: Ball bearing
G S2: Main bracket of the upper body
G S3: Main energy absorbing device
G S4: Spline
G S5: Screw of the telescopic motor
G S6: Tube-in and tube-out energy absorbing part
G S7: Bracket of the lower body
G S9: Universal joint
G S8, S10: Mounting parts
S11: Energy absorbing part by ejection of the bolt
G S12: Support for the telescopic housing
S1 is the spring–damper that models the ball bearing between the upper column
3 - p o in t su psupport
Three-point p o r t ( S 1 2(S12)
) F - D F-D
c u r vecurve
The energy
T h e en e rgy ababsorbing component
so rb in g co m p o n e n t (S 3 )(S3)
F -D F-D
cu rvecurve
value)
value)
͢ ͢ ͣ͟
Force/(nominal value)
Force/(nominal value)
͡ ͩ͟ ͢
Force/(nominal
Force/(nominal
͡ ͩ͟
͡ ͧ͟
͡ ͧ͟
͡ ͥ͟
͡ ͥ͟
͡ ͣ͟ ͡ ͣ͟
͡ ͡
͡ ͡ ͣ͟ ͡ ͥ͟ ͡ ͧ͟ ͡ ͩ͟ ͢ ͡ ͡ ͣ͟ ͡ ͥ͟ ͧ͟͡ ͡ ͩ͟ ͢
Displacement/(nominal
D eflec tio n /(n o m in al va lu value)
e) Displacement/(nominal
D e fle c tio n /( n o m in a l va luvalue)
e)
B olt
Bolt ejec tion(S11)
ejection (S 1 1 )F-D
F -D curve
cu rve ΆΡΡΖΣ͠ͽΠΨΖΣ͑ΔΠΝΦΞΟ͙͑΄͚ͧ͑ͷ͞͵͑ΔΦΣΧΖ
Upper/lower column (S6) F-D curve
value)
value)
͢ ͢
ͷΠΣΔΖ͙͠ͿΠΞΚΟΒΝ͑·ΒΝΦΖ͚
Force/(nominal value)
ͩ͟͡ ͩ͟͡
Force/(nominal
Force/(nominal
ͧ͟͡ ͧ͟͡
ͥ͟͡ ͥ͟͡
ͣ͟͡ ͣ͟͡
͡ ͡
͡ ͡ ͣ͟ ͡ ͥ͟ ͡ ͧ͟ ͩ͟͡ ͢ ͑͑͑͑͑͑͑͑͑͑͑͑͑͑͑ͣ͑͑͑͑͑͑͑͑͑͑͑͑͑ͥ͑͑͑͑͑͑͑͑͑͑͑͑͑ͧ͑͑͑͑͑͑͑͑͑͑͑͑͑͑ͩ͑͑͑͑͑͑͑͑͑͑͑͑͑͑͑͟͟͟͟͢͡͡͡͡͡
͟͡͡ ͟͡͡ ͤ͟͡ ͥ͟͡ ͩ͟͡ ͩ͟͡ ͪ͟͡ ͢͟͡
͵ΖΗΝΖΔΥΚΠΟ͙͠ͿΠΞΚΟΒΝ͑·ΒΝΦΖ͚
Displacement/(nominal value)
Displacement/(nominal
D efle ction /(n om in a l va luvalue)
e)
Figure 9.24. Force displacement curves for energy absorbing devices (S3, S6, S11, S12)
Application of Design Methods to Automobile Safety Problems 523
and the tilt-tube and its deformation is negligible. S2 connects the main bracket of
the steering system with the vehicle body. S3 is the main energy absorbing (EA)
device. When impact occurs, it absorbs the impact energy by elongation.
S4 connects male and female splines. S5 represents the screw for transmitting
power from the motor. It connects the telescopic housing with the motor mounted
on the tube-out and the deformation is negligible. S6 is an EA device to absorb the
1 2 3 4 5
S3 0.85 0.93 1.0 1.07 1.15
S6 0.85 0.93 1.0 1.07 1.15
S11 0.85 0.93 1.0 1.07 1.15
S12 0.85 0.93 1.0 1.07 1.15
1 2 3 4 5
S3 4.9286 4.9415 4.9938 4.9969 5.0510
S6 4.9454 4.9511 5.0048 4.9917 5.0188
S11 4.9195 4.9610 5.0265 5.0119 4.9929
S12 4.8185 4.9752 4.9602 5.0536 5.1043
1 2 3 4 5
S3 0.3 0.5 0.7 1.0 1.2
S6 0.3 0.5 0.7 1.0 1.2
S11 0.3 0.5 0.7 1.0 1.2
S12 0.3 0.5 0.7 1.0 1.2
impact energy by the relative movement between the tube-in and the tube-out. S7
and S9 are connections between the lower bracket of the steering system and the
vehicle body. S8 and S10 make the universal joint connected to front wheels. S11
is an EA device ejected from the main bracket to absorb the impact energy at the
beginning of impact. S12 is a three-point EA device connected between the main
bracket and the telescopic housing. It enhances the stability of the system.
The data for the analysis model has been tuned using the FMVSS 203 test,
which is standard for the steering system. The test is carried out at 24.14 km/h
Orthogonal
0.8
0.6
0.4
0.2
0
0 10 20 30 40 50 60 70 80
Time (ms)
Time (msec)
(15 mph) and the impact load acting on the body block must not exceed 11,125 N
(1,134 kgf). The analysis model is established by tuning the force-displacement
curves for the spring–dampers. The data are made from experiments and modified
to match the results between the FMVSS 203 test and the numerical analysis.
The objective of the design is to minimize the impact load on the body block.
The characteristics of the spring–dampers are the most important for the evaluation
of the impact load. Thus, the scale factors for the spring–dampers are selected as
design variables to minimize the impact load. Some important spring–dampers are
chosen for design variables through sensitivity analysis. They are the EA devices
such as S3, S6, S11 and S12 shown in Table 9.12. The number of levels for the
design variables is five where the stiffness is varied from í15% to 15%. The
stiffness is varied by the scale factor of the force displacement curves as illustrated
in Figure 9.24. Since we have four design variables with five levels, the L25 (5 6 )
orthogonal array in Table 9.13 is utilized. Then 25 analyses are conducted for the
design.
1 2 3 4 5
S3 4.6941 4.5404 4.6297 4.6932 4.8120
S6 4.7325 4.6647 4.6158 4.6586 4.6978
S11 4.6676 4.6217 4.5813 4.7453 4.7535
S12 4.5450 4.4723 4.5537 4.8227 4.9757
After the analysis model is established, the analysis for each row in Table 9.13
is carried out. The results in Table 9.13 are normalized. When the one-way table
in Table 9.14 is made, the solution is found to have the first levels. This is the first
row of Table 9.13 and also the best row in Table 9.13. The initial design is
compared with the solution in Figure 9.25. The peak is reduced by 7.7%.
The lowest value of the stiffness is selected. The ranges of the design variables
are expanded as shown in Table 9.15. New experiments are shown in Table 9.16
and the one-way table for Table 9.16 is shown in Table 9.17. The confirmation
experiment with the solution of Table 9.17 shows that the impact load is 0.8790
and this is compared with the results in Table 9.16. It is found that the tenth row
of Table 9.16 is the best and the impact load is 0.8499.
The values in Table 9.15 are arbitrarily expanded. It is found that the real parts
cannot be manufactured with the values in Table 9.15. Therefore, the usable
solution is the one from Table 9.13.
Part 4
Part 3
Part 1
Part 1
Part 1
Part 2
Part 3
Design variable
Tilt-tube, tilt-bracket, main-bracket
Size variables Base-bracket, housing, motor-bracket,
Column-out
Shape variables Housing, main-bracket, spline shaft male
bL d b d bU (9.12e)
Size and shape optimizations are carried out simultaneously. The first natural
frequency is increased to 27.1 Hz from the optimization process and the constraint
in Equation 9.12c is violated. In other words, a feasible region is not found from
the current configuration.
3
Initial
Magnitude (displacement)
Optimum
2
0
0 20 40 60 80 100 120 140
Frequency (Hz)
Figure 9.28. Harmonic response analysis results for the initial and optimum designs
Application of Design Methods to Automobile Safety Problems 529
Optimum value
Weight (kg) 11.1
Frequency (Hz) 30
It is noted that the peaks are considerably reduced although they are not included
in the optimization process.
9.6.5 Summary
The structure of a belt integrated seat (BIS) is shown in Figure 9.29. The
conventional seat is typically composed of the seat back, the headrest, the seat
cushion and the recliner. Traditionally, the seat back is a one piece circular frame
and is connected to the seat cushion frame by connecting screws. The recliner on
the left side of the seat cushion frame allows the seat back to rotate back and forth.
In a BIS, the seat belt is attached to the seat, while in a general passenger car it
is attached to the vehicle body. In a crash, the seat belt is located close to the
driver regardless of the occupant size and the injury level is lowered. On the other
hand, a large load is imposed on the seat by the seat belt. Figure 9.29 shows a belt
integrated seat. A belt integrated seat is composed of seven parts: the seat back
frame, the seat back plastic injection mold, the headrest, the headrest adjustment,
the recliner, the seat cushion and the seat sliding frame. The BIS has a different
structure in the seat back frame compared to the conventional seat of a general
passenger car.
The seat back frame has two hollowed rectangular frames on both sides. The
headrest frame, which supports the headrest, is inserted in the right hollowed
rectangular seat back frame allowing for easy height adjustment. Both seat back
frames are connected by a plastic injection mold to support the lower back of the
occupant. The seat cushion is constructed with a thicker panel than the
conventional one and is connected to the seat sliding frame. The lateral movement
of the seat is accomplished with electric motors located on either side of the seat
sliding frame. The back and forth rotation of the seat back is accomplished with
another set of electric motors driving the screw type gears, which provide
additional structural strength.
The specifications of both a conventional seat and a BIS are described in Table
9.22. The numbers in the table are taken from an actual product. The
specifications of height, width and thickness are based on the skeleton frame. The
weight of the BIS is approximately three times greater than that of the
conventional seat. Therefore, a method to reduce the weight of the seat should be
considered during the design phase.
According to FMVSS208, the sled test with an existing BIS is performed by the
manufacturers (FMVSS 2003). In the sled test, the Hybrid III 50% dummy is used
and the crash velocity is 27.8 mph (44.7 km/h) (Backaitis and Mertz 1994). The
deceleration pulse of the sled from the sled test is illustrated in Figure 9.30.
The multibody dynamic simulation program SAFE is used to model the sled
and the seat (SAFE 1999). In order to model the accurate behavior of the seat
during the crash, a segment element that has the mass and the moment of inertia as
illustrated in Figure 9.31 is used instead of the simple panel element. The panel
element has been used for the seat. This may be sufficient to evaluate the overall
behavior. The panel element can make a resistant force only on the orthogonal
direction of the panel. Thus, it is not accurate for seat behavior. The segment
element can generate resistant forces in any direction. Hyper-ellipsoid segments
are employed to describe the geometry of the seat more accurately. Figure 9.31
35
30
25
(G)
Deceleration(G)
20
Deceleration
15
10
0
0 10 20 30 40 50 60 70 80
-5
Time((ms)
Time ms )
Segment
Hyper-ellipsoid
The dummy and the seat are modelled by segment elements that have the mass and
ͩ͡͡
΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ͽΖΗΥ͚
Shoulder belt (left) ͩ͡͡ ΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ͽΖΗΥ͚
Shoulder belt (left)
ͨ͡͡ ΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ΚΘΙΥ͚
Shoulder belt (right) ΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ΚΘΙΥ͚
Shoulder belt (right)
ͨ͡͡
ͧ͡͡ ͽΒΡ͑ͳΖΝΥ͙ͽΖΗΥ͚
Lap belt (left)
ͧ͡͡ ͽΒΡ͑ͳΖΝΥ͙ͽΖΗΥ͚
Lap belt (left)
ͽΒΡ͑ͳΖΝΥ͙ΚΘΙΥ͚
Lap belt (right)
(kgf)
(kgf)
ͷΠΣΔΖ͙ΜΘ
ͦ͡͡ ͽΒΡ͑ͳΖΝΥ͙ΚΘΙΥ͚
Lap belt (right)
ͦ͡͡
ͷΠΣΔΖ͙ΜΘΗ͚
ͥ͡͡
Force
Force
ͥ͡͡
ͤ͡͡
ͤ͡͡
ͣ͡͡
ͣ͡͡
͢͡͡
͢͡͡
͡
͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪ͡ ͢͡͡ ͢͢͡ ͣ͢͡ ͤ͢͡ ͥ͢͡ ͦ͢͡ ͧ͢͡ ͡
͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪ͡ ͢͡͡ ͢͢͡ ͣ͢͡ ͤ͢͡ ͥ͢͡ ͦ͢͡ ͧ͢͡
΅ΚΞΖ͙͑ΞΤ͚͑
Time (ms) ΅ΚΞΖ͙͑ΞΤ͚͑
Time (ms)
the moment of inertia while the car body is modelled by simple panel elements.
Force–displacement functions are used for both elements to describe the
deformation characteristics in a crash environment. A crash pulse of the car is
obtained from a crash test and is used in the simulation. After validating the
dummy behavior with high speed photos from the sled test, the simulation model is
established.
The force curves on the anchor points for both the conventional seat and the
BIS are illustrated in Figure 9.32. The maximum forces with respect to times are
described in Table 9.23. The times to reach the maximum forces on each anchor
point are similar to those of the conventional seat. However, the time to reach the
maximum force on the left anchor point of the shoulder belt is slower by
approximately 30 ms, compared to other anchor points. This delay occurs because
the left anchor point of the shoulder belt follows the movement of the seat back
during the crash. Although the summation of forces from both anchor points is
similar in the conventional seat and the BIS, the forces are more evenly distributed
to both anchor points in the case of the BIS. Therefore, the energy is scattered in
the body of the dummy.
The acceleration curves on the head, the chest and the pelvis of the dummy for
both the conventional seat and the BIS are illustrated in Figure 9.33. The
ͦ͡
ͦ͡ ΖΒΕ͑ͲΔΔ͟
Head acc. ΖΒΕ͑ͲΔΔ͟
Head acc.
ͥͦ ͥͦ
ʹΙΖΤΥ͑ͲΔΔ͟
Chest acc. ʹΙΖΤΥ͑ͲΔΔ͟
Chest acc.
ͲΔΔΖΝΖΣΒΥΚΠΟ͙͚͑͑(G)
ͥ͡ ͥ͡
(G)
ͲΔΔΖΝΖΣΒΥΚΠΟ͙͚͑͑
ͤͦ ΖΝΧΚΤ͑ͲΔΔ͟
Pelvis acc. ͤͦ ΖΝΧΚΤ͑ͲΔΔ͟
Pelvis acc.
Acceleration
ͤ͡
Acceleration
ͤ͡
ͣͦ ͣͦ
ͣ͡ ͣ͡
ͦ͢ ͦ͢
͢͡ ͢͡
ͦ ͦ
͡ ͡
͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪ͡ ͣͤͥͦͧ͢͢͢͢͢͢͢͢͡͡͡͡͡͡͡͡ ͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪͣͤͥͦͧ͢͢͢͢͢͢͢͢͡͡͡͡͡͡͡͡͡
΅ΚΞΖ͙͑ΞΤ͚͑
Time (ms) ΅ΚΞΖ͙͑Ξ(ms)
Time Τ͚͑
maximum forces with respect to the times are described in Table 9.24. The
maximum head acceleration is similar to that of the conventional seat. However,
the time to reach the maximum is delayed by approximately 20 ms due to the
movement of the left anchor point of the shoulder belt. As a result, the maximum
chest acceleration is reduced by approximately 24% compared to that of the
conventional seat. In addition, the times of maximum accelerations for the head
and chest are more evenly distributed. Since the distance between the occupant's
chest and the steering wheel or the instrument panel is narrow, the reduction of the
maximum chest acceleration is quite important.
Figure 9.34 shows the relation between the forces and accelerations for both
the conventional seat and the BIS. The head acceleration and the force on the left
anchor points of the shoulder belt are similar. The chest acceleration and the force
on the right anchor point of the shoulder belt curves are also similar. These
similarities show that the seat belt plays an important role in the safety of the
occupant head and chest. It is important to select the proper type of the seat belt
for the BIS. An additional safety system, such as an airbag, would be more
efficient in protecting the occupants from potential injuries.
The seat belt is categorized into two types: (1) the emergency locking retractor
(ELR), which controls the web length in the belt spool only by the retractor, and
(2) the web locking retractor (WLR), which controls the web length in the belt
ͥ͡ ͨ͡͡
ͥ͡ ΖΒΕ͑ͲΔΔ͟
Head ͩ͡͡
Chest acc.
ʹΙΖΤΥ͑ͲΔΔ͟ ͤͦ ͧ͡͡
ͤͦ ͨ͡͡
Shoulder belt (left)
΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ͽΖΗΥ͚
ͲΔΔΖΝΖΣΒΥΚΠΟ͙͚͑͑(G)
ͤ͡
ͤ͡ ͧ͡͡ ͦ͡͡
(G)
΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ΚΘΙΥ͚
Shoulder belt (right)
ͲΔΔΖΝΖΣΒΥΚΠΟ͙͚͑͑
Acceleration
ͣͦ ΖΒΕ͑ͲΔΔ͟
Head
ͷΠΣΔΖ͙͑ΜΘΗ͚͑
ͣͦ ͦ͡͡
(kgf)
ͷΠΣΔΖ͙͑ΜΘΗ͚͑
(kgf)
ͥ͡͡
Acceleration
Chest acc.
ʹΙΖΤΥ͑ͲΔΔ͟
ͣ͡
ͣ͡ ͥ͡͡ Shoulder belt (left)
΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ͽΖΗΥ͚ ͤ͡͡
Force
Force
ͦ͢ ͤ͡͡ ͦ͢ ΄ΙΠΦΝΕΖΣ͑ͳΖΝΥ͙ΚΘΙΥ͚
Shoulder belt (right)
ͣ͡͡
͢͡ ͣ͡͡ ͢͡
ͦ ͢͡͡
ͦ ͢͡͡
͡ ͡ ͡ ͡
͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪ͡ ͣͤͥͦͧ͢͢͢͢͢͢͢͢͡͡͡͡͡͡͡͡ ͡ ͢͡ ͣ͡ ͤ͡ ͥ͡ ͦ͡ ͧ͡ ͨ͡ ͩ͡ ͪ͡ ͢͡͡ ͣͤͥͦ͢͢͢͢͢͢͡͡͡͡͡ ͧ͢͡
΅ΚΞΖ͙͑ΞΤ͚͑
Time (ms) ΅ΚΞΖ͙͑ΞΤ͚͑
Time (ms)
spool by the retractor and the lift bar. The modelling of a seat belt is dependent on
the web length in the belt spool: the slack of ELR between the chest and the belt is
initially 0 mm and the slack of WLR is initially í20 mm. The seat belt stiffness is
also categorized into two types: (1) low stiffness, and (2) high stiffness. The belt
stiffness is defined by the force–displacement function curve.
The types of seat belts are also divided into three types based on the slack
length: (1) í40 mm, (2) 0 mm and (3) 40 mm. The case of í40 mm slack length
has a pretensioner. A pretensioner reduces the slack at the initial stage of impact
by winding the seat belt. The fuse type belt is also considered and categorized.
The fuse type belt is constructed by sewing together certain lengths of the belt web.
As the certain stitched lengths are torn, the absorption of the load is controlled
during the crash. Four kinds of stitched length are considered. The fuse type belts
are defined by the force–displacement function curves as illustrated in Figure 9.35.
The severity of occupant injury for both the conventional seat and the BIS is
defined by the head injury criteria (HIC) and the chest severity index (CSI) for the
various seat belts. The simulations are performed for the different seat belts and
the results for occupant injuries such as HIC and CSI are shown in Tables 9.25 and
9.26. The results show a reduced injury to the head with 0 mm belt slack length
with the BIS compared to the results of the minimum (í40 mm) belt slack length.
Because the belt slack in the BIS is significantly less than that of the conventional
seat during the initial stage of the crash, the structure of the BIS is expected to
function similarly to that of the conventional seat belt with a pretensioner.
Reduced injury to the chest with minimum belt slack length occurs with the BIS.
As the stitched length in the fuse type seat belt is increased, HIC and CSI are
reduced in both cases for the conventional seat and the BIS. Therefore, it is
recommended that the travel distance of the dummy be maximized within the limit
1130 1130
Force (kgf)
Force (kgf)
272.35 272.35
(a) Belt with low stiffness (b) Belt with high stiffness
to avoid contact with the interior of the automobile. Lower injury of the occupant
occurs in the case of the low stiffness seat belt. In the case of the BIS, the ELR,
low stiffness and 0 mm slack are excellent. The WLR, low stiffness and í40 mm
slack are superior for the conventional seat.
In a rear crash, a stationary vehicle is hit by a moving vehicle from behind.
Injuries to the occupant’s spine and neck typically occur. To minimize these
injuries, the seat back frame should be stiff and easily collapse as it absorbs the
crash energy to reduce the applied load on the occupant.
A deceleration pulse curve as illustrated in Figure 9.36 is obtained from a test
and is used to establish the simulation model of a rear crash with the BIS. The test
538 Analytic Methods for Design Practice
was carried out by the BIS manufacturer. The test impact velocity is 18.4 mph
(29.6 km/h). Test regulations are not yet defined for BIS. The utilized impact
velocity is arbitrarily defined by the company for the preliminary design of the
product. The anchor points of the seat belt and stiffness are the same as the ones
used in the case of the frontal crash analysis. Figure 9.37 illustrates the behavior
of the dummy in the rear crash for both the conventional seat and the BIS. The
movement of the dummy in the conventional seat and the conventional seat itself
are similar to that of the BIS during the early stage of the collision. However, in
the conventional seat, the head displacement is greater than that of the chest during
80í160 ms. At approximately 160 ms, the chest is pushed forward as a result of
the elastic rebound of the seat back while the head continues to travel backward.
This causes the relative velocities between the head and the chest to increase,
thereby increasing the occupant’s chances of neck injury. However, in the BIS,
Application of Design Methods to Automobile Safety Problems 539
12
10
Deceleration( G )(G)
8
Deceleration
0
0 20 40 60 80 100 120 140 160
-2
Time
Time((ms)
ms )
0 ms 80 ms 160 ms
0 ms 80 ms 160 ms
the seat belt continuously restrains the dummy to the seat back and absorbs the
crash energy from the initial stage of the collision. The relative velocity between
the head and the chest is reduced due to this restraint. Thus, neck injury is
expected to be reduced.
Generally, the injury caused by a rear crash is influenced by the torque that
occurs in the neck while the upper body of the dummy is rotated backward during
the collision. The torque that occurs in the neck of the dummy during the
simulation is illustrated in Figure 9.38. The torques are similar from 0í80 ms in
both the conventional seat and the BIS. In the case of the conventional seat, the
torque occurs after 120 ms and reaches its maximum of approximately 310 kgfcm
540 Analytic Methods for Design Practice
at around 170 ms. In the case of the BIS, the torque is maintained from 80 ms and
reaches its maximum of 70 kgfcm at around 120 ms. Unlike the conventional seat,
the BIS restrains the dummy with the seat belt throughout the entire region during
rear crash. Thus, it has an important role in the safety of the occupant.
The finite element model to obtain the preliminary structural design trend is
established based on the conventional seat as illustrated in Figure 9.39 with the
commercial structural optimization program GENESIS (GENESIS 2004). The
z
y
Table 9.27. Element types and number used for the BIS model
Crash Anchor point Ave. belt load (kN) Max. belt load (kN)
direction location
x y z x y z
Left of shoulder belt 2.5 í0.9 í0.5 5.1 í1.9 í1.0
45q Right of shoulder belt 0.5 1.0 1.1 1.1 2.5 2.7
left Left of lap belt 1.2 0.6 2.3 3.3 1.5 6.1
45q Right of shoulder belt 1.0 í0.5 1.1 2.5 í1.1 2.7
right Left of lap belt 0.6 í1.2 2.3 1.5 í3.2 6.1
the right shoulder belt anchor point, the left lap belt anchor point and the right lap
belt anchor point. To consider the oblique crash of 45 degrees, the load of the
frontal crash is divided into the components in 45 degrees. Table 9.29 shows the
components of the average and maximum values. The displacement and stress
constraints for optimization are limited to 45 mm and 4.5 u 10 5 kN/m 2 .
Structural optimization as a preliminary study is performed for the four cases as
follows:
Case 1. Average belt loads for frontal impact.
Case 2. Average belt loads for multidirection impact.
Case 3. Maximum belt loads for frontal impact.
Case 4. Maximum belt loads for multidirection impact.
The optimization results are shown in Table 9.30. The design variables of the
BIS are slightly scaled to protect the manufacturer’s proprietary security and due
to the preliminary development stage of the product. However, the optimum
values are utilized to obtain the overall design trend, so the result values in Table
9.30 are sufficient. An extended member from the left seat back frame is
functionally linked with the left seat back frame using the design variable linking
Application of Design Methods to Automobile Safety Problems 543
Design variable
Initial Case 1 Case 2 Case 3 Case 4
(thickness /spring coeff.)
No. of iterations 7 11 22 16
method. All the mounting brackets for the seat are linked into one design variable.
The results of the structural optimization are summarized as follows:
(1) The thickness of the left seat frame is increased more than the others to
bear the force from the left shoulder belt.
(2) The thickness of the right recliner and the crossed member between the left
and right seat back frames are slightly increased in the case of a frontal
crash. However, these are increased remarkably in the case of a
multidirection crash. The right recliner and crossed member are expected
to be important in a practical situation such as the oblique crash.
(3) The spring coefficients of the left recliner are significantly increased. A
particular energy absorbing device would be useful for a practical design.
544 Analytic Methods for Design Practice
(4) The thickness of mounting brackets is increased under the maximum load
cases. It is insignificantly influenced by the direction of the crash because
the brackets are located on the car floor.
(5) The seat rail and the seat cushion frame have lower bounds. Therefore, if
the seat is appropriately attached to the bottom, a new design can be made
without increasing the thickness.
(6) The weight of the BIS is increased from 9.7 kg to 20.33 kg. This is
because the initial design is made from an existing conventional design.
9.7.5 Summary
(1) The structural configuration of the existing belt integrated seat (BIS) is
analyzed and the simulation models for the frontal and rear end collisions
are established through the occupant analysis program SAFE. The hyper-
ellipsoid element is used for the seat to represent the geometry and forces
more accurately.
(2) The movement of the dummy’s upper body occurs simultaneously with
that of the seat back of the BIS during both frontal and rear end crashes.
The seat belt slack is significantly small. The maximum values of the seat
belt force and the dummy accelerations with respect to times are
distributed. The right shoulder load is increased compared to that of a
conventional seat. This is because the left and right loads are evenly
distributed. The pelvis acceleration is increased. However, the chest
acceleration, which is more important, is considerably decreased.
Therefore, the safety performance of a BIS is generally excellent. There
are not any good performance indices for the BIS. In the future, research is
required for this aspect.
(3) The parametric study is performed on various seat belts to obtain the
proper type for the BIS. From the numerical results, the belt type with ELR,
high elongation and 0 mm slack is proposed.
(4) The finite element model of the BIS is established by the commercial
structural analysis and optimization software GENESIS. Structural
optimization is performed considering all of the seat frames and brackets as
the design variables including multidirection collision with the oblique
crash. The thicknesses of various frames including the left seat and the
recliner are significantly increased compared to those of the conventional
seat.
(5) Design specification is not well defined for the design of BIS. The design
environment is defined for this research. Therefore, the proposed design
example can only be used for a preliminary design of the BIS. The
industrial sponsor of this research is making a proto type product according
to the results of this research. In the future, establishment of design
specifications is required for the design of belt integrated seats.
Application of Design Methods to Automobile Safety Problems 545
Approximate
objective function and constraints
Const. ĺ Linear ĺ Quadratic
Satisfy No
termination criterion?
Yes
Stop
Section 6.7 is used for the approximation process. At the beginning, only a few
coefficients are obtained. For example, only the coefficients for linear terms can
be calculated. Optimization is performed with the approximated functions and the
optimum becomes a new candidate point. In other words, the number of candidate
points increases by one.
The process proceeds in an iterative manner until the new candidate point is
unchanged. As the iteration proceeds, constant, linear and quadratic terms are
sequentially produced for the approximated functions. This process is represented
in Figure 9.40. The interface between SAFE and VisualDOC is depicted in Figure
9.41.
GUI Input
Input (Graphic User Interface)
Define a design problem
Post-processor
Output
Crashworthiness
analysis
or Crash pulse
test
Crashworthiness analysis
Crash pulse Design car body and layout
dv
Fdt d(mv) m dt (9.13)
dt
where F is the external force, m is the mass, t is the time, and v is the velocity.
Because the mass is unchanged during a crash, the momentum of the body is as
follows:
dv dv
³ F dt ³m dt m³ dt m ³ a dt (9.14)
dt dt
Real pulse
0.6 5
2
0.4
3
0.2
7
1 8
0.0
0.0 0.2 0.4 0.6 0.8 1.0
–0.2
Time (ms/ms)
dv
where a.
dt
Suppose there are two different automobile body structures, model 1 and model
2. These represent two different body designs. In the same crash situation, each
model has to satisfy the law of conservation of momentum as follows:
T T
For model 1 ³ F model 1dt m model 1 ³ a model 1dt (9.15a)
0 0
T T
model 2
For model 2 ³F dt m model 2 ³ a model 2 dt (9.15b)
0 0
T T
model 1 model 2
³a dt ³a dt (9.16)
0 0
The crash pulse is the input for occupant analysis and is the discrete data for
the deceleration curve of the vehicle body with respect to time. The number of the
550 Analytic Methods for Design Practice
data points is 90 to 900. Therefore, all the data points cannot be defined for the
design variable set. The linear interpolation can be utilized as illustrated in Figure
9.43. Then a few data points can be used as design variables. The optimization
problem is formulated as follows:
The deceleration curve for a frontal crash is determined for a car with an airbag.
Eight data points are connected by linear functions as illustrated in Figure 9.43.
The times and accelerations of the eight points are selected as design variables.
The injury indices for the frontal crash are HIC, CSI and the femur load (FMVSS
2003). Here HIC is chosen for the objective function. The optimization
formulation is the same as Equation 9.18 except for replacing Equation 9.18b by
HIC. Table 9.31 shows the lower and upper bounds of the design variables.
RSAO is applied to find the optimum value of the above problem. The
optimum solution is shown in Tables 9.32 and 9.33. The optimum curve is
illustrated in Figure 9.44. HIC is improved by 6.15%. In Figure 9.44, the peaks
are reduced and the distance between the peaks is enlarged.
Initial Optimum
Point
Time Deceleration Time Deceleration
1 0.000 −0.003 0.000 −0.003
2 0.071 0.459 0.071 0.527
3 0.143 0.344 0.143 0.332
4 0.286 1.000 0.286 0.932
5 0.357 0.663 0.357 0.595
6 0.500 0.850 0.500 0.782
7 0.714 0.051 0.786 0.043
8 1.000 −0.048 1.000 −0.048
HIC Remarks
The deceleration curve for rear end impact is determined. When a rear end impact
occurs, the occupant body and the seat move forward together. The head is
supported by the body and moves more slowly than the body. Therefore, the
relative rotational movement of the head causes neck injuries. The objective
function is defined by the extension moment of the neck, which has the limit value
in the safety regulation.
The deceleration curve of the rear end impact is presented in Figure 9.45 and
11 data points are selected for linear interpolation. Constraints are defined from
the conservation of momentum. The bounds of the design variables are shown in
Table 9.34. The optimization formulation is the same as in Equation 9.18 with n =
11 except for the objective function in Equation 9.18b. The objective function is
defined as follows:
Table 9.35. Initial and optimum values for rear end impact
Figure 9.46. Initial and optimum curves for rear end impact
An airbag system is designed using the crash pulse for a frontal crash, which was
determined in the previous section. The design variables are defined as in Section
9.3. The design variables are the mass flow rate, the vent hole area, the firing time
of the inflator and the stiffnesses of the seat belt and the knee bolster. They are
shown in Table 9.36. There are no constraints and the objective function is HIC.
554 Analytic Methods for Design Practice
An orthogonal array is used for the design. The L25 (5 6 ) orthogonal array is
selected as shown in Table 9.37. The levels are normalized as in Section 9.3. The
first level is 0.7, the second level is 0.85, the third level is 1.0, which is the current
Application of Design Methods to Automobile Safety Problems 555
Level mean
Factor Mean
1 2 3 4 5
1 (A) 0.745 0.696 0.693 0.781 0.771 0.737
2 (B) 0.790 0.772 0.695 0.748 0.679 0.737
3 (C) 0.714 0.732 0.682 0.763 0.793 0.737
4 (D) 0.684 0.705 0.745 0.747 0.803 0.737
5 (E) 0.732 0.747 0.815 0.667 0.724 0.737
value, the fourth level is 1.15 and the fifth level is 1.3. The one-way table is made
as shown in Table 9.38 and the solution is A3 B5 C 3 D1 E 4 . A confirmation
experiment is carried out and HIC 0.69. The 20th row of Table 9.37 is better
with HIC 0.577. Thus, it is considered as the final solution. This phenomenon
appears because the interactions are ignored.
Now an optimization problem is defined. The design variables are the same as
the above DOE. The first and fifth levels are the upper and lower bounds,
respectively. A gradient-based method and RSAO are applied. Figure 9.47
presents the history of the objective function from each method. The values are
normalized by the initial objective functions. The gradient-based method obtains
an inferior solution even though it is mathematically rigorous. The reason is that
the functions in occupant analysis are quite noisy and the local sensitivities
frequently differ from the overall trend.
The results are compared in Table 9.39. All the methods can obtain an
improved solution. The RSAO method obtains the best solution. It is noted that
1.0
0.8
Optimum
0.6
0.4
the solution from the orthogonal array is obtained from a discrete space. Thus, it
may be more reasonable because an available design can be used.
A seat is designed based on the previously determined crash pulse. In rear end
impact the neck injuries depend on the safety performance of the seat because the
seat is the only safety device. Rear end impact generally occurs at a low speed.
However, it generates a relative motion between the neck and the body, and the
relative motion causes neck injuries.
Neck injuries result from the sudden extension moment, shear force, tension
and compression force at the occipital condyles. Especially, the limit value of the
extension moment is low and the extension moment is the main cause of neck
Design variable
Level
A B C D E
1 5 cm í50% 0. 30 10¶ í50%
2 7 cm 0 0. 45 20¶ 0
3 9 cm 50% 0. 60 30¶ 50%
Application of Design Methods to Automobile Safety Problems 557
injuries. Thus, a design should be carried out to decrease the extension moment.
Design variables are shown in Table 9.40. An optimization problem is formulated
here. The objective function is the extension moment and there are no constraints.
An orthogonal array is utilized for the seat design. The design variables and
1.0
function
0.8
Optimum
0.6
Objective
0.4
levels are defined in Table 9.41 and the utilized orthogonal array is shown in Table
9.42. The one-way table for Table 9.42 is shown in Table 9.43. The solution is
A1 B3C3 D1 E1 and the objective function is 19.335 from the confirmation
experiment. It is compared with all the rows of the orthogonal array. It is found
that the eighth row of the orthogonal array is the best with 11.731. Thus, the
eighth row is selected as the final solution. The phenomenon occurs because the
interactions are ignored. The optimization results are shown in Figure 9.48 and
Table 9.44. The trend of the results is similar to that of the airbag example.
9.8.9 Summary
The crash pulse for occupant analysis is obtained first and this differs from the
conventional process. An optimization problem is defined to obtain the crash
pulse and it is applied to frontal crash and rear end impact. The safety devices are
Table 9.44. Optimum values for the seat from various methods
Orthogonal
í5.0 1.5 0.300 70.0 1.5 11.731
array
Gradient
í5.0 0.5 0.6 70.0 1.25 11.446
method
designed based on the crash curves, and the crash performances are considerably
improved. The overall results are shown in Table 9.45. It is noted that the safety
performances are considerably improved by the optimized crash pulses.
References
ATB User’s Guide (1988) Wright Patterson Air Force Base, Dayton, OH
Backaitis SH, Mertz HJ (eds.) (1994) Hybrid III: The First Human-like Crash Test
Dummy. Society of Automotive Engineers, Warrendale, PA
Cook RD, Malkus PS, Pleska ME (1989) Concepts and Application of Finite
Element Analysis. Wiley, New York
Federal Motor Vehicle Safety Standards (FMVSS) (2003) Code of Federal
Regulations: Title 49–Transportation. Part 571, National Archives and Records
Administration, Washington D.C.
GENESIS User's Manual Version 7.5 (2004) Vanderplaats Research &
Development, Colorado Springs, CO
Lee WI, Park GJ, Park YS (1993) A Study on the Second-order Approximation
Method for Engineering Optimization. Proceedings of the IFIP WG7.5 Fifth
Working Conference, pp 213í220
Lee JW, Yoon KH, Park GJ (2003) A Study on Occupant Neck Injury in Rear End
Collisions. Proceedings of the Institution of Mechanical Engineers, Part D, Journal
of Automobile Engineering 217(1):23í29
Lim JM, Choi JW, Park GJ (1997) Automobile Side Impact Modeling Using ATB
Software. International Journal of Crashworthiness 2(3):287í297
Lim JM, Wu DH, Park GJ (2000) Analysis and Design Consideration of an Energy
Absorbing Steering System Using Orthogonal Arrays. International Journal of
Crashworthiness 5(3):271í278
LS/DYNA User’s Manual (2003) Livermore Software Technology, CO, USA
MADYMO User’s Manual Version 6.1 (2003) TNO Automotive, The Netherlands
560 Analytic Methods for Design Practice
Park YS, Lee JY, Lim JM, Park GJ (1996) Optimum Design of a Steering Column
to Minimize the Injury of a Passenger. International Journal of Vehicle Design
17(4):398í414
Park YS, Park GJ (2001) Crash Analysis and Design of a Belt Integrated Seat for
Occupant Safety. Proceedings of the Institution of Mechanical Engineers, Part D,
Journal of Automobile Engineering 215(8):875í889
SAFE User’s Manual (1999) Hanyang University, Korea
Shin MK, Hong SW, Park GJ (2001) Axiomatic Design of the Motor Driven
Tilt/Telescopic Steering System for Safety and Vibration. Proceedings of the
Institution of Mechanical Engineers, Part D, Journal of Automobile Engineering
215(2):179í187
Shin MK, Park KJ, Park GJ (2003) Occupant Analysis and Seat Design to Reduce
Neck Injury from Rear End Impact. International Journal of Crashworthiness
8(6):573í581
VisualDOC User’s Manual Version 4.0 (2004) Vanderplaats Research &
Development, Colorado Springs, CO
10 Multidisciplinary Design Optimization
10.1 Introduction
At present, engineering systems are fairly large and complicated. Design
requirements are rigorous and stringent for such systems. Accordingly, design
engineers are seeking new methods and one of them is multidisciplinary design
optimization (MDO). MDO is a design optimization method. Generally,
optimization is applied by considering only one discipline. However, it is fairly
difficult to solve modern engineering problems with a single discipline. Therefore,
we need design methodologies that can cover multidisciplines. Diverse MDO
methods have been proposed to overcome these difficulties.
Structural optimization is a sophisticated method that considers a single
discipline of structural analysis. On the other hand, multiple disciplines are
included in MDO. MDO usually has many design variables, objective functions
and constraints. Analysis can be independently performed for a discipline, and the
disciplines can be linked through analyses. That is, output from a discipline can be
input to other disciplines. This phenomenon is called the coupled process. In
MDO, it is important to efficiently manage the coupled process and to optimize
large-scale problems.
The need for MDO has emerged in the design of airplane wings. Pressures are
imposed on airplane wings by the flow of the surrounding air. The pressures act
on the wings as the lift force and the drag force. The pressures are external forces
on the wing structure and the shape of the wing is changed due to these external
forces. Then the air flow changes and the pressures change accordingly.
Equilibrium is obtained when the coupled process converges. This phenomenon is
investigated in aeroelasticity and computer simulation is adopted to solve the
problem. The pressure distribution is solved by computational fluid dynamics and
the deformation of the wing is analyzed by the finite element method. The
analysis method considering the coupled process is called multidisciplinary
analysis (MDA).
When we design airplane wings, the above two disciplines should be
considered simultaneously. A design problem is separately defined for each
discipline. For example, the design to minimize the drag force is performed by the
562 Analytic Methods for Design Practice
output from computational fluid dynamics and the weight of the wing structure is
minimized by using the output of the finite element method. When optimization
technology is employed in this case, it is called MDO. The two disciplines have
common design variables, objective functions and constraints. They can be
independently defined in each discipline.
Coupling is predominant in the analysis and design of airplane wings. The
problem has not yet been thoroughly solved from an MDO viewpoint, so designers
are examining various methods. Sometimes, a discipline of the control theory is
involved in the design of airplane wings and there are many references for the
design of wings. Therefore, the details of airplane wings are not explained in this
chapter. Instead, general methods for MDO are discussed and their application is
investigated (Sobieszczanksi-Sobieski and Haftka 1996).
The optimization formulation for a target problem requires the use of MDO.
Based on the formulation, the overall flow and utilized optimization methods are
determined and can be modified accordingly. Gradient-based methods are usually
used for the optimization process. Recently, some MDO research has adopted
nongradient-based methods such as genetic algorithms, neural networks, simulated
annealing, etc. Meanwhile, some software companies have announced various
MDO frameworks such as ModelCenter (ModelCenter 2004), iSIGHT (iSIGHT
2004), Optimus (Optimus 2003), MIDAS (Team Midas 1995), VisualDOC
(VisualDOC 2004), etc. In this chapter, the formulation of MDO is explained and
the methods to solve it are described.
Overall, there are seven methods for MDO. The formulation of the
multidisciplinary feasible (MDF) method is basic. It is easy to use even though
complicated system analysis should be carried out for each step. The coupled
relationship is solved in the system analysis. The formulation of the individual
discipline feasible (IDF) method has been developed to eliminate the system
analysis, and each discipline is independently solved. In the all-at-once (AAO)
method, even the analysis process is eliminated in the disciplines. None of the
disciplines determine the design in the above methods. Analysis is only carried
out in the disciplines (Cramer et al. 1993).
Various methods have been developed to allocate the design process to
disciplines. They are concurrent subspace optimization (CSSO) (Sobieszczanski-
Sobieski 1988), collaborative optimization (CO) (Braun 1996), bilevel integrated
system synthesis (BLISS) (Sobieszczanski-Sobieski et al. 1998) and
multidisciplinary optimization based on independent subspaces (MDOIS) (Shin
and Park 2005).
In CSSO, sensitivity information is obtained by solving the global sensitivity
equation (GSE), which is a derivative of the system analysis. Each discipline
considers the design conditions of the other disciplines indirectly. The overall
process is controlled by the coordination optimization problem (COP) in the upper
level. BLISS also uses GSE. Each discipline of the lower level has its own local
design variables and the common design variables are determined in the upper
level. MDOIS proposes an MDO method for a problem that has an independent
design problem for each discipline. These three methods conduct the system
analysis at the beginning of each cycle. On the other hand, CO does not have the
Multidisciplinary Design Optimization 563
system analysis. The upper level determines target values for disciplines, and each
discipline solves its own design problem while a given target value is satisfied.
Here, the characteristics of MDO are investigated and the above seven methods
are compared. The methods are constantly evolving and being improved upon. In
this chapter, they are explained based on the initial proposition of the methods.
The methods are derived mathematically. They can still solve small-scale
problems or problems with weak couplings. Approximation methods are
frequently utilized for large-scale problems or problems with strong couplings.
Currently, it seems that no method can solve all the MDO problems. For large-
scale problems, an improved solution is sought with specific approximation
methods instead of a mathematically optimum solution. Therefore, none of the
introduced methods is said to be the ultimate one.
In MDO, there are multiple disciplines that can have independent analysis and
design procedures. Suppose we have two disciplines. The coupling for analysis
can be expressed as
z1 h1 (b1 , b c , z c2 ) (10.1a)
z2 h 2 (b 2 , b c , z1c ) (10.1b)
where h1 and h 2 are the analyzers for disciplines 1 and 2, respectively. Each
analyzer can have its own design variables and coupling variables. b1 R nd 1 and
b 2 R nd 2 are local design variable vectors for disciplines 1 and 2, respectively.
b c R ndc is the common design variable vector shared by the two disciplines.
The state variable vectors z1 R ns1 and z 2 R ns 2 are the results (responses) from
the two disciplines, respectively.
When we have couplings between disciplines, some or all of the state variables
of a discipline are used as input to the other discipline. Such state variables are
called coupling variable vectors z1c R nc1 and z c2 R nc 2 . The state variables that
are not coupling variables are the noncoupling variables z1nc R nnc1 and
nnc 2 T T
z nc
2 R . The state variable vectors are expressed by z1 [z1nc z1c ]T and
T T
z2 [z nc
2 z c2 ]T .
Equation 10.1 can be expressed as
z nc
2 h 21 (b 2 , b c , z1c ) (10.2c)
z c2 h 22 (b 2 , b c , z1c ) (10.2d)
Equations 10.1 and 10.2 have the same meaning and they represent analyzers.
If we have couplings in analysis, the coupling variables should be evaluated. A
representative method is the fixed point iteration method. It is difficult to obtain
coupling variables in practice. For example, aeroelasticity is utilized in the wing
design. There are still many difficulties in the analysis of a wing with MDO
application. Approximation methods are frequently adopted for calculation of the
coupling variables. The state where all the coupled relationships are satisfied is
called multidisciplinary feasibility (MF). That is, the equilibrium between
disciplines is satisfied. The state where the equilibrium of a discipline is satisfied
Multidisciplinary Design Optimization 565
is called individual feasibility (IF). In IF, the coupled relationship may not be
satisfied. Achieving MF or IF is a fairly important problem in MDO.
The above three coupled relationships of analysis involved in MDO are
represented in Figure 10.1. In Figure 10.1, the design variable vectors b1 , b 2 and
b c are fixed and the state variable vectors z1 and z 2 are obtained from h1 and
h2 .
In Figure 10.1a, analysis is independently performed for each discipline. In
b1 , b 2 , b c
z1 h1 (b1 , b c ) z 2 h 2 (b 2 , b c )
z1 z2
b1 , b 2 , b c
z1 h1 (b1 , b c )
z1c
z2 h 2 (b 2 , b c , z1c )
z1 z2
b1 , b 2 , b c
z c2
z1 h1 (b1 , b c , z c2 ) z2 h 2 (b 2 , b c , z1c )
z1c
z1 , z 2
MDO formulation is defined based on the coupled aspects of the design variables,
the objective function and constraints. An MDO problem is formulated as follows:
z1 h1 (b1 , b c , z c2 ) (10.3d)
z2 h 2 (b 2 , b c , z1c ) (10.3e)
where b1 , b 2 and b c are local design variable vectors of disciplines 1 and 2, and
the common design variable vector shared by the two disciplines, respectively.
f [ f1 f 2 f c ] consists of the local objective functions for disciplines 1 and 2,
and the common objective function. g R m is the constraint vector.
The formulation in Equation 10.3 can be rewritten as follows:
Find b1 , b 2 , b c (10.4a)
ª f1 (b1 , b c , z1 ) º
« f (b , b , z ) »
to minimize « 2 2 c 2 » (10.4b)
«¬ f c (b1 , b 2 , b c , z1 , z 2 )»¼
g 2 (b 2 , b c , z 2 ) d 0 (10.4d)
g c (b1 , b 2 , b c , z1 , z 2 ) d 0 (10.4e)
z c2 h 22 (b 2 , b c , z1c ) (10.4i)
Discipline 4
Discipline 1 Discipline 3
Discipline 2
Upper level
Lower level
Discipline 1 Discipline 2 Discipline 3 Discipline 4
Design optimization
b1 , b 2 , b c z1 , z 2
z1c
z1 h1 (b1 , b c , z c2 ) z2 h 2 (b 2 , b c , z1c )
z c2
(z1c s1c ) 0, (z c2 s c2 ) 0
z1c z c2
z1c h1 (b1 , b c , s c2 ) z c2 h 2 (b 2 , b c , s1c )
e1 e 2
e1 h1' (b1 , b c , s c2 ) s1 e2 h '2 (b 2 , b c , s1c ) s 2
during the optimization process of the upper level. The target values are conveyed
to the lower level.
The classification is made based on the existing MDO methods. For example,
there is no multilevel method that uses evaluators. Thus, it is not included. Each
method will be explained later in detail.
(1) Using the fixed point iteration method, find a solution that satisfies MF.
(2) Use [ s1c s 2c ] [1 1] of Figure 10.4 and find a solution that satisfies IF.
(3) Using the condition of (2), find e1 and e 2 of Figure 10.5.
Solution
The result of Equation 10.5a is substituted into Equation 10.5b and the
result of Equation 10.5b is substituted into Equation 10.5a. The process is
repeated until it converges. Then a solution satisfying MF is obtained.
The process is shown in Table 10.1. The final solutions are z1c 3.0 and
z 2c 4.0.
(2) The solutions satisfying IF are
z1c h1 (b1 , b3 , s 2c )
z 2c h2 (b2 , b3 , s1c )
No. z1c z 2c
1 1.00000 1.00000
2 4.50000 6.00000
3 2.00000 2.50000
39 3.00000 3.99999
40 3.00000 4.00000
The optimization problem in Equation 10.3 may be difficult to solve for large-scale
problems. A large-scale problem can be decomposed with respect to each
discipline. The decomposition is usually carried out based on the type of analysis.
Using linear addition for objective functions in Equation 10.3 ( f f1 f 2 f c ) ,
Equation 10.3 can be modified as follows:
572 Analytic Methods for Design Practice
Find b1 , b 2 , b c (10.10a)
subject to g (b 1 , b 2 , b c , z 1 , z 2 ) d 0 (10.10c)
In Equation 10.10, the objective function and constraints are coupled with
respect to the design variables. The objective function and constraints can be
linearly expanded as follows:
df df df
f # f0 'b1 'b 2 'b c (10.11a)
db1 db 2 db c
dg j dg j dg j
g j # g j0 'b1 'b 2 'b c (j 1, 2,..., m) (10.11b)
db1 db 2 db c
d ª d d d º T
where « », 'b l ['bl1 'bl 2 'blr ] l {1, 2,
db l db
¬« l1 d b l2 db »
lr ¼
c} and r {nd1, nd 2, ndc}. The subscript 0 denotes the values of the current
design and ' is perturbation from the current design.
Using Equation 10.11, an optimization problem for each discipline can be
formulated by its own design variables. Suppose b1 and b c are allocated to
discipline 1 and b 2 is allocated to discipline 2. Then the optimization problem for
each discipline is defined as follows:
(1) Discipline 1
Find b1 , b c (10.12a)
df df
to minimize f 0 'b1 'b c (10.12b)
db1 db c
dg j dg j
subject to g j0 'b1 'bc d 0 ( j 1, 2, ..., m) (10.12c)
db1 db c
(2) Discipline 2
Find b2 (10.12d)
df
to minimize f 0 'b 2 (10.12e)
db 2
Multidisciplinary Design Optimization 573
dg j
subject to g j0 'b 2 d 0 ( j 1,2,..., m) (10.12f)
db 2
df wf wf dz nc wf dz c wf dz nc wf dz c
nc 1 c 1 nc 2 c 2 (10.13a)
db1 wb1 wz1 db1 wz1 db1 wz 2 db1 wz 2 db1
df wf wf dz nc wf dz c wf dz nc wf dz c
nc 1 c 1 nc 2 c 2 (10.13b)
db 2 wb 2 wz1 db 2 wz1 db 2 wz 2 db 2 wz 2 db 2
df wf wf dz nc wf dz c wf dz nc wf dz c2
nc 1 c 1 nc 2 c (10.13c)
db c wb c wz1 db c wz1 db c wz 2 db c wz 2 db c
dg j wg j wg j dz1nc wg j dz1c wg j dz nc
2
wg j dz c2
(10.13d)
db1 wb1 wz1nc db1 wz1c db1 wz nc
2 db1 wz c2 db1
dg j wg j wg j dz1nc wg j dz1c wg j dz nc wg j dz c2
nc
c nc 2 c (10.13e)
db 2 wb 2 wz1 db 2 wz1 db 2 wz 2 db 2 wz 2 db 2
dz qk ª dz qk dz qk dz qk º
where « » (l {1, 2, c}, k {1, 2}, q {nc, c}) and
db l «¬ dbl1 dbl 2 dblr »¼
r {nd1, nd 2, ndc}, j = 1, 2,…,m. The partial derivatives on the right hand side
of Equation 10.13 are explicit ones and the ordinary derivatives are obtained from
the system analysis.
Equation 10.2 can be derived as follows:
wh11 wh wh
dz1nc db1 11 db c 11c
dz c2 (10.14a)
wb1 wb c wz 2
wh12 wh wh
dz1c db1 12 db c 12c
dz c2 (10.14b)
wb1 wb c wz 2
wh 21 wh wh
dz nc
2 db 2 21 db c 21c
dz1c (10.14c)
wb 2 wb c wz1
574 Analytic Methods for Design Practice
wh 22 wh wh
dz c2 db 2 22 db c 22 dz1c (10.14d)
wb 2 wb c wz1c
wh pq ª wh pq wh pq wh pq º wh pq
where « » (l {1,2, c}, p, q {1,2}),
wb l «¬ wbl1 wbl 2 wblr »¼ wz lc
ª wh pq wh pq wh pq º
« c » (l , p, q {1,2}), s {nc1, nc 2} and r {nd1, nd 2,
«¬ wzl1 wzlc2 wzlcs »¼
ndc}. From Equation 10.14, derivatives of the state variables are obtained. For
instance, derivatives of the state variables with respect to b1 are
dz 2nc wh 21 dz1c
(10.15c)
db1 wz1c db1
dz c2 wh 22 dz1c
(10.15d)
db1 wz1c db1
ª wh11 º ª dz1nc º
«I 0 0 »« »
« wz 2 c » « db1 » ª wh11 º
wh « c » « wb »
«0 I 0 12c » « dz1 » « 1»
« wz 2 » « db1 »
« » « wh12 » (10.16)
«0 wh 21
nc « wb1 »
I 0 » «« dz 2 »» « 0 »
« wz1c » db
« »
« wh » « c1 »
«¬ 0 »¼
«0 22 0 I » « dz 2 »
«¬ wz1c »¼ « db1 »
¬ ¼
The terms related to the coupling variables in Equation 10.16 are expressed as
Multidisciplinary Design Optimization 575
ª wh12 º ª dz1c º
« I « » ª wh12 º
« wz c2 »» « db1 » « wb » (10.17)
« wh 22 «
I » « dz 2 »
c » « 1»
« wz c » db ¬ 0 ¼
¬ 1 ¼¬ 1 ¼
Equations 10.16 and 10.17 are called global sensitivity equations (GSE). Their
solution is the derivative of coupling variables with respect to design variables. It
can be used in Equation 10.13 and various MDO methods.
Oj !0 (10.18c)
where J represents the active set. Since equality constraints are always active, the
equality constraints can be considered in the same manner except for the sign of
the Lagrange multiplier vector Ȝ in Equation 10.18c. Therefore, the derivation in
this section is valid for problems with equality constraints.
When the parameter p is changed, Equation 10.18 should not be changed at the
optimum point with respect to p. In other words, the derivatives of Equation
10.18a with respect to p are zeroes. It is expressed as
w 2 g j (b)
¦ Oj 0, i 1,..., n (10.19a)
j J wbi wp
576 Analytic Methods for Design Practice
where
w2 f w2g j
Aik ¦ Oj (10.20b)
wbi wbk jJ wbi wbk
wg j
Bij , jJ (10.20c)
wbi
w2 f w2 g j
ci ¦ Oj (10.20d)
wbi wp jJ wbi wp
wg j
dj , jJ (10.20e)
wp
T
ª wb1 wb2 wbn º
Gb « ... » (10.20f)
¬ wp wp wp ¼
T
ª wO1 wO2 wO J º
GȜ « ... » (10.20g)
¬ wp wp wp ¼
Gb and GȜ are obtained from Equation 10.20a. Thus, the optimum sensitivity
of the objective function is
df (b) wf (b)
T f (b)Gb (10.21a)
dp wp
or
Equation 10.21b is derived from Equations 10.18a and 10.19b. Generally, the
df (b)
optimum sensitivity means Gb or .
dp
Find b1 , b2 , bc (10.22a)
ª dz c dz 2c º ª dz1c dz 2c º
Evaluate the global sensitivities « 1 » and « ».
¬« db1 db1 ¼» ¬« db2 db2 ¼»
Solution
ª dz c dz 2c º ª dz1c dz 2c º
Using Equation 10.15 « 1 » and « » are included in the
«¬ db1 db1 »¼ «¬ db2 db2 »¼
following simultaneous equations:
dz 2c wz 2c wz 2c dz1c dz1c
0 ( 0.6) u (10.23b)
db1 wb1 wz1c db1 db1
dz 2c wz 2c wz 2c dz1c dz1c
1 ( 0.6) u (10.23d)
db2 wb2 wz1c db2 db2
ª dz c dz 2c º
The solutions of Equation 10.23 are « 1 » [1.31579 0.78947] and
«¬ db1 db1 »¼
ª dz1c dz 2c º
« » [ 0.52632 1.31579] .
«¬ db2 db2 »¼
Find b1 , b2 (10.24a)
0 d b1 , b2 d 5 (10.24d)
df (b)
When p 2, evaluate optimum sensitivity .
dp
Solution
The optimum point is b1 0.0 and b2 0.64575 . Each component of the
optimum sensitivity is obtained from Equation 10.21b as follows:
wf (b)
3(b1 p) 2 3(0 2) 2 12 (10.25a)
wp
O 0.18892 (10.25b)
wg
2b2 2 u 0.64575 1.2915 (10.25c)
wp
df (b)
Therefore, the optimum sensitivity is
dp
df (b)
12 (0.18892)(1.2915) 11.75601 (10.26)
dp
Multidisciplinary Design Optimization 579
Find b1 , b 2 , b c
to minimize f (b1 , b 2 , b c , z1 , z 2 )
subject to g1 (b1 , b c , z1 ) d 0
g 2 (b 2 , b c , z 2 ) d 0
g c (b1 , b 2 , b c , z1 , z 2 ) d 0
b1 , b 2 , b c z1 , z 2
z c2
z1 h1 (b1 , b c , z c2 ) z2 h 2 (b 2 , b c , z1c )
z1c
The IDF method is a single level method that decomposes the disciplines. It does
not have system analysis. Complimentary vectors ( s1c and s c2 ) are adopted for the
coupling variable vectors ( z1c and z c2 ). Instead of MF, compatibility conditions
are utilized and they are included as equality constraints in the optimization
process. The formulation for the IDF method is as follows:
g 2 (b 2 , b c , z 2 ) d 0 (10.27d)
g c (b1 , b 2 , b c , z1 , z 2 ) d 0 (10.27e)
c2 s c2 z c2 0 (10.27g)
z1 h1 (b1 , b c , s c2 ) (10.27h)
z2 h 2 (b 2 , b c , s1c ) (10.27i)
An analyzer of a discipline does not wait for the output from the other
analyzers. Instead, it performs the analysis (IF) by using the complimentary
vectors (s1c and s c2 ). Eventually, the coupling variables (z1c and z c2 ) should be
the same (MF) as the complimentary vectors (s1c and s c2 ). Thus, compatibility
Multidisciplinary Design Optimization 581
Find b1 , b 2 , b c , s1c , s c2
to minimize f (b1 , b 2 , b c , z1 , z 2 )
subject to g1 (b1 , b c , z1 ) d 0
b1 , b c , s c2 b 2 , b c , s1c
g 2 (b 2 , b c , z 2 ) d 0
g c (b1 , b 2 , b c , z1 , z 2 ) d 0
MF c1 s1c z1c 0
c2 s c2 z c2 0
z1 z2
z1 h1 (b1 , b c , s c2 ) z2 h 2 (b 2 , b c , s1c )
IF IF
conditions (c1 and c 2 ) are defined as in Equations 10.27f and 10.27g. The
overall flow of the IDF method is presented in Figure 10.7.
When the IDF method is used, the number of design variables is increased by
the number of complimentary variables. The number of design variables is
nd1 nd 2 ndc plus N (s1c ) N (s c2 ) . N (s1c ) means the number of the elements
in vector s1c . The number of equality constraints is the number of coupling
variables. IF is always satisfied during the optimization process and MF is
satisfied when the optimum design is obtained.
Complimentary variables s1c and s 2c are defined for state variables z1c and z 2c .
The optimization problem is formulated as follows:
c2 s 2c z 2c 0 (10.28f)
The AAO method is a single level method that does not perform either the system
analysis or individual analysis for each discipline. Analysis is generally carried
out many times for a design. The analysis is quite expensive and the analysis
results during the design process are meaningless when the final solution is
obtained. The AAO method was developed to abolish the expensive analysis
process. Evaluators are utilized instead of analyzers in the AAO method.
State variables are calculated with respect to design variables in analysis. The
analysis process can be viewed as illustrated in Figure 10.8. Figure 10.8 represents
the analysis process of discipline 1. State variable vector z1 is obtained from
design variable vectors b1 and b c and coupling variable vector z c2 .
The analysis process can be viewed from a different perspective. Figure 10.9
presents an evaluator. The design and state variable vectors are inputs. When the
results of the evaluator are the zero vector, the analysis is regarded as finished.
The concept in Figure 10.9 is expanded to MDO by using the SAND concept.
Input Output
Analyzer
b1 , b c , z c2 z1
z1 h1 (b1 , b c , z c2 )
Input Output
Evaluator
b1 , b c , z1 , z c2 0
e1 e1 (b1 , b c , z1 , z c2 )
h1' (b1 , b c , z c2 ) z1
Complimentary variable vectors (s1c and s c2 ) are adopted for state variable vectors
(z1c and z c2 ). The process of the evaluator is considered in the same way as
equality constraints. This is similar to the IDF method. In the IDF method, the
complimentary variables for coupling variables are used as design variables in the
optimization process. On the other hand, the complimentary variables for coupling
and noncoupling variables are included in the design variable set. That is, all the
state variables are regarded as design variables.
The optimization formulation of the AAO method is as follows:
g 2 (b 2 , b c , s nc c
2 ,s2 ) d 0 (10.29d)
e1 0 (10.29f)
e2 0 (10.29g)
e2 h c2 (b 2 , b c , s1c ) s 2 (10.29i)
e1 e1 (b1 , b c , s1 , s c2 ) e1 e2 e2 e 2 (b 2 , b c , s 2 , s1c )
h1' (b1 , b c , s c2 ) s1 h '2 (b 2 , b c , s1c ) s 2
In the IDF method, complimentary variables s1nc , s1c , s2nc and s 2c are added to the
design variable set. The formulation is as follows:
1 ªm º
C ln « ¦ expUg i » (10.31)
U ¬i 1 ¼
System analysis
Converge? Yes
Stop
No
Sensitivity from global sensitivity equations
Update
design
Optimization for each discipline variables
find b1 , b c (10.32b)
df df
to minimize f # f 0 'b1 'b c (10.32c)
db1 db c
~ dC 2 dC 2
C 2 # C 20 'b 1 'b c d C 20 [ s 2 (1 r12 ) (1 s 2 )t12 ]
db 1 db c
(10.32e)
~z nc nc dz 2nc dz nc
2 z10 'b1 2 'b c (10.32h)
db1 db c
~z c dz c2 dz c
2 z c20 'b1 2 'b c (10.32i)
db1 db c
Multidisciplinary Design Optimization 587
(2) Discipline 2
find b2 (10.33b)
df
to minimize f # f 0 'b 2 (10.33c)
db 2
~ dC
subject to C1 # C10 1 'b 2 d C10 [ s1 (1 r21 ) (1 s1 )t12 ] (10.33d)
db 2
C 2 d C 20 [ s 2 (1 r22 ) (1 s 2 )t 22 ] (10.33e)
~z nc nc dz1nc
1 z10 'b 2 (10.33f)
db 2
~z c c dz1c
1 z10 'b 2 (10.33g)
db 2
z c2 h 22 (b 2 , b c , ~z1c ) (10.33i)
s p , rkp and t kp are constants in the lower level and are determined in COP of
the upper level. k and p represent the discipline number, and s p , rkp and t kp are
given by the user at the first cycle. When the objective function of a discipline is a
function of the local variables of the corresponding discipline, the above
approximation is not needed. Otherwise, the objective function is linearly
approximated. A cumulative constraint is employed to include all the constraints.
The cumulative constraint of a discipline is considered through approximation by
another discipline. This process enables on to consider the constraints of the other
discipline because a change of the design in a discipline may affect the other
discipline through coupling. This is shown in Equations 10.32e and 10.33d.
dC
(i = 1, 2, c) is the information obtained from the global sensitivity equation.
db i
The constraints of the other discipline are considered by s p , rkp and t kp .
The coefficients are illustrated in Figure 10.12, which explains rkp . Suppose
the constraint of discipline 1, C1 is violated. It is corrected by r11 in discipline 1
and by r21 in discipline 2. Thus, r21 is the amount of responsibility of discipline 2
for the constraint violation in discipline 1. This is called the responsibility
coefficient and has the following relationship:
588 Analytic Methods for Design Practice
C1
r11
r11 r21 1
r21
C1
t11 t 12 0
t 12
t11
NSS
p
¦ rk 1 (p 1, 2,..., NSS ) (10.34)
k 1
where NSS is the number of disciplines. The meaning of Equation 10.34 is that
“if the constraint violation of the pth discipline C p is 1, rkp is the amount of
responsibility for the kth discipline.”
The trade-off coefficient t is schematically drawn in Figure 10.12b. Suppose
constraint C1 in discipline 1 is satisfied. t11 is negative, thus, discipline 1 tries to
satisfy C1 that is oversatisfied by t11 . t 12 is positive and discipline 2 tries to
~
further reduce the objective function although C1 is violated by t12 . t kp is the
trade-off coefficient of constraint C p in the kth discipline. The trade-off
coefficients have the following relationship:
NSS
p
¦ tk 0 (p 1, 2,..., NSS ) (10.35)
k 1
Multidisciplinary Design Optimization 589
The meaning of Equation 10.35 is that “if constraint C p of the pth discipline is
satisfied, the constraint is oversatisfied or violated by t kp in the kth discipline.”
s p is a switch coefficient for the responsibility and trade-off coefficients.
When C p is violated, s p 1 and the trade-off coefficient is not activated. When
p
C p is satisfied, s 0 and the responsibility coefficient is not activated. The
switch coefficient is automatically determined according to the status of the
constraint violation. On the other hand, the responsibility and trade-off
coefficients can be defined by the user. A method exists to initialize the
responsibility and trade-off coefficients (Sobieszczanski-Sobieski 1988). From the
second cycle, they are determined from the COP.
The COP is formulated as follows:
2
p
subject to ¦ rk 1 (0 d rkp d 1) ( p 1, 2) (10.36c)
k 1
2
p
¦ tk 0 (p 1, 2) (10.36d)
k 1
Equation 10.36b needs the optimum sensitivity of the objective function with
respect to the responsibility and trade-off coefficients. Since the coefficients are
treated as constants in each discipline, the optimum sensitivity is obtained at each
discipline.
CSSO is a two-level optimization method. Each discipline of the lower level
(discipline) needs an optimizer and the upper level also needs an optimizer. IF and
MF are satisfied by the system analysis. Constraints are indirectly shared by all
the disciplines and common design variables are specifically handled by one
discipline.
dz 2c wz 2c dz1c dz1c
(0.6) u (10.37d)
db1 wz1c db1 db1
dz1nc wz1nc dz 2c dz 2c
(0.5) u (10.37e)
db2 wz 2c db2 db2
dz1c wz1c dz 2c dz 2c
(0.4) u (10.37f)
db2 wz 2c db2 db2
dz 2c wz 2c wz 2c dz1c dz1c
1 (0.6) u (10.37h)
db2 wb2 wz1c db2 db2
dz 2c wz 2c wz 2c dz1c dz1c
1 (0.6) u (10.37l)
dbc wbc wz1c dbc dbc
dz 2c
0.52632. These are utilized for the formulation of each discipline. The
dbc
common variable is handled within discipline 1. The formulations are as follows:
(1) Discipline 1
find b1 , bc (10.38b)
df df
to minimize f1 f0 u 'b1 u 'bc
db1 dbc
§ wf dz nc wf dz nc ·
f 0 ¨¨ nc 1 nc 2 ¸¸ u 'b1
© wz1 db1 wz 2 db1 ¹
§ wf dz nc wf dz nc ·
¨¨ nc 1 nc 2 ¸ u 'bc
¸
© wz1 dbc wz 2 dbc ¹
~ dC2 dC
C2 # C20 ǻb1 2 ǻbc d C20 [ s 2 (1 r12 ) (1 s 2 )t12 ]
db1 dbc
(10.38e)
1
C1 ln[exp( U1 g1 )] g1 (10.38f)
ȡ1
~z c c dz 2c dz c
2 z 20 u 'b1 2 u 'bc
db1 dbc
c
z 20 (0.78947) u 'b1 (0.52632) u 'bc (10.38i)
~ dz 2nc dz nc
z 2nc nc
z 20 u 'b1 2 u 'bc
db1 dbc
592 Analytic Methods for Design Practice
nc
z 20 (0.92105) u 'b1 (0.44737) u 'bc (10.38j)
find b2 (10.39b)
wf § wf dz nc wf dz nc ·
to minimize f 2 f0 u 'b2 f 0 ¨¨ nc 1 nc 2 ¸¸ u 'b2
wb2 © wz1 db2 wz 2 db2 ¹
(10.39c)
C 2 d C 20 [ s 2 (1 r22 ) (1 s 2 )t 22 ] (10.39f)
1
C2 ln[exp( U 2 g 2 )] g2 (10.39g)
U2
~z c c dz1c c
1 z10 u 'b2 z10 (0.52632) u 'b2 (10.39j)
db2
~z nc nc dz1nc nc
1 z10 u 'b2 z10 (0.65789) u 'b2 (10.39k)
db2
2
p
subject to ¦ rk 1 (0 d rkp d 1) ( p 1, 2) (10.40c)
k 1
Multidisciplinary Design Optimization 593
2
p
¦ tk 0 (p 1, 2) (10.40d)
k 1
df df
where p
and p are the optimum sensitivity of each discipline in the
drk dt k
lower level. The solutions are b1 3.08234, b2 3.08524, bc 2.97466
and f 0.50243 .
Given b1 , b c , z c2 (10.41a)
df
to minimize 'b1 (10.41c)
db1
(2) Discipline 2
df
to minimize 'b 2 (10.42c)
db 2
594 Analytic Methods for Design Practice
subject to g 2 (b 2 , b c , z 2 ) d 0 (10.42d)
z c2 h 22 (b 2 , b c , z1c ) (10.42f)
Given bc (10.43a)
df
to minimize 'b c (10.43c)
db c
The objective functions in the lower level are linearized with respect to the
local variables as shown in Equations 10.41c and 10.42c. The global sensitivity is
utilized for the linearization. For the upper level, the optimum sensitivity of the
objective function with respect to common variables is required as shown in
Equation 10.43c. The optimum sensitivity is obtained from the lower level. Since
there are two disciplines in the lower level, different optimum sensitivities can be
obtained from different disciplines. Therefore, the optimum sensitivity is
calculated based on satisfaction of the system analysis. BLISS uses GSE/OS
(GSE/optimized subsystems) for this. GSE/OS is a union of the global sensitivity
equation and the optimum sensitivity equations in Equations 10.41 and 10.42. It is
as follows:
nc
ª wh11 wh11 º ª dz1 º ª wh11 º
«I 0 0 0 » « db »
wz c2 wb1 « wb »
« » « cc » « c »
wh wh « » « wh12 »
«0 I 0 12 12 0 » « d z1 »
« wz c2 wb1 » db « wb c »
« »« c » « wh »
«0 wh 21 I 0 0
wh 21 » « dz nc »
2 « 21 »
« wz1c wb 2 » « db » « wb c »
« « c» (10.44)
wh wh » c « wh 22 »
«0 22 0 I 0 22 » « dz 2 » « »
« wz1c wb 2 » « db » « wb c »
« wb1 »« c » « wb1 »
«0 0 0 I 0 » « db1 » « wb »
« wz c2 » « db c » « wb c »
« wb 2 »« » « 2 »
«0 wz c 0 0 0 I » « db 2 »
«¬ wb c »¼
¬ 1 ¼ ¬« db c ¼»
Multidisciplinary Design Optimization 595
wb1 wb1 wb 2
where and are the optimum sensitivities of discipline 1, and and
wz c2 wb c wz1c
wb 2
are those of discipline 2.
wb c
The optimum sensitivities can be obtained by an optimization problem derived
by using the global sensitivity of the objective function and constraints as follows:
df
to minimize f 0 'b i (10.45c)
db i
dg i
subject to gi0 ǻb i d 0 (10.45d)
db i
where i is the discipline number. The results of Equation 10.45 are utilized to
calculate the approximate optimum sensitivity as follows:
wb i 'b i
| (10.46)
wp 'p
The optimum sensitivity can also be obtained from the other methods in Section
10.3.2. The results of Equation 10.46 are substituted into Equation 10.44 and the
solution of Equation 10.44 is evaluated. The other partial derivative terms in
Equation 10.44 are the same as those of Equation 10.16. The solution of Equation
10.44 is adopted for the linearization of Equation 10.43c.
The steps of BLISS are as follows:
Step 1. Design variable vectors b1 ,b 2 and b c are initialized.
Step 2. The system analysis is performed to calculate z1 and z 2 .
Step 3. Convergence is checked from the second cycle.
Step 4. Global sensitivities are evaluated.
Step 5. Problems in Equations 10.41 and 10.42 are solved in the lower level.
Step 6. Equation 10.44 for GSE/OS is established and solved.
Step 7. In the upper level, the problem in Equation 10.43 is solved.
Step 8. The design variables are updated and Step 2 is redone.
In BLISS, IF and MF are satisfied during optimization because the system analysis
is performed first.
596 Analytic Methods for Design Practice
Given b1 , bc , z 2c (10.47a)
df § wf dz1nc wf dz 2nc ·¸
to minimize u 'b1 ¨ u 'b1
db1 ¨ wz nc db wz 2nc db1 ¸¹
© 1 1
(2) Discipline 2
df § wf dz1nc wf dz 2nc ·¸
to minimize u 'b2 ¨ u 'b2 (10.48c)
db2 ¨ wz nc db wz 2nc db2 ¸¹
© 1 2
ª dz1nc º
« »
« dbc »
ª1 0 0 0.5 1 0 º « dz1c » ª 1 º
«0 1 0 0.4 1 0 »» «« dbc »» « 1 »
« « »
«0 0.7 1 0 0 1» « dz 2nc » « 1 »
« »« » « »
« 0 0.6 0 1 0 1» « dbc » « 1 » (10.49)
«0 wb1 c
« wb1 »
0 0 c 1 0 » « dz 2 »
« wz 2 »« » « wbc »
« wb » « dbc » « wb »
«0 2 0 0 0 1 » « db1 » « 2»
«¬ wz1c »¼ « dbc » «¬ wbc »¼
« db »
« 2»
«¬ dbc »¼
Given bc (10.50a)
df § wf dz1nc wf dz 2nc ·¸
to minimize u 'bc ¨ u 'b c (10.50c)
dbc ¨ wz nc dbc wz nc dbc ¸
© 1 2 ¹
The local variables are obtained from Equations 10.47 and 10.48, and the
common variables are determined from Equation 10.50. The solution converges to
b1 3.02503, b2 3.75753, bc 3.02503 and f 0.501046.
The first step of CO is to define the complimentary variables. This can vary
according to the user. Here, they are defined for local variables (b1 and b 2 ),
common design variables (b c ) and coupling variables as follows:
The left terms of Equation 10.51 are design variable vectors of the lower level or
the coupling variable vectors. The right terms are vectors determined in the upper
level and used as target values in the lower level.
The problem in Equation 10.4 is modified for each discipline as follows:
(1) Discipline 1
find b1 , b c , z c2 (10.52b)
to minimize c1 || b1 s1 || 2 || b c s c || 2 || z 1c s1c || 2 || z c2 s c2 ||
(10.52c)
subject to g1 (b1 , b c , z1 ) d 0 (10.52d)
(2) Discipline 2
to minimize c 2 || b 2 s 2 || 2 || b c s c || 2 || z 1c s1c || 2 || z c2 s c2 ||
(10.53c)
subject to g 2 (b 2 , b c , z 2 ) d 0 (10.53d)
z c2 h 22 (b 2 , b c , z1c ) (10.53f)
dc1
subject to c1 c10 's 0 (10.54c)
ds
dc2
c2 c20 's 0 (10.54d)
ds
z c2 h 22 (s 2 , s c , s1c ) (10.54h)
The sensitivity of the objective function in the upper level can be obtained
analytically or by the finite difference method as follows:
The equality constraints in Equations 10.54c and 10.54d are objective functions
of the disciplines. The derivative in Equations 10.54c and 10.54d is the optimum
sensitivity. Using Equation 10.21b, the optimum sensitivity is
dc j wc j wg j
Ȝ jT ( j 1,2, s {s1 , s 2 , s c , s1c , s c2 }) (10.56)
ds ws ws
dc1
2(d1 s) (d1 {b1 , b c , z1c , z c2 }, s {s1 , s c , s1c , s c2 }) (10.57a)
ds
600 Analytic Methods for Design Practice
Find s1 , s 2 , s c , s1c , s c2
s1 , s c , s1c , s c2 s 2 , s c , s1c , s c2
to minimize f (s1 , s 2 , s c , z1c , z1nc , z c2 , z nc
2 )
subject to ci 0 i 1,2
dc2
2(d 2 s) (d 2 {b 2 , b c , z1c , z c2 }, s {s 2 , s c , s1c , s c2 }) (10.57b)
ds
this example, the complimentary variables are defined for common design
variables and common coupling variables.
The optimization formulation for each discipline is as follows:
(1) Discipline 1
find b1 , bc , z 2c (10.58b)
(2) Discipline 2
dc1 dc dc
subject to c1 c10 u 'sc 1c u 's1c 1c u 's2c 0 (10.60d)
dsc ds1 ds 2
dc2 dc dc
c2 c20 u 'sc 2c u 's1c 2c u 's2c 0 (10.60e)
dsc ds1 ds 2
New target values are generated from Equation 10.60 and transmitted to
the lower levels. The final solutions are b1 6.04083, b2 6.76783,
bc 0.0, s1c 1.10539, s 2c 1.0854 and f 0.49223.
g 2 (b 2 , z 2 ) d 0 (10.61d)
z1 h1 (b1 , z c2 ) (10.61e)
z2 h 2 (b 2 , z1c ) (10.61f)
The system analysis should be performed to solve Equation 10.61. The global
sensitivity information may or may not be needed. The design problem for each
discipline is formulated as follows:
(1) Discipline 1
Given z c2 (10.62a)
find b1 (10.62b)
z1 h1 (b1 , z c2 ) (10.62e)
(2) Discipline 2
find b2 (10.63b)
to minimize f 2 (b 2 , z 2 ) (10.63c)
subject to g 2 (b 2 , z 2 ) d 0 (10.63d)
z2 h 2 (b 2 , z1c ) (10.63e)
Equations 10.62a and 10.63a represent the coupling variables obtained from the
system analysis. If the global sensitivity is available, the coupling variables in
Equations 10.62 and 10.63 can be replaced to consider the changes of the coupling
variables with respect to the changes of the design variables by the following
equations:
dz1c
z1c c
z10 'b 2 (10.64a)
db 2
dz c2
z c2 z c20 'b1 (10.64b)
db1
604 Analytic Methods for Design Practice
System analysis
Design variables
c
where z10 and z c20 are the outputs of the system analysis for z1c and z c2 .
Equation 10.64 may not be used. That is, the coupling variables are considered as
constants in the optimization process of a discipline.
The flow of MDOIS is illustrated in Figure 10.14, which represents the case
where global sensitivities are utilized. MDOIS is a practical method in design
practice. The design proceeds until IF and MF are satisfied. Design is carried out
for each discipline.
From the system analysis, the coupling variables ( z1c and z 2c ) are found. The
global sensitivities are not used in this example. The common design variable
vector is considered in discipline 1.
The disciplines have the following formulations:
(1) Discipline 1
Given z 2c (10.65a)
find b1 , bc (10.65b)
(2) Discipline 2
find b2 (10.66b)
The results of the disciplines are used by the system analysis. The final solution is
b1 3.02838, b2 3.75068, bc 3.02838 and f 0.50149.
In the previous sections, various MDO methods were introduced. MDO methods
basically solve the nonhierarchical system in Figure 10.2a directly or by
transforming it to the hierarchical system in Figure 10.2b. The problem
formulation and the solution process can be newly defined according to the
characteristics of the problem. Therefore, the methods introduced are several
methods selected out of many methods. There are many other MDO methods.
MDO methods are classified into single level methods and multilevel methods.
MDF, IDF, AAO and MDOIS are single level methods, while CSSO, BLISS and
CO are the multilevel methods. The single level methods are simple. However,
the characteristics of multidisciplines are not exploited in single level methods
such as MDF, IDF and AAO methods. The multilevel methods are complicated to
use.
MDO methods are also classified according to the existence of the system
analysis. The system analysis finds the coupling variables between analyses.
MDF, CSSO, BLISS and MDOIS have the system analysis. They have the
advantage that MF (multidisciplinary feasibility) is always satisfied during the
design process. However, many system analyses are required in MDF. In some
multilevel methods, the results of the system analysis and the global sensitivity in
the upper level are transmitted to disciplines in the lower level. The global
sensitivity is generally fairly expensive to calculate.
IDF and CO have the advantage that they do not need the system analysis. IF
(individual feasibility) is satisfied, but MF is not satisfied during the design
process. Moreover, the analysis cost in each discipline can be considerably high.
To overcome this difficulty, AAO uses evaluators instead of analyzers. This is the
606 Analytic Methods for Design Practice
u
MDF single O O O always always
u
u
u
IDF single O O always At optimum
u
u
u
AAO single O O At optimum At optimum
u
u
CSSO multi O O O always always
u
BLISS multi O O O O always always
u
u
u
u
u
u
MDOIS single O O always always
Multidisciplinary Design Optimization
607
608 Analytic Methods for Design Practice
10.5 Discussion
Multidisciplinary design optimization (MDO) was created to manage design
problems defined throughout multidisciplines. The formulation of MDO can be
relatively easily defined. However, it is very difficult to find a universal method to
solve MDO problems. Efforts are still needed to solve such problems.
The first difficulty in MDO is the coupling aspect between analyses. For
example, fluid mechanics and solid mechanics are coupled in aeroelasticity
analysis for airplane wings. Solving this coupling may not be impossible, but
difficult. Moreover, it is extremely difficult to include it in optimization since it
requires a large number of analyses. Some MDO methods need a complicated
mathematical process such as global sensitivity or optimum sensitivity. Therefore,
MDO methods are rarely applied to aeroelasticity problems.
There are engineering problems that have weak couplings between disciplines.
It is relatively easy to apply an MDO method to such problems. Some problems
need large-scale analysis. Even single disciplinary optimization is not easy to use
in such problems. As analysis tools are developing significantly, many problems
are enormous in size. Therefore, the exact process of the introduced MDO
methods may not be easy to use in practical design.
Although there are no coupled aspects between analyses, multiple analyzers
can be utilized in some problems. Recently, such problems have been considered
as MDO problems. Also a problem that has common design variables or common
constraints can be regarded as an MDO problem.
Due to the above difficulties, a different MDO method may be utilized for
different applications. Sometimes, a combination of methods is used or a new
method is defined using the ideas of existing methods. Approximation is
frequently adopted to solve complicated problems. The response surface method
introduced in Chapter 6 is frequently employed especially when analysis is
complicated. An example is introduced in the Appendix.
MDO is used to solve large-scale design problems. There is still no universal
method and, therefore, more research is anticipated in this area.
Multidisciplinary Design Optimization 609
Stiffness
Structure Occupant
Belt load
Weight HIC
Displacement Acceleration of the chest
x
S3 S4
S1
S5 S6
y
z S7
S2
S9
S8
S10
Figure 10.A.3. Energy absorbing steering system and its model for occupant analysis
Start
Given GS
find force–displacement (f–d) curve profile
Update of the seat back frame
f–d curve T
S 0 2
to minimize E ³ [G (t ) G (t )] dt
0
subject to the results of occupant analysis
No Converged?
(f–d curve)
Yes
Stop
Figure 10.A.4. Flow chart for the system analysis in MDOIS applied to a BIS design
taking crashworthiness into consideration
Multidisciplinary Design Optimization 613
Figure 10.A.5. Models of the BIS for system analysis and discipline optimization
614 Analytic Methods for Design Practice
ͤͦ
ͤ͡
ͣͦ
Force (kN)
ͣ͡
GZ
Iteration 3
G[
Iteration 4
ͦ͢
͢͡
͡
͡ ͦ ͢͡ ͦ͢ ͣ͡ ͣͦ ͤ͡
Displacement (mm)
6
5
4
Force (kN)
3
2
1
0
0 50 100 150
Time (ms)
Figure 10.A.7. Determined belt load in the system analysis at the first cycle of MDOIS
Objective function
1.2 Structure
1 Occup ant
function
0.8
Objective
0.6
Objective
0.4
0.2
0
0 1 2 3 4 5 6
Iteration
where b 0 is a design variable vector whose elements are the scale factors of the
predefined force–displacement curve profiles for a seat belt, an airbag mass flow,
and an energy absorbing steering column.
Because the results of the involved analyses are very noisy, local sensitivity
information is not as meaningful as that of the smooth response problem. Thus, it
is quite difficult to apply a gradient-based optimization technique to raw responses
of nonlinear structural analysis and occupant analysis. DOE or some nongradient-
8000
7000
6000
5000
No. 1
Load
4000
No. 2
3000 No. 3
2000 No. 4
1000 No. 5
0
0 0.5 1 1.5 2
Displacement
Figure 10.A.9. Force–displacement curve profile history of the seat back through the entire
cycles of MDOIS
Multidisciplinary Design Optimization 617
2
0 1 2 3 4 5 6
Iteration
based optimization techniques could be candidates for the design tool for the
present example. We chose the RSAO algorithm in Section 9.8.2 to apply the
gradient-based optimization technique to the present example. The reason is that
gradient-based optimization is very fast once the problem is well formulated and
the responses are smooth enough to be differentiated.
Each discipline is independently designed after the interdisciplinary
relationship is solved by the system analysis. The solution is converged in six
cycles. Figure 10.A.8 is a design history of the objective functions for each
discipline. They are normalized by the initial values because the orders of the
1.2
1
Design variable value
0.8
0.2 DV5
0
0 1 2 3 4 5 6
Iteration
Table 10.A.1. Details of the design variables in the structural discipline and the occupant
discipline
Summary
One of the MDO methods, MDOIS is applied to the design of the belt integrated
seat. In the original MDOIS, all the processes are mathematically defined by using
the system analysis, sensitivity analysis, etc. However, appropriate approximation
methods can be used in design practice.
References
ANSYS User’s Manual Version 8.1 (2004) ANSYS Inc., Canonsburg, PA
Balling RJ, Sobieszczanski-Sobieski J (1996) Optimization of Coupled Systems: A
Critical Overview of Approach. AIAA Journal 34(1):6í17
Barthelemy JF, Sobieszczanski-Sobieski J (1983) Optimum Sensitivity Derivatives
of Objective Function in Nonlinear Programming. AIAA Journal 21:913í915
Multidisciplinary Design Optimization 619
A
ABAQUS .............................................458 axiomatic design ................................7, 17
acceleration constraint..........................285 axiomatic design (related to robust design)
active constraint ...........................124, 148 .........................................................428
active set strategy .................................147
additive statistical model..............317, 331 B
additivity backward difference method ................201
- axiomatic design..............................33 barrier function method........................153
- design of experiments (DOE) ........329 basic feasible solution ..........................131
adjoint variable method................206, 218 basic solution .......................................131
admissible function ..............................275 beam adjuster .........................................82
airbag ...................................................502 beam element .......................................176
all-at-once (AAO) ................................582 belt integrated seat (application of MDO)..
all-in-one ..............................................579 .........................................................609
aluminum space frame (ASF) ..............474 belt integrated seat (BIS)......................530
analysis of mean (ANOM) ...........329, 405 bilevel integrated system synthesis
analysis of variance (ANOVA) ............312 (BLISS)............................................593
ANSYS ..................................................97 Brachistochrone problem .....................107
antisynergistic interaction ............320, 340 Broyden–Fletcher–Goldfarb–Shanno
artificial variable .................. 227, 256, 302 (BFGS) Method ...............................143
approximation
- force approximation ......................224 C
- global approximation in dynamic canonical form .....................................132
response.......................................276 central difference method.....................201
- Hessian matrix.......................144, 152 characteristic function
- inverse Hessian matrix ..................143 - design of experiments (DOE) .......311,
- linear approximation..... 146, 166, 188, 312, 331, 351
223, 585 - robust design .........................407, 413
- local approximation in dynamic chest injury...........................................501
response.......................................277 chest severity index (CSI) ............501, 536
- quadratic approximation................166 collaborative optimization (CO)...........597
- response surface method................360 common constraint ...............................567
- structural optimization...................223 common design variable ......................566
augmented Lagrangian method ...154, 263, common objective function..................566
274 common range........................................36
automobile body...........................498, 549 compatibility condition ................580, 597
automobile door ...................................513 complexity .............................................33
axiom .....................................................17 complimentary variable................581, 597
622 Index
sum of squares......................................312 U
surplus variable ....................................126 ultra light steel auto body (ULSAB) ....244
switch coefficient .................................589 unconstrained optimization
synergistic interaction ..........................320 - Broyden–Fletcher–Goldfarb–Shanno
system analysis.....................................566 (BFGS) method ...........................143
system range...........................................36 - conjugate gradient method ............141
- Davidon–Fletcher–Powell (DFP)
T method.........................................142
Taguchi method....................................399 - quasi-Newton method....................142
tailor welded blank (TWB) ..................244 - steepest descent method ........124, 138
target value uncontrollable factor ....................395, 402
- collaborative optimization (CO)....599 uncoupled design........................18, 22, 25
- Taguchi method.............................400 uniform distribution .............................430
Taylor expansion..................................146 Unigraphics ............................................97
Taylor series.........................................165 unit step function....................................43
t-distribution.........................................384 UNIX .....................................................95
time domain .................................256, 281 upper bound .........................................108
time finite element method...................274 upper level
time grid point..............................257, 260 - axiomatic design........................23, 24
time-dependent constraint ............257, 263 - multidisciplinary design optimization
tolerance (MDO).........................................567
- axiomatic design........................31, 35
- robust design .................................401 V
tolerance design ...........................401, 404 variance........................................382, 394
tolerance range ...............................38, 421 variance of the error .............................312
top down approach .................................70 variational method................................217
topology optimization ............10, 237, 467 vertex point ..........................................131
total potential energy............................107 virtual variable .....................................127
trade-off coefficient..............................588 VisualDOC...................................545, 562
transformation matrix...........................180 V-model .................................................68
transformation method .........................153
transient analysis..........................276, 281 W
transient load........................................256 web locking retractor (WLR) ...............535
TRIZ...................................................6, 15 weighting factor ...................................407
two-step optimization...................404, 413 worst case approach .............................258
two-way factorial design ......................317
Y
YAG laser ..............................................82
Z
zigzagging process ...........................23, 66