Vous êtes sur la page 1sur 10

1523_Frame_C01.

fm Page 1 Thursday, June 5, 2003 9:55 AM

1
Introduction

1.1 Forward and Inverse Problems Encountered in Structural


Systems
In engineering, computer-aided design (CAD) tools are used to design
advanced structural systems. Computational simulation techniques are often
used in such tools to calculate the displacement, deflection, strains, stresses,
natural frequencies, and vibration modes, etc. in the structural system for
given loading, initial and boundary conditions, geometrical configuration,
material properties, etc. of the structure. These types of problems are called
forward problems and are often governed by ordinary or partial differential
equations (ODE or PDE) with unknown field variables. For structure mechanics
problems, the field variable is basically the displacements; the constants in
the ODE or PDE and problem domain are known a priori. The source or the
cause of the problem or phenomenon governed by the ODE or PDE and the
relevant initial and boundary conditions are also known. To solve a forward
problem is, in fact, to solve the ODE or PDE subjected to these initial and
boundary conditions.
Many solution procedures, especially the computational procedures, have
been developed, such as:

• Finite deference method (FDM; see e.g., Hirsch, 1988; Anderson, 1995)
• Finite element method (FEM; see e.g., Zienkiewicz and Taylor, 2000;
Liu and Quek, 2003)
• Strip element method (Section 10.3)
• Boundary element method (BEM; see e.g., Brebbia et al., 1984)
• FEM/BEM (see e.g., Liu, Achenbach et al., 1992)
• Mesh-free methods (see e.g., Liu, 2002a; Liu and Liu, 2003)
• Wave propagation solvers (see e.g., Liu and Xi, 2001)

These methods for solving forward problems have been well established,
although the mesh-free methods are still in a stage of rapid development.
Using these methods, the displacements in the structure and then the strains

© 2003 by CRC Press LLC


 

1523  Page 2  Tuesday, June 3, 2003  2:14 PM

and stresses (outputs) can be obtained — as long as the material property,


the geometric configuration of the structure, and the loading, initial, and
boundary conditions (inputs) are given.
Another class of often encountered practical problems is called inverse
problems. In an inverse problem, the effects or outputs (displacement, velocity,
acceleration, natural frequency, etc.) of the system may be known (by exper-
iments, for example), but the parameters of the loading profile (inputs),
material property, geometric feature of the structure, boundary conditions,
or a combination of these may need to be determined. Solving this class of
problem is obviously extremely useful for many engineering applications.
One of the earliest inverse problems in mechanical engineering is the
inverse problem in wave propagation. These problems are formulated based
on the fact that mechanical (elastic) waves (Achenbach, 1973; Liu and Xi,
2001) traveling in materials are scattered from the boundaries and interfaces
of materials, and propagate over distance to “encode” the information on
their path such as the domain boundaries, martial properties, and the wave
source (loading excitation, etc.). It must be possible to “decode” some of the
information encoded in the waves that are recorded as wave responses. A
systematic method to decode the information is to formulate and solve
inverse problems. Problems of this nature arise from nondestructive evalu-
ation (NDE) using waves and ultrasounds, ocean acoustics, earth and space
exploration, biomedical examination, radar guidance and detection, solar
astrophysics, and many other areas of science, technology, and engineering.
The nature of inverse problems requires proper formulations and solution
techniques in order to perform the decoding successfully. In this book,
approaches to formulating inverse problems, inverse analysis procedures, and
computational techniques are discussed. Many engineering inverse problems
are formulated and investigated using these techniques and many important
issues related to inverse problems are examined and revealed by using simple
examples. Methods for dealing with these issues are also presented.
Note that many types of inverse problems exist in engineering. Some of
them can only be formulated in an under-posed form (see Chapter 2), due
to the difficulty or cost of obtaining more experimental data or observations.
Solving this class of under-posed inverse problems will be discussed but is
not the major focus of this book. This book focuses on inverse problems of
even- and over-posed problems because, for many engineering systems,
sufficient experimental readings can be produced, at least in numbers, to
formulate the problem in even- or over-posed forms.

1.2 General Procedures to Solve Inverse Problems


The general procedure of solving an inverse problem is illustrated in Figure
1.1. The details are as follows:

© 2003 by CRC Press LLC


 

1523  Page 3  Tuesday, June 3, 2003  2:14 PM

Define the problem

Create the forward model

Sensitivity analysis between the inputs and outputs

Experiment design

Minimize measurement error (e.g., filtering)

Inverse analysis
General inversion or optimization or NN
(Chapters 4-6); regularization techniques
(Chapter 3) may be used

No
Solution verification

Yes

END

FIGURE 1.1
General procedure to solve inverse problems.

• Define the problem — define the purpose and objectives of the


project with an analysis on the available budget, resources, and
timeframe. An overall strategy and feasible schedule should be
determined for later effective execution. Efforts must be made at all
times to (1) reduce the number of unknowns to be inversely identi-
fied and (2) confine all the parameters in the smallest possible region.
Made at the very first step, these two efforts can often lead to an
effective reduction on possibilities of ill-posed inverse problems,
and, thus, drastically increase the chance of success and improve the
efficiency and accuracy of the inversion operation.
• Create the forward model — a physical model should be established
to capture the physics of the defined problem. The outputs or effects
of the system should be as sensitive as possible to the system param-
eters to be inversely identified. The parameters should be indepen-
dently influential to the outputs or the effects of the system.
Enforcing more conditions can help to well-pose the inverse prob-
lems. Mathematical and computational models should be developed

© 2003 by CRC Press LLC


 

1523  Page 4  Tuesday, June 3, 2003  2:14 PM

for the underlined forward problem. Possible standard computa-


tional methods are FEM, FDM, FVM (finite volume method), mesh-
free method, wave solvers, etc.
• Analyze sensitivity between the effects or outputs and the parame-
ters — make sure that the outputs of the problem and parameters
(including the inputs) to be inversely identified are well correlated.
Ensuring high sensitivity of the outputs to the parameters is one of
the most effective approaches to reducing ill-posedness in the later
stage of inverse analysis. The analysis should be done using the
forward model created without the need of experiments that may
be expensive. Based on the sensitive analysis, modification to the
forward model and to the choices of parameters may be made.
• Design the experiment — decide on proper measurement methods,
type of equipment for testing and recording, and data analysis. The
number of the measurements or readings should be at least more
than the number of unknowns to be inversely identified, which can
lead to at least an even-posed problem. An over-posed system (using
more outputs) is usually preferred so as to improve the property of
the system equation and reduce the ill-posedness of the problem.
An over-posed formulation can usually accommodate higher levels
of noise contamination in the experimental data. However, too
heavily over-posed systems may result in a poor output reproduc-
ibility that can be checked later by computing the output reproduc-
ibility after obtaining the inverse solution.
• Minimize measurement noise (e.g., through filtering) — errors in the
measurement data should be eliminated as much as possible because
they can trigger the ill-posedness of the problem and can be mag-
nified in the inverse solution, or even result in an unstable solution.
Properly designed filters can be used to filter out the errors before
the measurement data are used for the inverse analysis. The princi-
ple is to use a low pass filter to filter out all the noise with frequency
higher than the frequency or wavelength shorter than the wave-
length of the effects of the problem. The frequency or the wavelength
of the effects of the problem can often be estimated by the forward
solver. Details will be covered in Chapter 3.
• Apply the inverse solver — if the system can be formulated in an
explicit matrix form, general inversion of the system (or transforma-
tion) matrix can be performed to obtain the inverse solution. For
complex systems that cannot be formulated in an explicit matrix
form, a functional of error can always be established using a proper
norm, and optimization/minimization techniques should be used to
search for the solution that minimizes the error norm. These optimi-
zation techniques will be discussed in detail in Chapter 4 and Chap-
ter 5. Proper regularization techniques may be used for ill-posed
inverse problems. The regularization techniques are very important

© 2003 by CRC Press LLC


 

1523  Page 5  Tuesday, June 3, 2003  2:14 PM

for obtaining stable solutions for the ill-posed inverse problems. Note
also that the use of some of the regularization techniques should be
the last resort to remedy the ill-posedness of the problem. Some side
effects will occur in using many of the regularization methods and
misuse of regularization techniques can also lead to erroneous
results. Regularization methods will be detailed in Chapter 3.
• Verify the solution — this is important to ensure that the inverse
solution obtained is physically meaningful. All possible methods
with proper engineering judgments should be employed to make
sure that the solution obtained is reliable. Checking on the output
and input reproducibility matrices can give some indications on the
quality of the solution. Modifications of the inverse and experimen-
tal strategy may be needed, and the preceding steps may be repeated
until the inverse solution is satisfactory. Note that many of the ver-
ifications can be done computationally, and experimental verifica-
tions need to be done at the final stage.

1.3 Outline of the Book


This book details the theory, principles, computational methods and algo-
rithms, and practical techniques for inverse analyses using elastic waves
propagating in solids and structures or the dynamic responses of solids and
structures. These computational inverse methods and procedures will be
examined and tested numerically via a large number of examples of force/
source reconstructions, crack detection, flaw characterization, material char-
acterization, heat transfer coefficients identification, protein structure pre-
diction, interatomic potential construction, and many other applications.
Some of these techniques have been confirmed with experiments conducted
by the authors and co-workers in the past years. Discussions of regulariza-
tion methods for the treatment of ill-posed inverse problems will be easy to
understand.
The book also discusses many robust and practical optimization algo-
rithms that are very efficient for inverse analysis and optimization, espe-
cially algorithms developed through the combination of different types of
the existing optimization methods such as gradient based methods with
genetic algorithms, intergeneration projection genetic algorithms, real
number coded microgenetic algorithms, and progressive neural networks.
The efficiency and features of all these optimization algorithms will be
demonstrated using benchmark objective functions as well as actual
inverse problems. Table 1.1 gives a concise summary of the applications
of those computational inverse techniques for actual inverse problems
studied in this book.

© 2003 by CRC Press LLC


1523  Page 6  Tuesday, June 3, 2003  2:14 PM
© 2003 by CRC Press LLC

TABLE 1.1
Summary of Applications of Computational Inverse Techniques for Actual Inverse Problems Studied in This Book
Computational Inverse Techniques Applications
Golden section search method (Section 4.3.1) Section 13.1: Coefficient identification of electronic cooling system
Conjugate gradient method (Section 4.4.3) Section 7.4.7: Identification of the time history of the force
1. Section 7.4.6: Identification of the time function of the force
2. Section 9.4: Identification of material property of functionally graded
materials
Conventional Nonlinear least square method (Section 4.5)
2.1 Section 9.4.1: Transversely FGM plates
optimization
2.2 Section 9.4.2: SiC FGM plates
techniques (Chapter 4)
3. Used frequently in the combined optimization methods

 
Newton’s root finding method (Section 4.6.1) 1. Section 12.4.3: Identification of stiffness factors of cantilever beams
2. Section 12.4.4: Identification of stiffness factors of plates
Levenberg-Marquardt root finding method 1. Section 12.5: Flaw detection in cantilever beams
(Section 4.6.2) 2. Section 13.5: Flaw detection in truss structures
1. Section 8.3.4: Identification of material constants of laminates
1.1 Glass/epoxy [0/45/-45]s laminate
1.2 Carbon/epoxy [0/45/-45/90/-45/45]s laminate
2. Section 8.3.5: Identification of fiber orientation in laminates
2.1 Eight-ply symmetrical laminates
2.2 Ten-ply symmetrical laminates
Genetic algorithms (GA) 2.3 Complex case study
Binary micro-GA (µGA) (Section 5.3.1)
(Chapter 5) 3. Section 8.3.6: Identification of material constants of laminated cylinders
3.1 Glass/epoxy [0/45/-45/90/-45/45]s laminate
3.2 Glass/epoxy [0/-30/30/90/-60/60]s laminate
3.3 Carbon/epoxy [0/-30/30/90/-60/60]s laminate
4. Section 9.5: Material characterization of FGMs
4.1 Section 9.5.1: FGM plates
4.2 Section 9.5.2: FGM cylinders
5. Section 10.6: Crack detection in beams
5.1 Section 10.6.1: Using SEM simulated displacement
5.2 Section 10.6.2: Using experimental displacement
1523  Page 7  Tuesday, June 3, 2003  2:14 PM
6. Section 11.3: Delamination detection in laminates
© 2003 by CRC Press LLC

6.1 Section 11.3.1: Horizontal delamination


6.2 Section 11.3.2: Vertical crack
7. Section 12.3: Flaw detection in sandwich structures
7.1 Sandwich beams
7.2 Sandwich plates
8. Section 13.4: Crack detection using integral strain measured by optical fibers
Real µGA (Section 5.3.2) 1. Section 8.4: Identification of material constants of composite laminate
Intergeneration projection GA (IP-GA) 1. Section 11.4: Delamination detection in laminates
(Section 5.4) 2. Section 13.3: Identification of material property of thin films
3. Section 13.8: Parameter identification in valueless micropumps
Improved IP-GA (Section 5.5) 1. Section 11.5: Delamination detection in laminates
2. Section 13.7: Fitting of interatomic potentials
IP-GA with three parameters (IP3-GA) 1. Section 13.2: Identification of the material parameters of PCBs
(Section 5.6) 2. Section 13.6: Protein structure prediction
GA with search space reduction (SR-GA) 1. Section 13.1.2.2: Thermal coefficient identification of electronic cooling
(Section 5.7) system
2. Section 13.1.2.3: Thermal coefficient identification of PCB
Genetic algorithm combined with the gradient-based methods (Section 1. Section 8.5: Identification of material constants of composite laminate
5.8) 2. Section 9.6: Identification of material property of functionally graded
materials
3. Section 11.6: Delamination detection in composite laminate
Plain neural network 1. Section 13.1.3: Coefficient identification of a telephone switch model
1. Section 8.6: Identification of material constants of composite laminate
2. Section 9.7: Identification of material property of functionally graded
materials
2.1 Section 9.7.1: FGM plate
Neural network 2.2 Section 9.7.2: FGM cylinder
(Chapter 6) Progressive neural network 3. Section 10.7: Crack detection in beams
3.1 Sections 10.7.2-10.7.4: Using SEM model
3.2 Section 10.7.5: Using beam model and harmonic excitation

 
3.3 Section 10.7.6: Using beam model and impact excitation
3.4 Section 10.7.7: Using FEM model
4. Section 11.7: Delamination detection in laminates
 

1523  Page 8  Tuesday, June 3, 2003  2:14 PM

The book is organized as follows:

• Chapter 1 provides a general description and procedure of inverse


analysis and backgrounds and motivations that led to the develop-
ment of these methods for nondestructive evaluation, as well as
development of this book.
• In Chapter 2 the general definition of forward problem as well as
inverse problem will be presented. Ill-posed inverse problems are
classified into three types. Issues related to these three types of ill-
posedness are revealed and discussed using very simple examples.
The formulation of inverse problems will be presented, and the
general procedure to solve inverse problems that can be formulated
in explicit matrix forms will be provided.
• Chapter 3 offers a brief introduction of five regularization methods
for ill-posed inverse problems. These regularization methods
include the Tikhonov regularization, regularization by singular
value decomposition, iterative regularization methods, regulariza-
tion by projection, and regularization by filtering.
• In Chapter 4 some conventional optimization techniques, including
direct search algorithms as well as gradient-based algorithms, are
introduced because engineering inverse problems are usually for-
mulated and solved as optimization problems. These techniques are
provided in a concise and insightful manner with the help of simple
examples.
• Chapter 5 describes the basic concept of genetic algorithms (GAs)
and some modified GAs, with an emphasis on the intergeneration
project GA (IP-GA) as well as the method that combines GAs with
gradient-based methods.
• In Chapter 6, the basic terminology, concepts, and procedures of the
neural network (NN) will be briefly introduced. A typical NN model
and multilayer perceptrons (MLP), along with the back-propagation
learning algorithm, will be detailed. Some practical computational
issues on NNs as well as the progressive NN model are also discussed.
• Chapter 7 through Chapter 12 present a number of computational
inverse techniques using elastic waves propagating in composite
structures or dynamic responses of structures. Practical complex
nondestructive evaluation problems of force function reconstruction,
material property identification, and crack (delamination, flaw)
detection have been examined in detail in the following order:
• Chapter 7 presents inverse procedures for identification of im-
pact loads in composite laminates. Traditional optimization
methods are employed for the inverse analysis and numerical
examples of identification of impact loads applied on beam and
plate types of structures are presented. Experimental studies

© 2003 by CRC Press LLC


 

1523  Page 9  Tuesday, June 3, 2003  2:14 PM

have also been presented for the verification of the inverse


solution.
• In Chapter 8 material constants include the elastic constants or
the engineering constants required in the constitutive law for
composites, and fiber orientation of composite laminates will be
inversely identified from the dynamic displacement responses
recorded at only one receiving point on the surface of composite
laminated structures.
• Chapter 9 discusses the computational inverse techniques for
material property characterization of functionally graded mate-
rials (FGMs) from the dynamic displacement response recorded
on the surface of the FGM structures.
• In Chapter 10, numerical analysis and experimental studies on
the use of flexural waves for nondestructive detection of cracks
and delaminations in beams of isotropic and anisotropic materi-
als are introduced. Computational inverse procedures employing
the GAs and NNs are detailed for determining the geometrical
parameters of the crack and delaminations.
• Computational inverse techniques using elastic wave responses
of displacement for delamination detection in composite lami-
nates are introduced in Chapter 11. Horizontal delaminations as
well as vertical cracks will be considered. GAs and NNs are
employed for the inverse analysis; the strip element method is
used as the forward solver to compute the wave response. Ex-
amples of practical applications are presented to demonstrate the
efficiency of computational inverse techniques for delamination
detection in composite laminates.
• Chapter 12 considers the detection of flaws in beams or plates;
special considerations and treatment for the detection of flaws
in sandwich structures are also provided. The finite element
model is used for forward analysis. GAs and Newton’s root find-
ing method, as well as the Levenberg–Marquardt method, are
used for the inverse analysis. A number of numerical examples
are provided to demonstrate the application of these computa-
tional inverse techniques.
• Several other application examples of the computational inverse
techniques are presented in Chapter 13. These topics range from
the electronic system (heat transfer coefficient identifications), use
of integral optical fibers, MEMS, and interatomic potential to the
protein structure. These applications provide a landscape view on
the broadness of the applications of the inverse techniques.
• Chapter 14 introduces a concept of total solution for engineering
mechanics problems as an extension of the inverse analysis. The
approach for obtaining a total solution is to formulate practical engi-

© 2003 by CRC Press LLC


 

1523  Page 10  Tuesday, June 3, 2003  2:14 PM

neering problems as a parameter identification problem. All the


parameterized unknown information is determined through an iter-
ative procedure of conducting alternately forward and inverse (or
mixed) analyses. This chapter suggests a new approach to formulate
and deal with practical engineering problems.

The background of and many terminologies used in this book are defined
in Chapter 1 through Chapter 3. These chapters will be useful in understand-
ing Chapter 7 through Chapter 14, and therefore should be read first before
proceeding to other chapters. Chapter 4 through Chapter 6 can be read
separately. In fact, these materials are useful not only for inverse problems
but also for general optimization problems. Readers who are familiar with
these optimization techniques may skip these chapters. Chapter 7 through
Chapter 14 can be read in any order, based on the interest of the reader,
because proper cross references for commonly used materials are provided.
The book is written primarily for senior university students, postgraduate
students and engineers in civil, mechanical, geographical and aeronautical
engineering, and engineering mechanics. Students in mathematics and com-
putational science may also find the book useful. Anyone with an elemen-
tary knowledge of matrix algebra and basics of mechanics should be able
to understand its contents fairly easily.

© 2003 by CRC Press LLC

Vous aimerez peut-être aussi