Piezoelectric Actuator Design via

Multiobjective Optimization Methods

zur Erlangung des akademischen Grades eines

DOKTORS DER INGENIEURWISSENSCHAFTEN (Dr.-Ing.)
der Fakultät für Maschinenbau
der Universität Paderborn

genehmigte
DISSERTATION

von
M. Eng. Bo Fu
aus Sichuan, VR China

Tag des Kolloquiums: 09.06.2005

Referent: Prof. Dr.-Ing. Jörg Wallaschek
Korreferent: Prof. Dr. Michael Dellnitz
Foreword

Piezoelectric actuators find widespread applications in almost all fields of engineering.

Ultrasonic welding, traveling wave motors or ultrasonic scalers are examples of systems
operated in resonance. Diesel injection valves, optical scanners and atomic force microscope
are examples of systems operating in a quasistatic mode.

The performance of piezoelectric actuators in a given application is mainly determined from

the tuning of the actuator characteristics to the load characteristics. Except for some very
limited special cases, no simple general design guidelines can be given today. Mathematical
methods for the optimization of piezoelectric actuators are therefore a very important tool for
the system designer.

In most engineering design tasks, multiple optimization criteria must be met. Despite
tremendous progress in the mathematical sciences, knowledge on methods and algorithms for
multicriteria optimization problems is limited in the engineering community.

Bo Fu was a scholar at the Paderborn Institute for Scientific Computation (PaSCo) and his
thesis work was funded within the DFG-Graduiertenkolleg “Scientific Computation:
Application-oriented Modeling and Development of Algorithms”. His thesis concentrates on
the study of multicriteria optimization problems arising in the design of piezoelectric
actuators and on mathematical methods which were developed for this class of optimization
problems. Classical and novel methods are studied in detail and applied to various typical
design problems of piezoelectric transducers. The results are of interest for all engineers
working on the design of piezoelectric actuators, which are interested in mathematical
background of multicriteria optimization, as well as for all mathematicians working in
multicriteria optimization, which are interested in engineering applications.

Paderborn, September 2005

(Prof. Dr.-Ing. Jörg Wallaschek)

Acknowledgments

First of all, I would like to express my sincere gratitude and appreciation to my thesis advisor,
Prof. Dr.-Ing. Jörg Wallaschek, for his continuous encouragement, support and patience
throughout my Ph.D. study. His observations and comments helped me to establish the overall
direction of my research and to move forward with investigation in depth, and his technical
and editorial advice was essential to the completion of this dissertation. I thank him for giving
me the opportunity to work with him.

I would also like to thank my co-advisor, Prof. Dr. Michael Dellnitz, for providing many
valuable comments and suggestions during the work. I am furthermore grateful to the other
members of my examining committee, Prof. Dr.-Ing. Roland Span and Prof. Dr.-Ing. Ansgar
Trächtler, for their valuable comments.

I would additionally like to thank all members of the Mechatronics and Dynamics research
group at the Heinz Nixdorf Institute for their help and support over the last few years.
Particular thanks go to Dr.-Ing. Tobias Hemsel for the valuable comments that were very
helpful in improving the presentation and contents of this dissertation. Thanks also to Dipl.-
Ing. Reinhard Böer, Dr.-Ing. Thomas Sattel, Dipl.-Ing. Michael Brökelmann, Dipl.-Ing. Maik
Mracek, Dipl.-Ing. Rafal Krol und Dipl.-Ing. Christian Potthast for their discussions and
support. Special thanks to Mrs. Marina Kassühlke and Mrs. Kerstin Hille for their help and
support throughout my Ph.D. study.

All the members of the Graduiertenkolleg of the Paderborn Institute for Scientific
Computation (PaSCo) are also due many thanks for the pleasant environment, assistance and
friendship. Particular thanks to [Link]. Fang Wang for several disscussions and Dipl.-Ing.
Nicolai Neumann for proof-reading some of the chapters.

I gratefully acknowledge the China Scholarship Council, the Graduiertenkolleg of the PaSCo
and the Heinz Nixdorf Institute for providing scholarships to pursue doctoral studies in
Germany.

Finally, last, but not least, I would like to thank my family, my wife Ling and son Yiheng, for
their love, encouragement, patience and understanding during the past few years.

Thank you all.

To my mother, my father, my wife and my son
CONTENTS I

Contents

1 Introduction .......................................................................................................1

1.1 Motivation ............................................................................................................. 1

1.2 Objective ............................................................................................................... 2
1.3 Scope ..................................................................................................................... 2

2 Piezoelectric Actuators ....................................................................................5

2.1 Piezoelectric Effect................................................................................................ 5

2.2 Piezoelectric Actuators.......................................................................................... 8
2.3 Models of Piezoelectric Actuators ........................................................................ 9
2.3.1 Nonparametric Models .............................................................................. 9
2.3.2 Continuum Models .................................................................................. 12
2.3.3 Finite Element Method ............................................................................ 18
2.3.4 Lumped Parameter Models...................................................................... 18
2.4 Typical Design Goals .......................................................................................... 23
2.4.1 One Stroke Driving.................................................................................. 23
2.4.2 Resonant Driving..................................................................................... 23
2.5 State of the Art of Optimization of Piezoelectric Actuators ............................... 24

3 Multiobjective Optimization Methods ............................................................29

3.1 Basic Concepts of Multiobjective Optimization ................................................. 29

3.2 Traditional Multiobjective Optimization Methods.............................................. 31
3.2.1 Weighted Sum Method............................................................................ 32
3.2.2 ε-Constraint Method ................................................................................ 33
3.2.3 Weighted Metric Methods ....................................................................... 33
3.2.4 Value Function Method ........................................................................... 34
3.3 Multiobjective Evolutionary Algorithms ............................................................ 35
3.3.1 Basic Principles of Evolutionary Algorithms.......................................... 35
3.3.2 Fitness Assignment and Fitness Sharing ................................................. 39
3.3.3 General Procedures of Multi-objective Evolutionary Algorithms .......... 41
3.3.4 Multiobjective Genetic Algorithm ......................................................... 42
3.3.5 Non-dominated Sorting Genetic Algorithm ........................................... 43
3.3.6 Strength Pareto Evolutionary Algorithm ................................................ 43
II CONTENTS

3.3.7 Elitist Non-dominated Sorting Genetic Algorithm.................................. 45

3.4 Constraint Handling in Multiobjective Evolutionary Algorithms....................... 45
3.4.1 Methods based on Preserving Feasibility of Solutions............................ 46
3.4.2 Methods based on Penalty Functions ...................................................... 46
3.4.3 Methods based on Feasibility and Domination of Solutions................... 47
3.5 Performance Metrics for Evaluating MOEAs ................................................... 48

4 Multiobjective Optimization of Piezoelectric Transducers ..........................51

4.1 Langevin Transducers ......................................................................................... 51

4.2 Performance Criteria ........................................................................................... 53
4.3 Determination of Optimal Prestress .................................................................... 57
4.3.1 Freely Vibrating Transducers .................................................................. 61
4.3.2 Transducers with a Mechanical Load...................................................... 62
4.4 Modeling of Langevin Transducers using Transfer Matrix Methods ................. 65
4.5 Optimization of Symmetrical Langevin-type Transducers ................................. 70
4.5.1 Derivation of the Whole Transfer Matrix................................................ 71
4.5.2 Problem Formulation of the Symmetrical Transducer without Loads .... 73
4.5.3 Implementation of the Optimization Process .......................................... 78
4.6 Optimization of Langevin-type Transducers with a Stepped Horn..................... 91
4.6.1 Derivation of the Whole Transfer Matrix................................................ 91
4.6.2 Problem Formulation of Transducers without Loads .............................. 92
4.6.3 Implementation of Optimization ............................................................. 95
4.6.4 Problem Formulation of Transducers with a Mechanical Load ............ 105
4.6.5 Implementation of Optimization ........................................................... 108

5 Results and Discussions..............................................................................113

5.1 Discussion of the Results for the Symmetrical Transducer ............................. 113
5.1.1 Analysis of the Results of the Optimization.......................................... 114
5.1.2 Determination of the Preferred Solution ............................................... 116
5.2 Discussion of the Results for Stepped-horn Transducers without Load ........... 122
5.2.1 Analysis of the Results of the Optimization.......................................... 124
5.2.2 Determination of the Preferred Solution ............................................... 125
5.3 Discussion of the Results for Stepped-horn Transducers with Load ................ 132
5.3.1 Analysis of the Results of the Optimization.......................................... 134
5.3.2 Determination of the Preferred Solution ............................................... 135
CONTENTS III

5.3.3 Load Characteristics of Stepped-horn Transducers............................... 139

6 Summary and Outlook ..................................................................................141

Appendix A Derivation of Individual Transfer Matrix ........................................145

Appendix B Developed Programs.......................................................................149

Bibliography...........................................................................................................151
LIST OF SYMBOLS V

List of Symbols

α electromechanical transformation factor

Ap cross area of piezoelectric elements

Ab cross area of metal blocks

βS , β T dielectric impermeability matrix

c modal stiffness

cL load stiffness

cD , cE elastic modulus (Young’s modulus) matrix

cp, cm wave speed

C electric capacitance

d modal damping

dL load damping

d piezoelectric charge constant matrix

Dp diameter of piezoelectric elements

Dt outer diameters of back blocks, front blocks, input side of horns and piezo-rings

Dt2 outer diameter of the output end of horns

D dielectric displacement vector

εS , εT absolute dielectric constant matrix

e piezoelectric constant matrix

f frequency

E elastic modulus (Young’s modulus)

E electric field strength vector

Ec coercive field strength

F force
VI LIST OF SYMBOLS

g piezoelectric voltage constant matrix

hp thickness of piezoelectric rings/discs

H piezoelectric constant matrix

j −1

k, keff coupling factor

kp wave number

Kϕϕ dielectric matrix

Kuu stiffness matrix

Kuϕ
piezoelectric matrix
,Kϕ u

Lb length of back and front blocks

Lf1, Lf2 length of the first part of horns, length of the second part of horns

Lp length of piezoelectric elements

Lm mechanical inductance

λp power efficiency

m modal mass

M piezoelectric quality number

Mp mass matrix

Ν number of piezoelectric rings, population size

η efficiency

ηe electrical loss factor

ηm mechanical loss factor

ηp piezoelectric loss factor

ϕe , ϕm phase difference

ϕ electric potential field

pe electrical power
LIST OF SYMBOLS VII

P polarization

Pa, Pma apparent electrical and mechanical power

Pe, Pm effective electrical power and mechanical power

Pr remanent polarization

Pr prestress

ρ density

Q electric charge

Qm mechanical quality factor

R electric resistance

sD , sE elastic compliance matrix

S strain vector

t time

tan δ loss factor

Typb material of back and front blocks

Typf material of horns

Typp material of piezo-rings

T stress vector

u, u3 displacment

U voltage

v velocity

ω angular frequency

w1 , w2 pseudo-weight of objective 1, pseudo-weight of objective 2

Ω exciting angular frequency

ζ diameter transformation ratio

In this dissertation, complex quantities are characterized by underline. Amplitudes are

characterized by a hat “^”.
LIST OF PUBLICATIONS IX

Parts of this work have already been published in terms of conference contributions:

Fu, Bo; Hemsel, Tobias; Wallaschek, Jörg: Model-based Diagnosis for Sandwiched
Ultrasonic Transducers. Proceedings of the 18th International Congress on Acoustics, ICA
2004, Kyoto, Japan Bd. 3, 2004, S. 2243-2246
Fu, Bo; Hemsel, Tobias; Wallaschek, Jörg: Multiobjective Optimization of Piezoelectric
Transducers using Evolutionary Algorithms. Proceedings of 2nd International Workshop on
Piezoelectric Materials and Applications in Actuators, Paderborn, 2005, (in print)
Fu, Bo; Hemsel, Tobias; Wallaschek, Jörg: Piezoelectric Transducer Design via
Multiobjective Optimization. World Congress Ultrasonics 2005, Beijing, P. R. China, (to be
published in 2006)
INTRODUCTION 1

1 Introduction

1.1 Motivation

Over the last decades the demand for piezoelectric actuators has been increased significantly.
Different types of piezoelectric actuators such as high strain multilayers, ultrasonic
transducers, converters and ultrasonic motors etc. have been developed. They have been
widely used in engineering fields like e.g. micro-positioning, active vibration control,
ultrasonic welding and machining. Piezoelectric actuators, like almost all mechanical systems,
are always expected to be “optimal”. However, the design of piezoelectric actuators is a
difficult task, because the overall characteristics of piezoelectric actuators are affected by
various factors such as dimensions of active and passive parts, inherent properties of
piezoelectric materials, electrical and mechanical boundary conditions, etc. Obviously,
empirical and intuitive design methods cannot well accomplish the design task. Systematic
optimization techniques are the most appropriate tool to be used in the design task of
piezoelectric actuators.

The design of a piezoelectric actuator can be described as a process of finding optimal design
variables like e.g. the dimension and material types of piezoelectric elements and mechanical
parts, which minimize or maximize a certain number of objectives like e.g. output amplitude,
input power, output power, subject to a certain set of specified requirements, like e.g.
resonance frequency, limits on input voltage and pre-stress, geometric constraints, etc. This is
a constrained multiobjective optimization problem (MOP) involving continuous and discrete
design variables. Commonly, in this problem some of the multiple objectives are conflicting
to each other and it is impossible to find a solution at which each objective function gets its
optimal value simultaneously. In this case, one tries to find a set of optimal compromises
namely Pareto-optimal solutions from which the designer can select one.

MOPs are often solved by traditional methods based on scalarization techniques. The
common ground of these methods is to convert a multiobjective (vector) optimization
problem into a single (scalar) optimization problem. After transformation the widely
developed theory and methods for single objective optimization can be used. However,
traditional methods show some difficulties. For example, they require some problem
knowledge before optimization is performed. Some techniques may be sensitive to the shape
of the Pareto-optimal front. Moreover, they require several optimization runs to obtain an
approximation of the Pareto-optimal set. Evolutionary algorithms (EAs), on the other hand,
are able to find multiple Pareto-optimal solutions in a single simulation run due to their
population-approach. They are well suited for multiobjective optimization and the problems
with discrete variables. Indeed, from the point of view of engineering only one best solution
needs to be implemented. If one knows the exact trade-off among objectives before the
problem is solved, there is no need to find multiple solutions. The exact trade-off between
2 CHAPTER 1

objective functions, however, is usually not clear before optimization is performed. Therefore,
it is better to find the Pareto-optimal set and then select one solution by some criteria.

Over the past decades many multiobjective optimization methods in particular multiobjective
evolutionary algorithms (MOEAs) have been developed. It is meaningful to apply them in the
design of piezoelectric actuators in order to improve performances of products and reduce the
cost of the product development. Although some studies concerning optimization of
piezoelectric actuators have been published and some optimization methods have been used,
these studies mainly concentrate on single objective optimization and continuous variables.
Multiobjective optimizations of piezoelectric actuators including continuous and discrete
design variables, especially the formulation of problems and the application of various
MOEAs have not been reported. Obviously, there is a need for a contribution to this topic.

1.2 Objective

The objective of this thesis work is to present an integrated procedure for piezoelectric
actuator design via multiobjective optimization methods. The work concentrates on the
formulation of MOPs, the application of different multiobjective optimization methods, the
evaluation of optimized results and the determination of the preferred design. There exist a
large variety of multiobjective optimization methods and it is not possible to apply every
existing method. The thesis mainly concentrates on the application of MOEAs. Regarding
classical scalarization methods the algorithms of single objective optimization will not be
stressed but the ways to converting the MOP to a single objective problem will be pointed
out. As far as optimization problems of piezoelectric actuators are concerned, one practical
MOP, namely the two-objective optimal design of Langevin-type transducers involving
continuous and discrete design variables, is considered. The transfer matrix method based on
continuum models of piezoelectric and mechanical rods and the lumped parameter method
based on the electromechanical analogies are mainly applied in formulation of optimization
problems.

1.3 Scope

The present thesis work is organized as follows:

In chapter 2, fundamentals of piezoelectric actuators are presented. First, the piezoelectric

effect and typical piezoelectric actuators are introduced. Then, four models of piezoelectric
actuators namely nonparametric models, continuum models, finite element models and
lumped parameter models based on the electromechanical analogies are described. After that,
typical design goals for typical piezoelectric actuators are presented. Finally, the state of the
art of optimization of piezoelectric actuators is reviewed.
INTRODUCTION 3

Chapter 3 presents multiobjective optimization methods. The basic concepts of multiobjective

optimization are first introduced. Classical multiobjective optimization methods (scalarization
methods) are then described briefly. Finally, five MOEAs are studied.

Chapter 4 presents the multiobjective optimization of Langevin-type transducers. First, the

Langevin transducer is introduced. Then performance criteria for the transducer are described.
After that, optimal prestress of the Langevin transducer for multiple objectives is discussed.
The modeling of the transducer using the transfer matrix method based on continuum models
is then described. Finally, MOPs for the design of Langevin-type transducers are formulated.
These problems are then solved by two-level optimization using MOEAs described in chapter
3. The first level optimization is performed in this chapter.

In chapter 5, the second level optimization is preformed and the preferred design is
determined. First, the non-dominated solutions are searched in the results of the first level
optimization obtained in chapter 4 for the transducers with 2, 4 and 6 piezo-rings. Second, the
results of the second level optimization are analyzed and the preferred design is determined
using various methods. Finally, the load characteristics of the Pareto-optimal transducers are
discussed.

The last chapter gives a summary to this thesis work and outlook for future work.
PIEZOELETRIC ACTUATORS 5

2 Piezoelectric Actuators

For the optimization of piezoelectric actuators two different kinds of techniques are needed.
One is the modeling of the actuators; the other is an appropriate optimization method. In this
chapter the most important fundamentals concerning piezoelectric actuators are described.
First, the piezoelectric effect and piezoelectric actuators are introduced. The models of
piezoelectric actuators are then described. After that, the typical design goals for different
types of piezoelectric actuators are present and the state of the art of optimization of
piezoelectric actuators is reviewed.

2.1 Piezoelectric Effect

The piezoelectric effect was first discovered by Pierre and Jacques Curie in 1880. They found
that when mechanical stress T was applied to a particular crystal material, electrical charges Q
were generated and a voltage U between the surfaces of the material was generated. About
one year later the inverse piezoelectric effect, which is the fundamental for applying
piezoelectric elements as actuators, was also examined by the Curies [Ike96].

The piezoelectric effect can appear only in crystals which have polar axes, i.e. in the crystals
which possess no center of symmetry. Many naturally occurring crystals, e.g. quartz,
tourmaline and Rochelle salt have this character. Piezoelectric ceramics (PZT-based
ceramics), which are probably the most important piezoelectric materials, may be considered
as a mass of randomly oriented piezoelectric crystallites. These materials have a perovskite
crystal structure. In the high temperature no-polar phase there is no spontaneous polarization
because the crystallites are cubic symmetric. Below the transition temperature known as the
Curie point, i.e. in a non-centrosymmetric ferroelectric phase, spontaneous polarization occurs
and the crystallites exhibit the tetragonal or rhombohedral structures. In the unpolarized state,
the polarizations of grains (crystallites making up a polycrystalline ceramic) or domains are
randomly oriented so that no overall polarization or piezoelectric effect appears. The
piezoelectricity of the polycrystalline ceramics may, however, be aligned in any chosen
direction by a poling treatment which involves applying a strong electric field (>3 kV/mm) to
align the polarization direction of each grain or domain as much as possible, heating the
material to beyond its Curie point and cooling the material below this point to “lock” the
domain structure. The polycrystalline ceramic therefore becomes anisotropic and has a
permanent polarization as if it were a single-domain crystal [Cul96] and [Uch97]. Details
concerning the configuration and fabrication of piezoelectric ceramics are described in
[JC71], [Phi 88] and [Set02].

The piezoelectric effect couples the electrical and mechanical behavior. It can be
approximately described by the following linear piezoelectric constitutive fundamental
equations [Ike96], [Phi91] and [Hem01]:
6 CHAPTER 2

S = sET + d tE
(2.1)
D = dT + ε T E

In the above equations the full tensor notation has been abbreviated into simpler matrix
notation. D, E, S and T are the vectors of the dielectric displacement, the electric field
strength, the strain and the stress respectively. ε T , sE and d are the matrices of the absolute
dielectric constant, the elastic compliance and the piezoelectric charge constant. The
superscripts T and E refer to the quantities to be kept constant (T=E=0) when measuring
material constants [IE87], [DI483] and [DI324]. Similar relations are found for various
choices of independent variable sets. The constitutive equations can be arranged in several
ways and the four types of the piezoelectric constitutive equation set are shown in Table 2.1.
By calculations one set of constants can be expressed by another, see e.g. [DI483] and
[DI324].

As poled piezoelectric ceramics are anisotropic, the material constants depend on the
directions of the stimulation and reaction. These are described by a (right hand) Cartesian
coordinate system. Usually the direction of polarization is taken to be that of the 3-axis (z
axis), see Fig. 2.1. The material constants are complied in the following elastoelectric matrix,
see Table 2.2. The material constants are generally written with two subscripts. The first
subscript indicates the direction of the excitation, and the second gives the direction of of the
system response. It is worth noticing that not all elements of the elastoelectric matrix are
occupied.

These material constants are identified by the measurement on geometrically simple

specimens under a small electric excitation [DI324] and [Rus95]. They are only applied for
the description of the small signal behavior and subject to a high fluctuation range [DI324]
and [Hem01].

Table 2.1 Types of fundamental piezoelectric constitutive equations

Independent variable Piezoelectric constitutive equation
S = s E T + dE
T, E
D = dT + ε T E
T = c DS − hD
S, D (2.2)
E = −hS + β S D
T = c E S − eE
S, E (2.3)
D = eS + ε S E
S = s D T + gD
T, D (2.4)
E = −gT + β T D
PIEZOELETRIC ACTUATORS 7

3 (z)

polarization
5
2 (y)

4
1 (x)

Fig. 2.1 Coordinate System

For large mechanical strains and large electrical field strengths the material behaves
nonlinear. Fig. 2.2(a) shows the relationship between the applied electric field strength E and
the induced polarization P. There appears a hysteresis caused by the transition of the
spontaneous polarization between the positive and negative directions. Fig. 2.2(b) shows the
relationship between the applied electric field and the strain, i.e. the so-called butterfly
hysteresis curve. When the applied field is small, the induced strain is nearly proportional to
the field. The piezoelectric constant can be obtained from the slope. As the field becomes
larger (i.e. greater than about 100V/mm), the strain curve deviates from this linear trend and
significant hysteresis is exhibited. If the drive electric field passes the coercive field, the
hysteresis changes into a butterfly shape. Such strain hysteresis during an electric field cycle
is caused by the change of the ferroelectric domain status. The detailed physical
interpretations on hysteresis are shown in e.g. [Sch96] and [Uch97]. Commonly, the losses
can be approximately described by complex material constants [Wal00] and [Hem01].

Piezoelectric ceramics are also temperature-dependent and exhibit total depolarization above
the Curie point, which, depending on the composition, ranges from 150°C to 400°C. As the
Curie point is approached the values of the piezoelectric constants decrease.

Table 2.2 Elastoelectric matrix for the stress T and the electric field E as independent
variables
T1 T2 T3 T4 T5 T6 E1 E2 E3
S1 s11 s12 s13 d31
S2 s12 s11 s13 d31
S3 s13 s13 s33 d33
S4 s44 d15
S5 s44 d15
S6 2(s11-s12)
D1 d15 ε11
D2 d15 ε11
D3 d31 d31 d33 ε33
8 CHAPTER 2

P S
Pr

-E 0 +E
Ec

-E 0 +E

a) b)

Fig. 2.2 a) Polarization P as a function of field strength b) Mechanical deformation (strain)

as a function of field strength

2.2 Piezoelectric Actuators

Piezoelectric actuators convert voltage and charge into force and motion. According to the
electrical drive method, they can be divided into two categories: the resonant driven
piezoelectric actuators and the non-resonant driven piezoelectric actuators. The resonant
driven piezoelectric actuators include e.g. ultrasonic transducers (converters), ultrasonic
motors and piezoelectric transformators. The non-resonant driven piezoelectric actuators
include various one-stroke actuators. Their operating frequency range is from quasistatic up to
about half of the first resonant frequency of the mechanical system.

In practice a variation of the electric field strength will deform a piezoelectric body in
different directions with different intensities. According to the type of the utilized
piezoelectric effect (the directions of the applied electric field and the extension),
piezoelectric actuators can be divided into three main groups, see Fig. 2.3.

In longitudinal actuators, also called d33-actuators, the applied electric field, the utilized
extension and the polarization have the same direction. They are usually used for small
movements and high forces and generally have because of high stiffness. Transversal actuator
actuators, also called d31-actuators, have the electric field applied in the direction of the
polarization and the main deformation is in the direction perpendicular to the polarization.
They are also used for small movements and high forces and also have high stiffness. In
flexural actuators, also known as d15-actuators or bimorphs, the electric field is applied in the
direction perpendicular to the polarization and the flexural deformation is utilized. They are
mainly used for large movements and have low stiffness.
PIEZOELETRIC ACTUATORS 9

E P E P E P

a) b) c)
Fig. 2.3 Piezoelectric actuators a) d33 ; b) d31; c) d15

Typical strain levels that can be obtained by piezoelectric actuators are in the region of 0.1%,
occurring at field strength in the region of 1000V/mm. This maximum field strength is limited
to approximately 0.75 of the value of the coercive field. In order to increase the total
displacement capability of a piezoelectric ceramic actuator, multilayer stacks as well as
amplifying mechanisms are often used [Cul96] and [Hen01].

In addition to hysteresis, other effects such as zero point drift, creep and aging are also
encountered in piezoelectric actuators [Sch96] and [Uch97].

A piezoelectric actuator commonly consists of several components including piezoelectric

ceramics and metal housing, which are assembled to form an integral entity. Piezoelectric
actuators have been widely used in various fields such as micro-positioning of tools, active
vibration control, ultrasonic welding and machining, common rail diesel injection systems.

2.3 Models of Piezoelectric Actuators

In any optimization problem two phases can be distinguished: formulation and solution.
Broad experience in solving problems has shown that the time needed to formulate a problem
makes up 70-85% of the total time required for a complete treatment, from the formulation to
results [SM95]. An adequate mathematical model is the fundamental of the formulation of
optimization problems. This section presents four models which can be used to describe the
dynamic behavior of piezoelectric actuators. In order to interpret the models more clearly, a
typical longitudinal piezoelectric actuator as shown in Fig. 2.4 a) is used as an example for
modeling.

2.3.1 Nonparametric Models

The nonparametric model deals with the piezoelectric actuator as a type of “black box” and
directly estimates the impulse or the frequency response of the system. This model does not
impose any assumptions about the actuator, other than that of linearity. In this context, a
piezoelectric actuator can be regarded as an electromechanical four-pole network element,
which has electric input quantities (the voltage U and the current I ) and mechanical output
quantities (the force F and the velocity v) , see Fig. 2.4 (b).
10 CHAPTER 2

F,v F,v I v
F
U

I U
a) b)
Fig. 2.4 a) Piezoelectric actuator b) Piezoelectric actuator as a four-pole network element

For harmonic vibrations of the piezoelectric actuator, the relation between inputs and outputs
can be written as

 Iˆ  y11 y12  Uˆ 

  = y ⋅
y 22   Fˆ 
, (2.5)
vˆ   21

where Û , Iˆ , F̂ and v̂ are complex amplitudes. Y is the transfer matrix, which also is called
the conductance matrix. The elements of the transfer matrix are defined as follows:
Iˆ
y11 = : the short-circuit input admittance (2.6)
Uˆ
vˆ
y 21 = : the short-circuit core admittance (forward) (2.7)
Uˆ

Iˆ
y12 = : the short-circuit core admittance (backward) (2.8)
Fˆ
vˆ
y 22 = : the short-circuit output admittance (2.9)
Fˆ

Fig. 2.5 shows the typical variation of the amplitude magnitude and phase shift of the short-
circuit input admittance y11 as a function of frequency (Bode plot) for a piezoelectric
actuator. Fig. 2.6 gives the corresponding locus of y11 in the complex plane (Nyquist plot).
According to Fig. 2.6, a piezoelectric quality number is geometrically defined as follows
[LeI75].

Yr
M= (2.10)
Yc

where Yr = Ymax-Ymin is the diameter of the locus of y11 as shown in Fig. 2.6. Ymax and Ymin are
the values of y11 at the frequencies fm and fn , respectively.
PIEZOELETRIC ACTUATORS 11

y11

fm fn
( )
Arg y 11 Frequency

fr fa
Frequency
Fig. 2.5 Variation of admittance magnitude and phase angle with frequency

( )
Im y 11
f f

fm
Ymax
fs

fn
Yc Ymin
fp
0 fa Re y11 fr ( )
Fig. 2.6 Frequency response in the complex plane

There are three pairs of characteristic frequencies, which are characteristic for these plots,
namely fs and fp , fm and fn as well as fr and fa . The series resonant frequency fs is the
frequency at which the admittance in the equivalent model becomes infinite if the electrical
loss R and the mechanical loss d are neglected (see section 2.3.4). The parallel resonant
frequency fp is the frequency at which the admittance in the equivalent model becomes zero if
R and d are neglected (also see section 2.3.4). The frequency fm is the frequency at which the
admittance becomes maximum and the frequency fn is the frequency at which the admittance
becomes minimum. The resonant frequency fr and antiresonant frequency fa are the
frequencies at which the phase angle of y11 is equal to zero. These frequencies are
approximately equal for the piezoelectric ceramics with a piezoelectric quality number M>>2,
i.e. fs ≈ fr ≈ fm and fp ≈ fa ≈ fn .

The coupling factor k is a measure of the effective energy conversion in the piezoelectric
actuator. For quasistatic operation, the coupling factor k is defined as follows [Ike96]:
12 CHAPTER 2

transformed, stored mechanical energy

k2 = (2.11a)
supplied electrical energy
or
transformed, stored electrical energy
k2 = (2.11b)
supplied mechanical energy

This definition gives a conceptual picture of the coupling factor, but it is not effective in
determining the value of k 2 of an actual actuator. It is possible to express the coupling factor
based on the frequencies defined above. This results in [Phi91]

f p2 − f s2
k2 = (2.12)
f p2

As an approximation, when k << 1 one may write

f p − fs ∆f
k2 = 2 =2 (2.13)
fp fp

It is worth to point out that though a high k2 is usually desirable for efficient transformation,
the coupling factor must not be mistaken for efficiency, since the losses are not considered in
the coupling factor. The unconverted energy is not necessarily lost (converted into heat) and
can mostly be recovered.

The mechanical quality factor Qm is used as a measure for the resonance rise of the
piezoelectric actuator. It can be derived from the 3dB bandwidth of the admittance at fs if
k2Qm>10 [Rus95] resulting in
fs fs
Qm = = (2.14)
∆f s (3dB) f 2 − f 1

The frequencies f1 and f2 are frequencies that correspond to the admittances that are 3dB lower
than the maximal admittance, respectively.

The advantage of nonparametric models is that they are able to analyze the dynamic behaviors
of all types of piezoelectric actuators, without knowing the concrete structure of the system.
As no parameters are used in the models, obviously, they are not suitable to the optimization
problem. Nonparametric models, however, play an important role in the experimental
characterization of dynamic behaviors of piezoelectric actuators.

2.3.2 Continuum Models

In continuum models the piezoelectric actuator is considered as a mechanical continuum with

additional degrees of freedom for the electrical behavior. Its behavior is described by the
PIEZOELETRIC ACTUATORS 13

partial differential equations of motion, which involve derivatives with respect to spatial
coordinates as well as with respect to time. Moreover, the electromechanical coupling of the
piezoelectric material is approximately described by the constitutive equations [BCJ64] and
[WH00].

It is difficult to find an analytical solution of the partial differential equations of motion of

continuous systems having complex geometry and boundary conditions. Therefore,
continuum models are mostly used for piezoelectric actuators with simplified geometry (rod,
beam, plate etc.) and boundary conditions.

In the following the continuum model of the piezoelectric actuator shown in Fig. 2.4(a) will
be described. Fig. 2.7 gives its dimensions.

Assumptions: It is assumed that the actuator has the shape of a slender rod and only the
longitudinal vibration is taken into account. The problem is then reduced to the one-
dimensional case by following assumptions:

• Dp < λ/4 (slender rod, λ is wavelength)

• T1= T2= T4= T5=T6= 0 (uniaxial stress condition)

• E1= E2 = 0 (non leakage electrical field)

• D1= D2 = 0 (surrounding medium air)

Governing equations The fundamental piezoelectric constitutive relation (T, E)-type (see
Table 2.1) for loss-free piezoelectric material is
E
S 3 = s33 T3 + d 33 E 3 (2.15)
T
D3 = d 33 T3 + ε 33 E3 (2.16)

u3 X1
X2
Dp
X3(z)
Ap
z dz
∂F3
F3 + dz
F3 ∂z

dz
Fig. 2.7 Dimensions of the piezoelectric actuator
14 CHAPTER 2

This is rearranged to give the relation of (S, E)-type as follows

T3 = c33E S 3 − e33 E3 (2.17)

D3 = e33 S 3 + ε 33S E3 (2.18)

where c33E = 1 / s33E , e33 = d 33 / s33E , ε 33S = ε 33

T
− d 332 / s33E

Eliminating E3 gives
e33
T3 = c33D S 3 − D3 (2.19)
ε 33S
2
where c33D = c33E + e33 / ε 33S

When the energy losses in the piezoelectric materials should be taken into account, the
following complex moduli can be introduced into the model
E
s 33 = s33E (1 − jη m ) (2.20)

ε T33 = ε 33T (1 − jη e ) (2.21)

d 33 = d 33 (1 − jη p ) (2.22)

Equations (2.20), (2.21) and (2.22) describe mechanical loss, dielectric loss and piezoelectric
loss, respectively. The loss factors η m , ηe and η p describe the phase shifts between the stress
and the strain, between the electric field strength and dielectric displacement as well as
between the strain and the electric field strength, respectively. The application of complex
moduli in the modeling will be discussed further in chapter 4.

Considering an element of length dz, the application of Newton’s second law to the free body
of the element shown in Fig. 2.7 yields

∂T3 ∂ 2u3
=ρ (2.23)
∂z ∂t2

and the conservation of charge leads to

∂D3
=0 (2.24)
∂z

Substituting equations (2.17) and (2.18) into equations (2.23) and (2.24) as well as using
∂ u3
S3 = (2.25)
∂z

and
PIEZOELETRIC ACTUATORS 15

∂ϕ
E3 = − (2.26)
∂z

yield the following governing equations

∂ 2u3 2
2 ∂ u3
= c p (2.27)
∂t2 ∂ z2

∂ 2ϕ e33 ∂ 2u 3
= (2.28)
∂ z 2 ε 33S ∂ z 2

Equation (2.27) is the one-dimensional wave equation, in which the wave speed is given by
c p = c33D ρ .

Solutions of governing equations

The solution of the wave equation (2.27) is generally obtained by means of the separation-of-
variables method, where u3 (z, t) is expressed as the product of a function of time and a
function of the space coordinate. The solution ϕ (z, t) of the governing equation (2.28) can
also be obtained by means of the same method. If the input voltage is a harmonic vibration
with frequency Ω , the displacement u and the electric potential ϕ are also considered as
harmonic vibrations with the same frequency. Therefore, the function of time in the
separation-of-variables method can be directly expressed as e jΩ t . Using complex variables to
express the displacement and potential fields yield

u 3 ( z , t ) = uˆ 3 ( z )e jΩ t (2.29)

ϕ ( z, t ) = ϕˆ ( z )e jΩ t (2.30)

Substituting equations (2.29) and (2.30) into equations (2.27) and (2.28) yields the following
governing equations

d 2 uˆ 3 2
2
+ k p uˆ 3 = 0 (2.31)
dz

d 2 ϕˆ e33 d 2 uˆ 3
= (2.32)
dz 2 ε 33S d z 2

where kp=Ω / cp is called wave number.

Since two arbitrary constants are required in the solution of a second-order ODE, the general
solution of the equation (2.31) has the following complex form
jk p z − jk p z
uˆ 3 ( z ) = C 1e + C 2e (2.33)
16 CHAPTER 2

Introducing equation (2.33) into equation (2.32) the solution ϕ̂ can be expressed as follows

ϕˆ ( z ) =
e33
ε S
(C e
1
jk p z
+ C 2e
− jk p z
)+ C z + C
3 4 (2.34)
33

The undetermined constants C1, C2, C3 and C4 can be computed from electrical and
mechanical boundary conditions. For example, when the piezoelectric actuator shown in the
Fig. 2.4(a) and Fig. 2.7 is fixed at one side (z=0) and a harmonic force F 3 (t ) = Fˆ 3e jΩt is
applied at the other side (z = Lp), the boundary conditions are:

Fˆ
uˆ 3 ( z = 0) = 0, Tˆ 3 ( z = L p ) = 3
Ap (2.35)
ϕˆ ( z = 0) = 0, ϕˆ ( z = L p ) = −Uˆ

After introducing these boundary conditions into equations (2.33), (2.17), (2.34) and (2.34)
four linear equations in terms of C1, C2, C3 and C4 can be obtained. The unknown constants
are then determined based on solving the four linear equations ( F̂ 3 and Û are assumed as
given quantities). Therefore, the complex amplitudes of the vibration displacement and
electric potential are obtained according to the equations (2.33) and (2.34), respectively.

Similarly, the dielectric displacement D3 ( z , t ) can also be described by means of the

separation-of-variables method as follows

D 3 ( z , t ) = Dˆ 3 ( z )e jΩt (2.36)

Considering the equations (2.25) and (2.26) and introducing the equations (2.36), (2.29) and
(2.30) into the equation (2.18) it follows

d uˆ 3 ( z ) d ϕˆ ( z )
Dˆ 3 ( z ) = e33 − ε 33S (2.37)
dz dz

Substituting the equations (2.33) and (2.34) into the equation (2.37) and with simple
computations, the complex amplitude of the dielectric displacement is given by

Dˆ 3 ( z ) = −ε 33S C 3 (2.38)

The electric current can be derived through integrating the dielectric displacement over the
area Ap of the electrode and then calculating the derivative with respect to time. Considering
∂D3 ∂D3
= = 0 the electric current is given by
∂x1 ∂x2

Iˆ = jΩ Dˆ 3 Ap (2.39)
PIEZOELETRIC ACTUATORS 17

The vibration velocity at the free side (z = Lp) of the actuator can also be obtained through
differentiating the displacement with respect to time. Its complex amplitude is given by
vˆ = jΩuˆ 3 (2.40)

Consequently, the electromechanical transfer behaviors of the actuator as shown in Fig. 2.4
(a) are well-established. In the same way, a four-pole network element which has two input
quantities and two output quantities as shown in the Fig. 2.4 (b) can be used. Its transfer
functions can be described by the same form as expression (2.5). According to the equations
(2.39) and (2.40), the expressions of the elements of the conductance matrix are
2
jΩA p k p L p ε 33S (e33 + c33E ε 33S ) cos(k p L p )
y 11 (Ω) = 2 2 2
(2.41)
k p L p (e33 + c33E ε 33S ) cos(k p L p ) − e33 sin(k p L p )

jΩe33ε 33S sin(k p L p )

y 12 (Ω) = y 21 (Ω) = 2 2 2
(2.42)
k p L p (e33 + c33E ε 33S ) cos(k p L p ) − e33 sin(k p L p )

jΩL p ε 33S sin(k p L p )

y 22 (Ω) = 2 2 2
(2.43)
A p k p L p (e33 + c33E ε 33S ) cos(k p L p ) − A p e33 sin(k p L p )

Obviously, the analytical model makes the “black box” in the nonparametric model clear and
into a “white box”. In the present case a four-pole network element was obtained. For more
general boundary conditions, piezoelectric actuators can also be considered as a six-pole
network element. This will be discussed further in chapter 4.

Using the expression of the short-circuit input admittance y11 shown in equation (2.41) the
series and parallel resonant frequencies of the piezoelectric actuator fixed at one side and free
at the other side can be calculated. According to the definitions described in section 2.3.1, the
series resonance frequency is the frequency at which the admittance y11 becomes infinite if
the energy losses in the piezoelectric material are neglected. This can be achieved when the
denominator of the expression of y11 is equal to zero. Therefore, the series resonant frequency
can be determined by solving the characteristic equation
2 2 2
k p L p (e33 + c33E ε 33S ) cos(k p L p ) − e33 sin(k p L p ) = 0 (2.44)
Similarly, the parallel resonant frequency can be obtained by solving the equation
 ΩL p 
cos =0 (2.45)
 c 
 p 
As continuum models are based on material data and geometrical parameters, they can
generally be used already in the design process for the formulation of optimization problems
of piezoelectric actuators.
18 CHAPTER 2

2.3.3 Finite Element Method

The finite element method (FEM) is an approximation of the continuum models that is
particularly well suited to computation. Generally it is applied in the computation of the
vibration behavior of piezoelectric actuators with complex geometry and boundary
conditions. In the FEM, the continuous structure of the piezoelectric actuator is discretized
into a number of finite elements. Besides the mechanical degrees of freedom ui (translation
and rotation displacements), an electrical degree of freedom ϕi (electrical potential) is added
at each node. The FEM solution vector consists of the displacement values ui and electric
potential values ϕi at the nodes i. The displacement and voltage fields at arbitrary locations
within the element are assumed as a linear combination of the nodal values of these fields,
respectively. The coupled electromechanical field is described by means of the linear
piezoelectric constitutive equations. The analysis then proceeds to writing a set of differential
equations of motion and Maxwell’s equations [Ler90] and [Abb98]. For the nodes of finite
elements a coupled equation system can be obtained as follows [Ans01] and [HKL98]:

M p 0  u&& K uu K uϕ   u   F 
⋅ + ⋅ =
K ϕϕ  ϕ  Q 
 0 (2.46)
 0 ϕ&& K ϕu

where Mp, Kuu , Kϕϕ , Kuϕ (Kϕu) are the mass, stiffness, dielectric and piezoelectric matrices,
respectively. F and Q are the nodal mechanical force and nodal electrical charge vectors,
respectively. In the case of damped systems, a proportional damping, i.e., a damping matrix,
which can be expressed as a linear combination of the mass and stiffness matrixes, is
generally added [Gen95].

Considering time harmonic excitation the above system of equations can be transformed into
an algebraic equation system:

K uu − Ω 2 M p K uϕ   uˆ   Fˆ 
 ⋅  =   (2.47)
 K ϕu K ϕϕ  ϕˆ  Qˆ 

where û , ϕ̂ , F̂ and Q̂ are complex amplitudes. The displacement field and electrical field can
be first obtained by solving this algebraic equation system taking into account boundary
conditions. The eigenfrequencies and eigenfuctions can be obtained by solving eigenvalue
problems [WH00] and [SK99]. The time of numerical calculation will rise remarkably if
damping is considered in the FEM model. For more details concerning FEM applied to
piezoelectric actuators, refer to [Hem01], [HKL98], [Ans01] and [Abb98].

If optimization problems are formulated by means of the FEM, the number of the variables is
usually very large. The cost of numerical computation in the FEM is notably higher than the
cost in continuum models. Therefore, the FEM is not well suited for optimization problems
occurring in the early design stages where the optimal results usually need to be obtained in
PIEZOELETRIC ACTUATORS 19

short time and without high costs. The FEM, however, is usually used when the accurate
geometrical details of the actuator need to be specified.

2.3.4 Lumped Parameter Models

The dynamic behavior of a linear system subjected to a harmonic excitation can usually be
described accurately by superposing only a few of the normal modes. Each normal mode
dominates the vibration behavior of the system in the range of the respective eignenfrequency.
Therefore, if the system is driven in the range of one of its eigenfrequencies, its behavior can
be described with reasonable accuracy by a model with only one degree of freedom.

Based on electro-mechanical analogies the vibration behavior of a piezoelectric actuator

operating in the vicinity of one of it’s eigenfrequencies can be described by an equivalent
mechanical or electrical model as shown in Fig. 2.8 [Len75], [Wal00] and [Hem01]. In these
models m, c, and d are modal mass, modal stiffness and modal damping, respectively. C and
R are electric capacitance and electric resistance. α is the electromechanical transformation
factor that describes the transmission ratio of electrical and mechanical quantities. U and Q
are input voltage and charge, u and F are modal displacement and mechanical load.

According to these models, the dynamics of the system can be described by

mu&& + d u& + cu = αU + F (2.48)

1
C
( ) (
Q − αu + R Q& − αu& = U) (2.49)

The conductance matrix for oscillations with harmonic excitation U (t ) = Uˆ ⋅ e jΩt can be
obtained as

Q 1/C
U
Q u
R
C d 1/c m u
U F F α
R c
m
1:α 1
d

a) b)
Fig. 2.8 Equivalent models for a piezoelectric actuator
20 CHAPTER 2

 1 α2 α 
 + 
 Iˆ 1 − mΩ 2 + jdΩ + c 2
− mΩ + jdΩ + c  Uˆ 
 jRΩ +
  = jΩ  C  ˆ  (2.50)
vˆ   α 1 F 
 2
− mΩ + jdΩ + c − mΩ 2 + jdΩ + c 

where Iˆ = jΩQˆ and vˆ = jΩuˆ .
In particular the following admittance functions can be obtained:
Iˆ jΩ jΩ α 2
y 11 = = + (2.51)
Uˆ jRΩ +
1 − mΩ 2 + jdΩ + c
C
vˆ jΩα
y 21 = = (2.52)
Uˆ − mΩ + jdΩ + c
2

The frequency responses of the above admittance functions can be described by the Bode plot
similar to Fig. 2.5 as well as the Nyquist plot similar to Fig. 2.6. Fig. 2.9 shows the Nyquist
plot of the y11 given by equation (2.51) in the case of R=0. As described in section 2.3.1 there
exist three pairs of characteristic frequencies: ω s and ω p , ω m and ω n as well as ω r and ω a ,
which are approximately equal for a piezoelectric actuator with a piezoelectric quality number
M >> 2 .
These frequencies can be derived from the expressions shown in (2.51) as follows:
c
• The series resonant frequency ωs = (2.53)
m

α2
c+
The parallel resonant frequency ωp = C (2.54)
m

Im(y11) Ω

ωm

ωs
ωn
ωp
ωa ωr Re(y11)

Fig. 2.9 The frequency response of the admittance y11 in the complex plane
PIEZOELETRIC ACTUATORS 21

• The maximum admittance frequency ω m for y11 max

The minimum admittance frequency ω n for y11 min

• The resonance frequency ω r for Im( y11 ) = 0

The antiresonance frequency ω a for Im( y11 ) = 0

Several performance criteria of piezoelectric actuators can be readily expressed by means of

model parameters of the equivalent model. According to equation (2.52) the ratio between the
amplitude of the modal displacement û at the resonance frequency ω s and its static
counterpart can be calculated. This ratio represents the mechanical quality factor Qm. It can be
expressed as

cm c
Qm = = (2.55)
d ωs d

The equivalent models can be reduced by transforming mechanical quantities into the
electrical side of the models using the transformation factor α. Fig. 2.10 shows the reduced
equivalent electrical model, where Rm, Cm and Lm are the mechanical resistance, capacitance
and inductance respectively. They are as follows:

d α2 m
Rm = 2
, Cm = , Lm = (2.56)
α c α2
Then the admittance corresponding to the mechanical characteristic of the actuator is

 1 
Ym = 1  Rm + + jΩLm  (2.57)
 j ΩC m 

In general, the value of R representing the electric loss (mainly dielectric loss) of the
piezoelectric material is small and can be neglected [Uch04]. In the case of R =0, the
admittance corresponding to the electrical characteristic of the actuator is

Q Rm Cm Lm

C
U F
R

Fig. 2.10 The reduced equivalent electrical model

22 CHAPTER 2

Ye = jΩC (2.58)

Therefore the admittance function y11 is given by:

y11 = Ye + Ym (2.59)

The ratio between the magnitudes of the Ym and Ye at the resonance frequency
( Ω = ω s = c m ) defines a piezoelectric quality number

Ym 1 / Rm α2 C
M= at Ω = ω s = = = Qm m (2.60)
Ye ω s C ω s dC C

The piezoelectric quality number is an important performance parameter. It presents the

extent of the phase rise or phase drop of the admittance functions and is appropriate to
classifying piezoelectric actuators concerning electrical behaviors. Fig. 2.11 and Fig. 2.12
show the Bode plot and the Nyquist plot of the admittance function y11 for three groups of
special values of M (here the electric loss R is neglected), respectively. It is noted that M = 2
is a critical value, at which the 0°-line is tangential to the phase-frequency plot of the y11 (see
Fig. 2.11) and the real axis is tangent to the Nyquist plot (see Fig. 2.12). When M > 2 the 0°-
line intersects the phase-frequency plot and the real axis intersects the Nyquist plot
correspondingly. When M<2, the intersections do not exist any more. The values of M will
affect the design of the power electric device for driving the piezoelectric actuator. This will
be further discussed in Chapter 4.

Fig. 2.11 y11 as a function of the frequency for three groups of the quality number M
PIEZOELETRIC ACTUATORS 23

Ω
Ω
Im(y11) Im(y11) Im(y11) ωm
Ω ωm

ωm ωs
ωs
ωs
ωn ωn
ωn
ωp ωp ωp
ωa ωr Re(y11) Re(y11) Re(y11)

M>2 M=2 M<2

Fig. 2.12 y11 as a function of the frequency in the complex plane for three groups of the
quality number M

The above lumped parameter models can well describe the vibration behavior of the
piezoelectric actuator operating in the vicinity of one of it’s resonant frequencies. Fig. 2.13
shows comparisons of measured and calculated admittance frequency responses for a
piezoelectric actuator. As shown in the figure, the calculated results from the expressions
(2.51) and (2.52) do agree with experimental results quite well. Generally, the values of
model parameters can be estimated by experiments [Fu04]. They can also be obtained by the
analytical calculation or FEM. Therefore, the equivalent model can be considered as a “gray
box” model of the piezoelectric actuator. The advantage of equivalent models is that the
dynamic behavior of the actuator can be described by simple algebraic equations and it is
convenient to formulate the performance criteria of actuators.

Fig. 2.13 Comparisons of the frequency responses y11 and y 21 measured and calculated from
lumped parameter models
24 CHAPTER 2

2.4 Typical Design Goals

Piezoelectric actuators have been widely used in different fields. For different applications
there exist different design goals. It is very difficult to describe all design goals for all
application fields. In this section, only typical design goals of piezoelectric actuators for one
stroke driving and resonant driving will be discussed.

2.4.1 One Stroke Driving

The piezoelectric actuators for one stroke driving generally operate well below the
fundamental resonance frequency. They may be divided into two classes. One is usually used
in precision instrumentation, which requires precise but minute motion (<10µm), where the
positioning precision is a typical design goal and only small forces are required. The other
one, which often has displacement amplifying mechanisms, is used in the fields, where large
stroke (>200µm) is needed at large forces, like e.g. the control of helicopter rotors [DLL00].
Here the maximum stroke or maximum magnification factor is a typical design goal while the
accuracy is of minor interest. For piezoelectric actuators used for one stroke driving, the
typical design goals can be concluded as follows
• large stroke per unit of length
• positioning with high precision
• fast response time
• large mechanical energy density
• low power consumption
• compact size

2.4.2 Resonant Driving

The most widely used piezoelectric actuators for resonant driving are piezoelectric
transducers and ultrasonic motors. Piezoelectric transducers are extensively used in acoustic
wave generation devices such as power ultrasonic transducers used in industrial machining,
welding, etc. [Lit03]. For these applications the typical design goals are
• specified mechanical resonance frequency
• sufficient frequency separation between the primary and all secondary (undesired )
resonance frequencies
• high efficiency, low energy loss
• compact size
• low cost
PIEZOELETRIC ACTUATORS 25

In [Hem01] typical design goals for a piezoelectric linear motor were summarized. Different
design requirements for the vibrator, contact and rotor were proposed.

Most of the piezoelectric transducers used in ultrasonic machining and bonding has a
sandwich construction. They are known as Langevin transducers. In chapter 4, the practical
design goals for the special piezoelectric transducers will be further discussed.

2.5 State of the Art of Optimization of Piezoelectric Actuators

The performance of piezoelectric actuators can usually be remarkably improved, if

mathematical optimization methods are applied in their development. According to the
different applications, there exist various optimization problems. As piezoelectric actuators
couple the electric and mechanical quantities, their optimization characterizes
multidisciplinary optimization problems. In recent decades, a number of optimization
problems of piezoelectric actuators have been studied. Some optimization algorithms have
been used successfully. In this section, the state of the art of optimization of piezoelectric
actuators is briefly reviewed. The review will mainly concentrate on the formulation of
optimization problems and the applied optimization methods.

A large number of studies have been carried out into optimization problems of piezoelectric
actuators in intelligent structures. An intelligent structure is the structure that can actively
sense and react to its environment via onboard sensors, actuators and computational/control
capabilities [SC93]. Such a structure provides an efficient method to implement various tasks
such as active control, shape control and health monitoring, see e.g. [CL87], [Cra94],
[LOM00], [CTS02] and [Pre02]. Among the various available materials for actuators of
intelligent structures, piezoelectric materials are excellent since they are light weight and can
be easily incorporated into a structure by surface bonding or embedding. They produce only
less significant inertial loads [CSe94]. It is important to determine optimal dimension,
placement and controls of piezoelectric actuators in the intelligent structure in order to
improve the performances. Current applications of intelligent structures are mainly in the field
of the active vibration control of aerospace structures, which often do not possess sufficient
passive damping to reduce severe vibration often resulting from a minor disturbance force. In
[SC93], [CS94] and [CSJ99] the optimum designs of intelligent structures for the active
vibration control have been studied. These problems are mixed continuous-discrete
optimization problems. There exist continuous and discrete design variables. The continuous
design variables include dimensions and control gains of piezoelectric actuators. The location
of piezoelectric actuators is the discrete design variable. Design criteria included the energy
dissipated by the piezoelectric actuators, electric power requirements, vibration reduction and
the fundamental frequency. Constraints were placed on total energy, input voltage,
displacements and mass. The multiple design objectives and constraints were usually
combined into a single unconstrained function which was then minimized. A simulated
annealing approach was used for the discrete search. A deterministic gradient-based search
26 CHAPTER 2

was used for the continuous search. Gradients of the continuous design variables were
computed by a finite-difference technique. In these works, however, the conflicts between
objectives were not analyzed and Pareto-optimal solutions were not discussed. In [SWW93],
[HL99] and [ZLG00] genetic algorithms have been used to find the optimal placement of
piezoelectric actuators and sensors for active vibration control of a plate. The results showed
that genetic algorithms are effective optimization tools. But multiobjective optimization was
not mentioned in these studies. Besides the works described above, there are many
publications about optimization of actuators and sensors for active control, see e.g. [HM03],
[YV96] and [BMF00]. Active control using piezoelectric actuators has been the subject of
attraction.

Topology optimization of piezoelectric actuators is another subject widely studied. The

purpose of topology optimization is to find the optimal layout of a structure in a specified
domain, i.e., the optimal distribution of the material and void phases in the design domain, in
order to minimize (maximize) defined objective functions or achieve specified properties
[BS03]. Topology optimization comprises several steps. First, the geometry of the design
domain, boundary conditions and loading are prescribed. Second, the design domain is
discretized by finite elements, and each element is assigned a design variable, which
determines the absence or the presence of material at the particular location. Third, the finite-
element analysis and sensitivity analysis are used to give the function value and the first-order
sensitivity (gradient) of the objective and constraint. After that, the optimization algorithm is
used to solve the optimization problem. Finally, the optimum topology is interpreted and
refined. In the literature, there are two classes of methods to solve the topology optimization
problem. One class is to solve the topology optimization problem directly as an integer 0-1
problem [Bec99] and [BS03]. The more effective method to solve this problem is to relax the
design domain, i.e., to replace the integer variables with continuous variables. A material
model that relates the material properties such as the elastic modulus in each point of the
domain to some design variable must be defined, and this model must allow materials with
intermediate properties and not only zero or full material [SNF98]. The relaxation can be
realized by two approaches, namely the homogenization method and the SIMP method (Solid
Isotropic Material with Penalization), see [BK88] and [BS03]. In the former the concept of
perforated microstructure is introduced. The problem is posed as optimizing the material
distribution in a perforated structure with infinite microscale voids. The size of each
microvoid is the design variable. In the latter material properties are assumed constant within
each element and the relative density of each element is the design variable. After relaxation
the mathematical programming techniques with continuous design variables can be used to
solve the optimization problem. The optimum result is a material distribution over mesh with
some intermediate values of density (“gray scale”) that represent the presence of intermediate
materials (composite), which are difficult to fabricate in practice. They can be eliminated by
using some form of penalty in the optimization procedure, see [BS03], or by applying the
image-based technique (IBT), see [SNK00]. Topology optimization methods have been
applied to design amplifying mechanisms (also termed mechanical amplifiers) of the
PIEZOELETRIC ACTUATORS 27

displacement of piezoelectric actuators, see e.g. [LDG00], [DLL00], [SNK00], [CF00] and
[BhF02]. Various formulations can be posed to obtain the optimum topology of the
amplifying mechanisms. In [Sig97] an inverting displacement amplifier using the formulation
of maximum mechanical advantage, namely the ratio of output force and input force, for static
response was investigated. In [LDG00] and [DLL00] three formulations for the maximum
output stroke, magnification factor and mechanical efficiency respectively were used for both
static and dynamic operations. The design problem was solved as a 2D material distribution
problem using Method of Moving Asymptotes (MMA). In [SNF98] and [SNK00] a
flextensional actuator was designed for simultaneously maximizing mean transduction and
mean stiffness. A multiobjective function was defined as the weighted sum of these two
objectives. In [SK99] a topology optimization method was proposed for design piezoelectric
transducers. The problem is posed as an eigenvalue optimization problem in structural
optimization. Three multiobjective functions were defined. The first one consists of the
maximization of the eletromechanical coupling factor for a specific mode or set of modes.
The second one is related to the design of a transducer with specified resonance frequencies.
The third one is related to the design of narrowband or broadband transducers. The sequential
linear programming (SLP) was used to obtain the optimized solution. However, Pareto optima
were not discussed. Topology optimization requires efficient and reliable large-scale
nonlinear optimization algorithms, because a fine mesh is required and a large number of
design variables are included in the problem. Usually, dual methods such as Method of
Moving Asymptotes (MMA) are more efficient than primal methods such as sequential linear
programming (SLP) and sequential quadratic programming (SQP) [DLL00]. Although the
topology of the structure obtained by using topology optimization methods is generally
difficult to manufacture, the topology optimization technique is one rational systematic design
methodology for the structural design if the physical size and the shape and connectivity of
the structure are unknown.

Some researchers have studied optimization problems concerning ultrasonic motors,

transducers and piezoelectric transformers. In [PRF03] the optimization of the geometrical
and material parameters defining the motor stator was studied, based on a parametric model
of a traveling wave ultrasonic motor including stator, rotor and a simplified rotor-stator
interface model. Three performance indicators, namely the piezoelectric coupling coefficient,
the efficiency and the output power were used as objective functions. However, in this work
the optimization process was based on single variable and no mathematical optimization
algorithms were applied. In [Hem01] high efficiency, high shear force and velocity, small size
and low cost were postulated as optimization goals for a novel piezoelectric linear motor. For
each goal the corresponding design requirements for the oscillator, the contact, the rotor and
the drive were introduced. No mathematical optimization methods were used in this work. In
[JNL00] the optimization of the velocity of a piezoelectric ceramic actuator was performed,
with driving voltage and system parameters being design variables. The Sequential Quadratic
Programming (SQP) method was used. Each variable was explored individually. No
optimization was performed for multivariable. The optimization of a piezoelectric transformer
28 CHAPTER 2

for maximum power transfer was studied in [HLH03], but no optimization algorithm was
mentioned in this work. In [Lit03] the optimization of a power ultrasonic transducer based on
an equivalent model was investigated. The optimization objectives, namely the vibration
amplitude and efficiency were analyzed. However, no mathematical optimization methods
were applied in this work.

Although a number of studies concerning optimization of piezoelectric actuators have been

reported in the above reviewed literature, there is a necessity for further investigations,
because an important class of optimization methods, namely multiobjective optimization
methods (in particular MOEAs) have hardly been applied in the systematic design of
piezoelectric actuators. Multiobjective optimization of piezoelectric actuators involving
continuous and discrete design variables is scarcely reported. There is a need to present an
integrated procedure for the piezoelectric transducer design via multiobjective optimization
methods. This is concerned with the modeling, the problem formulation, the application of
optimization algorithms and the selection of optimum results. In the next sections of this
dissertation, the above mentioned several aspects will be discussed.
MULTIOBJECTIVE OPTIMIZATION METHODS 29

3 Multiobjective Optimization Methods

In Chapter 2 the modeling of the piezoelectric actuators has been introduced. The models of
the actuators provide the base for formulating the multiobjective optimization problems
(MOPs) for piezoelectric actuators. On the other hand, in order to solve MOPs, an appropriate
optimization method is needed. Many multiobjective optimization methods have been
developed over the past decades. A comprehensive summary of some of these methods can be
found in [Mie99], [Deb02] etc. In this chapter, the multiobjective optimization methods which
are most commonly used in engineering will be described. The study concentrates on those
methods which will be used in the next chapter for solving the MOPs of piezoelectric
actuators. The structure of this chapter is as follows: First, the basic concepts of
multiobjective optimizations are described. Then basic multiobjective optimization methods
are introduced. Finally, as the main part of this chapter, the most commonly used
multiobjective evolution algorithms (MOEAs) will be studied.

3.1 Basic Concepts of Multiobjective Optimization

The MOP is an extension of the single-objective (scalar) optimization problem. In the MOP, a
number of objective functions are to be optimized simultaneously. Losing no generality, a
MOP can be stated as follows:
min f v (x) v = 1, 2, ..., m (3.1)

subject to g j (x) ≥ 0, j = 1,2,..., J

hk (x) = 0, k = 1,2,..., K
xil ≤ xi ≤ xiu , i = 1,2,..., n

where x ∈ R n is a vector of n decision variables (design variables), which describe a solution.

The side constraints xil ≤ xi ≤ xiu , the inequality constraints g j (x) ≥ 0 and the equality
constraints hk (x) = 0 define a feasible region G in the decision variable space R n . For each
solution x in the feasible region G exists a point (objective vector) in the objective space R m ,
denoted by F (x) ∈ R m , where F(x) = z = ( z1 , z 2 , ..., z m ) , z v = f v (x) for all v = 1, 2 , ... , m . The
T

set of these objective vectors is called the feasible objective region Z. Fig. 3.1 represents the
mapping from decision variable space into objective function space for a two-dimensional
case. As a vector of objectives, instead of a single objective, is optimized, multiobjective
optimization is also called vector optimization.
30 CHAPTER 3

x2 f2
Decision space Objective space

x F :Rn → Rm z

x1 f1

Fig. 3.1 Mapping from decision space into objective space

In the single-objective optimization, one can find a solution, which is absolutely best with
respect to all other alternatives. In the MOP, because of incommensurability and conflict
among objectives, it is not always possible to find a solution at which each objective function
gets its optimal value simultaneously. The solutions of a MOP are therefore given in the non-
dominance or Pareto optimality sense. For such solutions, no improvement in any objective
function is possible without deterioration to at least one of the other objective functions. A
more formal definition of non-dominance and Pareto optimality is as follows [Mie99] and
[Deb02]:

Definition 3.1: A solution x (1) is said to dominate the other solution x ( 2) if

f i (x (1) ) ≤ f i (x ( 2) ) for all i = 1,..., m and f j (x (1) ) < f j (x ( 2) ) for at least one index j.
If solution x (1) dominates solution x ( 2) , the following terms are also commonly used:

• Solution x (1) is non-dominated by solution x ( 2)

• Solution x ( 2) is dominated by solution x (1)

• Solution x (1) is non-inferior to solution x ( 2)

Definition 3.2: A decision vector x̂ is non-dominated with respect to a set P if it is not

dominated by any other member of the set P. An objective vector ẑ is non-dominated if its
image point in the set P is non-dominated.

The set of non-dominated solutions is called the non-dominated set and the set of non-
dominated solutions in the objective space is called non-dominated front.

If the non-dominance concept is extended to the entire feasible space G, the Pareto-optimality
can be defined as follows:

Definition 3.3: A decision vector x̂ ∈ G is Pareto-optimal if there does not exist another
decision vector x ∈ G such that f i (x) ≤ f i (xˆ ) for all i = 1,..., m and f j (x) < f j (xˆ ) for at least
one index j. An objective vector ẑ ∈ Z is Pareto-optimal if its image point in the set G is
Pareto-optimal.
MULTIOBJECTIVE OPTIMIZATION METHODS 31

The set of Pareto-optimal solutions is called Pareto-optimal set and the set of Pareto-optimal
solutions in the objective space is called Pareto-optimal front. Because of G ⊇ P , obviously a
Pareto-optimal set is always a non-dominated set.

Fig. 3.2 shows the non-dominated and Pareto-optimal solutions in the objective space for a
two-objective minimization problem. The fat line represents the Pareto front. Points A and B
represent two Pareto-optimal points of the front. Although points C and D are non-dominated
with respect to each other locally, they are not Pareto-optimal with respect to entire feasible
region Z, because clearly point B dominates points C and D and point A dominates point C.
Therefore, the non-dominated set is Pareto-optimal set if only if it is a non-dominated set with
respect to the entire feasible space Z.

In many multi-objective optimization algorithms, Pareto-optimal solutions are usually found

based on non-domination concept. Many approaches have been suggested for finding the non-
dominated set of solutions from a given set of solutions. In [Deb02] three approaches (Naïve
and Slow, Continuously Updated and Efficient Methods) have been described in detail.

3.2 Traditional Multiobjective Optimization Methods

Basically, there are three kinds of methods for handling MOPs: (1) generating approaches; (2)
preference-based approaches (3) interactive methods [Coh85] and [Mie99]. In the generating
methods, the entire set of Pareto optimal solutions is identified first, and then the best
compromise solution is chosen from the obtained set by using higher-level information. No
prior knowledge of preference structure over objectives is used. On the other hand,
preference-based approaches require decision makers to give their preferences in a formal and
structured way and then a composite single objective function is constructed. In interactive
methods, the decision maker works with an interactive computer program and progressively
provides preference information during the optimization process. The generating method is
more practical and less subjective, while the preference-based method is more subjective and
not straightforward. Interactive methods can be presumed to produce the most satisfactory
results.

Most traditional methods convert the MOP into a single (scalar) or a family of single
objective optimization problems by using some user-defined preference structures, and then
solve the problem using the widely developed methods for single objective optimization. In
the following, some most commonly used methods will be introduced. Details can be found in
[Mie99], [GCh00] and [Deb02].
32 CHAPTER 3

f2

A D Pareto front
B
f1

Fig. 3.2 Pareto-optimal solutions

3.2.1 Weighted Sum Method

The weighted sum method scalarizes a set of objectives into a single objective by assigning
weights to each objective function. This method is the simplest and the most widely used
method. It can be formulated as follows:
m
min F (x) = ∑ wv f v (x) (3.2)
v =1

subject to g j (x) ≥ 0, j = 1,2,..., J

hk (x) = 0, k = 1,2,..., K
xil ≤ xi ≤ xiu , i = 1,2,..., n
m
The weight wv represents the relative importance of the objective and usually ∑w
v =1
v = 1 . If all

the weights are positive the optimal solution to the problem is Pareto optimal for the initial
MOP. By choosing more than one set of positive weights other optimal solutions can be
generated. If F(x) is convex then all the optimal solutions of F(x) are Pareto-optimal
solutions. However, if F(x) is non-convex certain Pareto-optimal solutions cannot be found.
This can be illustrated geometrically. Consider the two-objective case in Fig. 3.3 (a). In the
2
objective function space a straight line L1: ∑w
v =1
v f v (x) = F1 is drawn. The minimization of

equation (3.2) can be interpreted as finding the value of F1 for which the line L1 is tangential
to the boundary of objective space Z and lies in the bottom-left corner of the objective space
Z. If different weight vector is selected, then the different slope of the straight line L1 is
defined. This leads to the different optimal solution point where the line L1 is tangential to the
boundary of objective space Z. However, when the lower boundary of the search space Z is
non-convex, the Pareto-optimal solutions between A and B is not available, see Fig. 3.3 (b).
MULTIOBJECTIVE OPTIMIZATION METHODS 33

2
L1 : ∑ w v f v ( x ) = F1
f2 v =1 f2

A
L1
L1 B

f1 f1
(a) (b)
Fig. 3.3 The weighted sum method for (a) convex (b) non-convex part of objective functions

3.2.2 ε-Constraint Method

ε-Constraint method can overcome some of the convexity problems of the weighted sum
method. In this method the MOP is reformulated by just keeping one of the objectives and
transfer the others into constraints. This approach is able to identify a number of Pareto-
optimal solutions on a non-convex boundary. It is formulated as
min f q (x) (3.3)

subject to f v (x) ≤ ε v v = 1,2,..., m m ≠ q

g j (x) ≥ 0, j = 1,2,..., J
hk (x) = 0, k = 1,2,..., K
xil ≤ xi ≤ xiu , i = 1,2,..., n

3.2.3 Weighted metric methods

The weighted metric method (or compromise method) identifies solutions closest to the ideal
( )
T
point z * . The ideal point (objective vector) z * is defined as z * = F * = f1* , f 2* , ... , f m* , where
f v* , v = 1,2,..., m is the minimum value for the v-th objective function individually. For the
conflicting objectives the ideal point is not attainable because the conflicting objective
functions are impossible to be minimized simultaneously. However, it is obvious the closer to
the ideal point the solutions are, the better the solutions. Therefore the ideal point is usually
used as a reference point. For a point z ∈ Z the distance from z to the z * may be
approximated by a Lp-norm as follows
p 1/ p
m 
L p (z ) = z − z *
= ∑ z v − z v*  (3.4)
p
 v=1 
34 CHAPTER 3

where p ∈ [1, ∞] . If different degrees of importance for objectives are considered, a weight
vector w = ( w1 , w2 ,..., wm ) is used. Therefore, for non-negative weights, the weighted Lp-
norm is as follows
1/ p
m p

L p ,w (x) = ∑ wv f v (x) − z v 
*
(3.5)
 v =1 

where x ∈ G . The MOP is then transferred to a problem of minimizing the weighted Lp –

norm. The parameter p will affect the optimal results. When p=1 the resulting method is equal
to the weighted sum method. When p = ∞ the problem reduces to weighted Tchebycheff
problem:
{ }
min L∞ ,w (x) = max wv f v (x) − z v* , v = 1,2,..., m (3.6)

subject to g j (x) ≥ 0, j = 1,2,..., J

hk (x) = 0, k = 1,2,..., K
xil ≤ xi ≤ xiu , i = 1,2,..., n

The parameter p will affect the optimal results. With p = 1or 2 , not all Pareto optimal
solutions can be obtained. Although the weighted Tchebycheff method can find any Pareto-
optimal solution, as p increases the problem becomes complex.

3.2.4 Value Function Method

In the value function (or utility function) method, a preference structure is mathematically
represented using a value function U : R q → R , which maps solutions in the objective space
into a real number so that the greater the number, the more preferred the solution in the
objective space. The problem is then as follows:
max U (F(x)) (3.7)
subject to g j (x) ≥ 0, j = 1,2,..., J
hk (x) = 0, k = 1,2,..., K
xil ≤ xi ≤ xiu , i = 1,2,..., n

where F(x) = ( f1 (x),..., f q (x) ) . The solution to the problem entirely depends on the chosen
T

value function. Because of the difficulty in obtaining value function, this method is limited for
solving MOPs.

In addition to the above described methods, there are many interactive methods and other
conventional multi-objective optimization methods like e.g. NIMBUS (Nondifferentiable
Interactive Multiobjective BUndle-based optimization System). A detailed discussion on
these approaches can refer to [Mie 99].
MULTIOBJECTIVE OPTIMIZATION METHODS 35

3.3 Multiobjective Evolutionary Algorithms

In a MOP, different decision makers with different preferences may select different Pareto-
optimal solutions. Therefore, it is important to find as many Pareto-optimal solutions as
possible so that the decision maker can choose the best one according to his preference.
Therefore, there are two goals in a multi-objective optimization [Deb 02]:
1. to find a set of solutions as close as possible to the true Pareto-optimal front
2. to find a set of solutions as diverse as possible

Conventional multi-objective optimization methods solve the MOP by scalarization. For these
methods, there exist many difficulties in finding multiple Pareto optimal solutions. For
example, they require some problem knowledge before optimization is performed. Some
techniques such as weighted sum methods may be sensitive to the shape of the Pareto-optimal
front. Moreover, they require several optimization runs to obtain an approximation of the
Pareto-optimal set [Zit99]. Over the past decade, multi-objective evolutionary algorithms
(MOEAs) have been developed. The basic feature of MOEAs is multiple directional and
global search through maintaining a population of potential solutions from generation to
generation [GCh00]. Due to their population-approach, MOEAs can find multiple Pareto-
optimal solutions in one single simulation run. Moreover, by using a diversity-preserving
mechanism, MOEAs also can find widely different Pareto-optimal solutions. In the next
section, the basic principles of evolution algorithms are first described, and then the mostly
used MOEAs are introduced.

3.3.1 Basic Principles of Evolutionary Algorithms

EAs imitate natural evolutionary principles to constitute search and optimization procedures.
The best known EAs include genetic algorithms (GAs), evolution strategies and evolutionary
programming. GAs are the mostly widely used EAs in the field of optimization.

The terminology used in GAs is borrowed from natural genetics. Individuals (solutions) in a
population are called strings or chromosomes. Chromosomes are made of units called genes,
each of which encodes a particular feature of the organism. The position of a gene in a
chromosome also reflects a particular characteristic of organism. Each chromosome consists
of several segments. Each segment refers to a particular value of a variable. In the simplest
case a chromosome contains only a segment, which refers to a particular value of a variable
and by a binary string, for example, 0110 in which each digit is treated as a gene. If 0110 is
transformed into the decimal system as follows:

0 ⋅ 2 3 + 1 ⋅ 2 2 + 1 ⋅ 21 + 0 ⋅ 2 0 = 6

The value 6 may represent the value of a variable, which is used to calculate the value of an
objective function in optimization. In order to use GA to find the optimal decision variables,
various methods are used to encode them into a chromosome. Binary encoding and real-
36 CHAPTER 3

number encoding are two mostly used encoding methods [GCh00]. In the binary encoding,
each decision variable is encoded as a bit using standard binary coding. For a problem having
N variables, x=(x1,…,xn), a chromosome contains n segments of a bit string as follows:

11
12...3
01 11
12...3
01 L 10
12...3
11
x1 x2 xn

Therefore, if each variable xi is coded in mi bits, the length of a complete chromosome is

∑
n
i =1
mi . As the binary encoding of continuous variable usually results in many difficulties
[Deb02], a real-number presentation is usually used in the problems having continuous
variables. In the real-number encoding, each chromosome is represented as a vector of real
numbers (decimal system). For example, for a problem with n variables, the real-number
vector is

 
x =  23 { 
{ , 6{ ,L,168
 x1 x2 xn 

Fig. 3.4 shows a standard GA flowchart.

Initialization An initial population of N individuals (solutions) is generated randomly. Each

individual refers to a design, which contains n design variables. Each individual is represented
by a chromosome (string), which contains n segments corresponding to n design variables

Initialization Evaluation

Crossover

Mutation

Fig. 3.4 The standard GA flowchart

MULTIOBJECTIVE OPTIMIZATION METHODS 37

Evaluation During evaluation, each individual is evaluated to give some measure of its
fitness. In the unconstrained single-objective optimization problem, the fitness of a solution is
assigned a value which is a function of the value of the corresponding objective function. In
MOPs, special approaches are needed to determine the fitness value of a solution according to
multiple objectives. These approaches will be introduced in the next sections.

Selection The primary task of the selection operator is to select and duplicate good solutions
to create a new mating pool of N parent solutions. There exist a number of selection methods.
Tournament selection, proportionate selection and ranking selection are some common types
[Deb02].

In the tournament selection, tournaments are played between two randomly chosen
chromosomes and the better chromosome is chosen and placed in the mating pool. The
process is repeated N times until the mating pool is full.

In the proportionate selection, the selection probability or survival probability for each
chromosome is determined in proportion to its fitness value. Then a roulette wheel can be
constructed displaying these probabilities. Thereafter, the wheel is spun N (population size)
times, each time selecting a chromosome. The wheel features the selection method as a
random sampling procedure. Therefore, the chromosome that has higher fitness value will be
copied into mating pool more often than the chromosome with lower fitness. The roulette-
wheel concept can be simulated on a computer. A random and deterministic version of the
above method is stochastic remainder roulette-wheel selection (SRWS) operator. In this
operator, the selection probabilities are multiplied by the population size and the expected
number of copies is calculated. Each chromosome is allocated copies according to the integer
part of the expected number, and then the remaining slots in the mating pool is filled by using
the RWS operator on the entire population with the fractional part of the expected number as
fitness. In [Bak85] the stochastic universal sampling (SUS), another version of RWS is
proposed. In this method, N (population size) numbers, which have equal space distances, are
generated as follows:
 1 2 N − 1
LR = rn, rn + , rn + ,..., rn +  mod 1 (3.8)
 N N N 

where rn is a random number between 0 and 1. Each element of LR represents a location of

the roulette wheel, see Fig. 3.5. Therefore, only a random number is needed in order to select
N solutions.

In ranking selection, the population is sorted from the best (rank N ) to the worst (rank 1)
according to their fitness, and then the selection probability of each chromosome is assigned
according to the ranking but not its raw fitness. Thereafter a roulette wheel is applied with the
selection probability and N chromosomes are chosen for the mating pool.
38 CHAPTER 3

6
7
5

8
4 rn
1
1/8
3
2
Fig. 3.5 Roulette wheel with a population of 8 solutions

Crossover The crossover operator provides the search mechanism of the GA. It creates new
solutions (offspring) by exchanging features of two parent solutions picked from the mating
pool at random. For a binary-coded chromosome, the single-point crossover operator and the
two-point crossover operator are usually used. In a single-point crossover operator, the
crossover operation is performed by choosing a crossing site along the string at random and
exchanging all bits on the right side of the crossing site, as shown in the following example:
Parents Offspring

0 1 0 1 0 0 1 0 1 0 0 0 1 0
1 0 1 0 0 1 0 1 0 1 1 0 0 1

In a two-point crossover operator, two different crossing sites are chosen randomly and the
crossover operation is completed by exchanging middle substrings. Following this idea, n-
point crossover can also be implemented. With the increase of crossover sites the extent of
string preservation reduces. In order to preserve some good chromosomes, not all strings in
the mating pool are used in a crossover. Usually, a crossover probability of pc is used.

For a real-parameter GA, crossover operation is implemented by using a pair of real-

parameter decision variable vectors to create a new pair of offspring vectors. Linear
crossover, blend crossover (BLX-α), and boundary operator are some commonly used
crossover.

In linear crossover operator, three solutions, 0.5 xi(1,t ) + 0.5 xi( 2,t ) , 1.5 xi(1,t ) − 0.5 xi( 2,t ) and
− 0.5 xi(1,t ) + 1.5 xi( 2,t ) are generated from two parent solutions xi(1,t ) and xi( 2,t ) at generation t .
The best two solutions will be chosen as offspring [Deb02].

In BLX-α [ΕS93], parent solutions xi(1,t ) and xi( 2,t ) give birth to offspring xi(1, t +1) and xi( 2, t +1)
according to

xi(1, t +1) = γ i xi(1, t ) + (1 − γ i ) xi( 2, t )

(3.9)
xi( 2, t +1) = (1 − γ i ) xi(1, t ) + γ i xi( 2, t )
MULTIOBJECTIVE OPTIMIZATION METHODS 39

where γ i = (1 + 2α )ui − α . u i is a random number in [0,1] and α is user-specified parameter. It

is reported that BLX-α with α = 0.5 performs better than BLX operators with any other α
value [Deb02].

Boundary operator was proposed in [SM97]. Considering that the global solution for many
optimization problems usually lies on the boundary of the feasible region, for many
constrained optimization problems, it may be beneficial to search just the boundary of the
search space defined by the constraints. An example of such an approach is sphere crossover.
Two parent solutions xi(1,t ) and xi( 2,t ) create one offspring xi(1, t +1) according to

xi(1, t +1) = α xi(1, t ) + (1 − α ) xi( 2, t ) (3.10)

where i =1,…, n , α is a random number in [1, n].

There exist many other crossovers, such as simulated binary crossover (SBX), unimodal
normally distributed crossover (UNDX), fuzzy connectives based crossover and direction-
based crossover, etc. For details refer to [GCh00].

Mutation The mutation operator is used to perform a local search to find an improved
solution. For the binary-coded chromosome, the bitwise mutation operator changes a 1 to 0
and vice versa with a mutation probability of pm . For the real-parameter coded chromosome,
random mutation, non-uniform mutation and Gaussian mutation are the most commonly used
real-parameter mutation operators [Deb02].

The above three operators, namely selection, crossover and mutation operators constitute the
main part of a GA. Through them a new population is created in every generation (iteration).
Finally, termination conditions are checked. If the specified maximum number of generations
is arrived, or another stopping criterion is satisfied, the iteration stops, otherwise the surviving
population become the starting population for the next generation.

Preferred Solution Generally, a best solution needs to be selected for the implementation.

In addition to GAs, evolution strategies and evolutionary programming are also usually used.
Good introduction material for the latter two methods can be found in [Bäc96].

3.3.2 Fitness Assignment and Fitness Sharing

As EAs can find multiple optimal solutions in a single run due to their population-to-
population approach, it is natural to apply them in MOPs. According to two goals in a multi-
objective optimization stated before, two aspects must be addressed in order to adapt EAs to
MOPs. First, special approaches are needed for accomplishing fitness assignment and
selection respect to multiple objectives, respectively, in order to guide the search towards the
Pareto-optimal front. Second, appropriate techniques are needed for maintaining population
diversity in order to achieve a well distributed and well spread non-dominated set.
40 CHAPTER 3

During the past decade several fitness assignment methods have been developed. These
methods can be classified as follows [GCh00]:

• Vector evaluation approach

• Weighted-sum method
• Goal programming approach
• Pareto-based approach

According to these different fitness assignment methods, various MOEAs have been
proposed. These will be described in subsequent section.

In order to maintain population diversity, fitness sharing technique [GR87] is usually applied
in MOEAs. The basic idea is to reduce the reproduction ability of a solution crowded by
many solutions through degrading its fitness value f i using the sharing function concept.
Typically the following function is defined as the sharing function:

  d ij α
   if d ij < σ share
Sh (d ij ) = 1 −  σ (3.11)
 share 
 0 otherwise


where α is a constant, σ share is the niche radius, which represents the minimal distance
between two solutions desired by the designer. The parameter d ij is the metric of distance
between any two solutions i and j. Therefore, if d ij = 0 , then Sh (d ij ) = 1 . This means a
solution has full sharing on itself. If d ij ≥ σ share , then Sh (d ij ) = 0 . This means two solutions
which are at least a distance away from each other do not have any sharing effect on each
other. For any other distance d two solutions have partial effect. A niche count nci, which
gives an estimation of the extent of the crowding near the i-th solution, is then calculated as
follows:

nci = ∑ j =1Sh (d ij )
N
(3.12)

The sum includes the ith solution itself. Thus, nci is always greater than or equal to one
because at least Sh (d ii ) = 1 . Finally, the shared fitness value f i ' for the i-th solution is
calculated as follows

fi
fi' = (3.13)
nci
MULTIOBJECTIVE OPTIMIZATION METHODS 41

Thus, if the ith solution does not have any sharing effect on any other solution in the
population, namely nci = 1 , its fitness value will not be degraded. Otherwise, the sharing
function will degrade fitness according to the crowding extent near the i-th solution.

3.3.3 General Procedures of Multi-objective Evolutionary Algorithms

As stated before, the dual goals in a multi-objective optimization are to find a set of solutions
as close as possible to the Pareto-optimal front and simultaneously as diverse as possible.
According to different ways to maintain the two goals, various MOEAs have been proposed.
In [Deb02] and [GCh00] one can find the detailed descriptions for various MOEAs. Here only
the Pareto-based MOEAs, which will be applied in the next chapter, are introduced.

Except the fitness assignment (evaluation) method for multiple objectives, the basic structure
of a Pareto-based MOEA is similar to that of an GA described before. The general procedures
are as follows:
Step1. Initialization: randomly create an initial population
Step2. Evaluation: identify non-dominated individuals from the population, assign the
fitness of each individual using an appropriate fitness assignment method
Step3. Selection: select parents for genetic operations using an appropriate selection
approach
Step4. Generation: generate offspring using genetic (crossover and mutation) operators,
Step5. Elite Preservation: ensure that the elite individuals enter the next generation using
an elite-preserving strategy
Step6. Termination: if the prespecified maximal number of generations is reached, stop;
otherwise return to step2
Step7. Determination of the preferred solution: a best solution is selected for
implementation
It is noted that in most single-objective EA, the best α solutions (α=1 to 0.1N, N is the
population size) are used as the elitist solutions [Deb 02]. In MOEA implementations, the best
α non-dominated fronts (for example the fronts with rank 1 and rank 2) are used as the elitists.
The proportion of the elitists from the population must be chosen carefully. If the proportion
is too large, the population will lose its diversity, whereas the convergence performance can
not be improved if the proportion is too small. In the following, some salient Pareto-based
MOEAs including Pareto ranking or Pareto tournament are introduced.

In Pareto ranking methods, the fitness of individuals is assigned not according to the values of
the objective functions, but in terms of the Pareto ranking of individuals. The basic procedure
of the Pareto ranking (Goldberg’s ranking) is as follows [Gol89]: first, non-dominated
individuals with respect to the population are identified and assigned rank 1. Then these
solutions are removed from the population. Next, the non-dominated solutions with respect to
42 CHAPTER 3

the remaining population are found and assigned rank 2. This procedure is continued until the
entire population is ranked. Fig. 3.6 illustrates this approach for a two-objective minimisation
problem. The multi-objective GA [FF93] and the non-dominated sorting GA [SD95] are two
representative methods based on Pareto ranking.

3.3.4 Multiobjective Genetic Algorithm (MOGA)

In this method, the population is ranked according to the follow expression:

ri = 1 + ni (3.14)

where ri presents the rank of the i-th solution. ni is the number of the solutions which
dominate the solution i. Obviously, the minimum rank is 1 (for the non-dominated solutions
with respect to population) and the maximum rank is no more than the size of the population.
The smaller the rank ri , the better the solution is. The population is sorted according to the
ranks and a raw fitness is assigned to a solution by using a mapping function from the best
solution to the worst solution. And then the raw fitnesses of the solutions within the same
rank are averaged. In order to maintain diversity among non-dominated solutions, the shared
fitness as described in the previous section is applied. Here a normalized distance of solution i
from each solution j having the same rank is calculated in objective space as follows

2
 f n( i ) − f n( j ) 
d ij = ∑n=1  max
q
min
 (3.15)
 fn − fn 

where f nmax and f nmin are maximum and minimum objective function values of the n-th
objective. Except the fitness assignment method, the rest of this method is the same as that in
a standard GA.
f1
1 2

1
2
2
1
1

f2

Fig. 3.6 Goldberg’s ranking

MULTIOBJECTIVE OPTIMIZATION METHODS 43

3.3.5 Non-dominated Sorting Genetic Algorithm (NSGA)

In this method, the population is first ranked using the basic procedures of the Pareto ranking.
A large fitness value is assigned to the individuals in the first non-dominated front, namely
the set of the non-dominated individuals with rank 1. In order to maintain the goal of
diversity, the sharing strategy is applied and the shared fitness of each individual in front 1 is
obtained. Then a fitness value that is smaller than the minimum shared fitness value of the
previous front is assigned to the individuals in the next front. The sharing strategy is used and
the shared fitness of each individual in front 2 are obtained. This procedure is continued until
the shared fitnesses of individuals in all fronts are obtained. A stochastic remainder roulette-
wheel selection (SRWS) operator is used to generate new population. By means of this
method, the non-dominated solutions in front 1 are emphasized and simultaneously the
diversity is maintained.

In this method, the normalized Euclidean distance dij of the solution i from another solution j
in the front 1 (Fr1) is calculated in the parameter space as follows:

2
 xn( i ) − xn( j ) 
d ij = ∑n=1  max
Fr1
min
 (3.16)
 x n − xn 

Where xnmax and xnmin are the maximum and minimum objective decision variable values of
the n-th decision variable, respectively. Fr1 represents the number of non-dominated
solutions in front 1. The sharing function value and niche count are then calculated using
equations (3.11) and (3.12) respectively. Finally the shared fitness of the solution i is obtained
by using equation (3.13). Once again, the sharing is performed in the solutions in front 2 (Fr2)
and the corresponding shared fitness value computed. This process continues until the shared
fitnesses of all solutions are obtained.

In the above two methods, no mechanism is used to assure that the non-dominated solutions
generated during evolutionary process enter the next generation, therefore, some non-
dominated solutions obtained may get lost during evolution. In order to avoid this problem, in
the following MOEAs an elite-preserving strategy is used.

3.3.6 Strength Pareto Evolutionary Algorithm (SPEA and SPEA2)

SPEA [ZT99] introduces the elite-preserving strategy by using an archive P' containing the
non-dominated solutions found so far (the so-called external non-dominated set). At each
generation, all non-dominated individuals in the current population P are copied to the
archive P' . Then domination-check is performed among the members of the archive. Any
dominated individuals are removed from P' . If the size of P' still exceeds a given limit,
further non-dominated solutions stored in P' are deleted by means of clustering without
destroying the characteristics of the non-dominated front. The fitness of each member in the
44 CHAPTER 3

current population P and archive P' is then assigned. For each individual in the archive P' a
strength value si ∈ [0,1) is computed, which represents the fitness value of the corresponding
individual. The strength si is calculated as follows:
ni
si = (3.17)
N +1
Where ni is the number of current population members that are dominated by an archive
member i. N is the size of the current population P. The fitness of an individual j in the
current population P is calculated by summing the strengths of all archive members i that
dominate j. It is given by

f j = 1+ ∑s i
i∈P ' ∧ i p j
(3.18)

This definition guarantees that members of P' have better fitness than members of P. Here a
small fitness corresponds to a high reproduction probability. According to this definition,
individuals near the Pareto-optimal front are preferred and simultaneously distributed along
the Pareto-optimal front. This kind of fitness assignment provides a niching method based on
Pareto domination.

After fitness values are assigned, individuals for the mating pool are selected from P+ P' by
using a binary tournament selection operator. Finally the crossover and mutation operations as
usual can be performed. In original study, a clustering approach, the average linkage method,
has been used to reduce the size of the external non-dominated set.

In [ZLT02], an improved version of SPEA has been proposed. SPEA2 uses an improved
fitness assignment scheme, a nearest neighbor density estimation technique and a new archive
truncation method. In SPEA2, the size of archive P' (external non-dominated set) is fixed. All
non-dominated individuals in P + P' are copied to the archive. If the number of non-
dominated individuals exceeds the archive size, the truncation operator is used. If the number
of non-dominated individuals is less than the archive size, the archive is filled up by
dominated individuals according to the fitness values. In order to fill the mating pool, the
binary tournament selection is performed on the archive. Similarly, the crossover and
mutation operators are applied to the mating pool to generate the offspring population. Unlike
SPEA, in SPEA2 each individual i in the population P and archive P' is assigned a strength
value si , which represents the number of individuals in P + P' which are dominated by the
individual i. The raw fitness f i of an individual i is calculated by summing the strengths of all
individuals that dominate i. A nearest neighbor density estimation technique is introduced to
distinguish between individuals that have identical raw fitness values. The density di
corresponding to i-th individual is calculated as follows
MULTIOBJECTIVE OPTIMIZATION METHODS 45

1
di = k
(3.19)
σ +2
i

where σ ik represents the distance to the k-th nearest individual. Here h is equal to the square
root of the size of P + P' . The final fitness of an individual i is defined by

Fi = f i + d i (3.20)

In archive truncation, the non-dominated individual having the minimum distance is first
removed from the archive. If there are several individuals having minimum distance the
individual with the second smallest distances is deleted and so on.

3.3.7 Elitist Non-dominated Sorting Genetic Algorithm (NSGA-II)

In NSGA-II [Deb02], an offspring population Qt is first generated from the parent population
Pt . Then these two populations are combined to form a population Rt of size 2N and a non-
dominated sorting is performed. Then the new population of size N is filled by the first non-
dominated front, the second non-dominated front, and so on, until all slots in the new
population Pt+1 is filled up. The remaining fronts are deleted. When the solutions in the last
front that is used to fill the new population are more than the remaining slots in the new
population Pt+1, a niching strategy is used and the solutions that locate in the least crowded
region are selected. This algorithm provides not only an elite-preservation but also maintains
a better spread among the solutions. Fig. 3.7 illustrates the procedure of NSGA-II.

In this algorithm, a crowded tournament selection operator is used. In the tournament

selection, two solutions i and j are compared. If they have different ranks, the solution that has
a better rank wins. If they have same rank, the solution that has a better crowding distance
wins. The crowding distance represents a measure of the density of solutions in the
neighborhood. It estimates half of the perimeter of the cuboids formed by the nearest
neighbors as the vertices.

The distance-based Pareto genetic algorithm is another commonly used elitist MOEA [OK
95] and [OK 96]. The basic idea is to assign fitness to each solution according to a distance
measure with respect to the non-dominated solutions in the proceeding generation.

3.4 Constraint Handling in Multiobjective Evolutionary Algorithms

Most constraint handling methods in MOEAs can be classified into three categories: methods
based on preserving feasibility of solutions, methods based on penalty functions and methods
based on feasibility and domination of solutions.
46 CHAPTER 3

Non-dominated Crowding distance

sorting sorting

Pt

Pt+1
Rejected

Fig.3.7 Schema of NSGA-II [Deb02]

3.4.1 Methods based on Preserving Feasibility of Solutions

In this method, solution that violates any of the constraints is discarded. An equality
constraint can be handled using explicit or implicit method. In the explicit method, the
explicit expression of a decision variable is derived from the equality constraint and then
substituted into all objective functions and other constraint functions. In this way, the number
of decision variables is reduced and this equality constraint is automatically satisfied in the
optimization process. When a variable is difficult to be expressed explicitly, the implicit
method can be used. For example, for each solution vector x=(x1, x2,…,xn ), the value of a
variable x1 is determined by finding the root of an equality constraint h(x)=0 in terms of
variable x1. Therefore, the equality constraint is also automatically satisfied in the
optimization process.

3.4.2 Methods based on Penalty Functions

The penalty technique is probably the most common technique to for constrained MOPs. In
this method, a constrained MOP is transformed into an unconstrained MOP by penalizing
infeasible solutions, where a penalty term is added to each objective function for any
constraint violation. Based on the objective function f m ( x ) and the penalty term pm ( x ) , the
new objective function is given by
~
f m (x ) = f m (x ) + pm (x ) (3.21)

Once the new objective functions are formed, any unstrained multi-objective optimization
methods described before can be used.

Generally, there are two classes of penalty function: static penalty and dynamic penalty. The
static penalty function is usually given by
MULTIOBJECTIVE OPTIMIZATION METHODS 47

pm ( x ) = rm ∑ j =1 g~ j ( x )
J
(3.22)

where rm is a penalty parameter for the m-th objective function. g~ j ( x ) is constraint violation
of the j-th constraint for the solution x . For the minimization problem and the inequality
constraints of the form “ ≥ ”, the constraint violation is calculated as follows:

| g (x) |, if g j (x) < 0

g~ j (x) =  j j = 1,2,..., J
0, otherwise (3.23)

where g j (x) is the normalized constraint function of the constraint g j (x) . That is, J
constraint violations | g j (x) | ( j = 1,2,...J ) have the same order of magnitude.

In the dynamic penalty function method, the penalty parameter is changed with the
generation. There are many dynamic penalty methods. A commonly used method is as
follows [JH94]:
J
[ ]
p m ( x ) = (C n ⋅ t ) γ ∑ j =1 g~ j ( x )
β
(3.24)

where t is generation counter. Cn , γ and β are user-defined constants. It is suggested that γ=1
and β=2.

3.4.3 Methods based on Feasibility and Domination of Solutions

The basic idea of these methods is to determine the winner of a tournament selection mainly
according to three criteria: prefer the feasible solution to the infeasible solution, prefer the
non-dominated solution to the dominated solution and prefer the solution residing in a less
crowded region of the objective space or decision variable space to the solution residing in
more crowded region.

In [JVG99] feasible and infeasible solutions are evaluated and a niche strategy is used to
maintain the diversity of Pareto-optimal solutions. A binary tournament selection operator is
used and a basic procedure is proposed: first, two solutions are picked up from the parent
population. Then a tournament is played. If one solution is feasible and the other is infeasible,
the former wins. If two solutions are feasible, a subpopulation of feasible solutions is selected
from parent population and then a procedure similar to the niched Pareto GA described before
is used. If two solutions are infeasible, a subpopulation of infeasible solutions is selected from
parent population and a procedure similar to the niched Pareto GA is followed, where the
constraint violation or nearness to a constraint boundary can be used as a criterion for
comparison.

The constrained tournament method [Deb02] is another commonly used method. In this
method, a constrain-domination concept is used to sort the population and then a binary
tournament selection is performed. A solution x (i) is said to constrain-dominate the other
solution x (j) , if any of the following conditions is true:
48 CHAPTER 3

1. solution x (i) is feasible and solution x (j) is infeasible

2. two solutions x (i) and x (j) are both infeasible, but solution x (i) has a smaller constraint
violation

3. both solution x (i) and x (j) are feasible and solution x (i) dominates solution x (j)

Therefore, this allows a non-dominated sorting in the feasible region, and the solutions in the
infeasible region are classified according to their constraint violation values. Similarly, the set
of non-constraint dominated solutions can be defined as those that are not constrain-
dominated by any member of the population. On the analogy of the procedure of the non-
dominated sorting described in the previous section, the non-constrain-dominated sorting for
the population is performed. A binary tournament selection operator is then used. First, two
solutions are picked from the non-constrain-dominated population. If one solution belongs to
a better non-constrain-dominated front, this solution wins. If two solutions belong to the same
non-constrain-dominated front, the solution, which resides in a less crowded region based on
a niched distance measure, will be chosen. A niched distance can be calculated using niche
count metric based on sharing function, head count metric or crowding distance metric.

Instead of defining constraint-domination for infeasible solutions based on an overall

constraint violation obtained by simply adding all constraint violations together, the non-
domination check of constraint violations for infeasible solutions is suggested in [RTS01]. In
this method, three different non-dominated sorting procedures, namely a non-dominated
sorting only with respect to the objective functions, a non-dominated sorting only according
to the constraint violation values and a non-dominated sorting with respect to a combined
objective function and constraint violation values, are performed. And then solutions are
chosen with these three different rankings and a new population is constructed.

3.5 Performance Metrics for Evaluating MOEAs

In the multi-objective optimization of piezoelectric transducers, it is expected that the

obtained non-dominated solutions are not only closer to the true Pareto-optimal front but also
convenient to be selected. These are also two general goals from the point of view of
engineering.

As stated before, there are two goals in multi-objective optimization: 1. find solutions as close
to the true Pareto-optimal front as possible. 2. find solutions as diverse as possible in the
obtained non-dominated front. These are also two general goals from the point of view of
engineering. The former emphasizes the precision of the optimized results. The latter
emphasizes the convenience of the selection of the results. In order to evaluate an
optimization method in terms of these two aspects, a number of performance metrics have
been developed. These performance metrics can be classified into three groups [Deb02]: The
first group is used to measure the extent of the solutions close to the Pareto-optimal solutions.
The second group is used to measure the diversity of the obtained solutions. The third group
MULTIOBJECTIVE OPTIMIZATION METHODS 49

is used to measure both aspects above using two metrics in an implicit manner. In this work,
set coverage metric is used to evaluate the closeness to the Pareto-optimal front. The diversity
among obtained non-dominated solutions is evaluated using two metrics that take into account
the uniformity of distribution and the extent of spread of solutions, respectively.

If A and B represent two sets of no-dominated solutions obtained using two different
optimization methods, the set coverage metric C (A, B) calculates the proportion of solutions
in the set B, which are dominated by solutions in the set A as follows[Zit99]:

The number of solutions in B, which are dominated by solutions of A

C ( A, B ) =
The number of solutions of B

(3.25)

Unlike those comparisons of different multi-objective optimization algorithms where one or

several test problems are designed artificially and true Pareto-optimal front are known, the
true Pareto-optimal fronts for the multi-objective optimization problems of discussed here are
difficult to be known. Therefore, the set coverage metric C (A, B) can only be used to evaluate
the relative convergence of the two sets of solutions with respect to the Pareto-optimal front
of the optimization problem.

In order to evaluate the diversity of the non-dominated solutions obtained by using the three
MOEAs, the following metric is used [Deb02]:

1
∑
Ns
SP = i =1
(d i − d ) 2 (3.26)
Ns

where d i = min k∈Ns ∧k ≠i ∑v =1 f vi − f vk represents the minimum value of the sum of the
m

distances in the objective space between the i-th solution and any other solution.
d = ∑i =1 d i Ns represents the mean value of these distances. Ns represents the number of
Ns

the members in the non-dominated set Ns . m is the number of the objectives, here m = 2 . The
more uniformly the solutions are spaced, the smaller the metric SP.

In order to evaluate the extent of spread of the solutions the following metric is applied
[Zit99]:

DI = ∑
m
v =1
(max Ns
i =1
Ns
f vi − min i =1 f vi )
2
(3.27)

For the two-objective problem, the metric DI measures the length of the hypotenuse of a
triangle formed by the extreme solutions in the objective space. The larger the value of the
metric DI is, the larger the extent of spread of the solutions.

In this chapter, the basic concepts of multi-objective optimization were introduced. The basic
ideas behind various multi-objective optimization methods using scalarization technique and
50 CHAPTER 3

evolutionary algorithms were described. Some representative MOEAs have been introduced
in detail. In the next chapter, multi-objective optimization of piezoelectric actuators using
several MOEAs will be discussed.
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 51

4 Multiobjective Optimization of Piezoelectric Transducers

As the basis for multi-objective optimization of piezoelectric actuators two different kinds of
techniques have been studied in chapter 2 and chapter 3. In this chapter, an integrated
procedure for piezoelectric transducer design via multi-objective optimization is studied. The
structure of this chapter is as follows: First, the Langevin transducer and its performance
criteria are introduced. Then based on lumped parameter models the optimal prestress for
transducers considering multiple objectives is studied. After analytical models based on the
transfer matrix method are introduced, several multi-objective optimization problems of
piezoelectric transducers are studied. In these problems two types of mechanical boundary
conditions, the freely vibrating transducer and the transducer subjected to a mechanical load
are considered.

4.1 Langevin Transducers

In ultrasonic machining and bonding as well as in many other applications, ultrasonic

transducers are used to transfer high frequency electrical energy into high frequency
mechanical vibration of a tool. The ultrasonic process is performed by the tool, which vibrates
at a resonance frequency, generally between 20 kHz and 100 kHz. Piezoelectric transducers
are the most commonly used type of ultrasonic transducers. Fig. 4.1 gives the principle
description of an ultrasonic machining (or bonding) system. Fig. 4.2 shows two typical
applications of piezoelectric transducers.

Most of the piezoelectric transducers have a setup, which can be in principle reduced to a
sandwich construction, where piezoelectric material is sandwiched between metal end blocks
(back section and front section or horn). They are known as so-called Langevin transducers or
sandwiched transducers, see Fig. 4.3. Langevin transducers have the following advantages
[Hul73], [Nep73] and [Rus95]:

AC Ultrasonic machining unit

Power
electronics Transducer Booster Tool Work piece

Control
electronics

Energy flow
Electric control unit
Information flow

Fig. 4.1 Ultrasonic machining (bonding) system

52 CHAPTER 4

70mm

Transducer used in Ultrasonic piezo scaler

ultrasonic bonding
Fig.4.2 Typical piezoelectric transducers

Front section

Piezoelectric
rings/disks

Back section

Fig. 4.3 Basic construction of Langevin-type transducers [Phi91]

1. A mechanical prestress may be imposed on the piezoelectric elements by means of a

central bolt or peripheral sleeve. Thus, the admissible dynamic stress amplitude and
hence the maximum power intensity are considerably increased. The mechanical
contact between the parts is improved and hence the mechanical damping decreases.

2. The mechanical quality factor of metal is generally higher than that of piezoelectric
material. The metal end sections are good heat-sinks, so that the transducer can be
driven at higher vibration levels than other types of ultrasonic transducers.

3. As the manufacture of metal is much easier than that of piezoelectric materials, more
variations of the shape and dimension of transducers are available.

4. As a part of piezoelectric material is replaced by metal, material costs decrease, also

the influence of active material on characteristics of transducers.

As the displacement of the piezoelectric material is limited, a horn is used to amplify this
limited displacement. There are four general designations of horns: constant, linear,
exponential and stepped, which refer to the degree to which the area changes from the base to
the tip. The thesis work concentrates on the design of constant-horn (also symmetrical) and
stepped-horn transducers via multiobjective optimization methods. In the practical
applications, piezoelectric transducers, like almost all mechanical systems, are always
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 53

expected to be “optimal”. In order to achieve optimum performances, optimum pre-stress,

geometrical and material parameters of the transducer must be determined.

4.2 Performance Criteria

There exist different performance criteria for piezoelectric transducers in the different
applications. For the piezoelectric transducers used in ultrasonic machining and bonding
devices, the most commonly applied performance criteria (some of which have been already
described in chapter 2) are collected here.

Resonance frequency The resonance frequency of the transducer should be equal to the
specified working frequency. Commonly, this resonance frequency is that of the first
longitudinal vibration mode, which is often known as so-called λ/2 vibration mode. The cause
of designing a transducer based on the λ/2 vibration mode is that the transducer can be
combined with other λ/2 parts like booster and tool to form a whole ultrasonic machining
device without obvious changes of the eigenform compared with that of each part before
synthesis [Lit03]. The advantage of the λ/2-synthesis is that in the ideal case no forces act at
the interfaces between individual parts. Therefore, the boundary conditions of each part in the
whole synthesized system are the same of those of each part free at both sides. Under this
condition, each part (transducer, booster or tool) can be first developed according to the
specified resonance frequency and then they are synthesized into a whole device. Fig. 4.4
describes the λ/2-synthesis principle schematically.

Electrical input power The input power of the transducer is:

pe (t ) = U (t ) ⋅ I (t ) (4.1)

where U(t) and I(t) are the AC input voltage and current respectively. If the system is not
driven exactly at resonance it will show a capacitive or inductive behavior. The mean power
or effective power is then given by

Transducer Booster Tool

Starting Working
amplitude amplitude

Fig. 4.4 The principle of the λ/2-synthesis

54 CHAPTER 4

T
1
T ∫0
Pe = U (t ) I (t )dt (4.2)

In the case of harmonic U(t) and I(t) the effective power results as
1
Pˆe = Uˆ Iˆ cos ϕ e = Pˆa cos ϕ e (4.3)
2
1
where Pˆa = Uˆ Iˆ is the apparent power, ϕ e is the phase difference between the U(t) and I(t).
2
cos ϕ e is the power factor representing the proportion of the effective power which can be
obtained from the apparent power. Obviously the power factor should approach 1 as closely
as possible. The apparent power determines the size of the power electric device used to drive
transducers. Therefore, in order to reduce the size of the power electric device, the apparent
power should be minimized for a given output requirement such as a given amplitude or
mechanical power.

Mechanical output power By analogy with the definition of the electrical power the moment
mechanical output power of a transducer can be calculated as follows
pm (t ) = F (t )v(t ) (4.4)

In a work period T, the effective mechanical power of the transducer is then given by
T
1
T ∫0
Pm = F (t )v(t )dt (4.5)

For the harmonic vibration the effective power results as follows

1
Pˆm = Fˆ vˆ cos ϕ m = Pˆm a cos ϕ m (4.6)
2
1
where Pˆma = Fˆ vˆ is the apparent mechanical output power, ϕ m is the phase difference
2
between the F(t) and v(t). cos ϕ m is the power factor representing the proportion of the
effective mechanical output power which can be obtained from the apparent mechanical
output power. Similarly the power factor should approach 1 as closely as possible. For a given
input power or input voltage, the effective mechanical output power should be maximized.

Coupling factor A coupling factor (also called the electromechanical coupling coefficient) k
can be defined for each vibration mode in the piezoelectric transducer. Besides the definitions
described given in chapter 2 there exist several definitions for the coupling factor depending
on the manner in which the ratio of mechanical and electrical energy is computed [Ike96]. For
optimization problems of piezoelectric transducers the definition of the coupling factor based
on the equations of a four-pole network element may be preferred since they include the
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 55

transducer dimensions. Generally, the fundamental relation for a piezoelectric transducer can
be expressed in integral form as follows:

Iˆ = y11Uˆ + y12 Fˆ
(4.7)
vˆ = y 21Uˆ + y 22 Fˆ

Where Iˆ , v̂ , Û and F̂ are the complex amplitudes of electric current, vibration velocity,
voltage and force, respectively. Each admittance expression y ij is a function of the driving
frequency Ω. That is, the admittances depend on the excitation frequency Ω. From equation
(4.7) the coupling factor can be derived as follows
2
2
y12
k eff = (4.8)
y11 y 22

For quasi-static operation, the coupling factor keff results as

2 2
k = lim k eff (4.9)
Ω→0

The coupling factor is an overall measure for the excitation and energy conversion. It is noted
that the coupling factor does not represent conversion efficiency.

Power efficiency The power efficiency λp can be defined as the ratio between the mechanical
energy delivered and the electrical energy absorbed by the transducer, see [Cer00]. For the
harmonic vibration of the transducer, this results in

Pˆm Fˆvˆ cos ϕ m

λp = = (4.10)
Pˆa Uˆ Iˆ

i.e. the power efficiency is the quotient of the effective mechanical output power and apparent
electrical input power of the transducer. Obviously, the power efficiency of the transducer
should be maximized.

Efficiency The efficiency η is defined as a ratio of the output mechanical energy to the
consumed electrical energy or the output electrical energy to the consumed mechanical
energy. In a work cycle, the input electrical energy is transformed into mechanical energy
partially and the remaining is stored as electrical energy in a transducer if there is no electrical
loss. The stored energy can be returned to the source during discharge of the transducer. For
the harmonic vibration of the transducer, the efficiency η can be calculated as follows

Pˆm Fˆvˆ cos ϕ m

η= = (4.11)
Pˆe Uˆ Iˆ cos ϕ e
56 CHAPTER 4

The efficiency η refers to the quotient of the effective mechanical output power and effective
electrical input power of the transducer. It should be maximized.

The above performance criteria can also be explained according to the energy balance of the
transducer as shown in Fig. 4.5. Respectively, they can be defined as follows

WmA + Wm W Wm
k2 = , λp = m , η = (4.12)
Wa Wa Wa − WeA

Mechanical quality factor As described in Chapter 2, the mechanical quality factor Qm

describes the resonance rise of the piezoelectric transducer. It is defined as an inverse of the
loss factor tan δ m . Obviously, a large mechanical quality factor Qm corresponds to a large
efficiency of the piezoelectric transducer and a large resonant amplitude. The material and
structural damping of the piezoelectric transducer mainly determine the mechanical quality
factor.

Piezoelectric quality number According to the description in chapter 2, the piezoelectric

quality number M presents the phase rise or descend of the admittance functions of the
piezoelectric transducer. When M<2, the resonance and anti-resonance frequencies do not
exist any more, the transducer can not be driven with zero reactive power. More apparent
power is needed resulting a large power electric device. Therefore, the piezoelectric quality
number of the transducer used in ultrasonic bonding and machining should be larger than two.
Furthermore, the larger the value of M is, the better the phase reserve of the transducer,
compare Fig. 2.11 and Fig. 2.12 in chapter 2. M>2 assures that the resonance frequency exists
and that the transducer can be driven with zero reactive power even though the load damping
may be large.

Wa Wm
WeA WmA

WeL WmL

Wa: Electrical energy absorbed by the transducer

WeA:Electrical energy stored by the transducer
WeL:Electrical energy loss
Wm: Mechanical energy delivered by the transducer
WmA:Mechanical energy stored by the transducer
WmL:Mechanical energy loss

Fig. 4.5 The energy balance of the transducer

MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 57

Volume of piezoelectric materials In general, piezoelectric materials are expensive.

Therefore, the volume of piezoelectric materials should be minimized in order to reduce the
cost of the whole transducer. However, the small volume will result in the increase of the
strain in the piezoelectric material and further leads to the rise of the temperature of the
piezoelectric material when the transducer is driven at resonance. This will increase the
mechanical loss of the transducer. Therefore a trade-off for the volume of piezoelectric
material should be considered.

Mass and size For the ultrasonic hand machining or drilling device the light and small
piezoelectric transducer is expected. The material type of the metal end blocks mainly
determines the mass and size of the piezoelectric transducer. For a given vibration mode and
resonant frequency, different mass and size of the transducer will be obtained if different
material types are used. Therefore the material types must be selected carefully in order to
optimize the mass and size of the transducer.

There are two cases when the above performance criteria are used as objective functions in
Multi-objective optimization problems. One case is that all performance criteria used as
objective functions can be improved simultaneously in the optimization process. This means
that the minimum solution corresponding to any objective function is the same. Therefore,
one can find the optimum design variables for all considered performance criteria using only
one criterion as the objective functions. The other case is that all performance criteria used as
objective functions cannot be improved simultaneously in the optimization process. This
means that the objectives are conflicting to each other. There exist multiple Pareto-optimal
solutions. Indeed, the latter is more interesting and also the concentration of this thesis work.
In [Sat03] the variations of the values of some performance criteria with respect to some
geometrical and material parameters of a transducer have been studied qualitatively. This can
be used as reference for determining objective functions in the following optimization
problems.

In the following, the optimal prestress for the transducer is first determined. Then optimum
geometrical and material parameters for each type of the transducer are discussed
respectively.

4.3 Determination of Optimal Prestress

Prestress is a very sensitive parameter in the design of the Langevin transducer. As the
influence of the prestress on the resonance performance of the transducer can not be modeled
in theory, experimental study will be applied. The basic steps are: first, lumped parameter
models are applied and optimization problems are described using model parameters. Then,
model parameters are identified from experimental results according to different prestress.
Finally, the optimum prestress for multiple objectives is determined. Because the model
parameters are estimated from experiments, this approach can be applied for any transducer
configuration. Losing no generality, a bolt-clamped Langevin-type piezoelectric transducer of
58 CHAPTER 4

length 70 mm has been constructed as shown in Fig. 4.6. A piezoelectric stack consisting of 4
piezoelectric rings, 4 copper electrodes and a force washer are clamped between two same
steel rods by means of a bolt. The bolt is insulated from the center electrode using a thin PVC
sleeve. The force washer based on strain gauge techniques was used to monitor pres-stress.

An experimental setup including an impedance analyzer (HP4192) and a laser vibrometer was
used for measure of admittance frequency responses Iˆ Uˆ and vˆ Uˆ with respect to various
levels of the prestress between 20MPa and 100MPa. Fig.4.7 shows the measured admittance
function Iˆ Uˆ and vˆ Uˆ in the vicinity of the first longitudinal vibration mode according to
different pre-stress. Obviously, the locations of resonance and anti-resonance frequencies as
well as admittance maxima and minima vary when prestress varies.

When a piezoelectric transducer operates in the vicinity of a resonance frequency, an

equivalent electrical or mechanical model shown in Fig. 2.8 in chapter 2 can describe its
vibration behavior. Similarly, the electrical and mechanical admittance functions Iˆ Uˆ and
vˆ Uˆ can be approximated as follows:

Iˆ jΩ jΩα 2
Yˆ e = = + (4.13)
Uˆ jRΩ + 1 − mΩ + jdΩ + c
2

C
vˆ jΩα
Yˆ m = = (4.14)
Uˆ − mΩ + jdΩ + c
2

Model parameters can be calculated analytically from measured admittances [Kro99] and
[Hem01]. In this work, the nonlinear least-square estimation technique was used for the
parameter identification, based on the measurement of the electrical admittance frequency
response Yˆ e only.

Force washer

Fig. 4.6 Bolt-clamped Langevin transducer

MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 59

-1
10

-2
10
|Ye| e[ A/V ]

-3
10

-4
10

-5
10
2.1 2.2 2.3 2.4 2.5 2.6 2.7
Frequency [Hz] 4
x 10

100

20MPa
50 30MPa
Phase [Grad]

40MPa
50MPa
Phase [ °]

0 60MPa
70MPa
80MPa
-50 90MPa
100MPa

-100
2.1 2.2 2.3 2.4 2.5 2.6 2.7
Frequenz [Hz] ×1044
-1 Frequency [Hz]
10

-2
10
|Ym| m[ m/sV ]

-3
10

-4
10
2.1 2.2 2.3 2.4 2.5 2.6 2.7
Frequency [Hz] x 10
4

100

50 20MPa
[ °][Grad]

30MPa
40MPa
50MPa
0
Phase

60MPa
Phase

70MPa
80MPa
-50
90MPa
100MPa

-100
2.1 2.2 2.3 2.4 2.5 2.6 2.7

Frequency [Hz] ×104

Fig. 4.7 Admittances as functions of prestress

60 CHAPTER 4

The parameter identification problem can be formulated as a least-square optimization

problem as follows

min6 f (φ) =
φ∈R
1 q
∑
2 i =1
[ (
(Yi − Yi )2 + ∠Yi − ∠Y i )]
2
(4.15)

where φ = [c m α C d R ] is estimated parameter vector. Yi and Yi are the measured and

T

estimated magnitudes of admittance function Yˆ e at the i-th sample point frequency (i=1, 2,…,
q), respectively. ∠Yi and ∠Y i are the measured and estimated phases of admittance function
Yˆ e at the i-th sample frequency point frequency (i=1,2,…, q). Here q=200 sample frequency
points are used for parameter identification. Fig. 4.8 shows model parameters estimated with
respect to varying prestress levels between 20MPa to 110MPa. The functional expressions of
the model parameters with respect to the prestress Pr were then acquired by means of curve
fitting with polynomials in the least-square sense. The resulting expressions are:

N  
2
 1  1
c = 1.28 × 10 9 − 4 × 10 −5   Pr 2 + 0.0081 Pr + 0.5818  (4.16)
m   MPa  MPa 

 1 
m = 0.05 kg  0.0004 Pr + 0.9133  (4.17)
 MPa 

N  
2
 1  1
α = 0.5013 − 8 × 10 −5   Pr 2 + 0.012 Pr + 0.5008  (4.18)
V   MPa  MPa 

 1 
C = 2.85 × 10 −9 F 0.0024 Pr + 0.7391 (4.19)
 MPa 

Ns  
2
 1  1
d = 41.59 7 × 10 −5   Pr 2 − 0.0142 Pr + 1.2539  (4.20)
m   MPa  MPa 

R = 60.58 Ohm ×
 3 2

 − 3 × 10 −6  1  Pr 3 + 0.0008 1  Pr 2 − 0.0723 1 Pr + 2.1752 
(4.21)
  MPa   MPa  MPa 
 

Applying the above expressions, the performance criteria can be expressed as functions of
prestress Pr if the transducer is modeled as a lumped mass system. Next some optimization
problems related to prestress are discussed.

When only one performance criterion is considered as the optimization objective in every
optimization process, single-objective optimization is performed. However, in most practical
applications multiple performance criteria with conflicting objectives are encountered and
Pareto-optimal solutions need to be found.
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 61

1.2

1
c

Parameter values normalized to

m
α

the corresponding maxima

0.8
C
d
0.6
R

0.4

0.2

0
20 30 40 50 60 70 80 90 100 110

Prestress[MPa]

Fig. 4.8 Estimated model parameter values as functions of pre-stress. The values are
normalized to the corresponding maxima.

4.3.1 Freely Vibrating Transducers

From the point of view of ultrasonic engineering the coupling factor k 2 and piezoelectric
quality number M are two important performance criteria. Here they are considered as
objective functions that need to be maximized simultaneously. The expressions of these two
objectives can be obtained according to equations (2.12) and (2.60) in chapter 2, respectively.
The optimization problem can then be formulated as follows:
ω p 2 − ωs 2 α2
maximize f1 (Pr ) = k = 2
= (4.22a)
ωp2 c C +α 2

α2
maximize f 2 ( Pr ) = M = (4.22b)
ω s dC
subject to 20MPa ≤ Pr ≤ 110MPa

As there is only one design variable Pr, Pareto-optimal solutions can be readily found. Fig.
4.9(a) shows the two objectives k 2 and M as functions of pre-stress, respectively. The values
of the objectives are normalized with respect to the corresponding maxima. The maximum of
each objective function can be found using normal numerical optimization method. The
prestress corresponding to the maximum effective coupling factor k 2 = 0.068 is Pr = 55.5MPa,
to the maximum piezoelectric quality number M=22.3 is obtained for Pr = 80.7MPa.
Obviously, the Pareto-optimal set is the interval [55.5MPa, 80.7MPa]. Fig. 4.9(b) presents
values of objective functions in the objective space. Though this two-objective optimization
problem appears to be trivial, it gives some interesting results:
62 CHAPTER 4

1 1

Piezo quality number

Normalized values

0.9 0.9

0.8 0.8
M
0.7 0.7
2
0.6
k 0.6

0.5 0.5

20 40 60 80 100 0.7 0.75 0.8 0.85 0.9 0.95 1

Coupling factor
Prestress [MPa]

(a) (b)
Fig. 4.9 (a) k 2 and M as functions of prestress (b) Pareo-optimal solutions in the objective
space

1. Varying prestress will lead to the variation of the resonance frequencies of

piezoelectric transducers. This will result in the change of resonance performances.

2. The performance criteria k 2 and M with respect to prestress can not arrive at optimum
simultaneously. There exist Pareto-optimal solutions. For the transducer prototype here
studied, the set of the Pareto optimal prestress is the interval [55.5 MPa, 80.7 MPa].
Correspondingly, the set of resonance frequencies is [25.09 kHz, 25.86 kHz].

Similarly, other two or multiple performance criteria can be considered as optimization

objectives and the corresponding Pareto-optimal solutions with respect to prestress can be
searched for.

4.3.2 Transducers with a Mechanical Load

In ultrasonic bonding and machining, piezoelectric transducers operate against loads (the
bonded or machined work pieces). In general, for simplicity, the load can be modeled as a
spring-damping load with a stiffness cL and a damping dL. Fig. 4.10 shows the equivalent
mechanical model of the transducer with a mechanical load. Here, the effective coupling
factor k 2 and the power efficiency λp are considered as objective functions.

According to the equivalent model shown in Fig. 4.10, the following equations are obtained:
mu&& + du& + (c + c L )u = −d L u& + αU (4.23)

1
C
( )
(Q − αu ) + R Q& − αu& = U (4.24)

If the input voltage is U (t ) = Uˆ ⋅ e jΩt , the following harmonic responses are obtained:
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 63

Q I 1/C

R
cL
v
α
dL
c
m
1
d

Fig. 4.10 Equivalent mechanical model of the transducer with a mechanical load

α Uˆ
uˆ = (4.25)
c + cL + j (d + d L )Ω − mΩ 2

jΩα Uˆ
vˆ = jΩ ⋅ uˆ = (4.26)
c + c L + j (d + d L )Ω − mΩ 2

jΩUˆ jΩα 2 Uˆ
Iˆ = jΩ ⋅ Qˆ = + (4.27)
1 c + cL + j (d + d L )Ω − mΩ 2
jRΩ +
C

By analogy with equations (2.53) and (2.54) in chapter 2, the series and parallel resonant
frequencies in this case are calculated as follows:

c + cL
ωs = (4.28)
m

α2
c + cL +
ωp = C (4.29)
m

According to the definition of the coupling factor k 2 , the first objective function is

2
ω 2p − ω s2 α2
f1 ( Pr ) = k = = (4.30)
ω p2 C (c + c L ) + α 2

When the transducer is driven at the resonant frequency, i.e. Ω = ω r ≈ ω s ,

α Uˆ m
uˆ res = (4.31)
j (d + d L ) c + c L
64 CHAPTER 4

α Uˆ
vˆ res = (4.32)
d + dL

j (c + cL ) mUˆ α 2 Uˆ
Iˆ res = + (4.33)
1 d + dL
jR (c + cL ) m+
C

The output force of the transducer is given by

c Lα Uˆ m d Lα Uˆ
Fˆ res = c L uˆ res + d L vˆ res = + (4.34)
j (d + d L ) c + c L d + dL

Applying the equations (4.32) and (4.34), the effective mechanical power P̂m delivered at
resonance, with respect to Pr is given by
1
Pˆm = Fˆres vˆres cos ϕ m (4.35)
2
Where ϕ m = ∠vˆ res − ∠ Fˆ res is the phase difference between v̂ res and F̂ res . The apparent
electrical input power P̂a at resonance can be calculated as follows:
1
Pˆa = IˆresUˆ (4.36)
2

Applying equations (4.35) and (4.36), the second objective function, namely the power
efficiency λp at resonance, with respect to Pr is given by

Pˆm
f 2 ( Pr ) = λ p = (4.37)
Pˆa

In the following the typical values c L = 1.27 kN/µm and d L = 25.4 / ω s Ns/m have been
chosen. The optimization problem is then formulated as follows:
maximize f1 ( Pr )
maximize f 2 ( Pr )
subject to 20MPa ≤ Pr ≤ 110MPa
Uˆ = 36 V
where model parameters α, d, c, m, C, and R in the expressions of the objective functions are
determined by equations (4.16) to (4.21), respectively.
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 65

1 1

Power efficiency
Normalized values

0.95 0.95
0.9 0.9
k2
0.85 0.85
0.8 λp 0.8
0.75 0.75

20 40 60 80 100 0.8 0.85 0.9 0.95 1

Prestress [MPa] Coupling factor

(a) (b)

Fig. 4.11 (a) k 2 and λp as functions of pre-stress (b) Pareto-optimal solutions in objective
space

Similarly, as there is only one design variable, the Pareto-optimal solutions can be readily
described by the plots as shown in Fig. 4.11. The values of the objectives are normalized with
respect to the corresponding maxima. The maximum of each objective function (the inverse
of the minimum of the objective function) can be found using normal numerical optimization
method. The prestress corresponding to the maximum coupling factor k 2 =0.034 is Pr = 62.3
MPa, to the maximum power efficiency λp = 0.78 is Pr = 85.9 MPa. The Pareto-optimal set is
the interval [62.3MPa, 85.9MPa].

In the above optimization problems, Pareto-optimal prestress has been studied individually
based on experiments and lumped parameter models. Next, optimization of the piezoelectric
transducer with respect to the design variables other than prestress will be discussed. First,
modeling of the transducer using the transfer matrix method based on continuum models is
introduced. This modeling method will be used in all optimization problems discussed later.
Then, the optimization problems of the transducer with a constant horn (i.e. symmetrical
transducers) and the transducer with a stepped horn will be discussed in turn.

4.4 Modeling of Langevin Transducers using Transfer Matrix Methods

The goal of modeling is to construct the mathematical expressions between the input
quantities and output quantities of the transducer. As described in chapter 2, continuum
models are directly related to dimensions, material parameters and boundary conditions and
are suitable for the formulation of the optimization problem. However, they can generally
only deal with simple and homogeneous elements. In order to apply continuum models in the
design of Langevin transducers, here the transfer matrix method based on continuum models
is introduced.

If the lateral dimensions of the whole transducer are smaller than one quarter of the
wavelength in the frequency range of interest, the simple rod theory is well adequate for the
66 CHAPTER 4

approximation of their vibration behaviors. Therefore, the basic processes of the transfer
matrix method are: First, the whole transducer is split into elementary mechanical and
piezoelectric rod parts, which are homogeneous and geometrically simple and connected in
series or in parallel. Second, for each rod part the transfer matrix relation is derived by means
of the rod theory. Third, in terms of the deformation continuity and force equilibrium, all
transfer matrix relations are concatenated to form a whole transfer matrix model for
describing the electromechanical behavior of the whole transducer. In the following, the
transfer matrix method is first introduced, and then the expressions of transfer matrixes are
derived.

For each block the analytical model allows the computation of the vibration velocity and force
at the right side of the block with respect to the vibration velocity and force at its left side.
According to Fig. 4.12 a transfer matrix relation for each mechanical block can be written as
follows

 vˆ a  m  ve 
ˆ
 Fˆ  = A  Fˆ  (4.38)
 a  e

where A m is the 2× 2 transfer matrix , which is derived from analytical models. The
expression can be reduced to the following form

Xa = Am Xe (4.39)

where

 vˆ   vˆ 
Xa =  a  , X e =  ˆe  (4.40)
ˆ
F a  F e 

In the case of a piezoelectric block (see Fig. 4.13), the transfer matrix relation can be written
as

Rod element Mechanical schematic Four-pole

Description description
ve va
Lb ve va
Ab ve va
Fe Fa Fe Fa
Fe Fa

Fig. 4.12 Rod element, mechanical scheme and four-pole description

MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 67

ve va
hp
Ap ve va

Fe Fa Fe Fa Fe Fa

I U I U
I U
z

Fig. 4.13 Piezoelectric rod element, mechanical scheme and four-pole description

 vˆ a   vˆ e 
 Fˆ  = A P  Fˆ  (4.41)
 a  e
 I 
ˆ  Uˆ 

where Iˆ and Û are the electric current and voltage, respectively. A P is a 3× 3 transfer
matrix, which is derived from analytical models. Similarly, it can be rewritten as

 X a   A Pm A Pem   X e 
 ˆ  =  Pme   (4.42)
 I  A A Pe   Uˆ 

where A Pm , A Pem ( A Pme ) and A Pe are 2× 2 , 2 × 1 (1 × 2) and 1× 1 transfer matrices,

respectively. They describe the mechanical field, electric field and electrometrical coupling
effects of the piezoelectric block, respectively.

The whole transfer matrix relation of the transducer can be obtained by connecting transfer
matrices of the elementary blocks in terms of interface conditions between two blocks. There
are four fundamental connections shown in Table 4.1. Using these fundamental connections
and the corresponding transfer matrices, the whole transfer matrices for various Langevin-
type transducers can be assembled. The transfer matrices of these fundamental connections
are derived as follows:

Connection 1: connection of two mechanical blocks in series

According to equations (4.38) and (4.39), the transfer matrix relations for block 1 and block 2
can be respectively written as follows

X a1 = A 1m X e1 , X a2 = A m2 X e2 (4.43)

Assuming that the deformation is continuous at the interface of two block elements, there
exist the following interface conditions

vˆ a1 = vˆ a 2 , Fˆ a1 = Fˆ e 2 (4.44)
68 CHAPTER 4

Table 4.1 Basic connections and their transfer matrices

Mechanical schematic and four-pole Transfer matrix relations
description
ve1 va1 ve2 va2
 va 2  m m  e1 
v
1. Connection Fe1 1 2 Fa2 F  = A 2 A1  F 
of two  a2   e1 
mechanical X a2 = A m2 A 1m X e1
blocks ve1 va1 ve2 va2
Fe1 1 Fa1 Fe2 2 Fa2

 X a2   A Pm
w A Pem
w
  X e1 
2. Connection ve1 va1 ve2 va2  Iˆ  =  Pme Pe   ˆ 
 w  w A A w  U 
of two
Fe1 1
P P
2 Fa2
piezoelectric where
blocks U U A Pm Pm Pm
w = A 2 A1

ve1 va1 ve2 va2 A Pem

w = A Pm Pem
2 A1 + A Pem
2

Fe1 1 Fa1 Fe2 2 Fa2 A Pme

w = A 1Pme + A Pme
2 A 1Pm
A Pe Pe Pe Pme
w = A1 + A 2 + A 2 A 1Pem
I1 U U I2
Iˆ w = Iˆ1 + Iˆ 2

 X a2   A MPm
w A MPem
w
  X e1 
3. Connection ve1 va1 ve2 va2  ˆ  =  MPme MPe   ˆ 
of one  I  A w Aw  U 
mechanical Fe1 1 2 Fa2
A MPm
w = A Pm m
2 A1
block (left) and I U
a piezo block A MPem = A Pem
ve1 w 2
(right)
Fe1 1 Fa1 Fe2 2 Fa2 A MPme
w = A Pme
2 A 1m
I U A MPe = A Pe
w 2

 X a2   A PMm
w A PMem
w
  X e1 
4. Connection ve1 va1 ve2 va2  ˆ  =  PMme PMe   ˆ 
 I A
  w Aw  U 
of one piezo
block (left) and Fe1 1 2 Fa2
A PMm
w = A m2 A 1Pm
a mechanical I U
block (right) A PMem
w = A m2 A 1Pem
ve1 va1 ve2 va2
A PMme
w = A 1Pme
Fe1 1 Fa1 Fe2 2 Fa2
A PMe
w = A 1Pe
I U
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 69

From equations (4.43) and (4.44), the whole transfer matrix relation is obtained as follows

X a2 = A m2 A 1m X e1 (4.45)

Following this approach, the whole transfer matrix relation for N blocks in series can also be
derived.

Connection 2: connection of two piezoelectric blocks mechanically in series and

electrically in parallel

According to equation (4.42), the transfer matrix relations for piezoelectric block1 and block2
can be respectively written as follows

 X a1   A 1Pm A 1Pem   X e1   X a 2   A Pm
2 A Pem
2
 Xe 2 
 ˆ  =  Pme Pe   ˆ   ˆ 
, =  Pme  ˆ  (4.46)
 I 1  A1 A1   U   I 2  A 2 A Pe
2  U 

Similarly, using the interface conditions X a1 = X e 2 , the whole transfer matrix relation can be
obtained as follows

 Xa2   A Pm
w A Pem
w
  X e1 
 ˆ  =  Pme Pe   ˆ 
(4.47)
 I  A w Aw  U 

Where

A Pm Pm Pm
w = A 2 A1

A Pem
w = A Pm Pem
2 A1 + A Pem
2

A Pme
w = A 1Pme + A Pme
2 A 1Pm
A Pe Pe Pe Pme
w = A1 + A 2 + A 2 A 1Pem
I = I1 + I 2

Following this approach, the whole transfer matrix relation for a piezoelectric stack consisting
of N piezoelectric blocks is available.

Connection 3: a mechanical rod (left) + a piezoelectric element (right)

According to equations (4.39) and (4.42) the transfer matrix relations for the mechanical
block 1 and the piezoelectric block 2 can be written as follows

X a1 = A 1m X e1 (4.48)

 X a2   A Pm
2 A Pem
2
  X e2 
 Iˆ  =  Pme Pe   ˆ 
(4.49)
   2
A A2  U 

Applying the interface conditions X a1 = X e2 yields the whole transfer matrix relation as
follows
70 CHAPTER 4

 X a2   A MPm
w A MPem
w
  X e1 
 ˆ  =  MPme MPe   ˆ 
(4.50)
 I  A w Aw  U 

where

A MPm
w = A Pm m
2 A1

A MPem
w = A Pem
2

A MPme
w = A Pme
2 A 1m

A MPe
w = A Pe
2

Connection 4: a piezoelectric element (left) + a mechanical rod (right)

Applying a similar approach as in connection 3, the whole transfer matrix relation can be
obtained as follows

 X a2   A PMm
w A PMem
w
  X e1 
 ˆ  =  PMme PMe   ˆ 
(4.51)
 I  A w Aw  U 

where

A PMm
w = A m2 A 1Pm

A PMem
w = A m2 A 1Pem

A PMme
w = A 1Pme

A PMe
w = A 1Pe

The expressions of the transfer matrices of the fundamental block elements can be derived
using the analytical models based on rod theory, respectively. A detailed derivation can be
found in Appendix. In the next sections, the optimization problems of symmetric transducers
and transducers with stepped horn will be discussed.

4.5 Optimization of Symmetrical Langevin-type Transducers

The symmetric Langevin-type transducer is the simplest Langevin-type transducer. Fig. 4.14
shows a typical design. A piezoelectric stack consisting of N (for example N=4) piezoelectric
rings and copper electrodes is clamped between two same metal rods by means of a bolt. The
piezoelectric rings and metal rods have the same outer diameter. Table 4.2 gives those
parameters which affect the performances of the transducer.
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 71

4h p

Dt

hp
2Lb +4h p

Fig. 4.14 A symmetrical Langevin transducer

ve1 va1 ve2 va2 v e3 va3

Fe1 Fa1 Fe2 Fa2 Fe3 Fa3

I U

Fig. 4.15 Four-pole network element description of the symmetrical Langevin transducer

Table 4.2 Parameters that affect the performances of the transducer

Û : Input voltage
Lb: Length of the metal parts
Dt: Outer diameter of metal rods and piezo-disks/rings
Di: Inner diameter of piezo-rings and diameter of the bolt
Lbt: Length of the bolt
hp : Thickness of piezoelectric rings

N: Number of piezoelectric disks or rings

Typb: Material parameters of metal end blocks
Typp: Material parameters of piezo-rings

4.5.1 Derivation of the Whole Transfer Matrix

The whole transducer is considered as a combination of two mechanical rod elements and a
piezoelectric stack in series. Then the whole transfer matrix relation of the transducer can be
derived from the transfer matrices of the fundamental rod elements shown in Table 4.1. Fig.
4.15 gives the four-pole network element description. Considering the modeling in the earlier
design stage, the bolt has not been modeled as a separate element. The actual piezoelectric
rings are modeled as piezoelectric discs. The electrodes are not taken into account in the
model. These simplifications will not cause obvious deterioration of the model.
72 CHAPTER 4

Based on the transfer matrix relation of the fundamental connection 2 shown in Table 4.1, it is
easy to obtain the transfer matrix relation for the piezoelectric stack consisting of N
piezoelectric rings in series. It is noted that all piezoelectric rings here have the same
thickness hp. They are connected in parallel electrically and the whole current for the piezo-
stack is

N   Pe N  Pe i −2 Pme Pm r Pem  ˆ
Iˆ = ∑ A Pme A Pm ( )
i −1
 e2  A + ∑  A + ∑ A
X + (
A )
A  U (4.52)
 i =1   i=2  r =0 

The whole mechanical output (the vibration amplitude and force) on the right side of the
piezoelectric stack is

 N −1 Pm i Pem  ˆ
X a2 = A( Pm N
) X e2 + ∑ A (
A U ) (4.53)
 i =0 

According to equations (4.52) and (4.53), the whole transfer matrix relation for the
piezoelectric stack can be written as follows

 Xa2   A PSm A PSem   X e2 

 ˆ  =  PSme   (4.54)
 I  A A PSe   Uˆ 

Where

 vˆ   vˆ 
X a2 =  a 2  , X e2 =  e 2 
 Fˆ a 2   Fˆ e 2 
A PSm = A Pm( ) N

N −1
A PSem = ∑ A Pm A Pem ( )
i

i =0
N −1
A PSme = ∑ A Pme A Pm ( )
i

i =0
N
 i −2

A PSe = A Pe + ∑  A Pe + ∑ A Pme A Pm A Pem 
r
( )
i=2  r =0 

Applying the above equation (4.54) and the transfer matrix relations of the fundamental
connections 3 and 4 shown in Table 4.1 yields the whole transfer matrix relation for the
symmetrical Langevin-type transducer as follows:

 X a 3   A PTm A PTem   X e1 
 ˆ  =  PTme  
 I  A A PTe   Uˆ 
(4.55)

Where
 vˆ   vˆ 
X e1 =  e1  , X a 3 =  a 3 
 Fˆ e1   Fˆ a 3 
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 73

PTm m PSm m a11PT a12PT 

A =A A A =  PT PT 
a21 a22 

PTem m PSem a13PT 

A =A A =  PT 
a23 

A PTme = A PSme A m = a31PT [ a32PT ]

A PTe = A PSe = a33PT

It should be pointed out that the matrix elements aijPT (i, j = 1, 2, 3) are functions of the material
parameters, dimensional parameters and the vibration frequency of the transducer. Their
explicit mathematical expressions are not given here because the expressions are very
complex. They are derived automatically in the programs.

4.5.2 Problem Formulation of the Symmetrical Transducer without Loads

Optimization objectives For the resonant driven piezoelectric transducer, the free vibration
amplitude and input power at the resonance operation are two important performance criteria.
Here they are considered as objective functions. For a given exciting voltage, the vibration
amplitude should be maximized. In the meantime, the input power of the transducer should be
minimized. In the following, the mathematical descriptions of these two objectives will be
derived from the above model. In order to predict vibration amplitude and current at
resonance frequency, losses due to material properties are included in the model using
complex moduli instead of real moduli.

For the transducer free at both sides, there exist the following boundary conditions

Fˆ e1 =0, Fˆ a 3 = 0 (4.56)

Introducing (4.56) into (4.55) yield the following expressions

PT
a 23
vˆ e1 = − PT
Uˆ (4.57)
a 21

vˆ a 3 = a11PT vˆ e1 + a13PT Uˆ (4.58)

Iˆ = a31PT vˆ e1 + a33PT Uˆ (4.59)

After inserting equation (4.57) into equation (4.58), the solution of vˆ a 3 is obtained as follows:
74 CHAPTER 4

 a PT a PT ˆ
vˆ a 3 =  a13PT − 11 PT23 U (4.60)
 a 21 

where

 a PT a PT 
Im a13PT − 11 PT23  = 0 (4.61) is
 a 21 
the so-called characteristic equation representing the resonance of the transducer free at both
sides.

When the transducer is driven by the voltage Uˆ e jΩt , the vibration velocity vˆ a 3 at resonance
can be calculated from equation (4.60). Therefore, the vibration amplitude at the right side of
the transducer can be obtained from

1  PT a11PT a 23
PT
ˆ
uˆ a 3 =  a13 − U (4.62)
jΩ  PT
a 21 

Equation (4.62) gives the expression of the first objective function, which should be
maximized. As optimization methods applied in this work only handle minimization
problems, the inverse of the above expression will be used as the objective function in the
optimization problem. It is given by
PT
a 21 Ω
f1 = (4.63)
a13 a 21 − a11 a 23 Uˆ
PT PT PT PT

Introducing equation (4.57) into equation (4.59) yields the input current

 a PT a PT ˆ
Iˆ =  a33PT − 31 PT23 U (4.64)
 a 21 

Therefore the input electrical power (apparent power) for the harmonic vibration is

1 Uˆ 2 PT a31PT a 23
PT
Pˆa = Iˆ Uˆ = a33 − PT
(4.65)
2 2 a 21

Equation (4.65) gives the expression of the second objective function. It is given by

f 2 = Pˆa (4.66)

Design variables The parameters that affect the performances of the transducer with specified
resonance frequency and vibration mode have been shown in Table 4.2. There are two types
of the parameters. The input voltage U and the dimensions of the transducer are of continuous
type. The number of the piezoelectric rings and material of the backing rod, front rod and
piezoelectric rings are of discrete type. In this work, the number of the piezoelectric rings N is
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 75

assumed as one of the elements of the set {2, 4, 6}. The metal rods take one of 4 pre-specified
material types (represented by corresponding density ρ, elastic modulus E and loss factor
tanδ). Not all parameters need to be determined by the optimization process. Some parameters
can be given before optimization. Here the following parameters are considered as given
quantities: the number of the piezoelectric rings N ∈{2, 4, 6} , the input exciting voltage Û and
the material type of the piezo-rings. As stated before, the inner diameter of the piezo-rings
and the dimensions of the bolt will not be considered as design variables individually. The
remaining design variables, which will be determined by the optimization process, are shown
in Table 4.3. Table 4.4 gives material data.
Constraints The transducer should operate at the specified resonance frequency ω r . Equation
(4.61) gives the resonance condition of the transducer free at both sides. Under the condition
that the values of the above given parameter are known, equation (4.61) describes an equality
constraint g1 with respect to the design variables Lb, hp and the material type of the metal end
rods. It is noted that there exist multi-mode solutions to equation (4.61) according to Lb and hp
for a given material type of the end rods. This can be illustrated with Fig. 4.16, which shows
the locus of roots of equation (4.61) in the ranges 0.2 mm ≤ h p ≤ 5 mm and
5 mm ≤ Lb ≤ 200 mm for the following given parameter values:

• the number of the piezoelectric rings N=2

• the material type of the backing and front rods: steel

• the piezoelectric material: PIC181

• the input voltage Uˆ = 100V

• the resonance frequency f r = 20kHz

Table 4.3 Design variables for the symmetrical transducer

Lb: Length of the metal parts

Dt: Outer diameter of metal rods and piezo-rings
hp: Thickness of piezoelectric rings
Typb: Material of metal end blocks (represented by corresponding
density ρ, elastic modulus E and loss factor tanδ)
76 CHAPTER 4

Table 4.4 Material types and corresponding parameters

Material Steel Titanium Aluminum Brass PIC 181
characteristics bronze (PI Ceramic Gmbh)
Density ( ρ) 7900 4430 8500 8400 7850
[
kgm −3 ]
Elastic modulus ( E) 2.1 1.15 1.4 0.9 0.7
[
× 1011 Nm −2 ]
Loss 4.3 11.4 4.3 3.3 29
factor( tanδ )
[ ]
× 10 −4
Charge constant ( d33) -- -- -- -- 265
[× 10 −12 mV -1 ]
Voltage -- -- -- -- 25.2
constant (g33)
[× 10 −3 VmN -1 ]
Relative -- -- -- -- 1200
permittivity
( T
ε 33 ε0)
Binary String 00 01 10 11
No. 1 2 3 4

0.005
Thickness of piezo-rings [m]

0.004
2nd
mode
0.003

0.002
1st 4th
mode mode
0.001 3rd
mode

0.05 0.1 0.15 0.2

Length of the metal parts [m]

Fig. 4.16 Locus of roots of equation (4.61)

MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 77

The locus of roots shown in Fig. 4.16 contains the first four vibration modes of the transducer.
In order to obtain the solutions for a specified vibration mode (here the first vibration mode,
also λ/2-mode) and to avoid the “switching” of the vibration modes during search, the search
ranges of the variables hp and Lb must be defined accordingly. The corresponding bounds of
hp and L are determined as follows: The lower bounds of hp and Lb are specified by the
designer. The upper bound of hp is either set by the designer or obtained by searching the root
of equation (4.61) corresponding to the lower bound of Lb. Similarly, the upper bound of Lb is
determined by searching the root of equation (4.61) corresponding to the lower bound of hp.
As the derivative of the constraint function g1 can not be found in a closed form analytical
solution, the secant method is used to numerically search solutions of equation (4.61). The
starting values must be picked carefully and tried sufficiently so that the secant method can
find the solutions only for the λ/2 mode. The plot of solutions to equation (4.61) can be used
for specifying the starting values for searching the upper bounds of Lb. For example,
according to Fig. 4.16, the two starting values for searching the upper bound of Lb can be
selected as 5 mm and 50 mm. Similarly, for the transducer with other material of metal end
blocks and number of the piezo-rings, the corresponding locus of roots of equation (4.61) can
be obtained and be used for specifying the starting values.

Considering design requirements for the electric driving device of the transducer, the
maximum apparent input power P̂max and current Iˆmax need to be limited. The constraints can
be concluded as follows:

 a PT a PT 
g1: Im a13PT − 11 PT23  = 0 , where Ω = ω r = 2πf r
 a21 

g2: 10 −6 VA ≤ Pˆa ≤ Pˆmax

g3: 10 −6 A ≤ Iˆ ≤ Iˆmax

g4: LbD ≤ Dt ≤ UbD

g5: Lbh ≤ h p ≤ Ubh

g6: LbL ≤ Lb ≤ UbL

The first constraint assures that the transducer operates at the specified resonance
frequency f r . The second and third constraints limit the input power and input current. Iˆ and
P̂a are given by the equations (4.64) and (4.65) respectively. Obviously, they must be positive
quantities. Bounds of 10 −6 VA and 10 −6 A are imposed on P̂ and Iˆ to avoid division by zero.
a

The last three constraints give the region of search for the optimum. The values of LbD , UbD ,
L bh , U bh and LbL are determined according to manufactory specifications and the
requirements for the piezoelectric transducer. The value of the upper bound UbL is dynamic
78 CHAPTER 4

upper bound. It varies according to the selected material type of the end blocks. It is
determined using the search method described before.
Denoting the variable vector x = ( x1 , x2 , x3 , x4 ) = ( Dt , h p , Lb , Typb ) , the objective vector
F (x) = [ f1 (x), f 2 (x)] and the constraint vector G = [g i , i = 1,..., 6] , the two-objective
optimization problem is formulated as follows:
Minimize F (x)
(4.67)
Subject to G

4.5.3 Implementation of the Optimization Process

The optimization process can be divided into two levels. In the first level, optimization is
performed individually for each given number N of piezoelectric rings. The respective non-
dominated solutions are searched for. In the second level optimization, non-dominated
solutions are searched again in the non-dominated solutions obtained in the first level
optimization. Here only the first level optimization is performed. The second level
optimization will be performed in chapter 5. For the optimization problem, the following
parameter values are assumed:
• f r = 20 kHz, λ/2 vibration mode

• Uˆ = 100 V, Iˆmax = 4 A

• LbD = 8 mm , UbD = 50 mm

• Lbh = 0.2 mm, Ubh = 5 mm

• LbL = 5 mm

For the given Û and Iˆmax , the maximum electrical input power P̂max is determined
correspondingly. Therefore the constraint g2 can be deleted. In the following, four MOEAs,
namely MOGA, NSGA, SPEA and NSGA-II are first used to solve the constrained MOP of
the transducer with 2 piezo-rings respectively. Then the most appropriate MOEA is selected
from the four MOEAs according to their optimized results and is applied in optimization of
the transducers with 4 and 6 piezo-rings.

Two strategies have been used to handle constraints. In MOGA, NSGA and SPEA, the
constrained MOP is transferred into a non-constrained MOP by using the penalty function
approach, whereas the constrain-domination concept is applied in NSGA-II.

In the penalty function approach, the constraints g3 is normalized (converted into the “ ≥ „
form) as follows:
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 79

g 3u (x) = Iˆmax Iˆ − 1 ≥ 0
(4.68)
g 3l (x) = Iˆ 10 −6 A − 1 ≥ 0

For each solution, the constraint violation for each constraint is calculated as follows:

| g i (x) |, if g i (x) < 0

g~i (x) =  i = 3u , 3l (4.69)
0, if g i (x) ≥ 0

Then the new objective functions (penalized functions) are defined as follows:
~
f 1 (x) = f1 (x) + p1 [ g~3u (x) + g~3l (x)]
~ (4.70)
f 2 (x) = f 2 (x) + p 2 [ g~3u (x) + g~3l (x)]

The penalty parameter p1 and p 2 should be chosen so that both of f i (x) and
pi [ g~3u (x) + g~3l (x)] have the same order of magnitude, where i = 1, 2 . Many static and
dynamic strategies can be used to choose penalty parameters. Here static values of penalty
parameters are used. In order to determine penalty parameters, the optimization was first
performed for a small number of generations. The values of penalty parameters were then
determined according to the obtained values of the original objective functions and constraint
violations. Here p1 = 2 × 10 2 and p 2 = 4 × 10 4 are chosen.

The values of two penalty parameters must be chosen properly so that the algorithms work
well. If the penalty parameter is not chosen adequately, Pareto-optimal solutions may not be
found or a poor distribution of solutions may be occur. In order to avoid choosing any explicit
penalty parameter, the constrain-domination concept instead of penalty function approach is
used in NSGA-II.

Four MOEAs here used differ in the way each solution is evaluated and selected to fill the
mating pool for offspring production. The rest of the algorithms (crossover and mutation
operations) are the same. After the mating pool is filled up, a crossover operator and a
mutation operator are applied to the strings of the mating pool to create new solutions.

In each MOEA, a mixed coding scheme with a mixed crossover and a mixed mutation
operator is used. A mixed coding (chromosome) for a typical design (individual or solution)
of the symmetric transducer with 4 design variables is as follow:
2{
.0 0{1 {
15.8 37.6
{
Dt Typb hp Lb

The first variable Dt is a continuous variable, which takes a real value in the range of
8mm ≤ Dt ≤ 50mm . The second variable Typb is a discrete variable, which take a two-bit
binary string representing a set values of ρ, E and tanδ of a specified material type. The
strings and corresponding material types are shown in Table 4.4. The third variable hp is a
80 CHAPTER 4

continuous variable, which takes a value in the range of 0.2mm ≤ h p ≤ 5mm . The last
variable Lb is continuous type, which takes a value in the range of 5mm ≤ Lb ≤ UbL . The UbL
is dynamic upper bound for the variable Lb. It varies from one generation to another and its
value is determined by solving the equality constraint g1 using the secant method.

In each generation, two chromosomes are selected from the mating pool randomly. A blended
crossover (BLX-α) operation is performed on the continuous parts Dt and hp. For the discrete
part Typb a single-crossover is used. A bit-wise mutation is only performed on the discrete
part. The value of the design variable Lb is determined by solving the equality constraint g1
using the secant method. Hereto, one offspring solution consisting of 4 variable values has
been obtained. Similarly, the other offspring solution can also be obtained.

Table 4.5 Parameters and techniques that are used in the four MOEAs.
Parameters and MOEAs
techniques
MOGA NSGA SPEA NSGA-II
Constraint Penalty function Penalty Penalty function Constrain -
handling function domination
strategy
Approach to find Continuous updated Continuous Continuous Continuous
non-dominated method updated updated method updated
set method method
Fitness Pareto ranking ( the rank Pareto ranking Pareto-based Pareto ranking
assignment of a certain individual (Goldberg’s method (the (Goldberg’s
corresponds to the ranking) and fitness is ranking) and
number of individuals in the shared assigned the shared
the current population by fitness according to the fitness
which it is dominated) strength of the
and the shared fitness individual)
Niching Fitness sharing in Fitness sharing Pareto-based Fitness sharing
technique objective space in parameter method in in parameter
space objective space space
Elite – No No Yes Yes
preserving
Selection SRWS SRWS Binary SRWS and the
operation tournament crowded
selection tournament
selection
Crossover 0.9 0.9 0.9 0.9
probability
Mutation 0.5 0.5 0.5 0.5
probability
δ share 0.5 0.184 -- 0.184

Population size 50 50 50 50
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 81

Table 4.5 gives some important parameters and techniques that are used in the four MOEAs.
It is pointed out that in NSGA-II the crowded tournament selection operator here is only used
in the process of creating the new population Pt+1 of size N from the combined population Rt
of size 2N consisting of the parent population Pt and the offspring population Qt. The SRWS
operator is still used to select the solutions in Pt+1 into the mating pool.

Optimization is performed first for the transducer with 2 piezo-rings using the above four
MOEAs, respectively. Figs. 4.17 to 4.20 show the respective non-dominated solutions in
objective spaces after 300 generations. Tables 4.6 to 4.9 give the corresponding values of
design variables and objective functions for the non-dominated solutions.
35000

30000
1/Amplitude[1/m]

25000

20000

15000

20 30 40 50
Input power [VA]
Fig. 4.17 Non-dominated solutions obtained by using MOGA for the symmetrical transducer
with 2 piezo-rings

Table 4.6 Values of design variables and objective functions for non-dominated solutions
obtained by using MOGA for the symmetrical transducer with 2 piezo-rings

Design Dt [mm] Typb hp [mm] Lb [mm] uˆ a 3 [µm] P̂a [VA]

1 8. 2 0.2 63.4 33.44 16.23

2 8. 4 0.2 40.7 92.19 53.43
3 8. 4 0.3 40.5 88.09 51.01
4 8. 4 0.4 40.5 85.94 49.77
5 8. 2 0.4 63. 32.64 15.59
6 8. 4 0.4 40.4 83.22 48.2
7 8. 4 0.5 40.2 80.15 46.42
8 8. 4 0.5 40.2 79.82 46.24
9 8. 4 0.6 40.1 78.12 45.26
10 8. 4 0.6 40.1 77.57 44.95
11 8. 2 0.6 62.7 32.14 15.33
12 8. 4 0.7 40. 75.29 43.63
82 CHAPTER 4

Table 4.6 continued

Design Dt [mm] Typb hp [mm] Lb [mm] uˆ a 3 [µm] P̂a [VA]

13 8. 4 0.7 40. 75.14 43.55

14 8. 2 0.7 62.6 31.91 15.22
15 8. 2 0.7 62.5 31.85 15.19
16 8. 2 1. 62. 30.94 14.76
17 8. 4 1.5 39. 58.06 33.68
18 8. 2 1.4 61.4 29.91 14.27
19 8. 4 1.2 39.4 64.41 37.42
20 8. 2 1.9 60.6 28.61 13.66
21 8. 2 1.7 60.9 29. 13.85
22 8. 4 1.7 38.8 55.3 32.14
23 8. 4 1.9 38.5 52.84 30.71
24 8. 4 2.2 38.1 49.09 28.51
25 8. 4 1.8 38.6 53.15 30.9
26 8. 4 2.2 38.1 48.48 28.17
27 8. 4 2. 38.3 50.68 29.49
28 8. 4 2.2 38.1 48.38 28.14
29 8. 4 2.6 37.5 44.1 25.65
30 8. 4 3. 37.1 41.18 23.96
31 8. 4 2.9 37.2 41.77 24.31
32 8.1 4 3.6 36.3 36.54 21.25
33 8.1 4 3.7 36.1 35.83 20.84
34 8.1 4 3.6 36.3 36.5 21.24

40000
1/Amplitude[1/m]

35000

30000

25000

20000

15000

10 20 30 40 50
Input Power[VA]

Fig. 4.18 Non-dominated solutions obtained by using NSGA for the symmetrical transducer
with 2 piezo-rings
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 83

Table 4.7 Values of design variables and objective functions for non-dominated solutions
obtained by using NSGA for the symmetrical transducer with 2 piezo-rings

Design Dt [mm] Typb hp [mm] Lb [mm] uˆ a 3 [µm] P̂a [VA]

1 8. 2 0.2 63.4 33.44 16.23
2 8. 2 4.4 56.5 23.54 11.13
3 8. 4 0.2 40.6 91.49 53.01
4 8. 4 1. 39.7 69.06 39.93
5 8. 4 1.1 39.6 66.6 38.49
6 8. 4 1.3 39.3 62.65 36.2
7 8. 4 3.2 36.8 39.13 22.48
8 8. 2 1.6 61.1 29.44 13.98
9 8. 2 2.7 59.3 26.73 12.68
10 8. 4 0.4 40.4 85.52 49.81
11 8. 2 2.1 60.3 28.09 13.42
12 8. 2 2.7 59.4 26.84 12.84
13 8.1 2 0.9 62.3 31.35 15.1
14 8.1 4 1.1 39.4 64.75 37.99
15 8.1 2 1. 62. 30.92 14.93
16 8.1 4 2.8 37.3 42.4 24.89
17 8.1 4 2.1 38.2 49.63 29.27
18 8.1 4 0.2 40.7 92.19 54.76
19 8.1 4 0.9 39.7 69.27 41.07
20 8.1 4 2.4 37.9 46.75 27.76
21 8.1 2 1.5 61.2 29.58 14.55
22 8.2 4 0.7 40.1 76.67 46.32
23 8.3 2 0.2 63.4 33.44 17.53
24 8.4 4 0.2 40.7 92.19 58.38

45000
1/Amplitude[1/m]

40000
35000
30000

25000
20000
15000

10 20 30 40 50
Inputpower[VA]
Fig. 4.19 Non-dominated solutions obtained by using SPEA for the symmetrical transducer
with 2 piezo-rings
84 CHAPTER 4

Table 4.8 Values of design variables and objective functions for non-dominated solutions
obtained by using SPEA for the symmetrical transducer with 2 piezo-rings

Design Dt [mm] Typb hp [mm] Lb [mm] uˆ a 3 [µm] P̂a [VA]

1 8. 2 0.3 63.2 33.12 15.89

2 8. 2 0.7 62.6 31.92 15.19
3 8. 2 1.4 61.5 29.94 14.22
4 8. 2 1.6 61. 29.26 13.89
5 8. 2 5. 55.5 22.58 10.66
6 8. 4 0.2 40.7 92.19 53.43
7 8. 4 1.5 39. 58.08 33.55
8 8. 2 0.8 62.4 31.55 15.
9 8. 4 3.8 36.1 35.39 20.28
10 8. 4 0.3 40.5 87.19 50.47
11 8. 4 4. 35.8 34.09 19.52
12 8. 4 0.7 40.1 76.27 44.11
13 8. 4 4. 35.8 34.21 19.59
14 8. 2 4.7 56. 23.05 10.89
15 8. 2 1. 62.1 31.03 14.75
16 8. 2 1.8 60.8 28.9 13.72
17 8. 2 3.7 57.6 24.68 11.68
18 8. 4 3.7 36.1 35.81 20.53
19 8. 2 4.2 56.8 23.8 11.26
20 8. 4 2.5 37.7 45.56 26.25
21 8. 4 2.1 38.3 50.11 28.9
22 8. 2 3.1 58.6 25.92 12.28
23 8. 4 1.8 38.6 53.98 31.16
24 8. 4 2.7 37.5 43.76 25.2
25 8. 4 2.2 38.1 49.01 28.26
26 8. 2 0.9 62.2 31.13 14.8
27 8. 4 0.7 40. 75.12 43.45
28 8. 4 2.6 37.6 44.7 25.75
29 8. 4 2.7 37.5 43.84 25.25
30 8. 4 2. 38.3 50.9 29.37
31 8. 2 2.4 59.8 27.5 13.04
32 8. 2 2.4 59.7 27.31 12.95
33 8. 4 3.3 36.7 38.89 22.35
34 8. 2 3. 58.8 26.17 12.4
35 8. 2 2.3 59.9 27.59 13.09
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 85

45000

1/Amplitude[1/m]
40000

35000

30000

25000

20000

15000

10 20 30 40 50
Input power [VA]

Fig. 4.20 Non-dominated solutions obtained by using NSGA-II for symmetrical transducer
with 2 piezo-rings

Table 4.9 Values of design variables and objective functions for non-dominated solutions
obtained by using NSGA-II for the symmetrical transducer with 2 piezo-rings

Design Dt [mm] Typb hp [mm] Lb [mm] uˆ a 3 [µm] P̂a [VA]

1 8. 2 0.2 63.4 33.44 16.23

2 8. 2 1.4 61.4 29.82 14.16
3 8. 2 3. 58.7 26.05 12.34
4 8. 4 0.3 40.5 86.43 50.02
5 8. 4 3.6 36.4 37.04 21.26
6 8. 2 0.8 62.4 31.61 15.03
7 8. 2 1.8 60.8 28.84 13.69
8 8. 2 1.9 60.7 28.69 13.62
9 8. 2 2. 60.5 28.43 13.49
10 8. 2 3.6 57.9 25.01 11.84
11 8. 2 4. 57.1 24.2 11.45
12 8. 4 0.2 40.7 92.19 53.43
13 8. 4 0.8 39.8 71.56 41.37
14 8. 4 1.3 39.3 61.75 35.68
15 8. 4 1.6 38.8 56.32 32.52
16 8. 4 1.7 38.7 55.15 31.84
17 8. 4 2. 38.3 50.89 29.36
18 8. 4 2.9 37.2 41.77 24.03
19 8. 4 3.1 37. 40.54 23.31
20 8. 4 4. 35.8 34.43 19.72
21 8. 4 4. 35.8 34.17 19.57
22 8. 2 1.7 60.9 29.07 13.8
23 8. 2 1.7 60.8 28.97 13.75
24 8. 2 3. 58.7 26.02 12.33
25 8. 2 3.1 58.7 26.01 12.33
86 CHAPTER 4

Table 4.9 continued

Design Dt [mm] Typb hp [mm] Lb [mm] uˆ a 3 [µm] P̂a [VA]

26 8. 2 4.1 57.1 24.12 11.41

27 8. 2 4.1 57. 24.1 11.4
28 8. 2 4.2 56.8 23.83 11.27
29 8. 2 5. 55.5 22.58 10.66
30 8. 4 4. 35.8 34.26 19.62
31 8. 2 3.7 57.6 24.7 11.69

In the following, the above four MOEAs will be evaluated in terms of the relative
convergence of the non-dominated solutions, the diversity of the non-dominated solutions, the
computing cost and the ability of finding non-dominated solutions. Then the most appropriate
MOEA will be selected and used to solve the optimization problems of the transducer with 4
and 6 piezo-rings.
In order to compare the optimized results visually, the non-dominated solutions obtained
above are plotted again according to the same axes (see Fig. 4.21). Based on Fig. 4.21,
however, it is difficult to identify which algorithm is better in terms of the convergence of
non-dominated solutions. Therefore, the set coverage metric C (A, B) has been applied, see
the definition of C (A, B) in chapter 3. Table 4.10 gives the calculated values of the set
coverage metric C (A, B), where A represents a non-dominated set obtained by using a
MOEA in the column and B represents a non-dominated set obtained by using a MOEA in the
row correspondingly.
Since no member of the non-dominated set NSGA-ΙΙ is dominated by the members of the
others, the set NSGA-ΙΙ has a better convergence to the Pareto-optimal front than the rest
relatively. Compared with SPEA, however, NSGA-ΙΙ has not an obvious advantage in terms
of the convergence because the metric C(NSGA-ΙΙ, SPEA) is very small.

45000

40000
1/Amplitude[1/m]

35000 NSGA

30000
NSGA2
25000
MOGA
20000
SPEA
15000

10 20 30 40 50
Input Power [VA]

Fig. 4.21 Non-dominated solutions in objective space obtained by using four MOEAs after
300 generations for the symmetrical transducer with 2 piezo-rings
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 87

Table 4.10 The values of metrics C(A,B) for non-dominated solutions obtained by four
MOEAs after 300 generations for the symmetrical transducer with 2 piezo-rings
A B NSGA NSGA-II MOGA SPEA

NSGA - 0 0 0

NSGA-II 4/24 - 4/34 1/35

MOGA 6/24 0 - 0

SPEA 4/24 0 5/34 -

Table 4.11 Values of metrics SP and DI and the computation cost for non-dominated
solutions obtained by four MOEAs after 300 generations for the symmetrical transducer with
2 piezo-rings

Metric NSGA NSGA-II MOGA SPEA

SP 1095 683 398 479

DI 31637 33434 24106 33434

Computation time
1714 14656 5797 5146
[s]
Number of the non-
24 31 34 35
dominated solutions

In order to compare the generated solutions in terms of the diversity, the metric SP and DI,
which evaluate the uniformity of distribution of solutions and the extent of spread of solutions
respectively, are applied. Table 4.11 gives the calculated values of the metrics. According to
the definitions of SP and DI, the more uniformly the solutions are distributed, the smaller the
metric SP; the larger the value of the metric DI is, the larger the extent of spread of the
solutions. It is obvious that the algorithms MOGA and SPEA perform better than the
algorithms NAGA and NSGA-ΙΙ in terms of the uniformity of distribution of solutions in
objective space, whereas the non-dominated set obtained by NSGA and SPEA has the largest
extent of spread relatively.

The reason that the algorithms MOGA and SPEA perform better than the algorithms NAGA
and NSGA-ΙΙ in terms of the uniformity of distribution of solutions in objective space is that
the former performs the niching strategy in the objective variable space, whereas the latter
performs the niching strategy in the design variable (parameter) space. The better diversity of
the solutions in objective space is expected if the niching strategy is performed in objective
space. In general, a good diversity of the solutions in objective space may not result in a good
88 CHAPTER 4

diversity of the solutions in design variable space, and vice versa. The choice between
decision-space niching or objective-space niching depends on what is desired in the obtained
non-dominated set. If diversity in decision variable values is emphasized, decision-space
niching should be performed. On the other hand, if diversity in objective function values is
emphasized, objective-space niching should be performed.

Table 4.11 also shows the computation time and the number of the non-dominated solutions
after 300 generations using the four MOEAs. Obviously, the computation cost of the NSGA is
the lowest and the computation cost of NSGA-ΙΙ is the largest. Although the short
computation time is always expected, the closeness to the Pareto-optimal front and the
diversity of non-dominated solutions are more important in the optimal design of transducers.
In terms of finding non-dominated solutions, SPEA shows a better performance.

According to above results of the comparison, it is concluded that SPEA show the best overall
performance for the two-objective optimization problem of the symmetrical transducer.
Therefore, it has been applied for optimization of the symmetrical transducer with 4 and 6
piezoelectric rings, respectively. Figs. 4.22 and 4.23 show the respective non-dominated
solutions after 300 generations for the transducer with 2 and 4 piezo-rings. Tables 4.12 and
4.13 give the values of the design variables and objective functions for the non-dominated
solutions.

After all non-dominated solutions for the transducer with 2, 4 and 6 piezo-rings are obtained,
the second-level optimization can be performed. This will be described in chapter 5.

30000
1/Amplitude[1/m]

25000

20000

15000

10000

50 75 100 125 150 175

Input power[VA]

Fig. 4.22 Non-dominated solutions obtained by using SPEA for the symmetrical transducer
with 4 piezo-rings
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 89

Table 4.12 Values of design variables and objective functions for the non-dominated solutions
obtained by using SPEA for the symmetrical transducer with 4 piezo-rings

Design Dt [mm] Typb hp [mm] Lb [mm] uˆ a 3 [µm] P̂a [VA]

1 8. 2 5. 47.3 33.81 31.37

2 8. 2 3.9 50.8 37.88 35.49
3 8. 2 3.8 51.2 38.4 36.02
4 8. 2 3.4 52.5 40.22 37.84
5 8. 2 3.2 53.1 41.12 38.74
6 8. 2 2.3 56.1 46.22 43.75
7 8. 4 1.9 36.2 71.75 82.45
8 8. 2 1.9 57.5 49.14 46.61
9 8. 2 2.4 55.7 45.5 43.07
10 8. 4 2.1 35.7 67.2 77.11
11 8. 4 1.7 36.6 76.63 88.26
12 8. 2 2.2 56.6 47.21 44.75
13 8. 2 1.9 57.4 48.92 46.42
14 8. 4 1.9 36. 70.16 80.63
15 8. 2 1.9 57.5 49.14 46.64
16 8. 2 1.7 58. 50.28 47.75
17 8. 2 2. 57.1 48.27 45.8
18 8. 4 1.6 36.8 78.64 90.66
19 8. 4 0.8 39. 116.51 135.01
20 8. 4 1.7 36.5 75.22 86.64
21 8. 4 0.3 40.1 153.58 178.24
22 8. 4 0.4 40. 147.79 171.49
23 8. 2 1.1 60.2 56.17 53.48
24 8. 2 1. 60.5 56.83 54.11
25 8. 2 0.9 60.6 57.22 54.5
26 8. 4 0.9 38.5 105.04 121.66
27 8. 2 0.2 63. 65.57 62.85
28 8. 2 0.2 63. 65.41 62.67
29 8. 4 0.3 40.2 159.7 185.41
30 8. 4 0.8 38.8 112.31 130.16
31 8. 2 0.2 62.9 65.07 62.27
32 8. 4 0.2 40.3 163.97 190.41
33 8. 2 1.8 57.8 49.83 47.35
34 8. 2 0.7 61.2 59.22 56.44
35 8. 4 1.8 36.4 74.59 85.95
36 8. 4 1.7 36.5 74.81 86.21
37 8. 2 1.6 58.4 51.25 48.75
38 8. 2 0.4 62.4 63.22 60.37
90 CHAPTER 4

24000
22000

1/Amplitude[1/m]
20000
18000
16000
14000
12000
10000
60 80 100 120 140 160
Input power [VA]

Fig. 4.23 Non-dominated solutions obtained by using SPEA for the symmetrical transducer
with 6 piezo-rings

Table 4.13 Values of design variables and objective functions for the non-dominated solutions
obtained by using SPEA for the symmetrical transducer with 6 piezo-rings

Design Dt [mm] Typb hp [mm] Lb [mm] uˆ a 3 [µm] P̂a [VA]

1 8. 2 5. 39. 40.75 55.03

2 8. 2 5. 39.1 40.91 55.28
3 8. 2 4.1 43.4 45.44 62.48
4 8. 2 3.9 44.7 47.09 65.04
5 8. 2 4.3 42.6 44.5 61.02
6 8. 2 4.1 43.4 45.5 62.58
7 8. 2 4.1 43.6 45.73 62.94
8 8. 2 3.7 45.5 48.17 66.72
9 8. 2 3.6 45.9 48.71 67.56
10 8. 2 3.6 46.2 49.09 68.15
11 8. 2 3.5 46.7 49.84 69.29
12 8. 2 3.6 45.7 48.45 67.16
13 8. 2 3.6 45.8 48.6 67.39
14 8. 2 3.1 48.2 52.15 72.83
15 8. 2 2.6 50.7 56.58 79.5
16 8. 2 2.9 49.6 54.54 76.45
17 8. 2 2.8 49.7 54.74 76.74
18 8. 2 1. 59. 79.09 112.57
19 8. 2 2.4 51.9 58.92 83.02
20 8. 2 2.1 53.3 62.19 87.89
21 8. 2 1.2 57.7 74.58 106.06
22 8. 4 1.4 35.7 101.1 173.83
23 8. 2 1.2 57.9 75.1 106.85
24 8. 2 2. 53.9 63.43 89.76
25 8. 2 0.9 59.2 80.14 114.16
26 8. 2 2. 54. 63.78 90.28
27 8. 2 1.3 57.5 73.72 104.85
28 8. 2 1.1 58.2 76.32 108.64
29 8. 2 0.6 61. 87.47 124.82
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 91

4.6 Optimization of Langevin-type Transducers with a Stepped Horn

In ultrasonic machining and bonding, it is generally required that piezoelectric transducers

have a working frequency between 10 kHz and 100 kHz and in the meantime have a large
vibration amplitude (10 to 50 micros at 20 kHz). It is difficult to achieve such large vibration
amplitude using the transducer configurations described in sections 4.5. In order to magnify
the vibration amplitude, the front block is usually designed as a linear (cone), exponential or
stepped acoustic horn (sonotrode). Fig. 4.24 shows the three mostly used forms of horns.
Here, a stepped-horn transducer is studied.

Fig. 4.25(a) shows a typical configuration of a Langevin-type transducer with a stepped horn.
It has been shown that the impedance transformation ration in the case of both Lf1 and Lf2
being λ/4 long is given by (A f 1 A f 2 ) , and the amplitude transformation ratio is given by
2

(A f1 A f 2 ) . Af1 and Af2 are the profile areas of the input side and output side of the horn. The
larger the ratio is, the larger is the vibration amplitude at the end of the horn [BPF91]. As seen
before, in addition to the vibration amplitude there exist other important performance criteria.
Therefore, multiple design goals should be considered.

4.6.1 Derivation of the Whole Transfer Matrix

The whole transducer is considered as a connection of a piezoelectric stack element and three
mechanical elements in series, where the stepped horn consists of two mechanical elements.
Fig. 4.25(b) gives four-pole network element description. Similarly, applying equation (4.94)
and the transfer matrix relations of the fundamental connection 1, 3 and 4 shown in Table 4.1,
the whole transfer matrix relation for this transducer can be written as follows:
 Xa4   A Hm
w A Hem
w
  Xe1 
 Iˆ  =  Hme  ˆ  (4.71)
  A w A He
w 
U 
where
 vˆ   vˆ 
X e1 =  e1  , Xa4 =  a4 
ˆ
ˆ
 F e1  F a4 

Hm m m PSm ma11H a12H 

A w =A A A
4 3 2 A = H
1 H
a21 a22 
a H 
A wHem = A 4m A 3m A 2PSem =  13H 
a23 

[
A wHme = A 2PSme A1m = a 31H a 32H ]
A wHe = A PSe = a33H
92 CHAPTER 4

Fig. 4.24 Shape (-), vibration amplitude (-) and strain (--) on three horns: a) exponential
horn, b) cone horn, c) stepped horn

Back section Piezo-stack Horn

(a) Dt Dt 2

Lb N hp Lf1 Lf2

ve1 va1 ve2 va2 ve3 va3 ve4 va4

(b) Fe1 1 Fa1 Fe2 2 Fa2 Fe3 3 Fa3 Fe4 4 Fa4

I U

Fig. 4.25 Langevin transducer with a stepped horn (a) typical configuration (b) four-pole
element descriptions

Similarly, the detailed expressions of the matrix elements aijH (i, j = 1, 2, 3) are not given here
because the expressions are very complex. They are derived automatically in the programs.

4.6.2 Problem Formulation of Transducers without Loads

In this section, the objective functions, design variables and constraints are defined. The
multi-objective optimization for the transducer at free vibration is discussed.

Optimization objectives Similarly, the free vibration amplitude and input electrical power at
the resonance operation are considered as objective functions. For a given exciting voltage,
the former should be maximized and the latter should be minimized. By analogy with
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 93

equation (4.63), the first optimization objective, which represents the inverse of the vibration
amplitude at resonance, is formulated as follows
H
a 21 Ω
f1 = (4.72)
a13 a 21 − a11 a 23 Uˆ
H H H H

According to equations (4.64) and (4.65) the expressions of the input current Iˆ and apparent
power P̂a for this problem are given by

aH aH
Iˆ = a33H − 31 H 23 Uˆ (4.73)
a 21

1 Uˆ 2 H a31H a 23
H
Pˆa = IˆUˆ = a33 − PT (4.74)
2 2 a 21

Equation (4.74) gives the expression of the second objective function. It is given by

f 2 = Pˆa (4.75)

Design variables The parameters which affect the performances of the transducer with a
stepped horn are shown in Table 4.14. Similarly, there are two types of parameters. N and
material are the discrete type. The others are continuous type. The inner diameter of piezo-
rings and the dimensions of the bolt will not be considered as design variables individually. It
is also assumed that the piezoelectric transducer has been pre-stressed optimally. In this
problem, the following parameters are then considered as design variables:

The dimensional parameters: Dt, Dt2 ( ζ= Dt2 / Dt ), Lb, hp, Lf1, Lf2

The material type: Typb (ρb, Eb , tgδb) and Typf (ρf, Ef , tgδf)

The rest is considered as given quantities. N can take one of the elements of the set {2, 4, 6}.

Constraints On the analogy of equation (4.61), the resonance condition for the transducer
with a stepped horn free at both sides is given by

 H a11H a 23
H

Im a13 − H  = 0 (4.76)
 a 21 

The equation (4.76) describes an equality constraint for the design variables. Similarly, the
secant method will be used in searching the solutions to equation (4.76). As there are multi-
mode solutions to equation (4.76) in the specified variable ranges, the bounds of the variables
and two starting points must be picked carefully using similar method described before so that
only the solutions for the λ/2 mode are found.
94 CHAPTER 4

Table 4.14 Parameters that affect the performances of the transducer

Û : Input voltage
Lb: Length of back blocks
Lf1: Length of the first part of horns
Lf2: Length of the second part of horns
Dt: Outer diameter of back blocks, input end of horns and piezo-rings
Dt2: Outer diameter of the output end of horns
hp: Thickness of piezoelectric rings
Di: Inner diameter of piezo-rings and outer diameter of bolts
Lbt: Length of bolts

N: Number of piezoelectric disks or rings

Typb: Material of back blocks
Typf: Material of horns
Typp: Material of piezo-rings

In addition to equality constraint, there exist design requirements for electrical quantities and
dimensional quantities. All constraints for this optimization problem can be concluded as
follows:
 H a11H a 23
H

g1: Im a13 − H  = 0 , where Ω = ω r = 2πf r
 a 21 

g2: 10 −6 VA ≤ Pˆa ≤ Pˆmax

g3: 10 −6 A ≤ Iˆ ≤ Iˆmax
g4: Dtl ≤ Dt ≤ Dtu
g5: ζ l ≤ ζ ≤ ζ u
g6: Lbl ≤ Lb ≤ Lbu
g7: h pl ≤ h p ≤ h ph

g8: L f 1l ≤ L f 1 ≤ L f 1h

g9: L f 2l ≤ L f 2 ≤ L f 2 h
Similarly, the first constraint assures that the transducer operates at the specified resonance
frequency f r . The second and third constraints limit the input power and input current.
Obviously, they must be positive quantities. Bounds of 10 −6 VA and 10 −6 A are imposed on P̂a
and Iˆ to avoid division by zero. The last six constraints give the region of search for the
optimum. The values of the lower bounds and upper bounds of Dt , ζ , Lb and h p as well as
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 95

the values of lower bounds of L f 1 and L f 2 are determined according to manufactory

specifications and the requirements for the piezoelectric transducer. The upper bounds L f 1h
and L f 2 h are dynamic upper bounds. They vary according to the selected material type of the
end blocks and the values of other continuous design variables. They are searched using the
secant method in the process of optimization.

Denoting the variable vector

x = ( x1 , x2 , x3 , x4 , x5 , x6 , x7 , x8 ) = ( Dt , ζ , Lb , Typb , Typ f , h p , L f 1 , L f 2 ) , the objective vector
F (x) = [ f1 (x), f 2 (x)] as well as the constraint vector G = [g i , i = 1,..., 9] , the two-objective
optimization problem can be written as follows:
Minimize F (x)
(4.77)
Subject to G

4.6.3 Implementation of Optimization

Similarly, two-level optimization is performed. In the first level, the optimized solutions for
each given number N of piezoelectric rings are searched. In the second level optimization,
Pareto-optimal solutions are searched again in all optimized solutions which obtained from
the first level optimization. For all optimization problems in the first level, the following
parameter values are assumed:
• f r = 20kHz , λ/2 vibration mode

• Uˆ = 100 V , Iˆmax = 4 A

• Dtl = 8 mm, Dtu = 50 mm, ζ l = 0.2, ζ u = 1

• Lbl = 5 mm, Lbu = 58.5 mm, h pl = 0.2mm, h pu = 5mm

• L f 1l = 5 mm, L f 2l = 2 mm

The optimization problem for 2 piezo-rings is first solved using MOGA, NSGA, NSGA-ΙΙ
and SPEA2. Then the most appropriate MOEA is selected and applied in optimization of the
transducers with 4 and 6 piezo-rings. MOGA, NSGA and NSGA-ΙΙ are the same methods as
are used for symmetrical transducers before. SPEA2 is an improved version of SPEA. Unlike
SPEA, SPEA2 uses an improved fitness assignment scheme, a nearest neighbor density
estimation technique and an enhanced archive truncation method. Similarly, the parameters
and techniques shown in Table 4.5 are also used in these methods.

The constraints are handled using the following methods. For the given Û and Iˆmax , the
upper bound P̂max of the electrical input power is determined correspondingly. Therefore the
96 CHAPTER 4

constraint g2 can be deleted. The g3 is dealt with using the penalty function approach similar
to that in section 4.5.3. The side constraints g4 to g9 are handled by directly limiting the values
of the design variables in the regions given by the respective side constraints. When any value
of the design variables goes outside the bounds, this value is replaced by the corresponding
bound. The equality constraint will be dealt with in each generation using the following
method: First, determine the upper bound Lf1u of the variable Lf1 for the λ/2 mode using the
secant method according to the obtained values of the variables ζ, Lb, Typb , Typf , hp and the
lower bound Lf2l. Replace the value of Lf1 with the corresponding bound if the value of Lf1
goes outside the range L f 1l ≤ L f 1 ≤ L f 1u . Second, determine the value of the variable Lf2 using
the secant method.

As the design variables include continuous and discrete types, a mixed coding scheme with a
mixed crossover and a mixed mutation operator are used in each MOEA. The chromosome
for a typical design is as follows:
10
{ .3 0{
.27 22
{ .6 11
{ 01
{ 2{
.9 19
{ .6 63
{ .3
Dt ζ Lb Typb Typ f hp Lf 1 Lf 2

The first variable Dt is continuous variable, which takes a real value in the range of
8 mm ≤ Dt ≤ 50 mm . The second variable ζ takes a real value in the range of 0.2 ≤ ζ ≤ 1 .
The third variable Lb takes a real value in the range of 5mm ≤ Lb ≤ 58.5 mm . The fourth and
fifth variables Typb and Typf are discrete variables, which take a two-bit binary string and
represent a set values of ρ ,E and tgδ of a specified material type, respectively. The strings
and corresponding material types are shown in Table 4.4. The sixth variable hp is a continuous
variable, which takes a value in the range of 0.2 mm ≤ h p ≤ 5 mm . The last two variables L f 1
and L f 2 are continuous type, which take a value in the rang of 5 mm ≤ L f 1 ≤ L f 1u and
2 mm ≤ L f 2 ≤ L f 2u , respectively. L f 1u and L f 2u are dynamic upper bounds. They vary from
one generation to another and their values are determined using the secant method.

In order to create offspring of the discrete parts Typb and Typf , two two-bit strings are
connected into a four-bit string and the single-crossover operator is used. For the continuous
variables the blended crossover (BLX-α) is applied. It is noted that the crossover and
mutation operations will not be applied to the continuous variables Lf2, whose value is
determined by solving the equality constraint according to other values of the variables.

The two-objective optimization problem of the stepped-horn transducer with two piezo-rings
was solved by the four MOEAs, respectively. Figs. 4.26 to 4.29 show the respective
optimization results in the objective function space after 300 generation. Tables 4.15 to 4.18
give the corresponding variable values and objective function values.
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 97

35000
1/Amplitude[1/m]
30000

25000

20000

15000

10000

5000
0 20 40 60 80 100 120
Input power [VA]

Fig. 4.26 Non-dominated solutions obtained by using MOGA for the freely vibrating stepped-
horn transducer with 2 piezo-rings

Table 4.15 Values of design variables and objective functions for non-dominated solutions
obtained by MOGA for the freely vibrating stepped-horn transducer with 2 piezo-rings

Dt Lb hp Lf1 Lf2 uˆ a 4 P̂a

Design ζ Typb Typf
[mm] [mm] [mm] [mm] [mm] [µm] [VA]

1 8. 0.2 5. 1 1 5. 68.2 63.6 44.12 0.48

2 8. 0.2 5. 1 2 5. 69.4 62.1 36.09 0.43
3 8. 0.2 5. 3 4 5. 55.5 38.7 119.46 2.43
4 8.3 0.2 6.2 1 1 4.9 60. 63.9 46.21 0.58
5 8.3 0.2 6.2 2 2 4.9 64. 62.6 26.81 0.29
6 8.4 0.21 6.6 2 3 4.9 68.2 48.8 58.02 0.89
7 8.6 0.21 7.5 1 4 4.8 54.6 38.3 133.61 4.25
8 12.7 0.26 22.1 1 4 3.5 45.7 36.2 149.76 32.63
9 15.2 0.3 31. 2 4 2.7 40.7 38.1 137.21 28.08
10 15.6 0.3 32.6 1 4 2.5 35.7 37.1 166.81 56.37
11 17.9 0.33 40.7 1 4 1.8 32. 36.6 173.56 94.16
12 19.9 0.36 48. 1 4 1.1 21.4 38.4 176.31 99.14
13 21.5 0.38 52.8 1 4 0.7 19.4 38.2 181.57 132.65
98 CHAPTER 4

25000
1/Amplitude[1/m]
20000

15000

10000

5000
0 10 20 30 40
Input power [VA]

Fig. 4.27 Non-dominated solutions obtained by using NSGA for the freely vibrating stepped-
horn transducer with 2 piezo-rings

Table 4.16 Values of design variables and objective functions for non-dominated solutions
obtained by NSGA for the freely vibrating stepped-horn transducer with 2 piezo-rings

Dt Lb hp Lf1 Lf2 uˆ a 4 P̂a

Design ζ Typb Typf
[mm] [mm] [mm] [mm] [mm] [µm] [VA]

1 8. 0.2 5. 3 4 5. 5. 42. 97.32 1.23

2 8. 0.2 5.4 3 4 4.7 5. 42. 99.28 1.28
3 8. 0.2 5. 1 2 5. 5.3 65.4 37.18 0.44
4 8. 0.2 5. 2 4 5. 5.8 42.2 79.73 0.86
5 8. 0.21 5. 4 4 5. 6.4 42. 92.32 1.19
6 8. 0.21 14.7 4 4 2.3 5.9 41.7 135.65 2.54
7 8. 0.21 5. 4 4 5. 7.4 41.9 89.19 1.14
8 8. 0.27 5. 2 3 5. 19.3 52.3 39.34 0.62
9 9.2 0.22 9.2 4 4 4.7 8.5 41.6 106.27 2.28
10 10. 0.2 35.1 3 3 0.2 5. 51.1 140.94 8.04
11 10.3 0.2 34.7 3 4 0.2 5.5 41.1 181.13 8.62
12 12.3 0.22 21.7 1 4 3.6 8.7 41.2 139.06 7.19
13 15.8 0.26 35.1 4 4 1.9 18.4 39.5 192.13 37.77
14 17.9 0.25 38.6 4 4 1.5 15.9 39.6 203.49 47.31
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 99

35000

30000
1/Amplitude[1/m] 25000

20000

15000

10000

5000
2 4 6 8 10
Input power [VA]

Fig. 4.28 Non-dominated solutions obtained by using NSGA-II for the freely vibrating
stepped-horn transducer with 2 piezo-rings

Table 4.17 Values of design variables and objective functions for non-dominated solutions by
using NSGA-II for the freely vibrating stepped-horn transducer with 2 piezo-rings

Dt Lb hp Lf1 Lf2 uˆ a 4 P̂a

Design ζ Typb Typf
[mm] [mm] [mm] [mm] [mm] [µm] [VA]
1 8. 0.2 5. 2 4 5. 8.1 42. 73.74 0.7
2 8. 0.2 5. 3 3 5. 46.5 50.3 49.71 0.51
3 8. 0.2 5. 3 2 5. 28.2 64. 28.51 0.28
4 8. 0.2 5. 1 4 5. 9.7 41.7 83.12 0.89
5 8. 0.2 5. 1 2 5. 21.5 64.4 28.98 0.29
6 8. 0.2 7.1 3 4 3.5 5. 42.1 107.97 1.51
7 8. 0.2 5.6 2 4 4.9 41.9 40.3 62.98 0.51
8 8. 0.2 5.7 2 2 4.9 14.6 64.9 27.43 0.26
9 8. 0.2 10.2 3 4 3.5 9. 41.6 110.07 1.55
10 8. 0.2 6.2 3 4 4.9 25. 40.9 78.11 0.79
11 8. 0.21 17.7 4 4 3.7 23. 40.6 139.87 2.73
12 8. 0.2 8.4 2 3 4.6 37.2 50.7 42.33 0.39
13 8. 0.21 27.5 4 4 2.7 21.4 40.2 184.81 5.11
14 8. 0.21 7.4 3 4 5. 14.5 41.3 90.83 1.12
15 8. 0.2 15.7 2 4 3.9 13. 41.4 94.67 1.19
16 8. 0.2 10.8 2 4 4.4 13.2 41.5 81.04 0.88
17 8. 0.2 11.8 2 4 4.1 15.4 41.4 80.1 0.84
18 8. 0.2 13.6 3 4 3.5 10.7 41.3 120.76 1.91
19 8. 0.2 14.6 1 4 3.2 8.2 41.5 124.03 2.01
20 8. 0.2 16.3 1 4 2.4 5. 41.7 138.67 2.48
21 8. 0.21 14.5 2 4 4. 18.7 41.2 85.14 0.97
100 CHAPTER 4

Table 4.17 continued

Dt Lb hp Lf1 Lf2 uˆ a 4 P̂a

Design ζ Typb Typf
[mm] [mm] [mm] [mm] [mm] [µm] [VA]
22 8.1 0.21 20.7 2 4 3.3 19.7 41.1 100.77 1.38
23 8.1 0.21 17.8 1 4 3.7 12. 41.1 128.38 2.27
24 8.1 0.21 36.7 1 4 1.6 12.4 40.8 170.71 4.12
25 8.1 0.21 30.4 3 4 2.6 16. 40.5 171.48 4.42
26 8.1 0.21 37.9 4 4 1.6 15. 40.1 210.13 6.83
27 8.1 0.21 37. 2 4 1.5 5.8 41.4 151.31 3.35
28 8.1 0.22 55.9 3 4 0.4 16.3 39.6 245.08 10.15
29 8.1 0.21 45.3 2 4 1. 13.1 40.8 158.23 3.93
30 8.2 0.21 43.3 1 4 0.8 5. 41. 180.33 4.67
31 8.2 0.21 41.8 1 4 1. 6.5 41. 177.65 4.52
32 8.2 0.22 53.2 4 4 0.2 8.8 39.6 232.59 9.26
33 8.2 0.22 53.2 4 4 0.2 7.8 39.7 225.94 8.87
34 8.2 0.21 44.3 3 4 0.7 5. 40.8 188.06 5.18
1/Amplitude[1/m]

80000

60000

40000

20000

2 4 6 8 10
Input power[VA]

Fig. 4.29 Non-dominated solutions obtained by using SPEA2 for the freely vibrating stepped-
horn transducer with 2 piezo-rings

Table 4.18 Values of design variables and objective functions for non-dominated solutions
obtained by SPEA2 for the freely vibrating stepped-horn transducer with 2 piezo-rings

Dt Lb hp Lf1 Lf2 uˆ a 4 P̂a

Design ζ Typb Typf
[mm] [mm] [mm] [mm] [mm] [µm] [VA]

1 0.2 58.1 1 4 0.6 5.3 40.7 192.92 4.8

2 8. 0.25 52.9 4 4 0.2 9.1 39.2 224.91 11.1
3 8. 0.7 5. 2 1 2.7 47.3 68.8 14.32 0.88
4 8. 0.72 5. 2 2 2.6 48.9 64. 10.4 0.65
5 8. 0.64 5.3 2 1 3.2 42.3 69.7 17.35 1.02
6 8. 0.67 5. 2 2 3.1 44.6 65.1 11.84 0.68
8.
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 101

Table 4.18 continued

Dt Lb hp Lf1 Lf2 uˆ a 4 P̂a

Design ζ Typb Typf
[mm] [mm] [mm] [mm] [mm] [µm] [VA]
7 8. 0.65 5.1 2 1 3.1 42.7 69.8 16.56 0.95
8 8. 0.65 5. 2 1 3.1 42.9 69.8 16.27 0.93
9 8. 0.59 6.2 2 3 3.4 38. 51.7 23.02 1.21
10 8. 0.58 6.4 2 3 3.5 37.5 51.8 23.78 1.26
11 8. 0.22 55.7 4 4 0.2 7.1 39.4 222.9 8.98
12 8. 0.23 55. 4 4 0.2 7.6 39.3 223.51 9.4
13 8. 0.2 58.5 2 4 0.4 5. 40.8 175.81 4.39
14 8. 0.2 58.2 1 4 1.2 5.2 40.7 190.54 4.7
15 8.1 0.2 58.2 4 4 1.2 5.2 39.4 213.8 6.8
16 8.1 0.2 58.5 2 2 2.7 5. 63.3 65.41 1.34
17 8.1 0.2 58.5 4 2 2.7 5. 61.5 82.98 2.11

Similarly, the four MOEAs are evaluated in order to select the most appropriate MOEA,
which will be used to solve the optimization problems of the transducer with 4 and 6 piezo-
rings. Fig. 4.30 shows non-dominated solutions in objective space obtained above. In order to
evaluate their convergence and diversity, the performance metric C(A, B), SP and DI have
been calculated. The corresponding results are shown in Table 4.19 and Table 4.20.

According to Table 4.19, it is obvious that the set NSGA-ΙΙ has the best convergence to the
Pareto-optimal front than the rest relatively. According to the values of SP and DI shown in
Table 4.20, NSGA and NSGA-II distribute do better than the others in terms of the uniformity
of distribution of solutions, whereas the set of solutions obtained by SPEA2 has the largest
spread.

Table 4.20 also shows the computation time and the number of the non-dominated solutions
after 300 generations using the four MOEAs. Obviously, the computation cost of the NSGA is
the lowest and the computation cost of SPEA2 is the largest. In terms of finding non-
dominated solutions, NSGA-II shows the best performance.

Considering that NSGA-II shows the obvious advantage in terms of the convergence of
solutions and the ability of finding non-dominated solutions in this optimization problem, it
has been applied for optimization of the stepped horn transducer with 4 and 6 piezoelectric
rings, respectively. Figs. 4.31 and 4.32 show the respective non-dominated solutions after 300
generations for 4 and 6 piezo-rings. Table 4.21 and Table 4.22 give the values of the design
variables and objective functions for the non-dominated solutions.

After all non-dominated solutions for the transducer with 2, 4 and 6 piezo-rings are obtained,
the second-level optimization can be performed. The results will be described in chapter 5.
102 CHAPTER 4

1/Amplitude[1/m]
80000
NSGA
60000
NSGA2

40000
MOGA

20000 SPEA2

0 20 40 60 80 100 120
Input Power[VA]

Fig.4.30 Non-dominated solutions in objective space obtained by using four MOEAs for the
freely vibrating stepped-horn transducer with 2 piezo-rings

Table 4.19 Values of metrics C (A, B) for the freely vibrating stepped-horn transducer without
2 piezo-rings

A B NSGA NSGA-II MOGA SPEA2

NSGA - 0 8/13 10/17

NSGA-II 11/14 - 12/13 13/17

MOGA 1/14 0 - 8/17

SPEA2 5/14 4/34 6/13 -

Table 4.20 Values of metrics SP and DI and the computation cost for the freely vibrating
stepped-horn transducer with 2 piezo-rings

Metric NSGA NSGA-II MOGA SPEA2

SP 487 857 2734 3849

DI 21982 32372 31793 91739

Computation time [s] 4290 23373 17393 52200
Number of the non-
14 34 13 17
dominated solutions
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 103

14000

1/Amplitude[1/m] 12000

10000

8000

6000

4000

2 4 6 8 10 12 14
Input power [VA]

Fig. 4.31 Non-dominated solutions obtained by using NSGA-II for the freely vibrating
stepped-horn transducer with 4 piezo-rings

Table 4.21 Values of design variables and objective functions for the non-dominated solutions
obtained by NSGA-II for the freely vibrating stepped-horn transducer with 4 piezo-rings

Dt Lb hp Lf1 Lf2 uˆ a 4 P̂a

Design ζ Typb Typf
[mm] [mm] [mm] [mm] [mm] [µm] [VA]

1 8. 0.2 12.4 1 4 2.5 6. 41.5 242.77 7.77

2 8. 0.2 10.2 4 2 3.2 5.9 64.4 91.74 2.57
3 8. 0.2 5. 4 2 4.1 5.9 64.6 77.01 1.82
4 8. 0.2 8.5 2 2 3.8 5.6 64.8 75.07 1.75
5 8. 0.2 21.5 2 4 1. 5.8 41.9 251.75 8.18
6 8. 0.2 5. 2 2 3.2 6.2 65.2 65.81 1.37
7 8. 0.2 5. 1 4 3.3 6.1 41.7 188.67 4.65
8 8. 0.2 9.4 3 2 4.3 5.3 64.2 90.93 2.53
9 8. 0.2 11.5 2 4 3.2 5.7 41.7 203.43 5.48
10 8. 0.2 13.1 3 2 3. 5.6 64.3 98.5 2.95
11 8. 0.2 14.1 2 2 3. 5.4 64.8 84.6 2.21
12 8. 0.2 15. 2 4 2.8 5.4 41.7 220.03 6.44
13 8. 0.2 8.8 3 2 4.1 5.4 64.3 89.05 2.43
14 8. 0.2 11.3 1 2 3.8 5.3 64.3 92.88 2.64
15 8. 0.2 17.1 2 4 2.7 5.5 41.6 228.54 6.95
16 8. 0.2 18.7 3 2 1.9 5.5 64.3 109.33 3.62
17 8. 0.2 13.1 2 2 2.6 5.7 65. 82.4 2.11
18 8. 0.2 14.3 3 4 3. 5.4 41.3 259.74 9.22
19 8. 0.2 16.8 2 4 3. 5.6 41.5 226.91 6.9
20 8. 0.2 17. 2 2 2.8 5.3 64.7 89.27 2.45
21 8. 0.2 8.6 3 4 3.1 6.1 41.5 222.46 6.55
22 8. 0.2 24.8 4 2 1.6 5.2 63.9 120.12 4.35
23 8. 0.2 11. 4 2 3.8 5.2 64.2 94.65 2.74
24 8. 0.2 9.9 1 2 3. 6. 64.6 88.45 2.38
25 8. 0.2 8.9 2 4 3. 5.8 41.8 188.88 4.66
104 CHAPTER 4

Table 4.21 continued

Dt Lb hp Lf1 Lf2 uˆ a 4 P̂a

Design ζ Typb Typf
[mm] [mm] [mm] [mm] [mm] [µm] [VA]

26 8. 0.2 18.5 1 2 2.7 5.4 64.1 105.75 3.38

27 8. 0.2 26.6 4 4 1. 5.2 41.1 334.68 14.33
28 8. 0.2 26.5 4 4 1.1 5.2 41.1 332.49 14.26
29 8. 0.2 13.8 1 2 3.1 5.4 64.3 97.66 2.89
30 8. 0.2 15.2 1 2 2.6 5.8 64.3 100.06 3.03
31 8. 0.2 5.3 3 2 3.5 5.8 64.7 77.53 1.84
32 8. 0.2 12.4 2 4 3.4 5.2 41.6 210.05 5.89
33 8. 0.2 21.9 1 2 1.6 5.3 64.3 111.42 3.76
34 8. 0.2 13.6 1 2 3.3 5.4 64.2 97.27 2.87
35 8. 0.2 22.2 3 2 1.6 5.3 64.2 114.75 3.97
36 8. 0.2 12.7 2 4 3.6 5.4 41.5 211.34 5.99
37 8. 0.2 20.6 4 4 1.8 5.2 41.2 301.42 11.96
38 8. 0.2 13.6 3 2 3.2 5.2 64.2 99.86 3.02
39 8. 0.2 5.8 1 2 4.7 5.1 64.4 80.15 1.96
40 8. 0.2 5. 1 4 5. 5.2 41.4 199.08 5.47

12000
1/Amplitude[1/m]

10000

8000

6000

4000

2000
10 20 30 40 50
Input power [VA]

Fig.4.32 Non-dominated solutions obtained by using NSGA-II for the freely vibrating
stepped-horn transducer with 6 piezo-rings
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 105

Table 4.22 Values of design variables and objective functions for the non-dominated solutions
obtained by using NSGA-II for the freely vibrating stepped-horn transducer with 6 piezo-rings

Dt Lb hp Lf1 Lf2 uˆ e 4 P̂a

Design ζ Typb Typf
[mm] [mm] [mm] [mm] [mm] [µm] [VA]
1 8. 0.2 7.9 4 4 1.2 48.8 39.6 340.97 16.15
2 8. 0.2 50. 1 4 0.2 5.6 40.9 561.83 39.66
3 8. 0.2 20.4 4 4 0.7 21.8 40.7 441.49 24.83
4 8. 0.2 25. 1 4 0.2 5. 41.5 487.2 31.15
5 8. 0.2 18.6 3 4 0.8 24.7 40.6 412.6 21.82
6 8. 0.2 58.5 4 4 0.2 5. 39.6 648.77 57.59
7 8. 0.2 5. 2 4 1.3 42.4 40.4 153.78 3.36
8 8. 0.2 18.6 3 4 0.8 23.9 40.7 406.38 21.17
9 8. 0.2 5. 2 3 1.5 49.7 50.3 102.15 2.52
10 8. 0.2 5. 2 3 1.4 43.7 50.6 93.85 2.32
11 8. 0.2 17. 1 4 1. 27.8 40.6 380.69 18.72
12 8. 0.2 15.8 1 4 1.1 29.8 40.5 372.16 17.95
13 8. 0.2 16.8 2 4 1. 29.4 40.8 260.58 8.71
14 8. 0.2 5. 2 1 1.3 40.3 65. 81.81 2.23
15 8. 0.2 18.8 2 4 0.7 16.8 41.3 277.37 10.15
16 8. 0.2 5. 2 3 1.2 35.3 51. 87.8 2.3
17 8. 0.2 11.7 1 4 1. 27.9 40.8 289.66 10.78
18 8. 0.2 15.6 2 4 0.9 24.1 41. 239.88 7.52
19 8. 0.2 5.6 2 3 1.5 47.3 50.4 108.23 2.72
20 8.1 0.2 6.6 2 4 1.3 39.4 40.5 168.02 3.9
21 8.1 0.2 29.2 1 4 1. 20.6 40.5 499.12 33.44
22 8.1 0.2 26.2 3 4 1. 22.8 40.3 521.78 37.77
23 8.1 0.2 5. 4 4 1.3 31.3 40.8 191.9 4.9
24 8.1 0.2 5.2 2 4 1. 21.5 41.4 134.86 3.09
25 8.1 0.2 19.4 1 4 1. 23.1 40.7 398.55 20.83
26 8.1 0.2 8.9 3 4 1.4 34.9 40.4 296.52 11.53
27 8.1 0.2 8.4 3 4 1.6 36. 40.3 306.24 12.42
28 8.1 0.2 9.6 1 4 1.4 35.2 40.4 297.58 11.63
29 8.1 0.2 8.4 4 4 1.6 41.6 40. 343.39 16.34
30 8.1 0.2 5.7 1 4 1.4 39.2 40.4 227.89 6.86
31 8.1 0.2 6.3 4 4 1.5 37.5 40.4 253.2 8.43
32 8.1 0.2 9.2 1 4 1.6 30.5 40.6 290.1 11.02
33 8.1 0.2 5.5 2 4 1.8 35.8 40.6 183.69 4.43

4.6.4 Problem Formulation of Transducers with a Mechanical Load

Optimization objectives In ultrasonic machining and bonding, piezoelectric transducers

operate against loads. The load has much influence on the behavior and vibration form of the
transducer. In order to study the effect of loads on the resonance performances, a simple
spring-damping load model is studied here. In the case of loading, the mechanical output
power and electrical input power are two important performance criteria. Here they are
considered as optimization objectives. For a given exciting voltage, the former should be
maximized, and in the meantime the latter should be minimized
106 CHAPTER 4

For the transducer with a spring-damping load on one end (see Fig. 4.33), there exist the
following boundary conditions

Fˆ e1 = 0 (4.78)

c
Fˆ a 4 = ( L + d L )vˆ a 4 (4.79)
jΩ

Substituting the boundary conditions (4.78) and (4.79) into (4.71), with simple computations
the following equations can be obtained

 jΩa 23
H
− jΩd L a13H − c L a13H ˆ
vˆ e1 =  H H HT
U (4.80)
 jΩd L a11 − jΩa 21 + c L a11 

vˆ a 4 = a11H vˆ e1 + a13H Uˆ (4.81)

Iˆ = a31H vˆ e1 + a33H Uˆ (4.82)

Applying equation (4.80) into equation (4.81), the solution of vˆ a 4 is obtained as follows:

 jΩa 23 H H H H
a11 − jΩa 21 a13 ˆ
vˆ a 4 =  H H HT
U (4.83)
 jΩd L a11 − jΩa 21 + c L a11 

where

 jΩa23 H H
a13 
H H
a11 − jΩa21
Im  H H HT 
=0 (4.84)
 jΩd L a11 − jΩa21 + cL a11 

is the characteristic equation representing the resonance of the transducer free at one end and
with a load on the other end.

dL
(a)
cL

v e1 va1 ve2 va2 v e3 va3 ve4 va4

(b) Fe1 1 Fa1 Fe2 2 Fa2 Fe3 3 Fa3 Fe4 4 Fa4

I U

Fig.4.33 Scheme of a stepped-horn transducer with a mechanical load

MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 107

The effective mechanical output power is

1
Pˆm = Fˆa 4 vˆa 4 cos ϕ m (4.85)
2

where ϕ m is the phase difference between Fˆ a 4 and vˆ a 4 . Applying equations (4.79) and
equations (4.83) yields
H H H H
ˆ 1 cL jΩa 23 a11 − jΩa 21 a13
Pm = + dL H H HT
Uˆ 2 cos ϕ m (4.86)
2 jΩ jΩd L a11 − jΩa 21 + c L a11

Equation (4.86) gives the expression of the first optimization objective, which should be
maximized. Because only minimization problems will be handled in all algorithms here
applied, the above objective for maximization is transformed into an objective for
minimization by using the inverse of the original objective. Therefore, the first objective
function is defined as
1
f1 = (4.87)
Pˆm

Introducing equation (4.80) into equation (4.82) yields the input current

 jΩa 23
H H
a31 − jΩd L a31PT a13PT − c L a31PT a13PT ˆ
Iˆ =  H H HT
+ a33PT U (4.88)
 jΩd L a11 − jΩa 21 + c L a11 

Therefore the input apparent power can be expressed as

H H
1 1 jΩa 23 a31 − jΩd L a31PT a13PT − c L a31PT a13PT
Pˆa = IˆUˆ = H H HT
+ a33PT Uˆ 2 (4.89)
2 2 jΩd L a11 − jΩa 21 + c L a11

Equation (4.89) gives the expression of the second objective function, which should be
minimized. It can be written as

f 2 = Pˆa (4.90)

Design variables The given quantities and design variables are the same as those of the
stepped-horn transducer without load described before. It is assumed that the load is a typical
spring-damping load for ultrasonic bonding [Brö02], which has the following parameter
values:

cL=1.27 kN/µm, (4.91)

dL=25.4/Ω Ns/µm (4.92)

where Ω is the vibration frequency of the transducer.

108 CHAPTER 4

Constraints Equation (4.84) gives the resonance condition of the transducer with a load at the
given resonance frequency f r . It describes an equality constraint for the design variable. The
constraints g2 to g9 in section 4.6.2 are also here used.

The two-objective optimization problem for the stepped-horn transducer with a mechanical
load can be written as follows:
Minimize F (x)
(4.93)
Subject to G (x)

Where F (x) = [ f1 (x), f 2 (x)] ,

G (x) = [g i (x), i = 1,..., 9] ,

x = ( x1 , x2 , x3 , x4 , x5 , x6 , x7 , x8 ) = ( Dt , ζ , Lb , Typb , Typ f , h p , L f 1 , L f 2 )

4.6.5 Implementation of Optimization

Considering that NSGA-ΙΙ shows a better overall performance in the optimization of the
freely vibrating transducer with a stepped horn, NSGA-ΙΙ is here applied for the all
optimization problems of the transducer with 2, 4 and 6 piezo-rings. The optimization
procedures similar to those in section 4.6.3 are used.

Figs. 4.34 to 4.36 show the obtained optimized solutions in the objective space after 300
generations for the number of the piezo-rings N =2, 4 and 6 respectively. Tables 4.23 to 4.25
give the corresponding design variable values and objective functions values. Similarly, the
second level optimization can be performed in the obtained all non-dominated solutions for
the transducer with 2 , 4 and 6 piezo-rings. This will be described in chapter 5.

2.5
Electrical power[VA]

1.5

0.5

1 2 3 4 5
1/Mechanical power [s/Nm]
Fig. 4.34 Non-dominated solutions for the stepped horn transducer with a mechanical load
and 2 piezo-rings
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 109

Table 4.23 Values of design variables and objective functions corresponding to non-
dominated solutions for the stepped horn transducer with a mechanical load and 2 piezo-
rings

Design
Dt
ζ Lb [mm] Typb Typf
hp Lf1 Lf2 P̂m P̂a
[mm] [mm] [mm] [mm] [Nm/s] [VA]

1 8. 0.86 34.3 3 3 4.6 29.9 24.3 0.22 0.3

2 8. 0.96 18.3 4 1 5. 15.3 72.9 0.19 0.27
3 8.5 0.88 34.8 4 3 4.5 32.6 12.8 0.34 0.43
4 8.8 0.86 35.7 4 4 4.2 33.9 2. 0.27 0.36
5 9.2 0.85 28. 3 3 4.7 32.7 29.6 0.31 0.41
6 9.4 0.86 41.5 3 1 4. 42.4 16.7 0.72 0.85
7 9.7 1. 44.3 1 3 3.3 31.8 20. 0.62 0.77
8 10.2 0.98 51.3 2 3 4. 8.7 41.7 0.73 0.87
9 10.3 0.79 42.9 3 1 4.1 43.1 14.4 1. 1.16
10 10.4 0.8 45. 2 1 4. 32.6 48.9 0.51 0.67
11 10.4 0.91 34.8 4 3 3.8 41.8 4. 0.84 1.
12 10.5 0.9 34.6 4 1 3.8 42. 17.7 1.14 1.31
13 10.5 0.89 34.8 4 3 3.9 41.7 3.9 0.87 1.02
14 10.5 0.9 46. 3 1 3.5 31.4 23.4 1.04 1.22
15 10.7 0.97 43.7 1 2 3.5 30. 35.4 0.36 0.54
16 11. 0.89 44.3 3 1 3.5 31.9 26.1 1.25 1.45
17 11.3 0.93 40.8 1 3 3.4 39.3 16.9 1.05 1.26
18 11.4 0.99 53.1 2 1 3.9 6.9 53.5 1.73 1.93
19 11.7 0.8 40.7 3 1 3.7 39.2 26.6 1.49 1.72
20 11.8 0.92 36.7 4 1 3.4 49.2 5.9 1.91 2.17
21 11.8 0.81 46.1 3 1 3.5 28.3 30.6 1.35 1.58
22 11.9 0.86 41.4 1 1 3.5 37.9 36.7 1.39 1.63
23 11.9 0.86 40.2 2 3 3.5 39.3 30.2 0.76 0.98
24 11.9 0.78 41.7 4 3 3.7 34.2 2. 1.26 1.47
25 12. 0.78 42. 3 1 3.7 38.1 26.4 1.57 1.82
26 12.8 0.66 47.8 3 1 3.9 42.3 9.1 2.37 2.66
27 13.2 1. 40.9 3 3 2.6 36.2 14.2 2.2 2.56

17.5
Electrical power[VA]

15

12.5

10

7.5

2.5

0 0 2 4 6 8
1/Mechanical power [s/Nm]

Fig. 4.35 Non-dominated solutions for the stepped horn transducer with a mechanical load
and 4 piezo-rings
110 CHAPTER 4

Table 4.24 Values of design variables and objective functions corresponding to non-
dominated solutions for the stepped horn transducer with a mechanical load and 4 piezo-
rings

Design
Dt
ζ Lb [mm] Typb Typf
hp Lf1 Lf2 P̂m P̂a
[mm] [mm] [mm] [mm] [Nm/s] [VA]

1 8. 0.56 20.3 3 1 3.9 48.2 45.9 0.34 0.68

2 8. 0.64 24.9 2 4 3.7 44.3 16. 0.3 0.59
3 8. 0.76 31.6 4 4 3.1 33.3 2. 0.71 0.92
4 8. 0.8 24.2 4 1 2.7 69.2 5.7 1.39 1.68
5 8. 0.81 16.2 3 4 3.7 47.7 2. 0.49 0.69
6 8. 0.89 5. 3 2 5. 32.2 53.8 0.11 0.33
7 8.1 0.61 27.4 1 3 3.5 45.5 17.5 0.83 1.1
8 8.2 0.66 22.6 1 3 4. 27.5 38.9 0.29 0.58
9 8.5 0.71 10.2 4 3 5. 28.4 42.6 0.22 0.46
10 8.7 0.62 27.3 4 1 4. 28.2 40.7 0.82 1.09
11 9.2 0.89 53.3 2 4 2.1 9.7 30.7 0.86 1.3
12 10.5 0.66 22.2 3 3 4.2 31.1 32.5 1.26 1.64
13 11.6 0.9 18. 2 1 4.2 32.9 60.4 1.72 2.11
14 11.6 0.51 44.7 1 4 3.8 29.8 2. 3.12 3.59
15 11.6 0.35 33. 4 3 3.6 33.9 9.7 4.27 4.94
16 11.8 0.54 50.1 3 4 3. 20.6 17. 2.01 2.6
17 12.1 0.6 32.9 3 1 3.5 36.1 34.4 4.16 4.82
18 12.4 0.32 57.7 4 4 4.1 6.8 2. 1.15 1.48
19 12.8 0.69 24.8 3 3 3. 53.5 5.3 6.89 7.72
20 12.9 0.79 32. 4 4 4.1 27.8 2. 4.84 5.44
21 13. 0.32 39.2 4 3 3.4 28.3 2. 6.23 6.97
22 13.2 0.82 29.8 2 4 4.4 28.2 21.8 2.89 3.42
23 13.4 0.65 28.4 3 4 3.2 39.3 2. 5.69 6.48
24 13.6 0.44 27.8 2 4 4.1 42.8 19.7 2.36 3.15
25 13.7 0.2 44.9 3 3 3.1 32.9 2. 8.79 9.8
26 14. 0.24 51.1 2 4 4.5 27.8 2. 6.8 7.51
27 14.1 0.73 35.5 3 4 3. 34.7 2. 7.29 8.23
28 14.1 0.49 34.4 4 3 3.8 28.8 12.5 8.66 9.6
29 15. 0.73 31.1 2 4 4.2 26.3 26.5 3.38 4.08
30 15. 0.8 33.8 4 2 3.9 33. 21.2 5.4 6.17
31 16.3 0.76 37.1 4 3 3.8 27.1 2. 16.89 18.06
MULTIOBJECTIVE OPTIMIZATION OF PIEZOELECTRIC TRANSDUCERS 111

17.5

Electrical power[VA]
15
12.5

10

7.5

2.5

0.5 1 1.5
1/Mechanical power [s/Nm] 2

Fig.4.36 Non-dominated solutions for the stepped horn transducer with a mechanical load
and 6 piezo-rings

Table 4.25 Values of design variables and objective functions corresponding to non-
dominated solutions for the stepped horn transducer with a mechanical load and 6 piezo-
rings

Design
Dt
ζ Lb [mm] Typb Typf
hp Lf1 Lf2 P̂m P̂a
[mm] [mm] [mm] [mm] [Nm/s] [VA]

1 8. 1. 58.5 2 1 5. 6.6 4.4 0.96 1.11

2 8. 1. 46.6 2 2 4.9 12.8 19.4 1.01 1.2
3 8. 1. 46.8 4 2 5. 3.1 2. 0.49 0.62
4 8. 1. 52.1 2 1 4.8 13. 3.9 1.32 1.51
5 8. 1. 53.7 2 3 5. 5. 8.2 1.12 1.29
6 8. 1. 58.5 3 1 4.7 3.4 2. 0.59 0.73
7 8.1 1. 40.6 3 4 5. 10.1 2. 1. 1.17
8 8.1 0.94 40.1 1 1 5. 5.8 13.6 1.46 1.67
9 8.3 0.99 39.9 1 3 5. 5. 11.6 1.49 1.7
10 8.3 0.68 40.6 4 3 5. 5. 5.1 0.86 1.02
11 8.5 0.92 54.9 3 2 5. 5. 5.3 0.78 0.94
12 8.7 0.95 48.3 2 4 4.9 10.6 7.4 1.62 1.87
13 8.7 0.95 48.1 1 2 4.9 10.4 15.6 1.26 1.49
14 9.2 0.91 45.7 3 1 4.9 8.1 3.7 1.68 1.93
15 9.2 0.91 45.7 2 3 4.9 8. 14.1 2.57 2.9
16 9.3 0.91 41.5 1 1 4.9 7.8 11.8 2.48 2.81
17 9.4 0.92 47.7 1 4 4.8 8.9 6. 1.98 2.28
18 10. 0.89 48.6 3 1 4.7 8.9 2. 2.18 2.5
19 10. 0.85 45.4 2 4 4.9 6.4 16.3 2.59 2.98
20 10.5 0.82 45.2 3 4 4.9 5. 7.9 2.5 2.87
21 10.5 0.94 27. 3 4 4.7 7.2 16.7 3.68 4.17
22 10.6 0.85 43.6 3 1 4.8 5.8 10.2 3.26 3.69
23 10.7 0.77 51.9 1 3 4.3 12.3 6.3 3.99 4.51
24 10.8 0.97 33.3 3 2 5. 5. 25.4 3.17 3.62
25 11.2 0.92 49.6 1 3 4. 7.3 13.3 5.28 5.93
112 CHAPTER 4

Table 4.25 continued

Design
Dt
ζ Lb [mm] Typb Typf
hp Lf1 Lf2 P̂m P̂a
[mm] [mm] [mm] [mm] [Nm/s] [VA]

26 11.2 0.78 36.5 2 3 4.3 11.1 29.9 4.85 5.5

27 11.3 0.91 40.6 3 1 4.8 6.7 9.7 4.75 5.31
28 11.3 0.92 28.4 1 1 4.6 8.6 25.8 8.58 9.35
29 11.5 0.9 41.9 1 1 4.4 18.5 4.9 7.44 8.2
30 11.5 0.92 37.7 3 3 4.8 6.3 10.3 5.27 5.88
31 11.8 0.66 45.7 2 3 3.8 15.4 25.6 5.64 6.45
32 12. 0.77 34. 2 1 4.1 25.3 31. 11.02 11.99
33 12. 0.94 32.4 1 1 4.6 13.1 16.2 10.74 11.62
34 12. 0.93 32.8 1 4 4.6 13.3 10.3 7.02 7.8
35 12.2 0.96 27.4 3 4 5. 5. 16.2 6.84 7.59
36 12.5 0.9 47.3 2 3 3.6 7.6 24.8 10.42 11.44
37 12.7 0.92 23.7 2 4 4.9 5.2 30.4 6.52 7.31
38 13.2 0.79 49.8 1 1 3.6 8.3 26.1 8.84 10.52
39 13.3 0.48 58.1 2 1 3.2 21.5 17.3 13.03 14.3
40 13.4 0.93 46.8 2 1 4.1 10.5 21.2 17.69 18.73

In this chapter, several two-objective optimization problems for Langevin-type transducers

have been studied. Based on experiments, applying lumped parameter models and parameter
identification the optimal pre-stress for multiple objectives have been found. Using transfer
matrix methods, the optimization problems for the symmetrical transducer and the transducer
with a stepped horn were formulated. Optimizations were performed using MOGA, NSGA,
NSGA-II, SPEA and SPEA2 and the results have been obtained. In next chapter, the results
obtained will be analyzed and discussed.
RESULTS AND DISCUSSIONS 113

5 Results and Discussions

In chapter 4, non-dominated solutions for transducers with 2, 4 and 6 piezoelectric rings have
been obtained. In order to identify the final Pareto-optimal solutions of each optimization
problem, the non-domination check (the second level optimization) needs to be performed in
these solutions. After the second level optimization is finished, an important problem is how
to determine the preferred solution for the implementation of the optimal design. In this
chapter, first, the results of the second level optimization are presented and analyzed. Then
the preferred solution is selected by means of different selection methods. Finally, the load
characteristics of the Pareto-optimal stepped-horn transducer are studied.

5.1 Discussion of the Results for the Symmetrical Transducer

The non-dominated solutions for the symmetrical Langevin transducer with 2, 4 and 6 piezo-
rings under the condition of no load have been given in chapter 4 (refer to Figs. 4.19, 4.22 and
4.23 as well as Tables 4.8, 4.12 and 4.13). Here they are plotted in Fig. 5.1. The Pareto-
optimal solutions were searched again in these non-dominated solutions. Fig 5.2 shows the
non-dominated solutions in objective space after the search is finished.

40000
1/Amplitude[1/m]

2 Piezo
30000

4 Piezo
20000

6 Piezo
10000

25 50 75 100 125 150 175

Input Power [VA]

Fig. 5.1 Non-dominated solutions in objective space for the symmetrical transducers with 2, 4
and 6 piezo-rings
114 CHAPTER 5

40000
1/Amplitude[1/m]

30000

2 Piezo
20000

4 Piezo
10000

25 50 75 100 125 150 175

Input Power [VA]

Fig. 5.2 Non-dominated solutions in objective space for the symmetrical transducer obtained
after the second level optimization

5.1.1 Analysis of the Results of the Optimization

It is obvious that the set of non-dominated solutions for the transducer with 6 piezo-rings are
dominated by members of the non-dominated sets of the transducers with 2 and 4 piezo-rings.
Therefore, they do not appear in the final results. The final set of Pareto-optimal solutions
consists of all 35 members of the non-dominated set for 2 piezo-rings and 7 members for 4
piezo-rings. Table 5.1 gives the values of the design variables and objective functions for
these Pareto-optimal solutions. It is noted that the value of the diameter of the transducer for
all solutions is equal to the lower bound of this design variable. This situation can be
explained from the point of view of engineering. The smaller the diameter of the transducer
is, the smaller the capacitance of the piezo-rings and the loss of the material. This will result
in the decrease of the input power and the increase of the vibration amplitude. Table 5.1 also
shows that the front and back metal sections of the symmetrical transducer should be made of
titanium or brass when the electrical input power and the vibration amplitude at the resonance
are considered as objective functions. In addition, compared with the transducers with 2
piezo-rings, the transducers with 4 piezo-rings have the larger electrical input power and
vibration amplitude. It is also noticed that the total length of the transducers with two end
blocks made of brass is shorter than that of the transducers with two end blocks made of
titanium. The end blocks of the all transducers with 4 piezo-rings are made from brass.
RESULTS AND DISCUSSIONS 115

Table 5.1 Values of design variables and objective functions as well as the corresponding
pseudo weights after the second level optimization for the symmetrical transducer
uˆ a 3 P̂a
Dt hp L
Design
[mm]
Typb
[mm] [mm]
N [µm] [VA] w1 w2
(Obj1) (Obj2)

1 8. 2 0.3 63.2 2 33.12 15.89 0.28 0.72

2 8. 2 0.7 62.6 2 31.92 15.19 0.26 0.74
3 8. 2 1.4 61.5 2 29.94 14.22 0.23 0.77
4 8. 2 1.6 61. 2 29.26 13.89 0.21 0.79
5 8. 2 5. 55.5 2 22.58 10.66 0. 1.
6 8. 4 0.2 40.7 2 92.19 53.43 0.53 0.47
7 8. 4 1.5 39. 2 58.08 33.55 0.45 0.55
8 8. 2 0.8 62.4 2 31.55 15. 0.25 0.75
9 8. 4 3.8 36.1 2 35.39 20.28 0.31 0.69
10 8. 4 0.3 40.5 2 87.19 50.47 0.52 0.48
11 8. 4 4. 35.8 2 34.09 19.52 0.29 0.71
12 8. 4 0.7 40.1 2 76.27 44.11 0.5 0.5
13 8. 4 4. 35.8 2 34.21 19.59 0.29 0.71
14 8. 2 4.7 56. 2 23.05 10.89 0.02 0.98
15 8. 2 1. 62.1 2 31.03 14.75 0.24 0.76
16 8. 2 1.8 60.8 2 28.9 13.72 0.21 0.79
17 8. 2 3.7 57.6 2 24.68 11.68 0.09 0.91
18 8. 4 3.7 36.1 2 35.81 20.53 0.31 0.69
19 8. 2 4.2 56.8 2 23.8 11.26 0.06 0.94
20 8. 4 2.5 37.7 2 45.56 26.25 0.39 0.61
21 8. 4 2.1 38.3 2 50.11 28.9 0.41 0.59
22 8. 2 3.1 58.6 2 25.92 12.28 0.13 0.87
23 8. 4 1.8 38.6 2 53.98 31.16 0.43 0.57
24 8. 4 2.7 37.5 2 43.76 25.2 0.38 0.62
25 8. 4 2.2 38.1 2 49.01 28.26 0.41 0.59
26 8. 2 0.9 62.2 2 31.13 14.8 0.25 0.75
27 8. 4 0.7 40. 2 75.12 43.45 0.5 0.5
28 8. 4 2.6 37.6 2 44.7 25.75 0.39 0.61
29 8. 4 2.7 37.5 2 43.84 25.25 0.38 0.62
30 8. 4 2. 38.3 2 50.9 29.37 0.42 0.58
31 8. 2 2.4 59.8 2 27.5 13.04 0.17 0.83
32 8. 2 2.4 59.7 2 27.31 12.95 0.17 0.83
33 8. 4 3.3 36.7 2 38.89 22.35 0.34 0.66
34 8. 2 3. 58.8 2 26.17 12.4 0.14 0.86
35 8. 2 2.3 59.9 2 27.59 13.09 0.18 0.82
36 8. 4 0.8 39. 4 116.51 135.01 0.75 0.25
37 8. 4 0.3 40.1 4 153.58 178.24 0.94 0.06
38 8. 4 0.4 40. 4 147.79 171.49 0.9 0.1
39 8. 4 0.9 38.5 4 105.04 121.66 0.7 0.3
40 8. 4 0.3 40.2 4 159.7 185.41 0.97 0.03
41 8. 4 0.8 38.8 4 112.31 130.16 0.73 0.27
42 8. 4 0.2 40.3 4 163.97 190.41 1. 0.
116 CHAPTER 5

5.1.2 Determination of the Preferred Solution

In the practical design of piezoelectric actuators, generally, only one solution needs to be
implemented. Therefore, how to select a particular solution from the set of non-dominated
solutions obtained is an important task from point of view of engineering.

The methods to determine the preferred solution from the obtained non-dominated solutions
can be divided into two classes. One class is to choose the preferred solution from the whole
set of non-dominated solutions according to high-level information. Since every solution is
considered in these methods, the best solution according to the high-level selection
information will not be lost. However, the cost to evaluate solutions in these methods is high
because in general the size of the non-dominated set is large. The other class is first to reduce
the size of the set of non-dominated solutions and then to determine the preferred solution
from the remaining solutions according to high-level information. The advantages of this
class of methods are the low cost to evaluate solutions and the convenience to determine the
preferred solution. However, the best solution may be lost in the process of reducing the size
of the non-dominated set.

If certain high level information is known, the preferred solution can be chosen according to
this high level information. As far as optimization problems of piezoelectric actuators are
concerned, any performance except those performance criteria used as optimization objectives
in the optimization problem can be considered as the high level information. From the point
of view of the ultrasonic technique, the piezoelectric quality number, the power efficiency and
the effective coupling factor are some important performance criteria. In addition to using
performance criteria of piezoelectric actuators, many techniques for multiple criteria decision-
making can also be adopted to select the preferred solution [Mie98]. One of the techniques is
using pseudo-weight vectors

After the non-dominated solutions were obtained, a pseudo-weight vector can be calculated
for each solution according to the following equation:

wi =
(f i
max
− fi ) (f i
max
− f i min ) (5.1)
m

∑(f
j =1
j
max
− f ) (f
j j
max
−f j
min
)
where f i max and f i min represent the maximum (worst) and minimum (best) values of the i-th
objective function in the obtained non-dominated solutions. m is the number of the
optimization objectives. The weight vector wi gives a relative important factor for the i-the
objective corresponding to the solution. Obviously, the weight wi is maximum for the best
solution for the i-th objective function.

The clustering technique is a technique usually used to reduce the size of the set of non-
dominated solutions. Here, for the two-objective optimization problem the following
procedures of the clustering are used [Zit99]: first, sort the obtained non-dominated solutions
RESULTS AND DISCUSSIONS 117

in ascending order of the value of objective function 1. Second, consider each of the solutions
as a separate cluster and find the centroid of each cluster, then calculate the Euclidean
distance between all pairs of the clusters in order. Third, combine the two clusters which have
a minimum distance. Next, find the centroid of each cluster, calculate the Euclidean distance
between all pairs of the clusters, and combine the two clustering having a minimum distance
into a cluster. Continue these steps until the required number of the clusters is achieved.
Finally, in each cluster the solution closest to the centroid is kept and the rest is abandoned. In
the whole process of clustering the two extreme solutions are always taken into account as
separate clusters.

In the following, the preferred design for implementation will be determined using the
methods described above.

Selection according to weight vectors The preferred solution can be determined according to
pseudo-weight vectors. Table 5.1 gives the calculated weights w1 and w2 , where w1 is for the
vibration amplitude and w2 for the electrical input power. A solution can be selected
according to a designer-preferred weight vector. For example, if a weight value of 70% is
assigned to the objective 1 (the vibration amplitude) and a weight value of 30% to the
objective 2 (the input power), solution (design) 39, which has a similar weight vector, can be
chosen. Design 5 and Design 42 are two extreme solutions among the obtained non-
dominated solutions. The former represents the Pareto-optimal design of a symmetrical
transducer which has the smallest electrical input power. The latter represents the Pareto-
optimal design of a symmetrical transducer which has the largest vibration amplitude. It is
noted that the larger the pseudo-weight of the objective function 1 (namely the larger the
vibration amplitude) is, the smaller the thickness of the piezo-stacks. On the other hand, the
larger the pseudo-weight of the objective function 2 (namely the smaller the input electrical
power) is, the larger the thickness of the piezo-stacks.

Table 5.2 Four designs used for explaining how to determine the preferred solution for the
symmetrical transducer

uˆ a 3 [µm] P̂a [VA]

Design Dt [mm] Typb hp [mm] Lb [mm] N w1 w2
(Obj1) (Obj2)

3 8 2 1.4 61.5 2 29.94 14.22 0.23 0.77

5 8 2 5 55.5 2 22.58 10.66 0 1

23 8 4 1.8 38.6 2 53.98 31.16 0.43 0.57

42 8 4 0.2 40.3 4 163.97 190.41 1 0

118 CHAPTER 5

Selection according to performance criteria In the optimization problem of the

symmetrical transducer, the input electrical power and the vibration amplitude at resonance
have been considered as optimization objectives. The rest of performance criteria can be used
as high-level information for the selection of optimized results. As stated before, the
piezoelectric quality number, the power efficiency, the effective coupling factor are important
performance criteria. Here the effective coupling factor k is first considered.

As described in chapter 2 and chapter 4, the effective coupling factor k describes the effective
energy conversion in piezoelectric transducers. The larger the value of k, the more electrical
energy is transferred into mechanical energy. The effective coupling factor k corresponds to
the distance between resonance frequency and anti-resonance frequency for a vibration mode.
Therefore, the design corresponding to the largest distance between resonant and anti-
resonant frequencies should be chosen. The resonance and anti-resonance frequencies can be
readily obtained if the frequency response of the admittance function Iˆ Uˆ is plotted. The
frequency response of the admittance function Iˆ Uˆ for the symmetrical transducer can be
calculated from equation (4.64). Therefore, the frequency responses of the admittance
functions Iˆ Uˆ for every members of the set of the final Pareto-optimal solutions shown in
Table 5.1 can be plotted and the preferred design having the best coupling factor k can be
determined. Here, as an example to explain how to determine the preferred selection
according to the high level information k, 4 designs in Table 5.1 have been studied. These
designs are shown in Table 5.2. Fig. 5.3 shows the frequency responses of the admittance
functions Iˆ Uˆ for the 4 designs. It is shown in Fig. 5.3 that the distance between the
resonance and anti-resonance frequencies increases with the increase of the thickness N ⋅ h p of
the piezoelectric stack. Design 5, which has the largest thickness of the piezo-stack
( N ⋅ h p = 2 × 5 = 10mm ), has the largest distance between resonance and anti-resonance
frequencies. Therefore, according to the coupling factor k Design 5 is the preferred design
among these 4 Pareto-optimal designs.

Selection by means of clustering technique In order to perform the required selection more
easily, the size of Pareto-optimal solutions can be reduced by means of the above described
clustering technique before the selection is performed. Fig. 5.4 shows the results after
clustering operation with a choice of 6 final solutions in the objective space. Table 5.3 gives
the values of the design variables and objective functions for the 6 final solutions. The
configurations of the transducers for the 6 final solutions are shown in Fig. 5.5.
RESULTS AND DISCUSSIONS 119

© ¦I/U© ¦ [ A/ V] 0.01

0.001

0.0001

Design3
0.00001
Design5
1.×10-6
Design23
19000 19500 20000 20500 21000 21500 22000
Frequency [ Hz] Design42

75

50
Phase[¡ ]
ã

25

-25
-50

-75

19000 19500 20000 20500 21000 21500 22000

Frequency[Hz]

Fig.5.3 Admittance frequency responses for the 4 designs given by Table 5.2

40000 Design 5
1/Amplitude[1/m]

Design 22

30000
Design 2

20000 Design 25
Design 39
Design 42
10000

25 50 75 100 125 150 175

Input Power [VA]

Fig. 5.4 Six final Pareto-optimal solutions after clustering for the symmetrical transducer
120 CHAPTER 5

Table 5.3 Values of design variables and objective functions as well as weights for 6 Pareto-
optimal solutions after clustering for the symmetrical transducer

Dt hp Lb uˆ a 3 [µm] P̂a [VA]

Design Typb N w1 w2
[mm] [mm] [mm] (Obj1) (Obj2)
2 8 2 0.7 62.6 2 31.39 15.19 0.26 0.74
5 8 2 5 55.5 2 22.58 10.66 0 1
22 8 2 3.1 58.6 2 25.92 12.28 0.13 0.87
25 8 4 2.2 38.1 2 49.01 28.26 0.41 0.59
39 8 4 0.9 38.5 4 105.04 121.66 0.7 0.3
42 8 4 0.2 40.3 4 163.97 190.41 1 0

Design 2

Design 5

Design 22

Design 25
5mm

Design 39 PIC181
Brass

Design 42 Titanium

Fig. 5.5 Configurations of the symmetrical transducers for 6 Pareto-optimal solutions after
clustering

After clustering is finished, similarly, the high level information described before can then be
applied to determine a preferred design from the results of clustering for implementation. It is
pointed out that in the optimization problems in chapter 4 the set of Pareto-optimal solutions
is searched under the condition of a given voltage Uˆ = 100 V . If the range of the working
voltage is given, the ranges of the objective function values are also determined. This
information is available for determining the preferred solution. The variations of two
objective function values with respect to different values of the input voltage between 10V to
300V for six Pareto-optimal designs obtained after clustering (see Table 5.3) have been
RESULTS AND DISCUSSIONS 121

calculated. Fig. 5.6 shows the results for the input power at resonance smaller than 100 VA
and the vibration amplitude at resonance larger than 10µm. The number of the piezo-rings
will affect the range of the input power and the vibration amplitude of the transducer for the
given range of voltage. For example, the vibration amplitude of Design 42 with 4 piezo-rings
is not smaller than about 16 µm (refer to point A). The vibration amplitude of Design 5 with 2
piezo-rings will not be larger than about 57 µm (refer to point B). Its input power will not be
more than about 67 VA. Obviously, Design 42 is the best design because the others do not
dominate the set of its Pareto-optimal solutions with respect to the input voltage. That is, in
order to arrive at a given vibration amplitude (here not smaller than about 16 µm), the
transducer corresponding to Design 42 needs the minimum input power among the 6 Pareto-
optimal designs. Therefore Design 42 is the preferred design for implementation. It is noticed
that the material of the end blocks and the total thickness of the piezo-stack affects the
performances of transducers for a given range of the input voltage. However, the
performances are first determined by material and then by the thickness. The transducers with
the end blocks made of brass have better performance than those with the end blocks of
titanium. Under the condition of the same material, the transducers with a thin piezo-stack
have better performances than the others.

In the above sections, the Pareto-optimal solutions for the symmetrical freely vibrating
transducer have been analyzed. The methods to determine the preferred design have been
discussed. Some important rules for designing the symmetrical Langevin-type transducers via
multiobjective optimization are concluded as follows:

100000
1/Amplitude[1/m]

80000
A
60000
B Design2
40000
Design5

20000 Design22

Design25

20 40 60 80 100 Design39
Input power [VA] Design42

Fig. 5.6 Values of the input power and the vibration amplitude of the symmetrical transducers
for 6 Pareto-optimal designs obtained after clustering for a given range of the input voltage
from 10V to 300V
122 CHAPTER 5

1. Brass and titanium are two kinds of preferred material of the end blocks for Pareto-
optimal designs. If a Pareto-optimal transducer is expected to have good performance
(lower electric input power and higher vibration amplitude) in a given range of the
voltage, brass is the preferred material for the end blocks. Furthermore, the total
thickness of the piezo-stack should be as small as possible. However, if a larger
effective coupling factor is expected, a thicker piezo-stack needs to be used.

2. The number of the piezoelectric rings will affect the ranges of the values of objective
functions. For a given voltage range, the maximum input power of the transducer with
2 piezo-rings is smaller than that of the transducer with 4 piezo-rings. On the other
hand, the minimum vibration amplitude of the transducer with 2 piezo-rings is lower
than that of the transducer with 4 piezo-rings.

It should be pointed out the implemented Pareto-optimal solution here is selected from many
known candidates which are obtained using MOEAs. This is different from the process of
obtaining a Pareto-optimal solution by using a traditional scalarization method, where a
preference structure over objectives is given in advance and then an implemented Pareto-
optimal solution is found. Obviously, the approach in which the entire set of Pareto-optimal
solutions is first identified and the preferred solution is then chosen is more practical and less
subjective.

5.2 Discussion of the Results for Stepped-horn Transducers without Load

Fig. 5.7 shows three sets of non-dominated solutions for the stepped-horn transducers with 2,
4 and 6 piezo-rings obtained in the first level optimization (refer to Figs. 4.28, 4.31 and 4.32
as well as Tables 4.17, 4.21 and 4.22). Similarly, the Pareto-optimal solutions were searched
again in these non-dominated solutions. Fig 5.8 shows the results after non-domination check.
Table 5.4 gives the values of the design variables and objective functions for these Pareto-
optimal solutions after the second level optimization.

35000
30000
1/Amplitude[1/m]

25000 2 Piezo
20000
15000 4 Piezo
10000
5000
6 Piezo

0 10 20 30 40 50
Input Power [VA]
Fig. 5.7 Non-dominated solutions in objective space for the freely vibrating stepped-horn
transducer with 2 , 4 and 6 piezo-rings
RESULTS AND DISCUSSIONS 123

1/Amplitude[1/m] 20000
2 Piezo
15000

10000 4 Piezo

5000 6 Piezo

0 10 20 30 40 50
Input Power [VA]

Fig. 5.8 Non-dominated solutions in objective space for the freely vibrating stepped-horn
transducer obtained after the second level optimization

Table 5.4 Values of design variables and objective functions for the non-dominated solutions
obtained after the second-level optimization for the freely vibrating stepped-horn transducer

Dt Lb hp Lf1 Lf2 uˆ a 4 P̂a

Design ζ Typb Typf N
[mm] [mm] [mm] [mm] [mm] [µm] [VA]

1 8. 0.2 5. 2 4 5. 8.1 42. 2 73.74 0.7

2 8. 0.2 5. 3 3 5. 46.5 50.3 2 49.71 0.51
3 8. 0.2 5. 1 4 5. 9.7 41.7 2 83.12 0.89
4 8. 0.2 7.1 3 4 3.5 5. 42.1 2 107.97 1.51
5 8. 0.2 5.6 2 4 4.9 41.9 40.3 2 62.98 0.51
6 8. 0.2 10.2 3 4 3.5 9. 41.6 2 110.07 1.55
7 8. 0.2 6.2 3 4 4.9 25. 40.9 2 78.11 0.79
8 8. 0.21 17.7 4 4 3.7 23. 40.6 2 139.87 2.73
9 8. 0.2 8.4 2 3 4.6 37.2 50.7 2 42.33 0.39
10 8. 0.21 7.4 3 4 5. 14.5 41.3 2 90.83 1.12
11 8. 0.2 15.7 2 4 3.9 13. 41.4 2 94.67 1.19
12 8. 0.2 10.8 2 4 4.4 13.2 41.5 2 81.04 0.88
13 8. 0.2 11.8 2 4 4.1 15.4 41.4 2 80.1 0.84
14 8. 0.2 13.6 3 4 3.5 10.7 41.3 2 120.76 1.91
15 8. 0.2 14.6 1 4 3.2 8.2 41.5 2 124.03 2.01
16 8. 0.2 16.3 1 4 2.4 5. 41.7 2 138.67 2.48
17 8. 0.21 14.5 2 4 4. 18.7 41.2 2 85.14 0.97
18 8.1 0.21 20.7 2 4 3.3 19.7 41.1 2 100.77 1.38
19 8.1 0.21 17.8 1 4 3.7 12. 41.1 2 128.38 2.27
20 8.1 0.21 36.7 1 4 1.6 12.4 40.8 2 170.71 4.12
21 8.1 0.21 30.4 3 4 2.6 16. 40.5 2 171.48 4.42
22 8.1 0.21 37. 2 4 1.5 5.8 41.4 2 151.31 3.35
23 8. 0.2 12.4 1 4 2.5 6. 41.5 4 242.77 7.77
24 8. 0.2 21.5 2 4 1. 5.8 41.9 4 251.75 8.18
25 8. 0.2 5. 1 4 3.3 6.1 41.7 4 188.67 4.65
26 8. 0.2 11.5 2 4 3.2 5.7 41.7 4 203.43 5.48
124 CHAPTER 5

Table 5.4 continued

Dt Lb hp Lf1 Lf2 uˆ a 4 P̂a
Design ζ Typb Typf N
[mm] [mm] [mm] [mm] [mm] [µm] [VA]

27 8. 0.2 15. 2 4 2.8 5.4 41.7 4 220.03 6.44

28 8. 0.2 17.1 2 4 2.7 5.5 41.6 4 228.54 6.95
29 8. 0.2 8.6 3 4 3.1 6.1 41.5 4 222.46 6.55
30 8. 0.2 8.9 2 4 3. 5.8 41.8 4 188.88 4.66
31 8. 0.2 26.6 4 4 1. 5.2 41.1 4 334.68 14.33
32 8. 0.2 26.5 4 4 1.1 5.2 41.1 4 332.49 14.26
33 8. 0.2 12.4 2 4 3.4 5.2 41.6 4 210.05 5.89
34 8. 0.2 12.7 2 4 3.6 5.4 41.5 4 211.34 5.99
35 8. 0.2 20.6 4 4 1.8 5.2 41.2 4 301.42 11.96
36 8. 0.2 5. 1 4 5. 5.2 41.4 4 199.08 5.47
37 8. 0.2 7.9 4 4 1.2 48.8 39.6 6 340.97 16.15
38 8. 0.2 50. 1 4 0.2 5.6 40.9 6 561.83 39.66
39 8. 0.2 20.4 4 4 0.7 21.8 40.7 6 441.49 24.83
40 8. 0.2 25. 1 4 0.2 5. 41.5 6 487.2 31.15
41 8. 0.2 18.6 3 4 0.8 24.7 40.6 6 412.6 21.82
42 8. 0.2 58.5 4 4 0.2 5. 39.6 6 648.77 57.59
43 8. 0.2 5. 2 4 1.3 42.4 40.4 6 153.78 3.36
44 8. 0.2 18.6 3 4 0.8 23.9 40.7 6 406.38 21.17
45 8. 0.2 17. 1 4 1. 27.8 40.6 6 380.69 18.72
46 8. 0.2 15.8 1 4 1.1 29.8 40.5 6 372.16 17.95
47 8. 0.2 16.8 2 4 1. 29.4 40.8 6 260.58 8.71
48 8. 0.2 18.8 2 4 0.7 16.8 41.3 6 277.37 10.15
49 8. 0.2 11.7 1 4 1. 27.9 40.8 6 289.66 10.78
50 8. 0.2 15.6 2 4 0.9 24.1 41. 6 239.88 7.52
51 8.1 0.2 6.6 2 4 1.3 39.4 40.5 6 168.02 3.9
52 8.1 0.2 29.2 1 4 1. 20.6 40.5 6 499.12 33.44
53 8.1 0.2 26.2 3 4 1. 22.8 40.3 6 521.78 37.77
54 8.1 0.2 5. 4 4 1.3 31.3 40.8 6 191.9 4.9
55 8.1 0.2 19.4 1 4 1. 23.1 40.7 6 398.55 20.83
56 8.1 0.2 8.9 3 4 1.4 34.9 40.4 6 296.52 11.53
57 8.1 0.2 8.4 3 4 1.6 36. 40.3 6 306.24 12.42
58 8.1 0.2 9.6 1 4 1.4 35.2 40.4 6 297.58 11.63
59 8.1 0.2 8.4 4 4 1.6 41.6 40. 6 343.39 16.34
60 8.1 0.2 5.7 1 4 1.4 39.2 40.4 6 227.89 6.86
61 8.1 0.2 6.3 4 4 1.5 37.5 40.4 6 253.2 8.43
62 8.1 0.2 9.2 1 4 1.6 30.5 40.6 6 290.1 11.02
63 8.1 0.2 5.5 2 4 1.8 35.8 40.6 6 183.69 4.43

5.2.1 Analysis of the Results of the Optimization

The final set of Pareto-optimal solutions for the stepped-horn transducer consists of 22
members from the set of the non-dominated solutions for 2 piezo-rings, 14 members from the
set for 4 piezo-rings and 27 members from the set for 6 piezo-rings. Unlike the final results
after the second level optimization for the symmetrical transducer, the final results for the
stepped-horn transducer include some members of the set of non-dominated solutions for 6
piezo-rings. This is because not all members of the set of non-dominated solutions of the
RESULTS AND DISCUSSIONS 125

stepped-horn transducer with 6 piezo-rings are dominated by the members of the set of non-
dominated solutions of the transducer with 2 and 4 piezo-rings. It is noted that the diameter of
the transducers for all solutions is equal or nearly equal to the lower bound of this design
variable. The reason for this phenomenon has been given in section 5.1.1. Furthermore, all
amplitude transformation ratios ζ, namely the ratio of the diameter between the output side
and input side of the horn, are equal or nearly equal to the lower bound of this design variable.
This is because a horn with the ratio ζ < 1 is a mechanical amplifier. For a given vibration
amplitude at the input side of the horn, the smaller the ratio ζ is, the larger the vibration
amplitude at the output side. Therefore, in order to obtain the maximum vibration amplitude
at the output side of the horn, ζ needs to be as small as possible. It should be noted that a large
transformation will result in a “sensitive” transducer, i.e. a transducer where the output load
has a large influence on the system’s behavior. In the present optimization problem, however,
this effect has not been taken into account in the formulation of the objective functions. In all
Pareto-optimal designs, there are four types of material (steel, titanium, aluminum bronze and
brass) for the back section but only two types (aluminum bronze and brass) for the horn. Most
horns are made of brass. Since a horn is used as mechanical amplifier, the stepped-horn
transducer can produce the larger vibration amplitude for a given input power compared with
the symmetrical transducer.

5.2.2 Determination of the Preferred Solution

Similarly, there are two classes of methods to determine the preferred design for
implementation from the results of the second level optimization. One class is to determine
the preferred solution directly from the whole set. The other is to determine the preferred
solution from the reduced set of Pareto-optimal solutions. In the following, several methods
are discussed respectively.

Selection according to weight vectors The pseudo-weight vectors of two design objectives
for all Pareto-optimal solutions shown in Table 5.4 have been calculated using equation (5.1).
They are shown in Table 5.5. Then the preferred solution can be selected according to a
designer-preferred weight vector. For example, if a weight value of 40% is prepared for
objective 1 (the vibration amplitude) and a weight value of 60% for the objective 2 (the input
power), Design 4 can be chosen. Design 9 and Design 42 are two extreme solutions among
the obtained non-dominated solutions. The former represents the Pareto-optimal design of a
stepped-horn transducer that consumes the smallest electrical input power among all designs.
The latter represents the Pareto-optimal design of a stepped-horn transducer that has the
largest vibration amplitude.
126 CHAPTER 5

Table 5.5 Values of the pseudo-weight vectors of the objective function for the non-dominated
solutions obtained after the second level optimization for the freely vibrating stepped-horn
transducer
Design w1 w2 Design w1 w2 Design w1 w2

1 0.314 0.686 22 0.448 0.552 43 0.45 0.55

2 0.137 0.863 23 0.504 0.496 44 0.601 0.399
3 0.346 0.654 24 0.507 0.493 45 0.583 0.417
4 0.399 0.601 25 0.473 0.527 46 0.578 0.422
5 0.26 0.74 26 0.482 0.518 47 0.512 0.488
6 0.402 0.598 27 0.491 0.509 48 0.522 0.478
7 0.33 0.67 28 0.496 0.504 49 0.528 0.472
8 0.438 0.562 29 0.493 0.507 50 0.502 0.498
9 0. 1. 30 0.473 0.527 51 0.46 0.54
10 0.367 0.633 31 0.553 0.447 52 0.699 0.301
11 0.375 0.625 32 0.552 0.448 53 0.739 0.261
12 0.34 0.66 33 0.486 0.514 54 0.475 0.525
13 0.337 0.663 34 0.487 0.513 55 0.598 0.402
14 0.417 0.583 35 0.536 0.464 56 0.532 0.468
15 0.42 0.58 36 0.48 0.52 57 0.539 0.461
16 0.435 0.565 37 0.564 0.436 58 0.533 0.467
17 0.352 0.648 38 0.759 0.241 59 0.565 0.435
18 0.387 0.613 39 0.628 0.372 60 0.496 0.504
19 0.426 0.574 40 0.679 0.321 61 0.509 0.491
20 0.463 0.537 41 0.606 0.394 62 0.529 0.471
21 0.464 0.536 42 1. 0. 63 0.47 0.53

It is noticed that the weight vectors are not uniformly distributed, because the final Pareto-
optimal solutions are not distributed evenly in the objective space. Therefore, not all designer-
preferred weight vectors can find corresponding designs in the set of the Pareto-optimal
solutions. In this case, other high level information is needed. It needs to be pointed out that
the pseudo weight of the electrical input power doesn’t increase with the increase of the
thickness of the piezo-stack any more. For example, the total thickness of the piezo-stack for
Design 36 is larger than the total thickness for Design 5 and the weight of the input power for
the former is smaller than that for the latter. This is different from the situation for the
symmetrical transducer. Moreover, in most cases, the larger the number of the piezo-rings is,
the larger the input power.

Selection according to performance criteria Any performance criterion which is not used as
the objective function can be considered as high level information to determine the preferred
design. For the symmetrical transducer, the effective coupling factor k has been considered.
Here, however, another important performance parameter of transducers, namely the
piezoelectric quality number M is considered as high-level information.

As described in chapter 2 and 4, the piezoelectric quality number M presents the phase
descend of the admittance function of the transducer. The larger the value of M is, the higher
RESULTS AND DISCUSSIONS 127

the extent of the phase reserve. A large phase reserve ensures that a resonance frequency
exists in the true sense of the meaning, and the transducer can be driven with zero reactive
power even though the damping of the load may be large. The frequency response of the
admittance function Iˆ Uˆ for the stepped horn can be calculated according to equation (4.73).
Figs. 5.9 and 5.10 show the Bode plot and the Nyquist plot of the admittance function Iˆ Uˆ
for 6 Pareto-optimal designs (Design 8 to Design 13). According to Fig. 5.9, it can be
obtained that Design 8 has a largest phase descend. That is, it has the largest piezoelectric
quality number M. This can be explained more clearly by Fig. 5.10. As stated in section 2.3.1,
the piezoelectric quality number is geometrically defined as the ratio between the diameter
Yr of the locus of Iˆ Uˆ and the offset Yc of its center from the real axis, namely
M = Yr Yc (see Fig. 5.9). Obviously, Design 8 has the largest M and should be preferred for
implementation.

0.0005
|I/U| [A/V]

0.0001 Design8
0.00005
Design9

Design10
0.00001
5.×10
-6 Design11
Design12
19900 19950 20000 20050 20100
Frequency [Hz] Design13

75
Phase[¡ ]
ã

50
25
0
-25
-50
-75
19900 19950 20000 20050 20100
Frequency [Hz]

Fig. 5.9 Admittances Iˆ Uˆ of 6 designs from Table 5.4 as functions of frequencies (the Bode
plot)
128 CHAPTER 5

0.0003
Im( Iˆ Uˆ )
0.0002

0.0001
Design8
Yc Yr
0 Design9

Design10
-0.0001
Design11
-0.0002
Design12
0 0.0001 0.0002 0.0003 0.0004 0.0005 Design13

Re( Iˆ Uˆ )

Fig. 5.10 Admittances Iˆ Uˆ of 6 designs from Table 5.4 as functions of frequencies in the
complex plane (the Nyquist plot)

It is noticed that the 0°-line in Fig. 5.9 is almost tangential to the phase-frequency plots of
Design 9. Correspondingly the real axis in Fig. 5.10 is nearly tangential to the Nyquist plot of
Design 9. That is, the piezoelectric quality number of Design 9 is nearly equal to 2. If the load
damping is large, it is possible that the intersections do not exist any more. The transducer can
not be driven with zero reactive power. Therefore, it is avoided to select Design 9 as the
optimal design of the transducer for ultrasonic bonding and machining.

It can be seen that Design 8 also has the largest distance between resonance and anti-
resonance frequencies. That is, Design 8 has the largest coupling factor. It should be pointed
out that for the Pareto-optimal stepped-horn, increasing the thickness of the piezo-stack
doesn’t mean the coupling factor will increase. Here, the material of the end blocks will affect
the coupling factor. For these 6 Pareto-optimal designs, the designs whose end blocks are
made of brass have better coupling factor than the others.

Similarly, these selection methods can be applied in other Pareto-optimal designs shown in
Table 5.4. Therefore, all solutions can be compared according to high level information and
the preferred design can be determined.

Selection by means of clustering technique There are 63 members in the set of Pareto-
optimal solutions obtained after the second level optimization. If all Pareto-optimal solutions
are evaluated according to high level information, the computation cost is large. In order to
reduce the size of the set of Pareto-optimal solutions, the clustering technique is used here.
Fig. 5.11 shows the results after performing the clustering operation similar to that used for
the symmetrical transducer with a choice of 5 final solutions in the objective space. Tables 5.6
and 5.7 give their corresponding values of the design variables, objective functions and
weights.
RESULTS AND DISCUSSIONS 129

20000
Design 2
17500

1/Amplitude[1/m]
15000 Design 5
12500
Design 10
10000

7500
Design 16 Design 48
5000

2 4 6 8 10
Input Power[VA]
Fig. 5.11 Five final Pareto-optimal solutions after clustering for the stepped-horn transducer

Table 5.6 Values of 5 Pareto-optimal solutions in design variable spaces after clustering for
the stepped-horn transducer

Dt Lb hp Lf1 Lf2
Design ζ Typb Typf N
[mm] [mm] [mm] [mm] [mm]

2 8 0.2 5. 3 3 5 46.5 50.3 2

5 8 0.2 5.6 2 4 4.9 41.9 40.3 2

10 8 0.21 7.4 3 4 5 14.5 41.3 2

16 8 0.2 16.3 1 4 2.4 5. 41.7 2

48 8 0.2 18.8 2 4 0.7 16.8 41.3 6

Table 5.7 Values of objective functions for 5 Pareto-optimal solutions and corresponding
weights for the stepped-horn transducer

Amplitude [µm] Input Power [VA]

Design w1 w2
(Obj1) (Obj2)

2 49.71 0.51 0.14 0.86

5 62.98 0.51 0.26 0.74
10 90.83 1.12 0.37 0.63
16 138.67 2.48 0.43 0.57
48 277.37 10.15 0.52 0.48
130 CHAPTER 5

The configurations of the stepped-horn transducers for the above 5 Pareto-optimal solutions
are shown in Fig. 5.12. It is noticed that the diameter and amplitude transformation factor ζ of
the transducers for all 6 designs are equal to the lower bounds of the design variables Dt
and ζ , respectively. That is, for this two optimization objectives (maximizing the vibration
amplitude and minimizing the input power of the transducer) the lower bounds of design
variables Dt and ζ will affect the optimized results. They should be defined according to
practical design requirements. Design 2 is the minimum input power design, but the vibration
amplitude is the smallest. Design 48 is the maximum vibration amplitude design. However its
input electrical power is the largest. The material types of the back section and horn affect the
length of the whole transducer. Among the 5 designs, the transducer with the back section of
steel and the horn of brass has small length. If a short transducer is expected, the transducer
corresponding to Design 16 is preferred. A finally implemented design can also be selected
from these 5 designs according to a designer-preferred weight vector or other high level
information such as piezoelectric quality number M, coupling factor k or the like. In the
following, the coupling factor k is considered.

Design 2

Design 5

Design10

Design 16

Design 48

5mm PIC181
Steel Aluminium bronze
Titanium Brass

Fig. 5.12 Configurations of the stepped horn transducers without load for 5 Pareto-optimal
solutions after clustering
RESULTS AND DISCUSSIONS 131

As stated before, the effective coupling factor k can be represented by the distance between
the resonance and anti-resonance frequencies for a vibration mode of the transducer. The
resonance and anti-resonance frequencies can be determined by the frequency response of the
admittance function Iˆ Uˆ . Fig. 5.13 gives the frequency responses of the admittance function
Iˆ Uˆ for the 5 designs, where the frequency at which the admittance is maximum corresponds
to the resonance frequency. The frequency at which the admittance is minimum corresponds
to the anti-resonance frequency. Obviously, Design 10 and 16 have larger distances between
the resonance and anti-resonance frequencies. Considering that the length of the transducer
with Design 16 is shorter than that of the transducer with Design 10, therefore, Design 16 is
preferred.

0.001
0.0005
|I/U| [A/V]

0.0001 Design2
0.00005
Design5

0.00001 Design10
5.×10-6
Design16
19900 19950 20000 20050 20100
Frequency[Hz] Design48

75
50
25
]
Phase[¡ ã

-25
-50

19900 19950 20000 20050 20100

Frequency [Hz]

Fig. 5.13 Admittance Iˆ Uˆ as functions of frequencies for 5 Pareto-optimal designs of

stepped-horn transducers without load obtained after clustering.
132 CHAPTER 5

5.3 Discussion of the Results for Stepped-horn Transducers with Load

The non-dominated solutions of the transducers with 2, 4 and 6 piezo-rings under a

mechanical load have been obtained in chapter 4 (refer to Figs. 4.34, 4.35 and 4.36 as well as
Tables 4.23, 4.24 and 4.25). Here they are plotted in Fig.5.14. The non-dominated solutions
were searched for again in these solutions and the results are shown in Fig.5.15 and Table 5.8
Electrical input power[VA]

17.5
15 2 Piezo
12.5
10
4 Piezo
7.5
5
6 Piezo
2.5
0
0 2 4 6 8
1/Mechanical output power[s/Nm]

Fig. 5.14 Three sets of non-dominated solutions of the stepped-horn transducer with load for
2, 4 and 6 piezo-rings
Electrical input power[VA]

17.5
15 2 Piezo
12.5
10
4 Piezo
7.5
5
6 Piezo
2.5
0
0 1 2 3 4 5
1/Mechanical output power[s/Nm]

Fig. 5.15 Non-dominated solutions of the stepped-horn transducer with load for 2, 4 and 6
piezo-rings after the second level optimization
RESULTS AND DISCUSSIONS 133

Table 5.8 Values of design variables and objective functions for the non-dominated solutions
obtained after the second level optimization for the stepped-horn transducer with load

Design
Dt ζ
Lb
Typb Typf
hp Lf1 Lf2
N
P̂m P̂a
[mm] [mm] [mm] [mm] [mm] [Nm/s] [VA]

1 8. 0.86 34.3 3 3 4.6 29.9 24.3 2 0.22 0.3

2 8. 0.96 18.3 4 1 5. 15.3 72.9 2 0.19 0.27
3 8.5 0.88 34.8 4 3 4.5 32.6 12.8 2 0.34 0.43
4 8.8 0.86 35.7 4 4 4.2 33.9 2. 2 0.27 0.36
5 9.2 0.85 28. 3 3 4.7 32.7 29.6 2 0.31 0.41
6 9.4 0.86 41.5 3 1 4. 42.4 16.7 2 0.72 0.85
7 9.7 1. 44.3 1 3 3.3 31.8 20. 2 0.62 0.77
8 10.2 0.98 51.3 2 3 4. 8.7 41.7 2 0.73 0.87
9 10.3 0.79 42.9 3 1 4.1 43.1 14.4 2 1. 1.16
10 10.4 0.8 45. 2 1 4. 32.6 48.9 2 0.51 0.67
11 10.4 0.91 34.8 4 3 3.8 41.8 4. 2 0.84 1.
12 10.5 0.9 34.6 4 1 3.8 42. 17.7 2 1.14 1.31
13 10.5 0.89 34.8 4 3 3.9 41.7 3.9 2 0.87 1.02
14 10.5 0.9 46. 3 1 3.5 31.4 23.4 2 1.04 1.22
15 10.7 0.97 43.7 1 2 3.5 30. 35.4 2 0.36 0.54
16 11. 0.89 44.3 3 1 3.5 31.9 26.1 2 1.25 1.45
17 11.3 0.93 40.8 1 3 3.4 39.3 16.9 2 1.05 1.26
18 11.4 0.99 53.1 2 1 3.9 6.9 53.5 2 1.73 1.93
19 11.7 0.8 40.7 3 1 3.7 39.2 26.6 2 1.49 1.72
20 11.8 0.92 36.7 4 1 3.4 49.2 5.9 2 1.91 2.17
21 11.8 0.81 46.1 3 1 3.5 28.3 30.6 2 1.35 1.58
22 11.9 0.86 41.4 1 1 3.5 37.9 36.7 2 1.39 1.63
23 11.9 0.78 41.7 4 3 3.7 34.2 2. 2 1.26 1.47
24 12. 0.78 42. 3 1 3.7 38.1 26.4 2 1.57 1.82
25 12.8 0.66 47.8 3 1 3.9 42.3 9.1 2 2.37 2.66
26 13.2 1. 40.9 3 3 2.6 36.2 14.2 2 2.2 2.56
27 11.6 0.51 44.7 1 4 3.8 29.8 2. 4 3.12 3.59
28 11.6 0.35 33. 4 3 3.6 33.9 9.7 4 4.27 4.94
29 12.1 0.6 32.9 3 1 3.5 36.1 34.4 4 4.16 4.82
30 12.8 0.69 24.8 3 3 3. 53.5 5.3 4 6.89 7.72
31 12.9 0.79 32. 4 4 4.1 27.8 2. 4 4.84 5.44
32 13. 0.32 39.2 4 3 3.4 28.3 2. 4 6.23 6.97
33 13.2 0.82 29.8 2 4 4.4 28.2 21.8 4 2.89 3.42
34 13.4 0.65 28.4 3 4 3.2 39.3 2. 4 5.69 6.48
35 13.7 0.2 44.9 3 3 3.1 32.9 2. 4 8.79 9.8
36 14. 0.24 51.1 2 4 4.5 27.8 2. 4 6.8 7.51
37 14.1 0.49 34.4 4 3 3.8 28.8 12.5 4 8.66 9.6
38 15. 0.73 31.1 2 4 4.2 26.3 26.5 4 3.38 4.08
39 15. 0.8 33.8 4 2 3.9 33. 21.2 4 5.4 6.17
40 16.3 0.76 37.1 4 3 3.8 27.1 2. 4 16.89 18.06
134 CHAPTER 5

Table 5.8 continued

Design
Dt
ζ
Lb
Typb Typf
hp Lf1 Lf2
N
P̂m P̂a
[mm] [mm] [mm] [mm] [mm] [Nm/s] [VA]

41 8. 1. 58.5 2 1 5. 6.6 4.4 6 0.96 1.11

42 8. 1. 46.6 2 2 4.9 12.8 19.4 6 1.01 1.2
43 8. 1. 46.8 4 2 5. 3.1 2. 6 0.49 0.62
44 8. 1. 52.1 2 1 4.8 13. 3.9 6 1.32 1.51
45 8. 1. 53.7 2 3 5. 5. 8.2 6 1.12 1.29
46 8. 1. 58.5 3 1 4.7 3.4 2. 6 0.59 0.73
47 8.1 0.94 40.1 1 1 5. 5.8 13.6 6 1.46 1.67
48 8.3 0.99 39.9 1 3 5. 5. 11.6 6 1.49 1.7
49 8.5 0.92 54.9 3 2 5. 5. 5.3 6 0.78 0.94
50 8.7 0.95 48.3 2 4 4.9 10.6 7.4 6 1.62 1.87
51 9.2 0.91 45.7 2 3 4.9 8. 14.1 6 2.57 2.9
52 9.3 0.91 41.5 1 1 4.9 7.8 11.8 6 2.48 2.81
53 9.4 0.92 47.7 1 4 4.8 8.9 6. 6 1.98 2.28
54 10. 0.89 48.6 3 1 4.7 8.9 2. 6 2.18 2.5
55 10. 0.85 45.4 2 4 4.9 6.4 16.3 6 2.59 2.98
56 10.5 0.82 45.2 3 4 4.9 5. 7.9 6 2.5 2.87
57 10.5 0.94 27. 3 4 4.7 7.2 16.7 6 3.68 4.17
58 10.6 0.85 43.6 3 1 4.8 5.8 10.2 6 3.26 3.69
59 10.7 0.77 51.9 1 3 4.3 12.3 6.3 6 3.99 4.51
60 10.8 0.97 33.3 3 2 5. 5. 25.4 6 3.17 3.62
61 11.2 0.92 49.6 1 3 4. 7.3 13.3 6 5.28 5.93
62 11.2 0.78 36.5 2 3 4.3 11.1 29.9 6 4.85 5.5
63 11.3 0.91 40.6 3 1 4.8 6.7 9.7 6 4.75 5.31
64 11.3 0.92 28.4 1 1 4.6 8.6 25.8 6 8.58 9.35
65 11.5 0.9 41.9 1 1 4.4 18.5 4.9 6 7.44 8.2
66 11.5 0.92 37.7 3 3 4.8 6.3 10.3 6 5.27 5.88
67 11.8 0.66 45.7 2 3 3.8 15.4 25.6 6 5.64 6.45
68 12. 0.77 34. 2 1 4.1 25.3 31. 6 11.02 11.99
69 12. 0.94 32.4 1 1 4.6 13.1 16.2 6 10.74 11.62
70 12. 0.93 32.8 1 4 4.6 13.3 10.3 6 7.02 7.8
71 12.2 0.96 27.4 3 4 5. 5. 16.2 6 6.84 7.59
72 12.5 0.9 47.3 2 3 3.6 7.6 24.8 6 10.42 11.44
73 12.7 0.92 23.7 2 4 4.9 5.2 30.4 6 6.52 7.31
74 13.2 0.79 49.8 1 1 3.6 8.3 26.1 6 8.84 10.52
75 13.3 0.48 58.1 2 1 3.2 21.5 17.3 6 13.03 14.3
76 13.4 0.93 46.8 2 1 4.1 10.5 21.2 6 17.69 18.73

5.3.1 Analysis of the Results of the Optimization

There are 76 individuals in the final set of Pareto-optimal solutions for the stepped-horn
transducer with load, where 26 individuals come from the set of the non-dominated solutions
for 2 piezo-rings, 14 individuals from the set for 4 piezo-rings and 36 individuals from the set
for 6 piezo-rings. It is noticed that not all diameters of the transducers are equal or nearly
equal to the lower bound of this design variable. Unlike the amplitude transformation ratio ζ
under the condition of no load, here the ζ of most transducers are near to the lower bound of
RESULTS AND DISCUSSIONS 135

this design variable, but rather have values around ζ = 0.8. Some of them are near to 1, which
refers to a constant horn, a special type of the stepped horn. Furthermore, four material types
have been applied for both the back section and horn. The thickness of all piezo-rings is more
than 3mm.

5.3.2 Determination of the Preferred Solution

In the front sections, two classes of methods and some high-level information have been used
to determine the preferred design for implementation. Here the similar methods and high level
information can be adopted. In the following, first, the size of the set of solutions is reduced
by clustering operation. Then, the coupling factor, the piezoelectric quality number and the
power efficiency as high-level information are applied in determining a preferred design.

Reduction of size of the set of non-dominated solutions There are 76 solutions in the set of
Pareto-optimal solutions obtained after the second level optimization. In order to reduce the
size of the set of Pareto-optimal solutions, a clustering technique similar to that used before is
applied here. Fig. 5.16 shows the results with a choice of 5 final solutions in the objective
space after clustering. Tables 5.9 and 5.10 give their corresponding values of the design
variables, objective functions and pseudo weights. Fig. 5.17 shows their respective
configurations.
Electrical input power[VA]

17.5 Design 76
15
Design 75
12.5
Design 74
10
Design 67
7.5
Design 21
5
2.5
0.1 0.2 0.3 0.4 0.5 0.6 0.7
1/Mechanical output power[s/Nm]

Fig. 5.16 Five Pareto-optimal designs obtained after clustering operation for the stepped-
horn transducer with load
136 CHAPTER 5

Table 5.9 Values of 5 Pareto-optimal designs in design variable spaces after clustering for
the stepped-horn transducer with load
Dt Lb hp Lf1
Design ζ Typb Typf Lf2 [mm] N
[mm] [mm] [mm] [mm]
76 13.4 0.93 46.8 2 1 4.1 10.5 21.2 6
75 13.3 0.48 58.1 2 1 3.2 21.5 17.3 6
74 13.2 0.79 49.8 1 1 3.6 8.3 26.1 6
67 11.8 0.66 45.7 2 3 3.8 15.4 25.6 6
21 11.8 0.81 46.1 3 1 3.5 28.3 30.6 2

Table 5.10 Values of objective functions for 5 Pareto-optimal designs after clustering and
corresponding weights for the stepped-horn transducer with load

Mechanical output Electrical Input power

Design w1 w2 Efficiency
power [Ns/m] (Obj1) [VA] (Obj2)
76 17.69 18.73 1 0 0.94
75 13.03 14.3 0.81 0.19 0.91
74 8.84 10.52 0.69 0.31 0.84
67 5.64 6.45 0.59 0.41 0.87
21 1.35 1.58 0.48 0.52 0.85

Selection according to performance criteria For the stepped-horn transducer under the
condition of load, the effective coupling factor k, and piezoelectric quality number M as well
as the power efficiency are important characteristics. They can be considered as high level
information to determine the preferred design from the 5 solutions obtained after clustering.

In order to compare the coupling factors and piezoelectric quality numbers of these 5 designs,
the frequency responses of the admittance functions Iˆ Uˆ for the 5 solutions were calculated
according to equation (4.88). The respective Bode plots and the Nyquist plots are shown in
Figs. 5.18 and 5.19.

Obviously, Design 76 has the largest distance between the resonance and anti-resonance
frequencies among those of the five designs. This is means the coupling factor k of Design 76
is the best. Furthermore, Design 76 has also the largest phase descend (the largest offset of the
center of the Nyquist plot from the real axis). Therefore the piezoelectric quality number M of
Design 76 is the largest among those of these designs. For the transducer operating against
load, the power efficiency is one of the most important characteristics as well. It can be used
as high-level information for determining a preferred solution. The power efficiency of the
RESULTS AND DISCUSSIONS 137

stepped-horn transducer under the condition of load can be readily derived from equations
(4.86) and (4.89). It is given by

Pˆm
λp = (5.2)
Pˆ
a

The power efficiencies of the above 5 Pareto-optimal transducers have been calculated and
the results are shown in Table 5.10. It is obvious that Design 76 has the highest power
efficiency.

Compared with the other designs, Design 76 has the advantages in terms of the coupling
factor, the piezoelectric quality number and the power efficiency. Therefore, it is the preferred
design for implementation.

Design 76

Design 75

Design 74

Design 67

Design 21

5mm

Steel PIC181
Titanium Aluminium bronze

Fig. 5.17 Configurations of the stepped horn transducers with load for 5 Pareto-optimal
solutions after clustering
138 CHAPTER 5

|I/U|[A/V]
0.001
0.0005
Design21

0.0001 Design67
0.00005 Design74

Design75
18000 20000 22000 24000 26000 Design76
Frequency [Hz]

75
50
25
]
Phase[¡ ã

0
-25
-50
-75
18000 20000 22000 24000 26000
Frequency[Hz]

Fig. 5.18 Admittances Iˆ Uˆ of 5 designs in Table 5.9 for the stepped-horn transducers with
load as functions of frequencies (the Bode plot)

0.002
Im( Iˆ Uˆ )
0.0015

0.001

0.0005

-0.0005

-0.001

-0.0015

0 0.0005 0.001 0.0015 0.002 0.0025 0.003 0.0035

Re( Iˆ Uˆ )

Fig. 5.19 Admittances Iˆ Uˆ of the 5 designs in Table 5.9 for the stepped-horn transducers
with load as functions of frequencies in the complex plane (the Nyquist plot)
RESULTS AND DISCUSSIONS 139

5.3.3 Load Characteristics of Stepped-horn Transducers

In the above multiobjective optimization problem of the stepped-horn transducer with load, a
spring-damping load has been assumed. For the given load, the set of Pareto-optimal designs
of the transducer has been obtained by means of MOEAs. In ultrasonic machining and
bonding, piezoelectric transducers operate against loads. It is then a natural question how the
load affects the behavior. In order to study the effect of loads on the resonance performances,
in the following, the damping is not a fixed value, whereas the other parameters are not
changed.

The power efficiency and mechanical output power of the 5 Pareto-optimal transducers as
functions of the damping were calculated according to equations (5.2) and (4.86) and plotted
in Figs. 5.20 and 5.21, respectively. The efficiencies of the 5 solutions increase rapidly in the
range of less than about dL=80Ns/m (for Design 21, Design 67, Design75 and Design 76) as
well as dL=500Ns/m (for Design 74) and decrease with increasing the load damping
gradually. It is noted that the efficiencies have maxima at the region of about 25-80 Ns/m for
Design 21, Design 67, Design75 and Design 76. For Design 74 the range is around 400-500
Ns/m. The mechanical output powers of Designs 67, 75 and 76 increase rapidly in the range
of about less than 40Ns/m and rapidly decrease with the increase of the load damping,
whereas for Designs 21 and 74 the variation of mechanical power is not obvious. It is noticed
that Design 76 has larger power efficiency and mechanical power than the others in a wide
range of the damping load.

If the damping of the transducer matches the load damping, the maximum power efficiency
and mechanical power can be obtained. Generally, for a given load damping, the match of
damping can be achieved by changing the structure of the transducer. It should be pointed out
that the power efficiencies and mechanical powers of 5 Pareto-Optimal designs for the
stepped-horn transducer didn’t arrive at their respective maximum values in the
multiobjective optimization problem before, where the load damping dL is equal to about 202
Ns/m (see section 4.6.4). Since two trade-off objectives must be considered simultaneously, it
is not possible to achieve a match of the transducer to the load damping by changing the
structure of the transducer. However, the electrical input power and mechanical output power
of the 5 designs are Pareto-optimal, respectively.

The load characteristics of other Pareto-optimal transducer designs can be analyzed similarly.
Obviously, load characteristics can be used a high level information to determine the
preferred transducer for implementation.

In this chapter, the optimized results of the optimization problems for Langevin-type
transducers have been evaluated and discussed. In order to search for the preferred solutions
various selection methods and high level information can be applied. The optimized results
show that the performances of the Pareto-optimal transducer are obviously affected by the
material of the end blocks. Titanium and brass are preferred materials. The load has much
influence on the behavior and vibration form of the Pareto-optimal transducer obtained under
140 CHAPTER 5

the condition of no load. In all multiobjective optimization problems discussed here, only
two-objective optimization problem considering the vibration amplitude and electrical input
power at a given resonance frequency as design objectives have been studied. For other
optimal design objectives, similar methods can be used. The methods and procedures for
designing piezoelectric transducers via multiobjective optimization can be applied in other
MOPs.

1
Power efficiency

0.8

0.6
Design21
0.4
Design67
0.2 Design74

0 Design75
0 100 200 300 400 500 Design76
Load damping[Ns/m]

Fig. 5.20 Power efficiencies of 5 designs shown in Table 5.9 as functions of the load damping
for the stepped-horn transducer
Mechanical power[Nm/s]

60

40 Design21

Design67
20
Design74

0 Design75
0 100 200 300 400 500
Load damping[Ns/m] Design76

Fig. 5.21 Mechanical output powers of 5 designs shown in Table 5.9 as functions of the load
damping for the stepped-horn transducer
SUMMARY AND OUTLOOK 141

6 Summary and Outlook

The design of piezoelectric actuators is usually based on single-objective optimization only.

In most practical applications of piezoelectric actuators, however, there exist multiple design
objectives which often are contradictory to each other by their very nature. This dissertation
has studied various multiobjective optimization methods and applied them in the design of
piezoelectric transducers. The main contribution has been to formulate the design of a
piezoelectric transducer as a constrained multiobjective optimization problem involving
continuous and discrete design variables and to find Pareto-optimal solutions using
multiobjective evolutionary algorithms. The determination of the preferred designs using high
level information was also addressed.

As the basis for the optimization of piezoelectric actuators the most important fundamentals
concerning piezoelectric actuators were first described. In order to formulate optimization
problems, the behaviors of piezoelectric actuators were modeled. The fundamental equations
for describing electromechanical behaviors of piezoelectric actuators are piezoelectric
constitutive equations. Four models that can be applied to describe the vibration behavior of
the piezoelectric actuator were described by using the modeling of a longitudinal piezoelectric
actuator as an example. These models are nonparameter models, continuum models, finite
element models and lumped parameter models. Nonparameter models are predominantly
applied in the experimental characterization of dynamic behaviors of piezoelectric actuators.
As no parameters are used in the models, they are not suitable to the optimization problem.
Continuum models can well describe the vibration behavior of piezoelectric actuators with
simple geometry. As material data and geometrical parameters are taken into account in
models, continuum models can be used in the design process for the formulation of
optimization problems. Finite element models can describe the vibration behavior of
piezoelectric actuators with complex geometry and boundary conditions. However, the cost of
numerical computation in the finite element models is notably higher than the cost in
continuum models. They are not well suited for optimization problems occurring in the early
design stages. Lumped parameter models can well describe the vibration behavior of the
piezoelectric actuator operating in the vicinity of one of its resonant frequencies. Their
advantage is that the vibration behavior of the actuator can be described by simple algebraic
equations and it is convenient to formulate the performance criteria for actuators using these
models. Piezoelectric actuators have been widely used in different fields. For different
applications, there exist different design goals. The typical design goals of piezoelectric
actuators for one stroke driving and resonant driving were introduced. The state of the art of
optimization of piezoelectric actuators was reviewed. Although various optimization
problems have been studied in literature, piezoelectric actuator design via multiobjective
optimization was scarcely reported. A latent need for studying the integrated optimal design
procedure of piezoelectric actuators including modeling, mathematical formulation,
optimization algorithms and implementation was identified.
142 CHAPTER 6

The models of the piezoelectric actuators provide the base for formulating MOPs. In
combination with an appropriate optimization method the underlying MOP can be solved. In a
MOP, because of conflict among objectives, it is not always possible find a solution for which
each objective arrives at its optimal value simultaneously. Therefore, the concepts of non-
domination and Pareto-optimality were introduced. The solutions of a MOP were given in the
non-domination or Pareto-optimality sense. To find solutions belonging to a better non-
dominated front and simultaneously to maintain diversity in the non-dominated solutions are
two important goals in a multiobjective optimization.

Most traditional multiobjective optimization methods convert the MOP into a single objective
problem and then solve the problem using widely developed methods for single objective
optimization. Some commonly used traditional multiobjective optimization methods
including weighted sum methods, ε-constraint methods, weighted metric methods and value
function methods were described in brief. These methods require some problem knowledge
before optimization is performed. Some techniques may be sensitive to the shape of the
Pareto-optimal front. Moreover, they require several optimization runs to obtain an
approximation of the Pareto-optimal set. Evolutionary algorithms (EAs), on the other hand,
are able to find multiple Pareto-optimal solutions in a single simulation run due to their
population-approach. They are well suited for multiobjective optimization and the problems
with discrete variables. By using a diversity-preserving mechanism, moreover, MOEAs can
also find widely different Pareto-optimal solutions. The basic structure of a MOEA is similar
to a single-objective GA. Since a scalar fitness is needed for reproduction in an EA, the
essence of a MOEA is how to assign a scalar fitness to each individual from multiple
objective functions. According to different fitness assignment methods, various MOEAs have
been developed over the past decades. In this dissertation, the multi-objective GA (MOGA),
the non-dominated sorting GA (NSGA), the elitist NSGA (NSGA-II), the strength Pareto EA
(SPEA) and the improved SPEA (SPEA2) were studied. In MOGA and NSGA, the fitness of
each individual is first assigned according to their respective Pareto-ranking methods. Then
the shared fitness is calculated in order to maintain diversity among non-dominated solutions.
Proportionate selection operators are used in the two GAs. In the two GAs mentioned above,
no elite-preservation is performed. NSGA-II, SPEA and SPEA2 are three MOEAs with elite-
preservation. In NSGA-II, non-dominated sorting is first performed among a population of
size 2N consisting of offspring and parent populations. A new population of size N is then
filled by individuals first according to their ranks and then according to the crowding
distances. This provides an elite-preservation. The crowded tournament selection operator is
used for obtaining the mating pool. In SPEA, the elites are preserved by using an external
population. The fitness (strength) of each elite individual and the fitness of each individual in
the current population are assigned according to respective expressions related to the number
of dominated solutions. A clustering technique is used to reduce the size of external
population when its size is larger than the specified size. A binary tournament selection
operation is performed in this MOEA. SPEA2 is an improved version of SPEA. An improved
fitness assignment scheme, a nearest neighbor density estimation technique and a new archive
SUMMARY AND OUTLOOK 143

truncation method are used in this MOEA. The methods for handling constraints in MOPs
were studied. The penalty function approach and the constrained tournament method are two
commonly used methods.

Based on modeling and multiobjective optimization techniques, multiobjective optimal design

problems of Langevin-typed piezoelectric transducers were studied. For the transducers used
in ultrasonic bonding and machining, the most important design criteria are the resonance
frequency, electrical input power, mechanical output power, coupling factor, power
efficiency, efficiency, mechanical quality factor, piezoelectric quality number, volume of
piezoelectric materials and mass and size of the transducer. These performance criteria can be
considered as objective functions in MOPs of transducers. In this dissertation, two-objective
optimization problems were studied. As the influence of prestress on the resonance
performance of the transducer can not be modeled in theory, experimental study was applied
in determining optimal prestress. The coupling factor and piezoelectric quality number were
considered as design objectives. Lumped parameter models and parameter identification were
used in the formulation of optimization problems. The Pareto-optimal pre-stresses for the
freely vibrating transducer and the transducer with a mechanical load were found respectively.
Two kinds of Langevin-typed transducers (the symmetrical transducer as well as the
transducer with a stepped horn) were designed via multiobjective optimization. The transfer
matrix method based on continuum models was applied in formulations of MOPs. For the
MOPs of freely vibrating transducers, the minimum electrical input power and the maximum
resonant amplitude were considered as two design objectives. For the MOPs of transducers
with a mechanical load, the maximum mechanical output power and the minimum electrical
input power were considered as design objectives. In all problems, an equality constraint is
used to assure that the transducer operates at the specified resonance frequency and several
inequality constraints are used to specify the ranges of the input power, input current and
dimensions of the transducer respectively. Each MOP of the freely vibrating transducer with 2
piezoelectric rings was solved by using MOGA, NSGA, NSGA-II and SPEA (SPEA2), where
penalty functions were used to transfer constrained MOPs into non-constrained MOPs. Each
MOP of the freely vibrating transducer with 2 piezoelectric rings was also solved by using
NSGA-II where constrained tournament method instead of the penalty function approach was
used. Unlike the common NSGA-II, the crowded tournament selection operation was
performed only for creation of the new population. In MOGA and SPEA (SPEA2), the
niching strategy was performed in the objective space. The other MOEAs used parameter-
space niching. As the design variables include continuous and discrete types, a mixed coding
scheme was used in each MOEA. The blended crossover (BLX-α) was applied to the
continuous variables. For the discrete variables a single-crossover was used. The mutation
operation was only applied to discrete variables.

Optimization results were evaluated in terms of the extent of the obtained non-dominated
solutions close to the true Pareto-optimal front and the diversity in the obtained non-
dominated solutions. For MOPs of symmetrical transducers, the SPEA did better than the
others relatively. It was then applied in the MOPs of symmetrical transducers with 4 and 6
144 CHAPTER 6

piezoelectric rings, respectively. For MOPs of the transducer with a stepped horn, NSGA2
using the penalty function method did better than the others. It was then used to solve the
MOPs of the transducer with a stepped horn. In each MOP, after all non-dominated solutions
for transducers with 2, 4 and 6 piezoelectric rings were obtained, the second level
optimization was performed and Pareto-optimal solutions were searched among all obtained
non-dominated solutions. In the practical design of piezoelectric transducers, generally, only
one solution is needed to be implemented finally. Methods for determining preferred solutions
from obtained Pareto-optimal solutions were studied. The preferred solution was chosen
according to the specified weight vector and the performance criteria which were not used as
design objectives, respectively. The coupling factor, the piezoelectric quality number and
power efficiency as high-level information have been applied in determining the preferred
designs. In order to reduce the size of obtained Pareto-optimal solutions, a clustering
technique was applied. The NSGA2 using the penalty function method was also applied in the
MOP of the stepped-horn transducers with a mechanical load. The second level optimizations
and clustering were performed .The preferred solutions were determined using similar
methods. The load characteristics were discussed.

This dissertation has presented an integrated procedure for piezoelectric transducer design via
multiobjective optimization methods. The integrated procedure can be applied in other two-
objective or Mutiobjective optimization problems of piezoelectric actuators. In order to obtain
better optimized results, further investigations can be performed in two aspects. First, the
quality of models of piezoelectric actuators can be improved if fewer simplifications are used.
In this dissertation, some simplifications have been introduced into the modeling. For
example, the bolt was not modeled as a separate element, but the characteristics of the bolt
were incorporated in the backing and front rods. The actual piezoelectric rings were modeled
as piezoelectric discs. The electrodes were not taken into account in the model. The nonlinear
characteristic of piezoelectric material under high electrical field strength was not considered.
Second, in addition to MOEAs used here there exist other MOEAs and other multiobjective
optimization methods such as set oriented subdivision algorithms (see [DSH04] and
[SMD03]) and multi-objective particle swarm optimization (MOPSO) methods [MT03]. They
can also be applied to MOPs of piezoelectric actuators.
APPENDIX A 145

Appendix A

Derivation of Individual Transfer Matrix

In this section, the expressions of the transfer matrices of the fundamental block elements are
derived using the analytical models based on rod theory, respectively. In the following
expressions, all material moduli are considered as real values. If material loss is taken into
account, complex moduli can be applied instead of real moduli.

Mechanical Rod Elements The vibration behavior of a mechanical rod element (see Fig.
4.12) can be described by the one-dimensional wave equation
2
u&&( z , t ) = cm u" ( z , t ) (A.1)

with motion and force boundary conditions. Where cm is the wave propagation speed

E
cm = (A.2)
ρ

The axial force and vibration velocity in the rod element are
F ( z , t ) = Ab ⋅ T ( z , t ) = Ab ⋅ Eu ' ( z , t ) (A.3)

v( z , t ) = u& ( z , t ) (A.4)

There are the following motion and force boundary conditions

u& (0, t ) = ve (t ), u& ( Lb , t ) = va (t ) (A.5)

Ab Eu ' (0, t ) = Fe (t ), Ab Eu ' ( Lb , t ) = Fa (t ) (A.6)

As the complex expressions will be used in the computation, the velocity and force quantities
at both sides of the rod element for harmonic vibrations are written as

v e (t ) = vˆ e e jΩt , v a (t ) = vˆ a e jΩt , (A.7)

F e (t ) = Fˆ e e jΩt , F a (t ) = Fˆ a e jΩt , (A.8)

The above boundary value problem can be solved by means of the separation-of-variables
method described in chapter 2 for the case of harmonic vibrations. That is

( )
u ( z , t ) = C1e jkm z + C 2 e − jk m z e jΩt (A.9)

where k m = Ω cm is the wave number.

146 APPENDIX A

Substituting the equation (A.9) into the equations (A.5) and (A.6) and considering the
equations (A.7) and (A.8) yield the following transfer matrix relation

 Ω 
 a 1
vˆ  (
e − jk m Lb + e jkm Lb ) Ab Ek m
( )
e jk m Lb − e − jk m Lb  ˆ
 ve 
 Fˆ  = 2  A Ek ⋅ ˆ  (A.10)
 a (
 b m e jkm Lb − e − jkm Lb ) ( )
e − jkm Lb + e jk m Lb  F e 
 Ω 

Using Euler’s equation this may be written as

 jΩ 
cos(k m Lb ) sin( k m Lb ) ˆ
 vˆ a   Ab Ek m  ve 
 Fˆ  =  jA Ek ⋅ ˆ  (A.11)
 a   b m sin( k L ) cos(k m Lb )   e 
F
 Ω 
m b

Therefore, the transfer matrix of the mechanical rod element is obtained as follows

 jΩ 
 cos(k m Lb ) sin(k m Lb )
1 Ab Ek m
Am =   (A.12)
2  jAb Ek m
sin(k m Lb ) cos(k m Lb ) 
 Ω 

Piezoelectric Elements The transfer model matrix for a piezoceramic element can be derived
by means of the method described in chapter 2. Here the brief derivation is given.

According to equations (2.17) and (2.18) in chapter 2 the axial force and the current applied in
the piezoceramic element (see Fig. 4.13) can be described as follows

F ( z , t ) = Ap ⋅ T ( z , t ) = Ap c33E u ' ( z , t ) + Ap e33φ ' ( z , t ) (A.13)

I ( z , t ) = A p ⋅ D& ( z , t ) = e33 u& ' ( z , t ) − ε 33S φ&' ( z , t ) (A.14)

Besides mechanical boundary quantities shown in equations (A.7) and (A.8), for the
piezoceramic element there exist electrical boundary quantities

U (t ) = Uˆ e jΩt , I (t ) = Iˆe jΩt (A.15)

According to equations (2.27) and (2.28) the governing equations for describing the vibration
behaviors of the piezoceramic element are
2
u&&( z , t ) = c p u" ( z , t ) (A.16)

e33
ϕ "( z, t ) = u" ( z , t ) (A.17)
ε 33S
APPENDIX A 147

According to equations (2.33) and (2.34), solutions of the above governing equations can be
written as
jk p z − jk p z
u ( z , t ) = (C1e + C2 e )e jΩt (A.18)

 e33 jk p z − jk p z 
ϕ ( z, t ) =  S
(C1e + C2 e ) + C3 z + C 4  e jΩt (A.19)
ε 33 

There are the following mechanical and electric boundary conditions (see Fig. 4.13)
u& (0, t ) = ve (t ), u& (h p , t ) = va (t ) (A.20)

Ap c33E u ' (0, t ) + Ap e33ϕ ' (0, t ) = Fe (t ), Ap c33E u ' (h p , t ) + Ap e33ϕ ' (h p , t ) = Fa (t ) (A.21)

Ap e33u& ' (0, t ) − Apε 33S ϕ& ' (0, t ) = I (t ), ϕ (0, t ) − ϕ (h p , t ) = U (t ) (A.22)

Substituting the equations (A.18) and (A.19) into (A.20), (A.21) and (A.22) it follows
jΩ(C1 + C 2 ) = vˆ e (A.23)
jk p h p − jk p h p
jΩ(C1e + C2 e ) = vˆ a (A.24)
2
e33
jk p Ap c33E (C1 − C 2 ) + jk p Ap (C1 − C 2 ) + Ap e33C3 = Fˆ e (A.25)
ε 33S
2
− jk p h p e33 − jk p h p
) + Ap e33C3 = Fˆ a
jk p h p jk p h p
jk p Ap c33E (C1e − C2e ) + jk p Ap S
(C1e − C2e (A.26)
ε 33

− jApε 33S ΩC3 = Iˆ (A.27)

e33 e33 jk p h p − jk p h p
S
(C1 + C 2 ) − S
(C1e + C2 e ) − C3 h p = Uˆ (A.28)
ε 33 ε 33

Eliminating C1, C2, and C3 and getting v̂ a , F̂ a and Iˆ in terms of v̂ e , F̂ e and Û from the above
equations, the transfer matrix relation can be obtained as follows

 vˆ a   A11 A13p   vˆ e 
p
A12p
 Fˆ  =  A p A22p

A23p   Fˆ e  (A.29)
 a   21
 Iˆ   A31p A32p A33p   Uˆ 

where
2 2
h p k p (e33 + c33E ε 33S ) cos(h p k p ) − e33 sin(h p k p )
A11p = (A.30)
h p k p (e332 + c33E ε 33S ) − e332 sin(h p k p )
148 APPENDIX A

jΩh p ε 33S sin(h p k p )

A12p = (A.31)
Ap h p k p (e332 + c33E ε 33S ) − Ae33
2
sin(h p k p )

p
jΩe33ε 33S sin(h p k p )
A =13 (A.32)
h p k p (e332 + c33E ε 33S ) − e33
2
sin(h p k p )

p
Ap k p (e332 + c33E ε 33S )(−2e33
2
+ 2e332 cos(h p k p ) + h p k p (e332 + c33E ε 33S ) sin( h p k p ))
A =21 2
(A.33)
jΩε 33S (h p k p (e33 + c33E ε 33S ) − e332 sin(h p k p ))
2 2
p
h p k p (e33 + c33E ε 33S ) cos(h p k p ) − e33 sin( h p k p )
A = 22 2 2
(A.34)
h p k p (e33 + c33E ε 33S ) − e33 sin(h p k p )

p
Ap e33 k p (e332 + c33E ε 33S )(−1 + cos(h p k p ))
A =23 2
(A.35)
h p k p (e33 + c33E ε 33S ) − e332 sin(h p k p )

p
Ap e33 k p (e332 + c33E ε 33S )(−1 + cos(h p k p ))
A =
31 2 2
(A.36)
h p k p (e33 + c33E ε 33S ) − e33 sin(h p k p )

p
jΩe33ε 33S sin(h p k p )
A =
32 2 2
(A.37)
h p k p (e33 + c33E ε 33S ) − e33 sin(h p k p )

p
jΩAp k p ε 33S (e332 + c33E ε 33S )
A =33 2
(A.38)
h p k p (e33 + c33E ε 33S ) − e332 sin(h p k p )

Similarly, (A.29) may be written as

 X a   A Pm A Pem   X e 
 ˆ  =  Pme   (A.39)
 I   A A Pe   Uˆ 
APPENDIX B 149

Appendix B

Developed Programs

B.1 Multiobjective Optimization of Symmetrical Langevin Transducers

1. Optimal design of symmetrical Langevin transducers using MOGA

2. Optimal design of symmetrical Langevin transducers using NSGA

3. Optimal design of symmetrical Langevin transducers using SPEA

4. Optimal design of symmetrical Langevin transducers using NSGA-II

B.2 Multiobjective Optimization of Stepped-horn Transducers

1. Optimal design of stepped-horn transducers without load using MOGA

2. Optimal design of stepped-horn transducers without load using NSGA

3. Optimal design of stepped-horn transducers without load using SPEA2

4. Optimal design of stepped-horn transducers without load using NSGA-II

5. Optimal design of stepped-horn transducers with load using NSGA-II

BIBLIOGRAPHY 151

Bibliography

[Abb98] Abboud, N. N.; Wojcik, G. L.; Vaughan, D. K.; Mould, J.; Powell, D. J. and
Nikodym, L.: Finite element modelling for ultrasonic transducers. Proceedings
of the International Symposium on Medical Imaging, SPIE, 1998
[Ans01] ANSYS 6.0 User Manual, 2001
[Bak85] Baker, J. E.: Adaptive selection methods for genetic algorithms. In Proceedings
of an International Conference on Genetic Algorithms and Their Application,
pp.101-111, 1985
[Bäc96] Bäck, T.: Evolutionary Algorithms in Theory and Practice. Oxford University
Press, 1996, ISBN 0-19-509971-0
[BCJ64] Berlincourt, D. A.; Curran, D. R. and Jaffe, H.: Piezoelectric and piezomagnetic
materials and their function in transducers. Chapter 3 in Mason, W.P.: Physical
Acoustics. Vol. 1A, 1964
[Bec99] Beckers, M.: topology optimization using a dual method with discrete variables.
Structural and Multidisciplinary Optimization, Vol.17, No.1, pp.14-24, 1999
[BhF02] Bharti, S. and Frecker, M.I.: Topology optimization and detailed finite element
modeling of piezoelectric actuators: effect of external loads and detail geometers
on actuator output. Proceedings of SPIE, Vol.4693, pp.124-135, 2002
[BK88] Bendsøe, M. P. and Kikuchi, N.: Generating optimal topologies in optimal
design using a homogenization method. Computer Methods in Applied
Mechanics and Engineering, Vol.71, pp.197-224, 1988
[BMF00] Barboni, R.; Mannini, A.; Fantini, E. and Gaudenzi, P.: Optimal plancement of
PZT actuators for the control of beam dynamics. Smart Materials and Structures,
Vol.9, pp.110-120, 2000
[BPF91] Bangviwat, A.; Ponnekanti, H. and Finch, R.D.: Optimizing the performanece of
piezoelectric drivers that use stepped horns. Journal of the Acoustic Society of
America, Vol.90, pp.1223-1229, 1991
[Brö02] Brökelmann, M. : Modellierung und Simulation piezoelektrischer Aktoren zum
resonanten Betrieb bei nichtlinearen mechanischer Randbedingungen.
Diplomarbeit, Heinz Nixdorf Institut, Universität Paderborn, Mechatronik und
Dynamik, 2002
[BS03] Bendsøe, M. P. and Sigmund, O. : Topology Optimization – Theory, Methods
and Applications. Springer-Verlag Berlin Heidelberg, 2003, ISBN 3-540-42992-
1
[Cer00] Monolithic Multilayer Actuators Operation and Applications. CeramTec AG,
Lauf, Germany, 2000. [Link]
152 BIBLIOGRAPHY

[CF00] Canfield, S. and Frecker, M.: Topology optimization of compliant mechanical

amplifiers for piezoelectric actuators. Struktural and Multidisciplinary
Optimization, Vol.20, pp.269-279, 2000
[CL87] Crawley, E. F. and Luis, J.: Use of piezoelectric actuators as elements of
intelligent structures. AIAA Journal, Vol.25, pp.1373-1385, 1987
[Coh85] Cohon, J. L.: Multicriteria programming: Brief review and application. Design
Optimization, pp.163-191. Academic Press, New York, 1985
[Cra94] Crawley, E. F.: Intelligent structures for aerospace: a technology overview and
assessment. AIAA Journal, Vol.32, No.8, pp.1689-99, 1994
[CSe94] Chattopadhyay, A. and Seeley, C.E.: A multiobjective optimization for the
design of rotating composites box beams with discrete piezoelectric actuators.
Proceedings-SPIE the International Society for Optical Engineering, Vol.2190,
pp.38-51, 1994
[CS94] Chattopadhyay, A. and Seeley, C.E.: A simulated annealing technique for
multiobjective optimization of intelligent structures. Smart Materials and
Structures, Vol.3, pp.98-106, 1994
[CSJ99] Chattopadhyay, A.; Seeley, C. E. and Jha, R.: Aeroelastic tailoring using
piezoelectric actuation and hybrid optimization. Smart Materials and Structures,
Vol.8, pp.83-91, 1999
[CTS02] Chee, C.; Tong, L. and Steven, G. P.: Piezoelectric actuator orientation
optimization for static shape control of composite plates. Composite structures,
Vol.55, pp.169-184, Elsevier, 2002
[Cul96] Culshaw, B.: Smart structures and materials. Artech House, INC., 1996, ISBN 0-
89006-681-7
[Deb02] Deb, Kalyanmoy: Multi-Objective Optimization using Evolutionary Algorithms.
John Wiley & Sons, Ltd, 2002, ISBN 0-471-87339-X
[DI483] Leitfaden zur Bestimmung de dynamischen Eigenschaften von piezoelektrischer
Keramik mit hohem elecktromechanischen Koppelfaktor. DIE IEC 483, Beuth
Verlag GmbH, Berlin, 1988
[DI324] Piezoelektrische Eigenschaften von keramischen Werkstoffen und
Komponenten. DIN EN 50324, VDE VERLAG GMBH, 2002
[DLL00] Du, H.; Lau, G. K.; Lim, M. K. and Qui, J.: Topological optimization of
mechanical amplified for piezoelectric actuators under dynamic motion. . Smart
Materials and Structures, Vol.9, pp.788-800, 2000
[DSH04] Dellnitz, M.; Schütze, O. and Hestermeyer, T.: Covering Pareto Sets by
Multilevel Subdivision Techniques. Journal of Optimization, Theory and
Applications, 2004
BIBLIOGRAPHY 153

[ES93] Eschelman, L. J. and Schaffer, J. D.: Real-coded genetic algorithms and interval
schemata. In Foundations of Genetic Algorithms 2, pp.187-202, 1993
[FF93] Fonseca, C. M. and Fleming, P. J.: Genetic algorithm for multiobjective
optimization: Formulation, discussion, and generalization. In Proceedings of
Fifth International Conference on Genetic Algorithms, pp.416-423, 1993
[Fu04] Fu, B.; Hemsel, T. and Wallaschek, J.: Model-based Diagnosis for Sandwiched
Ultrasonic Transducers. In: Proceedings of the 18th International Congress on
Acoustics, ICA 2004, 4-9 April 2004, Kyoto, Japan Bd. 3, 2004, pp. 2243-2246
[GCh00] Gen, M. and Cheng, R.: Genetic algorithms and engineering optimizaiton.
Wiley-Interscience Publication, USA, 2000, ISBN 0-471-31531-1
[Gen95] Genta, G.: Vibration of structures and machines: practical aspects. Springer-
Verlag, New York, 1995, ISBN 0-387-94403-6
[Gol89] Goldberg, D.: Genetic Algorithms in Search, Optimization and Machine
Learning, Addison-Wesley, Reading, MA, 1989
[GR87] Goldberg, D. and Richardson, J.: Genetic algorithm with sharing for multi-modal
function optimization, in Proceedings of the 2nd International Conference on
Genetic Algorithms, Lawrence Erlbaum Associates, Hillsdale,N.J.,pp41-49,1987
[Hem01] Hemsel, T. : Untersuchung und Weiterentwicklung linearer piezoelecktrischer
Schwingungsantriebe. Dissertation, Heinz Nixdorf Institut, Universität
Paderborn, 2001, ISBN 3-935433-10-7
[Hen01] Henke, A.: Modellbildung, Simulation und Optimierung piezoelektrischer
Stellsysteme. Dissertation, Heinz Nixdorf Institut, Universität Paderborn, 2001
[HKL98] Hemsel, T.; Krome, J. W.; Littmann, W. and Wallaschek, J.: Modellierung
piezoelektrischer Stapel- und Vielschichtaktoren für dynamischen Betrieb. 2.
Polnisch-deutscher Workshop: Werkzeuge der Mechatronik, Ilmenau, 1998,
ISBN 3-00-002556-1, pp.224-233
[HL99] Han, J. H. and Lee, I.: Optimal placement of piezoelectric sensors and actuators
for vibration control of a composite plate using genetic algorithms. Smart
Materials and Structures, Vol.8, pp.257-267, 1999
[HLH03] Hsu, Y.; Lee, C. and Hsiao, W.: Optimizing piezoelectric transformer for
maximum power transfer. Smart Materials and Structures, Vol.12, pp.373-383,
2003
[HM03] Halim, D. and Moheimani, S.O.: An optimization approach to optimal placement
of collocated piezoelectric actuators and sensors on a thin plate. Mechatronics,
vol.13, pp.27-47, 2003
[Hul73] Hulst, A. P.: On a family of high-power transducers. In Proceeding of Ultrasonic
International Conference, pp.285-294, 1973
154 BIBLIOGRAPHY

[IE87] IEEE Standard on Piezoelectricity. American Standards National Institut, New

York, 1987
[Ike96] Ikeda, T.: Fundamentals of Piezoelectricity. Oxford University Press, 1996,
ISBN 0-19-856339-6
[JC71] Jaffe, B. and Cook, W. R.: Piezoelectric Ceramics. Academic Press, 1971, ISBN
0-12-379550-8
[JH94] Joines, J.A. and Houck, C. R.: On the use of nonstationary penalty functions to
solve nonlinear constrained optimization problems with GAs. In Proceedings of
the International Conference on Evolutionary Computation, pp.579-584
[JNL00] Jiang, T. Y.; Ng, T. Y. and Lam, K. Y.: Optimization of a piezoelectric ceramic
actuators. Sensors and Actuators, Vol.84, pp.81-94, 2000
[JVG99] Jiménez, F.,Verdegay, J. L. and Goméz-Skarmeta, A. F.: Evolutionary
techniques for constrained multiobjective optimization problems. In proceedings
of the workshop on Multi-Criterion Optimization Using Evolutionary Methods
held at Genetic and Evolutionary Computation Conference (GECCO-1999), pp.
115-116
[Kro99] Krome, J. W.: Modelle zur Untersuchung des Schwingungsverhaltens von
Statoren für piezoelektrische Ultraschallwanderwellenmotoren. Dissertation,
Heinz Nixdorf Institut, Universität Paderborn, HNI-Verlagsschriftenreihe, Bd.59,
1999, ISBN 3-931466-58-2
[LDG00] Lau, G. K.; Du, H.; Guo, N. and Lim, M. K. : Systematic design of
displacement-amplifying mechanisms for piezoelectric stacked actuators using
topology optimization. Journal of Intelligent Material Systems and Structures,
Vol.11, pp.685-695, 2000
[LeI75] Lenk, A. and Irrgang, B.: Elektromechanische Systeme. Band 2:Systeme mit
verteilten Parametern. VEB Verlag Technik, Berlin, 1975
[Len75] Lenk, A.: Elektromechanische Systeme. Band 1: Systeme mit Konzentrierten
Parametern. VEB Verlag Technik, Berlin, 1975
[Ler90] Lerch, R.: Simulation of piezoelectric devices by two- and three-dimensional
finite elements. IEEE Transactions on Ultrasonics, Ferroelectrics and Frequency
Control, Vol.37, pp.233-247, 1990
[Lit03] Littmann, W.: Piezoelektrische, resonant betriebene Ultraschall-
Leistungswandler mit nichtlinearen mechanischen Randbedingungen.
Dissertation, Heinz Nixdorf Institut, Universität Paderborn, 2003, ISBN
[LOM00] Li, Y.; Onoda, J. and Minesugi, K.: Simultaneous optimization of piezoelectric
actuator placement and feedback for vibration suppression. Acta Astronautica,
Vol.50, No.6, pp.335-341, Elsevier, 2002
[Mie99] Miettinen, K.: Nonlinear multiobjective optimization. Kluwer Academic
Publishers, 1999, ISBN 0-7923-8278-1
BIBLIOGRAPHY 155

[MT03] Mostaghim, S. and Teich, J: Strategies for finding good local guides in Multi-
objective Particle Swarm Optimization (MOPSO). Proceedings of IEEE Swarm
Intelligence Symposium, April 24-26, 2003, Indianapolis, Indiana, USA
[Nep73] Neppiras, E. A.: the pre-stressed piezoelectric sandwich transducer. In
Proceeding of Ultrasonic International Conference, pp.295-302, 1973
[OK95] Osyczka, A. and Kundu, S.: A new method to solve generalized multicriterion
optimization problems using genetic algorithm, Structual Optimization, Vol.10,
No.2, pp.94-99,1995
[OK96] Osyczka, A. and Kundu, S.: A modified distance method for multicriterion
optimization problems using genetic algorithm, Computers and Industrial
Engineering, Vol.30, No.4, pp.871-882,1996.
[Phi88] Valvo Unternehmenbereich Bauelemente der Philips GmbH: PIEZOXIDE
(PXE) Eigenschaften und Anwendungen. [Link] Hüthig Verlag Gmbh,
Heidelberg, 1988, ISBN 3-7785-1755-4
[Phi91] Waanders, J. W.: Piezoelectric Ceramics-Properties and Applications. Philips
Components, Eindhoven, 1991
[Pre02] Preumont, A.: Responsive Systems for Active Vibration Control. Kluwer
Academic Publishers, Dordrecht, 2002, ISBN 1-4020-0898-8
[PRF03] Pons, J. L.; Rodriguez, H.; Fernández, J. F.; Villegas, M. and Seco, F. :
Parameter optimization of ultrasonic motors. Sensors and actuators, Vol.107,
pp.169-182, Elsevier, 2003
[RTS01] Ray, T.; Tai, K. and Seow, K. C.: An evolutionary algorithm for multiobjective
optimization. Engineering Optimization, Vol.33, 399-424, 2001
[Rus95] Ruschmeyer, K.: Piezokeramik- Grundlagen, Werkstoffe, Applikationen.
Renningen-Malmsheim, Expert Verlag, 1995, ISBN 3-8169-1152-8
[Sat03] Sattel, T.: Mechatronisches System zur schwingungsüberlagerten Bearbeitung
von Baustoffen. III . Quartalsbericht 2003
[SC93] Seeley, C.E. and Chattopadhyay, A.: The development of an optimization
procedure for the design of intelligent structures. Smart Materials and Structures,
Vol.2, No.2, pp.135-146, 1993
[Sch96] Schäufele, A.: Ferroelastische Eigenschaften von Blei-Zirkoat-Titanat-
Keramiken. Dissertation, VDI Verlag Gmbh, Düsseldorf, 1996, ISBN 3-18-
344505
[SD95] Srinivas, N. and Deb, K.: Multi-Objective function optimization using non-
dominated sorting genetic algorithms. Evolutionary Computation, Vol. 2, pp.
221-248, 1995
156 BIBLIOGRAPHY

[Set02] Setter, N.: Piezoelectric materials in devices: extended reviews on current and
emerging piezoelectric materials, technology, and applications. EPFL, Swiss
Federal Inst. of Technology, 2002, ISBN 2-9700346-0-3
[Sig97] Sigmund, O.: On the design of compliant mechanisms using topology
optimization. Mechanics of Structures and Machines, Vol. 25, pp.493-524, 1997
[SK99] Silva, E. C. N. and Kikuchi, N.: Design of piezoelectric transducers using
topology optimization. Smart Materials and Structures, Vol.8, pp.350-364, 1999
[SM95] Statnikov, R. B. and Matusov, J.: Multicriteria Optimization and Engineering.
Chapman & Hall, 1995, ISBN 0-412-99231-0
[SMD03] Schütze, O.; Mostaghim, S.; Dellnitz, M. and Teich, J.: Covering Pareto Sets by
Multilevel Evolutionary Subdivision Techniques. In Proceeding of Evolutionary
Multi-Criterion Optimization: Second International Conference, EMO 2003,
Faro, Portugal, April 8-11, pp. 118-132, 2003
[SNF98] Silva, E.; Nishiwaki, S.; Fonseca, J.; Crumm, A; Brady, G.; de Espinosa, F.;
Halloran, J. and Kikuchi, N.: Topology optimization applied to the design of
piezocomposite materials and piezoelectric actuators. Proceedings of SPIE,
Vol.3323, pp.216-227, 1998
[SNK00] Silva, E.; Nishiwaki, S. and Kikuchi, N.: Topology optimization design of
flextensional actuator. IEEE Transaction on Ultrasonics, Ferroelectrics, and
Frequency Control. Vol.47, No.3, pp.657-671, 2000
[SWW93] Sadri, A.M.; Wright, J.R. and Wynne, R.J.: Modelling and optimal placement of
piezoelectric actuators in isotropic plates using genetic algorithms. Smart
Materials and Structures, Vol.8, pp.490-498, 1993
[Uch97] Uchino, K.: Piezoelectric Actuators and Ultrasonic Motors. Kluwer Academic
Publishers, USA, 1997, ISBN 0-7923-9811-4
[Uch04] Uchino, K.: Materials, design, drive/control, modelling and application. In
Proceeding of Actuator 2004, 9th International Conference on New Actuators,
14-16 June 2004, Bremen, Germany, pp.38-47
[Wal00] Wallaschek, J.: Systementwurf piezoelektrischer Aktoren. VDI Bildungswerk,
Mechatronik Workshop, Düsseldorf , 2000
[WH00] Wallaschek, J. and Hemsel, T.: Piezoelektrische Werkstoffe und ihre technische
Anwendung. Heinz Nixdorf Institut, Universität Paderborn, Mechtronik und
Dynamik, 2000
[YV96] Yousefi-Koma, A. and Vukovich, G.: Shape and location optimization of
piezoelectric actuators in active control systems. Proceedings of the ASME
Aerospace Division, AD-vol.52, pp.559-566, 1996
[Zit99] Zitzler, E.: Evolutionary algorithms for multiobjective optimization: methods
and applications. Dissertation, Technische Wissenschaften ETH Zürich, Nr.
13398, Shaker Verlag, 1999, ISBN 3-8265-6831-1
BIBLIOGRAPHY 157

[ZLG00] Zhang, H.W.; Lennox, B.; Goulding, P.R. and Leung, A.Y.T.: A float-encoded
genetic algorithm technique for integrated optimization of piezoelectric actuator
and sensor placement and feedback gains. Smart Materials and Structures, Vol.9,
pp.552-557, 2000
[ZLT02] Zitzler, E.; Laumanns, M. and Thiele, L.: SPEA2: Improving the strength Pareto
evolutionary algorithm. Evolutionary Methods for Design, Optimization and
Control with Applications to Industrial Problems, 2002
[ZT99] Zitzler,E. and Thiele,L.: Multiobjective Evolutionary Algorithms: A
Comparative Case Study and the Strength Pareto Approach. IEEE Transactions
on Evolutionary Computation, Vol.3, No.4, pp.257-271, 1999