Professional Documents
Culture Documents
by
Youhua Liu
Dissertation Submitted to the Faculty of Virginia Polytechnic Institute and State University
in partial fulfillment of the requirements for the degree of
DOCTOR OF PHILOSOPHY
In
Aerospace Engineering
Approved:
Romesh C. Batra
Eric R. Johnson
Zafer Grdal
Efstratios Nikolaidis
April 2000
Blacksburg, Virginia
Keywords:
by
Youhua Liu
(ABSTRACT)
The aerospace industry is increasingly coming to the conclusion that physics-based highfidelity models need to be used as early as possible in the design of its products. At the preliminary
design stage of wing structures, though highly desirable for its high accuracy, a detailed finite
element analysis(FEA) is often not feasible due to the prohibitive preparation time for the FE
model data and high computation cost caused by large degrees of freedom. In view of this situation,
often equivalent beam models are used for the purpose of obtaining global solutions. However, for
wings with low aspect ratio, the use of equivalent beam models is questionable, and using an
equivalent plate model would be more promising.
An efficient method, Equivalent Plate Analysis or simply EPA, using an equivalent plate
model, is developed in the present work for studying the static and free-vibration problems of builtup wing structures composed of skins, spars, and ribs. The model includes the transverse shear
effects by treating the built-up wing as a plate following the Reissner-Mindlin theory (FSDT). The
Ritz method is used with the Legendre polynomials being employed as the trial functions.
Formulations are such that there is no limitation on the wing thickness distribution. This method is
evaluated, by comparing the results with those obtained using MSC/NASTRAN, for a set of
examples including both static and dynamic problems.
ii
The Equivalent Plate Analysis (EPA) as explained above is also used as a basis for generating
other efficient methods for the early design stage of wing structures, such that they can be
incorporated with optimization tools into the process of searching for an optimal design. In the
search for an optimal design, it is essential to assess the structural responses quickly at any design
space point. For such purpose, the FEA or even the above EPA, which establishes the stiffness and
mass matrices by integrating contributions spar by spar, rib by rib, are not efficient enough.
One approach is to use the Artificial Neural Network (ANN), or simply called Neural Network
(NN) as a means of simulating the structural responses of wings. Upon an investigation of
applications of NN in structural engineering, methods of using NN for the present purpose are
explored in two directions, i.e. the direct application and the indirect application. The direct method
uses FEA or EPA generated results directly as the output. In the indirect method, the wing innerstructure is combined with the skins to form an "equivalent" material. The constitutive matrix,
which relates the stress vector to the strain vector, and the density of the equivalent material are
obtained by enforcing mass and stiffness matrix equities with regard to the EPA in a least-square
sense. Neural networks for these material properties are trained in terms of the design variables of
the wing structure. It is shown that this EPA with indirect application of Neural Networks, or
simply called an Equivalent Skin Analysis (ESA) of the wing structure, is more efficient than the
EPA and still fairly good results can be obtained.
Another approach is to use the sensitivity techniques. Sensitivity techniques are frequently used
in structural design practices for searching the optimal solutions near a baseline design. In the
present work, the modal response of general trapezoidal wing structures is approximated using
shape sensitivities up to the second order, and the use of second order sensitivities proved to be
yielding much better results than the case where only first order sensitivities are used. Also
different approaches of computing the derivatives are investigated. In a design space with a lot of
design points, when sensitivities at each design point are obtained, it is shown that the global
variation in the design space can be readily given based on these sensitivities.
iii
Acknowledgments
This work would not have been accomplished without the support and guidance of my advisor
and committee chairman, Dr. Rakesh K. Kapania. Dr. Kapania's professional attitude influenced me
a lot, and his prompt responses to my questions and submitted work, encouragement during all
phases of my work, and his understanding are greatly appreciated. I am grateful to Dr. Romesh C.
Batra, Dr. Zafer Grdal, Dr. Eric R. Johnson, and Dr. Efstratios Nikolaidis for serving in my
committee. I would like to thank the financial support of NASA Langley Research Center on this
research through Grant NAG-1-1884 with Dr. Jerry Housner and Dr. John Wang as the Technical
Monitors. I am also thankful to other students for the helps I have received, especially Dr. Daniel
Hammerand, Dr. Luohui Long, and Mr. Erwin Sulaeman.
Finally, I would say this work could not have got started, let alone been finished, without the
unconditional support, trust and love of my wife, Ting, and my daughter, Lisa. I owe them a lot.
Contents
List of Tables
List of Figures
xi
Nomenclature
xvi
1. Introduction
11
vi
21
32
42
5.6.4
75
6.6
95
149
viii
References
153
160
164
167
Vita
171
ix
List of Tables
Table 3.1 Comparison of Continuum Model Properties for a Lattice Repeating Cell. . . . . . . . . . 31
Table 3.2 Comparison of Continuum Model Properties for a Lattice Repeating Cell. . . . . . . . . . 31
Table 5.1 Natural frequencies (Hz) of the cantilevered swept-back box wing. . . . . . . . . . . . . . . . 64
Table 5.2 Displacement (in) of the cantilevered swept-back box wing. . . . . . . . . . . . . . . . . . . . . 69
Table 5.3 Comparison of FEA and EPA in terms of DOF and Number of Elements. . . . . . . . . . .74
Table 7.1 Differences between the natural frequencies by EPA and ESA. . . . . . . . . . . . . . . . . . 101
Table 7.2
Table 7.3
List of Figures
Fig. 2.1 A feed-forward multi-layer neural network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .13
Fig. 2.2 Details of a neuron. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .13
Fig. 2.3 Transfer functions. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .14
Fig. 2.4 Radial basis function neural network. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .15
Fig. 3.1 Geometry of repeating cells of a single-bay double laced lattice structure. . . . . . . . . . . . 22
Fig. 3.2 Evaluating continuum model properties for a repeating cell. . . . . . . . . . . . . . . . . . . . . . .24
Fig. 3.3 Training data for GA = f ( Ac , Lc ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Fig. 3.4 Distributions of training and testing points. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
Fig. 3.5 Feed-forward NN simulation for GA = f ( Ac, Lc ) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Fig. 3.6 Feed-forward NN simulation errors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
Fig. 3.7 Radial-basis function NN simulation for GA = f ( Ac, Lc ) . . . . . . . . . . . . . . . . . . . . . . . . . .28
Fig. 3.8 Radial-basis function NN simulation errors. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .28
Fig. 3.9 Training history of a 3-10-1 feed-forward NN by trainbp. . . . . . . . . . . . . . . . . . . . . . . . .30
Fig. 3.10 Training history of a 3-10-1 feed-forward NN by trainbpa. . . . . . . . . . . . . . . . . . . . . . .30
Fig. 3.11 Training history of a 3-10-1 feed-forward NN by trainlm. . . . . . . . . . . . . . . . . . . . . . . .31
Fig. 4.1 The coordinate system and its transformation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
Fig. 4.2 The Legendre polynomials. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .36
Fig. 4.3 The Chebyshev polynomials. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 37
Fig. 5.1 Wing skin. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .44
xi
Natural frequencies using equivalent plate analysis with mode tracking. . . . . . . . . . . . .84
Fig. 6.3
xii
Fig. 6.4 The 2nd natural frequency w.r.t. wing plan area
using 1st and 2nd order sensitivities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Fig. 6.5 The 3rd natural frequency w.r.t. wing sweep angle
using 1st and 2nd order sensitivities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 88
Fig. 6.6 Comparison of the natural frequencies of the first 6 modes for wing structures
randomly chosen inside the box of design space, as obtained by the NN and ST
w.r.t. those obtained using a full-fledged EPA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 92
Fig. 6.7 Comparison of the natural frequencies of the first 4 modes for wing structures
along a path inside the box of design space (n 1 =0.945, n 2 =8.200, n 3 =3.203) using
only the 1st order sensitivities. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 93
Fig. 6.8 Comparison of the natural frequencies of the first 4 modes for wing structures
along a path inside the box of design space (n 1 =0.945, n 2 =8.200, n 3 =3.203) using
sensitivities up to the 2nd order. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
Fig. 7.1 An example of mass density distribution generated using Eq. (7.8) . . . . . . . . . . . . . . . .101
Fig. 7.2 The stiffness matrix given by EPA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .102
Fig. 7.3 The stiffness matrix given by ESA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .103
Fig. 7.4 Difference between stiffness matrices given by EPA and ESA. . . . . . . . . . . . . . . . . . . .104
Fig. 7.5 The mass matrix given by EPA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
Fig. 7.6 The mass matrix given by ESA. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 106
Fig. 7.7 Difference between mass matrices given by EPA and ESA. . . . . . . . . . . . . . . . . . . . . . 107
Fig. 7.8 49 randomly chosen wing plan forms in design space I. . . . . . . . . . . . . . . . . . . . . . . . . .110
Fig. 7.9 Comparison of the first 10 frequencies by EPA and ESA. . . . . . . . . . . . . . . . . . . . . . . .111
Fig. 7.10 The relative errors in Fig. 7.9. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .112
Fig. 7.11 25 wing plan forms systematically varying through design space I. . . . . . . . . . . . . . . .113
Fig. 7.12 Comparison of the first 6 frequencies by EPA and ESA. . . . . . . . . . . . . . . . . . . . . . . .114
Fig. 7.13 25 randomly chosen wing plan forms in design space II. . . . . . . . . . . . . . . . . . . . . . . .117
Fig. 7.14 Comparison of the first 10 frequencies by EPA and ESA. . . . . . . . . . . . . . . . . . . . . . .118
Fig. 7.15 The relative errors in Fig. 7.14. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .119
xiii
Fig. 7.16 16 wing plan forms systematically varying through design space II. . . . . . . . . . . . . . .120
Fig. 7.17 Comparison of the first 6 frequencies by EPA and ESA. . . . . . . . . . . . . . . . . . . . . . . .121
Fig. 7.18 An arbitrarily chosen wing plan form in design space II. . . . . . . . . . . . . . . . . . . . . . . .122
Fig. 7.19 Comparison of displacements by EPA and ESA for 1 lb tip force . . . . . . . . . . . . . . . .123
Fig. 7.20 Comparison of the Von Mises stress at wing root by EPA and ESA
under 1 lb tip force . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .124
Fig. 7.21 Comparison of the Von Mises stress along central spar by EPA and ESA
under 1 lb tip force . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .125
Fig. 7.22 25 randomly chosen wing plan forms in design space III. . . . . . . . . . . . . . . . . . . . . . . 128
Fig. 7.23 Comparison of the first 10 frequencies by EPA and ESA. . . . . . . . . . . . . . . . . . . . . . .129
Fig. 7.24 The relative errors in Fig. 7.23. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .130
Fig. 7.25 16 wing plan forms systematically varying through design space III. . . . . . . . . . . . . . 131
Fig. 7.26 Comparison of the first 6 frequencies by EPA and ESA. . . . . . . . . . . . . . . . . . . . . . . .132
Fig. 7.27 An arbitrarily chosen wing plan form in design space III. . . . . . . . . . . . . . . . . . . . . . . 133
Fig. 7.28 Comparison of displacements by EPA and ESA at 1 lb tip force . . . . . . . . . . . . . . . . .134
Fig. 7.29 Comparison of the Von Mises stress at wing root by EPA and ESA
under 1 lb tip force . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .135
Fig. 7.30 Comparison of the Von Mises stress along central spar by EPA and ESA
under 1 lb tip force . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .136
Fig. 7.31 16 randomly chosen wing designs in design space IV. . . . . . . . . . . . . . . . . . . . . . . . . .139
Fig. 7.32 Comparison of the first 10 frequencies by EPA and ESA. . . . . . . . . . . . . . . . . . . . . . .140
Fig. 7.33 The relative errors in Fig. 7.26. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .141
Fig. 7.34 16 wing designs systematically varying through design space IV. . . . . . . . . . . . . . . . .142
Fig. 7.35 Comparison of the first 6 frequencies by EPA and ESA. . . . . . . . . . . . . . . . . . . . . . . .143
Fig. 7.36 An arbitrarily chosen wing design in design space IV. . . . . . . . . . . . . . . . . . . . . . . . . .144
Fig. 7.37 Comparison of displacements by EPA and ESA at 1 lb tip force . . . . . . . . . . . . . . . . .145
Fig. 7.38 Comparison of the Von Mises stress at wing root by EPA and ESA
under 1 lb tip force . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .146
xiv
Fig. 7.39 Comparison of the Von Mises stress along central spar by EPA and ESA
under 1 lb tip force . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .147
Fig. B.1 Sketch for a wing composed of main-body and wing-let. . . . . . . . . . . . . . . . . . . . . . . . 164
Fig. C.1 The Karman-Trefftz transformation. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 168
Fig. C.2 Airfoils shapes obtained using Karman-Trefftz transformation. . . . . . . . . . . . . . . . . . . 170
xv
Nomenclature
a
Ac , Ad , Ag
ANN
{B}
chord-length
c0 , cr
chord-length at root
c1
chord-length at tip
[C]
bi j
b1,b2,b3
[D]
constitutive matrix
{d }
displacement vector
DOF
D pq
Youngs Modulus
EA
axial rigidity
EI
bending rigidity
xvi
EPA
ESA
f ()
transfer function
FEA
FEM
FF
feed-forward
FSDT
GA
shear rigidity
[G]
[H]
h1, 2
i, j
integers
I,J,K,L,M,N,
P,Q,R,S
integers
initff
[J]
Jacobian matrix
J 11 , J 12 , J 21 , J 22
integer
[K]
~
[K ]
Lc , Ld , L g
'logsig'
l1, 2
[M]
xvii
~
[M ]
MAC
m, n
integers
NN
Neural Network
Nz
n1, n2
N 1~ 4 ( , )
transformation functions
n rib
number of ribs
n spar
number of spars
{P}
Pi (x )
Legendre polynomials
p, q
integers
'purelin'
Px , Py , Pz
force components
{q}
ri j
RBF
simuff
simurb
solverb
ST
sensitivity techniques
xviii
t0
skin thickness
t1, 2
kinetic energy
[T]
'tansig'
trainbp
trainbpa
trainlm
Ti (x )
Chebyshev polynomials
{x}
eigenvector
x,y,z
Cartesian coordinates
x ij
x1~ 4
y1~ 4
[ZZ]
strain energy
u,v,w
u0 , v0 , w0
{v }
velocity vector
wkij 1
weight between node k of the ( j 1) -th layer and node i of the j -th layer
wijK , wijM
w1,w2,w3
wi ( p)
xix
w.r.t.
with regard to
x , y , z ,
x , y
{ }
strain vector
{ }
x , y , z ,
strain tensors
xy , yz , zx
{i }
{ j }
r ( )
eigenvalue
Poisson's ratio
{ }
stress vector
frequency, rad/sec
s ( )
xx
Chapter 1
Introduction
1.1 The Trend of Early Analysis in Product Design
To reduce product development cycle is essential to a nowadays manufacturing enterprise not
only on economic savings in the process itself, but also to a broad business advantage in getting
product innovations to customers faster, and thereby increasing the company's market share 1 .
One of the most valuable CAE (Computer Aided Engineering) tools is finite-element analysis
(FEA), which assists in analyzing structures to detect areas that might undergo excessive stress,
deformation, vibration, or other potential problems. Yet, instead of assisting in reducing time to
market, the traditional, full-blown FEA actually became a bottleneck and was often done only
toward the end of product design.
The experience of manufacturers in many industries has shown that 85~90% of the total time
and cost of product development is defined in the early stages of product development, when only
5~10% of project time and cost have been expended 1 , 2 . This is because in the early concept stages,
fundamental decisions are made regarding basic geometry, materials, system configuration, and
manufacturing processes.
The process, however, can be re-oriented so that analysis is performed much earlier to shorten
the product development cycle. This moves CAE/analysis forward into conceptual design, where
changes are much easier and more economical to make in correcting poor designs earlier. The
CHAPTER 1
INTRODUCTION
major benefits of up-front analysis includes giving designers the ability to perform "what-if?"
simulations that enable them to evaluate alternative approaches and explore options early in the
design cycle to arrive at a superior design. This methodology employs CAE to help avoid "fires" in
the early design stage, rather than uses CAE to put out "fires" in the later design stage as the
traditional practice does 1 .
Therefore, instead of being the last thing to do, CAE is now one of the first things for a
designer to do to make sure that the best design possible is to be obtained 3 .To facilitate this
methodology of early analysis in product design, there have emerged the following two issues
concerning the development of CAE.
The first issue is the lack of integration between CAD and analysis programs. The need to
translate, clean up, and further process design data for use in analysis has limited the effectiveness
of both CAD and CAE software. Over the past few years, software vendors have been moving to
tightly couple CAD and CAE software programs by tying them into suites using a shared database
and a single user interface. Sharing database means that engineers no longer have to translate
design data to formats that the analysis program can recognize, and vice versa. It also allows
updates in one system to be reflected immediately in the other. CAD and CAE sharing the same
user interface makes it easier for a user to switch from one program to the other.
The second issue is the inappropriateness of FEA as the tool of CAE in many cases. Usually
FEA can only be integrated in the early design stage of structurally simple products or components
of a structurally complex product. For instance, at the preliminary design stage of built-up wing
structures, though highly desirable for its high accuracy, a detailed finite element analysis(FEA) is
often not feasible because: (i) the preparation time for the FEM model data may be prohibitive,
especially when there is little carry-over from design to design; (ii) for complex structures
composed of large number of components, a detailed FEA involves huge number of degrees of
freedom, and needs large amount of CPU time and computation capacity, which makes the cost too
high. For such cases, unconventional methods that are more efficient than FEA are needed.
CHAPTER 1
INTRODUCTION
People have employed continuum models, assuming the complex structures to behave
similarly, for analysis at the early stage of the design process of a complex product. This includes
using beam, plate or shell models to simulate complex structures. In the present work,
methodologies are developed in employing the first-order shear deformation theory (the Mindlin
plate) to simulate the structural responses of built-up wing structures, incorporating neural
networks and other tools to further enhance analysis efficiency. It is hoped that the methodologies
developed in the present work can be used in the early design stages of aerospace wings and other
plate-like complex structures, therefore a superior design can be obtained in a development process
of shorter cycle and less expenses.
CHAPTER 1
INTRODUCTION
distributed processor that has a natural propensity for storing experiential knowledge and making it
available for future use by resembling the brain in two aspects: (a) Knowledge is acquired by the
network through a learning process; (b) Inter-neuron connection strengths known as synaptic
weights (or simply weights) are used to store the knowledge 4 .
With a history traced to the early 1940s, and two periods of major increases in research
activities in the early 1960s and after the mid-1980s, ANNs have now evolved to be a mature
branch in the computational science and engineering with a large number of publications, a lot of
quite different methods and algorithms and many commercial software and some hardware. They
have found numerous applications in science and engineering, from biological and medical
sciences, to information technologies such as artificial intelligence, pattern recognition, signal
processing and control, and to engineering areas as civil and structural engineering.
1.2.2 Applications in Structural Engineering
In the field of structural engineering, there have been a lot of attempts and researches making
use of NN to improve efficiency or to capture relations in complex analysis or design problems.
The following are a few examples. Abdalla and Stavroulakis 5 applied NN to represent
experimental data to model the behavior of semi-rigid steel structure connections, which are related
to some highly nonlinear effects such as local plastification etc. Several cases of neural network
application in structural engineering can be found in Vanluchene and Sun 6 . All the problems
treated in Ref. 6 had been reproduced in Gunaratnam and Gero 7 with a conclusion that
representational change of a problem based on dimensional analysis and domain knowledge can
improve the performance of the networks. There is a summary of applications of NN in structural
engineering in Ref. 8. In Liu, Kapania and VanLandingham 9 , methodologies of applying Neural
Networks and Genetic Algorithms to simulate and synthesize substructures were explored in the
solution of 1-D and 2-D beam problems.
CHAPTER 1
INTRODUCTION
CHAPTER 1
INTRODUCTION
The complex nature of the various methods and the large number of problems encountered in
determining the equivalent models are not surprising given the fact that determining these models
for a given complex structure (a large space structure or a wing) belongs to a class of problems
called inverse problems. These problems are inherently ill-posed and it is fruitless to attempt to
determine unique continuum models. The present work deals with investigating the possibility that
a more rational and efficient approach of determining the continuum models can be achieved by
using artificial neural networks.
The following are examples of work on using beam or plate models to simulate repetitive
lattice structures: Noor, Anderson, and Greene 10 ; Nayfeh, and Hefzy 11 ; Sun, Kim, and
Bogdanoff 12 ; Noor 13 ; Lee 14 ~16 . Specifics of these methods will be discussed in Chapter 3.
In the area of analyzing aerospace wing structures, a number of studies have been conducted on
using equivalent beam models to represent simple box-wings composed of laminated or anisotropic
materials, which include Kapania and Castel 17 , Song and Librescu 18 , and Lee 19 . They have given
some fine results for the specific problems. However, for wings with low aspect ratio, the use of
equivalent beam models is questionable, and using an equivalent plate model would be more
promising.
CHAPTER 1
INTRODUCTION
Deformation Theory (FSDT), or the Higher-order Shear Deformation Theory (HSDT) etc. The
CPT is based on the Kirchhoff-Love hypothesis, that is, a straight line normal to the plate middle
surface remains straight and normal during the deformation process. This group of theories work
well for truly thin isotropic plates, but for thick isotropic plates and for thin laminated plates they
tend to overestimate the stiffness of the plate since the effects of through-the-thickness shear
deformation are ignored 22, 23 . The FSDT is based on the Reissner-Mindlin model 24, 25 , where the
constraint that a normal to the mid-surface remains normal to the mid-surface after deformation is
relaxed, and a uniform transverse shear strain is allowed. The FSDT is the most widely used theory
for thick and anisotropic laminated plates owing to its simplicity and its low requirement for
computation capacity. For more accurate results or more realistic local distributions of the
transverse strain and stress, one should use the HSDT 26 , or the CFSDT (Consistent First-order
Shear Deformation Theory) proposed by Knight and Qi 27 .
Methods of solving the CPT, FSDT or HSDT mainly include finite element, Galerkin, and
Rayleigh-Ritz methods 20, 21 . In the context of using equivalent plate to represent the behaviors of
wing structures at the conceptual stage at least, it is obvious that while the computationally costly
finite element method is to be avoided, the Rayleigh-Ritz method is attractive.
There have been several studies using equivalent plate models to model wing structures.
Giles 28, 29 developed a Ritz method based approach, which considers an aircraft wing as being
formed by a series of equivalent trapezoidal segments, and represents the true internal structure of
aircraft wings in the polynomial power form. In Giles 28 the CPT was used, but this shortcoming
was removed subsequently 29 . Tizzi 30 presented a method whose many aspects are similar to that
of Giles. In Tizzi's work several trapezoidal segments in different planes can be considered, but the
internal parts of wing structures (spars, ribs, etc.) were not considered. Livne 31 formulated the
FSDT to be used for modeling solid plates as well as typical wing box structures made of cover
skins and an array of spars and ribs based on simple-polynomial trial functions, which are known to
be prone to numerical ill-conditioning problems. Livne and Navarro then further developed the
method to deal with nonlinear problems of wing box structures 32 .
CHAPTER 1
INTRODUCTION
CHAPTER 1
INTRODUCTION
CHAPTER 1 INTRODUCTION
10
assess the structural responses quickly at any design space point. For such purpose, the FEM or
even the above EPA, which establishes the stiffness and mass matrices by integrating contributions
spar by spar, rib by rib, are not efficient enough.
One approach is to use Neural Networks as a means of simulating the structural responses of
wings. This is the so called direct application of neural networks, as discussed in Chapter 2.
Another approach is to use the sensitivity techniques. Sensitivity techniques are frequently used in
structural design practices for searching the optimal solutions near a baseline design 33,34 . In the
present work, the modal response of general trapezoidal wing structures is approximated using
shape sensitivities up to the 2nd order, and the use of second order sensitivities proved to be
yielding much better results than the case where only first order sensitivities are used. Also
different approaches of computing the derivatives are investigated. These two approaches of direct
simulation of modal wing responses are described in Chapter 6, along with an example showing
results giving by both approaches.
Finally, in Chapter 7, a method more efficient than the EPA with indirect application of neural
networks is developed. Instead of evaluating the matrices over all components of the wing
structure, evaluation is performed only over the skins, whose "equivalent" material constitutive
matrix and mass density distribution are changed accordingly to incorporate the effects of spars and
ribs. The new skin material properties are simulated using Neural Networks in terms of the wing
design variables. As it is shown, while the Neural-Network-aided EPA, which can be called
Equivalent Skin Analysis (ESA), gives almost equally good results, it uses only a fraction of the
CPU time spent in the ordinary EPA in generating the matrices.
Major parts of the present work are published. They include Chapter 4 and 5 in Ref. 48 and 49,
Chapter 6 in Ref. 50 and 51, and Chapter 7 in Ref. 52.
Chapter 2
Basics of Neural Networks
In this chapter a brief description is given to the most extensively used neural network in civil and
structural engineering, Multi-Layer Feed-Forward NN, and another kind of NN, Radial Basis
Function NN, which is very efficient in some cases. Some conceptual features of NN are listed.
Several functions of MATLAB NN Toolbox are introduced, which will be used as the major tools in
the present work. At the end of this chapter a brief discussion is made on approaches of application
of neural networks.
11
CHAPTER 2
12
In the following a brief description is given to two kinds of extensively used neural networks
and some of the pertinent concepts.
2.1.1 Feed-Forward Multi-Layer Neural Network
An example of feed-forward multi-layer neural network is shown in Fig. 2.1, where the
numbers of input and output are 3 and 2 respectively, and there are two hidden layers with 5
neurons in the first hidden layer, and 3 neurons in the second hidden layer. The details of a neuron
is illustrated in Fig. 2.2. As shown in Fig. 2.2, in the j -th layer, the i -th neuron has inputs from
the ( j 1) -th layer of value x kj 1 (k = 1, , n j 1 ) ,and has the following output
x ij = f ( ri j )
(2.1)
where
ri j =
n j 1
w
k =1
j 1
ki
x kj 1 bi j
(2.2)
in which wkij 1 is the weight between node k of the ( j 1) -th layer and node i of the j -th layer,
and bi j is the bias (also called threshold). The above relation can also be written as
ri =
j
n j 1
w
k =0
j 1
ki
x kj 1
(2.3)
CHAPTER 2
13
hid d en
layers
inp ut
layer
o utp ut
layer
Inp ut
signals
S ynap tic
weights
x 1 j-1
w 1 i j- 1
x2
j-1
.
.
.
x k j-1
w 2 i j- 1
k= n j-1
.
.
.
T hresho ld
-1
S umming
junc tio n
bi j
ri j
f ( .)
T ransfer
func tio n
w k i j- 1
xi j
O utp ut
CHAPTER 2
14
The transfer function (also called activation function or threshold function) is usually specified
as the following Sigmoid function
f (r) =
1
.
1 + e r
(2.4)
Other choices of the transfer function can be the hyperbolic tangent function
1 e r
f (r) =
,
1 + e r
(2.5)
(2.6)
(2.7)
Piecewise-linear
0.75
Hyperbolic tangent
0.5
f(r)
0.25
Sigmoid
-0.25
-0.5
Linear
-0.75
-1
-5
r
Fig. 2.3 Transfer functions
CHAPTER 2
number of
neurons:
n1
input
layer
15
n2
hidden
layer
n3
output
layer
x k2 = wk2, j r j + bk2
(2.8)
j =1
n1
r j = G ( x i1 , w1j , bi1, j )
(2.9)
i =1
where w 2 , b 2 are the weights and bias respectively, and the Gaussian function is used as the
transfer function:
G ( xi1 , w1j , bi1, j ) = exp( {bi1, j }2 {xi1 w1j }2 )
where w1 is the center vector of the input data, and b1 is the variance vector.
(2.10)
CHAPTER 2
16
Many NN methods are universal approximators, in the sense that, given a dimension (number
of hidden layers and neurons of each layer) large enough, any continuous mapping can be
realized. Fortunately, the two NNs we are most interested in, the multi-layer feed-forward NN
and the radial basis function NN, are examples of such universal approximators 53,54 .
The input and output relationship of NN is highly nonlinear. This is mainly introduced by the
nonlinear transfer function. Some networks, e.g. the so-called "abductive" networks, use double
even triple powers besides linear terms in their layer to layer input-output relations 55 .
A NN is parallel in nature, so it can make computation fast. Neural networks are ideal for
implementation in parallel computers. Though NN is usually simulated in ordinary computers
in a sequential manner.
A NN provides general mechanisms for building models from data, or give a general means to
set up input-output mapping. The input and output can be continuous (numerical), or not
continuous (binary, or of patterns).
CHAPTER 2
17
numerical modeling, which is of our major concern for the present study, there is a great
similarity between NN training and some kind of least-square fitting or interpolation.
Simulation using NN gives better results in interpolation than in extrapolation, the same as any
other data fitting or mapping methods.
Where and when to use NN depend on the situation, and NN is not a panacea. The following
comment on NN application on structural engineering seemingly can be generalized in other
areas:
"The real usefulness of neural networks in structural engineering is not in reproducing existing
algorithmic approaches for predicting structural responses, as a computationally efficient
alternative, but in providing concise relationships that capture previous design and analysis
experiences that are useful for making design decisions" 7 .
Despite the above features and wide application in a lot of areas, there seems to be no evidence
for neural networks to claim superiority over some other mapping tools. For instance, in a recent
paper of Nikolaidis, Long, and Ling 73, it is claimed that the response surface polynomials with
stepwise regression and the neural network models appear to be almost equally accurate, but it took
considerably less time to develop the polynomials than the neural networks.
CHAPTER 2
18
[w1,b1,w2,b2,w3,b3]=initff(p,n1,'logsig',n2,'logsig',t,'logsig');
where w1, w2, and w3 are initial values for the weight matrices of the 1st (hidden), 2nd (hidden)
and 3rd (output) layer respectively, b1, b2, and b3 are the bias (threshold) vectors, n1 and n2 the
number of neurons in the 1st and 2nd hidden layer respectively, and 'logsig' means that the Sigmoid
transfer function is used.
The present version of MATLAB NN Toolbox can support only 2 hidden layers, but the number
of neurons is only limited by the available memory of the computer system being used. For the
transfer function, one can also use other choices, such as 'tansig' (hyperbolic tangent sigmoid),
'radbas' (radial basis) and 'purelin' (linear) etc.
Experiences of using initff indicated that it seems to be a random process since it is found that
the result of the execution of this algorithm each time is different. And other conditions kept the
same, two executions of this function usually give quite different converging histories of training
by the training algorithm 8 .
Shown in the following is the MATLAB algorithm for training feed-forward network with backpropagation:
[w1,b1,w2,b2,w3,b3,ep,tr]=trainbp(w1,b1,'logsig',w2,b2,'logsig',w3,b3,'logsig',p,t,tp);
where most of the parameters which the user should take care of have been mentioned in the above
paragraphs. The only parameter that the user sometimes need to specify is the 1 4 vector tp,
where the first element indicates the number of iterations between updating displays, the second the
maximum number of iterations of training after which the algorithm would automatically terminate
the training process, the third the converging criterion (sum-squared error goal), and the last the
learning rate. The default value of tp is [25, 100, 0.02, 0.01].
Other algorithms for training include trainbpa (training feed-forward NN with backpropagation and adaptive learning), solverb (designing and training radial basis network), and
trainlm (training feed-forward NN with Levenberg-Marquardt algorithm) etc.
trainbpa and trainlm have very similar formats for using as that of trainbp. The radial basis
network designing and training algorithm has the following format
CHAPTER 2
19
[w1,b1,w2,b2,nr,err]=solverb(p,t,tp);
where the algorithm chooses centers for the Gaussians and increases the neuron number of the
hidden layer automatically if the training cannot converge to the given error goal. So it is also a
designing algorithm.
After the NN is trained, one can predict output from input by using simulation algorithms in
terms of the obtained parameters w1, b1, w2, b2, etc. For feed-forward network one use
y=simuff(x,w1,b1,'logsig',w2,b2,'logsig',w3,b3,'logsig');
where x is the input matrix, and y the predicted output matrix. Similarly, after a radial basis
network has been trained one uses
y=simurb(x,w1,b1,w2,b2);
to predict the output.
Once a NN is trained, we can use the formulations in 2.1 or 2.2 together with the obtained
parameters (weights etc.) to setup the network to do prediction anywhere and not necessarily within
the MATLAB environment.
CHAPTER 2
20
network would not oscillate around the training data. Once the networks are trained, structural
responses at any design point can be recalled in a fraction of a second and this is really favorable in
a design situation 51 .
2.4.2 Indirect Application
Here it is desired to find a way of incorporating NN into the application of the equivalent plate
analysis (EPA) of complex wing structures, other than just making use of results generated by EPA
as the training data base. Note that in the EPA of a complex wing, the computational effort is
mainly spent on integrals for generating the contribution from the inner-structural components of
the wing, i.e. the spars and the ribs, in the stiffness and mass matrices. If an anisotropic material
can be found to replace the inner components, in terms of an equivalent skin, such that the new
composite wing has very close global properties as the original one, then the EPA can be
performed more efficiently. Solution of the adequate material properties of the anisotropic material
is the major obstacle here. The role of NN will be relating the material properties to all kinds of
wing design parameters, and it can be trained when there exists enough data base for training. This
way of applying NN has been claimed to be the best use of the Neural Networks in structural
engineering 7 . This is the path that is to be followed in Chapter 7.
Chapter 3
Continuum Model Approaches
3.1 Methods of Obtaining Continuum Models
A lot of methods have been used to develop continuum models to represent complex structures.
Many of these methods involve the determination of the appropriate relationships between the
geometric and material properties of the original structure and its continuum models. An important
observation is that the continuum model is not unique, and determining the continuum model for a
given complex structure is inherently ill-posed therefore diverse approaches can be used. This can
be clearly shown in the following example of determining continuum models for a lattice structure.
The single-bay double-laced lattice structure shown in Fig. 3.1 has been studied in Ref. 10, 12,
and 14 with different approaches to the continuum modeling. This lattice structure with repeating
cells can be modeled by a continuum beam if the beam's properties is properly provided.
Noor et al's method include the following steps 10 : (1)introducing assumptions regarding the
variation of the displacements and temperature in the plane of the cross section for the beamlike
lattice, (2)expressing the strains in the individual elements in terms of the strain components in the
assumed coordinate directions, (3)expanding each of the strain components in a Taylor series, and
(4)summing up the thermoelastic strain energy of the repeating elements which then gives the
thermoelastic and dynamic coefficients for the beam model in terms of material properties and
geometry of the original lattice structure.
21
CHAPTER 3
22
Length of battens: L g
1
Areas: Ac , Ag , Ad
Lc
batten
Lg
diagonal bar
longitudinal bar
Fig. 3.1 Geometry of repeating cells of a single-bay double laced lattice structure
In Sun et al 12 , the properties of the continuum model is obtained respectively by relating the
deformation of the repeating cell to different load settings under specified boundary conditions. For
example, the shear rigidity GA is obtained by performing a numerical shear test in which a unit
shear force is applied at one end of the repeating cell and the corresponding shear deformation is
calculated by using a finite element program. The mass and rotatory inertia are calculated with a
averaging procedure.
Lee put forward a method that he thought to be more straightforward 14 . He used an extended
Timoshenko beam to model the equivalent continuum beam. By expressing the total strain and
kinetic energy of the repeating cell in terms of the displacement vector at both ends of the
continuum model, and equating them to those obtained through the extended Timoshenko beam
CHAPTER 3
23
theory, he got a group of relations. The number of these relations, 2N(1+2N), where N is the degree
of freedom of the continuum model, is usually larger than that of the equivalent continuum beam
properties to be determined. Lee then introduced a procedure in which the stiffness and mass
matrices for both the lattice cell and the continuum model are reduced and so is the number of
relations. Yet how to reduce the number of relations to be equal to the number of unknowns seems
to depend on luck.
All the above three methods give close results for the continuum model properties, and the
continuum models also generate promising global results for the lattice structure.
CHAPTER 3
y, v
1/3
1/3
y, v
1/2
1/3
24
y, v
1/2
1/3
z, w
1/3
z, w
z, w
x, u
1
1/3
2
x, u
x, u
3
2
3
2
Lg = EI
Lg Lc
3
4
Lc
L2g Lc
v5 v4
1
4
Lc
w4 + w5 + 14 w6
1
2
CHAPTER 3
25
Figs. 3.5 and 3.6. Results of a radial-basis-function (RBF) NN doing the same job are shown in
Figs. 3.7 and 3.8. In both cases the training error criteria were set to be 0.4 10 3 .
From Figs 3.5 and 3.7 we can see that both the FF NN and RBF NN give a very good
simulation of the relation GA = f ( Ac , Lc ) , except at points outside the training data range of
variable Lc ( cf. Fig. 3.4). At the "inside" points, or positions where interpolations are made, the
abstract values of the relative errors are well below the 1%. On the other hand, at points outside the
training data range of variable Lc , the relative errors can be as high as 3~5%. This provides another
proof to the fact mentioned before, that interpolation using NN will give results more accurate than
extrapolation.
26
GA Training Data
x10 6
2.5
2
1.5
0
0
2
4
1
Lc
(m
)
(m
Ac
8
10
x10-4
12
2.4
2.2
x10
-4
2
1.8
1.6
1.4
1.2
1
0.8
0.6
0.4
0.2
0
10
12
Lc(m)
GA
Ac(m )
CHAPTER 3
27
x10 6
GA(FF)
0
2
0
4
0.5
Lc( 6
m)
1
8
1.5
10
x10 -4
2
(m
Ac
12
FF Simulation Errors
X
0.4
9
0.4
-0.0
x10 -2
2
-0.0
2
-0.68
-1.1
7
-0.0
85
-2.
0
-0.
07
- 0.
-2
07
-0.
-4
2
0.5
Lc
(m 6
)
1
8
1.5
10
(m
Ac
x10 -4
{GA-GA(FF)}/GA
CHAPTER 3
28
2
0
GA(RBF)
x10
2
0
4
Lc 6
(m
)
1
8
10
(m
Ac
x10
-4
12
0.9
.96
8
2.4 96 1
1. 1.43 .90 8
5
0
0.3 -0.1
3.43
1.41
0
0.9
0
0.9
x10 -2
4
5
-0.1
5
-0.1
-0.1
0
0
0
5
-0.1
.1
5
2
4
Lc( 6
m)
8
10
(m
Ac
12
x10 -4
{GA-GA(RBF)}/GA
1.9
-0
CHAPTER 3
CHAPTER 3
29
Sum-Squared Error
10
10
10
30
-1
5000
10000
15000
20000
Epoch
10
Sum-Squared Error
CHAPTER 3
101
10
10-1
0
5000
10000
15000
20000
Epoch
CHAPTER 3
10
31
Sum-Squared Error
100
10-1
10
-2
10-3
10
-4
10-5
50
100
150
200
250
Epoch
EA( 10 7 N )
2.659
present (FF
simulated, trained
by trainbpa, 3
variables)
2.60
present
(RBF
simulated,
3vaiables)
2.66
present (FF
simulated, trained
by trainlm, 3
variables)
2.656
U. Lee
Noor et
al
2.71
2.53
GA( 10 6 N )
2.183
2.24
2.17
2.186
2.2
2.2
EI( 10 7 N m 2 )
8.147
8.08
8.17
8.148
8.20
8.01
Table 3.2 Comparison of Continuum Model Properties for a Lattice Repeating Cell
Features of the cell: E = 7.17 1010 N / m 2 , Lg = 5.0m
Variables specified: Lc = 7.5m, Ac = 8 10 5 m 2 , Ad = 4 10 5 m 2 , Ag = 6 10 5 m 2
present (FF simulated, trained
by trainlm, 4 variables)
2.682
U. Lee
Noor et al
EA( 10 7 N )
Present
(FEA)
2.659
2.71
2.53
GA( 10 6 N )
2.183
2.134
2.2
2.2
EI( 10 7 N m 2 )
8.147
8.181
8.20
8.01
Chapter 4
An Approach for the Solution of Mindlin
Plates
In this chapter a FSDT (the Reissner-Mindlin theory) method, using the Ritz method with the
Legendre or the Chebyshev polynomials being employed as the trial functions, is derived to solve
plate problems. Formulations are such that there is no limitation on the plate thickness variation
and therefore can be used to deal with real-life wings in the next chapter.
v ( x, y , z, t ) = v0 ( x, y , t ) + z y ( x, y , t )
w( x, y , z , t ) = w0 ( x, y, t )
(4.1)
32
CHAPTER 4
33
where as shown in Fig. 4.1, u, v, w are displacements in the x, y , z direction respectively, subscript
0 refers to quantities associated with the plane z = 0 , x and y are the rotations about the y and
x direction respectively. It is assumed here that the middle surface of the plate is without or with
a very small curvature, therefore z = 0 can be considered to be the middle surface.
From Eq. (4.1) we can get the strains:
u u0
=
+z x
x x
x
v v0
y =
=
+z y
y y
y
w
z =
=0
u v u0 v0
= 2 xy =
+
=
+
+ z( x + y )
y x y
x
y
x
w0
v w
yz = 2 yz =
+
= y +
z y
y
w0
w u
zx = 2 zx =
+
= x +
x z
x
x =
xy
(4.2)
y, v
z, w
(-1,-1)
(-1,1)
z
2
x, u
(1,-1)
(1,1)
CHAPTER 4
34
x = N i ( , ) xi
i =1
4
y = N i ( , ) yi
i =1
(4.3)
where
N 1 ( , ) = 14 (1 )(1 )
N 2 ( , ) = 14 (1 + )(1 )
N 3 ( , ) = 14 (1 + )(1 + )
N 4 ( , ) = 14 (1 )(1 + )
(4.4)
x
y
(4.5)
= 4 [(1 + )( x3 x4 ) + (1 )( x2 x1 )]
x 1
= 4 [(1 + )( x3 x2 ) + (1 )( x4 x1 )]
.
y 1
= 4 [(1 + )( y 3 y 4 ) + (1 )( y 2 y1 )]
y 1
= 4 [(1 + )( y3 y 2 ) + (1 )( y 4 y1 )]
(4.6)
y
[ J ] 1 =
1
J
J 22
J
21
J 12 J 11
=
J 11 J 21
J 12
J 22
(4.7)
CHAPTER 4
x y x y
J = .
35
(4.8)
We express the terms on the plane z = 0 in Eq. (4.1), i.e. u0 , v0 , w0 , x and y , in the
following forms
I
J
u 0 = {B IJ }T {qU } = U ij (t ) Bi ( ) B j ( )
i =1 j =1
K
L
T
v 0 = {B KL } {qV } = Vkl (t ) Bk ( ) Bl ( )
k =1 l =1
M N
w0 = {B MN }T {qW } = Wmn (t ) Bm ( ) Bn ( )
m =1 n =1
P Q
T
x = {B PQ } {q X } = X pq (t ) B p ( ) B q ( )
p =1 q =1
R S
y = {B RS }T {qY } = Y rs (t ) Br ( ) B s ( )
r =1 s =1
(4.9)
or
u0
{qU }
v
{q }
0
V
T
T
T
T
T
T
w0 = {{B IJ } , {B KL } , {B MN } , {B PQ } , {B RS } }{qW } = {B} {q}
{q }
x
X
{qY }
y
(4.10)
{q X } = { X 11 , , X PQ }T , {qY } = {Y11 , , Y RS }T
(4.11)
{B } = {B1 ( ) B1 ( ), , B ( ) B ( )}T
= IJ , KL, MN , PQ, RS
(4.12)
CHAPTER 4
36
is the Ritz base function vector, in which Bi (x ) can either be chosen to be the Legendre
polynomials or the Chebyshev polynomials:
Bi ( x ) = Pi 1 ( x ) or Bi ( x ) = Ti 1 ( x )
where
P0 ( x ) = 1
P1 ( x ) = x
2n + 1
n
Pn +1 ( x ) = n + 1 xPn ( x ) n + 1 Pn 1 ( x ), n = 1, .
(4.13)
and
T0 ( x ) = 1
T1 ( x ) = x
T ( x ) = 2 xT ( x ) T ( x ), n = 1, .
n
n 1
n +1
(4.14)
The foremost 10 non-constant items of the Legendre polynomials and the Chebyshev
polynomials are shown in Figs. 4.2 and 4.3 respectively.
First 10 Lengendre polynomials
1
0.75
Pn(x) (n=1,...,10)
0.5
0.25
0
-0.25
-0.5
-0.75
-1
-1
-0.5
0.5
CHAPTER 4
37
Tn(x) (n=1,...,10)
0.5
0.25
0
-0.25
-0.5
-0.75
-1
-1
-0.5
0.5
1
2
{ } { }dV
T
(4.15)
Introducing the stress-strain relations by the generalized Hookes law, we have for Eq. (4.15)
U=
1
2
{ } [ D ]{ }dV
T
(4.16)
Note that { } = [ D ]{ } and [ D ]T = [ D ] is assumed, and the integration domain V in Eqs. (4.15)
and (4.16) includes all and only the spaces the components of the wing occupy.
Using Eqs. (4.5) and (4.7), we can write
CHAPTER 4
38
y f
f
x
= [ J ]T fx
f
y
y
y
f
f
f
J 11
x
f = ([ J ]T ) 1 = ([ J ]1 )T =
f
f
J 12
f
J 21
J 22 f
(4.17)
x J 11
y 0
{ } = xy = J 12
0
yz
zx 0
J 21
0
J 22
0
0
0
J 12
J 11
0
0
= [T ]{ }
0
J 22
J 21
0
0
0
0
0
J 12
J 11
0
0
0
J 22
J 21
z J 11
0
z J 12
0
0
z J 21
0
z J 22
0
0
0
z J 12
z J 11
0
0
0
z J 22
z J 21
0
0
0
0
0
0
1
u0
u0
v0
v0
0 w0
0 w0
0 x
1 x
0 y
y
x
y
(4.18)
CHAPTER 4
I J
U ij (t ) Bi( ) B j ( )
i =I 1 jJ=1
U ij (t ) Bi ( ) B j ( )
i =1 j =1
K L
U kl (t ) Bk ( ) Bl ( )
u0 k =1 l =1
u0 K L
U kl (t ) Bk ( ) Bl( )
v0 Mk =1 Nl =1
v0
U mn (t ) Bm ( ) Bn ( )
m =1 n =1
w0 M N
w0 U mn (t ) Bm ( ) Bn ( )
{ } = x = mP=1 nQ=1
= [C ]{q}
U pq (t ) B p ( ) Bq ( )
x p =1 q=1
y P Q
U pq (t ) B p ( ) Bq ( )
y p =1 q=1
R S
x U rs (t ) Br ( ) Bs ( )
rR=1 sS=1
y
U
(
t
)
B
(
)
B
(
)
rs
r
s
r =1 s =1
P Q
U pq (t ) B p ( ) Bq ( )
p =1 q=1
R S
U rs (t ) Br ( ) Bs ( )
r =1 s =1
39
(4.19)
where
0
0
0
0
[ B IJ , ]
0
[ B KL , ]
0
0
0
0
0
[ B MN , ]
0
0
[C ] = 0
0
0
[ B PQ , ]
0
0
0
0
0
[ B RS , ]
T
0
0
{B PQ }
0
0
0
0
0
0
{B RS }T
B1 ( ) B1 ( ) B1 ( ) B2 ( ) B ( ) B ( )
[ B , ] =
B1 ( ) B1 ( ) B1 ( ) B2 ( ) B ( ) B ( )
{B } = {B1 ( ) B1 ( ) B1 ( ) B2 ( ) B ( ) B ( )}T
= IJ , KL, MN , PQ, RS .
and {q} is the general displacement vector shown in Eq. (4.10).
(4.20)
CHAPTER 4
40
1
2
{q} [C ]
T
(4.21)
If we write
U = 12 {q}T [ K ]{q}
(4.22)
(4.23)
This is the stiffness matrix of the plate in terms of {q} . Details for the constitutive matrix [D ]
for different parts of the wing structure can be found in Appendix A.
1
2
v dV = {v} {v}dV
2
1
2
(4.24)
d v
{v} = = 0 + z y = 0 1 0 0 z w0 t = [ Z ][ H ]{q}
t
t t
0 0 1 0 0 x t
w
y t
t
where
1 0 0 z 0
[ Z ] = 0 1 0 0 z
0 0 1 0 0
u
{d } = v is the displacement vector, and
w
(4.25)
CHAPTER 4
{B IJ }T
0
[H ] = 0
0
0
0
0
{B KL }
0
0
0
0
0
{B MN }T
0
{B PQ }
41
{B RS }T
0
0
0
0
(4.26)
1
2
(4.27)
where
1
0
[ ZZ ] = [ Z ]T [ Z ] = 0
z
0
0
1
0
0
z
0 z
0 0
1 0
0 z2
0 0
0
z
0
0
z 2
(4.28)
Compare
T = 12 {q}T [ M ]{q}
(4.29)
(4.30)
which is the mass matrix of the plate in terms of the general velocity vector, {q}
It should be pointed out that the present formulation can deal with quadrilaterals of any shape
(see Fig.4.1), which include a skewed trapezoid, the usual shape for a wing plan form. Also, the
formulation can be extended to more complicated cases. When the wing plan form is composed of
several quadrilaterals or trapezoids, one can obtain the whole stiffness and mass matrices by
assembling the corresponding matrices of the component elements, as has been done in Tizzi 30 . A
brief description of how to do in this kind of scenario can be found in Appendix B.
Chapter 5
Equivalent Plate Analysis of Built-Up Wing
Structures
Now we want to solve a wing problem by assuming that the wing behaves like a plate. This
assumption is very reasonable as long as the wing has a small thickness-chord ratio. The
formulations for a general quadrilateral plate derived in Chapter 4 will be used to deal trapezoidal
built-up wing structures in this chapter. This is accomplished by evaluating Eqs. (4.20) and (4.27)
for all the wing components. How to deal with the boundary conditions and the convergence
problem are discussed briefly. Once the total stiffness and mass matrices are determined, various
static and dynamic problems can be solved based on their utilization.
42
CHAPTER 5
I = F ( x, y , z )dV =
1 1
G ( , )dd
43
(5.1)
where
Nz
G ( , ) = F [ x ( , ), y ( , ), z ] J dz
zi 2
i =1
(5.2)
zi 1
here N z is the number of integration zones in z -direction, and zi1 and zi 2 are integration limits of
the i -th zone.
Using the Gaussian quadrature, we can get the numerical value of integral in Eq. (5.1) as
Mg Ng
I g i
(M g )
( Ng )
gj
G[ i
(Mg )
, j
(Ng )
(5.3)
i =1 j =1
(M g )
where g i
(Ng )
,gj
(M g )
, j
( Ng )
represent the number of sampling points used in the and directions respectively.
For a wing composed of skins, spars and ribs, the integrals in Eq. (5.1) can be detailed as
follows:
5.1.1 Skins
The sketch of skins at a wing section is shown in Fig. 5.1. Particulars of integration for skins:
z L + 2 t L F J dz + zU + 2 tU F J dz dd
zU 12 tU
1 1 z L 12 t L
F ( x, y, z )dV =
V
(5.4)
where t L,U = t 1 + tan 2 L,U , subscript L, U indicate the lower and upper skin respectively. It
should be noted that for skins composed of laminated layers the skin contributions can be obtained
by simply adding up the efforts of all the layers, with the material constitutive matrix of each layer
being treated separately (see Appendix A).
CHAPTER 5
44
ZL
ZU
t0
Calculation of Eqs. (4.23) and (4.30) using Eq. (5.4) give the stiffness and mass matrices of the
skins: [ K skin ] and [ M skin ] .
5.1.2 Spars
Their contribution to the stiffness and mass matrices can be calculated by performing the
integrals for each spar. Representative dimensions of a spar are shown in Fig. 5.2. Thus, for a spar
cap, we have:
1
= d (l1 c )d 12
1
1
zL + 2 tL
1
zL + 1 tL + h
l1
c
z L + 2 t L + h + zU 2 t L
l1 d + 1
zU 12 tU h1 F [ x ( , ), y ( , ), z ] J dz
zL 2 tL
s ( )
F ( x , y , z )dxdydz = d
s ( ) +
zU 1 t L
2
zU 1 tU h1
2
(5.5)
F {x [(l c ) + ( ), ], y [(l c ) + ( ), ], z} J dz
1
s
1
s
F ( x, y, z )dxdydz =
V
s +
zU 12 tU h1
z L + 12 t L + h
d (t1 c )d
F [ x ( , ), y ( , ), z ] J dddz
F {x [(t1 c ) + s ( ), ], y [(t1 c ) + s ( ), ], z} J dz
(5.6)
CHAPTER 5
45
Calculation of Eqs. (4.23) and (4.30) using Eqs. (5.5) and (5.6) give the stiffness and mass
matrices of the skins: [ K spar ] and [ M spar ] .
l1 , l 2
h1 ,h 2
Skin
Spar or rib cap
Spar or rib web
t 1,t 2
Rib, r ()
x,
Spar, s()
3 Wing
Spar
orspar
Rib and rib
Fig. 5.2Fig.
Sketches
for
wing
5.1.3 Ribs
The contributions of the ribs to the stiffness and mass matrix can be calculated in a manner
similar to the one used for spars. The dimensions of a rib are also given in Fig. 5.2.
For a rib cap:
l2
zU + 2 t L + h2 + zU 2 tU
l2
zU 12 tL h2 F [ x( , ), y( , ), z ] J dddz
1 r z L + 12 t L
s
F ( x, y, z )dxdydz =
V
= d (l 2 s )d 1
1
1
zL + 2 tL
1
r +
z L + 12 t L + h2
zU 12 t L
zU 12 tU h2
(5.7)
CHAPTER 5
46
F ( x, y , z )dxdydz =
V
r +
zU 12 tU h2
z L + 12 t L + h2
d (t 2 s )d
F [ x ( , ), y ( , ), z ] J dddz
(5.8)
F {x [ , (t 2 s ) + r ( ) ], y [ , (t 2 s ) + r ( ) ], z} J dz
The same as for the spars, integration on ribs can be obtained by summing up contributions from all
the ribs.
Calculation of Eqs. (4.23) and (4.30) using Eqs. (5.7) and (5.8) give the stiffness and mass
matrices of the skins: [ K rib ] and [ M rib ] .
1
2
12
u 02 + y v 02 + z w02 + x x2 + y y2 )dl
(5.9)
CHAPTER 5
47
(5.10)
Considering Eqs. (4.9) and (4.11), and comparing Eq. (5.10) with U spring = 12 {q}T [ K spring ]{q} ,
we obtain the stiffness matrix for the springs as
[ K spring ] = diag ([ K x ], [ K y ], [ K z ], [ K x ], [ K y ])
(5.11)
where
1
[ K x ] = 12 c r x {B IJ ( ,1)}{B IJ ( ,1)}T d
1
1
T
1
[ K y ] = 2 c r y {B KL ( ,1)}{B KL ( ,1)} d
1
1
[ K z ] = 2 c r z {B MN ( ,1)}{B MN ( ,1)}T d
1
1
1
[ K x ] = 2 c r x {B PQ ( ,1)}{B PQ ( ,1)}T d
1
1
1
[ K y ] = 2 c r y {B RS ( ,1)}{B RS ( ,1)}T d
1
(5.12)
The magnitudes of x , y , z , x , and y must be large enough such that the boundary
conditions are properly simulated. But they cannot be too large, or else all the stiffness values other
than those of these springs will be pushed beyond the significant figures of the computation system.
This is to say, the spring magnitudes need to be within a range in order that the boundary condition
is properly simulated. This range depends on the specifics of the computation environment and the
computer system. In Figs. 5.3 and 5.4 the first 10 natural frequencies of a wing that we shall
specify in 5.4 , obtained using a method to be explained in 5.3 , are shown with regard to the spring
magnitudes. In these cases, it is assumed x = y = z = x = y = Spring value , and 6 and 8
terms of the Lengendre polynomials are used respectively. Since the computation is performed
using MATLAB 5.2, double precision is used. It can seen from both Figs. 5.3 and 5.4 that within the
spring value of 10 8 ~16 (lb / in 2 or lb / in ) , all of the natural frequencies are stable. In all the
following cases, the spring value will be specified as 1012 .
CHAPTER 5
48
2500
K=6
W1,...,W10
2000
1500
1000
500
10
15
20
CHAPTER 5
49
2500
K=8
W1,...,W10
2000
1500
1000
500
10
15
20
CHAPTER 5
50
q
j
=0
q
j
j = 1, , , = IJ , KL , MN , PQ , RS .
(5.13)
(5.14)
where V is the potential energy, and by using Eqs. (4.22), (4.29) and (5.13), we can find the
natural frequencies and mode shapes for the free vibrating wing by solving the following
eigenvalue problem
[K total
M total ]{x} = 0
(5.15)
where
[ K total ] = [ K strain ] + [ K spring ], [ K strain ] = [ K skin ] + [ K spar ] + [ K rib ] ,
(5.16)
(5.17)
CHAPTER 5
51
Wing I: sweep angle = 30 $ , span = 192 in , chord length at root = 72 in , chord length at tip
= 36 in ;
Wing II: sweep angle = 24.891$ , span = 72.857 in , chord length at root = 49.627 in , chord
length at tip = 24.332 in .
Figs. 5.5 and 5.6 show the first 10 natural frequencies of Wing I and II when 4 to 10 terms of
the Legendre polynomials are used. More cases of convergence tests for plates are reported in
Lovejoy and Kapania 20, 21 . We can have the following rule of thumb: when K terms of trial
functions are employed, the first K natural frequencies will have converged or nearly converged
values. Based on this rule, either K = 6 or K = 8 is used in all calculations in following sections
and chapters.
CHAPTER 5
52
4000
N=1
N=2
N=3
N=4
N=5
N=6
N=7
N=8
N=9
N=10
3000
2000
1000
10
CHAPTER 5
53
10000
N=1
N=2
N=3
N=4
N=5
N=6
N=7
N=8
N=9
N=10
9000
8000
7000
6000
5000
4000
3000
2000
1000
0
10
CHAPTER 5
54
Q = Q x + Q y + Q z
(5.18)
where
Q x = Px ( x, y , t ) u( x, y , z, t ) dx dy
Q y = Py ( x, y , t ) v ( x, y , z, t ) dx dy
Q z = Pz ( x, y , t ) w( x, y , z , t ) dx dy
(5.19)
Q x = Px (u0 + z x ) dx dy
Q y = Py (v0 + z y ) dx dy
Q z = Pz w0 dx dy
(5.20)
Using Eq. (4.6), that is, approximating the displacements u 0 , v 0 , w 0 , x , y in terms of the Ritz
functions, the total work done by the external force on the whole wing surfaces is given by
Px {B IJ }T {qU } + Py {B KL }T {qV } + Pz {B MN }T {qW }
Q =
dxdy
T
T
+ zPx {B PQ } { x } + zPy {B RS } { y }
= {P}T {q}
(5.21)
(5.22)
in which
{P1, 2 ,3 } = Px , y , z ( x, y, t ){B IJ , KL , MN }T dxdy
{P4 ,5 } = z Px , y ( x, y , t ){B PQ , RS }T dxdy
(5.23)
If the external force is a concentrated force, the above derivations can be simplified. For instance
for the first component of the generalized load vector we have
CHAPTER 5
P1 = Px ( x ( a , a ), y ( a , a ), t ){B IJ ( a , a )}T
55
(5.24)
where ( a , a ) is the transformed coordinates of the point where the load is applied.
Using the principle of virtual work, we have, for the static case, the following relation for the
generalized displacement vector {q} and generalized load vector {P}
[ K ]{q} = {P}
(5.25)
CHAPTER 5
56
agreement with those obtained using the FEA. The relative differences of the natural frequencies
for the first 8 modes are within -0.62~2.12%.
by FEA
Fig. 5.7 Mode Shapes and Natural Frequency f ( rad / s ) for a Trapezoidal Plate
CHAPTER 5
57
CHAPTER 5
by FEA
58
CHAPTER 5
59
CHAPTER 5
by FEA
60
Fig. 5.9 Mode Shapes and Natural Frequency f ( rad / s ) for the Solid Wing
CHAPTER 5
61
CHAPTER 5
62
CHAPTER 5
by FEA
63
CHAPTER 5
64
Table 5.1 Natural frequencies (Hz) of the cantilevered swept-back box wing
Mode
No.
1
2
3
4
5
Description of
Mode Shape
1st bending
In plane
1st torsion
2nd bending
2nd torsion
FEA
(Livne 31 )
115.6
317.6
418.4
576.4
1086
FSDT
(Livne 31 )
114.7
312.4
428.9
575.3
1125
FEA
(present)
116.6
327.9
409.4
572.1
1064
FSDT
(present)
118.0
349.7
419.1
571.1
1090
CHAPTER 5
65
CHAPTER 5
66
-3
x 10
0.5
-0.5
-1
-1.5
100u(FEM)
100v(FEM)
w(FEM)
100u(present cal.)
100v(present cal.)
w(present cal.)
-2
-2.5
-3
50
100
150
Y (inch)
x 10
0.5
0
-0.5
-1
-1.5
-2
100u(FEM)
100v(FEM)
w(FEM)
100u(present cal.)
100v(present cal.)
w(present cal.)
-2.5
-3
-3.5
50
100
150
Y (inch)
CHAPTER 5
67
0.01
-0.01
-0.02
-0.03
100u(FEM)
100v(FEM)
w(FEM)
100u(present cal.)
100v(present cal.)
w(present cal.)
-0.04
-0.05
-0.06
50
100
150
Y (inch)
0.01
0
-0.01
-0.02
-0.03
-0.04
100u(FEM)
100v(FEM)
w(FEM)
100u(present cal.)
100v(present cal.)
w(present cal.)
-0.05
-0.06
-0.07
50
100
150
Y (inch)
x 10
68
-5
20
100u(FEM)
100v(FEM)
w(FEM)
100u(present cal.)
100v(present cal.)
w(present cal.)
15
10
-5
50
100
150
Y (inch)
-4
8
7
100u(FEM)
100v(FEM)
w(FEM)
100u(present cal.)
100v(present cal.)
w(present cal.)
CHAPTER 5
5
4
3
2
1
0
-1
50
100
150
Y (inch)
CHAPTER 5
69
Measured 31,60
1.81 10 4
2.21 10 4
FEA(present)
FSDT(present)
1.79 10
2.19 10 4
1.74 10 4
2.20 10 4
CHAPTER 5
70
0.25
0.5
0.75
2
Lower Skin (FEM)
Upper Skin (FEM)
Lower Skin (present cal.)
Upper Skin (present cal.)
0.25
0.5
0.75
CHAPTER 5
71
6
4.7
3.65
4.7
4.70
60
52.
5.9
2.92 .25
2
3.
30
50
0.
Re
3 .6
3.6
4.35
3 .6 5
2.2
2.60
4
2
5
0.85
2.60
2.25
lat 0.2
1.55
ive
dis 0.4
tan
ce 0.6
0.50
to
lea
din 0.8
g-e
1
dg 1
e
5
0.8
t
roo
e
th
0.2
0.4
to
ce
n
ta
dis
0.6
0.8
la
Re
tive
5.
04
CHAPTER 5
72
4.00
5
4.3
3 .3
2.9
2.95
1.90
90
1.
Re
0
0
1.9
1.55 0
lat 0.2
1.20
ive
dis 0.4
tan
ce 0.6
to
lea
din 0.8
g-e
dg 1
e
t
roo
e
th
0.2
0.4
to
ce
n
ta
dis
0.6
0.8
1
la
Re
tive
CHAPTER 5
73
CHAPTER 5
74
needed for calculating the element matrices. It is based on this reasoning that a more efficient
approach is pursued in Chapter 7.
It is worth noting that for the FEA of the built-up wing the ratio of number of elements to DOF
is the largest, more than 3 times higher than the other cases. This indicates a relative larger
computational effort on calculating and assembling the element matrices in the FEA, also more
efforts for calculating the contributions from the structure components in the EPA. Therefore it
seems that the difference in efficiency of FEA and EPA is more reliably reflected in the difference
between the DOFs.
Table 5.3 Comparison of FEA and EPA in terms of DOF and Number of Elements
FEA
Example
DOF
No. of
Elements
DOF
k=6
1350
3300
600
200
250
370
180
180
180
EPA
DOF
No. of
k=8
Elements
k=6
320
36
320
36
320
36
No. of
Elements
k=8
64
64
64
Chapter 6
Modal Response Using Sensitivity Techniques
and Direct Application of Neural Networks
The modal response of wing structures is very important for assessing their dynamic response
including dynamic aeroelastic instabilities. Moreover, in a recent study 61 an efficient structural
optimization approach was developed using structural modes to represent the static aeroelastic
wing response (both displacements and stresses).
In this chapter, the natural frequencies of general trapezoidal wing structures are to be
approximated using shape sensitivities up to the 2nd order, and different approaches of computing
the derivatives are investigated. The baseline design and shape sensitivities are calculated based on
the equivalent plate-model analysis (EPA) method developed by Chapter 4 and 5. For
comparison, an efficient method that employs the artificial neural networks to relate the natural
frequencies of a wing to its shape variables is also established. An example of a 4 3 full factorial
experimental design, i.e., 4 levels in 3 variables, is treated by these methods to display their
respective merits.
75
CHAPTER 6
76
(6.1)
where s is the length of semi-span, a and b are the chord-length at wing tip and root
respectively, as shown in Fig. 6.1.
b
s
a
x
The sensitivities for the design parameters at a baseline design point indicate trends in the
response of the baseline design if the parameters are perturbed. Usually, only the first order
derivatives are used:
n
f 0 i
( x x0i )
i
x
(6.2)
CHAPTER 6
where
77
f 0
f
= i
i
x
x
is the sensitivity at the baseline point with respect to the i -th design
parameter. For a more accurate approximation, we can use higher-order derivatives in the Taylors
expression:
n
n
+ 12 ( x i x0i ) i f ( x01 , x02 , , x0n )
x
i =1
where besides the first order derivatives, second order derivatives
(6.3)
2 f0
(i, j = 1, n ) are also
x i x j
used.
CHAPTER 6
78
In the present study, a simple yet effective method is used. In this method, any ordinary
eigenvalue solver can be used, and the modes of the baseline structure are chosen as the
benchmarks. By using the modal assurance criterion (MAC) defined as
MAC ji =
({ j }T { i }) 2
({ j }T { j })({ i }T { i })
(6.4)
where { j } and {i } are the eigenvector for the perturbed and the baseline design respectively, if
MAC ji = max( MAC li ) , we say that the j -th mode of the perturbed design corresponds to the i -th
l
f ( x + x ) f ( x x )
+ O ( x 2 )
2x
(6.5)
CHAPTER 6
f ( x ) =
79
f ( x + x ) 2 f ( x ) + f ( x x )
+ O ( x 2 )
2
x
(6.6)
where
x = x
(6.7)
in which is the relative step size, but herein it is simply called the step size. Eq. (6.5) can be
applied twice for evaluating the mixed second order derivatives such as
2 f0
(i j ) .
x i x j
The analytical approaches for shape sensitivities of modal response can be based on the
following equations
i
[ M ]
[ K ]
= { i }T
i
{ i }
(6.8)
{ i } n
= ij { j }
j =1
(6.9)
where
ij =
1
[ M ]
[ K ]
{ i }T
i
{ i }, j i
( i j )
1
T [ M ]
ii = { i }
{ i }
(6.10)
here is the shape variable, i and { i } are the i -th eigenvalue and eigenvetor, and { i } is massnormalized such that { i }T [ M ]{ i } = 1 . Eqs. (6.8) and (6.9) were first derived by Wittrick 66 and
Fox and Kapoor 67 respectively. One can find more on this topic in Ref. 68.
The major difficulty of applying Eqs. (6.8) and (6.9) lies in the calculation of
[ K ]
and
[M ]
[ K ]
. For instance, consider
. According to Chapter 4, the stiffness matrix [K ] is
formulated as an integral
CHAPTER 6
z2
T
[
]
[T ]T [ D ][T ] J dz [C ]dd
C
1 1
z
1
1 1
80
(6.11)
[C ] [G ][C ]dd
T
z2
where only the inner part [G ] = [T ]T [ D ][T ] J dz is a function of the shape variables, and the
z1
[ K ] 1 1
[G ]
= [C ] T
[C ]dd
1
1
in which
(6.12)
[G ]
can either be determined analytically or numerically.
People often make use of the advantages of both the finite difference and analytical approaches
in different stages of obtaining some complicated sensitivities. While trying to use the analytical
approach as much as possible, in other parts of the process the finite difference is used, as in the
cases of Refs. 40 and 42. This kind of approach is usually called semi-analytical.
In summary, there are three approaches to calculate the first order modal sensitivities:
(i) analytical approach: Eqs.(6.8)~(6.10) are used, and
[M ]
[K ]
and
are determined
analytically.
(ii) semi-analytical approach: Sensitivities
[M ]
[K ]
and
in Eqs.(6.8)~(6.10) are
[K ]
[G ]
, Eq. (6.12) is used where
is calculated
{ i }
i
and
are determined using Eq. (6.5) directly.
For the second order sensitivities, there can still be three approaches as specified above. While
the formulation for the analytical approaches is becoming more complicated, a scheme as simple as
Eq. (6.6) can be used for the finite difference approach.
CHAPTER 6
81
R 1
Ri ( p ) R ( p i ) + ( p p i )
+ ( p p i ) T
R
pi 2
p i
T
(6.13)
where p = ( v 1 , v 2 , , v n ) T is an arbitrary point in the design space, p i = (v i1 , v i2 , , v in ) T is the i th node point in the design space, Ri ( p) is the response at p estimated by using the response and
its sensitivities at pi , R( p i ) is the response at the i -th node point p i , and
= 1 , 2 ,, n
pi v v
v
.
p = pi
Once there are enough estimates for the response at p using Eq. (6.13), a more accurate
evaluation of response at p can be determined using the following weighting procedure involving
the so-called exponentially decaying influence function 35 :
R ( p ) = wi ( p ) R i ( p )
(6.14)
where i ranges through the N w design points which are closest to p , and the weight coefficients
wi ( p) are determined such that its sum is unity:
wi ( p ) =
exp( C p p p i )
exp( C
(6.15)
p pi )
(v
j =1
w ( p) = 1 .
i
CHAPTER 6
82
CHAPTER 6
83
some kind of interaction between the two modes. If an eigenvalue solver that can work more
accurately with repeated eigenvalues is made use of, the situation can be improved.
The effect of step size on the finite difference approach for sensitivities was investigated for all
the four shape variables. The case with the taper ratio is shown in Fig. 6.3. From all the cases, it is
seen that for the best results for both the 1st and 2nd order sensitivities, the step size
Eq. (6.7), should be between 0.005~0.015, and for fairly accurate results
defined in
can be between
0.0017~0.045.
To evaluate
[G ]
analytically proved to be formidable except only in some simplified cases.
In order to compare the sensitivities using the analytical, semi-analytical and finite difference
approach, a special case of the above baseline wing with a constant thickness was considered so
that the analytical derivation of
[G ]
in Eq. (6.12) is not formidable. When is specified as
0.005, it is found that for the 1st order sensitivities to the four shape variables ( , , , and A ) the
relative difference (averaged for the first 10 modes) between the finite difference and analytical
approach is 0.003%, 0.003%, 0.002% and 0.003% respectively. The relative difference between the
semi-analytical and analytical approach is 0.14%, 0.04%, 0.02% and 0.01% respectively. Therefore
in this case the finite difference approach is more accurate than the semi-analytical one, however
both the approaches yield quite accurate results.
CHAPTER 6
84
3500
3000
2500
2000
1500
1000
500
Aspect Ratio
Fig. 6.2
CHAPTER 6
85
2000
step size =
0.0017-0.045
1500
1000
step size =
0.005-0.015
Sensitivities
500
0
-500
-1000
-1500
1st Order
(2nd Order)/200
-2000
-2500
-3000
-4
-3
-2
-1
Fig. 6.3
CHAPTER 6
86
For the original baseline wing, since the derivation of the analytical derivatives for
[G ]
is
too formidable, only the comparison of the 1st order sensitivities using the finite difference and the
semi-analytical approach was made. It is found in this case the sensitivities to the aspect ratio ,
taper ratio and plan area A using both approaches are quite close, the average difference for the
first 10 modes being in the range of 0.5~1.4%. As an example, the 2nd natural frequency w.r.t. A is
shown in Fig. 6.4, where it can be seen that the 1st order sensitivities using the finite difference and
the semi-analytical approach almost coincide with each other. On the other hand, sensitivities to the
sweep angle using the two approaches have had some quite large relative differences especially
for modes whose sensitivity to is small. One such example, the 3rd natural frequency w.r.t. , is
shown in Fig. 6.5, where attention should be paid to the scale for the vertical coordinate to see how
small the sensitivity to is.
It is observed in Fig. 6.5 that, as in the case of the constant-thickness wing, the finite difference
approach has a better performance than the semi-analytical one. In fact, in some extreme cases, the
linear approximation using the first order sensitivity obtained using the semi-analytical approach is
not at all tangent to the actual variation at the baseline point. This is not the case for that using the
finite difference approach, if the step size chosen is not too large. However, the computational
effort for the semi-analytical approach might be less than that for the finite difference approach,
since in the former case the eigenvalue problem needs to be solved only once while in the latter it
needs to be solved twice.
CHAPTER 6
87
780
Actutal Variation
Linear Approx. (fd)
2nd-Order Approx.
Linear Approx. (semi-an)
760
740
720
700
680
Baseline
660
640
620
600
4500
5000
5500
60002
6500
Fig. 6.4 The 2nd natural frequency w.r.t. wing plan area
using 1st and 2nd order sensitivities
7000
CHAPTER 6
88
776
Actual Variation
Linear Approx. (fd)
2nd-Order Approx.
Linear Approx. (semi-an)
775
774
773
772
771
Baseline
770
769
768
767
766
25
30
35
Fig. 6.5 The 3rd natural frequency w.r.t. wing sweep angle
using 1st and 2nd order sensitivities
CHAPTER 6
89
It is obvious from observing Figs. 6.4 and 6.5 that the approximation using sensitivities up to
the second order has much improved the results compared with the case where only the first order
sensitivity is used. Similarly it has been shown in Haftka and Gurdal 69 that, for the stress-ratio in a
three-bar truss, the quadratic approximation is much more accurate than the linear one. Also it can
be seen that the second order sensitivities using the finite difference scheme of Eq. (6.6) are fairly
accurate, at least for the purpose of engineering application. Another advantage of this scheme is
that it shares the perturbation data with the first order sensitivity scheme Eq. (6.5), therefore its
evaluation has no increase in the computational effort at all.
Using the finite difference approach based on Eq. (6.5) the mixed second order sensitivities
2 f0
(i j ) can be readily determined. As an example, the mixed second order sensitivity on
x i x j
and A for the first five natural frequencies were calculated, and the results are listed as follows:
0.0099, 0.0153, 0.0353, 0.0494 and 0.0156.
6.6.2 Application of Sensitivity Technique (ST) and Neural Networks (NN)
For a trapezoidal wing, there are four major independent shape variables, i.e. the sweep angle
, the aspect ratio , the taper ratio , and the plan area A . As an example , a 4 3 full factorial
experimental design with 4 levels in , , and respectively, was used. Particulars of the levels
of every variable are as follows: = [0 $ ,10 $ ,20 $ ,30 $ ] , = [1.0,1.5,2.0,2.5] , and
= [0.3,0.4,0.5,0.6] . The plan area is chosen to be a constant: A = 3500in 2 . The other particulars
are the same as in 6.7.1.
The natural frequencies of the wing structure at the 64 node points in the design space were
calculated using EPA, and the 1st and 2nd order sensitivities at these points were also determined by
finite difference using EPA 3 . For each mode, a feed-forward neural network with a structure of
3 15 10 1 , i.e. 3 inputs, 15 neurons in the first hidden layer, 10 neurons in the second hidden
layer, and 1 output, is trained using the MATLAB NN Toolbox function trainlm that trains feedforward network with the Levenberg-Marquardt algorithm 8 . There are 64 sets of training data,
CHAPTER 6
90
which are non-dimensionalized before the training process. Once the networks are trained, the
input-output relationships can be readily retrieved by using the function simuff.
For the application of sensitivity technique, the major task is to evaluate the sensitivities, and
to generate responses at an arbitrary design point using Eqs. (6.13) and (6.14) does not need large
amount of CPU time. The constant C p in Eq. (6.14) was specified to be 10, and N w = 10 was used.
Shown in Fig. 6.6 are the first 6 natural frequencies of 20 randomly chosen wing structures
inside the box defined in terms of lower and upper bounds on the design variables specified above.
From the figure it can be seen that both of the results given by NN and ST are in very good
agreement with the desired values (those given by the EPA) except for a few cases where there are
some differences. These cases might be caused by the unstable performance of the algorithm used
for extracting eigenvalues in the EPA near the mode-crossing points, as shall be shown in Figs 6.7
and 6.8. In order to see the effects of sensitivity order, a randomly chosen path inside the design
space box is defined as
v j = v 0j (1 a j ) + v1j a j , j = 1,2,3
v 1 = , v 2 = , v 3 = ,
nj
j
a = s , n j = r j /(1 r j ).
(6.16)
where v 0j and v1j are lower and upper bounds of variable v j , for instance, v 01 = 0 $ , v11 = 30 $ etc.,
s [0,1] is the range of a shape variable, and r j ( j = 1,2,3) are randomly determined values between
0 and 1. Results of natural frequencies of the first 4 modes for wing structures defined by points
along a path with n 1 =0.945, n 2 =8.200, and n 3 =3.203 are shown in Fig. 6.7, where only the 1st
order sensitivities were used, and in Fig. 6.8, where sensitivities up to the 2nd order were used. It
can be seen that when sensitivities up to the 2nd order are used, results are effectively improved.
Generally speaking, neural networks and sensitivity technique can give equally good results,
and the former uses less time than the latter. But both methods, once the NNs are trained or the
sensitivities are obtained, are much more efficient than the EPA . For instance, the CPU times
CHAPTER 6
91
consumed by the EPA, the sensitivity based method and the NN based method are in the ratio of
55 : 1 : 0.06 .
The example used above has only three variables. For design problems with more variables, the
method of NN and ST can still be applied in general, only at the expense of more computing time.
We can expect that similar conclusions to those obtained above still apply to these cases. For a
design problem with very large number of variables, in combination with the NN or ST method,
methodologies to shrink the design space, such as the reasonable design space approach described
in Balabanov et al 70 , can be used. This can make the search of optimal design easier and at the
same time the application of NN or ST more accurate, just as the case in Ref. 26 where the
response surface approximation was used to simulate high-fidelity models. Also for this kind of
high dimensionality design problems, a full multi-level factorial experimental design is almost
impossible to use hence the methods of either NN, or ST, or even response surface are hard to
apply because the cost would be too high. In such a case, an incomplete block statistical
experimental design using the D-optimal criterion 71,72 can be used, which, with much reduced
number of design node points, makes the application of NN or ST possible.
CHAPTER 6
92
4000
by Neural Networks
by Sensitivity Technique
3500
3000
2500
2000
1500
1000
500
1000
2000
3000
4000
Comparison of the natural frequencies of the first 6 modes for wing structures
randomly chosen inside the box of design space, as obtained by the NN and ST
w.r.t. those obtained using a full-fledged EPA
CHAPTER 6
4000
93
3500
3000
by EPA
by Neural Networks
by ST, 1st mode
by ST, 2nd mode
by ST, 3rd mode
by ST, 4th mode
2500
2000
1500
1000
+++++++++++++++++++++++++++++
++++
++
++
++
++
++
++
++
++
+
500
0
0.25
0.5
0.75
Fig. 6.7
Comparison of the natural frequencies of the first 4 modes for wing structures
along a path inside the box of design space (n 1 =0.945, n 2 =8.200, n 3 =3.203)
using only the 1st order sensitivities
CHAPTER 6
4000
94
3500
3000
by EPA
by Neural Networks
by ST, 1st mode
by ST, 2nd mode
by ST, 3rd mode
by ST, 4th mode
2500
2000
1500
1000
+ ++ + ++ ++ + ++ ++ ++ + ++ + ++ ++ + ++ ++ +
++ ++
++
+ +
+ +
++
+ +
+ +
+ +
+ +
+
500
0
0.25
0.5
0.75
Fig. 6.8
Comparison of the natural frequencies of the first 4 modes for wing structures
along a path inside the box of design space (n 1 =0.945, n 2 =8.200, n 3 =3.203)
using sensitivities up to the 2nd order
Chapter 7
Equivalent Skin Analysis Using Neural
Networks
Since the calculation of various integrals in Eqs. (4.23) and (4.30) are time-consuming, it is desired
to replace the actual wing structure by an equivalent continuum model, that is, one that is
composed of a skin-like material, whose constitutive matrix [D ] and distribution of mass in
Eqs. (4.23) and (4.30) respectively are to be decided.
95
CHAPTER 7
96
(7.1)
~
= [ K ij ] = [ D pq ( G pqij ,mn )] = D pq [G ijpq ]
p
p ,q
where g m and g n are the Gauss quadrature weights; the constitutive matrix [D ] relates the stress
and strain vectors by { } = [ D ]{ } , and D pq is the p -th row, q -th column term of the constitutive
matrix;
m ( m = 1, , K ) corresponds to the m -th Gauss integration position in the x-direction,
K is an integer with a usual value of 6 or 8;
n ( n = 1, , K ) corresponds to the n -th Gauss integration position in the y-direction;
p = 1, ,5 is the row number of [D ] ;
q = 1,
G pqij ,mn
,5
5
0
1 q 5
(7.2)
~
E K = wijK K ij ( D pq ) K ij
(7.3)
i, j
CHAPTER 7
97
and noting
~
K ij
D pq
we can obtain the constitutive matrix term [ D pq ] by solving the following linear equation set:
( w
p ', q '
i, j
K
ij
(7.4)
i, j
( p ', q ' )
i, j
K
ij
{G ijp ' q ' + (1 p ' q ' )G qij' p ' }G ijpq D p ' q ' = wijK K ij G ijpq
i, j
(7.5)
where ( p ' , q' ) and ( p, q ) have the followings 15 combinations instead of 25: (1,1), (1,2), (1,3),
(1,4), (1,5), (2,2), (2,3), (2,4), (2,5), (3,3), (3,4), (3,5), (4,4), (4,5), and (5,5).
7.1.2 Mass distribution
Let's write [ M ] = [ M total ] = [ M ij ] as the target matrix, and
~
[ M ] = mn g m g n [ H ]Tmn [ Z ]Tmn [ Z ] mn [ H ] mn = mn [ F ] mn
m
(7.6)
as the mass matrix of the continuum model, where m ( m = 1, , K ) corresponds to the m -th Gauss
integration position in the x-direction, n ( n = 1, , K ) corresponds to the n-th Gauss integration
position in the y-direction, mn is the mass density of the equivalent model at position ( m , n ), m
and n are integration weights, and [ F ] mn = [ H ]Tmn [ Z ]Tmn [ Z ] mn [ H ] mn is a N N matrix varying
with position ( m, n ) ( N = 5K 2 ).
By constructing an error function
CHAPTER 7
98
~
E M = wijM M ij ( mn ) M ij
(7.7)
i, j
~
M ij
E M
M ~
= 2 wij M ij ( mn ) M ij
=0 ,
mn
mn
i, j
and noting
~
M ij
mn
ij
ij
is the i, j -th term of matrix Fmn ),
( Fmn
= Fmn
we can obtain the mass distribution mn by solving the following linear equation set with K 2
unknowns:
( w
m ', n '
i, j
M
ij
ij
ij
) m'n ' = wijM M ij Fmn
, m, n = 1, , K , m' , n' = 1, , K .
Fmij'n ' Fmn
(7.8)
i, j
M
Skin
Skin
max( M ij ) + 1
wij = 10 M ij
i, j
(7.9)
The basic idea behind this choice is that we want to form the equivalent matrices more in the way
of the skin's, which is more like a plate than the other components of the wing, i.e. spars and ribs.
Several choices about the variation of and [D ] have been tried, but it is found that the
present assumptions give the best results in terms of feasibility and accuracy. For instance, to be
consistent with the assumption that each term of [D ] is a constant throughout the wing area, can
also be assumed a constant. This certainly decreases the accuracy of the method due to the loss of
flexibility in varying to simulate the target mass matrix [M ] , but the resultant reduction in
computational effort is small since in the first place, forming Eq. (7.8) and training the -related
neural networks do not need much CPU time. In other cases, [D ] was assumed to be variable in the
span-wise direction or throughout the wing area, but it is found that although the equivalent
~
material is more flexible to simulate the target stiffness matrix [K ] , the resultant [K ] usually has a
CHAPTER 7
99
larger abstract error and the solution of the free vibration problem usually gives worse natural
frequencies. Moreover, the CPU time needed for generating Eq. (7.5), which requires the major
computational effort in our method, increases in a factor of about K (number of Gauss integration
points, usually with a value of 6) in the case of [D ] being variable in the span-wise direction. In the
case of [D ] being variable throughout the wing area, the increase can be as large as K 2 times. As
we shall see in the following examples, these kinds of increase in CPU time are formidable.
CHAPTER 7
100
2
A = 3500 in
n spar = 4
n rib = 10
The "equivalent skin" constitutive matrix and mass density distribution solved using Eqs. (7.5)
and (7.8) are shown in the following and in Fig. 7.1.
1.2515 0.2849 0.0040 0.0035
0.2849 1.1956 0.0091 - 0.0203
D
= 0.0040 0.0091 0.3836 0.0041
E
0.0035 - 0.0203 0.0041 0.0294
0.0157 - 0.0008 0.0123 0.0024
=
0
0.9969
1.0852
1.0649
1.1551
1.2637
1.3450
1.1337
1.2629
1.2084
1.1939
1.1490
1.4151
1.1451
1.2890
1.2303
1.2193
1.1415
1.3742
1.1230
1.2436
1.1869
1.1783
1.2346
1.5221
0.0157
- 0.0008
0.0123 .
0.0024
0.1481
1.1124
1.2368
1.2606
1.4135
1.4525
1.4531
1.1687
1.3352
1.2941
.
1.2507
1.1338
1.2034
Comparison of the target stiffness and mass matrices, [ K t ] and [ M t ] , with their simulated
counterparts, [ K s ] and [ M s ] , is shown in the following and Figs. 7.2~7.7. The relative differences
for the first 10 natural frequencies by EPA and ESA are shown in Table 7.1.
max[ K t K s ]
= 2.49%,
max[K t ]
max[ K t + K s ]
= 4.95%.
max[ K t ]
max[ M t M s ]
= 1.20%,
max[ M t ]
max[ M t + M s ]
= 0.75%.
max[ M t ]
CHAPTER 7
101
Table 7.1 Differences between the natural frequencies by EPA and ESA
Mode number
10
f ESA f EPA
100
f EPA
0.07
2.56
-0.08
1.26
5.41
-1.37
4.24
10.05
2.50
1.40
16
2.32 .2
1.1 6
1.12.91.39
26
1.3
-1
-1
3
-0.5
29
1 . 1 .1 9
6
1.1
1.
16
-0.5
1.26
.193
.3
11.42
1.2
3
1 .3 6
1.25
1.3
61
.3
1.1
13.1
1.106
1.1
1.3
1.5
6
.1
119
1.
Normalized Density
0.5
0.5
1
Fig. 7.1 An example of mass density distribution generated using Eq. (7.8)
1.49
1.46
1.42
1.39
1.36
1.33
1.29
1.26
1.23
1.19
1.16
1.13
1.10
1.06
1.03
CHAPTER 7
102
1E+09
8E+08
6E+08
2E+08
0
-2E+08
0
-4E+08
50
-6E+08
J 100
50
150
150
100
[K]t
4E+08
CHAPTER 7
103
1E+09
8E+08
6E+08
2E+08
0
-2E+08
0
-4E+08
50
-6E+08
J 100
50
150
150
100
[K]s
4E+08
CHAPTER 7
104
0.02
-0.02
0
-0.04
50
100
0
50
150
100
150
Fig. 7.4 Difference between stiffness matrices given by EPA and ESA
([K]t-[K]s)/max([K]t)
CHAPTER 7
105
0
0
-0.5
50
-1
100
0
50
150
100
150
[M]t
0.5
CHAPTER 7
106
0
0
-0.5
50
-1
100
0
50
150
100
150
[M]s
0.5
CHAPTER 7
107
0.02
0.01
0
0
-0.01
50
-0.02
100
0
50
150
100
150
Fig. 7.7 Difference between mass matrices given by EPA and ESA
([M]t-[M]s)/max([M]t)
CHAPTER 7
108
= {1.0,1.75,2.5},
= {0.3,0.45,0.6}.
For each point in this design space, the EPA is carried out, then Eqs. (7.5) and (7.8) are used to
generate the 15 constitutive matrix terms and mass densities at 36 ( 6 6 ) Gauss sampling points.
Upon the obtained "equivalent skin" constitutive matrix [D] and mass density distribution [ ] , the
ESA is performed based on the simulated stiffness and mass matrices. For each of these
parameters, a feed-forward neural network with a structure of 4 15 10 1 , i.e. 4 inputs, 15
neurons in the first hidden layer, 10 neurons in the second hidden layer, and 1 output, is trained
using the MATLAB NN Toolbox function trainlm that trains feed-forward network with the
Levenberg-Marquardt algorithm 8 . Therefore, there are totally 15+36=51 networks to be trained.
There are 81 ( 34 ) sets of training data, which are non-dimensionalized before the training process.
Once the networks are trained, the input-output relationships can be readily retrieved by using the
function simuff.
The major computational effort was spent in generating the 81 sets of training data, with about
15 hours of CPU time being spent on a PII/350 personal computer, while less than 1 hour of CPU
time being used in training the neural networks. A set of results are given in Figs. 7.8 to 7.10 where
49 points, which mean 49 new designs, were randomly chosen within the design space box. Upon
each new design both the EPA and the ESA are performed. The plan forms of the new design are
shown in Fig. 7.8. The first 10 natural frequencies by the EPA and the ESA are compared in Fig.
CHAPTER 7
109
7.9 and their relative differences (based on the EPA results) are shown in Fig. 7.10. It can be seen
that except for a very few cases (2 out of 490), the relative difference is within -10%~10%.
Fig. 7.11 shows 25 new designs through a randomly chosen path inside the design space box
which is defined as
v j = v 0j (1 a j ) + v1j a j , j = 1, ,3
v 1 = , v 2 = , v 3 = ,
nj
j
a = s , n j = r j /(1 r j ).
(7.10)
where v 0j and v1j are the lower and upper bounds of variable v j , for instance, v 01 = 0 $ , v11 = 30 $
etc., s [0,1] is the range of a shape variable, and r j ( j = 1,2,3) are randomly determined values
between 0 and 1. Results of natural frequencies of the first 6 modes for wing structures defined by
points along a path with n 1 =0.945, n 2 =8.200, n 3 =3.203 and n 4 =1.778 are shown in Fig. 7.12,
where it can be seen that results by the EPA and the ESA agree with each other quite well.
CHAPTER 7
110
CHAPTER 7
111
7000
6000
5000
4000
3000
2000
1000
1000
2000
3000
4000
5000
6000
Frequency by EPA
Fig. 7.9 Comparison of the first 10 frequencies by EPA and ESA
7000
CHAPTER 7
112
0.15
Relative error
0.1
0.05
-0.05
-0.1
1000
2000
3000
4000
5000
Frequency by EPA
Fig. 7.10 The relative errors in Fig. 7.9
6000
7000
CHAPTER 7
Fig. 7.11 25 wing plan forms systematically varying through design space I
113
CHAPTER 7
114
6000
+
x
5000
by EPA
1st mode by NN-aided EPA
2nd mode by NN-aided EPA
3rd mode by NN-aided EPA
4th mode by NN-aided EPA
5th mode by NN-aided EPA
6th mode by NN-aided EPA
4000
3000
2000 x x x x x x x x
x x x x x x x
x x
x x
+
+
+
+
+
+
+
+
+
x
+
+
+
+
+ +
+ +
+ +
x
+ + x
+ x
+ +
x +
x
1000
0.25
0.5
0.75
CHAPTER 7
115
= {1.0,1.75,2.5},
2
A = {2000,3500,5000}in .
For each point in this design space, the EPA is carried out, then Eqs. (7.5) and (7.8) are used to
generate the 15 constitutive matrix terms and mass densities at 36 ( 6 6 ) Gauss sampling points,
and the ESA is performed. For each of these parameters, a feed-forward neural network with a
structure of 4 15 10 1 , i.e. 4 inputs, 15 neurons in the first hidden layer, 10 neurons in the
second hidden layer, and 1 output, is trained using the MATLAB NN Toolbox function trainlm that
trains feed-forward network with the Levenberg-Marquardt algorithm. Therefore, there are totally
15+36=51 networks to be trained. There are 81 ( 3 4 ) sets of training data, which are nondimensionalized before the training process. Once the networks are trained, the input-output
relationships can be readily retrieved by using the function simuff.
The major computational effort was spent in generating the 81 sets of training data, with about
45 hours of CPU time being spent on a PII/350 personal computer, while less than 1 hour of CPU
time being used in training the neural networks. A set of results are given in Figs. 7.13 to 7.15
where 25 points, which mean 25 new designs, were randomly chosen within the design space box.
Upon each new design both the EPA and the ESA are performed. The plan forms of the new design
are shown in Fig. 7.13. The first 10 natural frequencies by the EPA and the ESA are compared in
Fig. 7.14 and their relative differences (based on the EPA results) are shown in
Fig. 7.15. It
can be seen that except for a very few cases (3 out of 250), the relative difference is within 10%~10%.
Fig. 7.16 shows 16 new designs through a randomly chosen path inside the design space box
which is defined as
CHAPTER 7
116
v j = v 0j (1 a j ) + v1j a j , j = 1, ,4
v 1 = , v 2 = , v 3 = , 4 = A
nj
j
a = s , n j = r j /(1 r j ).
(7.11)
where v 0j and v1j are the lower and upper bounds of variable v j , for instance, v 01 = 0 $ , v11 = 30 $
etc., s [0,1] is the range of a shape variable, and r j ( j = 1, 4) are randomly determined values
between 0 and 1. Results of natural frequencies of the first 6 modes for wing structures defined by
points along a path with n 1 =0.945, n 2 =8.200, n 3 =3.203 and n 4 =1.778 are shown in Fig. 7.17,
where it can be seen that results by the EPA and the ESA agree with each other quite well.
While the former results are about free vibration frequencies, Figs. 7.18 to 7.21 show some
static results. For an arbitrary new design whose plan form is shown in Fig. 7.18, a down-ward (- z
direction) point force of 1 lb is applied at the mid-point of the wing tip (actually the force is
divided into components acting on the two spar tips close to the mid-point). Fig. 7.19 shows
displacements along the leading-edge by the EPA and the ESA, where u, v, w are displacement
components in the chord-wise, span-wise, and vertical directions respectively. Figs. 7.20 and 7.21
show the Von Mises stress distributions at the wing root and the central spar respectively. It also
can be seen that the EPA and the ESA give very compatible static results.
CHAPTER 7
117
CHAPTER 7
118
7000
Frequency by ESA
6000
5000
4000
3000
2000
1000
1000
2000
3000
4000
5000
6000
7000
Frequency by EPA
Fig. 7.14 Comparison of the first 10 frequencies by EPA and ESA
CHAPTER 7
119
0.15
Relative error
0.1
0.05
-0.05
-0.1
1000
2000
3000
4000
5000
6000
Frequency by EPA
Fig. 7.15 The relative errors in Fig. 7.14
7000
CHAPTER 7
Fig. 7.16 16 wing plan forms systematically varying through design space II
120
CHAPTER 7
121
6000
by EPA
1st Bending mode by ESA
1st Torsion mode by ESA
2nd Bending mode by ESA
In plane mode by ESA
2nd Torsion mode by ESA
3rd Bending mode by ESA
+
x
5000
4000
3000
x
2000
x
+
x
+
x
+
x
+
x
+
x
+
x
+
x
+
x
+
1000
0.25
0.5
x
+
x
+
+
x
0.75
+
x
CHAPTER 7
122
CHAPTER 7
123
5.0E-05
0.0E+00
-5.0E-05
-1.0E-04
-1.5E-04
-2.0E-04
10u (EPA)
10v (EPA)
w (EPA)
10u (ESA)
10v (ESA)
w (ESA)
-2.5E-04
-3.0E-04
-3.5E-04
10
20
30
40
50
60
70
80
Fig. 7.19 Comparison of displacements by EPA and ESA for 1 lb tip force
CHAPTER 7
124
3.5
2.5
1.5
0.5
0.25
0.5
0.75
CHAPTER 7
125
3.5
2.5
1.5
0.25
0.5
0.75
CHAPTER 7
126
= {1.0,1.75,2.5},
2
A = {2000,3500,5000}in ,
n rib = {7,10,13}.
Similar to case I, for each point in design space II, the EPA is carried out and Eqs. (7.5) and
(7.8) are used to generate the 15 constitutive matrix terms and the 36 mass densities which are then
used to perform the ESA. 51 feed-forward neural networks with the structure of 6 15 10 1 are
trained using the MATLAB NN Toolbox function trainlm. There are 729 ( 36 ) sets of data that could
be used for training, but it was found that at some design points the differences between the natural
frequencies by the EPA and the ESA become too large. Therefore a screening process was
introduced, in which any point where the maximum relative difference between the first 10 natural
frequencies by the EPA and the ESA surpasses 20% will be discarded. 28 points were removed
through the process, therefore 701 sets of data were used for training.
Generating the 729 sets of pre-training data used about 152 hours of CPU time on the Crunch
(SGI Origin 2000 with eight R10000 processors) of the College of Engineering, Virginia Tech, and
training the neural networks spent about 2 hours on a PII/350 PC. A set of results are given in Figs.
7.22 to 7.24 where 25 points were randomly chosen within the design space box. The plan forms of
the new designs are shown in Fig. 7.22, where dashed lines indicate the spar or rib positions. The
first 10 natural frequencies by the EPA and the ESA are compared in Fig. 7.23 and their relative
differences (based on the EPA results) are shown in Fig. 7.24. It can be seen that the relative
difference is within -5%~15%.
Fig. 7.25 shows 16 new designs through a randomly chosen path inside the design space box
which is defined as
CHAPTER 7
127
v 1 = , v 2 = , v 3 = , 4 = A, 5 = n spar , 6 = n rib
n
a j = s j , n j = r j /(1 r j ).
v j = v 0j (1 a j ) + v1j a j , j = 1, ,6
(7.12)
where r j ( j = 1, ,6) are randomly determined values between 0 and 1, and see Eq. (7.10) for the
definition of other symbols. Results of natural frequencies of the first 6 modes for wing structures
defined by points along a path with n 1 =0.2243, n 2 =0.8591, n 3 =0.2064, n 4 =3.0700, n 5 =2.2196
and n 6 =0.9440 are shown in Fig. 7.26, where it can be seen that results by the EPA and the ESA
agree with each other quite well.
Now some static results. For an arbitrary new design whose plan form is shown in Fig. 7.27, a
down-ward (- z direction) point force of 1 lb is applied at the mid-point of the wing tip. Fig. 7.28
shows displacement components along the leading-edge by the EPA and the ESA, compared FEA
using MSC/NASTRAN. Figs. 7.29 and 7.30 show the Von Mises stress distributions at the wing
root and the central spar respectively together with the FEA results. Comparison of the natural
frequencies of this wing as given by the EPA, the ESA and the FEA is shown in Table 7.1. It can
be seen that the EPA and the ESA results are close, and they all agree quite well with the FEA
results.
CHAPTER 7
Fig. 7.22 25 randomly chosen wing plan forms in design space III
128
CHAPTER 7
129
7000
Frequency by ESA
6000
5000
4000
3000
2000
1000
1000
2000
3000
4000
5000
6000
7000
Frequency by EPA
Fig. 7.23 Comparison of the first 10 frequencies by EPA and ESA
CHAPTER 7
130
0.2
Relative error
0.15
0.1
0.05
-0.05
1000
2000
3000
4000
5000
6000
Frequency by EPA
Fig. 7.24 The relative errors in Fig. 7.23
7000
CHAPTER 7
Fig. 7.25 16 wing plan forms systematically varying through design space III
131
CHAPTER 7
132
5000
by EPA
1st Bending mode by ESA
1st Torsion mode by ESA
2nd Bending mode by ESA
In plane mode by ESA
2nd Torsion mode by ESA
3rd Bending mode by ESA
+
x
4000
3000
x
2000
+
x
+
x
+
x
+
x
+
1000
0.25
x
+
x
+
x
+
+
x
+
x
+
x
0.5
+
x
+
x
+
x
+
x
0.75
+
x
CHAPTER 7
Fig. 7.27 An arbitrarily chosen wing plan form in design space III
133
CHAPTER 7
134
5.0E-05
0.0E+00
-5.0E-05
-1.0E-04
-1.5E-04
10u (EPA)
10v (EPA)
w (EPA)
10u (ESA)
10v (ESA)
w (ESA)
10u (FEM)
10v (FEM)
w (FEM)
-2.0E-04
-2.5E-04
-3.0E-04
-3.5E-04
10
20
30
40
50
60
70
80
CHAPTER 7
135
3.5
2.5
1.5
Upper skin (EPA)
Lower skin (EPA)
Upper skin (ESA)
Lower skin (ESA)
Upper skin (FEM)
Lower skin (FEM)
0.5
0.25
0.5
0.75
CHAPTER 7
136
3.5
2.5
Upper skin (EPA)
Lower skin (EPA)
Upper skin (ESA)
Lower skin (ESA)
Upper skin (FEM)
Lower skin (FEM)
1.5
0.25
0.5
0.75
CHAPTER 7
137
v 1 = t 0 t , v 2 = a rt , v 3 = h1 , 4 = h2 , 5 = n spar , 6 = n rib
n
a j = s j , n j = r j /(1 r j ).
v j = v 0j (1 a j ) + v1j a j , j = 1, ,6
(7.13)
where r j ( j = 1, ,6) are randomly determined values between 0 and 1, and see Eq. (7.10) for the
definition of other symbols. Results of natural frequencies of the first 6 modes for wing structures
defined by points along the path with n 1 =0.0031, n 2 =0.9999, n 3 =0.2089, n 4 =64.7024, n 5 =0.9067
CHAPTER 7
138
and n 6 =0.5325 are shown in Fig. 7.35, where it can be seen that results by the EPA and the ESA
agree with each other quite well.
For an arbitrary new design whose plan form is shown in Fig. 7.36, a down-ward (- z direction)
point force of 1 lb is applied at the mid-point of the wing tip. Fig. 7.37 shows displacement
components along the leading-edge by the EPA and the ESA, compared with FEA using
MSC/NASTRAN. Figs. 7.38 and 7.39 show the Von Mises stress distributions at the wing root and
the central spar respectively together with the FEA results. Comparison of the natural frequencies
of this wing as given by the EPA, the ESA and the FEA is shown in Table 7.2. Again, it can be
seen that the EPA and the ESA results are close, and they all agree quite well with the FEA results.
It is noted that a coarser design space III does not worsen the accuracy of the ESA.
CHAPTER 7
139
CHAPTER 7
140
2000
1750
Frequency by ESA
1500
1250
1000
750
500
250
0
500
1000
1500
2000
Frequency by EPA
Fig. 7.32 Comparison of the first 10 frequencies by EPA and ESA
CHAPTER 7
141
0.15
Relative error
0.1
0.05
-0.05
500
1000
1500
Frequency by EPA
Fig. 7.33 The relative errors in Fig. 7.32
2000
CHAPTER 7
142
CHAPTER 7
143
1500
1400
1300
by EPA
1st Bending mode by ESA
2nd Bending mode by ESA
In plane mode by ESA
1st Torsion mode by ESA
3rd Bending mode by ESA
2nd Torsion mode by ESA
+
x
1200
1100
1000
900
800
700
600
500
400
300 x
200 +
x
+
100
0
0.25
0.5
0.75
CHAPTER 7
144
Skin thickness
at root (x10)
Rib cap height (x10)
Skin thickness
at tip (x10)
CHAPTER 7
145
0.0E+00
-1.0E-04
-2.0E-04
-3.0E-04
10u (EPA)
10v (EPA)
w (EPA)
10u (ESA)
10v (ESA)
w (ESA)
10u (FEM)
10v (FEM)
w (FEM)
-4.0E-04
-5.0E-04
-6.0E-04
-7.0E-04
50
100
150
200
CHAPTER 7
146
1
0.9
0.8
0.7
0.6
0.5
0.4
Upper skin (EPA)
Lower skin (EPA)
Upper skin (ESA)
Lower skin (ESA)
Upper skin (FEM)
Lower skin (FEM)
0.3
0.2
0.1
0
0.25
0.5
0.75
CHAPTER 7
147
1.5
0.5
0.25
0.5
0.75
CHAPTER 7
148
7.3 Conclusion
CPU time savings using the ESA are obvious. For instance, when 6 terms of the Legendre
polynomials ( K = 6 ) are used, about 85% less CPU time is spent in evaluating the stiffness and
mass matrices compared with the EPA, where matrix evaluating takes about 68% of the total CPU
time when solving the free vibration problem. When K = 8 , about 83% less CPU time is spent in
evaluating the matrices compared with the EPA, where matrix evaluating takes about 65% of the
total CPU time. Generally speaking, the results given by the ESA in design space II and III are as
good as those in design space I although the number of variables increases from 4 to 6.
Table 7.1
Mode No.
Mode Shape
EPA
ESA
FEM
1
1 bending
279.3
274.5
279.9
st
Table 7.2
Mode No.
Mode Shape
EPA
ESA
FEM
4
In plane
1447.4
1440.3
1454.4
5
2 torsion
1945.5
1936.3
1830.8
nd
1
1 bending
71.9
70.8
66.0
st
2
2 bending
982.8
984.1
965.6
nd
2
2 bending
233.9
239.4
222.6
nd
3
In plane
358.1
358.4
377.0
4
1 torsion
452.2
469.4
413.1
st
5
3 bending
479.9
504.8
468.0
rd
Chapter 8
Conclusions and Future Work
8.1 Conclusions of the Present Work
(1)
An efficient method capable of static and vibration analyses of the built-up wing structures,
Equivalent Plate-model Analysis (EPA) method, has been developed and comparisons for a
series of examples with commercial FEA calculations have shown the accuracy of the
method for design purposes. On the assumption that the wing structure behaves like a plate
whose deformation can be modeled by the FSDT of Reissner-Mindlin, the Rayleigh-Ritz
method is applied to solve the plate problem, with the Legendre polynomials being used as
the trial functions. The stiffness matrix and mass matrix are determined by applying the
Lagranges equations, and can be calculated numerically by using the Gaussian integration
quadrature. Then static analysis can be readily performed and the natural frequencies and
the mode shapes of the wing can be obtained by solving an eigenvalue problem.
Comparison of results by the present method with those by the commercial finite
element analysis code MSC/NASTRAN for a series of 5 vibration problems, 4 static
loading problems, and 1 stress distribution problem showed an overall good agreement
between the two approaches with different methodologies. Mode shapes and natural
frequencies for cases from a thick wing-shaped plate, the same plate with a camber, a solid
wing, to built-up wing structures composed of skins, spars and ribs, have all shown that the
149
CHAPTER 8
150
present method has a fairly good correlation to the FEA, although results for simpler cases
seem to be more accurate. It is also shown that static displacements and stress variation
trends of wing structure can be predicted by the present method quite accurately.
The EPA is formulated mostly in matrix form and calculation can be readily carried out
in the MATLAB environment. It is suitable to be used for the early stages of wing design.
Due to the efficiency of the method, it can also be used as a means to analyze the shape
sensitivity of wing structures.
(2)
(3)
The Equivalent Plate Analysis (EPA) of built-up wing structures is coupled, in an indirect
way, with Neural Networks (NN) to make an even more efficient method, the Equivalent
Skin Analysis (ESA). In the EPA, major part of computational effort is spent on calculating
contributions to the stiffness and mass matrices from each spar and rib. This can be avoided
by replacing the wing inner-structure with an "equivalent" material that combines to the
CHAPTER 8
151
skin and whose properties are simulated by neural networks. The constitutive matrix, which
relates the stress vector to the strain vector, and the density of the equivalent material are
obtained by enforcing mass and stiffness matrix equities with regard to the EPA in a leastsquare sense. Neural networks for the material properties are trained in terms of the design
variables of the wing structure. Examples show that ESA takes off more than 80% of the
CPU time that is spent in the EPA on computing the total stiffness and mass matrices, and
still fairly good results can be obtained. Therefore, the ESA is very promising to be used at
the early stages of wing structure design.
(2)
To extend the present series of efficient methods to deal with steady-state and transient
response problems. Since the stiffness and mass matrices have been given, solution of these
kinds of problems without dissipation should be straightforward. For problems where
structural dissipation is to be considered, work should be done to set up the dissipation
matrix.
CHAPTER 8
(3)
152
To extend the present series of efficient methods to include geometrical nonlinearity, and
material nonlinearity (plasticity, strain hardening etc.), for problems of large deformation
and in extreme material conditions.
(4)
To extend the present series of efficient methods to deal with all spectrum of problems that
are structure-concerned and need to be addressed in wing structure design, such as
aeroelasticity (divergence and flutter speeds), global buckling, and composites/structures
with imperfections or damages.
References
1. Gregory Roth, "The Value of Early Analysis", Computer-Aided Engineering, Oct. 1999, pp. 3638
2. Housner, J., Draft input to White Paper on Next Generation Revolutionary Analysis and Design
Environment (NEXTGRADE) for Aerospace Systems, Version 3, Feb. 27, 1996
3. Jean Thilmany, "CAD Meets CAE", Mechanical Engineering, Oct. 1999, pp. 66-69
4. Haykin, S., Neural Networks: A Comprehensive Foundation, Macmillan College Publishing
Company, 1994
5. Abdalla, K. M. and Stavroulakis, G. E., "A backpropagation Neural Network Model for SemiRigid Steel Connections", Microcomputers in Civil Engineering, Vol. 10, 1995, pp77-87
6. Vanluchene, R. D. and Sun, R., "Neural Network in Structural Engineering", Microcomputers in
Civil Engineering, Vol. 5, 1990, pp. 207-215
7. Gunaratnam, D. J., and Gero, J. S., "Effect of Representation on the Performance of Neural
Networks in Structural Engineering Applications", Microcomputers in Civil Engineering, Vol. 9,
1994, pp. 97-108
8. Kapania, R. K. and Liu, Y., Applications of Artificial Neural Networks in Structural
Engineering with Emphasis on Continuum Models, Technical Report, Virginia Polytechnic
Institute and State University, Blacksburg, VA, June 1998.
9. Liu, Y., Kapania, R. K. and VanLandingham, H., "Simulating and Synthesizing Substructures
Using Neural Network and Genetic Algorithms", Modeling and Simulation Based Engineering, ed.
S. N. Atluri & P. E. O'Donoghue, Tech Science Press, 1998, Vol. I, pp. 576-581
153
REFERENCES
154
10. Noor, A. K., Anderson, M. S., and Greene, W. H., Continuum Models for Beam- and Platelike
Lattice Structures, AIAA Journal, Vol. 16, No. 12, 1978, pp. 1219-1228.
11. Nayfeh, A. M., and Hefzy, M. S., Continuum Modeling of the Mechanical and Thermal
Behavior of Discrete Large Structures, AIAA Journal, Vol. 19, No. 6, 1981, pp. 766-773.
12. Sun, C. T., Kim, B. J., and Bogdanoff, J. L., On the Derivation of Equivalent Simple Models
for Beam- and Plate-Like structures in Dynamic Analysis, AIAA Paper 81-0624, 1981, pp. 523532.
13. Noor, A. K., Continuum Modeling for Repetitive Lattice Structures, Applied Mechanics
Review, Vol. 41, No. 7, 1988, pp. 285-296.
14. Lee, U, Dynamic Continuum Modeling of Beamlike Space Structures Using Finite-Element
Matrices , AIAA Journal, Vol. 28, No. 4, 1990, pp. 725-731.
15. Lee, U, Dynamic Continuum Plate Representations of Large Thin Lattice Structures, AIAA
Journal, Vol. 31, No. 9, 1993, pp. 1734-1736.
16. Lee, U, Spectral Element Approach for the Homogeneous Continuum Representations of a
Periodic Lattice Structure, International Journal of Space Structures, Vol. 12, No. 1, 1997, pp.1-8.
17. Kapania, R. K. and Castel, F., A Simple Element for Aeroelastic Analysis of Undamaged and
Damaged Wings, AIAA Journal, Vol. 28, No. 2, 1990, pp. 329-337.
18. Song, O. and Librescu, L., Free Vibration and Aeroelastic Divergence of Aircraft Wings
Modeled as Composite Thin-Walled Beams, AIAA Paper 91-1187, 1991.
19. Lee, U, Equivalent Dynamic Beam-Rod Models of Aircraft Wing Structure, Aeronautical
Journal, Dec. 1995, pp. 450-457.
20. Lovejoy, A. E. and Kapania, R. K., Natural Frequencies and Atlas of Mode Shapes for
Generally-Laminated, Thick, Skew, Trapezoidal Plates, CCMS(Center for Composite Materials
and Structures)-94-09, Virginia Polytechnic Institute and State University, Blacksburg, VA, Aug.
1994.
REFERENCES
155
21. Lovejoy, A. E., Natural Frequencies and Atlas of Mode Shapes for Generally-Laminated,
Thick, Skew, Trapezoidal Plates, MS Thesis, Department of Aerospace and Ocean Engineering,
Virginia Polytechnic Institute and State University, Blacksburg, VA, Aug. 1994.
22. Dawe, D. J., Buckling and Vibration of Plate Structures Including Shear Deformation and
Related Effects, Aspects of the Analysis of Plate Structures, Ed. by Dawe, D. J., Horsington, R.
W., Kamtekar, A. G. and Little, G. H., New York, Oxford University Press, 1985, pp. 75-99.
23. Reddy, J. N. and Miravete, A., Practical Analysis of Composite Laminates, CRC Press, Inc.,
Boca Raton, 1995, pp. 52-62
24. Reissner, E., The Effect of Transverse Shear Deformation on the Bending of Elastic Plates,
Journal of Applied Mechanics, Vol. 12, 1945, pp. A-69~77.
25. Mindlin, R. D., Influence of Rotatory Inertia and Shear on Flexural Motions of Isotropic,
Elastic Plates, Journal of Applied Mechanics, Vol. 18, 1951, pp. 31-38.
26. Kapania, R. K. and Raciti, S., Recent Advances in Analysis of Laminated Beams and Plates,
Part I: Shear Effects and Buckling, AIAA Journal, Vol. 27, No. 7, 1989, pp. 923-934.
27. Knight, N. F. and Qi, Y., On a Consistent First-Order Shear-Deformation Theory for
Laminated Plates, Composites Part B: Engineering, 28 B (1997), pp. 397-405.
28. Giles, G. L., Equivalent Plate Analysis of Aircraft Wing Box Structures with General
Planform Geometry, Journal of Aircraft, Vol. 23, No. 11, 1986, pp. 859-864.
29. Giles, G. L., Equivalent Plate Modeling for Conceptual Design of Aircraft Wing Structures,
AIAA Paper 95-3945, September 1995.
30. Tizzi, S., Numerical Procedure for the Dynamic Analysis of Three-Dimensional Aeronautical
Structures, Journal of Aircraft, Vol. 34, No. 1, 1997, pp. 120-130.
31. Livne, E., Equivalent Plate Structural Modeling for Wing Shape Optimization Including
Transverse Shear, AIAA Journal, Vol. 32, No. 6, 1994, pp. 1278-1288.
32. Livne, E. and Navarro, I., Nonlinear Equivalent Plate Modeling of Wing Box Structures, 40th
AIAA/ASME/ASCE/AHS/ASC SDM Conference, April 12-15, 1999, St. Louis, MO, pp. 346-372.
Also Journal of Aircraft, Vol. 36, No. 5, 1999, pp. 851-865.
REFERENCES
156
33. Haug, E. J., Choi, K. K. and Komkov, V., Design Sensitivity Analysis of Structural Systems,
Academic Press, Inc., Orlando, Florida, 1986.
34. Adelman, H. M. and Haftka, R. T., "Sensitivity Analysis of Discrete Structural Systems", AIAA
Journal , Vol. 24, No. 5, 1986, pp. 823-832.
35. Barthelemy, J.-F. M. and Haftka, R. T., "Approximation Concepts for Optimum Structural
Design FD5HYLHZStructural Optimization, Vol. 5, 1993, pp. 129-144.
36. Kapania, R. K., Bergen, F. D. and Barthelemy, J.-F. M., "Shape Sensitivities Analysis of Flutter
Response of a Laminated Wing", AIAA Journal , Vol. 29, No. 4, 1991, pp. 611-612.
37. S. Singhvi, and R. K. Kapania, "Analytical Shape Sensitivities and Approximations of Modal
Response of Generally Laminated Tapered Skew Plates", Journal of Aircraft, Vol. 30, No. 3, 1993,
pp. 423-426.
38. Eldred, L. B., "Sensitivity Analysis of the Static Aeroelastic Response of a Wing", Ph. D.
Dissertation, Virginia Polytechnic Institute and State University, Blacksburg, VA, Feb. 1993
39. Kapania, R. K. and Issac, J. C., "Sensitivity Analysis of Aeroelastic Response of a Wing in
Transonic Flow", AIAA Journal , Vol. 32, No. 2, 1994, pp. 350-356.
40. Bhardwaj, M. and Kapania, R. K., "Shape Sensitivity Analysis of Divergence Dynamic
Pressures", Journal of Aircraft, Vol. 32, No. 4, 1995, pp. 898-901.
41. Issac, J. C. and Kapania, R. K., "Sensitivity of Flutter Response of a Wing to Shape and Modal
Parameters", AIAA Journal , Vol. 33, No. 10, 1995, pp. 1983-1986.
42. Issac, J. C. and Kapania, R. K., "Aeroelastic Analysis of Wings Using Automatic
Differentiation", AIAA Journal , Vol. 35, No. 3, 1997, pp. 519-525.
43. Kapania, R. K. and Singhvi, S., Free Vibration Analyses of Generally Laminated Tapered
Skew Plates, Composites Engineering, Vol. 2, No. 3, 1992, pp. 197-212.
44. Singhvi, S., Analysis, Shape Sensitivities and Approximations of Modal Response of
Generally Laminated Tapered Skew Plates, Master Thesis, Virginia Polytechnic Institute and
State University, Blacksburg, VA, Sept. 1991.
REFERENCES
157
45. Kapania, R. K. and Lovejoy, A. E., Free Vibration of Thick Generally Laminated Cantilever
Quadrilateral Plates, AIAA Paper 95-1350, 1995; AIAA Journal, Vol. 34, No. 7, 1996, 1474-1480.
46. Kapania, R. K. and Lovejoy, A. E., Free Vibration of Thick Generally Laminated
Quadrilateral Plates with Point Supports, AIAA Paper 96-1346, 1996; Journal of Aircraft, Vol.
35, No. 6, 1998, pp. 958-965.
47. Cortial, F., Sensitivity of Aeroelastic Response of Wings Using Equivalent Plate Models,
Technical Report, Department of Aerospace and Ocean Engineering, Virginia Tech, 1996.
48. R. K. Kapania, Y. Liu, "Wing Vibrations Using FEM & Equivalent Plate Models", 1999 ASME
Mechanics and Materials Conference, June 27-30, 1999, Virginia Tech, Blacksburg, VA 24061
49. Kapania, R. K. and Liu, Y., Static and Vibration Analyses of General Wing Structures Using
Equivalent Plate Models, to appear in AIAA Journal. Also, as Paper 2000-1434 at 41st
AIAA/ASME/ASCE/AHS/ASC Structures, Structural Dynamics, and Materials Conference &
Exhibit, 3-6 April 2000, Atlanta, Georgia.
50. Liu, Y. and Kapania, R. K., Modal Response of Trapezoidal Wing Structures Using Second
Order Shape Sensitivities, AIAA Journal, Vol. 38, No. 4, April 2000, pp. 732-735.
51. Kapania, R. K. and Liu, Y., Efficient Simulation of Wing Modal Response: Application of 2nd
Order Shape Sensitivities and Neural Networks, submitted to 8th AIAA/NASA/USAF/ISSMO
Symposium on Multidiscipline Analysis and Optimization, 6-8 September 2000, Long Beach, CA.
52. Liu, Y. and Kapania, R. K., Equivalent Skin Analysis of Wing Structures Using Neural
Networks", submitted to AIAA Journal and also to be presented at 8th AIAA/NASA/USAF/ISSMO
Symposium on Multidiscipline Analysis and Optimization, 6-8 September 2000, Long Beach, CA.
53. Hornik, K., Stinchcombe, M. and White, H., "Multilayer Feedforward Networks are Universal
Approximators", Neural Networks, Vol. 2, 1989,pp. 359-366.
54. Hartman, E. J., Keeler, J. D., and Kowaski, J. M., "Layered Neural Networks with Gaussian
Hidden Units as Universal Approximators", Neural Computation, Vol. 2, 1990,pp 210-215.
REFERENCES
158
55. Rabelo, L. C., Shastri, V., Onyejekwe, E. and Vila, J., "Using Neural Networks and Abductive
Modeling for Device-Independent Color Correction", Scientific Computing and Automation, May
1996
56. Ross, C. T. F., Finite Element Methods in Structural Mechanics, Ellis Horwood Ltd., 1985
57. Jang, J.-S. R., Sun, C.-T. and Mizutani, E., Neuro-Fuzzy and Soft Computing: A Computational
Approach to Learning and Machine Intelligence, Prentice Hall, Inc., 1997.
58. Hurty, W. C. and Rubinstein, M. F., Dynamics of Structures, Prentice-Hall, Inc., New Jersey,
1964.
59. Karamcheti, K., Principles of Ideal-fluid Aerodynamics, R. E. Krieger Pub. Co., New York,
1980, pp. 488-490
60. Issac, J. C., Sensitivity Analysis of Wing Aeroelastic Responses, Ph. D. Dissertation,
Department of Aerospace and Ocean Engineering, Virginia Polytechnic Institute and State
University, Blacksburg, VA, Nov. 1995.
61. Karpel, M., Moulin, B. and Love, M. H., "Modal-Based Structural Optimization with Static
Aeroelastic and Stress Constraints", Journal of Aircraft, Vol. 34, No. 3, 1997, pp. 433-440.
62. Eldred, M. S., Venkayya, V. B. and Anderson, W. J., "Mode Tracking Issues in Structural
Optimization", AIAA Journal, Vol. 33, No. 10, 1995, pp. 1926-1933.
63. Eldred, M. S., Venkayya, V. B. and Anderson, W. J., "New Mode Tracking Methods in
Aeroelastic Analysis", AIAA Journal, Vol. 33, No. 7, 1995, pp. 1292-1299.
64. Rall, L. B., "Automatic Differentiation: Techniques and Applications", Lecture Notes in
Computer Science, Vol. 120, Springer-Verlag, Berlin, 1981
65. Automatic Differentiation of Algorithms: Theory, Implementation, and Applications, Ed.
Griewank, A. and Corliss, G. F., SIAM, Philadelphia, 1992
66. Wittrick, W. H., "Rates of Change of Eigenvalues, with Reference to Buckling and Vibration
Problems", Journal of the Royal Aeronautical Society, Vol. 66, 1962, pp. 590-591.
67. Fox, R. L. and Kapoor, M. P., "Rates of Change of Eigenvalues and Eigenvectors", AIAA
Journal, Vol. 6, No. 12, 1968, pp. 2426-2429.
REFERENCES
159
68. Friswell, M. I. and Mottershead J. E., Finite Element Model Updating in Structural Dynamics,
Kluwer Academic Publishers, the Netherlands, 1995.
69. Haftka, R. T. and Gurdal, Z., Elements of Structural Optimization, Kluwer Academic
Publishers, the Netherlands, 1992
70. Balabanov, V. O., Giunta, A. A., Golovidov O., Grossman, B., Mason, W. H., Watson, L. T.
and Haftka, R. T., Reasonable Design Space Approach to Response Surface Approximation,
Journal of Aircraft, Vol. 36, No.1, January-February 1999
71. Box, M. J. and Draper, N. R., Factorial Designs, the |X'X| Criterion and Some Related
Matters, Technometrics, Vol.13, No. 4, 1971, pp. 731-742.
72. Mitchell, T. J., An Algorithm for the Construction of D-Optimal Experimental Designs,
Technometrics, Vol.16, No. 2, 1974, pp. 203-210.
73. Nikolaidis, E., Long, L. and Ling, Q., Neural Networks and Response Surface Polynomials for
Design of Vehicle Joints, Computers and Structures, 75 (2000), pp. 593-607.
Appendix A
The Constitutive Matrix for Various Cases
A.1 Rotation along z -axis
Assume that the coordinate system ( x, y , z ) is rotated along the z -axis an angle and becomes
( x ' , y ' , z ' ) . If we write
{ } = { x , y , xy , yz , zx }T
{ } = { x , y ,2 xy ,2 yz ,2 zx }T
{ ' } = [ D' ]{ ' } , { } = [ D ]{ }
it can be derived that
[ D' ] = [T ][ D ][Te ] 1
(A.1)
where
[Te ]1 = [Te ( )] 1 = [Te ( )]
cos2
2
sin
[T ] = 12 sin 2
sin 2
cos2
1
2 sin 2
sin 2
sin 2
cos 2
0
0
0
0
0
0
0
cos
sin
0
0
sin
cos
160
(A.2)
APPENDIX A
CONSTITUTIVE MATRIX
cos 2
2
sin
[Te ] = sin 2
0
0
sin 2
cos 2
sin 2
sin 2
12 sin 2
cos 2
0
0
0
0
0
0
0
cos
sin
1
2
161
sin
cos
0
0
0
(A.3)
0
[T ] = 0
0
12 sin 2
0
0
1
0
0 cos
0 sin
0
0
0
0
sin
cos
cos 2
0
[Te ] = 0
0
sin 2
0
0
1
0
0 cos
0 sin
0
0
sin
cos
A.3 Skin
The local constitutive matrix is
sin 2
sin 2
0
0
cos 2
(A.4)
12 sin 2
1
2 sin 2
cos 2
(A.5)
APPENDIX A
CONSTITUTIVE MATRIX
E
1 2
E
2
[ D ] = 1
0
0
E
1 2
E
1 2
0
0
0
where D33 = G , G =
D33
0
0
0
D 44
0
0
0
D55
162
(A.6)
E
, E is the Young's modulus, is the Poisson's ratio, and
2(1 + )
D44 = kG , D55 = kG , k =
2
5
or
for thick plates and D44 = 0 , D55 = 0 for thin skins.
6
12
If the wing has a high aspect ratio and a small swept angle, then we can see ( x0 , y 0 , z0 ) (the
local coordinates) to rotate an angle = (see Fig. 2 for the definition of )along the y -axis
to coincide with ( x, y, z ) (the global coordinates), therefore the global constitutive matrix is
[ D ] = [T ( )][ D ][Te ( )]1
where [T ] and [Te ] are found in Eqs. (A.4) and (A.5). If the skin is composed of laminated
layers, for the i -th layer the global constitutive matrix is
[ Di ] = [T ( i )][ D ][Te ( i )] 1
where [T ] and [Te ] are found in Eqs. (A.2) and (A.3), and i is the orientation angle of the i th layer.
(A.7)
After a rotation of an angle along the z -axis, the global constitutive matrix becomes
[ D ] = [T ( )][ D ][Te ( )] 1
APPENDIX A
CONSTITUTIVE MATRIX
163
where [T ] and [Te ] are found in Eqs. (A.2) and (A.3). The angle is the angle between the local
orientation of the spar or rib and the x -axis.
(A.8)
Appendix B
Formulation for Multi-Plane Problem Using
EPA
As shown in Fig. B.1, a wing is composed of a main-body and a wing-let. What is given in this
appendix can be used as a basis to solve wing structures with more than one wing-let or composed
of more than two planes.
( 2 ,2 )
( 1 , 1 )
z
y,
x,
164
APPENDIX B
165
(B.1)
where [ K 1 ] and [ K 2 ] are the stiffness matrices for the main-body and the wing-let respectively,
whose formulation can be found in Ref. 1, [ K BC ] is the stiffness matrix for the large springs
simulating the boundary conditions at the root 2 , and U JT is the strain energy relating to the joint
between the main-body and the wing-let, which will be treated as follows.
In the ( 1 ,1 ) and ( 2 , 2 ) planes, the relationships between the displacement vectors and the
general displacement vectors 1 can be written as
{u1 } = [ H 1 ]{q1 }
{u 2 } = [ H 2 ]{q 2 }
(B.2)
where [ H 1 ] and [ H 2 ] are functions of ( 1 ,1 ) and ( 2 , 2 ) respectively. For the joint joining the
main-body and the wing-let, we can have
1 1 1, 1 = 1; 1 2 1, 2 = 1
and by expressing the displacement vector in plane ( 2 , 2 ) in terms of plane ( 1 ,1 ) , we get
{u 2 ' } = [ R ]{u 2 }
(B.3)
(B.4)
where [ K JT ] is the stiffness matrix for the joint, and springs with very large magnitude can be used
if the joint is rigid. Using Eqs. (B.2) and (B.3), Eq. (B.4) can be extended as
U JT =
1
2
(B.5)
APPENDIX B
166
{q 2 }
1
2
1
2
[ K 11 ] [ K 12 ] {q1 }
[ K ] [ K ] {q }
22
2
21
(B.6)
Comparing Eq. (B.6) with Eqs. (B.1) and (B.5), we can obtain the stiffness matrix of the whole
[ K 11 ] [ K 12 ]
structure [ K ] =
in terms of {q} :
[ K 21 ] [ K 22 ]
[ K 22 ] = [ K 2 ] + [ H 2 ]T [ R ]T [ K JT ][ R ][ H 2 ]
T
T
T
T
1
[ K 12 ] = [ K 21 ] = 2 [ H 1 ] ([ K JT ][ R ] + [ K JT ] [ R ])[ H 2 ]
[ K 11 ] = [ K 1 ] + [ H 1 ]T [ K JT ][ H 1 ] + [ K BC ]
(B.7)
0 {q1 }
{q } [ M ]
T = {q1 } [ M 1 ]{q1 } + {q 2 } [ M 2 ]{q 2 } = 1 1
[ M 2 ] {q 2 }
{q 2 } 0
1
2
1
2
1
2
(B.8)
where [ M 1 ] and [ M 2 ] are the mass matrices for the main-body and the wing-let respectively 1 .
Therefore the mass matrix of the whole structure is
0
[ M 1 ]
[M ] =
[ M 2 ]
0
(B.9)
Appendix C
Airfoil Sections Generated with the KarmanTrefftz Transformation
The Karman-Trefftz transformation 59 is defined as
z nc ( c ) n
=
z + nc ( + c ) n
(C.1)
z
=
or
=
+
z + 2c ( + c ) 2
(C.2)
As shown in Fig. C.1, the Karman-Trefftz transformation in Eq. (C.1) maps a circle in the plane:
= ae i +
(C.3)
where a = m
sin
sin
=c
is the radius,
sin
sin( + )
167
APPENDIX C
KARMAN-TREFFTZ TRANSFORMATION
168
y
(, ) >(x, y)
C
T
O
C'
T'
(nc, 0)
(-nc, 0)
-plane
z-plane
= OTC , = COT , = OC , m = OC , c = OT , a = CT
= tan 1
Re( )
= arg( ) ,
Im( )
APPENDIX C
KARMAN-TREFFTZ TRANSFORMATION
1 r 2n
2n
n
1 2 r cos n + r
2r n cos n
y = y ( , ) = nc
1 2r n cos n + r 2 n
1
r = r1 / r2 , r1, 2 = {( # c ) 2 + 2 } 2
= 1 2 , 1, 2 = TAN 1
#c
169
x = x ( , ) = nc
(C.4)
in which, a general arc tangent function TAN 1 is defined according to the common arc tangent
function tan 1 [
, ]:
2 2
1 y
tan x , if y 0 and x > 0;
y
1 y
TAN = + tan 1 , if ( y 0 and x 0) or ( y < 0 and x < 0);
x
x
2 + tan 1 y , if y < 0 and x 0.
(C.5)
If the common arc tangent function tan 1 instead of TAN 1 is used in Eq. (C.4), there would be
some abnormal kinks in the curve in the z plane transformed from a circle in the plane.
It can be proved that the trailing edge angle of the airfoil is = ( 2 n) . When n = 2, = 0, so
the Joukowski airfoils have cusped trailing edges.
The camber of the airfoil is determined by Im( ) = m sin = a sin , and the thickness by
Re( ) = m cos = c a cos . The chord-length is
(2 + ) n + n
l = nc 1 +
n
n
(2 + )
(C.6)
where = 2 cos 1 is a quantity of small value having a close relationship with the
c
APPENDIX C
KARMAN-TREFFTZ TRANSFORMATION
170
II
III
When there is an incoming flow, it is ready to obtain the velocities and pressures at any point,
streamline patterns, and the lift coefficient etc., by using Eq. (C.1) or (C.4) and the velocity
potential of the flow past a circle with circulation.
Vita
Youhua Liu was born on August 22, 1963, in Jing County, Hunan Province, China. He attended
the No. 1 Middle School of Jing County for high school education, from September, 1978 to July,
1980, among the first group of students enrolled county-wide since the Cultural Revolution. In
September, 1980, he enrolled at Huazhong (Central China) University of Science and Technology
(HUST), in Wuhan, capital of Hubei Province. Majoring in Naval Architecture and Ocean
Engineering, he earned a Bachelor's degree in engineering in July, 1984. He continued to study and
got a Master's degree in engineering in July, 1987. From 1991 to 1996 he worked at HUST as a
faculty member and research engineer. In 1994, he worked for two months as a visiting scientist at
Yokohama National University, Japan.
In August 1996, Youhua Liu began his study toward a PhD with the Department of Aerospace
and Ocean Engineering, Virginia Tech. Blacksburg is a wonderful place to live in and being a
Hokie is a fantastic experience to him. He is expected to obtain his PhD in aerospace structures in
early 2000.
171