Professional Documents
Culture Documents
Doyle-Wave Propagation in Structures PDF
Doyle-Wave Propagation in Structures PDF
Doyle-Wave Propagation in Structures PDF
Frederick F. Ling
Series Editor
Wave Propagation in
Structures
Spectral Analysis Using Fast
Discrete Fourier Transforms
Second Edition
i Springer
James F. Doyle
Department of Aeronautics and Astronautics
Purdue University
West Lafayette, IN 47906
USA
Series Editor
Frederick F. Ling
Emest F. Gloyna Regents Chair in Engineering
Department of MechanicaI Engineering
The University of Texas at Austin
Austin, TX 78712-1063 USA
and
William Howard Hart Professor Emeritus
Department of MechanicaI Engineering,
AeronauticaI Engineering and Mechanics
Rensselaer Polytechnic Institute
Troy, NY 12180-3590 USA
987654321
ISBN 978-1-4612-7304-2
This book is dedicated to my
father and mother,
Patrick & Teresa Doyle,
on my mother's seventieth birthday.
Thanksfor the dreams.
Advisory Board
Biomechanics V.C.Mow
Columbia University
Energetics J. Welty
University of Oregon. Eugene
The study of wave propagation seems very remote to many engineers, even to
those who are involved in structural dynamics. One of the reasons for this is that
the examples usually taught in school are either so simple as to be inapplicable to
real world problems, or so mathematically abstruse as to be intractable. This book
contains an approach, spectral analysis or frequency domain synthesis, that I have
found to be very effective in analyzing waves. What has struck me most about this
approach is how I can use the same analytical framework to predict experimental
results as well as to manipulate the experimental data itself. As an experimentalist,
I had found it very frustrating having my analytical tools incompatible with my ex-
periments. For example, it is experimentally impossible to generate a step-function
wave and yet that is the type of analytical solution often available.
Spectral analysis is very encompassing - it touches on analysis, numerical
methods, and experimental methods. I want this book to do justice to its versatility,
so many subjects are introduced. As a result some areas may seem a little thin but I
do hope, nonetheless, that the bigger picture, the unity, comes across. Furthermore,
spectral analysis is not so much a solution technique as it is a different insight into
the wave mechanics, consequently, in most of the examples an attempt is made to
make the connection between the frequency domain and time domains.
In writing this second edition, I have strived to keep what was good about
the first edition - that combination of experimental and analytical results -
but incorporate more recent developments and extensions. The question not fully
articulated in the first edition is: What should be different about a book on waves in
structures? This is the question that has guided my reorganization of the existing
material as well as the selection of new topics. It is clearer to me now that the
essence of a structure is the coupling of systems and this should be the central
theme of a book on waves in structures. There are two readily recognized forms of
coupling: mechanical coupling "at the ends" such as when two bars are joined at an
angle, and differential coupling as when two bars are connected uniformly along
their lengths by springs. Both couplings are intimately related to each other as seen
from the example of a curved beam: it can be modeled as a collection of small
straight segments connected end to end, or directly in terms of coupled differential
equations. The two approaches, ultimately, give the same results, but, at the same
time, give quite different insights into the system behavior. The former leads to
richer system response functions and its ultimate form is in the spectral element
method. The latter leads to richer differential equations which is manifested in
x Preface
http://aae.www.ecn.purdue.edu/~doyle
Look under the section on Source Code. In a similar vein, I have tried to supple-
ment each chapter with a collection of pertinent problems plus specific references
that can form the basis for further study.
A book like this is impossible to complete without the help of many people,
but it is equally impossible to properly acknowledge all of them individually.
However, I would like to single out Brian Bilodeau, Albert Danial, Sudhir Kamle,
Lance Kannal, Matt Ledington, Mike Martin, Steve Rizzi, Gopal Srinivasan, and
Hong Zhang, for their input and help over the years. Any remaining errors and
inaccuracies are purely my own doing.
a radius
A cross-sectional area
A,B,C,D frequency dependent coefficients
b thickness, depth
B bulk modulus
CK group speed
Co longitudinal wave speed, .; EA/ pA
Cp, CS, CR Primary, Secondary and Rayleigh wave speeds
CE specific heat
D plate stiffness, Eh 3 /12(1 - v 2 )
E,E Young's modulus, viscoelastic modulus
EI beam flexural stiffness
F member axial force
g;(x) element shape functions
G(t), G structural unit response, Frequency response function
h beam or rod height, plate thickness
Hn Hankel function, Hn = in ± i Yn
complex .J=T
1 second moment of area, 1 = bh 3 /12 for rectangle
In modified Bessel functions of the first kind
in Bessel functions of the first kind
k, kl' k2 waveguide spectrum relations
k" ky, k, 2-D wavenumbers
K stiffness, thermal conduction
[ k ], [K ] stiffness matrix
Kn modified Bessel functions of the second kind
L length of element, distance to boundary
M, Mx moment
n frequency counter
N number of terms in transform
pet), p acoustic pressure
pet), P applied force history
q distributed load, heat flux
r radial coordinate
R radius
t time
T time window, temperature
u(t) response; velocity, strain, etc.
xiv Notation
u, v, w displacements
v member shear force
w space transform window
x, y, Z rectilinear coordinates
YII Bessel functions of the second kind
Greek letters:
at coefficient of thermal expansion
[w2pA/EI]1/4, [w2ph/D]I/4
permutation symbol
Kronecker delta, small quantity
determinant
viscosity, damping
plate curvature
angular coordinate
v Poisson's ratio
Shear modulus
Lame constant
mass density
a, E stress, strain
space transform variable
rotation
Helmholtz functions
lateral contraction
angular frequency
cut-off, coincidence, critical frequency
Special Symbols:
nil random noise
.. a2 a2
'il 2 differential operator, 3x"! + ]i";!
v volume .
[ ] square matrix
{ } vector
Subscripts:
a acoustic medium
m space wavenumber counter
n frequency counter
1,2 sensors, modes
P,S,R Primary, Secondary, Rayleigh waves
Superscripts:
complex conjugate
* bar, local coordinates
dot, time derivative
frequency domain (transformed) quantity
wavenumber domain (transformed) quantity
prime, derivative with respect to argument
Introduction
wave solutions for structures more interesting than simple rods and beams can be
pieced together successfully. Moreover, the way to solve problems of structures
with many members and boundaries is then available.
A connection not often investigated by those studying structural dynamics is
the relationship between wave propagation and vibrations. For many engineers,
these are two separate areas with quite distinct methods of analysis. However, an-
other advantage of the spectral approach to dynamics is that the close connection
between waves and vibrations becomes apparent. Even the same language can be
used, terms such as power spectral density, filtering, spectral estimation, convolu-
tion, and sampled waveforms have the same meaning. Consequently, many of the
technologies developed over the last thirty years for vibrations and modal analysis
are directly applicable to the spectral analysis of waves.
An exciting possibility (and one of the motivations for writing this book) is to
facilitate the reverse process, that is, to transport many of the waves ideas into
vibration analysis. This should lead to a richer understanding of such topics as
impulse testing and transient vibrations. A number of examples throughout the
book demonstrate the "evolution" of resonance as multiple reflections are included
in the analysis.
by spectral analysis, the kernel solutions and spectrum relations are obtained.
Having initiated a wave, how it interacts with structural discontinuities is then
investigated. Each chapter includes an example of coupling both at the differential
and mechanical levels. Chapter 4 then addresses the question of the adequacy of
the waveguide models for rods and beams. A foundation in the construction of
higher order waveguides is provided that will readily allow extensions to be made
to other specialized problems.
Chapter 5 introduces the spectral element method as a matrix method approach
to structural dynamics that combines aspects of the finite element method with
the spectral analysis method. In a way, this chapter is the culmination of the con-
nected waveguide approach and clearly makes the study of wave propagation in
complicated structures practical.
The next three chapters extend the spectral analysis to the difficult area associ-
ated with plates and plated structures. The extended nature of the plate waveguide
introduce a number of different challenges: Chapter 6 looks at orthotropic material
properties and wave scattering; Chapter 7 deals with the problem of plate/fluid in-
teraction; Chapter 8 attempts to develop a matrix methodology for plated structures
including cylindrical shells.
A number of threads run through all the chapters. The spectral methodology is an
ideal companion for experimental analysis, and throughout the chapters summaries
of some of its experimental applications are given. The emphasis is on how spectral
analysis can extend the type of information extracted from the experimental data-
for example, how a structural response can be used to infer the force history causing
it. Another thread is the role played by elastic constraints. This is first introduced
in connection with the simple rod but reappears with increasing complexity in the
subsequent chapters. Its ultimate expression is found in the coupled modes of the
curved plate.
Admittedly, a large range of problems have been left out even though most
of them are treatable by the spectral methods. Consequently, an effort is made
to supplement each chapter with a collection of pertinent problems plus specific
references that indicate extensions of the theory and the applications. These can
form the basis for further study.
A Question of Units
The choice of any particular set of units is bound to find disfavor with some read-
ers. Nondimensionalizing all of the plots is an unsatisfactory solution especially
since experimental results are included. Furthermore, working with dimensional
quantities help give a better "feel" for the results. The resolution of this dilemma
adopted here is to do most of the examples using a nominal material with nominal
properties. In the two major systems of units, these are given in the following table:
6 Introduction
Property Imperial SI
Material aluminum aluminum
Young's Modulus, E 10 x 106 psi 70GN/m2
Mass Density, p 0.25 x 1O- 3 Ib·s2/in. 4 2800kglm 3
Poisson's Ratio, v 0.33 0.33
Height, h I in. 25mm
Depth, b I in. 25mm
Length, L 100 in. 2500mm
It has long been known that an arbitrary time signal can be thought of as the super-
position of many sinusoidal components, that is, it has a distribution or spectrum of
components. Working in terms of the spectrum is called spectral analysis. In wave
analysis, the time domain for a motion or response is from minus infinity to plus
infinity. Functions in this domain are represented by a continuous distribution of
components which is known as its continuous Fourier transform (eFT). However,
the numerical evaluation and manipulation of the components requires discretizing
the distribution in some manner - the one chosen here is by way of the discrete
Fourier transform (DFT). This has the significant advantage that it allows the use
of the very efficient fast Fourier transform (FFT) computer algorithm.
The goal of this chapter is to introduce the discrete Fourier transform for the
efficient computation of the spectral content of a signal. Possible sources of errors
in using it on finite samples and ways of reducing their influences will be discussed.
The Fourier transform is then applied to wave analysis. The crucial step is to set up
the connection between the spectral responses at different space locations - we
do this through the governing differential equations. In doing so, certain key ideas
emerge which will recur throughout this book; central among these ideas are that
of a wave mode, its spectrum relation, and its phase and group speeds as depicted
in Figure 1.1.
Time Position
~
Figure 1.1: Response showing the difference between the group and phase speeds. Move-
ment of the envelope corresponds to group behavior while movement of the zero crossings
corresponds to phase behavior.
7
8 Chapter 1. Spectral Analysis of Wave Motion
Definition
The continuous Fourier transform pair of a function F(t), defined on the time
domain from -00 to +00, is given as
(1.1)
where C(w) is the continuous Fourier transform (eFT), w is the angular frequency,
and i is the complex R. The first form is the inverse transform while the second
is the forward transform - this arbitrary convention arises because the signal to be
transformed usually originates in the time domain. The factor of 2;rr is necessary
so that a sequential use of the forward and inverse transforms recovers the original
function. However, it should be pointed out that other forms for this factor can be
found in the literature.
The process of obtaining the Fourier transform of a signal separates the wave-
form into its constituent sinusoids (or spectrum) and thus a plot of C(w) against
frequency represents a diawam displaying the amplitude of each of the constituent
sinusoids. The spectrum C(w) usually has nonzero real and imaginary parts. We
will consistently use the superscript 'hat' to indicate the frequency domain spec-
trum of a function.
By way of a simple example of the application of Fourier transforms, consider
a rectangular pulse where the time function is given by
F(t) = Fo - al2 ~ t ~ al2
and zero otherwise. Substituting into the forward transform and integrating gives
~
C(w) =1+ G
-a12
/
2 .
Foe-1wtdt
wal2
~
= Foa { sin (waI2) } == Co(w)
In this particular case the transform is real only and symmetric about w = 0 as
shown in Figure 1.2. The term inside the braces is called a sine function, and
has the characteristic behavior of starting at unity magnitude and oscillating with
decreasing amplitude as its argument increases. It is noted that the value of the
transform at w = 0 is the area under the time function. This, in fact, is a general
result as seen from
When the pulse is displaced along the time axis such that the function is given
by
F(t) = Fa
which has both real and imaginary parts and is not symmetric with respect to w = O.
On closer inspection, however, we see that the magnitudes of the two transforms
are the same; it is just that the latter is given an extra phase change of amount
w(ta + aI2). Figure 1.2 shows the transform for different amounts of shift. We
therefore associate phase changes with shifts of the signal along the time axis.
- - real
imag
~
o
t =-a/2
o
o=-a/5
t
t =+a/l0
" " .. __ '' 0 ..... " .,.,, __
Frequency [kHz]
I I
pair
F(t) <:} C(w)
will be referred to, where the symbol <:} means "can be transformed into." The
double arrowheads reinforce the idea that the transform can go in either direction
and that its properties are symmetric.
If the functions FA(t) and FB(t) have the transforms CA(w) and CB(w), respec-
tively, then the combined function [FA(t) + FB(t)] has the transform [CA(w) +
CB(W)]. That is,
(1.3)
This is an essential property of the transform and it means that if a signal is
composed of the simple sum of two contributions (sayan incident wave and a
reflected wave) then the transform is also composed of a simple sum of the separate
transforms. This linearity property is at the heart of superposition.
The function F(at) (where a is a nonzero constant) has the transform pair
I A
(1.5)
This property was already seen in the last example where the rectangle was dis-
placed from the origin. Of course, the property refers to any change at any position.
There is a corresponding relation for frequency shifting.
The transform pairs, Equations (1.1), are valid for both F(t) and C(w) being
complex, but the functions of usual interest in wave analysis are when F(t) is real.
To see the effect of this, rewrite Equation (1.1) in terms of its real and imaginary
parts as
From the above it is apparent that when F(t) is real-only, CR is even and C/ is
odd. Mathematically, this is expressed as
CR(-w) = CR(w) , C,(-w)=-C/(w) or C(-w)=C*(w) (1.7)
1.2 Discrete Fourier Transform 11
which says that the functions are symmetrical and antisymmetrical, respectively,
about the zero frequency point. This can also be expressed as saying that the
negative frequency side of the transform is the complex conjugate of the positive
side.
A very interesting property arises in connection with the products of functions.
Consider the transform of two time functions
(;(w) = f FA(t)FB(t)e-iwtdt
(;(w) = ff (;A(w)e+iwtdwFB(t)e-iwt dt
This particular form is called a convolution. It will have use in understanding the
effects of sampling and filtering on the computed transforms. For example, a signal
truncated in the time domain can be thought of as the product of the original signal
with the truncating function. The result in the frequency domain is then no longer
a simple representation of either.
There is a similar relation for products in the frequency domain, namely,
(1.9)
This shows that a time domain convolution can be performed as frequency domain
multiplication. Even though this involves both a forward and inverse transform it is
computationally faster than performing the convolution directly. All the mechan-
ical systems considered in the later chapters will be represented in the frequency
domain as products of the input times the system response. In fact, as will be seen,
the reason why the frequency domain is so useful for analyzing these systems is
because the complicated convolution relations become simple algebraic relations.
Fourier Series
Let the time function F(t) be known only over a period T; to apply the continuous
Fourier transform to it, we must extend it somehow to the infinity limits. In the
Fourier series representation, the function is assumed extended to plus and minus
infinity as a periodic function with the period T. We can view this either as separate
functions of duration T placed one after the other; or as the superposition of
separate functions of infinite duration but with nonzero behavior only over the
period T. We adopt this latter view as shown in Figure 1.3.
-00 +00
n .(J.
n
-00 +00
+
-00 +00
+
-00 ~+oo
Figure 1.3: Periodic extended signal as a superposition of infinite signals.
We saw from the rectangular pulse example that if Coo(w) is the transform of a
pulse, the transform of the same pulse shifted an amount T is Coo(w)e- iwT • It is
obvious therefore that the transform of the periodic signal can be represented as
This transform will show infinite peaks whenever the frequency is one of the
following discrete values Wn = 2rr:n/T. Under this circumstance, each of the
exponential terms is unity and there is an infinity of them. (This result should not
be surprising since the C(O) component is the area under the curve, and for the
periodic rectangle of Figure 1.3 we see that it is infinite.) At other frequencies, the
exponentials are as likely to be positive as negative and hence their sum will be
relatively small. Therefore our first conclusion about the transform of a periodic
extended signal is that it will show very sharp spectral peaks. We can go further and
say that the transform is zero everywhere except at the discrete frequency values
Wn = 2rr:n/T where it has an infinite value. We represent this behavior by use of
the delta function, 8(x); this special function is zero everywhere except at x = 0
where it is infinite. It has the very important additional property that its integral
over the whole domain is unity. Thus
i: i:
The remainder of the transform pair is given by
Interchanging the summation and the integration, and using the properties of the
delta function gives
n
Integrate both sides of this equation over a time period T, and realizing that all
terms on the right-hand side are zero except the first, gives that
Except for a normalizing constant, this is the complex Fourier series representation
of a periodic signal.
Consider the transform of the rectangular pulse of the last section. The coeffi-
cients are given by
This result is the product of three terms. The first, Foa, is the size of the pulse
as represented by the area. The third, the exponential, is a phase change due to
the shifting of the pulse relative to the time origin. For example, if the pulse is
symmetric about t = 0, then t" = -all and there is no phase change of the
transform. The second term, sinc function, is the core of the transform and its
amplitude is shown plotted in Figure 1.4 for various periods. Notice that the spacing
of the coefficients is at every 1I T Hertz in the frequency domain. For the periods
T = 500, 200, 100 JLS this gives spacings of t:.f = 2, 5, 10 kHz, respectively.
Notice that in all the cases of Figure 1.4, the Fourier series gives the exact
values of the continuous transform, but it does so only at discrete frequencies. The
discretization of the frequencies is given by
2:rrn
wn=r
Therefore, for a given pulse, a larger period gives a more dense distribution of
coefficients, approaching a continuous distribution in the limit of an infinite period.
This, of course, is the continuous transform limit. The finite time integral causes
the transform to be discretized in the frequency domain.
14 Chapter 1. Spectral Analysis of Wave Motion
o Fourier series
- - Continuous
Figure 1.4: Fourier series coefficients for a rectangular pulse of nonzero duration a = 50 J-LS
extended with different periods T.
F2
F I,""
Fo:' b..T F8 F9
Time
to tl t2 tm t4 t5 t6 t7 t8 t9
We see that this is the sum of the transforms of a series of rectangles each shifted
1.2 Discrete Fourier Transform 15
in time by t = tm + tlT /2. The contribution of each of these will now be evaluated
more closely.
First look at the summation term. If n > M, that is, if n = M + n *. then the
exponential term becomes
M-\
L Fme-in·wotm
m=O
. sin(x) wntlT n T n
smc(x) == - - , x=----7r---7r-
x 2 TM M
The sinc function is such that it decreases rapidly with increasing argument and
is negligible beyond its first zero. The first zero occurs where x = 7r or n .. M;
if M is made very large, that is. the integration segments are made very small,
then it will be the higher order coefficients (i.e., large n) that are in the vicinity of
the first zero. Let it be further assumed that the magnitude of these higher order
coefficients are negligibly small. Then an approximation for the coefficients is
L Fme-iw,l
M
Dn ~ tlT{I} m
L L
N-\ N-\
tlT Fme-iw"lm = l!.T Fme-i2rtnm/N (1.11)
m=O m=O
where both m and n range from 0 to N - 1. These are the definition of what is
called the discrete Fourier transform (DFT). It is interesting to note that the expo-
nentials do not contain dimensional quantities; only the integers n, m, N appear. In
this transform, both the time and frequency domains are discretized, and as a con-
sequence, the transform behaves periodically in both domains. The dimensional
16 Chapter 1. Spectral Analysis of Wave Motion
scale factors f).T, liT have been retained so that the discrete transfonn gives the
same numerical values as the continuous transfonn. There are other possibilities
for these scales found in the literature.
The discrete transfonn enjoys all the same properties as the continuous trans-
fonn; the only significant difference is that both the time domain and frequency
domain functions are now periodic. To put this point into perspective consider the
following: A discrete Fourier transfonn seeks to represent a signal (known over a
finite time T) by a finite number of frequencies. Thus it is the continuous Fourier
transfonn of a periodic signal. Alternatively, the continuous Fourier transfonn it-
self can be viewed as a discrete Fourier transfonn of a signal but with an infinite
period. The lesson is that by choosing a large signal sample length, the effect due
to the periodicity assumption can be minimized and the discrete Fourier transfonn
approaches the continuous Fourier transfonn.
Numerical Example
To help better illustrate the properties of the discrete transfonn consider the fol-
lowing numerical example. This can also serve as the test case for any numerical
implementation of the DFT.
Let a real-only function be given by the following sampled values:
with f).T = 1, N = 8. This has the shape of a rectangular pulse if the points are
connected by straight lines - remember, however, that the data is sampled and
hence no infonnation is actually known between the sampling points. Eight points
are given, thus it is implicit that the function repeats itself beyond that. That is, the
next few values are 0,1,1,0, and so on. The transfonn becomes
Dn - L
7
Fme-i21rnm/S ... FJe-iTln/4 + F2e- iTln / 2
m-o
The first ten transfonn points, in explicit fonn, are
Do 2.0
DJ 0.707 - 1.707i
D2 -1.0 - 1.0i
D3 -0.707 + 0.293i
D4 0.0
Ds -0.707 - 0.293i
D6 -1.0 + 1.0i
D7 0.707 + 1.707i
Ds 2.0
D9 0.707 - 1.707i
1.2 Discrete Fourier Transform 17
The obvious features of the transform is that it is complex and that it begins to repeat
itself beyond n = 7. Note also that D4 [the <4N + l)th value] is the Nyquist value.
The real part of the transform is symmetric about the Nyquist frequency, wh!le the
4
imaginary part is anti-symmetric. It follows from this that the sum [Dn + D N - n]
gives only the real part, that is,
These two functions are the even and odd decompositions, respectively, of the
transform. Also note that Do is the area under the function.
L Fme-i2rrnm/N ,
N-I
Sn = n = 0, 1, ... , N - 1
m-o
We write this in the expanded form
So {Fo + FI + F2 + ... }
SI {Fo + Fle-i2rrl/N + F2e-i2rr2/N + ... }
S2 {Fo + Fle-i2rr2/N + F2e-i2rr4/N + ... }
and so on. For each sum Sn, there are (N - 1) complex products and eN - 1)
complex sums. Consequently. the total number of computations (in round terms)
is on the order of 2N 2 • The purpose of the FFf is to take advantage of the special
form of the exponential terms to reduce the number of computations to less than
N 2•
18 Chapter 1. Spectral Analysis of Wave Motion
The key to understanding the FFr algorithm lies in seeing the repeated forms
of numbers. This will be motivated by considering the special case of N being 8.
First consider the matrix of the exponents -i2rr(";.;'):
0 0 0 0 0
0 1 2 3 (N - 1)
-i2rr 0 2 4 6 2(N - 1)
0 3 6 9 3(N - 1)
N
It is apparent that for an arbitrary value of N, 2rr will not, in general, be multiplied
by an integer number. These exponents, however, can be made quite regular if N
is highly composite. For example, if N is one of the following
N = 2Y = 2, 4, 8, 16,32,64, 112,256, 512, 1024, ...
then the effective number of different integers in the matrix is decreased. Thus if
N = 8 we get
0 0 0 0 0 0 0 0
0 1 2 3 4 5 6 7
0 2 4 6 0 8+2 8+4 8+6
-i2rr 0 3 6 8+1 8+4 8+7 16+2 16+5
8 0 4 8+0 8+4 16+0 16+4 24+0 24+0
0 5 8+2 8+7 16+4 24+ 1 24+6 32 +3
0 6 8+4 16+2 24+0 24+6 32+4 40+2
0 7 8+6 16+5 24+4 32+3 40+2 48 + 1
which will effectively evaluate to
0 0 0 0 0 0 0 0
0 1 2 3 4 5 6 7
0 2 4 6 0 2 4 6
-i2rr 0 3 6 1 4 7 2 5
8 0 4 0 4 0 4 0 0
0 5 2 7 4 1 6 3
0 6 4 2 0 6 4 2
0 7 6 5 4 3 2 1
This comes about because the exponentials take on the following simple forms
e- i211"[O] = e- i211"[1] = e- i211"[2] = e- i211"[3] = ... = 1
We see that many of the computations used in forming one of the summations
is also used in the others. For example, So, S2, S4, S6 all use the sum (Fo + F4).
Realizing that e-i2rr4/8 = -1, then we also see that all the odd summations contain
common terms such as (Fo - F4). This re-use of the same computations is the reason
a great reduction of computational effort is afforded by the FFf. The algorithm
sets up the bookkeeping so that this is done in a systematic way.
The number of computations with and without the FFf algorithm is given by
~Nlog2N versus
When N = 8, this gives a speed factor of only 3.5:1, but when N = 1024, this
jumps to over 100: 1. It is this excellent performance at large N that makes the
application of Fourier analysis feasible for practical problems. On a benchmark
machine of 1 MFIops, say, a 1024-point transform takes significantly less than one
second.
The fast Fourier transform algorithm is so efficient that it has revolutionized
the whole area of spectral analysis. It can be shown quite simply that it enjoys
all the same properties of the continuous transform. Therefore, in the subsequent
analysis, we will assume that any time input or response can be represented in the
spectral form
L
N-J
F(t) = Cne+iwnt
and the tasks of forward and inverse transforms are accomplished with a computer
program.
o FFT
o samples -eFT
that N real points are transformed into N /2 complex points and no information
is gained or lost. Therefore, the useful frequency range extends only up to the
Nyquist, given by
I
fNyquist = 2D-.T
This range is increased only by decreasing D-. T. Thus for a fixed number of points,
fine resolution in the time domain (small D-.T) means course resolution in the
frequency domain. Finer resolution in the frequency domain is achieved only by
increasing the sample length T.
It is also seen from the figure that the match between the FFT amplitude and
the continuous Fourier transform is very good at low frequencies but gets worse
at the higher frequencies. As mentioned before, the discrete and the continuous
transforms match closely only if the highest significant frequency in the signal is
less than the Nyquist. The rectangular pulse can be represented exactly only by
using an infinite number of sinusoids but the contributing amplitudes get smaller
as frequency increases and therefore a finite number of sinusoids can suffice.
For a given sample rate D-. T, the number of samples only determine the density of
transform points. Thus the top two FFTs in Figure 1.6 are numerically identical at
the common frequencies. Increasing the sample rate for a given number of samples
increases the Nyquist frequency, and therefore the range of comparison between the
discrete and continuous transforms. It is seen that by the fourth plot, the amplitudes
in the vicinity of the Nyquist are negligible. Using the discrete transform puts an
upper limit on the maximum frequency available to characterize the signal. If
the signal is not smooth, the amplitudes of the high-frequency sinusoids used to
describe the signal are high. Any attempt at capping the high-frequency sinusoids
introduces a distortion in the amplitudes of the lower-frequency sinusoids. This is
called aLiasing and the next section discusses it in more detail.
1.3 Examples Using the FFf Algorithm 21
__J\__
Time ij.Ls] Frequency [kHz]
..................................~................................I......~'-
. ........
' ! I ", I"" I !" I, II I I I,.! I
o 200 400 600 BOO O. 5. 10. 15. 20. 25. 30. 35. 40.
Figure 1.7: Comparison of some pulse type signals in both the time and frequency domains.
The remaining two plots are for modulated signals - in this case a sinusoid
modulated (multiplied) by a triangle. The sinusoid function on its own would give
a spike at 15 kHz, whereas the triangular modulations on their own are as shown
22 Chapter 1. Spectral Analysis of Wave Motion
for the top traces. The product of the two in the time domain is a convolution in
the frequency domain. The convolution tends to distribute the effect of the pulse.
A point of interest here is that if a very narrow-banded signal in the frequency
domain is desired. then it is necessary to extend the time domain signal as shown
in the bottom example.
The modulated waves are called wave packets or wave groups because they
contain a narrow collection of frequency components. It may have been thought
that the triangular pulse is a more natural group since in the time domain it is
highly localized. but what is also important to us is the frequency content of the
wave and we see that the modulated wave is localized in both the space and time
domains.
Effect of Superpositions
The signals we will deal with will often have multiple reflections. A sense of their
effect can be gauged from Figure 1.8 where the traces are arranged so as to have
a different number of reflections.
~~-------------
1v~
reflections, the spectral peaks become very sharp and this situation would then be
called resonance.
Measurement Techniques
Testing associated with waves is somewhat different from that of vibrations. Some
of the characteristics of the testing to record transient signals associated with impact
and stress wave propagation in structures are:
• High frequency content (1 kHz - 1 MHz),
• Initial portion of the signal is very important,
• Only a small number of "runs" are performed,
• The number of data recording channels are usually limited (2 - 8),
• Only a short duration of the trace is analyzed (0 - 5000 /-Ls).
There is quite a range of measurement techniques available for studying wave
propagation problems. Of course, some are more suitable for certain situations
than others, but since the emphasis in this book is on structures, then only those
techniques most commonly used for structural analysis will be surveyed. A typical
setup is shown in Figure 1.9.
Most types of electrical resistance strain gages are usually adequate and since
the events are of short duration, temperature compensation is usually not a serious
issue. The largest gage size allowable is dependent on the highest significant fre-
quency (or shortest wavelength) of the signal, but generally gages equal to or less
than 3 mm (0.125 in.) are good for most situations. A rule of thumb for estimating
the maximum length is
L= l~ft (1.12)
where L is the gage length, f is the highest significant frequency, and E and p are
the Young's modulus and density, respectively, of the structural material. A typical
gage of length 3 mm (0.125 in.) is good to about 160 kHz. Other aspects of the
use of strain gages in dynamic situations are covered in References [8, 98]. Either
24 Chapter 1. Spectral Analysis of Wave Motion
y Force
Transducer
Force
Transducer
accelerometers
--'-------+----r-r-r-l~'___'__~
PL--CB-303A-'------03
Power
... _ __.__r_r_l,:"
Supplies
TekAM502
computer waveform recorder pre-amps
Figure 1.9: Typical experimental setup for a plate (left) or frame (right).
has a resolution of 11256 of the full scale. If the signal can be made to fill at least
half full scale, then this resolution of about 1% of the maximum signal is ade-
quate. This situation is best achieved by having good preamplifiers. Many of the
newer digital recorders combine the features of an oscilloscope with those of a
small stand-alone microcomputer. But if the data are to be eventually transferred
to a computer for analysis, then it seems reasonable that this computer should also
perform all the processing. However, the preprocessing capability of the recording
machine is very useful during the setup stage.
When we compare the truncated measured signal to the (imagined) infinite trace,
we say the fonner was "windowed" - only a small portion of the infinite trace is
visible through the window. Examples of time windows commonly used are the
rectangle and the Hanning window (which is like a rectangle with smoothed edges);
see References [9, 13] for more examples of windows. None of the windows usually
used in vibration studies seem particularly appropriate for transient signals of the
type obtained from propagating waves. The window used in most of the studies
to follow is the rectangular window tapered in some way on the large time side.
This window is thought to be adequate and is certainly the easiest to implement. A
problem with rapid truncation is that the reconstructed signals exhibit significant
disturbances at the point of truncation. If the sample size is large compared to the
length of interest, then this is not a serious problem because the disturbances can
be easily identified since they are localized in time and will not be confused with
the true signal.
Sampled Waveforms
By necessity, actual signal records are finite in length, and this can add some
difficulties to the scheme of estimating its spectral content. This section reviews
the major effects that should be considered.
An actual signal of very long duration must be truncated to a reasonable size
before the discrete Fourier transfonn is evaluated. The transfonn of the signal will
therefore be dependent on the actual function F(t), say, and the way it is sampled.
Note that if the sampling rate is every ll. T, then the highest detectable frequency
using the DFT is 1/(2ll.T). Thus, if ll.T is too large, the high frequency appears as
a lower frequency (its alias) as shown in Figure 1.11. This phenomenon of aliasing
can be avoided if the sampling rate is high enough. By "high enough" is implied
a rate commensurate with the highest significant frequency in the signal. From a
practical point of view, this is also avoided by recording with the analog filters set
so as to remove significant frequencies above the Nyquist.
• data
-sampled
The estimated spectrum is given only at discrete frequencies and this can cause a
problem known as leakage. For example, suppose there is a spectral peak at 17 kHz
and the sampling is such that the spectrum values are only at every 2 kHz. Then the
energy associated with the 17 kHz peak "leaks" into the neighboring frequencies
and distorts their spectral estimate.
Leakage is most significant in instances where the signal has sharp spectral
peaks and is least significant in cases where the signal is a flat, broad-banded
signal. In the analysis of the impact of structures, the signals generated generally
do not exhibit very sharp spectral peaks (as in the case of vibrations). However, it
can become a problem if there are very many reflections present in the signal.
Since discrete Fourier analysis represents a finite sample of an infinite signal on
a finite period, then schemes for increasing the apparent period must be used. The
simplest means of doing this is by padding. There are various ways of padding
the signal and the one most commonly used is that of simply adding zeros. Alter-
natively, Reference [37] shows how a long term estimate of the signal behavior
can also be used to pad the signal. The main consequence of padding is that since
a larger time window is used then higher resolution in the frequency domain is
achieved.
As regards padding with zeros, it should be pointed out that it really does not
matter if the zeros are added before the recorded signal or after. Thus even if the
signal never decreases to zero, padding is justified because it represents the early
quiescent part of the signal.
Herein lies the advantage of the spectral approach to solving differential equations
- algebraic expressions in the Fourier coefficients replace the time derivatives.
That is, there is a reduction in the number of derivatives occurring.
Similarly, the spatial derivatives are represented by
au
- = -
a L une
~ iw t
n = L -aUen " iw t
ax ax ax
and in shorthand notation
au au" au
-::::}- or (1.15)
ax ax ax
In this case there appears to be no reduction, but as will be seen later, with the re-
moval of time as an independent variable, these derivatives often become ordinary
derivatives, and thus more amenable to integration.
1.5 Spectral Analysis of Wave Motion 29
Spectrum Relation
Consider the following general, linear (one-dimensional), homogeneous differen-
tial equation for the dependent variable u(x, t):
au au a2 u a2 u a2 u
u+a- +b- +c-+d- +e-- + ... =0
ax at ax 2 at 2 axat
The coefficients a, b, c, ... are assumed not to depend on time, but could be func-
tions of position. If, now, the solution is given the spectral representation
Since each eiw"t term is independent, then this equation must be satisfied for each
n. That is, there are n simultaneous equations of the form
(1.16)
where AI, A2, ••• depend on frequency, possibly on position, and are complex. The
subscript n is dropped, but it will be understood that an equation of the above
form must be solved at each frequency. We see that the original partial differential
equation becomes a set of ordinary linear differential equations in the Fourier
coefficients Un.
Consider the special case when the coefficients a, b, c, ... in the partial differen-
tial equation are independent of position and lead to linear differential equations
with constant coefficients. Reference [63] considers in good detail the solution
of equations of this form and should be consulted for the details not given here.
These equations have solutions of the form eAx , where A is obtained by solving the
algebraic characteristic equation (obtained after substituting into Equation (1.16»
(1.17)
au + - = 0
A du => [a - ik]C = 0
dx
This gives k and the solution as
k = -ia, u(x) = Ce- ax
Since this has constant coefficients, assume Ce- ikx is a solution which (on substi-
tution) results in the characteristic equation
k2(W) -= -wa
The set of coefficients CI(W), C2 (w) are called the amplitude spectrums.
This solution correspond to sinusoids (infinite wave trains) moving to the left
and right, respectively. A line in the x - t plane connecting a common point is
a straight line. That is, points of common phase (position on the sinusoid) travel
along straight lines in x - t as shown in Figure 1.12 for a forward moving train. If
the slope of this line is thought of as a speed, then it can be called the phase speed
and is given by
x 1 W
c=-=-=-
t a k
The relation between phase speed and frequency is called the dispersion relation.
We see that different frequency components travel at the same speed and that the
superposition also travels at the same speed. When the phase speed is constant with
respect to frequency, the signal is said to be nondispersive and it will maintain its
superposed shape. That is, at each different position in space, the sinusoids have
the same relative phase and so will superpose to give the same shape.
Consider now the following differential equation (which actually represents
flexural wave motion in beams)
(1.23)
32 Chapter 1. Spectral Analysis of Wave Motion
--JI/VI/\/'vVWVVWI/I/VVVI/V'\}\-- --,/l!v\NWI/VWWJWWJ\jV\~-
. " ,", :\/",/\/",j\/\,:"'./\/\' :',",;',,", ;----
--~,N'vVI/V\ANvvVWVVWNv~,----
,":',' ',,''.;',,'':',,1-
Figure 1.12: Segments of an infinite wave train at different positions. Left: nondispersive
system, individual sinusoids travel at same speed. Right: dispersive system, individual
sinusoids travel at different speed.
where, again, a is real and positive. Since the coefficients are constant, assume
C e-i(kx-wtj is a solution giving the characteristic equation as
+L C3e-afox+iwt +L C4eafox+iwt
The first two are wave terms, as in the previous example, and again travel with a
constant phase speed given by
w 1
c == - =-../W
k a
Different frequency components, however, travel at different speeds. As a result,
the superposition of the components changes at the different locations and the wave
will appear to change shape as it propagates. When the phase speed is not constant
with respect to frequency, the signal is said to be dispersive. This is demonstrated
for two wave train segments shown in Figure 1.12. Observe that the resultant signal
travels at a speed different from the individual sinusoids.
1.5 Spectral Analysis of Wave Motion 33
The last two terms in the solution are spatially damped vibrations. While not
wave solutions per se (they are usually referred to as evanescent waves), they
nonetheless play a crucial role in being able to satisfy the imposed boundary
conditions when, say, a wave interacts with a joint.
The two previous examples had the familiar double time derivative associated
with inertia. Consider, as a different case, the following differential equation related
to Fourier's law of heat conduction
a2 u
-=a-
au (1.24)
ax 2 at
where a is real and positive. Again, since the coefficients are constant, assume
Ce-i(kx-wt) is a solution giving the characteristic equation as
[_k 2 - iwa] C = 0
There are two modes having the spectrum relations k(w) - ±,J-iaw or
k\(w) - -(1 - i),jaw/2 == -(1 - i)f3, k2 (w) = +(1 - i)Jaw/2 == +(1 - i)f3
This is comprised of left and right propagating terms, but both are also spatially
damped. Further, the solutions are dispersive in the sense that f3 is not a linear
function of w.
Although this solution does not fall neatly into a simple wave description, the
spectral approach solves it with ease. This is very fortunate, because as waves
interact with boundaries and discontinuities, they lose their "wave nature" causing
a more complicated dynamic motion and the spectral approach can still handle
these cases.
Group Speed
The dynamic solution is comprised of the superposition of many harmonics and it
was shown that some are waves, some are damped waves, while some others are
vibrations. All these combine to give the observed motion. It is of interest to know
how to describe the propagating part of the disturbance since that is what arrives
at a remote location. It must be reiterated, however, that for all practical problems,
all types of solutions are present, and it may be unwarranted to attempt to isolate
the so-called "wave."
In any local region of interest the solution (for a typical mode) can be written
in the exponential form
34 Chapter I. Spectral Analysis of Wave Motion
At a typical frequency, the wavenumber can be written in terms of its real and
imaginary parts as
This resulting wave is comprised of two parts besides the average amplitude spec-
trum. There is a sinusoid (called the carrier wave) of average frequency w* and
wavenumber k*, and travels with average speed c* = w* / k*. But this is modu-
lated by another wave (called the group wave) of wavenumber 4/).k, frequency
4/).k dw* /dk*, and travels at a wave speed of dw* /dk*. The phase velocity of the
modulation is called the group speed. That is,
dw ( dC) dc (1.27)
cg == dk = C / I - k dw = c + k dk
In general, the group speed is numerically different from the phase speed.
While the above analysis is for only two harmonics, it can be imagined that many
harmonics interact in a similar manner giving rise to a carrier wave modulated by
a group wave. This is what is depicted in Figure 1.12 and also in Figure 1.1; the
group corresponds to the envelope of the disturbance and the phase to where the
zeros occur. In general, however, it is not possible to "see" the individual sinusoids
~ only the group superposition is observed. It cannot be emphasized enough the
importance of the group speed behavior because this is what is actually there. The
ability to synthesize or assemble the group behavior is a unique feature of spectral
1.6 Propagating and Reconstructing Waves 35
analysis coupled with the FFf algorithm. Put another way, the FFf allows both
the assembling and disassembling of the group response.
It is possible to have very complicated group behavior. For example, it is possible
to have a solution for which the group speed is positive (the disturbance moves
forward) and yet the phase speed is negative. According to Reference [48], this
is the situation as a caterpillar moves forward, the ripples in its body move from
head to tail.
The amplitude A can be complex and the above form is for a forward moving
wave. The following is a collection of terms often used in wave analysis:
Basic Algorithm
In its simplest terms, the solution to a waves problem is represented as
where Gis the analytically known transfer function of the problem. It is a function
of position x and has different numerical values at each frequency. Pn is the ampli-
tude spectrum; this is known from the input conditions or from some measurement.
Thus Pn G is recognized as the Fourier transform of the solution. Of course, it is
different at each position, but once it is evaluated at a particular position then its
inverse immediately gives the time history of the solution at that point. Figure 1.13
is a flow diagram for the basic algorithm to propagate a wave.
r--------------,
Time function
~ P(tm)
L, ______________ ~
FFT
(forward) r-------------- ..
• Transform
. "'
P(wn )
_ _ _ _ _ _ _ _ _ _ _ _ _ _ .1
Do Loop
Un = G(wn)Pn
Files
r--------------,
• Transform
OIl u(wn)
FFT
L ______________ ~
(inverse) r--------------,
• Time function
Programs u(tm)
Briefly, the time input p(t) is converted to its spectrum Pn through a use of the
forward FFf. The transformed solution is then obtained by evaluating the product
Un = PnG(kmnx)
at each frequency and some position. This is finally reconstructed in the time
domain by use of the inverse FFf. In the process, it is necessary to realize (when
using the FFf to perform the inversion) that Pn Gis evaluated only up to the Nyquist
frequency and the remainder is obtained by imposing that it must be the complex
conjugate of the initial part. This ensures that the reconstructed time history is real
only.
An additional feature of note is that when doing propagations the dc (zero fre-
quency) component is undetermined since it does not propagate. It is advantageous
to remove its arbitrariness by imposing that the first value of the reconstruction be
1.6 Propagating and Reconstructing Waves 37
Uo = -
A
~U"
' " A
This is consistent with the idea that the system is quiescent before the arrival of the
wave. Furthermore, it is good practice to give all signals a zero header (a beginning
portion of zeros before the actual signal) to emphasize this point.
--J \
X= o
x= 5m
\ x=l Om
\ x=l 5m
~ X= 20m
\
~~~~~~~~~~~~~~~~-L~~~~!Time~l
X= 25m
kn -_ ..jwnwo
Co
where Wo is a constant. (Strictly speaking the transfer function should also in-
clude the damped vibration terms, but the present form is adequate for far-field
predictions.) The amplitude spectrum is obviously as given before but the recon-
structions shown in Figure 1.15 are very different. What starts out as a well-defined
triangular shape disperses as it propagates. Not only does it change its shape, but
it is no longer possible to identify the so-called pulse that is propagating. This is
characteristic of dispersive signals and is due to the relation between k and w being
nonlinear. If the speed is slow for low frequencies, then it takes a very long time
for the low-frequency part of the signal to arrive. Thus it will have a long tail.
Figure 1.15 also shows the results of propagating a dispersive signal a large
distance. The curious feature is that it is the lower frequencies that propagate
from the left into view first. To minimize the effect of this it is a good idea to
pad the window with zeros. Thus the creeping components must first traverse the
zeros before they come into view. When the higher frequencies arrive sooner than
unwanted lower frequencies, then (depending on the circumstance) the unwanted
components can often simply be filtered out. There are two ways of doing this.
The first method is simply by band-pass filtering, that is, every component below a
certain frequency is blocked. The second scheme uses an n-point moving average
but retains the difference and not the average. This latter scheme can do an excellent
job in reconstructing the propagating wave. The most effective scheme, however,
is simply to increase the window size at the price of adding more terms. This is
feasible because of the efficiency of the FFT algorithm and its essentially linear
dependence on N.
1.6 Propagating and Reconstructing Waves 39
x=O
x=O.Sm
x=1.0m
x=1.Sm
x=2.0m
x=2.Sm
Integration of Signals
Inherent in the spectral analysis approach is the assumption that all functions of
interest can be both integrated and differentiated. However, during the integration
stage care must be taken to include the proper constants of integration. This is of
special importance when, for example, obtaining displacement from strain since
some unexpected side effects arise.
To illustrate some of the issues that arise due to the enforced periodicity of
the results, Figure 1.16 shows the displacement at the impact site of a beam. The
periodicity requires the beginning and final points of the reconstruction to be the
same. This figure shows the profound effect of the window size. For this particular
beam problem a nonzero value of displacement should persist for all time after
the passage of the pulse. However, periodicity of the result on the finite window
forces the end displacement to zero so that the trace is continuous from the end of
one window to the beginning of the next.
This problem is minimized when the output is at the same differentiation level as
the input. For example, when obtaining strain at one location from strain at another,
then padding with zeros can be used. But in the case of displacement from force it
is seen that padding does not help. What is required is a window size large enough
so that the slope of the trailing edge of the history is not significant. For a given
sample rate, the window can only be increased by increasing the number of points
in the transform. Conversely, for a fixed number of points, the sample ratel:l. T
must be increased, thus distorting the higher frequencies and running the risk of
aliasing.
To understand the effect in time, it is of interest to write the differential relations
in terms of the frequency and wavenumber for both beams and rods. (More details
40 Chapter 1. Spectral Analysis of Wave Motion
x=O.5m
Time [JlS]
, ,
o. 1000. 2000. 3000. 4000. 5000.
Figure 1.16: Integration of signals shows wraparound effect.
on the origin of the rod and beam equations can be found in the next two chapters.)
The simplest propagating solutions for rods and beams are, respectively,
ROD BEAM
dv
- -+ kv
dx
du du d2v 2 dv
- -+ ku -+ wu -+ - strain - -+ k v -+ wv -+ -
dx dt dx 2 dt
d v
3
- -+ k 3 v
dx 3
d4v d2v
loading - -+ k 4 v -+ w 2 v -+ -
dx 4 dt 2
Thus strain and velocity are at comparable levels of w, and likewise for the loading
and accelerations. This chart can be interpreted by noting, for example, that for
beams integration of strain twice with respect to x to find displacement is equivalent
to integration of velocity once with respect to time. The displacement is known,
therefore, to within a linear function of time:
v= ff edxdx+C]t+C2
Problems
1.1 Show that by introducing the real coefficients an and b n such that (an - ibn) ==
en2/ T, we get the usual Fourier series representation
That is, the real part is even and the imaginary part is odd. What this shows is
that when performing Fourier transforms, information is neither gained nor
lost. For example, 1024 pieces of real-only data are converted into 512 pieces
of complex data.
- Reference [11], p. 193.
1.3 Consider a function f(tm) that is described by 2N samples divided as follows:
It is seen that the summation terms are actually N - point transforms. This is a
completely general result and shows that if only an N -algorithm is available,
nonetheless, transforms of multiple N can be performed. More generally, if
the data size is 8N say, then the data can be broken down as 2 sets of 4N,
then 4 sets of 2N, and finally, 8 sets of N. The N - point algorithm can then
be used on each set and some post-sorting of the arrays is all that is required
to reconstruct the 8N transform.
- Reference [11], p. 193.
42 Chapter 1. Spectral Analysis of Wave Motion
F(n) i
HR(n) + R(N - n)] + i [len) - leN - n)]
+e-irrnIN n
[len) + leN - n)] - i ~ [R(n) - R(N - n)]}
i:
- Reference [11), p. 193.
1.5 Given the integral
let) = f(w)eih(w.tJdw
Show that in the case when few) is narrow banded about a frequency We> say,
then the integral simply becomes
dh(w,.) = 0
I (t) "'"
dw
Rods and struts are very important structural elements and form the basis of many
truss and grid frameworks. Since their load bearing capability is axial, then as
waveguides they conduct only longitudinal wave motion.
This chapter begins by using elementary mechanics to obtain the equations of
motion of the rod. As it happens, the waves are governed by the simple wave equa-
tion and so many of the results are easily interpreted. The richness and versatility
of the spectral approach is illustrated by considering such usually complicating
factors as viscoelasticity and elastic constraints. Particular emphasis is placed on
the waves interacting with discontinuities such as boundaries and changes in cross
section, as shown in Figure 2.1.
The basic waves in rods are described by a single mode; coupled thermoelasticity
is introduced in this chapter as a first example of handling multi-mode problem!!.
time
~
Figure 2.1: Velocity response of a semi-infinite two-material rod free at one end and im-
pacted at the junction. Note how the pulse is initially trapped in the upper material but
eventually leaks away after multiple reflections.
43
44 Chapter 2. Longitudinal Waves in Rods
Equation of Motion
The elementary theory considers the rod to be long and slender, and assumes it
supports only I-D axial stress as shown in Figure 2.2. It further assumes that the
lateral contraction (or the Poisson's ratio effect) can be neglected. Both of these
restrictions will be removed in Chapter 4 when the higher order rod theories are
developed.
-
q(x) - 1]Azi
EA,pA
tY
L
II...-_ _
x
----1...~_x...L-1
f----+
_-----I
u(x, t)
..-
F
GJ
~>-ii
~x F+ ~F
Following the assumption of only one displacement, u(x), the axial strain is
given by
au
E
xx
=-
ax
Let the material behavior be linear elastic, then the I-D form of Hooke's law gives
the stress as
E = Young's Modulus
F = f O'xx dA = EA ax
au
Let q(x, t) be the externally applied axial force per unit length. While not es-
sential, we will consider the cross section to be rectangular of depth b and height
h. With reference to Figure 2.2, the balance of forces gives
where pA is the mass density per unit length of the rod, I] is the damping (viscous)
per unit volume, and the super dot indicates a time derivative. If the ~ quantities
2.1 Elementary Rod Theory 45
are very small, then the equation of motion becomes (after dividing through by
~x)
ax
aF
=
a2 u
pA atT + TJAat - q
au
The independent variables are x and t. Substituting for the force in terms of the
displacement gives the equation of motion
-a [ EA- a2 u +TJA--q
au] =pA- au (2.1)
ax ax at 2 at
In the special case of uniform properties and no damping, all dependent variables
(stress, strain, etc.) have an equation of the form
(2.2)
for the homogeneous part. In this case, the waves in the rod are governed by the
simple wave equation and the general solution is that of D' Alembert, given by
This solution does not hold for the general rods of interest in this chapter and
therefore will not be pursued any further in its present form.
Spectral Analysis
Let the solution have the spectral representation developed in the last chapter. The
kernel solutions corresponding to this are obtained by considering the homoge-
neous equation
d [ EA-
- dflJ +w pAu~ - 2 iWTJAu~ =-q~ (2.4)
dx dx
This is the basic equation used in the remainder of the chapter. Assume that both
the modulus and area do not vary with position, then the homogeneous differential
equation for the Fourier coefficients becomes
d2~
U 2 ~
EA-2 + (W pA - iWTJA)u = 0
dx
This is an ordinary differential equation with constant coefficients - the frequency
is considered as a parameter.
Since this equation has constant coefficients, then it has the exponential solutions
e- ikx leading to
(2.5)
Chapter 2. Longitudinal Waves in Rods
where A and B are the undetermined amplitudes at each frequency. When combined
with the time variation, this solution corresponds to two waves: a forward moving
wave and a backward moving wave. That is,
(2.6)
It is understood that all quantities inside the summation (A, B, k, etc.) could depend
on the frequency w. The spectrum relation is shown plotted in Figures 2.3 .
.5 Wavenumber, kh
.4
-simple real
.3
- viscoelastic
.2
.1
imaginary
.0
-.1
,, , , !! !,,! , • t ,I" ,!, : , , !
Frequency [kHz)
! , I , , , I ,
Figure 2.3: Spectrum relations for a rod; comparison of elastic and viscoelastic materials.
kl =wJ pA
EA
(2.7)
which is linear in frequency. This gives, for the dispersion relation, constant phase
and group speeds of
dw {EA
cg = dk = VPA = Co
(2.8)
This simple behavior is shown in Figure 2.4 where it is seen that the phase and
group speeds are constant. Also shown are some other complicated behaviors (to
be treated later) and serve as a reminder that even the simple rod is capable of a
wide variety of behaviors: the goal of this chapter is to develop some of the tools
necessary for their description.
As a useful reference, the properties and speeds of some typical structural ma-
terials are given in Table 2.1. The units for density (and mass) are often a source
of confusion in the imperial system of units; in the table it is given as weight per
unit volume divided by gravity, i.e., p = W / g.
2.2 Basic Solution for Waves in Rods 47
1.2
1.0
.8
- - simple, group & phase
.6 - - viscoelastic, group
......... viscoelastic, phase
.4
.2
Frequency [kHz]
.0 ' !
t ! !, !,,' ,'!! ,I,! , f ! ! ! , , , , ! , , ! ! !
L y
x
pet) -+- D -+- F(x = 0) -IL--_~_>-_A_e_-_ik_x_ _ _ _---II ~ 00
~ u(x =0)
Let the end ofthe bar atx = 0 be subjected to a force history pet), then according
to the free-body diagram of Figure 2.5 we have
au(x,t)
atx = 0: F=EA =-P(t)
ax
Let all functions of time have the spectral representation; the boundary condition
becomes, in expanded form,
This has to be true for all time, hence the equality must be true on a term by term
basis giving
dUn '
EA-=-P
dx n
The problem stated will always reduce to establishing the relationships at a par-
ticular n (or frequency wn ), hence the subscript will usually be understood. The
condition at x = 0 now becomes (using the solution (2.5) and since there is only a
forward moving wave)
P
or A=--
iklEA
The complete solution for the forward moving wave is therefore
Knowing the spectrum of the force history, P, the displacement (and, consequently,
the stress, etc.) can be determined at any location x. Examples of the resulting
strain histories are shown in Figure 1.14 for the undamped case. As is apparent,
the disturbance travels at the now familiar constant speed of Co = J E A / p A.
The transfer function, C(x, w), for this case is
and is shown plotted in Figure 2.6 for two different x positions. We see that the
position x contributes the phase change. Note that the amplitude varies as 1/ kJ and
therefore also as 1/w. The displacement is related to force by a single integration
(that is why kJ occurs in the denominator). Its effect is to essentially amplify the
low-frequency components and may cause windowing problems as discussed in
Chapter 1. The transfer function for the velocity, on the other hand, is
and the amplitude is relatively insensitive to frequency since the damping is usually
small. Consequently, the windowing parameters used for transforming the force
will also be appropriate for reconstructing the velocity.
giving the following interesting interrelationships for the undamped case (when
k = w/co)
" Co Co .• Co Co
U·I = Ikc 0 U·I = --a'
E I = --
EAFI , U r = Ikcou r = +-ar = +-Fr
E EA
50 Chapter 2. Longitudinal Waves in Rods
Thus the history profiles of the particle velocity. force. and stress are the same.
Notice that a tensile stress moving forward causes a negative velocity (Le .• the
particles move backward). Also. note that the convention for the sign of the forces.
Fi and Fr. is that of the stress because they are internal forces.
Distributed Constraint
It can be imagined that as the wave moves down the rod. the response can be
influenced by the surrounding medium. Specifically. let there be retarding forces
(per unit length) proportional to displacement ( - K u) and the velocity (-rrit); the
latter we have already encountered as internal viscous damping but it is realized
that the same effect could arise from outside the rod. The differential equation of
motion becomes
a2 u au a2 u
EA ax 2 - Ku - 71Aat = pA ati (2.10)
d2 u 2 ~
EA- + [w pA - iW71A - K]u = 0 (2.11)
dx 2
Since the coefficients are constant. this has the same solution as Equation (2.5)
except that now the spectrum relation can be complex (even if the damping is zero)
and given by
(2.12)
The wave behavior is damped in x. Hence the generic solutions are damped pro-
portional to ~x. Generally, we will associate the complex part of the spectrum
relation with damping or dissipation. As an aside, it is advisable to add numerical
stability to all the calculations by incorporating some small damping in the system.
The range of "small" should typically be specified on the order of
Returning now to Equation (2.12), an interesting case arises when only the elastic
constraint K is present; Figure 2.7 shows a plot of the spectrum relation. This is a
rather complicated plot because we see the occurrence of a cut-offfrequency. That
is, there is a nonzero frequency for which the wavenumber k is zero. This is given
by
2PA K
W ---=0 or (2.13)
EA EA
The idea of a cut-off frequency will occur regularly in our later discussions and so
it is of value now to consider this simple case in a little more detail.
Below the cut-off frequency, the wavenumber is imaginary only and therefore the
wave attenuates; for a spectrum wave the components below We will not propagate
very far and a signal sampled at a large distance will be rich in only the high-
frequency components. This is illustrated in Figure 2.8.
The question must arise that if the rod is elastic and the constraint is elastic how
can there be dissipation? For the system as a whole (the rod and spring together)
52 Chapter 2. Longitudinal Waves in Rods
J -
x=O
x=250 mm
x=500 mm
Frequency [kHzl
" "',,,!' Time [/lSI' ... , . .,. .., .. ,. , . . , ,., , .,
O. 1000. 2000. O. 2. 4. 6. 8. 10. 12. 14. 16. 18. 20.
Figure 2.8: Reconstructions of velocity (and their frequency domain counterparts) for an
impacted rod with elastic constraint. The responses at x - 0 are scaled by 0.5.
Viscoelastic Rod
The purpose of this section is to show how waves in viscoelastic media may be
posed in terms of spectral analysis. A good review of uniaxial wave propagation
in viscoelastic rods (including some experimental results) can be found in Refer-
ence [69].
The derivation of the equation of motion follows the same procedure as before
except that the stress/strain behavior is time dependent. While there are many ways
of expressing this relationship the following form is adequate for present purposes:
Here the stress and strain are related through multiple derivatives in time. This can
now be expressed in the spectral form as
or simply
a = E(w)i, (2.15)
2.3 Dissipation in Rods 53
which resembles the linear elastic relation. The equation of motion also appears
similar to the elastic case and is
A d 2u 2 A A
(2.16)
which is obviously dispersive. A plot of this (for the standard linear solid considered
next) is shown in Figure 2.3. The nonzero imaginary component means there will
be some attenuation and the effect of this is seen in the time reconstructions of
Figure 2.9.
.8
x=O
real
.6 .!:; 50. [
!!!
x-1000 mm
Ci5 O. _ _..J
_--....JIA '-----_ _
.4
...... -_ ... _- ... -.... -....
imag
__
x=2000mm
........ _-- ..
, ////
.
.... _- ....... .
----~/
/'--...
~-----
x-3000 mm
To amplify on these results, consider the special case of the standard linear solid
as visualized by a parallel spring and dashpot in series with another spring and
described by
. a . E\E2
a + -[E\ + E 2 ] = EIE + - - E
TJ TJ
Unlike the linear elastic material, this requires three material properties, E I, E 2 ,
and TJ, to describe it. The viscoelastic modulus is written in the following forms:
A E2+iWTJ Eo + iWTJa EI -Eo
E(w) = EI . = E 1--'-----'-- a= (2.17)
E1+E2+1WTJ E\+iwTJa'
and has the very slow and very fast behavior limits of
A E\E2
E(O)~ =Eo,
E\ +E2
54 Chapter 2. Longitudinal Waves in Rods
Note that both of these limits are elastic. Consequently, the viscoelastic energy
dissipation occurs only in the middle frequency range as shown in Figure 2.9 for
the case of EI = 2E2. The dissipation is a maximum where the magnitude of E(w)
is a maximum and is given by
E2 EI +E2
w= 1
11(E 1 - Eo) 11
Reference [70] discusses this dissipation.
The phase and group speeds, in general, are given as
c= ~k =.jEA
pA '
c = dw =
g dk VPA{iiI [1 - !!.- dE]
2E dw
(2.18)
(Not that if the imaginary part of the wavenumber is very significant then the con-
cept of group speed has little meaning.) Both speeds have the same limiting values
and are shown plotted in Figure 2.3 where it is observed that the group speed has
a maximum where the damping is significant. Thus damping does not necessarily
cause the wave to "slow down" although its amplitude will decrease. The effect of
viscoelasticity on the propagation of a pulse is to decrease the amplitude (because
of the damping) and spread the pulse out (because of the spectrum of speeds).
Governing Equations
The derivation of the governing differential equation for the displacement is similar
to that as shown earlier in the chapter, except that the constitutive relation must
take into account the fact that the strain also changes because of the temperature.
That is,
F =aA = EA[e -aT] = EA [:: -aT]
EA - -aT )] =pA-
- a [(au au 2
(2.19)
ax ax at 2
2.4 Coupled Thennoelastic Waves 55
aq
- - = pCE-
aT + EaT.-
aE
ax at at 0
where To is the absolute temperature and C E is the specific heat. The final piece
of information is the Fourier law of heat conduction
q=-K-
aT
ax
where K is the thermal conductivity. The equation governing the temperature is
then
-a [aT] aT
K - =pCE-+EaT a2-
o- u (2.20)
ax ax at axat
Note that the coupling for u occurs through the constitutive relation while in the
temperature it occurs through the balance relation.
For simplicity, consider rods of uniform cross section with uniform properties,
the coupled differential equations governing the axial displacement u(x, t) and
temperature change T (x, t) are
(2.21)
Assume that both the displacement and temperature have spectral representations,
then on substitution into the differential equations get
d2 t ~
K--2 - iwpCET = iwaETo-
dii
dx dx
These are ordinary differential equations with constant coefficients and have so-
lutions of the form
ii = Ae- ikx ,
only if the following system of equations is satisfied
the eigenvalues of a matrix.) Multiplying the determinant out and imposing that it
be zero allows k to be determined from the following characteristic equation:
This can be solved as a quadratic equation in kl. Since this is a regularly occur-
ring requirement in spectral analysis, we take the opportunity now to state the
preferred (avoiding numerical round-off errors) way of doing this. Following Ref-
erence [102], consider the equation
where the sign of the square root is chosen such that the imaginary part is negative.
The two roots are then given by
q 2ao
Z]=-, Zl=-
2al q
u
But since and T are coupled according to Equation (2.22), then the amplitudes
for each mode (the corresponding barred and unbarred coefficients) are related by
- klEA
}
- pAwl
( )= ikE Aa () = Rj ( (2.24)
)
where kj for the appropriate mode "j" must be used. Consequently, there are a
total of four independent solution sets whose amplitude spectra are represented by
the coefficients A, B, C, D. We emphasize this by writing the temperature solution
as
Note that R3 and R4 are obtained by substituting -k] and -kl, respectively, into the
expression for the amplitude ratio. The coefficients are determined by the boundary
and radiation conditions. Before solving an explicit problem, we will first look in
more detail at the spectrum relation.
2.4 Coupled Thermoelastic Waves 57
kl J
pAu}
=± - -
EA '
It is clear that the second mode is highly dispersive. The first mode is obviously
that of the longitudinal stress wave in the rod, the other is associated with the
temperature change according to Fourier's Law. It is emphasized, though, that
all modes are necessary to describe the displacement response and all modes are
necessary to describe the temperature response. That is, aside from coupling of
the differential equations there can also be coupling of the initial (or boundary)
conditions, and all modes are necessary to satisfy these properly.
We can get a better appreciation of the coupling effect by considering an approx-
imate version of the spectrum relations. First re-write the characteristic equation
as
[Kki + iwpCEl
or more compactly
k2 = ±
J
-iwpCE(l + f3y)
K
where
1
y= ------~---------
[1 +pAw2 K/iwpC EEA]
The coefficient y is frequency dependent and ranges from 1 to O. We could replace
it with 1.
It turns out that the spectrum relations are very insensitive to the coupling pa-
rameter f3 -- changes of many orders of magnitude are necessary before an effect
is noticeable.
The mode 1 behavior has a negligible imaginary component and a nearly linear
relation between frequency and the real part. Consequently, no attenuation (in
space) is expected. Further, since the group speed is nearly constant, then this
58 Chapter 2. Longitudinal Waves in Rods
portion of the response will propagate unchanged in shape. The mode 2 behavior is
quite different in each of these regards. First, there is a sizable imaginary component
indicating attenuation. Second, since the relation is nonlinear in frequency, that
portion that is propagated will change its shape. Finally, in comparison to the first
mode, the magnitudes of the respective wavenumbers are vastly different, indeed,
nearly in a ratio of 100 to 1. Consequently, propagation speeds are exceedingly
slow and attenuation will be expected to be much more significant.
In summary, it can be said that irrespective of the particular initial boundary
value problem, the mode 2 contribution to the displacement and temperature will
be highly localized in space, whereas the other mode will send out a propagat-
ing nondispersive component. The relative magnitudes involved depend on the
geometry and the particular imposed conditions.
aT au(o, t) ]
K-(O, t) = q(t), a = E[E - aT] = E [ ax - aT(O, t) =0
ax
In terms of Equation (2.23), these become
where only that portion of the solution satisfying the radiation condition in x > 0
is considered. Solving for the coefficients and using Equation (2.24) gives
-q
B = -[aR J +ikd
Kll.
where
2.4 Coupled Thennoelastic Waves 59
(2.25)
The strain and temperature histories can now be reconstructed and some plots are
shown in Figure 2.10.
-
x=o.o
I\"
x=0.1 mm
2.
O.
I\,
x=0.2 mm
-2. 200.
x=O.O
I\,
x=0.3 mm
O. i~ x=0.03 mm
Time [IlS] I I I I I I I I I I I I I I I I I I I I I
.0 .5 1.0 1.5 O. 1. 2. 3. 4.
Figure 2.10: Strain (units of I.U) and temperature (units ofOC) responses. The x = 0 strain
response is scaled by 10- 4 •
These example plots show the behavior to be as expected: the strain signal is
propagating unattenuated at a constant speed, while the temperature variation is
highly localized to x = O. What might be surprising is the very small magnitude
of strain considering that at the end it is aT(O, t), which is on the order of 1000
J-LE. The explanation can be seen from the transfer function associated with the
strain where it is noted that at x = 0, G = a, but for nonzero x the second mode
attenuates leaving
That is, since k2 is so much larger than kJ, the overall amplitude diminishes con-
siderably.
Time reconstructions of the temperature histories at different locations show that
beyond a very short distance from the source there is no temperature change. They
also show that the propagating component of temperature could not be detected
because the coefficient for the mode 1 is nearly zero.
60 Chapter 2. Longitudinal Waves in Rods
L
(d) L-I~_>-_ _ _ _----'I =} FJ {=C=J =} F2 {= L-I_ _ _ _----'
More complicated problems, such as waves in curved rods [25,62], and helical
springs [58, 122], and waves in noncollinear rods [2, 33, 76], could also be treated,
but these generally create new wave modes (such as flexural waves) and so will be
left to Chapters 3 and 5.
where A is associated with the known incident wave and B with the unknown
reflected wave. This is an important point for later analyses: Since the differential
equations establish all the possible solutions, then the number of choices is auto-
matically prescribed - there is no possibility of "leaving something out." (This,
of course, assumes that the choice of structural model is adequate for the problem.)
The boundary conditions are set in tenns of
Force: EA-
au
ax
2.5 Reflections and Transmissions 61
A typical set of boundaries along with their equations is shown in Table 2.2. Note
how easily the time domain conditions are converted into conditions on the spectral
components U.
Free:
ilu(O, t)
EA--=O EA du(O) =0
ilx dx
ilu(O, t) reO)
Spring: E A - - = -Ku(O,t) EA_u_ = -K u(O)
ilx dx
ilu(O, t) ilu(O, t) du(O) •
Dashpot: EA--=-TJ-- EA-- = -1]iwu(O)
ilx ilt dx
ilu(O, t) il 2 u(O, t) du(O) •
Mass: EA---=-m
ax at 2
EA--
dx
= +m w2 u(O)
Assume that at the end of the rod there is an elastic spring as shown in Fig-
ure 2.11 (a). The resistive force is proportional to displacement, hence at the bound-
ary x = 0, we have F(t) = -K u(O, t) or
(2.26)
This simply gives B = A, which says that the reflected displacement pulse is the
same as the incident. By differentiating, it can be seen that the reflected stress pulse
is inverted. If there is no spring, i.e., K = 0, then
B=A, Ur = Ui,
62 Chapter 2. Longitudinal Waves in Rods
1\ 1\ small stiffness
/\ 1\
~
~/~
V large stiffness
O. 500.
, , , V
1000.
, , ,
1500.
, , Time [J.l.Sl
!
2000.
, , , ,
2500.
Figure 2.12: Velocity responses for a rod attached to springs of various stiffnesses. The
monitoring site is also the impact site.
which, of course, is the free-end condition. On the other hand, if the spring is very
stiff, i.e., K = 00, then
B=-A, U r = -Uj,
This gives the fixed-end condition. The stress is inverted in the first case and the
displacement in the other. These limiting results are independent of frequency and
hence the reflected pulse will retain the same shape. This is shown as the two
extreme cases of Figure 2.12.
For a given spring stiffness, the limiting behavior on the frequency gives rise
to the same limits as above. That is, the high frequencies experience a free end
and the low frequencies experience a rigid end. Consequently, for a pulse with a
spectrum of frequencies, the reflected signal is distorted as shown for the spring
of medium stiffness. (Parenthetically, medium in this situation means relative to
the amplitude spectrum of the wave.)
It is clear that for the medium stiffness spring there is a phase shift in the
occurrence of the maximum velocity. This is a typical example of how a wave
propagating in a nominally non dispersive medium behaves dispersively when in-
teracting with discontinuities. To amplify on this, consider the relation between
the coefficients to be a transfer function. That is, B = G(w)A and Figure 2.13(a)
shows the plot of G(w) against frequency. What is striking is that the peak in the
imaginary part while the real part goes through zero. This is almost like resonance
behavior; however, unlike resonance, the amplitude always remains unity.
2.5 Reflections and Transmissions 63
where U o is the displacement of the rod and U c is that of the mass. To these we add
the relations obtained from the wave behavior in the rod
F = EA{-ikA+ikB}
giving the reflected wave amplitude as
ikE A + m cw 2 Ct A A
(2.27)
A
What is most interesting about the relation is the presence of the single degree of
freedom oscillator in the denominator of Ct. If there is no dashpot this could become
zero at the natural frequency and therefore Ct would dominate the response. In fact,
at this frequency G = -1 indicating the behavior of a fixed end. What is striking
is that it has the form of a damped resonance, Le., a peak in one component while
the other component goes through zero. This is clearly seen in Figure 2.13(b).
!.~
/)"1
,
Figure 2.13: Transfer function for different end conditions. (a) Spring, (b) oscillator.
To explain these results, consider some of the limiting cases. If there is no mass,
Le., mc = 0, then
B=A, U r = Ui, a r = -ai
64 Chapter 2. Longitudinal Waves in Rods
which, of course, is the free end condition because the spring has nothing to react
against. On the other hand, if the mass is very large, i.e., me = 00, and the stiffness
very large then
B=-A, u, = -Ui, a r = ai
This gives the fixed end condition. The stress is inverted.in the first case and the
displacement in the other. These limiting results are independent of frequency and
hence the pulse will retain the same shape.
An interesting case is when K = 0, me = 00 and we take TJ = EAkjw = copA
then we get that B = 0, that is, there are no reflections. This is one of the few cases
where a single dashpot can absorb the complete wave spectrum.
(2.28)
UJ
where the subscript J refers to the joint. These become (since there is only a
forward moving wave in rod 2 and both rod segments are similar)
These relationships are complex and frequency dependent; thus distortion of the
wave can be expected. This is exhibited in the responses shown in Figure 2.14.
2.5 Reflections and Transmissions 65
/\ small mass
~ '------
-~--
~__~~__________~Ia~rg~e~m~a~s~s~
, , , , , ! , ,
Time [JlS]
, , • ! I
Figure 2.14: Responses for a rod with various concentrated masses. The monitoring site is
also the impact site.
It is clear that the mass acts as a frequency filter. For the low frequencies, for
example,
showing that the mass acts as a rigid end and does not transmit any of the wave.
Notice that the same result is effected by increasing the size of the mass.
The case of transmission of waves from one rod to another is handled the same
way as above. Balance of force at the (assumed massless) joint of Figure 2.11 (c)
and continuity of displacement gives
Solving for the reflected and the transmitted coefficients then gives
(2.30)
statically, the mass distribution of the rod makes a profound difference dynamically.
The expressions for the stresses are simply
(2.31)
Reference [107] gives some experimental validation of these stepped rod results.
(2.32)
The incident wave is Al and there is only one wave in the third rod. Note that
the rod forming the joint will have a standing wave since both forward and back-
ward moving waves will be present simultaneously in the finite length. This is
fundamentally different from the two waves in the semi-infinite first rod as will be
seen.
The boundary conditions at the joint interfaces, x = 0 and L, are
atx=O:
atx = L: (2.33)
1 1
(kEAh -(kEAh
e-ik2L eik2L
-(kEAhe- ik2L (kEAhe ik2L
While this can be solved explicitly, it is best left as part of the algorithm for solving
for the transfer function of the wave. If all the bars are of the same material, then
this relation simplifies considerably, but leaving it in its present form allows, for
example, a solution for the viscoelastic joint.
Figure 2.15 shows some of the responses on both sides of the insert for an incident
tensile triangular pulse. It is apparent that a pulse is propagating backward and
forward in the joint. Furthermore, each reflected and transmitted pulse is similar
2.5 Reflections and Transmissions 67
incident side
reflected side
, Time [J.1S] ,
o. 500. 1000. 1500. 2000. 2500.
Figure 2.15: Transmitted and reflected responses near a distributed elastic joint. The mon-
itoring site is also the impact site.
Figure 1.16: Transfer function for the responses near an elastic joint. (a) Reflected side. (b)
transmitted side.
68 Chapter 2. Longitudinal Waves in ROds
In this, the shape of the distribution, f(x), does not change over time - it is only
the scaling, pet), that changes. If the area under f(x) is unity then pet) can be
interpreted as the resultant force. Now represent the shape with the Fourier series
2nm
w
f(x) = 1 '""' -< X
L fme-r,m , ~m=-
W
m
where W is the space window. Note that since f(x) is real-only, we could use
the usual form of the Fourier series; however, the exponential form is easier to
manipulate. We now let the response have a similar representation so that the
2.6 Distributed Loading 69
- -plm
Um =
[;~EA - (w 2 pA - iW'1A)]
A( ) = - -p f,m .,
e -I,m x
ux
~ ;~EA - (2.35)
A [ _ ]
W (w 2 pA - iW'1A)
This can then be incorporated in the usual FFT inverse. Generally, however, it is
more efficient to compute the summation over m directly rather than use the FFT
algorithm because the response at only a few locations, x, are usually required.
Lx
q(x, t) q(x, t)
I~~ I I~~ I
II II
+- -00 t----+ X +00 -+ +- -00 t----+ X +00 -+
Figure 2.17: Two modelings of a rod with distributed load over the region 0 :::: x :::: L.
The remaining specification for the problem is the actual load distribution shape.
For simplicity, let us take it as being rectangular over a length L as shown in
Figure 2.17. This gives
Responses are shown in Figure 2.18 where the history is that of our smoothed pulse
and the parameters M = 1200, W = 150 m (6000 in.) were used. A feature to note
about the force responses is that at x = L /2, which is the middle of the distribution,
the axial force is zero. This-can be explained by considering material away from
the distributed load site: to the left, the material is being pulled in tension, to the
right, the material is being pushed in compression. Note, however, that the velocity
of all points is positive.
-waveguide
o transform
velocity
Time (J1s1
. I ..
Figure 2.18: Resultant axial force and velocity response for points under the distributed
load.
(2.36)
Suppose now the distributed load is a delta function so that j = 1 then we know
that the value of the integral is simply the sum of the residues [7]. The contours
of the integrand are shown in Figure 2.19, also shown is the integration path used
for the Cauchy integral.
real wavenumber
1.0
.8
.6
.4
.2
.0
·.2
-.4
-.6
-.8
-1.0
Figure 2.19: Contours of the determinant. On the left is a slice of Re(~) against frequency
for Im(~) = 0, and on the right is Re(~) against Im(~) at 15 kHz.
From the contours of Figure 2.19, we see there are two poles corresponding to
where the denominator is zero. These occur at
72 Chapter 2. Longitudinal Waves in Rods
which, in fact, are the waveguide spectrum relations. Furthermore, the residues at
each of these poles is proportional to
which are the kernel function solutions for the waveguide. There is an intimate
connection between what we will call the waveguide solution and integral solutions
of the form of Equation (2.36). The significant difference lies in the size of the
domains; as pointed out above the waveguide domain can be finite thus making it
useful for structural analysis.
We will not have any need for contour integration in the complex plane because,
generally, we will either do the integration (summations) numerically or replace
the integrals with their waveguide form. The matrix methodology for handling
connected waveguides is developed in Chapters 5 and 8.
Problems
2.1 Show that if the phase term in the generic solution can be written as
w
±kx + wt = -(±x + cot) = w~ or W1/
Co
is recovered.
- Reference [7], p. 45
2.2 Show that longitudinal disturbances in a gas within a long cylinder (acoustic
waves) are governed by an equation very similar to that of the rod.
- Reference [80], p. 177
2.3 Show that transverse disturbances propagation along a string under tension
are governed by an equation very similar to that of the rod.
- Reference [48], p. 42
2.4 Assume the area of the rod varies according to the following power law:
A(x) = Au
a+ x]m
[-a-
where An, a, m are constants. Show that the complete solution is given by
2.5 Continuing the previous problem, for the special case of m = 1, show that
This is similar in form to that for the uniform rod except for the decreasing
amplitude.
- Reference [56], p. 111
2.6 For the conical hom (or linear taper) m = 2, and noting the relation between
the Bessel functions of half-order and the trigonometric functions, show that
W/W e = 2 sin(ka/2),
Sketch the spectrum relation and discuss the ability of the string to support
waves of small wavelength. What happens if W > We ? What is the meaning
of the complex roots for w < We ?
- Reference [80], p. 223
2.9 A periodic layered structure is comprised of alternating layers of thickness
do and dh. Show that the characteristic equation is
A beam is a long slender member designed to support lateral loads. In doing so,
the displacement is predominantly transverse to the centerline, and internal shear
forces and bending moments are generated. The dynamic behavior of beams is
called flexural motion.
From the wave point of view the significant difference between a beam and a rod
is that the beam (even in its simplest form) does not have aD' Alembert solution
and therefore the solutions are dispersive (as shown in Figure 3.1) from the outset.
Also, because of the higher order governing equations, there are two fundamental
modes of motion, one of which propagates while the other is evanescent. The devel-
opment in this chapter is similar to that of the last: after deriving the basic solution,
its application to reflections and elastically constrained systems is demonstrated.
Since the analytical framework is the same as for rods, these applications are not
described in as much detail as before. Instead, the chapter concentrates on two new
topics which deal with coupling: the first is that of curved beams and the other is
of beams with distributed spring connections.
Time Position
Figure 3.1: Dispersion of waves in beams. Each line is a "snapshot" of the defonned shape
of the beam. Note how the shape changes over time.
74
3.1 Bernoulli-Euler Beam Theory 75
Equations of Motion
Consider along, narrow beam with the loads applied as in Figure 3.2. The Bemoulli-
Euler beam model assumes that the deflection of the centerline v(x, t) is small and
only transverse. While this theory assumes the presence of a transverse shear force,
it neglects any shear deformation due to it.
..
q(x) - TJV
.0 •
EI,pA
• _ ......... "';~~
. - / / / v(x.t)
....: : : : : : : : • • • • • • • • • • • • • • • 0 . 0 • • • • • • • • 0 • • 0 . . . . . . . . . ..
C~UtsaM
V ax v +a v
Figure 3.2: Slender beam in flexure and typical loaded element.
u(x, y) ~ au
u-(x, 0) + y oy Iy-o + ... = u(x) - y4J(x) + ...
_ oii
ii(x, y) ~ v(x, 0) + y oy Iy-o + ... = v(x) + yl/r(x) + ...
!
as
a2v
M = -au ydA = E I -2
ax ,
The cross sectional property 1 is called the second moment of area, and the com-
bination coefficient E 1 is called the flexural stiffness. For a beam of rectangular
cross section, we have 1 = bh 3 /12, where b is the width and h the height.
We consider two types of distributed load: qv(x, t) is the transverse load, while
q", is the distributed torque. Referring to Figure 3.2, the equations of motion in the
y direction and about the z axis are (if ~X is small), respectively,
av a2 v av
ax pAarz + '1Aat - qv
aM
-+V = -q", (3.2)
ax
where we have neglected any rotational inertia. Note that, in contrast to the rod,
the beam has two equilibrium equations. Substituting for the moment into the
equations of motion gives
-
a2 [a 2 v] a2 v au aq",
E I - +pA-+'1A-=q(x,t)=qv-- (3.3)
ax 2 ax 2 at 2 at ax
It can be easily verified that there is no D' Alembert solution [Equation (2.3)] to
the homogeneous part of this equation even if the cross sectional properties are
constant.
It is worth emphasizing that the moment resultant is related to the deformation
through the constitutive relation, whereas the shear force resultant is related to the
deformation through the eqUilibrium equation. We can rationalize this inconsis-
tency by conceiving of the beam as having a very large shear modulus.
3.1 Bernoulli-Euler Beam Theory 77
Spectral Analysis
Consider the beam to have constant properties along its length, then the homoge-
neous differential equation can be written in the spectral form
(3.4)
Particular solutions to this can be obtained from particular solutions of the follow-
ing two equations:
(3.5)
This form emphasizes that the beam has two fundamentally different modes -
this will become apparent later.
Since the equations have constant coefficients, then the solutions are obviously
of the exponential form e- ikx , which on substituting into the equations gives
k2 = ±if3
The spectrum relations for both modes are plotted in Figure 3.3 as the heavy lines;
the other plots show the effect of axial loads which we consider later. There are
four possibilities in all, allowing the complete solution to be written as
The first and third are wave solutions, while the second and fourth are predom-
inantly spatially damped vibrations (sometimes called the ringing or evanescent
terms).
1.2 Wavenumber, kh
1.0
.8
.6
.4
.2 - - mode 1 (real only)
.0 ........ mode 2 (imag only)
'~:~~~~~~:::::~~~i~:::<:::::::"":C:.:":.:.:2:
-.2
-.4
-.6
-.8
-1.0
-1.2 I ! , , t i t , , , , ! , , ! , , ! , ! I , , I , I
The mode 2 behavior is entirely imaginary for the undamped case, hence there is
no propagation behavior for this mode. Considering only the mode 1 wave motion
gives the phase and group speeds as
cg == -
dw
-2.jW -
[EI]1/4 .. 2c (3.7)
dk pA
Both of these are shown plotted in Figure 3.4 as the heavy lines. In contrast to the
elementary rod case, these speeds depend on the sectional properties of the beam
in terms of the area A = bh and second moment of area I = bh 3 /12. Note that
the group speed is twice that of the phase speed. Both speeds are also dispersive
and indicate speeds approaching infinity for very high frequencies. This is an
unreasonable limit, as we will see in the next chapter, but for now we can accept
that the derived beam theory is adequate at least for the lower frequencies .
Substituting for these quantities in terms of the displacement solution gives the
wave in x > 0, for example, as
(3.8)
(3.9)
The behavior of the strain at various positions along the beam is shown in Fig-
ure 3.6. The response is obviously dispersive. Note particularly that because the
low-frequency components travel the slowest, they will take the whole time win-
dow to arrive at the monitoring site. As a result, an extra-large window is usually
required when analyzing beams. In fact, the window used for the present analysis
is twice that shown. Since this is a recurring problem with beams, it is of value to
consider it in more detail.
.~ 1~~·E
______0-
!\
~~ 5:. :0~o~m: .:.m:- __
'V ~
A
Ci5 -100.
, , , , , I ,
o. 500.
Figure 3.6: Typical strain responses of an impacted beam.
and this is different from the rod in two significant respects. First, the presence
of e- flx indicates a large contribution at x = 0 that dies rapidly as x increases -
this is called evanescent behavior. Second, the transfer function depends on the
reciprocal of f33 and consequently also on w 312 • This fractional power dependence
on frequency does not really pose a problem in the spectral analysis approach per
se, although it is what makes the analysis in the time domain difficult. For example,
the relation between the force and strain at the impact site is
i(O w)
h
- ( 1 - i)-
P= h [EI]I/4 (l - i)-
P (3.10)
,
=
8EI f3
-
8EI
-
pA JW
The convolution theorem of the first chapter [Equation (1.8)] can be used to convert
this to time domain form as
1'(0 t)
,
= -
h
4E 1
[EI]I/4
-
pA
1
--
.j2ii
11 ..;t="T
0
P(r)
---dr (3.11)
This interesting equation (which also appears in Reference [114]) shows the strain
(and velocity) to be a singular hereditary integral of the force. Consequently, the
response will not be quiescent (even for a finite duration pulse) until the time
approaches infinity. This is where the fractional powers cause some windowing
problems - no matter what size the window is, some response will always persist
beyond it and thus move into the neighboring window. The negative effects of this
can be minimized by choosing a relatively large window size but this is at the cost
of more computation time. A cheaper approach is to include some damping in the
system so that all signals eventually die out. Actually, a judicious combination of
both will give the optimum results.
To aid in the analysis, assume that the wave propagates symmetrically outward,
then at x = 0:
aV
ax = 0, Vi = Vr
(3.12)
3.3 Bernoulli-Euler Beam with Constraints 81
The tenns in square brackets are the transfer functions for the waves and they
indicate a changing amplitude with respect to position. For the forward disturbance,
the other mechanical quantities of usual interest are:
Velocity:
Strain:
Stress:
When there is no damping, we have that 13 2 = w..j pAj E I, showing that the
velocity and stress and strain at y = hj2 are related through
2 [EI 2 [EI
h VPA Ei(O, t) = VPA ai(O, t) = .j3 Ei(O, t)
Co
Vi(O, t) = Eh (3.13)
That is, the particle velocity has a simple relationship to the maximum bending
strain and stress. This, of course, is not true at any other position and is not true
when damping is present.
is
(3.14)
Since the coefficients are constant, then the spectrum relation can be obtained
simply by taking solutions of the form v(x, t) = voe-i(kx-wl) to get
Elk4 - pAw2 + iWlJA + K = 0
giving the wavenumbers (in the undamped case) as
pA 2 K
-(J) - - =0 or (3.16)
EI EI
for both modes. The spectrum relations for the two modes can now be written as
(3.17)
3.3 Bernoulli-Euler Beam with Constraints 83
For frequencies greater than We there are two purely real and two purely imaginary
roots and these asymptote to
A]1/4 [PA]1/4
kl ~ ±.jW [~I ' k2 ~ ±i.jW EI
which are the usual spectrum relations for beams. For frequencies less than We
1 +1. [ P A 2
kl =±-- --(W -W)
2 ]1/4
../2 EI e '
which shows that all roots are complex. Thus all components associated with both
modes attenuate to some extent. Reconstructions are similar to Figure 2.8 in that
filtering of the lower frequency components occur.
Since all modes are complex below the cut-off frequency, the question then
arises as to which of the available solutions should be chosen to represent a for-
ward moving wave, say. The appropriate criterion is to choose those solutions that
decrease amplitudes (exhibit dissipation) in the direction of propagation; that is,
choose the wavenumbers that have negative imaginary parts. As a consequence
of this choice, there is a standing wave set up at frequencies less than We. That
is, both forward and backward moving waves are established simultaneously; the
concept of speed has no practical meaning in this case although we could formally
compute wi kreal and dwldkreal for each mode. Note that W = We is a branch point
and it is difficult to determine the continuation of each mode. This ambiguity is
removed by adding a little damping as discussed previously.
- - dv
Fo sintf>Ax ~ Fo-Ax
dx
(3.18)
Since the coefficients are constant, then the spectrum relation can again be obtained
by taking solutions of the form v(x, t) = voe-i(kx-wt) to get
84 Chapter 3. Flexural Waves in Beams
(3.19)
The sign of the pre-stress can make a significant difference to the character of the
solution as shown in Figure 3.3. For example, a compressive pre-stress, Fo = - p,
gives the mode I a nonzero wavenumber value of kJ = "j P / E I at zero frequency.
The interpretation of this diagram is that for the propagating mode the wavelengths
in the compressed beam are decreased (the wavenumber is increased), while those
in the stretched beam are extended compared to the unstressed beam. The effect
of this on the propagating speeds is shown in Figure 3.4. Notice that the group
speed is increased substantially in the compressed beam. Thus very low frequency
components will propagate very fast.
Although slightly premature, consider the consequences of this increase in speed
in a finite beam. The low frequency components take very little time to reflect,
indeed in any time duration many reflections will occur and hence resonance is
established. In the limit of zero frequency this occurs in negligible time and the
situation corresponds to static instability or buckling.
The first two terms are the incident wave and A, B are assumed specified. The
remaining terms are the generated reflections. The last term (De+ ik2X ), while not
a wave, is nonetheless necessary to satisfy the boundary conditions even if only
a propagating term is incident on the boundary. As emphasized before, the above
is the general solution to the governing differential equation and as such there
cannot be any others waves. That is, these are the totality of possible responses for
the given structural model and properly stated boundary conditions should then
determine the appropriateness of each.
3.4 Reflection of Flexural Waves 85
The boundary conditions are specified in terms of the different orders of deriva-
tive of v(x, w). These are collected as
Displacement: v=v(x,w)
A dv
Slope: ¢= dx
Moment:
d2
A
M=+EI-
v
dx 2
Shear Force:
A d3
V=-EI-
v
dx 3
Table 3.1 is a collection of some typical simple boundaries and the associated equa-
tions. Note that in each case two conditions must be specified in the time domain
and these become conditions on the various space derivatives in the frequency
domain.
av(O, t) = 0
Fixed: v(O,t)-O,
ax
a 2 v(O, t)
Pinned: v(O, t) = 0, =0
ax 2
a2 v(O, t) _ 0 El a3v (O,t) =0
Free: EI 2 - ,
ax ax 3
a2 v(O, t) _ 0 a3v(O, t)
Linear Spring: EI 2 -, EI 3 =+Kv(O,t)
ax Ox
a 3 v(O, t) av(O, t)
Torsion Spring: v(O, t) = 0, EI =-a---
ax 3 ax
EI a 2 v(O, t) _ 0 E a 3 v(O, t) av(O, t)
Oashpot: I =+1]---
ax 2 ' ax 3 at
a 2 v(O, t) _ 0 a 3 v(O, t) a2 v(O, t)
Mass: EI 2 - , EI 3 = +m 2
ax ax at
The simplest boundary condition is that of a pinned end. Here the deflection is
zero giving as one of the equations
A+B+C+D=O
There is also zero moment giving as the second equation
EI[A(-ikt>2 +B(-ikd + C(ikd +D(ikdJ = 0
These two equations allow C and D to be determined, they are simply
C=-A, D=-B (3.20)
86 Chapter 3. Flexural Waves in Beams
Thus, the reflected wave is the same as the incident wave but inverted; this is
analogous to what occurs for the rod. Note that there is no spatially decaying term
generated if none is incident.
An example of a more complicated boundary is that of a free end. It is necessary
to impose both zero moment, as above, and also zero shear, giving
EI[A( -ik t )2 + B(-ik2)2 + C(ik t )2 + D(ikd] 0
-EI[A( -ikt)3 + B( -ik2)3 + C(ik t )3 + D(ik2)3] 0
total
reflected
propagated
, , . , , Ti,me (J;lSJ,
o. 1000. 2000. 3000. 4000. 5000.
Figure 3.8 shows an example of the response of a beam impacted near a free
end. The impact site is also the monitoring site and the boundary is located at
L = 500 mm. Parenthetically, a side benefit of the spectral approach to waves is
that the decomposition of the response is a natural by-product of the analysis. That
is, the reflected wave can be reconstructed on its own, as shown, if required. Also
shown in the figure is a wave propagating a distance 2L in an infinite beam. The
comparison with the reflection alone clearly shows the phase shift.
Assume that at the end of the beam there are two elastic springs, one resisting
the deflection and the other resisting the rotation. The shear and moment balance
3.4 Reflection of Flexural Waves 87
equations require
d 3AV A d2v
E I - =-eL-
dv
E I - =+Kv
dx 3 dx 2 dx
giving, with K == KI EI, a == eLI EI,
[ (iki + K) (iki + K) ] {C} _ [ (iki - K) (iki - K) ]{A}
(ikl+a)kl (ik2+a)k2 D - (-ikl+a)kl (-ik2+a)k2 B
The coefficients depend on frequency, even in the simple beam case. As a conse-
quence, the response will exhibit the same type of dispersion effect as observed
in the rod case. That is, for given springs, the low frequencies experience a fixed
end, whereas the high frequencies experience a free end. The special cases of free
end (K = 0, eL = 0) and fixed end (K = 00, eL = (0) are easily recovered from the
above.
where the subscripts refer to the beams. It is assumed that only a propagating wave
is incident on the joint and its amplitude spectrum is represented by the A term.
This is reasonable since in most cases the generated evanescent term decays to
negligible size. Also retaining this evanescent term is cumbersome. The first beam
contains a reflected wave C and a ringing term D, and the second beam contains
corresponding propagating and ringing terms. As pointed out before, this is the
maximum number of possibilities - the boundary conditions should determine if
any are redundant.
Consider two beams of similar sectional properties connected by a rigid con-
centrated mass as shown in Figure 3.5. Continuity and balance of resultants at the
concentrated mass requires
VI = V2 VJ
aVI aV2
-=- ¢J
ax ax
-MI +M2 0
-VI + V2 mJvJ
88 Chapter 3. Flexural Waves in Beams
where the subscript J refers to the joint properties. This gives with m* == m, / E I f33
and assuming both beam segments to be the same
-1 -1
[J
I
(3.23)
1 -1
-l -i +m*ul 1 + m*w2
The presence of the mass makes the equation frequency dependent and thus the
mass (as in the case of rods) acts as a frequency filter. Low-frequency components
transmit through the joint unaffected, the high-frequency components are blocked.
The equations of motion of the three member joint in Figure 3.9 are
- FI + F2 cos e - V2 sin e mu j
- VI + F2 sin e + V2 cos e
-MI + M2 + ~L(VI + V2) (3.24)
where L is a size estimate of the joint. Displacement continuity for small joint
rotations are
UI Uj
U2 Uj cose + Vj sine
VI Vj - ~L¢,
(3.25)
3.4 Reflection of Flexural Waves 89
(3.26)
The axial force, shear force and bending moment, respectively, are related to the
displacements by
F=EA-
au
ax'
Finally, if the flexural displacements are expanded as Equation (3.2) and the lon-
gitudinal displacements as
II
where kLn is the longitudinal wavenumber, then a system of six equations can be
established so as to determine the six wave coefficients. Such a system of equations
is frequency dependent. This comes about not only from the mass but also from
the presence of mixed wavenumber terms. For example, both V and F occur in
the same equation, and since their wavenumbers have different dependencies on
frequencies, then the frequency does not cancel through.
As a first special case, consider when both beam segments are similar but just
oriented differently. Also, let the size of the joint L be considered negligible, then
the system of equations becomes
where K == k 3 EI / kLEA. The presence ofthe axial terms means there are longitu-
dinal waves generated in each beam section. When a longitudinal wave is incident,
then add to the right-hand side the vector
Figure 3.10 shows the variation of the transmitted flexural and longitudinal
amplitude with frequency when a longitudinal wave is incident. Similar results
are obtained for an incident flexural wave. These plots show that the joint has an
obvious frequency effect. If a curved rod is considered as a collection of angled
straight segments, then similar results are expected there also. It must be borne
in mind in both of these examples that beyond 100 kHz or so the Bernoulli-Euler
90 Chapter 3. Flexural Waves in Beams
2.0
Amplitude 100 Hz 10 kHz
1.8
-c -c
1.6 ···.·.D ........ D
1.4 -A -A
······B ........ B
1.2 -Q -Q
..... p ........ p
1.0
.8
.6
.4
.2
.0
O. 30. 60. 90. 120. 150. Angle 30. 60. 90. 120. 150. 180.
Figure 3.10: The effect of frequency on the reflected and transmitted amplitude ratios.
beam theory is no longer valid and one of the improved beam theories should be
used.
As a second special case, consider beams that have discontinuity of stiffness
and density, but remain collinear. This is the analog of the stepped rod, and the
free body diagram is similar to that shown in Figure 3.5. The joint is assumed to
be rigid, of very small mass and extent. A more complete exposition can be found
in References [44, 45].
Continuity conditions at the joint are the same as for the concentrated mass ex-
cept that the joint mass is considered zero. In terms of the displacement coefficients
l
these conditions become (with x = 0 being the location of the joint)
i~ R
(3.28)
-1 1
-i
where the coefficient ratios R == [E2/z/Elh]I/4 and Q == [P2A2/PIAdl/4 have
been introduced. Generally, it is best to solve this system numerically as part of the
general scheme for obtaining the transfer function. The explicit solution is given
here because it can be used to help verify the implicit solution schemes used on
some of the more general cases. Solving in terms of A gives
C -[(1- R 2 Q 2i+i2RQ(R 2 - Q2)]iA/~
D (l - R2Q2)(1 + R2Q2)[1 - i] A/ ~
A (l + R2Q2)(R + Q)2A/Q~
B -(1- R2Q2)(R+iQ)2A/Q~
(3.29)
3.5 Curved Beams and Rings 91
displacements by
_ ur 1 aUg
Ee(}=-+--
r r ae
1 aUr aUe UI}
--+---
r ae ar r
We will now replace these with an approximate set based on the assumption that
the dimension in the r direction is small.
Consider a small segment of beam as shown in Figure 3.II(b). We begin by
expanding the displacements in a Taylor series about the mid-plane (r = R) using
the variable I; == r - R. That is,
where ljIe is a rotation of the subscripted face in the direction of the curvature.
These approximations say that the deformation is governed by three independent
functions that are functions of the space variable e and not r. Derivatives with
respect to r are replaced with derivatives with respect to I; and evaluated at r = R!.
It is understood that all variables are also functions of time, but we make the
assumption that the above kinematic representations do not change under dynamic
conditions.
The shear strain corresponding to the above deformations is
_ 1 aUr Uo I;
2Ere ~ Rae -ljIe - Ii + RljIe
Because the beam is slender, we now make the usual slender beam approximation
that the transverse shear strains are negligible on average. This leads to
1 aUr Ue
ljIe ~ Rae - Ii
_
ue(r e) ~ UIJ(e) -I; ( -
I aUr
- - UIJ)
-
, R ae R
3.5 Curved Beams and Rings 93
Eee
This derivation has paralleled that of the straight beam derived earlier in the chapter;
essentially the same assumptions and approximations were used.
At this stage, it is worth our while to convert the above to a more usual form of
notation. It is typical in curved beam analysis to have a hoop coordinate s and a
transverse coordinated y pointed toward the origin of the circle. That is, we have
Other curved beam theories have slightly different expressions for this strain; the
present theory is closest to that of the Bernoulli-Euler straight beam. The most
significant aspect of this strain-displacement relation is the nonzero centroidal
strain (at y = 0) even if v is the only deflection. This will give rise to the coupling
of the longitudinal and flexural waves.
Equations of Motion
Under the assumptions as developed above, the beam is in a state of uniaxial stress.
That is, the only nonzero stress is the axial (hoop) stress given by
a = EE =
ss ss as _ ~]
E [au R
_ yE (aas2 v2 + ~R au)
as
This has a linear variation on the cross section just as for the straight Bernoulli-
Euler beam. This gives rise to two resultants: the normal force F, and the bending
moment M, given as
F == fa ss as - ~]
dA = EA [au R' M==- f a
ss
a [av u]
ydA=EI-
as -as+ -R
(3.30)
94 Chapter 3. Flexural Waves in Beams
where the integration is over the cross sectional area A. To these we append the
resultant shear force V.
Let the beam be slender with a fairly large radius of curvature and consider a
small segment under the action of these resultants. The equations of motion are
easily established as
av F a2 v
-+-=pA-
aM
-+V=O (3.31)
as R at 2 ' as
These relations are valid for arbitrary plane curved beams; only if R is constant
are they restricted to circular geometries. Obviously if R becomes very large the
straight beam results (Equations (3.3» are recovered. Substituting for the resultants
in terms of the deformations gives the coupled set of equations (for R = constant)
Spectrum Relation
The spectral form of the governing equations is
1 d 2u 2A 1 [ dv d 3 V] o
( EA+-EI)-+PAW u+- -EA-+EI-
R2 ds 2 R ds ds 3
6 [2k +R22J4
k - o
- k - [421
13 + k IJ2k + [2k - R2-IJ4
R2 - -R4
- 0
13 =0 0
This has six solutions in all, but since only k 2 terms appear, then there are three
basic modes: one associated with the longitudinal behavior and two associated
with the flexural. This can be seen by noting that for very large R the above can
be factored into
(k 2 - f32)(k 2 + f32)(k 2 - k~) = 0
In general, of course, the modes are coupled and it is not proper to speak of a
longitudinal mode or a flexural mode.
Before we solve for the spectrum relations, it is beneficial to check certain
features. First we see if there is a cut-off frequency; set k = 0 in the characteristic
equation to get
1 E A - pAw
[ R2 2J pAw2= 0,
The presence of a cut-off frequency is typical of elastically coupled systems. What
is interesting is that this cut-off is independent of section properties but depends
on the radius. Now look at when the frequency is zero, the characteristic equation
can be factored as
Only one root goes through zero, the others are a double root on the real axis.
Finally, look at what happens as the frequency becomes very large. If we assume
that EA » E 1/ R2 then we get
and
-q
S2 = -[r-+-..;t=r===2=_=q==3:-]i"":""ll
96 Chapter 3. Flexural Waves in Beams
(3.35)
Z = a + ib = Ae i </> ,
the phase ¢ = tan-I (b/a) has an ambiguity of N77:. We remove this ambiguity
by keeping track of the total phase. That is, starting with some value and with
reference to the unit circle in the complex plane, as the wavenumber goes from
the 4th to the 1st quadrant the total phase is increased by 277:. Conversely, if the
wavenumber goes from the 1st to the 4th quadrant the total phase is decreased by
277:. The nth root is then given by
This scheme works quite robustly when the frequency increment is not too large.
Figure 3.12 shows all six spectrum relations for a curved beam with exaggerated
bending contribution. That is, keeping E A and R fixed (which fixes the cut-off
frequency), we increase E I by increasing the thickness of the beam. The figure
is plotted for the case of h = R. This value is physically unrealistic but it gives
a clearer picture of what the spectrums are doing in the small frequency range. It
appears that there are five branch points in all.
One of the difficulties in using the above formulas for the cubic solution is
that we are not always certain which branch is chosen after a branch point. In the
quadratic case, the ambiguity is removed by adding damping, this does not work
satisfactorily in the present case because of the combination of square and cube
roots. We will now give an approximation to the roots which is valid for slender
beams, that is, when E I « E A R2. The three roots of the characteristic equation
can be approximated nicely as
pAu} 9
± --+--
EI 4R4
± (3.36)
3.5 Curved Beams and Rings 97
Figure 3.12: Spectrum relations for a curved beam with exaggerated bending coupling.
These approximations are shown plotted in Figure 3.13 for h = R/lO. We note
that by decreasing h (with fixed R) that the complicated region shifts down to the
very low frequencies.
The approximate spectrum relations help us to get a clearer picture of the true
behavior. The coupling, generally, does not cause drastic changes in the spectrums
of the flexural dominated modes (kJ, k 2 ); indeed if h ~ Rj 100 it is clear from the
figure that the major effect is in the longitudinal dominated mode. The deviation
of the two flexurally dominated modes from the uncoupled theory corresponds to
the effect due to an axial compression. The consequences of this for beams were
treated earlier in this chapter. For the longitudinal dominated mode, the behavior
is seen to be similar to a rod with elastic constraint; this situated was treated in
Chapter 2. Finally, notice how the spectra are relatively unaffected at the higher
frequencies.
We conclude that for a slender beam of fixed radius, the most significant aspect
of the spectrum relation is the presence of a cut-off frequency in the longitudinally
dominated mode. The effect of this is illustrated in the following problem.
.8 Wavenumber, kh
mode 1
.6
.4
.2 mode 3
.0
-'~:~~3~-"="~",,,,,,,,,,,_ ,_~,~,~hX,~;_
-.2
-.4
-.6 .ode 2
Figure 3.13: Spectrum relations for a curved beam: comparison with approximate forms.
coupled modes. The solution for the forward propagating terms is written as
(3.37)
where the amplitude ratios for each mode are obtained from Equation (3.34) as
The solution is arranged so that for the uncoupled case we could set R t = R2 =
R3 = 0 and recover the usual solutions.
The free body diagram for the impact region looks similar to that of Figure 3.5
except that there is also an axial force. At the impact site s = 0, we impose
au =0 u(O, t) = 0,
as '
The first two of these allow the solution to be written as
v(s) A [e-ik,S - ae- ik2S + /JR3e-ik3S]
u(s) A [Rte-ik,S - aR2e-ik2S + tJe-ik3S]
p
A=-------------------~
2EI {[iki - aiki + f3R3ik~] - ~[k~ RJ - akiR 2 + f3k~]}
- - transverse velocity
........ axial velocity
s=o
s=R
s=2R
s=3R
._-------------------------------------------------------
. , , , ! , ! Ti!m~ [J.;LS1,
o. 1000. 2000. 3000. 4000. 5000.
There are two points of interest. First, note how the initial zero axial velocity
eventually becomes significant; this is a direct consequence of the coupling. Sec-
ond, note the oscillatory behavior of the transverse velocity; a frequency analysis
of these responses show that the oscillation coincides with the cut-off frequency
given by Wc = Col R. Other aspects of this problem are developed in Chapters 5
and 6 when we consider curved plates.
I I K
(3.38)
where, for simplicity, we take the properties of each beam to be equal. Obviously,
if K = 0 we get the uncoupled waveguide behavior of each beam.
2nm
/;m ==--
W
where W is the space window. On substitution, and since the result must be true
for any m, we conclude that
[
(EIl;~ - pAu} + K)
-K
-K
(EI/;~ - pAu} + K) ]{~~ L !~ L = {
{ VI} =W~~m
V2
1" -1 [a K
(3.39)
(3.40)
The solutions for VI and Vl can be incorporated in the usual FFT inverse. Unlike the
time reconstructions, it is generally more efficient to compute the above summation
directly rather than use the FFT algorithm because only a few response locations
x are usually required.
3.6 Coupled Beam Structure 101
-1
-, imaginary 1 -, imaginary
Figure 3.16: Contours of detenninant as a function of wavenumber ~ h at two frequencies
4.5 kHz and 5.5 kHz.
The remaining specification for the problem is the actual load distribution shape
- for simplicity, let us take it as being a point force. This gives
Responses are shown in Figure 3.17 where the history is that of our smoothed pulse
and the parameters M = 600, W = 150 m (6000 in.) were used. The feature of
significance in the responses is the high frequency rider. This is usually indicative
of some resonance (oscillator) type interaction. Intuitively one can see that the two
beams could begin to behave as a coupled spring mass system; but why is there
only one resonance when the beams have a continuous mass distribution? Is it
because there is only one spring stiffness?
Since our coupled beam system is symmetric, we also show the sum and differ-
ence of the responses at a common x location. The resonance is completely absent
in the sum whereas the difference is almost stationary.
102 Chapter 3. Flexural Waves in Beams
-v1 - - transform
....... v2 o waveguide
(v1 + v2)/2
I! , ! , t, ,," !, ,!!!. , ! , ! , t ! ,
Time [J.1S]
, I , , , I
Waveguide Solution
To construct a waveguide solution, we begin by enquiring if free wave solutions
of the form
{}
~1 = { ~1} e-ikx
V2 V2 0
are possible. To answer this, we substitute into the homogeneous governing equa-
tions to get the system of equations
2
[ (E1e- PAW +K) -K ]{~1} =0 (3.41)
-K (E1k 4 - pAw2 + K) V2 0
pAw2 -2K
E1
The individual eight roots are then obtained as the ±1 and ±i fourth roots of each
of these. This indicates that one set of modes will have a cut-off frequency at
{2K
pAw2 -2K =0 or w =
c VPA
This coincides exactly with the resonance observed in Figure 3.17. In fact, this is
the same oscillatory behavior as observed in the curved beam solution and in the
3.6 Coupled Beam Structure 103
rod with elastic constraint. Again, we see that elastic coupling leads to at least one
mode that is dissipative at low frequencies - this is indicative of energy transfer
between the two systems.
The solution involving the eight modes is written as
VI (x) + Be-ik,x + Ce- ik3X + ... + He+ik4X
Ae-ik,x
The amplitudes for each mode are not independent but related through Equa-
tion (3.41) as
[Elk 4- PAW2+KJA
K VIO = R VIO
A AA
V20 =
where R is called the amplitude ratio. For our special case of similar beams this
is simply ± 1 allowing the solution to be written as
VI(X) Ae-ik,x + Be-ik,x + Ce- ik3X + ... + He+ ik4X
V2(X) = Ae-ik,x + Be-ik,x _ Ce- ik3X + ... _ He+ ik4X (3.43)
To gain further insight into this solution, particularly as it pertains to the waveguide
solutions, consider now the new variables defined as
For the forward moving waves, we then get the very simple representations
(3.44)
We have thus reduced the system to an equivalent waveguide for the average
vertical motion vex, t) and average lateral contraction t{I(x, t). The generic form
of these behaviors are shown in Figure 3.18. It is clear that a superposition of the
two would give an asymmetric motion.
~ ~
antisymmetric -- x symmetric -- x
~ ~
Figure 3.18: Generic shapes of the waveguide behavior. Left is v = (VI + v2)/2, right is
1/f = (VI - v2)/2.
We are now in a position to consider the impact of this coupled structure; note,
however, that if the load is applied equally to each beam then the system can be
treated as a "smeared beam." That is, treated as an ordinary beam but with the
appropriate spectrum relations used. Thus, for the transverse deflections get
(3.45)
104 Chapter 3. Flexural Waves in Beams
(3.46)
These are compared in Figure 3.17 to the wavenumber transform results; both sets
of results are completely comparable.
Although the two solutions give the same results, they both have different impli-
cations for our structural analysis. First, the waveguide solution is computationally
more efficient (since it does not have a summation over m). This, however, is bal-
anced by the computational effort to determine the spectrum relation - the present
case is relatively simple, but more complicated situations can be envisioned. The
main advantage of the waveguide solution is that it allows the domain to be seg-
mented. That is, the transform method requires a very large space domain W;
the waveguide, on the other hand, can be finite or semi-infinite. The significance
of this is that we can connect many such waveguides at different positions and
orientation; this is not feasible with the transform method. Thus, one could even
imagine forming a complex connected structure made of finite segments of the
couple beams. This is not conceivable using the wavenumber transform solution.
Problems
3.1 Let the thickness of a beam be given by h = ho (a + x/a) where the width is
assumed constant. Show that the solutions are
•
vex) =
I (A}z(z) + BY2(Z) + C h(z) + DK (z)}
,Jz 2
vex, 0) = 0
Show that the shape at any later time is given by
The waveguide theories of the last two chapters were based on some elementary
mechanics modeling; the question obviously arises as to the adequacy of these
models. We will try to clarify the issues involved by considering some exact solu-
tions of waves in waveguides; this is a difficult and mathematically cumbersome
area so only a limited number of aspects will be treated. The chapter begins by
looking at waves in extended media and waves in bounded media as illustrated in
Figure 4.1. Perhaps the most important insight into the functioning of the exact
waveguides is the number of modes they support - in fact, they support an infin-
ity of modes. We will show that the behavior of these modes was anticipated by
considering the effects of elastic constraint in the elementary models.
The wave modes play a role similar to that of modes in vibration theory in
that the higher-order ones become important only at the higher frequencies. We
therefore discuss the construction of higher-order models but using only a limited
number of modes. We do this in a rational manner via Hamilton's principle; a
side benefit of this approach is that it shows the way to construct models for more
difficult or geometrically complicated situations.
P-wave
Figure 4.1: Experimental photoelastic picture showing a Rayleigh wave propagating near
a free surface.
105
106 Chapter 4. Higher-Order Waveguides
J-x r" j;
t"", X
a Xl
Z au
aUj aUj
2Eij = -+- (4.1)
aXj aXi
au av aw
Exx = ax ' Eyy = ay , Ezz = -
az
au av av aw aw au
Yxy = ay + ax ' YyZ = az + ay , Yzx =~+az
The components of strain are symmetric. Note that the engineering strain Yij is
twice the corresponding tensorial value, Eij.
Let the material be linear elastic and isotropic, and governed by Hooke's law
taken in the form
211E- -
r IJ
= a-IJ- - 2J-L
A
'7: '
+ 3A 8-·lJ '""' akk aij = 2J-LEij + A8ij L
k
Ekk (4.2)
4.1 Waves in Infinite Media 107
where JL and A are the Lame constants and 8ij is the Kronecker delta. For future
reference, the relationship among the various material constants are
vE E (1 - v)E
A= , A + 211 = (43)
(1 + v)(1 - 2v) JL = G = 2(1 + v)' .- (1 + v)(1 - 2v) .
where Ji are the components of the body force. For example, the first of these in
expanded notation is
OC1xx OC1xy OC1xz 02u ou
- - + - - + - - +p/x = P -2 +1]-
ax oy az at at
The two other equations are obtained by cyclic permutation of the subscripts. Com-
bining the equations of motion and constitutive relation, and using the displacement
as the dependent variable gives the Navier's equations as
(4.7)
where Eipqis the pennutation symbol. This represents the three displacements
U 1. U2, U3with the four potential functions <1>, H\, H2 , H3; note that the extra con-
dition uniquely detennines Hq • On substituting for Ui into the Navier's equations,
get on rearranging
These equations are satisfied if the tenns inside the square brackets are zero, that
is, if <I> and Hq are chosen such that
2 .. •
/LV Hq - pHq -17Hq = 0 (4.9)
Each of these equations are wave equations and both are nondispersive when there
is no damping, TJ = o.
It will be shown in more detail later that <I> is associated with the P-wave that
travels with speed C p, given by
c 2p A +_
== _ 2/L = E(l - v) =c2 _ _1_ -_ v __
(4.10)
P p(l + v)(l - 2v) 0 (1 + v)(l - 2v)
The other three potentials, H q , are associated with the S-wave that travels with
speed Cs given by
2_/L G E 2
C =-=-= =c - - - (4.11)
s p p 2p(1 + v) 0 2(1 + v)
It is worth pointing out that it is interactions with boundaries that usually makes
the waves dispersive; in the construction of a waveguide theory, the boundary
conditions are usually already embedded in the theory and that is why the waves
are usually dispersive.
4.1 Waves in Infinite Media 109
Figure 4.3: Directions of the waves. P and SV lie in the x-y plane, SH lies perpendicular
to the x-y plane.
(4.12)
a<I> a<I>
U=-
ax '
v -ay
-, w=o (4.13)
_ AV2<1>
a2<1>
= 2/L-- a xz = a yz = 0 (4.14)
ax8y'
Note that although the strain in the z direction is zero, the corresponding stress
is not; this is a situation of plane strain. The behavior of the P-wave is the 3-D
analog to the one-dimensional longitudinal wave in a rod.
110 Chapter 4. Higher-Order Waveguides
Hz only: SV-waves
The S-waves are also called equi-voluminal, distortional, rotational, shear, or sec-
ondary (S) waves. The differential equation to be satisfied is
a2 a2 } .. .
II
fA'
{- +-
ax2 ay2
Hz - pHz- "'YIHz = 0 (4.15)
aHz aHz
U=- v == - - - w=o (4.16)
ay , ax '
(1xx
a
2H
2J-L-_z
axay
U yy
a2 Hz
... -2J-L--
axay
(1xy == J-L[- a 2 Hz + a 2 Hz]
ax 2 ay2
(1zz == U xz - (1yz - 0 (4.17)
The signs and magnitudes of the normal stresses indicate a state of pure shear stress
on planes oriented ±45° to the coordinate axes. The behavior of the SV-wave is
the 3-D analog to the shear behavior of flexural waves in a beam.
aHx aHy
U = v = 0, w= - - - + - - , (4.18)
ay ax
a2 2 0
J-L { - + -a } W -
..
pw - "1w =
. (4.19)
ax 2 ay2
4.2 Semi-Infinite Media 111
q(x, t)
(a) (b)
k 2 = poi - iTJw
p - A+2f..1,
112 Chapter 4. Higher-Order Waveguides
In Cartesian coordinates, we can expect exponential solutions of the form e-ik,y e- ikxx .
Substitute this into the governing equation to get the restriction on the wavenum-
bers kx and ky; that is,
k2 _ pw 2 - iT/w
S=
f..L
where, for phase matching reasons, the same kx appears in both relations. Fur-
thermore, the ky wavenumbers are different for each potential and are determined
from
(4.21)
Note that if kx is greater than k p , then kpy is imaginary and the solution is a wave
in x but evanescent in y. That is, it is localized to near y = O. Both of these spectra
have the characteristic behavior of a rod with an elastic constraint, but the spring
constant depends on the wavenumber kx • A plot of the spectra, for different values
of kx is shown in Figure 4.5.
The spectral form for the displacements corresponding to these potentials are
(4.22)
When imposing the boundary conditions at y = 0, say, it is required that the phases
of CI> and Hz be the same. It is clear from the displacement and stress expressions
that this is ensured by having (kxx - wt) common in both potential functions.
4.2 Semi-Infinite Media 113
-P-mode
--.. -- S-mode
I; = 2lrm
m- W
where W is the space window. Since f(x) is real-only we could use the usual form
of the Fourier series; however, the exponential form is easier to manipulate.
The tractions along y = 0 are given by
From this we conclude that we must represent both potentials as summations over
kx = I;m = 2rrmtW in order to match the tractions conditions. We now get the
system of equations to be solved as
(4.24)
where
114 Chapter 4. Higher-Order Waveguides
Recalling that k p and ks depend on frequency W n• we see that the solution will
entail summations over n and m; in the following we will take the frequency
dependence for granted and concentrate only on the ~m dependence. We get for
the displacements. for example.
This can then be incorporated in the usual FFf inverse; generally. however, it is
more efficient to compute the summations over m directly rather than use the FFf
algorithm because the response is usually required at only a limited number of
points (x, y) .
.5
.4
.3
.2
.1
.0
-.1
-.2
-.3
-.4
-.5
O. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. -.5 -.4 -.3 -.2 -.1 .0 .1 .2 .3 .4 .5
Frequency [kHz] imaginary wavenumber
Figure 4.6: Contours of the determinant. On the left is a slice of Re(~) against frequency
for Im(~) = 0, and on the right is Re(n against Im(~) at 5 kHz.
(4.25)
Figure 4.6 shows how this determinant depends on wavenumber and frequency. In
the left figure, the horizontal axis shows the frequency from 0 to 10kHz and the
vertical axis is Re(~)b varying from -0.5 to 0.5. Since most of the contributions
would come from where t. is a minimum, the contours show that these seem
to follow a line ~ = const w. This interesting result implies the possibility of a
nondispersive wave traveling in the balf-plane - we will take that issue up in the
next section but before then we check if the responses indicate something similar.
4.2 Semi-Infinite Media 115
The remaining specification for the problem is the actual load distribution shape.
For simplicity, let us take this as being rectangular over a length L. This gives
im = (ei~L - l)/(i~m)
Responses are shown in Figure 4.7 where the input history is that of our smoothed
pulse and the parameters M = 600, W = 150 m (6000 in.) were used. Along the
surface we see very little attenuation beyond the x = 200 mm mark.
V y=500mm
V x=500mm
V y=400mm
V x=400mm
V y=300mm
V x=300mm
t y=200mm
V x=200mm
"If>
(a) (b)
I
O.
I
500. 1000.
I I
1500.
y-100mm
. ITime[~l
2000.
I
o.
I
V
500.
, I
1000.
I I
1500.
I
x=100mm
I I I I
2000.
I I
2500.
Figure 4.7: Velocity responses for a point impact on a half-plane. (a) Penetration into the
plane at x = 0, (b) along surface y = o.
The displacement of the surface itself is shown in Figure 4.8. This rather in-
teresting behavior is the type of motion exerted on the foundations of buildings
during an earthquake.
Figure 4.8: Surface Rayleigh behavior at a distance x = 400 mm from the impact site.
116 Chapter 4. Higher-Order Waveguides
(4.26)
where y is
2 f..L c~ (1 - 2v)
y =--=-co---
)... + 2f..L c~ 2(1 - v)
The characteristic equation is cubic in k 2 , hence there are three possible modes.
Furthermore, since no explicit frequency occurs in this equation (there are only
ratios of wavenumbers which are simple numbers in the undamped case); the
characteristic roots are linear in frequency and the surface wave, if it exists, is
nondispersive.
For the special case of)... = f..L, (v = 0.25) we have y2 = 1/3 and the equation
can be factored as
While all three correspond to waves propagating in x, it is only the third that is
evanescent in y. This special wave is called a Rayleigh wave and the speed is
denoted by CR. The potentials and amplitude ratios are
B 2;}1 - (cRlcp)2
(4.28)
A 2 - (cRlcs)2
Antisymmetric Potentials
We first concentrate on constructing potentials that can satisfy the lateral boundary
conditions. To this end, we introduce
(4.30)
The spectral form for the displacements corresponding to these potentials are
Ii A[ -ikx]Spy + B[ -ksy]Ssy
v = A[kpy]Cpy + B[ikx]Csy ( 4.31)
q(x) = ip I(x)
If we impose that the area under I(x) is unity, then this corresponds to applying
half of the resultant load P on each of the lateral surfaces. Now represent the
distribution shape with the Fourier series
2nm
~m==-
W
where W is the space window. Note that since I(x) is real-only we could use
the usual form of the Fourier series; however, the exponential form is easier to
manipulate. Indeed, since I(x) is symmetric in the problem being considered we
could even use the cosine series expansion.
The tractions along y = ±h/2 are given by
q = ayy = {A[ -le(k;+ k~y) + 2f.L( -k~y)]Sp + B[2f.L( -i ksykx)]Ss }e- ikxX
a
0= xy = {A[2f.L(-ik py kx )]Cp + B[f.L(k; - k~y)]Cs }e- ikxx
where
After solving for A and B, we can obtain the displacements and stresses, and these
can then be incorporated in the usual FFf inverse to do the time reconstructions.
The responses, when reconstructed, look similar to those of an impacted beam
and so we will not pursue that aspect at the moment. The aspect that this solution
allows us to investigate is the distributions of displacement and tractions on an
arbitrary cross section. This is precisely the information we will require later in
constructing approximate waveguide theories.
Position, y/h
.50
.25
.00
-.25
-.50
Figure 4.9: Exact displacement and traction distributions, on the cross section at
x = 500 mrn, for the antisymrnetric response of a deep waveguide.
Figure 4.9 shows the distributions of the displacements and tractions when a
20kHz narrow-banded pulse was input to a waveguide of h = lOOmm. This
relatively deep waveguide was chosen so as to exaggerate some of the 2-D effects.
Each line in the figure corresponds to a different time and generally seven points
through the thickness were used. The purpose of these plots is to get a sense of how,
dynamically, the distributions on the cross sections vary over time. While there is
some "jaggedness" in the plots, a number of features are quite apparent. The two
features of note are: the u displacements indicate that similar to the elementary
beam theory "plane sections remain plane", but contrary to the elementary beam
theory there is a parabolic shear stress distribution.
Symmetric Potentials
The development for the symmetric problem closely follows that of the previ-
ous subsection. We first construct potentials that can satisfy the lateral boundary
120 Chapter 4. Higher-Order Waveguides
conditions. Introduce
(4.34)
The only difference, in comparison to the anti symmetric case, are the distributions
with respect to y. The spectral form for the displacements corresponding to these
potentials are
u A[-ikx]C py + B[ksy]Csy
v A[ -kpy]Spy + B[ikx]Ssy (4.35)
From this we conclude that we must represent the potentials as summations over
kx = ~m = 2n m / W. This leads to the system of equations to be solved as
[-)..(~;' + k~y) + 2J.t( -k~y)]C p
[
[2J.t(ikPy~m)]Sp
where
~m = 2nm/W
4.4 Doubly Bounded Media: Lamb Waves 121
Position. y/h
Characteristic Equations
For the antisymmetric waves, we ask if the potentials
represent waves that satisfy zero traction on the lateral boundaries. Note that at this
juncture, the wavenumber kx = k is as yet unspecified. The traction free boundary
condition now becomes
where
This is the same as Equation (4.33) with the difference that kx is unspecified and
the system of equations is homogeneous. We can get a nontrivial solution for the
homogeneous system only if the determinant is zero. That is,
The wavenumber k can be chosen so that this equation is true. However, this is
a complicated transcendental equation with the relation between wavenumber k
and frequency w (through k p and ks ) obviously multivalued; it therefore must be
solved numerically.
For the symmetric waves, we can proceed as before using the potentials
This too is complicated and requires numerical methods for its solution.
Generally, in the exact solution of waveguides, we will encounter characteristic
equations that are transcendental and therefore generally cannot be solved in closed
form. As a consequence, some numerical scheme is usually used for solution. The
next section outlines a simple numerical scheme for obtaining complex roots; this
is the scheme used to obtain the plots.
4.4 Doubly Bounded Media: Lamb Waves 123
which is just the sum of the absolute values; in some cases, using the logarithm
of this is better. The problem is therefore reduced to one of finding the minimum
value of the real function w(x, y) in the real (x, y) space.
A straightforward scheme for finding the minimum of a function is to start at
some point (xo, Yo), say, and by estimating the slopes at that point (by function
evaluations at neighboring points (Xl, YI), (X2, Y2» a direction for marching toward
a minimum can be chosen. This is repeated many times with decreasing step size
until a function value as small as desired is achieved. A few extra features in order
to check convergence and to choose the step size complete the algorithm.
This method will find a local minimum; good initial guesses are required in order
for it to find the global minimum. There are no hard and fast rules about obtaining
good initial guesses; however, once a value of a root is obtained at a particular
frequency, then this can be used as the initial guess at the next frequency of the
same spectrum. There is generally no problem until two modes overlap or come
very close to each other. In these cases, it may be necessary to start the iteration at
a higher value of frequency and arrive at the same point from the other side of the
branch point. Once a single spectrum has been determined, then the others can be
obtained in a similar manner (by starting with a different smart guess) or by using
a deflation scheme.
sinh(kh) ± kh = 0 (4.39)
124 Chapter 4. Higher-Order Waveguides
(a)
I ! I ! Ii ! I ! ! I ! ! ! ! I ! ! ! ! I ! ! ! ! I ! ! ! ! I ! ! f ! I ! ! ! ! I ! ! , ! I ! , ! ! I ! ! ! t! ! ! ! t!
-6. -5. -4. -3. -2. -1. O. 1. 2. 3. 4. 5. 6.
Figure 4.12: Spectrum relations for the first four antisymmetric mode pairs. The wavenum-
bers range over -2 < krealb < 2, -2 < kimagb < 2.
which as outlined above follows the elastic constraint behavior. Again an important
feature is that the next two modes approach the axis simultaneously which means
that both the second and third modes need to be considered simultaneously. This
is an important distinction between the symmetric and anti symmetric modes: for
the former it is the second and third modes which are next in order of importance,
whereas for the latter it is the third and fourth.
Figure 4.13: Spectrum relations for the first three symmetric Lamb mode pairs. The
wavenumbers range over -2 < krealb < 2, -2 < kimagb < 2.
When external forces act on the structure, stresses are set up inside the body.
Consider a typical small volume of dimensions lix, li y, liz, as shown in Figure 4.2.
This volume is under the action of the following system of stresses:
A line that was originally of length lix in the undeformed state has a new length
lix(1 +Exx) in the deformed state. Now consider an additional infinitesimal defor-
mation added to the body; a typical change in length is lix(dExx). The work done
by the O'xx component of stress during this change is
where the symbol V represents the volume. Hence the work increment for the
whole body is obtained by allowing the typical volume to become infinitesimal
and then integrating over the total volume. For the complete system of stresses we
have the generalization
Because of perfect elasticity, the work expended can be regained if the loads are
gradually decreased. That is, the work is stored in the elastically distorted body
in the form of energy known as strain energy or elastic energy. We will use the
i
definition
B
liU - i{O'}T{dE}dV (4.41)
Using Hooke's law, this can be put in the alternate fonns of stress only or strain only
when needed. The above relations will now be particularized to some structural
systems of interest by writing the distributions of stress and strain in tenns of
resultants. We have already developed these structural members in the previous
chapters, so the assumptions and restrictions as stated there are still assumed to
apply here.
For the rod member, there is only an axial stress present and it is uniformly
distributed on the cross section. Let F be the resultant force; then a = F I A = E E
and
axial:
210 EA
r
u =! L ~dx =!
210
rL
EA (dU)2 dx
dx
For the beam member in bending, there is only an axial stress but it is distributed
linearly on the cross section in such a way that there is no resultant axial force. Let
M be the resultant moment; then a - -MYI 1= EE and
bending: u =!
2
r
10
L
M2 dx =!
EI 2
r
10
L
EI (d 2V)2 dx
dx 2
The shear forces in a beam can also do some work. Let the shear stress be assumed
to be uniformly distributed on the cross section, and the resultant shear force be
V, then r = V I A = Gy and
torsion: U =!
2
lL
0
T2
- d x =!
GJ 2
lL
0
d~ 2
GJ ( - ) dx
dx
where ~ is the twist per unit length, T is the resultant torque, G the shear modulus,
and J is the second polar moment of area.
State C. A deformable body has an infinite number of material points and hence has
an infinite number of degrees offreedom. We therefore must assume that all real and
virtual displacements are continuous (differentiable) functions of the coordinates
x, y, and z. Functions that have these properties are called kinematically admissible
functions.
The virtual work of a deformable solid body is divided into two parts as follows:
where aWe is the virtual work of external forces and can be broken further into a
contribution oW S of the surface forces and a contribution oW b of the body forces.
The term aWi is the virtual work of the internal forces, in this case the work due
to the elastic straining of the body. That is, the internal virtual work is given by
(4.42)
which states that a deformable body is in equilibrium if the total virtual work
is zero for every independent kinematically admissible virtual displacement. We
will interpret the symbol 0 as meaning a variation and the above equation as a
variational principle.
To apply these ideas to dynamic problems, we need to account for two new
aspects: first is the presence of inertia forces, and second is that all quantities are
functions of time. We will take care of the former by use of d'Alembert's principle
and the latter by time averaging. To make the developments concrete, we initially
analyze the one-dimensional rod; the extension to multiple dimensions will follow
easily since we work with scalar quantities.
D' Alembert's principle converts a dynamic problem into an equivalent problem
of static equilibrium by treating the inertia as a body force. For example, for a rod
we treat the total body force (per unit length) as comprised of fb - pAil where
-pAil is the inertia of an infinitesimal volume. The virtual work of the body force
is
oW b =Iv fb oU dY - Iv pAilou dY
In writing this relation, we suppose that the performance of the virtual displacement
consumes no time; that is, the real motion of the system is stopped while the virtual
displacement is performed. Consequently, the time variable is conceived to remain
constant while the virtual displacement is executed.
We will concentrate on the inertia term in the above virtual work expression.
Noting that
-
d(.~) ......
U IJU = U oU + U oU
dt
130 Chapter 4. Higher-Order Waveguides
(a) (b)~V+IlV
......................._._.
.......... t2
......... v(x, t)
v+llv
time
Hamilton disposed of the last term in the virtual work relation by integrating
the equation over time between the limits of the 1 and 2 configurations. The last
term is a time derivative and so may be integrated explicitly to give
-f u~udvl::
By assumption, the configuration has no variations at the extreme times and hence
the term is zero. Consequently, the virtual work relation becomes
f r,
r2 [8W" + 8W b + 8T - 8U]dt = 0 (4.43)
4.6 Modified Beam Theories 131
8 1.
I,
/2
[T - (U + V)]dt = 0 (4.44)
Among all motions that will carry a conservative system from a given
configuration at time t\ to a second given configuration at time t2,
that which actually occurs provides a stationary value of the integral.
were used. The displacements are also functions of time but that is not important at
the moment. Since we are interested in flexural deformations, we set u{x, 0) = O.
Looking at the deformations of Figure 4.9, it can be assumed that the vertical
deflection is nearly constant while the horizontal is almost linear. We therefore
retain only one term in each expansion and obtain the approximate displacements
as
u{x, y) ~ -ycf>(x), vex, y) ~ vex) (4.45)
This says that the deformation is governed by two independent functions, v{x) and
cf>(x), that depend only on the position along the centerline. We could, of course,
retain more terms in the expansion and develop an even more refined theory, but
as will be seen, this expansion is adequate for our purpose.
The axial and shear strains corresponding to the above deformation are
ali au av ( av)
EV\' = - = 0, = ay + ax = -cf> + ax
.. ay Yxy
acf>
a = -yE- ax)' = G (-cf> + av)
n ax ' . ax
4.6 Modified Beam Theories 133
It is worth comparing these to the distributions of Figure 4.9. The most significant
difference is in the shear stress distribution; the present distribution can, at best,
only represent the average or resultant shear behavior. Substituting for the stresses
into the strain energy density function, we have
We assume the same kinematic description applies even in the dynamic case,
therefore the total kinetic energy is
rLq(x)vdx -
- 10
L
M¢lo -vvl o
L
8 1 10r
II
12 {
L[ . a¢ 2 av 2 ]
HPAv 2 + PI¢2]_4[ E1 Cx) +GA( -¢+ ax) ]+qV dx
+M¢I>vvl~} dt = 0
Taking the variation inside the integrals and using integration by parts, we get
134 Chapter 4. Higher-Order Waveguides
The fact that the variations 8v and 84J can be varied separately and arbitrarily, and
that the limits on the integrals are also arbitrary, lead us to conclude from the terms
in square brackets that
a [av
GA- --4J ] pAv - q
ax ax
a24J
EI-+GA [av
--4J ] (4.46)
ax 2 ax
The associated boundary conditions (at each end of the beam) are obtained from
the remaining terms and are specified in terms of any pair of conditions selected
from the following groups:
{v or V=GA[:~-4J]}, (4.47)
Adjustable Parameters
Because we started with an approximation for the deformation distributions, this
gave us an approximate potential function from which we derived a governing
differential equation and a set of boundary conditions most consistent with that
approximation. We can imagine, therefore, proposing a different potential and
deriving natural boundary conditions and governing equations that are different
but nonetheless consistent. This is a key point: we use the derivation of the strain
and kinetic energies to establish the form of the potential, but then we are free to add
adjustable parameters because the variational principle will correctly implement
these parameters.
Rather than associate the parameters with the deformation, it is a little more
convenient to modify the potential itself. Reasoning that the shear is represented
least accurately in our modeling, we modify the energy terms associated with it as
where the adjustable parameters Kl and K2 have been introduced. The modified
differential equations become
GAK 1 - a [av
--4J ] pAv - q
ax ax
--4J]
a -4J +GAK 1 [av
2
EI p IK2¢ (4.48)
ax 2 ax
4.6 Modified Beam Theories 135
{q> or M = EI :~ } (4.49)
Spectrum Relation
Since there are two dependent variables v, q> and the coefficients are constant, then
assume solutions of the form
v = voe-i(kx-wt) ,
GAK\k2 - pA{J}
[ (4.50)
ikGAK\
PAW2]\/4 .[PAW2]1/4
k):::::l± [ - - k2 :::::l±1 - -
EI ' EI
These are the Bernoulli-Euler beam spectrum relations. Note that they do not
depend on either GAK 1 or pI K2 and that is why this higher-order theory does
136 Chapter 4. Higher-Order Waveguides
2.0
- Bernoulli-Euler
1.6 - Timoshenko
1.2
.8
.4
real
.0
imaginary
-.4
, Frequency [kHz}
-.8 I! , , , '-T:;! ! , I , , , , , , ,
1.2
- Bernoulli-Euler
1.0 - Timoshenko
o Lamb
.8
.6
.4
.2
.0
Frequency [kHz}
, ,
-.2 , ! , , , , ! , , , ,
not introduce additional modes. On the other hand, as the frequency becomes very
large
k2 => ±w - -2 IK
JP EI
Both are propagating, nondispersive, modes. Therefore, the imaginary branch must
have turned real. Look for a possible cut-off frequency where k is zero and this is
found to be where
(4.51)
The limits on the phase and group speeds give for high frequencies
cg"'c- JGAKI
-- and
pA
KI = - =
c1 (0.87 + 1.12V)2 ~0.86
c~ 1+ v
Either deformation can be chosen as the unknown since the amplitude coefficients
are related by
For the impact problem, it is simplest to choose the rotation as the working variable.
138 Chapter 4. Higher-Order Waveguides
input
~ ___________________________________ x=O
x=17.Sm
, ! ! I , I " ! I ! ,
Time [~]
o. 1000. 2000. 3000. 4000. 5000. 6000. 7000.
Slope: ¢= ¢
Moment:
Shear Force:
(4.53)
Thus, the procedure is essentially the same as used for the elementary theory. The
only real difference is that the resultants are slightly more complicated functions
of the dependent variable.
So as to put the central frequency of the input pulse (30 kHz) above the cut-
off frequency, the beam thickness was increased to 100 mm. (This gave a cut-off
frequency of about 15 kHz.) Such an input force generates responses that also have
many zero crossings. The reconstructions of Figure 4.18 show the propagation of
two separate waves emanating from a single impact. The reconstructions are for
the slope ¢(x, t); notice that at x = 0, it is zero.
4.7 Modified Rod Theories 139
EA, pA, v, L
Fo ---1'-______--'~ FL
t-+ Uo
q(x, t)
.. . au
Ut = TE t = -VTE = - V T -
ax
In this, we have assumed that the transverse displacement is proportional to the
distance T from the centroid of the cross section. The total kinetic energy of the
140 Chapter 4. Higher-Order Waveguides
T= 1~p[U(X,t)2+UI(X,t)2]dv=~lL ip[u2+v2r2(:~fJdAdX
Since uis a function only of x (and time) then we can perform the integration with
respect to the cross section to give
where J is the polar moment of area. The total strain energy is given by
U = ~ lL EA (::Y dx
The final term we need is the potential of the applied forces. With reference to
Figure 4.19, the potential of these are given by
v=-lLqdX-(-Fouo+FLud=-lLqdX-FUI~
Substitute these energies into Hamilton's principle to get, after taking the variations
inside,
1 10r
II
12 {
L [PAU8U + v2pJau a8u _ EA au a8uJ dx + F8ul~} dt = 0
ax ax ax ax
Integrate this by parts so as to have all terms multiplied by a common variation
8u:
1II
12
0
[a
2
{lL E A ---;
ax
2 a
+ v 2pJ ---;·· - pAu ] 8u dx
ax
- a a··
[EA-.!:!..+v2pJ-.!:!..-F ] 8ui L} dt=O
ax ax 0
Since the time limits and space limits in the integrations are arbitrary, then the first
integrand is zero giving the governing differential equation for u(x, t) as
a2 u a2 ii
EA-+v2pJ- -pA- =0
a2 u (4.54)
ax 2 ax 2 at 2
It is clear that if either Poisson's ratio or the polar moment of area is negligibly
small then we recover the elementary rod equation.
The remaining terms must also be zero and thereby they specify the boundary
conditions; at either end of the rod, we specify
u or
au 2 aii
F=EA-+v pJ- (4.55)
ax ax
4.7 Modified Rod Theories 141
pA
kl =±w
This has the odd behavior in that it has the opposite of a cut-off frequency - it has
a frequency were the wavenumber goes to infinity. After this critical frequency, the
wavenumber is purely imaginary. What this means is that the transverse motion is
absorbing all the input energy.
.8
- Elementary
.6 ....... Love
....... Mindlin-Herrmann
.4
-Three mode
o Exact Lamb
.2
.0 Frequency [kHz]
! , ,I " I! ",! , , , , , / I ! I !! ,'!! ! I
=.j [1 _
3
C = dw EA V2P]W2J2
g dk pA EA
This goes to zero at the critical frequency. Figure 4.20 shows a comparison of the
dispersion behaviors of the Lamb symmetric modes and of the Love theory. (The
effective] and A were computed for a cross section of h x h.) As was done with
the Timoshenlco theory, it is possible to introduce an adjustable parameter. This
142 Chapter 4. Higher-Order Waveguides
ends up appearing in the identical positions as the Poisson's ratio term, that is, as
v2 pJ Kr. The figure shows the dispersion relation when K\ = 1.2.
This plot is complicated in the vicinity of 25 kHz because the exact Lamb modes
indicate the occurrence of two additional propagating modes.
This accounts for the lateral displacements but ignores the nonuniform distribution
of u on the cross section. Using these displacements to get the stresses and strains
and after substituting into the strain energy density function and integrating on the
cross section, we get
u = ~lL[(2JL+A)A[(::r +Y,2J+2AA::Y,+JLI[~:r]dX
T = ~ Io [pAu + pl1P] dx
L
2
If the applied surface tractions and loads on the rod are as shown in Figure 4.19,
then the potential of these loads is
{L L
V = - 10 q(x)vdx - Fvl o -Q\y,loL
Proceeding as before, but introducing adjustable parameters associated with the
shear and inertia terms, we get the Mindlin-Herrmann equations as
a2u ay,
(2JL + A)A ax 2 + AA ax pAii - q
a2 Y, au (4.57)
JLI K\ ax 2 - (2JL + A)Ay, - AA ax
The associated boundary conditions (at each end of the rod) are specified in terms
of any pair of conditions selected from the following groups:
These are a set of coupled equations in the longitudinal displacement u(x, t) and
lateral contraction 1/1 (x , t), and were first presented in Reference [93]. Note that
the equation for "" corresponds to elastic constraint. The elementary theory is
recovered by setting /-LI K\ ... 0, pI K2 = 0 and replacing 1/1 in terms of u in the
first of the above equations.
Since there are two dependent variables u and 1/1 and the coefficients are constant,
then to get the spectrum relation assume solutions of the form
On substitution into the differential equations these give the system to be satisfied
as
2
[ -(A + 2/-L)Ak + pAai -ikAA ] { Uo } = 0
ikAA -/-LI K\k2 - (2A + /-L)A + pI K2W2 CPo
Setting the determinant to zero gives the characteristic equation for determining k
as
[(211-+A)A/-LIKtJe+ [4/-L(I1-+A)A 2 - p IK2w2 (2/-L+A)A - pAw2/-LIK\]k 2
- [(211- + A)A - pI K2W2] pAw2 - 0
This characteristic equation is quadratic in k 2 and therefore, in all, there are two
mode pairs in contrast to the single pair of the elementary solution. Figure 4.20
shows the dispersion behaviors.
The mode 1 behavior is real only, but it shows a decreasing speed with frequency.
In the limit of low frequencies, this mode reduces to that of the elementary theory.
The second mode behavior is quite different. First, it exhibits a cut-off frequency
given by
(211- +A)A
pIK2
Below this frequency the second mode attenuates. Where this occurs is also in-
versely proportional to the height h or radius a. Thus slender rods give a large We
and therefore the second propagating component of the wave is usually not ob-
served. Furthermore, even though the mode is present at the low frequencies, the
large value of k2 means that it is highly attenuated for any nonzero x; consequently,
it is usual to neglect it.
Beyond the cut-off frequency, the second mode is a propagating mode but it does
not resemble either of the exact modes. It seems that the addition of a second mode
has improved the estimate of the first mode behavior but only crudely approximates
that of the second mode.
There are a variety of ways to determine K I and K 2. Consider the limiting cases
of low and high frequency (or what is the same thing, small dimensions a or h),
respectively. That is, for low frequencies
{EA
kl =±VPA'
144 Chapter 4. Higher-Order Waveguides
The first mode is the elementary solution. Notice that the second mode is imag-
inary only which contrasts with the Lamb mode of being complex. It is as if
this is approximating the average of the second and third Lamb modes. For high
Rh
frequencies
A
kz=±w - -
2/1 + A
The first of these should coincide with the Rayleigh wavenumber
or (4.59)
Since the Rayleigh wave speed is very close to the shear wave speed then we can
take this ratio as unity, or as
-
KJ
=
(0.87 + 1.12V)2 ~ 0.86
Kz 1+ v
The separate values of the parameters can then be chosen to satisfy the cut-off
frequency condition, that is, KJ = 12/rr2. The dispersion relations of Figure 4.20
were obtained using the parameters KJ = 1.68, K z = 1.80. These were chosen so
that the cut-off frequency coincided with the minimum group speed for the Lamb
modes.
Three-Mode Theory
While the Mindlin-Herrmann theory extended the range of the Love theory, it is
clear from Figure 4.20 that it breaks down in the vicinity of the cut-off frequencies.
The main reason is that the Lamb modes have a pair of modes intersecting in this
frequency range and the Mindlin-Herrmann theory is approximating this by only
a single mode. We will now add a third mode.
We begin by expanding the deformations in a Taylor series and retain one addi-
tional term in each expansion to obtain the approximate displacements
vex, y) ~ 1{I(x)y
In this form, u(x) has the meaning of average axial displacement. The normal and
shear strains corresponding to these deformations are
au au a¢ yZ aji
Exx - = - + -h(1 - 12-), Eyy =- = 1{1
ax ax ax h2 ay
au + av = (_ ¢24~ + a1{l)y
ay ax h ax
The corresponding stresses are
Y + -a1/!] y
/-t [ -fjJ24-
h ax
where 'A, /-t are the Lame constants. It is worth comparing these to the distributions
of Figure 4.10. The most significant difference is in the shear stress distribution;
the linear distribution of the present formulation can be interpreted as representing
the average shear behavior. Substituting stresses and strains into the strain energy
density function and integrating on the cross section, we get
where we have associated resultants Q\ and Q2 with the degrees of freedom 1/!
and fjJ, respectively. After using Hamilton's principle, we get
a2 u a1/!
(2/-t + 'A)A-2 + 'AA- pAu - q
ax ax
a2 1/! au afjJ
/-t I ax 2 - (2/-t + )")A1/! -)"A ax - 2/-tAh ax pIlf
with the associated natural boundary conditions (at each end of the rod)
(4.61)
The differential equations for 1/! and fjJ correspond to elastic constraint, so before
we proceed to consider the full spectrum relation, it is worth looking at the system
146 Chapter 4. Higher-Order Waveguides
composed of just '1/1 and t/J. The characteristic equation is obtained from
(2/L + J...)A
w = .jS/L A
C pl' pI
These two modes certainly capture the behavior of the second and third Lamb
modes as depicted in Figure 4.13.
This equation is cubic in k 2 and the formulas of Section 3.5 can be used for its
solution. Figure 4.21 presents the full spectrum relation and it seems to capture
all of the relevant behaviors of Figure 4.13. Furthermore, the direct comparisons
of the group speeds in Figure 4.20 shows that the model can be adjusted to give
accurate representation. The figure, in fact, shows the dispersion relations with no
adjustments made to the spectrum relations.
w 3cs
f=-~--
21l" 4h
that a simple modified rod theory such as the Love theory, is all that we need in
structural analysis.
Problems
4.1 Consider the possibility of propagating free surface waves propagating with
the boundary conditions at y = 0 of v = 0, axv = O.
- Reference [56], p. 329.
4.2 Consider the possibility of free waves in the doubly bounded medium when
the boundary conditions at y = ±h12 are v = 0, ax)" = O. Show that the
characteristic equation is
There are two families of modes, show that they are given by
H- =
Pk,
-2--e
-lks \' -ik,x
)'e
• 2ksJ.tksv
- Reference [72], p. 8.
Problems 149
4.7 SH-waves propagating in a half-space overlaid by a thin solid layer give rise
to the possibility of Love waves. Show that for a layer of thickness T and
properties indicated by a prime, their characteristic equation is
The only way to efficiently handle wave propagation problems in structures with
complicated boundaries and discontinuities is to develop a matrix methodology
for use on a computer. In this chapter, we develop such a method.
Our approach is similar in style to that of the finite element method but has the
very significant difference that the element stiffness matrix is established in the
frequency domain. This allows the inertia of the distributed mass to be described
exactly. As a consequence, these spectrally formulated elements exactly describe
the wave propagation dynamics and (in contrast to the conventional element) this
means that elements can span all the way from one joint to another without loosing
fidelity. Furthermore, the system of equations to be solved is very small. The
approach provides all the advantages of the spectral analysis method plus the
efficiency and organization of the finite element method. A typical structure with
spectral element modeling in shown in Figure 5.1.
o • nodes
hole
(a) (b)
Figure 5.1: A general structure modeled as connected waveguides. (a) Physical structure,
(b) spectral element model.
150
5.1 Structures as Connected Waveguides 151
(c)
,pI
:::;.
TI
GJ, pI, L ~
.:::::
T2
,pI
:::;::
TI
. II
Figure 5.2: N odalloads and degrees of freedom for three spectral elements. (a) Longitudinal,
(b) flexural, (c) torsional.
Many frame structures can be treated as a collection of the basic elements. For
those cases involving local discontinuities such as holes, cracks, or complicated
joints, we introduce a spectral super-element.
152 Chapter 5. The Spectral Element Method
Shape Functions
The general longitudinal displacement for a rod can be written as
u(x) = Ae-ik,x + Be-ik,(L-x) (5.1)
where A, B are constants determined from the boundary conditions on the element.
Note that the only restriction here is that the rod theory be a one mode theory; thus
it could represent the refined Love theory or a simple rod with elastic constraint.
The displacement end conditions for the two-noded element, shown in Fig-
ure 5.2(a), are
The functions gl (x) and g2(X) are the frequency-dependent rod shape functions.
Figure 5.3 illustrates the frequency dependence of the shape functions. Note that
the functions occur in pairs, where one member of the pair is the "mirror image"
of the other.
The significance of the shape functions is that the complete description of the
element is captured in the two nodal degrees of freedom u1, U2. That is, we saw in
the previous chapters how the forces and other quantities are obtained by differ-
entiation of the displacements, now since the shape function is predetermined for
a particular structural element, we see that this formulation reduces the unknowns
to just the nodal values.
Since the spectral element can be very long then the ability to compute the
response between nodes is necessary. This can be done conveniently using the
shape function and the nodal values. Thus, the displacements and member forces
are obtained simply as
U(X) gj(X)Ul + g2(X)U2
Fcx) E~A[g~(X)Ul + g~(X)U2j (5.4)
5.2 Spectral Element for Rods 153
[X/L]
.0 .2 .4 .6 .8 1.0
Figure 5.3: Frequency dependence of the rod shape functions; gl (x) top. g2(X) bottom.
(5.5)
The shape function g, (x) is frequency dependent and is complex even in the simple
rod case with no damping.
The dynamic stiffness of the two-noded element can now be written in the matrix
form
(5.6)
154 Chapter 5. The Spectral Element Method
This is shown plotted against the lower frequencies in Figure 5.4. In general.
the dynamic stiffness is complex but in the special case of a simple rod with nO
damping. the matrix is real-only. This is the case shown in the figures.
f--=====-----
1.5
1.0
.5
.0 - - dynamic stiffness
--..... -- conventional
-.5
-1.0
Frequency [kHz]
-1.5 ~........-'-'-'-.......................~~.......-'-.....................................................L........................
O. 1. 2. 3. 4. 5. 6. 7.
Figure 5.4: Comparison of conventional and spectral element stiffnesses. Normalized as
kijJ(EAJ L).
For comparison. the form for the conventional element with the consistent mass
matrix [40] can be written as
A EA [
[k ] = -
L -1
I -1 ]_
1
pALai
6
[2 1]1 2
It is worth nothing that this equation is recovered exactly in the limit as w (and
hence k l ) goes to zero.
The stiffness relation for the throw-off element is simply
(5.7)
Note that it is complex and therefore the structural stiffness matrix is also complex
after assembling. This implies that the system experiences dissipation - a not
unexpected result since the throw-off elements conduct energy out of the system.
There is no corresponding conventional element.
A special characteristic of the two-noded spectral elements is that they have
a very large dynamic range which. in the above. is easily identified as coming
from the term in the denominator. This may be disconcerting initially. but the
same behavior is actually implicit in the conventional finite element formulation.
To elaborate on this. look at Figure 5.4; the conventional stiffness is monotonic in
frequency and kll • for example. goes through zero only once. The spectral dynamic
stiffness. On the other hand. has many zeros as seen from the broader frequency
plot of Figure 5.5. A stiffness that has such wild behavior may seem strange but.
as we show next. this is also implicit in the conventional formulation.
5.2 Spectral Element for Rods 155
Normalized stiffness _ _ 11
.......... 12
1.
Frequency [kHz]
, , , , , , , , , , , , , , , I , , , , I , , , , , , ! , I , , , , ! I
Replace this by two elements each of length L and assemble them (further details 4
on assembling is given later). The resulting stiffness relation is
{ Ft}
~ ,. L
2EA [ (1 - ~~2)
-(1 + 24~2)
-(1 + p~2)
(1 - q~21 -(1 +
0
p;2)
] { UI }
~e
F2 0 -(1 + 24~) (1 - "i2~ ) U2
The middle force is zero because there are no applied loads at that node. Now
u
solve for e in terms of the other displacements and remove it from the equations
to give the reduced system
{ FJ} EA
F2 = T (1 -
1
-h~2)
[(1 - -&~2 + t~2~4)
-(1 + ~~2)2
-(1 + ~~2)2 ] { UI }
(1 - -&~2 + 1~~4) U2
The behavior of kll' for instance, is such that it crosses zero near kL = 1.57
compared to kL = 1.73 before. This is closer to the spectral value of 7r /2 = 1.57.
What is more important, however, is that it has a second zero and it also goes
through an infinity at kL = ,JU = 3.46. Thus the apparently odd behavior of
Figure 5.5 is also implicit in the conventional formulation - if a sufficient number
of elements are used. The spectral approach is equivalent to an infinite number of
conventional elements.
The significant difference. from a practical point of view. between the two for-
mulations is that the number of rod segments in the present formulation need only
coincide with the number of discontinuities. Thus. only one element need be used
per uniform segment. This can result in an enormous reduction in the size of the
matrices to be solved.
156 Chapter 5. The Spectral Element Method
A Simple Example
Before we formally develop the assemblage procedure, it is useful to consider the
following simple problem: A semi-infinite rod, fixed at one end and impacted as
shown in Figure 5.6. We will model this as two elements where the second element
is considered as extending to infinity.
>.
:g 10
o CD ® •
®
~ 0
Time [J.ls]
! I ! ! I ! ! I ! ! It, I ! I
EA iklL [ 1 + e- i2kJL
T (1 - e-i2kJL) _2e- ikJL
EA[ik!l{U2}
An important point is that both elements share the same nodal displacements at
Node 2 and therefore continuity (compatibility) of displacements is automatically
satisfied. Equilibrium at the (imaginary) joint where the applied loads act, gives
Node 1: Node 2:
or in matrix form
Substituting for the nodal forces in terms of the stiffness matrices gives
Note that the boundary conditions to be imposed are quite sitnilar to those of the
conventional element and are imposed in the usual way [40]. That is, since UI = 0,
we can delete its corresponding rows and columns to give the reduced stiffness
relation
5.2 Spectral Element for Rods 157
The reduced stiffness matrix gives nonzero behavior at Node 2 which is determined
to be
{U2} = i2EIAkl [1 - e-i2kIL]{p}
This represents two pulses shifted in time by the amount needed to travel a distance
2L. The magnitude of the pulse is one-half that of the input.
The velocity response at the impacted node is shown in Figure 5.6. The meaning
of the figure is that the initial pulse causes a forward-moving compressive wave and
a backward-moving tensile one. This latter pulse, on reflection from the fixed end,
retains its sign and is picked up later as a compressive pulse (with negative particle
velocity) at the impacted node. What is particularly significant about this example
is that although the two-noded element exhibits singularities in the stiffness, the
resultant structural stiffness (in this case) does not exhibit any such singUlarity.
The term inside the square bracket is zero since gi(X) satisfy the homogeneous
equation. Particularizing gi to its two cases, and realizing that gl (0) "" I, gl (L) = 0,
g2(O) - 0, g2(L) = I, we get that the first term is the axial force at each end. The
two sets of equations are
But since g~(L) = -g;(O), we recognize the square matrix as the dynamic stiff-
ness for the rod element. Consequently, the stiffness relation for the rod with a
158 Chapter 5. The Spectral Element Method
In other words, a known distributed load can be replaced with an equivalent set
of nodal loads. In this way, the matrix formulation described so far need not be
changed.
Shape Functions
We write the general transverse displacement for a Timoshenko beam in the form
v(x) RtAe-iklx + R2Be-ik2X - RtCe-ikl(L-X) _ R2De-ik2(L-X)
~(x) = Ae- iklx + Be- ik2X + Ce-ikl(L-x) + De- ik2 (L-x) (5.9)
The functions gi (x) are the frequency dependent beam shape functions and are
given by
where l:!. == (RI - R2). The complete description of this element has now been
captured in the two nodal degrees of freedom VI and ~I. Note that, in all, there
are four shape functions even though there are only two degrees of freedom. This
comes about because, just as for the amplitudes, the shape functions are related
through the amplitude ratios.
The development of the two-noded element, shown in Figure 5.2(b), follows
along the same lines. The end conditions on the element are
where {t/>}i are the respective amplitude ratios. So as to make the matrix notation
a little more accessible, we will take the developments of the rod as the archetype
and use its notation except presented as matrices. The displacement is written as
{A} == { ! }, {B} == { ~ },
We have defined the amplitude ratio matrices as
The matrix [ ell ] is referred to as the modal matrix. It is a fully populated matrix
r
of order [2 x 2] and typically is not symmetric. The matrix e(x) J is diagonal and
is the unit matrix when the argument is zero.
We wish to replace the vectors {A} and {B} in terms of the nodal displacements
at x = 0 and x = L. That is,
-[eIlB]re(L}J
-[eIlB1re(O)J
i }-
Solve this as a system of matrices to get
{ ~ }- [G 1{ [G ][u1 (5.11)
where {lV}T[ G] and (N}T r i, J[G] give the shape functions for 4>(x) and vex),
respectively. The {4 x I} vector {N} is a function of x given by
(5.13)
Figure 5.7 illustrates the frequency dependence of the shape functions. The shape
functions occur in pairs where 83 and 84 are the mirror images of 8, and 82,
respectively.
5.3 Spectral Element for Beams 161
O~~-----;~~~:=---T~~~----~~~~~~O~ZHZ
·····1 kHz
-1 -10 kHz
.0 .2 .4 .6 .8 1.0
Figure 5.7: Shape functions for beam; g, (x) top, g;(x) bottom.
The nodal loads are expressed in terms of the displacement degrees of freedom
through the relationships for the structural resultants
~ d¢ ~ dcp
2~
2~
M(x) = +EI dx ' V(x) = - E I - - plw cp
dx 2
Evaluation of these at x = 0 leads to the dynamic stiffness as
{ -~(O)
-M(O)
}={ ~\ }=EI[~\1
M\ k2\
(5.16)
This symmetric throw-off stiffness matrix is complex and therefore the structural
stiffness matrix will also be complex after assembling. Keep in mind that a complex
stiffness will exhibit a damped behavior in the time domain. In actuality, what is
happening is that the bar is conducting energy out of the system and this appears
as a damping effect.
The two-noded or finite length spectral element looks the same as the conven-
tional element except that the mass is distributed instead of being concentrated at
the ends. Although it is not required in the following, it is assumed that there are
no applied loads between the ends. The development will be summarized using
the more general matrix notation.
162 Chapter 5. The Spectral Element Method
Consider a beam waveguide of finite length L with loads applied only at its
ends. We write the shape functions as
where {u} is the collection of four degrees of freedom. This leads to the moments
and shear expressions
(5.17)
+EI{N"(o)f +plw2{N(O)friIJ
-EI{N'(O)f
-EI{N"(L)}T - plw2{N(L)}T riIJ
+EI{N'(L)}T
The relation between the loads and degrees of freedom can be written in the form
where
5.4 General Frame Structures 163
and the remaining terms are obtained using symmetry. Some of these stiffness
terms are shown plotted over the low-frequency range in Figure 5.8.
12.~=......."",,",,"~""
B.
b:::::::==::::====~~~2=~~~
4. I=' __~~~2;2___
O. 12
-4.
- - dynamic stiffness
-B. 11
" "" conventional
-12.
-16.
Frequency [Hz)
-20.~~~~~~~~~~~~~~~~~~~~~
This figure also shows a comparison between the spectral and conventional
dynamic stiffnesses where the form forthe conventional element with the consistent
mass matrix [40] is
12 6L -12 22L 54
A EI [ 6L 4L2 -6L 6L] _
2L' w2 pAL [ 156
22L 4L' 13L -13L]
-3L'
[ k ] =L3
- -12 -6L 12 -6L 420 54 13L 156 -22L
6L 2L' -6L 4L' -13L -3L' -22L 4L'
It is apparent that they both have very similar limiting behavior for small frequency
- the slight differences are due to the use of Timoshenko beam theory for the
spectral element. In fact, when the Bernoulli-Euler beam theory is used, they
agree exactly to order w 2 terms. As can be seen from Figure 5.9, however, for
higher frequencies there are significant differences because the conventional form
behaves monotonically while the spectral element exhibits multiple zeros. For the
conventional element to have these zeros it is necessary to piece many of them
together as was demonstrated for the rod element.
100.
10.
1.
Frequency [kHz]
! , ! , , ! , !
" " !! ! , ! , , ! , ,
o. 1. 2. 3. 4. 5. 6. 7. 8.
Figure 5.9: Dynamic stiffness at higher frequencies normalized as kll I(E I I L 3),
k 12 /(E I I L 2), k 22 /(E I I L).
The global stiffness relation is between these two sets of global quantities.
The member forces are related to the member end displacements by the following
partitioned matrices:
lu}
kab kae kad
[ k""
{F} = ~ba kbb kbe kbd
kea keb kce ked
kda kdb kde kdd
where a bar implies that the term is referred to the local member axes and each [k]
is a [3 x 3] matrix. The expanded form of the first quadrant of the local stiffness
5.4 General Frame Structures 165
matrix is
kfl 0 0 0 0 0
AB AB
0 kll 0 0 0 k12
AB A8
[ kaa
kba
kab
kbb ] 0
0
0
0
kll
0
AB
0
AS
kll
k12
0
AB
0
0
0 0 k12 0 k22 0
AB AB
0 kl2 0 0 0 k22
The rest ofthe stiffness matrix is filled out just as for the static case. The superscripts
R, B, and S refer to Rod, Beam, and Shaft, respectively.
The matrix required to transform one Cartesian coordinate system rotated with
respect to a second Cartesian coordinate system, sharing a common origin, is
where Lx, m x, and n x are the cosines of the angles that the local i axis makes
with the global x, y, z axes, respectively, as shown in Figure 5.10. The problem
here is to find the remaining elements of [R], as lx, m x , and nx define only the
orientation of the member i axis. A space frame member may have its principal
axes in general directions and this too is to be taken into account. To visualize
how the rotation matrix is constructed, consider three successive rotations from
the structure axes to the member axes. The first rotation is about the z axis and
the second about the y axis. The third transformation consists of a final rotation
z
through the angle y about the i axis, causing the y and axes to coincide with the
principal axes of the cross section. Multiplication of the three successive rotations
gives (with D ~ )1 - n;)
Lx mx
-m x cos y -Lxnx sin y Ix cos y - mxnx sin y nx ]
[R] ~ [ D . D
Dsiny
mx sin y - lxnx cos y -Lx sm y - mxnx cos y
Dcosy
D D
This is the general form of the rotation matrix. It is expressed in terms of the
direction cosines ofthe member (which are readily computed from the coordinates
of the joints) and the angle y, which must be given as part of the description of
the structure itself. A further exposition of transformations relevant to structures
can be found in References [126, 3].
As the translational displacements are independent of rotational displacements,
the transformation of a nodal vector from the local to the global coordinate axes
system uses the matrix
member
~'"'-----+-------,,_ x 9 loba I
z
fx
Figure 5.10: Direction cosines for a general member.
{F}=[k]{u}, (5.19)
where [ k ] is a square, symmetric, and complex [12 x 12] matrix; the throw-off
members are also put in this form by augmenting with zeros.
(5.20)
where [k ] is called the dynamic structural stiffness matrix. Note that since the
individual augmented stiffnesses have many zeros, the resulting matrix is banded.
Furthermore, although each element has different forces, they all share common
displacements at the nodal points. That is, continuity of displacements from mem-
ber to member is automatically assured. A more detailed description of this as-
sembling procedure for static problems is given in References [3, 40].
External forces or displacements may be prescribed at the nodes ofthe structural
system. However, a force and corresponding displacement cannot both be given at
the same node. The global system of equations must be rearranged and renumbered
5.4 General Frame Structures 167
to account for these boundary conditions. If the known nodal displacements are
zero (such as at a fixed boundary) then we simply "scratch the rows and columns"
corresponding to these degrees of freedom and the remaining stiffness matrix
is then called the reduced structural stiffness matrix relating only the unknown
displacements to only the given external loads.
The stiffness matrix [K] has the symmetry and bandwidth properties of the
corresponding static stiffness matrix and in general, is complex.
or
the latter form showing, in essence, that we compute the system transfer function.
Consequently, the force history need not be specified (only its location) at assembly
time and therefore the same structural solution can be used for many different input
histories.
A big advantage of the spectral formulation is that since there are a small number
of nodes, then it is feasible to store the complete solution. For example, a plane
frame with 11 members can have, say, 7 nodes, giving a system size of 21. The
storage requirements based on a 1024-point FFT transform and 8-byte words is
about
21 x 512 x 8 x 2 = 168kB
This very small size for the complete solution means it is now feasible to add a post-
processing capability that allows the reading of the nodal solution and obtaining
the member loads or distributions at any desired point without having to redo the
complete solution. This is a very useful feature since the elements can be very
long. Thus the procedure for obtaining the spectral response is to first obtain the
generalized displacements at all the nodes and store them to disk; then, as part of the
postprocessing, the response at particular nodes or arbitrary locations are obtained
by reading the appropriate nodal displacements and performing the conversion.
168 Chapter 5. The Spectral Element Method
Two key points to note are that accuracy is improved for the spectral approach
by increasing N (the number of terms in the spectral representation), and that
execution time is essentially linear with N. This contrasts with the conventional
approach which requires, simultaneously, an increase in the number of time steps
and an increase in the system size. Depending on the bandwidth change, the latter
can have a very severe effect on performance. Furthermore, since the computations,
at each frequency, are completely independent of each other but execute the same
instructions, this approach is eminently suited for mUlti-processor computers of
the single instruction multiple data (SIMD) type. See References [26,27,28] for
examples on a massively parallel computer using 16,384 processors.
Curved Joint
It was shown in Chapter 3 that the analysis of curved beams introduced coupled
modes and complicated the analysis. Consequently, a great deal of time and effort
must be expended in obtaining the spectrum relation. The present example shows
an alternative approach - the curve is modeled as a collection of straight elements,
each of which has a simple description. The complexity of connecting them all
together is handled by the matrix methodology. To illustrate what happens, we
will consider the sequence of bends shown in Figure 5.11. They correspond to
increasing smoothness in a 90° bend.
p p p
-+- -+- -+-
L L L L
Figure 5.11: Sequence of bends of increasing smoothness.
The responses at the impact node are shown in Figure 5.12. The sharp bend
shows that the incident pulse is reflected as if it met a free end, and very little
of the energy is transmitted. When a 45° transition is introduced, distortion is
5.5 Structural Applications 169
observed in the reflected signal. This indicates mode conversion to flexural waves
as well as more transmission of energy. It is also apparent that there are multiple
reflections within the 45° member. The final case is a fairly smooth somewhat
circular bend. This exhibits the smallest reflected initial pulse and a frequency
analysis of it shows substantial filtering. The smoothness of the reflected portion
of the signal indicates that the higher-frequency components are passing through
unaffected.
- longitudinal
....... flexural
, I t , t I , , ~im~ [~)
o. 1000. 2000. 3000. 4000. 5000.
Figure 5.12: Velocity response at the impact site for various bends.
It should be pointed out that while this matrix approach is useful for finite bends
it would not do at all for, say, a helical spring. The analysis of Chapter 3 is then
necessary since it would allow the study of very long bends.
Modal Analysis
While the spectral approach is good for dynamic wave propagation problems,
it also has its use in conventional vibration analysis. This section shows a brief
example.
There is a difference in doing a modal analysis and doing the forced response
analysis. For the modal analysis, the frequency spacing are entirely arbitrary. In
fact, only one frequency need be analyzed if that is all that is desired. For the forced
response, however, the frequencies are dictated by the requirements of the inverse
FFf transform. For finite structures such as the portal frame shown in Figure 5.13,
the propagating pulse has many reflections. Indeed, a true elastic (nondissipative)
system would continue this ad infinitum without diminish of amplitude. Unfor-
tunately for the FFf formulation, this would mean components propagating into
neighboring time windows (as explained in Chapter 1). This can be alleviated
somewhat by including material damping and having a very large time window.
The net result of this is that a finer frequency resolution must be used and this is at
170 Chapter 5. The Spectral Element Method
the expense of a large N. The problem is less severe in more complex structures
because the energy from a pulse input is naturally being diffused and will not be
trapped so readily. The portal frame is our the example structure because it is one
often used to demonstrate the vibration of structures.
p
-~===~
L
L
-x-disp
....... y-disp
-- z-rotn
• ! ! , I I ,
Frequency [Hz]
, I ! , ! , I
The spectral element method could be used to perform a modal analysis; one
approach simply uses frequency scans such as Figure 5.14. Once the resonances
have been located then precise values can be obtained by zooming in on the fre-
quencies. Doing this is only a matter of inputting a narrower range of frequencies
with a finer increment. The ability to zoom is especially useful for the higher res-
onances because it does not require a finer element model nor does it require the
determination of the lower resonances. This is a decided advantage over the con-
ventional finite element approach. The shapes are reconstructed using the shape
functions at the determined frequency.
5.5 Structural Applications 171
!p - L
2
Figure 5.16 shows the lengthwise velocity response when the truss is impacted in
this direction. It exhibits some dispersive characteristics but predominantly seems
to behave like a rod. Actually, looking at the transverse displacement, we see
that it is close to the Love theory and the type of dispersion (affecting the higher
frequencies) is typical of that model.
x=o
x=L
---:..:-=------' f\
.. ..: .--..--.,.'7,,;...:..""
.. ----....---...,............-~.......- - -...-.....~-
x=2L
_ _ _ _ _ _ _--J. .. / " " '.. "
Figure 5.17 shows how the axial force varies away from the impact site. AlthOugh
not shown, it is antisymmetric in the loads with predominant tension on the top and
predominant compression on the bottom. In loose terms, these axial tensile and
compressive loads give rise to a net moment on the cross section which is exactly
how a beam would behave if struck in a similar manner. Furthermore, the variation
of axial force with distance exhibits the dispersive characteristics of a beam with
the load at the center experiencing the maximum value and longest duration.
172 Chapter 5. The Spectral Element Method
x-4L
t ,
Comments
There are many variations on these problems that can be pursued as well as many
phenomena that can be investigated. For example, it is apparent from Figures 5.16
and 5.17 that a good deal of energy is trapped in the horizontal members near the
impact. So an interesting question is how to design the joints so that this trapping
does not occur and the energy can diffuse throughout the structure. Such a design
would tend to minimize local impact damage. On the other hand, some structural
designs might require the impacted area to trap as much energy as possible so as
to minimize damage to the remainder of the structure.
An aspect of these problems not covered here (but should eventually be ad-
dressed) is that of describing the global flow of energy through the system. Keep
in mind that while the spectral approach is capable of describing (in detail) the
motion of every particle in the structure, the fundamental type of information
recorded are the spectral components of the nodal displacements. It would be of
interest, therefore, to investigate how this global information (which is already in
a convenient reduced form) could be made to give a "big picture" view of energy
propagation. This integrated level is somewhat similar to the modal analysis view
of structural dynamics except that, fundamentally, it is based on a transient event
and not the long-term steady-state behavior. In other words, the inter-component
phase relation is an important part of the picture.
Some preliminary aspects are considered in Reference [83]. The difficulty seems
to be in identifying appropriate global quantities that have heuristic values for
illustrating the flow of energy. The ideas developed in the next section on the
frequency domain formulation of some stationary energy principles could also
playa role.
5.6 Waveguides with Varying Cross Section 173
A Variational Principle
Spectral analysis converts a dynamic problem into a series of pseudo-static prob-
lems with the frequency as a parameter. That is, time no longer appears explic-
itly but is replaced instead by phase relationships among the variables. This is
manifested by all relevant variables being complex. From the frequency domain
statement of (dynamic) equilibrium, the definition of work and internal energy
quantities can be arrived at in a similar manner as is done in the static approach;
we give only a brief summary here.
Borrowing on the definition of virtual work for the time domain [60], we define
the work done by a force P in experiencing a virtual displacement 8ii as
8W == P8ii
This relation resembles the usual definition of virtual work but with the significant
difference that both P and 8ii are complex, frequency dependent quantities. Con-
sequently, the work is also complex. This work definition is also different from
the definition of power as used in power flow analysis [75] where the complex
conjugate of the displacement is used.
174 Chapter 5. The Spectral Element Method
We will use the example of the I-D rod to motivate the development of the
general equations. The loading equation is a force balance equation, hence the
total work done by it in the displacement flu (for a rod segment of length L) is
The last term is equivalent to work done at the ends of the rod volume, that is,
since the end loads are PI = - F(O) and Pz = F(L). We now view the integrated
load term as an equality between external applied loads and internal energies in
the form
Define the potential of the applied loads as the work done by the distributed loading
£l(x) and the end loads PI, Pz as
(5.21)
where llK is the strain energy, {;M the inertia energy. Hence, the integrated ex-
pression for the loading has the equivalent forms
or (5.23)
where if is called the total potential. This is our stationary principle in the frequency
domain; it says that the total potential of the external loads and internal energy
achieves a stationary value at the (dynamic) equilibrium state.
A general plane body can be described by two displacement components {u} T =
{u, ill, three stress components {a}T = {axx, a vv • a x \"}, and three strain components
{E} T = {E xx. Evy , Ytv}. If {q} is the vector containing distributed transverse loading
5.6 Waveguides with Varying Cross Section 175
and {p}T is the applied load vector, then the virtual work and energy terms can be
defined as
1 {af 8{E}dV
and the principle of virtual work (or the principle of stationary potential energy)
is written as
8[Us + UI + V] = 0
where V is the potential of the applied load, Us is the strain energy, UI is the
inertial energy, and the integration is over the volume V. In order to make the
equations more general, a viscous damping factor 1/ is also introduced explicitly
into the inertia energy expression. The word "energy" is not to be taken literally;
we use it only by analogy with the time domain counterparts - if need be, we can
interpret the approach as basically a Galerkin procedure [6]. A complementary
approach to defining the variational principle is given in Reference [52]. Also
worthy of note is the discussion in Reference [112] where a vibration problem
with viscous damping is transformed into an undamped vibration problem with a
negative elastic constraint.
(5.25)
where N is the total number of degrees of freedom. To ensure that the character-
istics of the system are taken into consideration, the Ritz functions must satisfy
the essential (geometric) boundary conditions. This assumed displacement field is
substituted in Equation (5.24) and the minimization of the total potential of exter-
nal loads and internal energy with respect to unknown coefficients will yield the
176 Chapter 5. The Spectral Element Method
(5.26)
with (; = Us + U/. Here, {u} is the displacement vector and [K] is the dynamic
stiffness matrix. The procedure as outlined is applicable to any structure; we will
apply it specifically to the simple structure representing a tapered element.
If the assumed displacement field is the exact solution for the governing differ-
ential equation, the Ritz method will yield an exact dynamic stiffness matrix [53].
Otherwise it is approximate, but the accuracy of the representation can be improved
by increasing the number of degrees of freedom. This is equivalent to increasing
the number of terms in Equation (5.25) or increasing the number of segments. In
other words, the length of the element, unlike for other spectral elements, is now
a factor in the performance of the element.
~~
V2
CD ® Dh(X)
EI(x), pA(x), L ~ M2
~2 b
Figure S.18: Nodal loads and degrees of freedom for a tapered Timoshenko element.
Let the displacement solution in the tapered beam be approximated by the uni-
form beam solution
where k/ and k2 are the two wavenumbers obtained by solving the characteristic
equation
5.6 Waveguides with Varying Cross Section 177
iGAoKlkl
RI= ,
GAKIk? - pA ow2
The cross sectional properties Ao and 10 are at some reference location. These
deformations can be made into Ritz functions by replacing the coefficients with
the nodal degrees of freedom. The displacements and rotations can then be written
in a compact matrix form as
(5.28)
where {N}T[G] and {Nf[id[G] give the necessary Ritz functions for 4> and v
and the explicit form of the [4 x 4] matrix [G) is given earlier in Equation (5.11).
The frequency domain strain energy and the inertial energy for the Timoshenko
beam can be written as
where we have retained the adjustable parameters K I and K 2 • The potential energy
of the system with only nodal applied loads is given by
(5.30)
where {F} contains the nodal loads rVI, MI, V2, M2}. The assumed displacement
field from Equation (5.28) is then ~ubs!itute~ into the above two energy expressions
and the sum of the total energy (U = Us + U I ) and the potential energy (V) is then
minimized with respect to the generalized degrees of freedom (u;) to obtain the
dynamic stiffness relation. This can be written symbolically as
A A 2
[k]=[kD]-w[mD]
and the square matrices [tD] and [mD] (which, in the limit oflow frequency, have
the interpretation of static stiffness and mass matrices, respectively) are given by
+[ G f[lLGA(X)KdN2HN2}T dx][ G)
After differentiating {!V} and simplifying, the dynamic stiffness matrix can finally
be written as
The square matrices [ t] and riA] contain the integrals arising out of the cross
section variation and have the following form:
where
e-i(k,+h)x e-ik,L e-ik2Le-i(k,-k2)X ]
e- 2ik ,x e-ik , L e i(k,-k2)x e-ik2L
e-2ik ,(L-x) e-i(k,+k,)(L-x)
e- 2ik, (L-x)
and lex) = I(x)/ 10 , A(x) = A(x)/ Ao. If A(x) and lex) are constant, then the above
leads to the previously derived stiffness, albeit in a different form.
In dynamic problems we do not have a simple separation of mass and stiffness;
the above designations are used nonetheless because in the limit of frequency the
above do indeed show the separation.
where E represents the slope or strength of the taper. This gives the variation of
cross sectional properties as
5.6 Waveguides with Varying Cross Section 179
All the integrals in [ iI ] (except i/l3 and i I24 ) are of the form JoL (I +E X / L)3 eax dx,
where a is a complex number. The explicit form of this integral is
{L (
Jo 1+ L
EX)3
eaxdx
Similarly, all the integrals in [I~] (except ~13 and i A24 ) are of the form JoL(l +
EX / L )e ax dx which can be written explicitly as
l o
L ( EX) ax
1+-
L
1
e dx=-[(l+E)e
a
aL
- l ] - - E[ eaL -1]
a2 L
CD
+
®
h=25mm
250mm
Cf (
®
50mm
0
h=75mm
®
(a)
Figure 5.19: Geometry for the tapered beam example. (a) Spectral element model, (b) part
ofFEMmesh.
In order to study the effect of length on the performance of the element, a spectral
element model with four elements are used. With reference to Figure 5.19, Element
(3-4) is tapered with depth increasing from 25 mm to 75 mm over a length of 50 mm.
This gives a slope of E = 2.0. The ends of the tapered element are connected to
180 Chapter 5. The Spectral Element Method
deep unifonn spectral elements. The equivalent finite element model (with the
tapered region shown in Figure 5. 19(b» had eight modules through the thickness
and the throw-off regions modeled as a rectangular mesh of length 625 mm.
The force is input transversely at Node 2 and the transverse velocity (iJ) at this
location recorded. For a single tapered element, the solution shown in Figure 5.20
indicates that the approximate nature of the element causes some spurious reflec-
tions. By modeling the tapered region with two elements, the response curves
(also shown in the figure) indicate that the finite element solution and the spectral
element solution virtually overlay each other.
- - Ritz solution
......... FEM
! I !
, , , ! ! ,
.--...,. , , Ti,m~ [~l ,
o. 100. 200. 300. 400. 500.
These results establish that good approximations to the true response (as repre-
sented by the finite element responses) can be recovered by using multiple tapered
elements. This shows that, as expected, the smaller the element size the better the
convergence for a given steep taper. References [53, 54] consider other situations
including longitudinal behavior.
(a) (b)
Figure 5.21: Super-element connected to waveguide spectral elements. (a) Arbitrary joint
meshed with triangular elements, (b) reduction of connection nodes.
where [K] is the stiffness matrix, [C ] is the damping matrix, [M] the mass
matrix, {u} the vector of nodal degrees of freedom, and {P} the vector of applied
loads. The spectral form of the equations of motion are
where iii = (J)2 - icx{J) and [K] is the dynamic stiffness matrix of the region. To
be consistent with the waveguide spectral element formulation, we have taken the
damping as Rayleigh damping [40] in the form [ C ] = cx[M]. The square matrices
are of size [N x N].
The dynamic stiffness matrix [K] is now partitioned according to the connection
and internal degrees of freedom as
(5.34)
182 Chapter 5. The Spectral Element Method
and {u e } and {ud are the vectors of connection and internal degrees of freedom,
respectively. Solving the second of Equations (5.34), we get the following for {u;}:
(5.35)
Using this in the first of Equations (5.34), we obtain the stiffness relation for the
super-element as
(5.36)
This equation relates the connection forces {Pc} to only the connection displace-
ments {u c }; all the internal degrees of freedom have been condensed out and in
this way it resembles an element. It is important to realize that so far we have not
introduced any approximation to the super-element stiffness; this contrasts, for
example, with the method of Guyan reduction [57]. The size of [Kss] is [c xc],
where c is the total number of connection degrees of freedom and is expected
to be considerably smaller than N, the total number of degrees of freedom. The
computation of [Kss] requires computation of [Kiir 1 (which is nearly of the order
[N x N]). Doing this efficiently is absolutely essential since it is done for each
frequency component. Moving this computation outside the frequency Do-Loop
is the main difficulty to be encountered and we give it special attention next.
(5.37)
where {¢} j and A j are the eigenvector and eigenvalue of the jth (undamped)
mode, respectively. This was arrived at by using the orthogonality relations
[<t>][M][<t>f = 11 J, [<t>](K][<t>f = I AJ
5.7 Spectral Super-Elements 183
We thus have the inversion of [Xii] while retaining the frequency explicitly. Further,
the product {<I> 1j {<I> }) need be performed only once outside the frequency Do-Loop.
The summation in Equation (5.37) is performed over the first m modes. But if
we follow what is usually done in vibration analyses and include only the lower
modes then we get very poor estimates for the inverse of [Xu]. On the other hand,
if we include all the modes, then the computational cost is a big dis-incentive to do
an accurate modeling of the joint region. We will accelerate the convergence with
respect to the number of modes by using a suggestion made in References [78, 119].
Noting that, without approximations,
1 1 x2 X4
-- = - + - + -::----::-
a - x2 a a2 a 2(a - x 2)
25 modes
o DOD o ODIE o
10 modes
-------ao------103-<~-------El-EHD---G---
V-
5 modes
1. [ ------- Inverse Exact
- Inverse Modal
o Modal frequencies
Freq [kHz]
O. I~"__...............__!._~-'--_'____'____"_'--_'_-'-........__'_.L..__'__'_I---1.1"'--"'--1'
the input, even less than five modes may be acceptable. The main achievement
of Equation (5.38) is to give the exact inverse at low frequencies; in this way, the
procedure is essentially that of the "mode acceleration method" commonly used
in modal analysis [19].
Using the superscript S for the equivalent spectral super-element, the displacement
field across the interface can be written in the local (barred) coordinates as
(5.39)
5.7 Spectral Super-Elements 185
where Yi is the distance of the ith node measured from the center of the connection.
This reduction process is shown schematically in Figure S.21(b). The resulting
equivalent spectral element stiffness matrix is of size [4C x 4C] where C is the
number of connecting interfaces (there are three indicated in the figure).
Consider one face of super-element connection oriented at an angle e to the
global axes. (Note that the super-element may have many such faces.) For the
super-element, the unbarred and the barred displacements are related by the usual
rotation matrix
This equation is for a single interface; if there are C connections, then the overall
transformation matrix [ T ] will look like
Tl 0 0 0
o T2 0 0
[T] ~ 0 0 T3 0
o 0 0 Tc
Using these transformation relations, finally allows us to obtain the stiffness re-
lation for the super-element in terms of an equivalent waveguide element. This is
written symbolically as
~ W _ ~ T
[Kss] = [T ][Kss][ T] (5.41)
In this form, the super-element can be assembled along with the other spectral
elements as if it were just another waveguide spectral element.
(5.42)
186 Chapter 5. The Spectral Element Method
where
[D] [Hd[MuHH1f
- L{G2}j[A2(Ajl_ w 2
J ]
J H 2}j - L{H2}{A~(Ajl_ W 2)JG 2}j
] ]
and
The matrices [A], [B], [Hd, and [Gd are frequency independent and hence eval-
uated only once outside the frequency Do-Loop. The vectors {G 2 lJ and {H2 }j,
which are involved in the generation of matrices [C], [D], and [E], are evaluated
and stored at each of the first m modes. Details of the computer implementation
of these equations can be found in Reference [53].
When the lumped mass model is used in the formulation, matrices [D] and [El
vanish and all other matrices simplify considerably. The stiffness becomes, simply,
(5.43)
where
with [Gd = [Kc;][Kul- 1 and {G 2 lJ = [Kc;]{q,}j. The matrices [Kl and [Ml have
a simple interpretation in terms of the effective stiffness and mass of the region
representing the super-element.
The matrix [ <i> ] is essentially what makes the spectral super-element unique.
This matrix has built into it the dynamic behavior of the region. This can be seen
in Figure 5.23. Here a segment of a waveguide 125 mm long and 25 mm deep
was meshed and replaced with a super-element and the figure shows the effect of
5.7 Spectral Super-Elements 187
1000~ _JJ~---f\==.. =
100. .. 8,9,10
10.
1.
Freq [kHz]
t , , , 1 I, , I
the number of modes on the kll stiffness term. It is clear that beyond five or so
modes that the behavior of the super-element and the waveguide would be indis-
tinguishable if the excitation was band limited to below 40 kHz. The reason why
the stiffness plots do not change with the addition of every mode is because some
of the modes may only affect the flexural behavior, say, and not the longitudinal
behavior.
Angle Joint
This example explores the behavior of waves as they interact with an angled joint.
The joint in Figure 5.24 has a two-dimensional distribution of mass and elasticity
and therefore will cause the waves to behave in a two-dimensional manner.
(a)
P(t)
® 127mm ® 62mm ® ®
Figure 5.24: Waveguides connected by an angled joint. (a) Spectral element model, (b)
super-element mesh.
- Spectral element
........ FEM
....... Rigid joint
vl2vo ulim
i:, ==2000mmls
------------~------------V·-IV-·----~~
u/~ 0
Figure 5.25: Angled joint responses at the impact site and transmitted responses on oblique
member.
The spectral element model of the structure (shown in Figure 5.24) consists
of five elements with the super-element having three nodes or interfaces (5-4-3).
Elements (2-1), (5-6), and (4-7) are throw-off elements and all have the same
depth of 25 mm. The super-element mesh is also shown in the figure. It has 440
internal degrees of freedom and 42 connection degrees of freedom. The structure
is impacted transversely at a point 127 mm from the joint and both transverse and
longitudinal velocities are reported at two points: at the impact site and at a point
127 mm from the center of the joint along the inclined member. Responses were
also computed for the case where the joint was treated as "rigid."
Figure 5.25 shows the responses for all three models. While the results predicted
by the super-element and the conventional finite element virtually overlay each
other, the rigid joint modeling shows significant deviations. The same is also true
for the transmitted responses. This example clearly illustrates the error that is
introduced by treating the joint as rigid. It also demonstrates the ability of the
super-element approach to handle structures having more than two interfaces or
connections.
(5.44)
5.7 Spectral Super-Elements 189
where {u I S is the vector containing the spectral element displacements at all nodes
and (for the lumped mass case)
All the matrices involved in [L] have already been computed (as part of the global
solution) and stored on disk. These arrays are also frequency independent and
hence most of the products are performed only once outside the frequency Do-
Loop. Normally, the displacements are required at only a few selected nodes and
using the above form, it is possible to do this conveniently. Consequently, the cost
of recovery of internal degrees of freedom is minimal and can be performed on an
as needed basis.
In this final example, we consider a waveguide with a transverse crack extend-
ing to half of the total depth of 25 mm as shown in Figure 5.26(a). Alternative
approaches and background information on this problem can be found in Refer-
ence [91].
r( •
(a)
t
crack=12.5mm
( ,,+ PIt)
~ V
~ h=25mm
CD ® 127mm
® 25mm @ 254mm
®
Figure 5.26: Waveguide with an open transverse crack. (a) Spectral element model, (b)
super-element mesh.
(a)
- - Spectral element
-------- FEM
5[
u/~) O.
'!
O.
-1.
Vlit
Model
__M_od_e_II_ _ _ _ _ _ _ _ _~
Time IllS]' ! ! ! ! , ! ! , ! ! , ! ! !
Figure 5.27: Waveguide with a crack. (a) Normalized velocity responses at the impact site,
(b) normalized Mode I and Mode II stress intensity factors.
K ll :::: ~ VLlV
E---
2LlaB
where Lla is the element length at the crack tip, Fx and Fy are the crack tip forces,
and Llu and Ll v are the corresponding crack opening displacements. The results
are normalized with respect to Ko = .j7i(iamax , where a max is max(P(t)f A). These
normalized (absolute) values are also shown plotted in Figure 5.27. The Mode I
response indicates two major peaks, the first due to the incident wave and the
second due to the reflection from the fixed boundary. The mUltiple trailing peaks
are due to the generated flexural waves that propagated to and then reflected from
the fixed boundary. The magnitude of the Mode II response is relatively small;
what is interesting is that it occurs only in response to the flexural waves. All these
results match up very well with the 2-D finite element solution.
(a)
P(t) ! 2
• •
A f--- 381 mm -----I
B
I-- 381 mm---i 762 mm - - - - - i
•
254mm
(b)
1397 mm
Figure 5.28: Two structural systems used for force identification. (a) Simple beam with
boundaries at locations A and B, (b) complex structure.
In this section, we attempt to layout the pertinent issues involved when multiple
reflections playa significant role. This includes cases where the physical bound-
aries of the structure are relatively close, and the response becomes a complex
superposition of the incident and reflected waves which can no longer be sepa-
rately distinguished. The approach taken is applicable to structures in general, but
to help focus ideas, we concentrate on frame structures. This will also show the
power of the spectral element method. More details and examples can be found in
References [37, 38, 41, 85, 86].
where G represents the structure. Note that in this, and the following, u(t) can
represent any response - displacement, velocity, strain, and so on - in the ex-
perimental results to be shown it is the acceleration.
The basic problem of force identification is: given u(t) (perhaps imperfectly) and
knowledge of the structure G (again perhaps imperfectly) find Pet). We emphasize
the imperfect nature of the given information as an essential aspect (and difficulty)
of the problem. That is, the purely mechanical inverse of Equation (5.45) is not
really the issue (see Reference [102] for a selection oftechniques), it is the factors
that exacerbate the ill-conditioning of the problem that are the main concern.
The structures of interest are shown in Figure 5.28. The frame members are made
of aluminum [12.7 mmx 12.7 mm] in cross section. The data recording techniques
and the details of the experimental set-up of Figure 1.9 are as given in papers such
192 Chapter 5. The Spectral Element Method
as References [38, 85]. A force transducer with a steel spherical cap of diameter
15 mm constituted the impactor. Accelerometers placed at the locations shown
in the figure were used to record the signals with the data collected at a rate of
2/Ls, this was then smoothed using a five-point moving average. Since obtaining
force from acceleration (for beams and frames) involves a partial integration, the
responses were integrated in time to determine any trends; this trend was then
subtracted uniformly from the measured data. We will refer to this smoothed and
de-trended data as the experimental data.
. • l
······i~"\)i~VN·JWWv,VWVVWiWi\0VWN,VVVV!!VV\{i'
~1~ finite
.Q 0 " semi-infinite
~ .. o·.f\.. ·---··f:.[\j"\v...··.·--·---,--··----·----------------------------------.----------------------------------------
~-1 " • ,. _ _ sensor 1
:i. ------------ sensor 2
-t'- infinite
----J<.--.--------------------------------------------------------------------------------------------------------
I, , ! , I I! f I ! I , ! , ! , ! , , , , , ! ! , ! , , , , ! !
Time
II , !
[~]
! , !
Figure 5.29 shows the acceleration responses at two locations for some simple
beam structures. In each case, the position of the impact site relative to the sensors
remained consistent, with the difference being the addition of free boundaries (at
locations labeled A and B in Figure 5.28(a» to the problem. For the infinite case,
each response consists of a forward wave only which shows a typical dispersive
behavior. In the semi-infinite case, the addition of a single boundary at position
A gives a response which contains a similar incident wave followed by a single
reflection; these reflections can either be separated from or superimposed on the
incident portion of the response. In fact, the latter is more likely for dispersive
systems such as frames.
One of the effects of the reflections is to make the response window larger.
This is evident in the finite beam case where the multiple reflections (between free
boundaries A and B) cause the response to persist for a very long time. This raises
the issue of whether or not the data collected in any finite sample is complete. This
is a complicated issue, but it is apparent in the semi-infinite beam case that, from
a wave propagation point of view, the reflections contain redundant information
that is already available in the incident portion of the wave. The same is true
5.8 Impact Force Identification 193
1000
-sensor 1
...... sensor 2
r--
, , I I I ,
infinite
, F~e9 [k,H31
0 5 10 15 20 25
== L
N
Gn(w) G(tk)e- iWn1, (5.46)
k-l
= -A-
Gn
That is, we deconvolve the structure G from the response U to get the frequency
domain force P. An inverse Fourier transform on P gives the force history. The
power of the Fourier methods is that they establish a one-to-one relation between
the frequency components - each frequency component of force is uniquely
related to each frequency component of response. The effectiveness of the approach
is enhanced when G can be obtained directly as when using the spectral element
method as outlined in the earlier sections.
Figure 5.30 shows the magnitude of G for the same simple structures of Fig-
ure 5.29; in each case Grelates the force to acceleration. For the infinite beam, G
is a simple monotonic function and performing the deconvolution is a straightfor-
ward matter of division. Addition of a single boundary to the problem produces
194 Chapter 5. The Spectral Element Method
a G that can have zeros, and this makes division problematic. In other words,
when G is zero (or near zero) the response according to Equation (5.46) is miss-
ing information at those frequencies, hence the reconstructed P is also missing
the information. To elaborate, suppose we are dealing with an N = 1024-point
discrete Fourier transform, then P is represented by N = 1024 unknowns and u
has N = 1024 knowns. However, a subset of these do not give useful information.
Although they may be few in number, their effect on the reconstructed force would
make the force history appear meaningless.
The finite beam presents an additional difficulty. The presence of sharp spectral
peaks is the complementary problem to loss of information - here the response is
overly rich in some frequency components. The deconvolution division can be per-
formed but any mismatch between Uand Gwould render P (at these frequencies)
very suspect. We will refer to this as an over-sensitivity problem.
Equation (5.46) shows that if G has zeros at certain frequencies then uwill also
u
be zero at these frequencies. In other words, has no information about P at these
frequencies. The loss of information can be mitigated by adding noise near the
zeros. That is, we rewrite the convolution and deconvolution problems as
(5.48)
where the asterisk indicates a complex conjugate and R is a small amount of ran-
dom noise. (The addition of noise to the denominator does not, of itself, replace
the lost information, but rather it removes the potential singularity at those fre-
quencies where G becomes quite small.) Figure 5.30 suggests that the placement
of the sensor affects where (and how many) zeros occur. Hence when two sensors
are available, we can rewrite the above as
(5.49)
This will minimize the possibility of zeros or near-zeros occurring in the denom-
inator. Although not explicitly indicated above, the system response functions G
are normalized, in this way the sensors can even be of different type such as ac-
celerometer and strain gage. We will refer to Equation (5.49) as the frequency
domain deconvolution method.
The figure shows the quality of the reconstructed force using one and two sen-
sors; from this it is apparent that the combination of two sensors effectively replaces
any information lost by a single sensor.
The loss of information is a problem specific to the frequency domain methods
and does not appear in any of the time domain deconvolution methods such as
mentioned in Reference [102]. Consequently, the use of multiple sensors should
be the usual practice in frequency domain deconvolution problems.
We finish by looking at the relatively complex structure shown in Figure 5.28(b).
The responses shown in Figure 5.32 contain multiple reflections between the joints
and the free ends, and suggests that this problem embodies some of the difficulties
already highlighted. It also has the difficulty of imprecise structural modeling.
The imprecise modeling is mainly due to the joints -- this is a mechanical joint
of extended mass and elasticity. In reality, the joint behaves somewhere between
the rigid model of connected waveguides and a joint with some elasticity. In the
196 Chapter 5. The Spectral Element Method
Frequency [kHz]
Figure 5.33: The system response function IG(w)1 at the two sensor locations.
sensor 2
sensors 1&2
present case, the model can never be determine exactly within the limitation of
connected waveguides.
An examination of the system response functions (shown in Figure 5.33) con-
firms the expectation of very many sharp spectral peaks and the likelihood of an
over-sensitivity problem.
Figure 5.34 shows the force reconstructions using an elastic joint model; the
properties of this model were obtained by a model search method (details can be
found in Reference [84]). This example illustrates the necessity of an adequate
structural model. It also indicates that the purely analytical modeling must be
supplemented by experimental data in order to determine the appropriate model. It
is gratifying to note that the method correctly identifies the presence of the second
impact at about 1100 J1,s.
This section dealt exclusively with the frequency domain deconvolution method.
By employing the spectral element method, the system transfer function, G, is
readily available at any location on the structure. While this is not essential in the
Problems 197
Problems
5.1 Show that the stiffness relation for a shaft is given by
_2e- ikL
1 + e- i2kL ]{ t }
- Reference [43], p. 227
5.2 Use the Taylor series expansion of the exact stiffness matrix to show that the
matrix representation of the elastic foundation is proportional to the mass
matrix.
- Reference [40], p. 269
5.3 Show that in the limit as w becomes very small, that the spectral shape func-
tions 8n(X) approach those of static problems.
- Reference [40], p. 265
6
Waves in Thin Plates
A plate is a body where one of the dimensions is substantially smaller than the other
two. Plates in flexure are the two-dimensional equivalent of beams and classical
plate theory is its equivalent of the Bernoulli-Euler beam theory, whereas the in-
plane or membrane behavior of plates is analogous to that of rods. Many of the
behaviors of plates and beams are quite similar, therefore only those aspects that
are significantly different will be covered in this chapter.
The two-dimensionality of the plate affects the solution in a number of ways.
Most significantly, because the plate is large in extent, the waves are no longer
confined to follow a I-D guide and are confined only within the lateral surfaces;
the properties of the plate and of the waves can be different in different directions as
shown in Figure 6.1. Additionally, an extended plate can have a localized boundary;
in that case we then refer to the reflected waves as scattered waves. What makes
this interesting is that the incident wave can now bend around the scatterer. We
conclude the chapter by considering a curved plate; this is a rich example of
differential coupling.
Figure 6.1: Contours showing the contrast in deformed shape of isotropic plate (left) and
orthotropic plate (right) 30 JLs after impact. Note the directional dependence of the waves
in the orthotropic plate.
198
6.1 Plate Theory 199
z z
where O/X and o/y are rotations of the subscripted faces in the directions of the
curvatures. These say that the deformation is governed by five independent func-
tions: u(x, y), vex, y) are the in-plane displacements; w(x, y) is the out-of-plane
displacement; and 0/Ax, y), 0/ y(x, y) are the rotations of the mid-plane. It is un-
derstood that all variables are also functions of time, but we make the assumption
that the above kinematic representations do not change under dynamic conditions.
The normal and shear strains corresponding to the above deformations are
au au ao/x
En - = - -z--
ax ax ax
au av ao/ y
Eyy - = - -z--'
ay ay ay
200 Chapter 6. Waves in Thin Plates
Yxy
Yxz
Yyz
Because the plate is thin and because the lateral surfaces are traction free, the stress
in the z direction cannot be very large. We therefore assume that it is approximately
zero; that is, we assume a state of plane stress with au ~ O. Substituting for the
strains in the Hooke's law for plane stress then leads to
Substitute for the stresses and strains and integrate with respect to the thickness to
get the total strain energy as
U= 1
2
rD[( a1/lx
JA ax
+ a1/l y)2_
ay
V)[4 a1/lx a1/ly _ (a1/ly + a1/lx )2]] dxdy
ax ay
1(1 _
2 ax ay
+1 rGh [(1/1 _ aW)2 + (1/1 _ aW)2] dx dy (6.3)
J2 A x ax ay Y
If the applied surface tractions and loads on the plate are as shown in Figure 6.2,
then the potential of these loads is
where the edge loads can be on each face. The equations de-couple into in-plane
and out-of-plane sets, hence we now find it convenient to treat them separately.
U = !f
2
E*h [(aU)2 + (aV)2 +2v au av +
ax ay ax ay
!(l _ V)(au + aV)2] dxdy
ay ax
i
2
a
2u
(6.6)
== phfii2
a2 v
phfii2 '
These are the Navier's equations and have the associated boundary conditions
where we specify one condition from either set:
{ u or Nxx == - Eh
- 2 [au
- + vav]}
- ,
1- v ax ax
We can give interpretation of the boundary conditions in terms of resultants on
the cross section. For example, the resultants of the normal and shear stresses are
defined on the cross section as
and leads to
Eh [au av]
(1 - v2) ax + v ay
Eh [av au]
(1 - v2 ) ay + v ax
Eh [au av] (6.7)
2(1 + v) ay + ax
202 Chapter 6. Waves in Thin Plates
That is, Nxx and so on are the resultant forces per unit length due to the stresses
acting on the edge faces.
We have already considered solutions to the Navier's equations in Chapter 4,
therefore they will not be pursued any further here.
-D
1 [ (1 - v)V 2 l/Ix + (1 + v)- - x + -al/ly )] + Gh
a (al/l [aw
- - l/Ix ]
2 ax ax ay ax
1
'iD [ (1 - v)V 2 l/Iv + (1 + v)- - x + -al/ly )] + Gh
a (al/l [aw
- - l/Iy ]
. ay ax ay ay
(6.9)
where Ip == h 3 /l2. These are the Mindlin equations of motion for a plate; this
theory accounts for the shear deformation as well as the rotational inertia. The
associated boundary conditions (on each edge face of the plate) are specified in
terms of any three conditions selected from the following groups:
These are specified for an x face, the other faces are similar.
We can give interpretation to the boundary conditions in terms of resultants of
the stresses on the cross section. For example, taking resultants for the shear stress
defined as
6.1 Plate Theory 203
leads to
We can also take a moment due to the stresses acting on the edge faces. For example,
D [a lfrx + v a lfry ]
ax ay
2Mxy D alfrv
[- "+ alfrx]
- (I-v) (6.11)
ax ay
In order to account for the truncation error of the expansions u and v, we could
add correction coefficients to the energies as was done in Chapter 4. We will not
pursue this here because our interest is to develop a theory for thin plates.
aw
- -lfrv = 0
ay "
It is important to realize that while these combinations are zero, their product
with Gh is nonzero (since it is related to the transverse shear resultant). The
displacements for the flexural motion is approximated as
aw aw
u(x, y, z) ~ -z-(x, y), vex, y, z) ~ -zay-(x, y), w(x, y, z) ~ w(x, y)
ax
The normal and shear strains corresponding to the above deformations are
204 Chapter 6. Waves in Thin Plates
-Ez [a 2w 02W]
1 - v2 ox2 + v oy2
-Ez [a 2w 02W]
1 - v 2 oy2 + V ox2
Substitute for the stresses and strains and integrate with respect to the thickness to
get the total strain energy as
u
T
where we have made our second assumption that the rotational inertia is negligible.
The potential of the applied loads is
V=- f ow
q(x,y)wdxdy-Mxx--Vxzw+···
ox
where the edge loads can be on each face. Using Hamilton's principle then leads
to the governing equation
(6.14)
The shear to be specified is called the Kirchhoffshear. This shear is not the resultant
Qxz but is actually given by
Kirchhoff shear:
This can be understood by first realizing that the shear moment Mxy can be inter-
preted as a couple comprised of vertical forces a small distance apart. Then, since
the moment is distributed, so too are the vertical forces, which consequently at any
given location will have an imbalance in the vertical forces. Alternatively, the clas-
sical plate theory has restrictive degrees of freedom where the shear strains Yxz and
Yyz are zero. That is, the shear resultants Qxz and QyZ do not have a relationship to
the corresponding deformation. While this can be rationalized in the constitutive
relation by saying that the shear modulus in the transverse direction is very large,
it means that the resultant force is associated with higher order derivatives of the
deformation.
In the later sections we will look at boundaries that are straight, let the boundary
be located at x = constant, then to summarize, the type of boundary conditions
to be satisfied are to be chosen from
Displacement : w = w(x, y, t)
aw
Slope : 4Jx = ""ih
Moment: Mxx = +D [a2~ + v a2~]
ax ay
Shear : VXZ = -D [aax3 w3 +(2 - v) axay2
a3 w ] (6.16)
Note that the Poisson's ratio venters the last two of these and acts to couple the
variations in x to those in y. The corresponding expressions for the y face are
obtained by permuting x and y.
Spectral Analysis
It is easy for us to add some viscous damping to the governing equations, it is just
a matter of modifying the inertia term to give
2 2 aw a2 w
DV V w + TJh at + ph at 2 = q (6.17)
(6.18)
The solution of this equation can be written as linear sums of solutions of the
following two differential equations:
phu} - i 1Jhw
(6.19)
D
These form the basic equations for further analysis and emphasize that there are
two fundamentally different modes just as there is for beams.
Consider a plane wave such that w(x, y) = w(x). the differential equation for
w becomes
d2 -
~2 ±fPw=O
dx
This has constant coefficients. hence e- ikx is a kernel solution. The spectrum
relations are then
k2 = ±J-{32
For the undamped case, the relation between the wavenumber and frequency is
given by
kl = ±vfw [ ';;
h] 1/4 , k2 = ±ivfw [ ';;
h]I/4
This is quite similar to that of the beam if the following associations are made
EI D E
-{}- or E {} 2
A h (l - v )
Thus the plate behaves as a beam in plane strain. This applies to the expressions
for the phase and group speeds also:
C =
W
k = vfw
[Dph ]1/4 ' dw
c g = dk = 2vfw
[D]I/4
ph = 2c (6.20)
The behavior of Figure 3.3 for beams can be taken as the behavior of thin plates
when subjected to a plane wave.
Axisymmetric Equations
The differential operator for the spatial variation under axisymmetric conditions
is
,,2=_+ d2 1 d
__
dr 2 r dr
The governing equations for the plate. when damping is neglected. are similar to
those for the static deflection of plates on an elastic foundation and Reference [123]
has shown that those problems have axisymmetric solutions of the form
The remaining coefficient A is determined from the applied load condition. Let
the force be concentrated and again following the approach in Reference [123]
(for a static load) consider the equilibrium of a small cylinder section of the plate
surrounding its point of application. The shear force per unit length Q. at the
cylinder. is related to the force by
p A P
Q=- or Q=-
21Tr 21Tr
But the shear is also related to the displacement by
In the above. the superscript prime indicates differentiation with respect to z = fJr.
Letting r approach zero and equating both relations for Qallows A to be determined
in terms of P. The solution for the displacement now becomes
208 Chapter 6. Waves in Thin Plates
w(r t)
,
= "
~ 8fPD
-iP {lO - iYo - i~Ko}eiwt
n:
(6.23)
Other aspects of this solution are developed in good detail in Reference [67].
An interesting relation is obtained from the above by looking at the velocity at
the impact site, that is,
W ~-2- {,
. (r= 0,t) = "wP .2 K } iwt = "~
JO-lIO-I- 0 e
·V P {I} eiwt
r;:r;n
8f3 D n: 8-v phD
Since the frequency does not occur in the transfer function, then performing the
summation over all frequency components gives the time domain relation
. 1
w(t) = r;:r;n P(t) (6.24)
8-vphD
That is, the contact force is directly related to the particle velocity at the impact
site. Compare this with the equivalent hereditary relationship determined for the
beam. Reference [90] shows some applications of this formula. This equation no
longer applies when there is damping in the plate.
Strain Responses
For variety, in this section we look at the strain responses for the impacted plate.
These are interesting because while the displacements are finite everywhere the
strains can be infinite.
Some experimentally recorded strain histories are shown in Figure 6.3. Since
these experimental strains will be used actively in the next examples, their origin
will be briefly summarized. The basic experimental set-up is detailed in Refer-
ence [45] and in Chapter 1. An aluminum plate 1.65 mm thick was used with a
steel ball (12.6 mm in diameter) as the impactor. The data was digitally recorded
every microsecond (1 f.,ts) for a total of 4096 points per channel. Once the data are
collected, they are transferred in digital form to the computer where it is de-coded
and scaled. It is also smoothed by use of moving averages. The traces of Figure 6.3
have most of the reflections removed so that the data resembles that of an infinite
plate. Since both the plate and sample window are finite, it is necessary to do some
pre-processing of the strain data. The three main operations as discussed in Chap-
ter 1 are: removing obvious reflections, appending the signal with its long-term
behavior, and padding with zeros. Appending the signal improves the estimate of
low-frequency content. The amount of padding needed depends on the distance of
the gage from the impact site. An estimate of the amount needed is given by
tpad = Xgage
JN!)'T [Ph]
~ Ii
1/4
For the above aluminum plate a gage located at 50 mm would require 250 f.,tS
padding for a window of 1000 f.,tS, but 500 f.,ts padding for a window 5000 f.,ts. In
6.2 Point Impact of a Plate 209
the present case there is ambiguity between some of the reflections and the multiple
impacts. The data are first given a five-point moving average smoothing. The data
from 600 /-LS onward are then given an 80 I-point average smoothing. The resultant
processed signals are the ones shown in Figure 6.3.
- experimental
------- predicted from 50mm
------- predicted from Omm
c:
E
.~
iii 100
§ 0
'E -100 x=O
x=50mm
Time [J.1S] ,
O. 200. 400. 600. 800. 1000.
Figure 6.3: Comparison of predicted radial strains with experimentally measured values
for an impacted plate.
The motivation for the five-point average is that the data will later be sampled by
the FFT every 2 /-LS and this ensures that there are no aliasing effects. The purpose
of the 801-point average is to remove reflections. This type of scheme can be
successful because the mean of the reflections is zero.
The strain histories have some very interesting features. First, the peak strain
observed decreases very rapidly with position. Second, they all tend to the same
limiting behavior which can be shown to be proportional to
The strain decreases in time much more rapidly than for a beam which has a l/..ji
dependence - the reason is that the energy occupies an increasing volume as it
propagates.
The radial and hoop strains are obtained by differentiation of the displacement
to give
h a2 w iPh i2
Err 2 ar2 = 32D {(10 - lz) - i(Yo - Y2 ) + -;-(Ko + K 2 )}
h aw iPh i2
2r a;: = 32D {(10 + lz) - i(Yo + Y2) + -;-(Ko - K2)} (6.25)
This solution has a peculiarity in that at r = 0 the strain is infinite (even though
the displacements are finite) and given by
210 Chapter 6. Waves in Thin Plates
-
Err
- =-
= Eoe i Ph
- [ 1 - -4i In(r) ]
32D 7C
A variety of schemes can be used to eliminate this singularity; it can be done most
directly by realizing that a strain gage gives the value of strain averaged over its
length. That is, if the gage is of length L and centered at r = 0, then its strain
reading is
-
Err =
iPh [ 4i
32D 1 - ~ In
(f3L)]
2e
This is nonsingular for all finite L. When strain measurements are made at the
impact site, it is this equation that is used in the analysis.
En (rl) Pn . Gn(rl)
Pn En(rl)/Gn(rl)
En (r2) En(rd· G n(r2)/G n(rl) (6.26)
These correspond, respectively, to the three cases: (1) given the force obtain the
strain at rl; (2) given the strain at rl, obtain the force; and (3) given the strain
at rl obtain the strain at r2. In each case, the force is obtained at the center of
the plate (r = 0) and the transfer functions are known fron Equation (6.25). Of
course, once the new spectral components are known the time histories are obtained
simply by using the FFf inverse. The latter two of Equations (6.26) illustrate the
active use of experimental data. For example, the third equation shows how we
can check the consistency of two experimentally measured strain histories. As
shown in Figure 6.3, we can use the strain measured at location 1 to predict the
strain at location 2 and compare that to the strain actually measured at location 2.
The figure confirms that the data and modeling are of good quality although small
fluctuations measured at x = 50 mm get exaggerated when propagated back to the
origin. This active approach shows how responses measured at a limited number
of points can be used to infer responses at a very large number of points.
Figure 6.4 shows the reconstructions of the force history. They show good agree-
ment with each other, the largest differences being where the reflections are sig-
nificant. Also shown in this figure is the force history predicted using the Hertzian
contact law [36, 37]. The agreement is also very satisfactory, indicating that the
Hertzian contact is good for steel on aluminum.
6.3 Wavenumber Transform Solution 211
.4
.2
.0 0°
Time [JlSJ,
-.2
O. 200. 400. 600. BOO. 1000.
Figure 6.4: Comparison of reconstructed force histories with Hertzian contact theory.
These results confirm the basic validity and accuracy of the spectral approach to
actively manipulating experimental data. There are two main areas of sensitivity.
First, the further away from the impact site the gage is located, the weaker the signal
is and consequently the greater the noise level in the predicted force. Second,
reflections tend to obscure the estimated long-term behavior of the strains thus
giving poorer estimates when the contact time is relatively long.
While the effects of reflections can be minimized by using a very large plate,
this often is inconvenient or expensive. The next few sections plus Chapter 8 will
take up the issue of handing reflections directly.
Basic Solutions
Since the waves, possibly, can have arbitrary shapes in space, consider a represen-
tation of the form
phwZ - i1Jhw
D
The coefficients of the differential equations are constant, hence the solutions are
exponentials of the form e- ikx . The characteristic equations associated with these
solutions are
-k~ - ~z + f32 = 0, -k~ - ~2 - f32 = 0
(6.28)
These spectrum relations are shown in Figure 6.5. It is noted that for a particular
~ the first mode shows a cut-off frequency with the lower-frequency components
being nonpropagating.
The solution for the plate becomes
That is, the actual solution is obtained by summing kernel solutions of the above
form for many values of Wn and ~m. In elementary terms, it can be thought of
as containing a forward-moving wave A, a backward-moving wave C, and corre-
sponding evanescent waves B and D. However, because the wavenumbers k\ and
kz can be complex then the above distinctions are somewhat blurred. To amplify
this idea, consider the solution as a plane wave in x (the bracketed term) modified
in y. Then, for a particular ~, the summation over w is similar to that for a beam
in Chapter 3. The corresponding spectrum relations, however, are modified by ~
as shown in Figure 6.5.
It is not always clear, a priori, as to the maximum M to use. The safest approach
is to run the problem for a number of M and check for convergence.
The present formulation is set up to solve problems whose boundary conditions
are specified along x = constant. At the boundary x = L, the solution and the four
conditions can be written as
-mode 1
........ mode 2 F1l'!ql
Figure 6.5: The spectrum relation for different values of the ~m wavenumber.
\
-Mxx
-
D
-\ - (6.30)
VVxz
A* == Ae-ik,L ,
have been used. The particular problems of interest will be constructed from com-
binations of these.
(6.3\)
Since this solution must be symmetric about x = 0, then the slope with respect
to x is zero, giving ik2B = -ikJA. The shear near x = 0 is determined from
Equation (6.30) to be
or
_ • pA 4
w(x, y, t) = ~
~ I ~
-ik,x
[e
-ik2X
e]
i~v iwt
8DP.2 W ~Em - k - - - k - e- . e (6.32)
n f' m ] 2
It is interesting to note that if ~ = 0, then k] = -ik2 = f3 and the solution for the
impact of a beam is recovered. Also of interest is the comparison with the solution
using Bessel functions given in Equation (6.23). It is seen that both solutions
are the same since the Bessel functions in polar coordinates have the cartesian
representation
{J""
.
'2
-IY - - K I}
0
4
:::}-
W
L [e k
E -----
m k2
-ik,x
e]
-ik2X
e -i~v.
TC m]
This is analogous to Lamb's (1904) relation (13) except that here the summation
is finite.
We can· gain further insight into the behavior of this solution by considering
the behavior of the bracketed term in Equation (6.32) as shown in Figure 6.6.
The figure is for m = 150 arid x = 300 mm. For nonzero x, the k2 exponential is
negligible. When f3 < ~ the k] wavenumber is imaginary and the wave completely
evanesces. When f3 > ~ the wavenumber is real and a component of a wave will
propagate. If we view each ~m as a waveguide, then all the lower frequencies will
be filtered in a manner analogous to that of a rod with elastic constraint. Hence for
a spectrum of ~m' we expect the low-frequency content of the arriving signal to be
dominated by the smaller wave numbers or larger radii.
Figure 6.6: Transfer function for the point impact of a plate; m = 150, x = 300 mm.
6.3 Wavenumber Transform Solution 215
-M=3200
------- M=1600
o experiment
c: ------- M=900
'i!! 100
g
'Iii 0
::!l -100
E -M=400
x=O
x=50mm
Time [J.1S] ,
o. 200. 400. 600. 800. 1000.
Figure 6.7: Comparison of experimental strains with predicted values using the wavenumber
transform.
Figure 6.7 shows the comparison of the experimental strains of Figure 6.3 with
those predicted by the above analysis. The results establish the equivalence of the
Bessel and double series solutions away from the impact site. The specifications
for the computations are time window T = 2048 J-LS, N = 1024, space window
W - 20 m and M is as given in the figure. At the impact site, the singularity in
strain is indicated by an increasing peak strain value as M is increased. Again, when
dealing with experimentally measured strains, this singularity can be removed by
integrating over the length of the strain gage.
Although the solutions involving Bessel functions and the wavenumber trans-
form are equivalent, the former is computationally more efficient since only a
single series is involved. Unfortunately, the Bessel form of solution is not suitable
for imposing boundary conditions along a straight edge and so the double series
form must be used. If the response is required at many instances of time and many
locations in space, then a double FFT algorithm should be used for the inversion.
In the present circumstances, however, the solution is required (almost) continuous
in time but only at a limited number of space locations (because the experimental
comparison is with a limited number of strain gage data sets). Consequently, the
summations over m are performed long form, but those over n are done using the
FFT.
(6.33)
and the elements Dij are the anisotropic stiffnesses. The equation of motion then
becomes
(6.34)
For convenience, assume the coordinate directions coincide with the principal
material directions, then we can set DI6 = D26 = 0 and the equation of motion
reduces to
(6.35)
kl=±y'W [ ~
h J 1/4
, k2 =±iy'W [ ~h J 1 4
/
Doo Doo
Formally, these are the same as for the isotropic plate except that the stiffness
appearing in the relation is a function of the orientation of the wave.
Consider now arbitrary waves of the form
6.3 Wavenumber Transform Solution 217
(6.36)
propagating in the plate. The characteristic equations associated with these solu-
tions are
Even though the material is orthotropic, there is still only two families of modes.
When the material is isotropic, we have D22 = D)) = DI2 + 2D66 = D and
DI2 = vD and we see that the above reduces to that of Equation (6.28).
The spectrum relations have an interesting behavior because of the terms inside
the radical; for certain materials, the square bracketed term could be negative
giving wavenumbers that are complex instead of being entirely real or entirely
imaginary. For example, the material properties for graphite/epoxy and glass/epoxy
are, respectively,
1 0.03 0 ] 1 0.07 0]
[ D ]gr/ep = D)) [ 0.03 0.46 0 , [ D ]gllep = D)) [ 0.007 0.57 0
o 0 0.05 o 0.14
These give terms inside the radical of
In both cases, the individual modes of the orthotropic plate behave like a beam on
an elastic foundation.
(6.37)
Since this solution must be symmetric about x = 0, then the slope with respect
to x is zero, giving ik2B = -ik]A. The shear near x = 0 is determined from
Equation (6.30) to be
218 Chapter 6. Waves in Thin Plates
or
q= 2Dik 1 A(kf - k~)
which formally looks the same as for the isotropic plate. Let the applied force
spectrum be represented by the Fourier series
L ijme-i~mV L Eme-i~mY
~ ~
q(y) = ~ = ; (6.38)
m m
Again, this solution looks very similar to that of the isotropic plate and to that of the
beam. Also of interest is the comparison with the solution using Bessel functions
given in Equation (6.23). Contours of displacement constructed from this solution
are shown in Figure 6.1. Note the different speeds of propagation in the different
directions.
where Ce == cos (J and Se == sin (J. This gives a wave generally moving in the y
direction, with the A term being incident, the C term being reflected, and the D
term being an x attenuated vibration. For simplicity, we will assume the incident
evanescent B wave is negligible. Since they all have the common y term then
------
the boundary conditions will be satisfied for any y. The geometry is shown in
Figure 6.8.
reflected
Lx
---
incident
edge
------
Figure 6.8: Wave reflections near a straight edge.
w =0= A+C+D
aU!
ax
- 0- -ifJCeA + ifJCeC + fJ) 1 + S~D
giving the coefficients of the generated waves as
C=-
)1 + S~ + iCe A, (6.40)
)1 +S~ -iCe
When the angle of incidence, (J, is zero, then the reflected waves are phase shifted
relative to the incident waves. On the other hand, when (J approaches 1r /2, then
the D term disappears and C = -A showing that the superposition is zero.
Consider now the situation where the edge of the plate is supported by springs
such that the boundary conditions become
aw
M xx =-(l-
ax • Vxz - -Kw
where K and a are the linear and torsional spring constants, respectively. In terms
of the wave coefficients these conditions become
-1+v*+iCoa* 1 + v* + J+
1 SJ a*
[
iCo(1 + v*) - K* J+ I SJ(1 - v*) + K*
{ I - v* + i Coa*
= iCo(l + v*) + K*
where
v* - (1 - v)SJ, a* = a/{JD, K* = K/{J3 D
Obviously, the special cases of K = 0 or 00 and a = 0 or 00 can easily be taken off.
It is apparent that the general case is frequency dependent. An interesting point to
note is that Poisson's ratio appears only in association with sin (). This means that
for normal incidence Poisson's ratio does not playa role.
w=o, Mxx =0
C = _Ae-ik,L, (6.41)
Thus, the wave appears to originate from an image source a distance 2L away from
the actual source. However, this is the only case where the image idea applies; in
general, the new waves will originate at the boundary itself - we will develop
this idea a little more in the final section of this chapter.
6.4 Waves Reflected from a Straight Edge 221
For a fixed edge condition both the deflection and slope are zero,
w=o, aw =0
ax
giving as the system to be solved
{ C} 1 [kJ + k2 (6.42)
D = kJ - k2 -2kJ
This is zero only when the frequency is zero. The presence of both e- ik , L and
e-ik,L in both C and D shows that the reflected waves cannot be written simply in
terms of image waves.
A free edge boundary requires the applied moment and shear to be zero,
where
The solution is
2k2k~Jk~2 (6.43)
-kJk~2 - k2k~J
(6.44)
real wavenumber
8.
6.
4.
2.
o.
-2.
-4.
-6.
o. 5. 10. 15. 20. 25. 30. 35. 40. -8.0 -6.0 -4.0 -2.0 .0 2.0 4.0 6.0 8.0
Frequency [kHz] imaginary wavenumber
Figure 6.9: Contours of the detenninant. On the left is a slice of Re(~) against frequency
for Im(~) = 0, and on the right is Re(~) against Im(~) at 20 kHz.
the determinant will become so small that significant responses can be observed.
Thus, in comparison to our previous double summation solution, this plot shows
that, of the [N x M) matrix of components, only those along the locus of zero
determinant contribute significantly.
The free edge boundary condition is different from the others in that it is possible
for the determinant to be zero. This is interesting enough that we consider it in
greater detail in the next section.
Free-Edge Waves
In the absence of damping, it is possible for the determinant of Equation (6.44)
to equal zero, therefore raising the possibility of a singular solution for {C, D}.
In other words, it is possible for a free wave to travel along the edge. This wave
is analogous to a Rayleigh wave as detailed in Chapter 4 and in Reference [125);
hence the following development attempts to accentuate the similarities as much
as possible.
In its simplest form, we are interested in knowing if the wave
tV = [Ee- ik1x + Fe-ik2X)e-iky
where
can propagate along the free edge and satisfy the boundary conditions. At present,
the wavenumber k is as yet undetermined. By imposing the free-edge conditions,
6.4 Waves Reflected from a Straight Edge 223
this gives
[i~11~2 i~:h] { : }
=0
(6.45)
(6.46)
is zero. (This expression for the determinant is obviously the same as occurred in
the last section.) The wavenumbers where this happens are (after expanding and
simplifying)
These eight values coincide with the eight zeros of Figure 6.9. We are interested
in free waves, that is when k is real only, hence of the eight possible solutions, we
are only interested in the following pair:
1/4
[ (1 - 3v) + 2Jl - 2v + 2v 2 ]
k = ±fJ
(l - v)2(3 + v)
The quantity inside the bracket is close to unity for all values of Poisson's ratio.
For example,
v = 0.00, k= 1.0000013, kJ = -O.OooOifJ, k2 = -1.4142ifJ
v 0.25,
= k= 1.0004113, kl = -0.0287ifJ, k2 = -1.4145ifJ
v = 0.30, k= 1.0009513, kJ = -0.0436ifJ, k2 = -1.4149ifJ
v = 0.35, k= 1.00196fJ, kl = -0.0626ifJ, k2 = -1.4156ifJ
Our value of v = 0.33 gives k = 1.oo148fJ, kl = -0.0544ifJ, and k2 = -1.4152ifJ.
Using this result in Equation (6.45) gives the amplitudes ratios for E and F as
W = L E[e-O.054fi"x + _V_e-1.415fi"x]e-l.OOJ48ifi"YeiWnf
2-v
(6.47)
n
This wave is characterized by a speed that is slightly slower than the corresponding
speed in the infinite sheet, and by an exponential decay as it penetrates into the
interior of the plate. We will refer to this wave as the "free-edge wave."
224 Chapter 6. Waves in Thin Plates
- double summation
free edge wave
x~
y=O.5m
y=1.0m
~
_______~_.,A!!~~~AL y=1.5m ~
_________v_.vii~~W.vv y=2.0m ~m
---------------------~~ ------~~----
x=200mm
o. 500. 1000. 1500. 2000. 2500. 3000. 2000. 2500. 3000. 3500.
Figure 6.10: Edge behavior.
Figure 6.10 shows transverse velocity responses at various distances from the
impact site along the edge. For comparison, we use the free-edge wave (with
E characterized at y ~ 1500 mm) and use spectral analysis to propagate it both
forward and backward as illustrated earlier in this chapter. The most striking feature
of this plot is that the amplitude of the edge wave diminishes very slowly over
space compared to waves which propagate in the interior of the plate. In addition,
our free-edge solution shows insignificant deviations from the double summation
solution, thus confirming that the edge wave responses behave very similarly to
Rayleigh waves.
Figure 6.10 also shows how the edge wave penetrates the interior of the plate.
The observation points are at y = 2500 mm, and various x locations inward from
the edge. The magnitude of the response decreases rapidly, showing that the edge
wave is localized to regions close to the edge. Furthermore, the close agreement
of the single and double summation responses again helps confirm that the edge
wave responses behave very similarly to Rayleigh waves.
The previous sections have looked at the reflection of waves from a straight edge
that was also total in the sense that the complete extended boundary was involved.
For the two-dimensional plate, however, it is possible for a boundary to be limited
in extent - a circular inclusion as shown in Figure 6.11; this has the significant
consequence that only some of the wave is affected by the boundary. This is a
difficult subject usually covered in specialty books such as Reference [99]; this
section gives only a short introduction to some of the ideas involved.
~ scattered wave
wave~___
incident f G~:=~:~ IJrJJL
.~:ow
*- scattered wave
forward scattered
One of the basic issues involved is that the coordinate system used to describe
the incident wave is different from the coordinate system ofthe boundary geometry
and consequently, different also from the coordinates used to describe the reflected
waves. To illustrate the consequences ofthis, we will consider a plane wave incident
on a fixed circular boundary, the schematic of the event is shown in Figure 6.11.
where x = 0 is the origin of our new coordinate system describing the boundary
and we have assumed negligible damping. For simplicity, we will take the incident
evanescent term to be negligible (B ;::::; 0).
The geometry of the boundary is circular as shown in Figure 6.11 hence we
need to represent the incident wave in this coordinate system. Let
where 1m are Bessel functions of the first kind. This representation is actually a
Fourier series with 1m as the Fourier coefficients; in that case we have the relation
226 Chapter 6. Waves in Thin Plates
W= LEm(-i)mA*Jm(k1r)cosme
m
Just as e-i~mY modified the plane waves in Equation (6.29) to give circularly crested
waves, so does cos me modify the circular waves to give the plane waves.
plane wave
~'""'-'''-/
plane wave
The first question we need to resolve is the number of terms necessary in the
expansion representing the incident wave. Figure 6.12 shows the reconstruction
of the plane wave at x = 0 and y = 300 mm for different number of M. The
symptom of not having enough terms in the expansion manifests itself as a sort of
wrap-around problem. At the distance y = 300 mm, it is necessary to have about
M = 60 to get good convergence. The figure also shows that the representation
deteriorates as the distance y is increased for a fixed number M.
Scattered Wave
The governing equations for flexural waves in cylindrical coordinates are
1 a2 ]
[ -a2+ -
I a
-+--
2
w±j3 w=O
A 2 A
ar2 r ar r2 ae
Let the displacement be represented by a Fourier series in the hoop direction, that
is,
w(r, e) = Wm cos(me) L
m
6.5 Scattering of Flexural Waves 227
These Bessel functions are shown plotted in the Appendix; note that both Ym and
Km go to infinity at the origin r ... 0 whereas 1m goes to infinity at large argument.
Hence, when we look at the exterior problem we take B = -i A so that the solution
is a Hankel function that coincides with an outward propagating wave, and we take
D=O.
A general wave emanating from a circular region is therefore represented by
where Hm - Jm - iYm is the Hankel function. We therefore have for the total
response in the plate surrounding the boundary
The coefficients Am, Bm are now determined from the boundary conditions.
For simplicity, we will consider only the fixed boundary case, References [71,
99] are good sources for the treatment of some other cases. For the fixed edge
condition at r .. a, we have
w(a)=O,
aw
-=0
ar
giving as the system to be solved
[ Hm
H'm
where the prime indicates differentiation with respect to the argument z - k1r.
Using Cramer's rule, this gives
For a fixed argument, this determinant approaches zero as the order m increases.
To avoid numerical difficulties, we use ratios of the Bessel functions and write the
solution as
228 Chapter 6. Waves in Thin Plates
where
Note that the derivatives are computed from the recurrence relations as given in
the Appendix.
Figure 6.13 shows some of the responses at various positions around the bound-
ary. First, along the line x = -2a we see the effect of the back-scattered wave,
that is we see the wave reflecting off the object at an angle.
Figure 6.13: Scattered responses near the boundary. (a) Near side at x - -2a, (b) far side
at y = o.
Since there are responses in the shadow region, we could say there is a diffrac-
tion effect, that is, the wave has "bent around" the boundary. Generally, however,
diffraction is associated with sharp edges and in the present case we say there
is a forward scattered wave. It should not be very surprising that there is a for-
ward scattered wave since, in the absence of the boundary, the incident plane wave
is present and therefore the scattered wave is necessary to cancel it. We expect
the total response to be relatively small in the shadow region, hence the forward
scattered wave must be large enough to cancel the incident wave.
To further elaborate on the scattered waves, we will consider the far-field (k I r »
1) response. Replace the Bessel functions with their asymptotic forms
6.6 Lateral Boundary Conditions 229
The scattered power is related to the quantity 11/1(0)1 2 and Figure 6.14 shows the
angle dependence of this quantity for a number of frequencies. The plots are
normalized with respect to their areas (or total power). As the frequency increases,
a greater proportion of the scattered power is directed in the forward direction; this
field interacts with the incident wave to give a long shadow region.
----.
--... wave
...... / direction
50kHz
Figure 6.14: Directivity plot of the power function 11/r(6)1 2 over a range of frequencies.
is general enough to satisfy the boundary conditions along two x =constant edges
simultaneously - it is the boundary conditions along y =constant that pose the
difficulty and this is what we consider in this section.
What we call the x axis is of course arbitrary, in the following we will choose
the x =constant boundaries to be the more significant boundary and that the
y =constant boundary is more remote and hence less significant. Should it happen
that both boundaries are equally significant, then the ideas to be presented are still
applicable but their implementation will be very cumbersome.
Pinned Boundary
When the lateral boundary condition corresponds to that of a pinned boundary
then we can effect a solution in a very simple fashion. Let the disturbance be at
230 Chapter 6. Waves in Thin Plates
image _ .
moments
lateral boundary
load j Lx
Figure 6.15: Geometry of a quarter plane.
The coefficient A is determined by the imposed force condition. From this we see
that the solution does not have enough parameters to simultaneously satisfy the
lateral condition. The superposition idea adds to this a second solution
and it is easy to verify that choosing A = - A will then satisfy the boundary
condition.
Figure 6.16 shows the effect of the pinned boundary - in many ways the results
could be anticipated from those of the pinned beam. Indeed, if this boundary is
remote enough then essentially plane waves impinge on it and the beam analysis
is applicable.
The second of the two plots shows the behavior of aw/ay along the boundary.
While the deflection is zero, we see that the slope has increased to twice the infinite
plate value.
6.6 Lateral Boundary Conditions 231
y=300mm
y=200mm
y = 100mm x=200mm
y =300mm x=300mm
I I I I I I I I Time [JISl I I I I I I I I
O. 500. 1000. 1500. 2000. O. 500. 1000. 1500. 2000. 2500.
Figure 6.16: Responses for pinned lateral boundary located at y = 300 mm. Left is
w(x = 0, y), right is ilw/ily(x, y = L).
Fixed Boundary
Consider now the case when the lateral boundary is fixed, that is, both the deflection
and slope is zero. A superposition similar to the previous section will satisfy the
zero deflection but in doing so the slope condition will be grossly violated. There
is not an image source that will set this slope to zero and simultaneously leave the
deflection unaffected. Consider the fixed boundary as being a pinned boundary but
with a distributed moment chosen precisely to set the slope the zero. If we can find
such a moment distribution then we have the required solution.
We begin by considering the solution for a concentrated moment located at
x = 0 and y = L. We can obtain this from the y derivative of the concentrated
force solution - the validity of this can be explained by visualizing the shear
distribution along x = 0 as being a bell shaped curve, the y derivative corresponds
to a shear distribution that goes from positive to negative giving a resulting couple
about the x axis. We express this solution as
(6.48)
This solution is capable of satisfying the slope condition at only a single point. Let
this be at x = 0, y = L, we then require
232 Chapter 6. Waves in Thin Plates
This, in conjunction with Equation (6.48), gives a solution that cancels the incident
wave everywhere not just at the boundary; in other words we cannot effect a solution
for arbitrary wavenumber ~m •
However, it is still possible to get a zero localized slope, but now we must work
on the frequency component. We know that the amplitude A is related to the applied
force in the manner
Am = PQm
(this can be seen from the impact solutions from earlier in the chapter). It is clear
that Amust have a similar representation
We can then solve for T (which now has the interpretation of the applied moment)
as
Lm Qm(l- ~)[-2i~e-i~mL]
T= P k2
Lm Qm(l- :~)(i~m)2
Figure 6.17 shows some of the responses at various positions along the line x = 0;
in comparison to Figure 6.16 we see that the reflections are diminished. This is not
surprising since the clamped condition generates evanescent waves which tends to
diminish the strength of the reflected propagating components.
Figure 6.17 also shows some of the slope responses near the point of application
of the concentrated moment. It is clear that the moment influences only a small
region in its vicinity - in the present case of about 25 mm radius. This situation
is identical to what we saw in the previous section on the scattering from a rigid
boundary; the present case is the limit as the radius of the rigid boundary goes to
zero.
For us to simulate an extended fixed boundary would require many concentrated
moments. When selecting their amplitude, not only must the phase of the incident
wave be considered, but also the presence of all the other concentrated moments.
In other words, the distributed amplitudes must be obtained by solving a set of
simultaneous equations. This adds significantly to the computational cost. How-
ever, if the lateral boundary is somewhat remote, then a good approximation can
be obtained from a sparse distribution of concentrated moments, and furthermore,
these need only be phase matched to the incident wave.
6.7 C lfved Plates and Shells 233
--'V V\./I~~-=-=
"A ~ x=25 mm
y=300mm X= 18mm
y =200mm
x= 12mm
-pinned
........ infinite -fixed
-pinned
y= 100mm x=6 mm
x=Omm
w........ . I! ! ! ! I ! ! ! ! I ! ! ! ! I Time [IISI I !!! !!!! ! I ! ! ! ! I ! ! ! ! I
Figw'e 6.17: Responses for a pinned lateral boundary with a single fixed point. Left is
w(x ,,0, y), right is (Jw/(Jy(x, y = L).
Figure 6.18: Free body diagram for curved shell segment. Left is cylindrical coordinates;
right is curved shell coordinates.
with ~ == (r - R) and where the third equation allows for bending about the z axis.
These give the nonzero strains as
(6.50)
w(s, y, z) ~ w(s, y)
Ess
Eyy (6.52)
2Esy -
Other shell theories have slightly different expressions for these strains; the present
theory is closest to that of Reisner [104, 95]. An excellent survey of the different
theories are given in References [77, 81]. The theory developed here is the shell
equivalent of the classical plate theory and the Bernoulli-Euler beam theory.
Equations of Motion
While it is possible to derive the equations of motion based on a free body diagram,
it is much more advantageous to use Hamilton's principal since we then get the
set of boundary conditions consistent with these equations.
The strain energy for a small segment of shell in plane stress is
where E* == E /(1 - v 2 ). Substitute for the stresses and strains and integrate with
respect to the thickness to get the total strain energy as
where
where we have neglected the rotational inertia. Let the potential of the applied
loads be
aw
l/I=--
as
These equations are grouped in terms of membrane E[.] and flexural D [ .] contri-
butions. This rather complicated collection of equations are a combination of the
the flat membrane. flat plate. and curved beam equations.
The associated boundary conditions on the side s =constant are specified in
terms of one each of the following pairs:
aw or Qm = D [-
a2w + va2w 1 au]
-+- - (6.56)
as as 2 ay2 R as
It remains now to interpret the resultants and relate them to these boundary con-
ditions.
Resultants
Referring to Figure 6.18. we can form the resultants per unit length as
Nss =- f ass dz •
After substituting for the stresses and strains in terms of our approximations leads
to
6.7 Curved Plates and Shells 237
Nsv 1
= -(1 - [av au]
v)E- - + - (6.57)
. 2 as ay
Again, after substituting for the stresses and strains in terms of our approximations
leads to
1
Ms = -D(l-
y 2
v)- - +2-
ay R as
a [u aw]
Comparing these expressions to those for the boundary conditions, we see that the
natural boundary conditions are equivalent to specifying
1
Qu Nss + liMss
Qv Nsy
Qw
_ aMss _ 2 aMsy = v:
as ay sz
Qm Mss (6.58)
The first of these resembles the resultant load expression used for curved beams,
while the third resembles the Kirchhoff shear stress relation.
Spectrum Relation
In order to proceed, we need to have the spectrum relation. Assume solutions of
the form
which, on substitution into the governing equations, gives the homogeneous system
of equations
-y
using
The ai, a2 and y terms alone define the flat membrane problem, while a3 alone
defines the flat plate flexural problem. All the other terms are couplings due to the
curvature. The curved beam result is obtained by setting ~ = O.
The characteristic equation, obtained from the determinant of this system, is the
most complicated spectrum relation we have encountered so far. It is quartic in
k 2 , which makes it extremely difficult to solve in simple form. Fortunately, it is
not necessary to resort to a purely numerical scheme for solution since an explicit
solution is available. Following Reference [1], we first write the characteristic
equation as
Z4 + a3z 3 + a2z 2 + alz + ao = 0
The formulas given in Section 3.5 are used to solve this cubic equation.
These equations can be easily programmed. The difficulty that arises is in choos-
ing the appropriate square or cube root in both the cubic and quadratic solvers, as
well as the appropriate root from the cubic solver. It is seen that while there are
only four z roots, the formulas give a multiplicity of 144 roots. This is especially
acute since we wish to compute a single Z j at a time. The issue is that in computing
the nth rootofacomplex number z = a+ib = Aei ¢, the phase q, = tan-I (b/a) has
an ambiguity of Nrr.
We remove this ambiguity by keeping track of the total phase. That is, starting
with some value and with reference to the unit circle in the complex plane, as the
wavenumber goes from the 4th to the 1st quadrant the total phase is increased by
2rr. Algorithmically, we compute complete modes separately at each m. Starting at
a large value of frequency, we work toward the lower frequencies, keeping track of
the phases. The root c is chosen as the one that had the largest real value initially.
Approximate spectrum relations are useful here in identifying the appropriate
modes. This scheme works quite robustly when the frequency decrement is not too
large. Periodic checks with an iterative root finder helps confirm the correctness
of the roots. It must be said that determining the spectrum relations is now a
significant operation in itself, and unlike the previous reported cases, they are
no-longer computed on-the-fly as part of the structural analysis.
An idea of the variety of behaviors is shown in Figure 6.19 for a value of
~ = 2rrm/ W with m = 50. Not surprising, there are many branch points in the
spectrum relation. Most of the complicated behavior occurs in the vicinity of low
frequency as was also the case for the curved beam in Chapter 3. We know this
6.7 Curved Plates and Shells 239
~
I.... :>..--·. . . · · · · . · · . · ·. ···-..· · ·. ·=~
_Kl
.......
.. ..... K4
................
..........
""'"
...•...•...
•......•..
.......
Figure 6.19: Spectrum relations for m = 50.
region is where the thickness is relevant and to exaggerate its effects, the plots are
for the case when h / R = 0.1 and R = 250 mm for an aluminum plate.
It is difficult to get a clear idea of the complete spectrum relations from Fig-
ure 6.19, but just as we did for the curved beam, we will introduce an approximation
that will help to delineate the separate contributions. An approximation that is rea-
sonable for thin shells, that is, when h « R, is given by
e I
2 1
kp - R2 - ~
2
k 22 k~ _ ~2
1 2 4 1 ~2
k 32 --~ + /3 + 4R4 - R2
2R2
e4
3
--~
2R2
2
-
4
/3 + 4R4 - R2
9 ~2
(6.60)
The 12 coefficients are not independent but are related through the amplitude ratios.
Let us write the solutions as
where the symbol ¢ indicates an amplitude ratio. The solution is arranged so that
we can set ¢ = 0 to recover the familiar form for the uncoupled cases. We can
regroup this solution into the somewhat more compact form
The free body diagram for the impact region looks similar to that of Figure 3.5
except that there is also an axial force. At the impact site s = 0, we impose
u(O,t)=O,
aw
-=0
as
and we specify that the in-plane shear traction is zero:
a =C[au+av]=O
sx ay as
This gives three equations for the four unknown coefficients. Let us choose C as
the reference unknown, then the system of 3 equations can be arranged as
6.7 Curved Plates and Shells 241
This system is solved at each value of frequency OJ, and each value of wavenumber
~. We are now in a position to write all of the displacements in terms of the single
unknown C. First introduce the subscript notation
{ ~
W
} (s) = [~: ~~ ~:] !~ }+ { ~:
~31 ~2 ~34
{
g4 ~33
} g3
The final piece of information comes from the impacting force. We have
or
'i1-
P = ,",{D[2
L - k j + 2~ 2
(1 - ] . j [2
v) ~Ijgj +' Dlk k j + ~ 2(2 - v) ] ~jgj }C
j R
where the summations are over the modes. This allows us to solve for C in terms
of P.
Some velocity reconstructions are shown in Figure 6.20. Note that the responses
at the large angles are scaled by 10 which indicates the very rapid decay of am-
plitude. The most significant point of interest is the oscillatory behavior of the
transverse velocity. This we observed also in the curved beam example. What is
interesting here is that there is not a single cut-off frequency (there is one for each
m mode) and yet a single dominant spectral peak establishes itself.
The cylindrical shell is a very complicated example of coupling - not only are
the number of variables cumbersome to work with, but evaluating and interpreting
the spectrum relation is also quite a challenge. It, therefore, can be imagined that
related problems such as conical or elliptical shells would begin to pose almost
insurmountable difficulties. In Chapter 8, we will introduce a matrix methodology
that will help alleviate some of these difficulties, and point a way to the spectral
analysis of complicated thin-walled structures. The method has much in common
with the spectral element method of Chapter 5.
242 Chapter 6. Waves in Thin Plates
360 xl0
270 xl0
180 xl0
90 xl0
Frequency [kHz]
, , 1 , 1 Time [~] ...,1.....................................~...........
, .........................
I ' ..........
, '....'..L'.... ' ........
' '..............
' 1
Figure 6.20: Out of plane velocity responses for an impacted curved plate. Left is time
responses w; Right is frequency response Iwi.
Problems
6.1 Show that the governing equation for a plate on an elastic foundation is
where Nu . Ny)" Nxy are the in-plane loads per unit length.
- Reference [116], p. 431
6.2 Show that the solution for the initial value problem w(r, t) = foe-"I"',
w(r, t) = 0 is given by
w(r, t) b = JD/ph
H(x)
There are many practical situations where the interaction between the dynamics of a
structure and a surrounding fluid is of great importance. The most obvious is noise;
noise is the propagation of acoustic energy through the fluid. The interaction can
also influence the response of the structure itself; examples include dams, chimney
stacks, ships, fuselages, propellers, and transmission cables.
A structure immersed in a fluid can experience three types of loading. The first
is the structure-borne excitation caused by the propagating structural waves. The
second is a pressure loading arising from some source within the fluid; this acts
as a distributed external loading on the structure. The third is the "self-loading" or
fluid loading caused by the moving structure interacting with the fluid. This also
acts as a distributed loading but since the magnitude depends on the motion of the
structure, it has the effect of coupling the structure and fluid motions. Our primary
interest in this chapter is of a problem where a typically finite structure is coupled
to a relatively extended second medium as shown in Figure 7.1. We briefly review
some of the basics of acoustic waves but more thorough treatments can be found in
References [51, 66, 97]; the finer points of structure/fluid interaction are covered
in the excellent summary paper by Crighton [22].
Figure 7.1: Moire of the real and imaginary parts of the fluid pressure near a finite panel
vibrating at 1 kHz. Note the nonzero pressure near the immobile baffle.
243
244 Chapter 7. Structure-Fluid Interaction
The time variation of the fluid density Pa and pressure Pa, due to the passage of
an acoustic wave, may be written as
where Pa, Pa are the mean (ambient) values, and p(r, t), p(r, t) are the fluctuations.
(We use the subscript 'a' to stand for 'acoustic medium'.) Substituting these into
the equations for fluid motion and treating the fluctuations as small quantities leads
to the following continuity and equilibrium equations:
ap au aw
- +Pa- +Pa- =0
at ax az
aaxx aaxz .. aaxz aazz ..
- - + - - =Pa W (7.2)
~+~=Pau,
ax az
where aij are the stress components. The constitutive behavior of the fluid is
described by
au aw II(~
a' a'
axx = -p+2j.t-,
ax a-- = -p+2j.t-,
az
a
xz
=
az +~)
f"" ax
P_ B!!...- = _B(au + aw) (7.3)
Pa ax az
where B is the bulk modulus and j.t is the viscosity. Note that the stress is related
to both the volumetric behavior as well as the strain rate.
7.1 Acoustic Wave Motion 245
These equations are more easily manipulated in the frequency domain, and in
their spectral form appear as
~ au
P +p,-+p,-=O
aw
a ax a aZ
aa aa aa aa
--a:;- + az =
xx xz 2~ xz zz 2 ~
-PaW -- + -- = -PaW W
U,
aX az
~
=-
~ . au
p + 2JL1W- , ~ = JLiw -(au + -aw)
O'xx
ax O'xz
az ax
~ ~ . aw ~ p (au aw)
O'zz - -p + 2JL1W- ,
az p = B Pa = -B ax + ih (7.4)
Rearranging these equations, the spectral form for the displacement becomes
(7.5)
~
U=-+-,
alP aiI ~
W=---
alP aiI
ax az az ax
which leads to
Note that k~s is entirely imaginary and depends linearly on the frequency - this
means that the response is similar to diffusion (or Fourier heat conduction), con-
sequently, the shear response will be localized to the boundaries or where the
disturbance is initiated. The other mode exhibits dissipation due to the viscosity,
but when the viscosity is negligible, the acoustic wave speed is nondispersive and
given by
a _ W
C = (B
ka yPo
Table 7.1 shows some typical property values for air and water. Later, we will
consider the interaction of acoustic waves with plates; for comparison, Figure 7.3
shows the phase speeds of waves in a thick steel plate and a thin aluminum plate.
246 Chapter 7. Structure-fluid Interaction
.4
phase speed, CICo
.3 1=-------------:::;::00.,...::;.'---------- water
.2
coincidence
2.Smm aluminum
/
.1 rL-~~========== air
.0 Freq [kHz]
LI.-...........'-'-'.......................I....:I.................................................................................................................................................
Inviscid Fluid
A common assumption for most fluids it that they are inviscid; this leads to certain
simplifications. The constitutive relations become
O'xx--p=-
A A P,
B- O'zz = - p = -
A A P,
B-
Pa Pa
leading to the momentum relations
ap
- = PaW
2A
ax U,
Combining these with the expression for P/ Pa gives the Helmholtz equation gov-
erning the pressure field as
or (7.7)
{p; := = Pa w2u } , -
ap
aZ = PaW
2
W
A}
7.2 Plate-Fluid Interaction 247
This set of equations will be our primary equations governing the dynamic response
of the fluid. Note that it exhibits only one mode of wave behavior.
In the subsequent sections, we will consider structures composed of multiple
panels modeled as collections of plane surfaces or plates. There a cartesian co-
ordinate system local to the plane will be used. Under these circumstances, for
example, a distributed pressure source can be represented by
The wavenumber behavior of kz is similar to the first plate mode in Figure 6.5. In
the case when P is unity, this corresponds to a concentrated pressure source; the
representation, however, is valid for more general pressure distributions. A similar
representation for the components of the motion gives the displacements in the
fluid as
_ -ikz_
w=--p, (7.9)
w2 Pa
Since our interest will be in fluids contacting plates, consider the fluid to be in
contact with a surface located at z = 0 and whose normal displacement is ill, then
the first of the above equations indicates that the displacement and the pressure
are related through convolution integrals since, at Z = 0,
In other words, the pressure in the fluid at any point is related to the displacement
response at every point on the surface. It is this convolution relation that is the
source of many difficulties in the structurelftuid interaction problem.
The loading on the plate has two parts: q (x, t) is the applied distributed transverse
loading and p(x, z = 0, t) is the distributed fluid pressure loading. These equations
are complemented by the fluid equations
248 Chapter 7. Structure-Fluid Interaction
L: IUid B, Pa
:
~IY t"tt
++ftt
q =-p
q = qe
plate: D, ph
~
Figure 7.4: Geometry of the plate with fluid loading.
(7.11)
z= 0) ,
a p(x, z = 0) 2 ~
(7.12)
Wplate(X) = Wfluid(X,
az
--=--,----'- = W PaWplate(X)
At this stage, we do not know the wavenumber k. Substituting for the displacement
and pressures into this plate equation gives
7.2 Plate-Fluid Interaction 249
There are three unknown amplitudes W, Pr, Pt and one unknown wavenumber
k; we need another equation. We get this from the boundary conditions at z = 0,
namely,
ap+
-- = PaW
2 A
W or
az
and
or
where continuity of displacement between the plate and fluid is assumed. When
we combine these equations together, they must all have a common phase behavior
with respect to x, hence we conclude that the wavenumbers are related by
-2Pi
W= ----------~----------~ (7.13)
2p w2
Dk:S: - phw2 + iWTJh + . a
-lkaCe
and the reflected and transmitted pressures obtained from
_ _ Pa W2W _ PaW 2 W-
Pr = Pi + -.-k-C- , PI = C
-:------k
-I a e 1 a 0
Note that if the fluid is on one side of the plate only, then the factor of 2 in the
w
denominator of is changed to unity.
Figure 7.5 shows the variation of transmitted and reflected pressures for an
incidence angle of f) = 40°. These responses are surprisingly complicated. To help
understand them we will consider some special cases. If the plate is very rigid then
w~ 0 and we get Pr = Pi and PI = 0, and the total pressure acting on the plate is
2 Pi. This is true for any angle of incidence. This is the so-called blocked pressure.
For flexible plates in the limit of W = 00 we also have w ~ 0 and Pr = Pi, Pt = O.
That is, the plate is behaving as if it is very rigid. Conversely, in the limit of W = 0
we have Pr = 0 and Pt = Pi; the plate is transparent to the wave.
Figure 7.5 shows that at a certain frequency there is a significant change in
behavior. To explain this, first consider when the incident angle is normal, then
k = 0 and the displacement is
-2p,
W= --~--~----
-phw2 + iWTJh + 2iwPaca '
250 Chapter 7. Structure-Fluid Interaction
- transmitted
........ reflected
Figure 7.5: Transmitted and reflected pressures for an incident angle of (j = 40°.
since W / ka = Ca' The responses are dominated by the inertia of the plate and the
stiffness has no influence. The reason for this is that the plate wavelength A = 2rr / k
is infinite and hence the plate is not in flexure. Consequently, the plate is moving as
a rigid body. Also note that the inertia of the fluid appears like a viscous damping
term.
For arbitrary incident angles, the mass and stiffness contributions are oppo-
site in sign and hence there are combinations when it is possible for the reactive
component to be zero. That is, when
or
The frequency, W co , at which this occurs is called the coincidence frequency. Thus
for a given angle of incidence there is a unique coincidence frequency Wco; however,
since So cannot exceed unity, there is a lower limiting frequency for the coincidence
phenomenon given by
We
2!Ph
= Cay
c~ /
D = coh y 12(1 - v 2) (7.14)
PaCa Pa Co J 2
E == - - = - - / 12(1 - v ),
phwc P Ca
The parameter E, called the intrinsic fluid loading parameter, is the same for all
plates of a given material embedded in a given fluid. This is typically small with
values such as
In the examples that follow, we will take aluminum/air as the case of light fluid
loading and steel/water as the case of heavy fluid loading.
The pressure has a similar representation, except that we must realize it is two-
dimensional:
(7.16)
m
ap 2-
"""ih = PaW W,
(a) (b)
L=100mm
x=O x=O
x=O.Sm
x=1.0m
x=1.0m --1'1111111111111111111
x=1.Sm x=1.5m
x=2.0m
" " I
o. 2000. 4000. 6000. 8000. o. 2000. 4000. 6000. 8000. 10000.
Figure 7.6: Time domain responses for the line loading of an infinite aluminum plate in air.
(a) Plate responses w(x, t), (b) fluid pressures p(x, z = L, t).
As with the case of a plate in a vacuum, q", is chosen to be 1.0 to represent a point
load at x = O. Figure 7.6 compares the w-velocity responses for an aluminum plate
in air and the resulting pressures 100 mm from the plate.
The plate responses look similar to those of the beam and plate encountered in
Chapters 3 and 6, respectively. The pressures, however, exhibit an unusual trailing
behavior. In addition, the period of the zero crossings is almost constant whereas
for the plate velocity they increase. Figure 7.7 compares the same responses but
in the frequency domain. The most striking feature of the pressure is the spectral
peak in the vicinity of 5 kHz - this corresponds to the coincidence frequency as
seen from Figure 7.3. The peak gets sharper further away from the impact site.
What we also notice is that the amplitudes in the vicinity below coincidence are
diminishing rapidly. This is seen through the exponential term
(a) (b)
L= 100mm
~
x=O x=O
~
±
x=O.5m x=O.5m
~
x=1.0m x=1.0m
x=1.5m x=1.5m
x=2.0m
I Freq [kHz]
~I, ,
x=2.0m
, ,I
III II II II I I II I, , ,I II I II I I II I I II IIII II
o. 2. 4. 6. 8. 10. 12. o. 2. 4. 6. 8. 10. 12. 14. 16.
Figure 7.7: Magnitude of the frequency domain responses. (a) Plate responses liww(x)l,
(b) fluid pressure lfi(x, z = L)I·
plates: D, ph t
I
I W2
Lx
L fluid: B, Pa
t
t
I Wj
P
Figure 7.8: Geometry of the double plate system.
q2 + p(z = L) (7.18)
254 Chapter 7. Structure-Auid Interaction
The pressure in the cavity has waves propagating in both z directions, hence we
represent it by
We wish to concentrate on the plates for the moment, so we will remove the
pressure in the following manner. Assuming there is continuity of displacement
between plate and fluid, the momentum relation becomes
-
ap = PaW W
2 A
or
aZ
This can be particularized to both interfaces so as to solve for the coefficients in
terms of the displacements. That is,
{1} _
B
Pa w2 jik z
-1+e 2
[1 -e] {~l} ,
e -1 W2
We are now in a position to write the relation between the pressure on the two
plates in terms of the displacements
{ - p(O) } _
+p(L)
[-I e
This, in fact, is an acoustic spectral element in the sense that the pressures at the
generalized nodes (the location of the plates) are related only to the displacements
at these nodes. Also, the form of the matrix is identical to that of the spectral
element for the rod, Equation (5.6).
We complete the solution by adding this pressure to the equations of motion for
the plates to get
(7.19)
x-O
x=1.0m
x=1.5m
x-2.0m
Figure 7.10 compares the frequency domain response and again we see the
spectral peaks in the vicinity of 5 kHz for the trim panel. These peaks are not
as crisp as in the single plate case; it seems that the cavity resonances may be
contributing to this. To help bracket this effect, consider the case of rigid walls and
only plane waves propagating in the cavity. The resonance frequencies are given
by
NCa
IN - 2L ::::: 1.7, 3.4, 5.1, 6.8, ...
where the frequency is in kHz. These resonances are occurring in the vicinity of the
frequency range of interest as well as covering the coincidence frequency, hence
we now consider the effect of these resonances a little further.
Reverberation Analysis
The thrust of the analysis is to decompose the result of Equation (7.19) and thereby
help explain the features of Figures 7.9 and 7.10. We begin by noting that, without
approximation, we can write
1
-I +e2
[1+e 2
-2e
-2e ]
l+e 2 =-
[I
0
0]
I
2e [e
+ -1+e2 -I
The plate equations become
256 Chapter 7. Structure-Fluid Interaction
~_::5m ~x=0.5m
\~ x=1.0m ~X=1.0m
x=1.5m x=1.5m
x=.2.0m x=2.0m
Figure 7.10: Frequency domain responses for the double panel. (a) Impacted plate, (b) trim
plate magnified by 10.
where
D.. = D~~ - (ai ph - iWl"/h) - PaW2/ ikz
We recognize the first matrix as describing uncoupled single plates with fluid
loading. The second matrix causes the reverberation where the acoustic wave
reflects multiple times inside the cavity.
Note that
1 2 4 e N +2 2 4
--=I+e +e +···+--~l+e +e + ...
l-e 2 l-e 2
where the approximation is most valid if there is some dissipative mechanism. We
can now write the reverberant pressures as
{ -~(O)}
+p(L) rev
= PaW22 [ e2
ikz -e
-e]
e2 (1 + e2 + e4 + ...) {WI}
W2
-p(O) =
The interpretation is that W2, for example, causes a pressure wave which impinges
on the first plate after it travels a distance L; this is reflected and travels a further
distance of 2L before it impinges again. And the process continues. The factor
7.4 Waveguide Modeling 257
of "2" is important because it is the pressure if the plate were rigid, this is the
so-called blocked pressure previously mentioned. A similar interpretation is given
for the pressure on the second plate.
A final point that is worth noting is that there are two attenuation effects occurring
in the solution. Since all materials have some damping then the waves associated
with the longer reflection paths will eventually disappear. Second, any initially
localized disturbance will expand to occupy a greater and greater volume over the
reflection paths and therefore will diminish. A useful approximation, then, is to
assume that only the very first incident wave causes a fluid loading, that is, for our
double panel we have
{-
+p(L)
~(O) }
rev
:::::: Pa W2 2 [0
ik z -e
{
~e] ~21 }
This is the blocked pressure approximation. It must be realized that this approx-
imation is separate from the self-loading effect which is already included. The
blocked pressure is a useful approximation of the fluid loading when the fluid is
light or the structure is relatively stiff.
d 4 ill(x) 2 A A
Assume now a solution for the plate of the form ill = we- ikx which leads to the
pressure response of
_ Pa w2 _
p = ----;-k w,
-I z
258 Chapter 7. Structure-Fluid Interaction
(7.20)
The third term in the first equation describes the effect of the fluid loading. A
detailed explanation of the significance of each term in Equation (7.20) is given
by Crighton [22] and the roots of this characteristic equation have been studied
extensively in References [21,23,24].
With the presence of the fluid loading, an analytical solution to this characteristic
equation is not obvious. Figure 7.11 show contours of the characteristic equation
below and above the coincidence frequency. We see that it is dominated by four
roots. It is desired to identify the roots that correspond to the plate flexural and
evanescent waves.
2.0
1.5
1.0
.5
.0
-.5
-2.0 -1.5 -1.0 -.5 .0 .5 1.0 1.5 .5 -1.0 -.5 .0 .5 1.0 1.5 2.0
Figure 7.11: Contours of the characteristic equation at 5 kHz and 15 kHz for a steel plate
in water.
!\
- - transform
- - waveguide
-------- no fluid
x=o
x=1m
x=2m
I , !
Time [f..LS] I
, , !
deviations occurring at the load site that are primarily in the low frequency range.
This is confirmed at the large x locations. We therefore conclude that in cases of
moderate fluid loading (and especially for aluminum plates in air), the branch cut
contribution can be neglected. Furthermore, the approximation is reasonable even
for steel in water if the responses away from the impact site are of interest.
It would be possible to include an approximate model for the branch cut contri-
bution, but that will not be pursued here.
Considering Equation (7.20), it can be noticed that if the fluid loading term is
set to zero, the characteristic equation of the plate in a vacuum is recovered. For
the vacuum response, the plate can be described in terms of two modes with the
corresponding spectrum relations given by
4.~---------------------------,
wavenumber, kh
3.
o exact
2.
real -approx'n
1.
···no fluid
O.t-----
-1.
real x400
-2.
imaginary x400
-3.
Freq [kHz]
4.~~~~~~~~~~~~~~
O. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
Figure 7.13: Wavenumber behavior for an aluminum plate in air. (Left) k 1 (w), (right) k2 (w ).
mode. We see that at coincidence and beyond, the effect is of increased viscous
damping -- this is consistent with the observation that the fluid is receiving more
of the energy at these frequencies. But otherwise the behavior is very similar to
the in vacuum behavior. That is, the pole contributions are associated with root k,
which corresponds predominantly to the propagating flexural wave, and with root
k2 which corresponds predominantly to an evanescent flexural wave.
There exists a third root to the characteristic equation that appears only above
the coincidence frequency, which is not seen in either of these plots. The possibility
of this root causing a structural wave is of interest and requires a more detailed
interrogation of the structure. The wavenumber transform solution for the impact of
a plate, as described in the previous section, is now used as a numerical experiment
to examine the effect of this third root. To make the wave more discernible, we
impact the plate with a force having a narrow-banded frequency range as shown
in Figure 1.7. The resulting waves disperse very slowly and the group behavior
of the waves are easier to observe. By inputting a force centered about a certain
frequency, a wave traveling at the corresponding group speed should be identifiable
if it exists.
The velocity responses for an input force centered about 15 kHz is shown in
Figure 7.15_ This figure is very interesting in that the presence of two waves can
be identified, one wave corresponds to the plate flexural wave mode and the other
corresponds to a wave traveling at the speed of sound in the fluid. Although the
fluid loading on the structure causes a third structural wave to be present, it can
be seen that this wave can only be distinguished at large distances with very large
magnification and only over a very narrow frequency range. It therefore seems
reasonable to neglect this wave in the modeling of connected structures. That is,
we assume there are two dominant structural waves resembling those for a structure
in a vacuum.
7.4 Waveguide Modeling 261
2.0--------------~
wavenumber, kh
o exact
-approx'n
.5
no fluid
-1.0
imaginary x10
-1.5
Freq [kHz) Freq [kHz)
-2.0 u........J..........J.............Lu..u.J............L.u.......J............L.u..u..L.........Ju......u u.......J...........J.........u.J............L.u.......J............L.u..u..L.........Ju......J................
O. 2. 4. 6. 8. 10. 12. 14. 16. 18. 20D. 2. 4. 6. 8. 10. 12. 14. 16. 18. 20.
Figure 7.14: Wavenumber behavior for a steel plate in water. (Left) kj(w), (right) k2(W).
Notice that if the fluid loading term is set to zero, kl and k2 expressed in Equa-
tion (7.21) can be found from the terms inside the first and second set of brackets,
respectively. By assuming that the appropriate roots will also come from these
terms and using kl and k2 as the first guess, approximations for the exact roots can
be written as
1
1 PaW
kl ~ ± kzl == Jk?; - tV
f3 + 2Dikzlf3 2 '
1
2 PaW
k2 ~ ±i
f3 + 2Di kzlfJ2 ' kzl == Jk?, + f32 (7.22)
These approximate solutions for kl and kl are also plotted in Figures 7.13 and
7.14, respectively. There is very good agreement between the approximations and
the exact numerical roots over the entire frequency range including the region near
coincidence.
By incorporating the fluid loading term directly into the modified spectrum re-
lations, we are now in a position to replace the double summation wavenumber
transform solution by a single summation over frequency. This is useful when an
262 Chapter 7. Structure-Fluid Interaction
x=1m
x4 x=2m
x16 x=3m
Figure 7.15: Response at large distances on the structure showing the presence of a wave
traveling at the sonic speed Ca ~ 1.5 km/s. Note the magnifications used.
enclosed structure is viewed as having two distinct regions; an interior region and
an exterior region. The interior region has loadings due to both the fluid and the
radiation from the vibration of other plates. However, the exterior region only ex-
periences loadings from the fluid. Using the modified spectrum relations, the fluid
loading for the exterior problem can be accounted for without considering the fluid
response. In this way, the solid/fluid interaction problem is partially decoupled.
We conclude this section by illustrating the difference the waveguide formulation
makes by reconsidering the example of the impact of an infinite plate.
Although we have solved essentially the same problem before in Chapter 3, we
will do it again in the new variables. The transverse displacement of an infinite
plate with two forward propagating waves are expressed as
w(x) = Ae-ik,x + Be- ik2X
where kl and k2 are the modified spectrum relations presented in Equation (7.22).
The boundary conditions for this problem are that at x = 0 the slope of the plate is
zero (ow/ox = 0) and the applied load is related to the shear by
J ~ ~ o3 W
--P=V=-D-
2 ox3
The response of the plate is then determined to be
W ~ (x ) = P [-ik,X
e - -kJe -ik2X] (7.23)
2DikJ (k~ - k~) k2
This solution is to be compared to that of Equation (7.17). The biggest difference
is that the present solution does not require the summation of m. The waveguide
responses of Figure 7.12 were computed using this solution.
7.5 Radiation from Finite Plates 263
It is clear that it is now possible to solve for the response of impacted finite
plates with a surrounding fluid.
fluid: B, Po Lx
plate: D, ph tp baffle
2L
dw(L) ~
--=""2
dx
Solving for the coefficients in terms of the nodal degrees of freedom allows the
transverse displacement of the plate to be written in the form
(7.24)
where
These shape functions, when plotted, are very similar to those as depicted in
Figure 5.7 for the beam. Indeed, these become the shape functions for the Bernoulli-
Euler beam model when we consider thin plates in vacuum. Again, the shape
functions occur in pairs where 83 and g4 are the mirror images of gl and 82,
respectively.
The deflection of a semi-infinite plate can be written in a similar manner as
(7.26)
with
The complete description of this plate segment has now been captured in the two
nodal degrees of freedom WI and ~I.
As an example of using this procedure, consider the response due to the impact
of a finite aluminum plate in air. Treating only half of the plate, we have that
WI = W O , 1/11 = 0, W2 = 0, 1/12 = 0, and the shear relation! P = - V = Dwo"'. This
gives the solution
These responses are shown in Figure 7.17(a); the presence of multiple reflections
are obvious. Figure 7.18(a) shows the corresponding frequency domain behavior
where the reflections give rise to multiple spectral peaks.
Also shown in these figurs are the undamped resonance frequencies for the
vibrating plate. It is clear that the impact has excited many of the symmetric
modes of vibration. We will now look at how these vibrations are transmitted into
the fluid.
7.5 Radiation from Finite Plates 265
Time [Il.5]
I I I I I I I I I I I I I I I I I I I I I ~I~I~~I~I~I~I~I~,.,~,~,~I~,~,~~I~,~,~,~,~I
Figure 7.17: Time domain responses for an impacted aluminum plate of length
2L = 500mm. (a) Plate responses w(x, t), (b) fluid pressures p(x, z = L, t).
w(x, z) = L Wme-ik,ze-i~mx ,
m
By extending the plate deflection over the full space window of the fluid, we can
then give it the spectral representation
W(x) = L wme-il;m x
m
(7.27)
266 Chapter 7. Structure-Fluid Interaction
o resonance, sym
• resonance, anti-sym
Freq [kHz]
1 .~......, .......
u...~,~,.L. I""~,,,,,,,,,,,,,,,,,,,,-,'..J.I~,...
, ....
' ,...I~,~,...
, '-',1 1
...,....,~,~,.L.I~',....,~,.l..I.....
, '~'..LI~,.....
' '~'..J.I....
,~,.....'ul.....
, .... ,.........J
,~,..J,lu,.....
Figure 7.18: Frequency domain responses for an impacted plate of length 2L = 500 mm.
(a) Plate responses liww(x)l, (b) fluid pressure lfi(x, z = L)I.
where
8jm = Lgj(x)e-i~mXdx
These integrals are easily evaluated; indeed we need only replace the exponential
terms of hj (x) in Equation (7.25) with
We have assumed that the shape functions gj(x) are zero outside the length of the
plate element; this is equivalent to assuming each finite plate segment is baffled to
infinity. The continuity of Wi and Vti between plate segments ensure continuity of
the fluid field. In the case of the semi-infinite plate segment, the displacement only
consists of the terms WI and .(j,1, then only the first two integrals II and h need be
evaluated. Furthermore, letting L --+ 00 gives the integrals as I j = i/(';m + k j ).
In piecing together the solution, we treat the plate as having two segments with
the boundary conditions
This leads to the solution for the fluid displacement written in terms of the central
deflection of the plate as
7.5 Radiation from Finite Plates 267
Note that the representation for the first plate segment must be shifted an amount
L. The pressure is given by
The pressure responses for the fluid can be seen in Figure 7.17 for a baffled alu-
minum plate in air. Similar to the plate response, these responses indicate the
presence of multiple reflections occurring in the plate. This figure also indicates
that the baffled finite plate not only excites the fluid directly in front of it but also at
a distance along the baffle. The pressure at x = 2L, z = 0 is nonzero even though
the plate is baffled at that location - this is further indication that, in fluids, the
relation between the pressure and displacement is nonlocal. This, perhaps, is seen
even more clearly in Figure 7.1 which shows the 2-D pressure distribution at 1 kHz.
The frequency domain depiction of these responses is shown in Figure 7.18.
The structural resonances again appear as spectral peaks. It can be seen that the
frequencies at which the structural resonances were present in Figure 7.18 are read-
ily transmitted into the fluid. Intuitively, we might have thought that the response
should be largest along a line normal to the plate. The responses shown in these
plots indicate that this is not so. Furthermore, the directivity plot of Figure 7.19
shows that the direction of maximum pressure is even frequency dependent.
10kHz
5kHz
1kHz
500Hz
100Hz
Figure 7.19: Directivity patterns for impacted finite plate. Also shown are the plate shapes
at each frequency.
Far-field Approximation
The pressure response can be written in the fonn
Consider the situation when W(O is somewhat broad banded, then the integral
is most affected by the exponential tenn. Since the exponential tenn is oscillatory,
most of the contribution to the integral comes from where the phase has a minimum
value since at large phase they will tend to self-cancel. Approximate the phase in
that vicinity as
We use the assumption that the first derivative is zero to detennine the wavenumber
i:
~o where the phase has a stationary value. The integral then becomes
p(x, z) ~ W(~)e-i(<t>v+<t>~~' d~ =
(2) /fifW(~(})e-i[<t>(~o)+7T(4J
with the stipulation that <P/(~O) = O. This approximation is the method of Stationary
Phase. It is generally accurate in the large phase region which means asymptotically
large time or asymptotically large distances. See References [56, 66] for more
details on its application.
Specializing the method to our case, we have
These are the values of wavenumber that will tend to make the phase a minimum
value and hence give the main contribution of the integral. Substituting these into
the integral approximation then gives
~(} = ka cos¢
7.6 Cylindrical Cavity 269
This interesting result shows that the far-field pressure depends on angle only
insofar as w(~o) depends on the angle. For example, suppose that the displacement
shape is that of a rectangle, then we have the angle dependence of the pressure
given by
_ sin(~o) sin(kaL cos ¢)
w(~o) = - - = -'---'--
~oL kaL cos¢
Realizing that ka = W / c u , we see that the angle would change dramatically as the
frequency is changed.
Figure 7.19 shows the directivity patterns at a number of frequencies; the near
field behavior was computed from the full solution with r = 2L. It is clear that
these patterns are very sensitive to direction when the frequency gets close to
coincidence. Also shown are the deflected shapes of the plate at each frequency.
These shapes indicate an almost sinusoidal plate deflection except at the center
and edges - these are the points of significant radiation.
2R
Pressure Field
With reference to Figure 7.20, we expect a variation of pressure in the hoop (or
e) direction when a general load is applied. The Helmholtz equation in cylindrical
coordinates is
a2 I a 1 a2 ] ~ 2~
[ -a 2 + --a + -2 ae 2 P +kaP = 0
r r r r
Let the pressure be represented by a Fourier series in the hoop direction, that is,
270 Chapter 7. Structure-Fluid Interaction
The Bessel functions of the second kind, Ym, approach infinity at the origin r = 0
hence for this particular problem we will take B = O. Note, however, that if we
were looking at the exterior problem then we would take B = -iA so that the
solution is a Hankel function that coincides with an outward propagating wave.
The pressure and displacements in the fluid are represented by
ap 2~
a;: = PaW Ur or
m m
In these, the pressure and displacement are written in terms of the single function
u~ which is the radial displacement of the outer shell. We can therefore interpret
the displacement expression as a shape function relation. The contours of some of
these shape functions are shown in Figure 7.21. These contours help visualize the
standing wave that is established inside the cylindrical cavity.
Shell Response
We looked at the response of curved plates in Chapter 6; we now adapt these
equations to the complete cylindrical shell. This will be straight-forward since
we are assuming there is no variation in the (global) axial y direction. Replacing
s = RO and making the associations U r = -W, UIJ = U we get that the equations
are essentially the same as for the beam - only the effective material properties
change. That is,
7.6 Cylindrical Cavity 271
- - d2u - dw - d 3w 2
(E + D ) - - E - + D - + phw U
d(J2 d(J d(J3
- d 4w - - du - d 3u
2
D - - + Ew + D - - E - - phw W
d(J4 d(J3 d(J
qr - p(r ~ R) (7.29)
[
-(E + D)m 2 + phw2
- - 3 - 4 -
Em + Dm 3 2
PaW J(kaR)
2
]{ Um }
- =
{qo}
-
- Em - Dm Dm + E - phw - Wm qr
kaJ'(kaR)
Consider the case where there is only a normal applied load qr = q, ql! ~ 0, the
solution is obtained directly by use of Cramer's rule. The two displacements are
" (Em + Dm 3 )
u(r, (J) - ~ Il qm sin(m(J)
" -(E + D)m 2 + phw 2 _
w(r, (J) ~ Il qm cos(m(J)
272 Chapter 7. Structure-Fluid Interaction
These displacements can be reconstructed in the usual manner. We will leave this
to the next chapter where the variation in the axial direction is accounted for. What
is of more direct interest here is the role of the fluid loading as it affects the cylinder
in comparison to its effect on the infinite plate and the two parallel plates.
Fluid Loading
This is our third generic fluid loading geometry. The relevant equation in each case
is:
single:
parallel:
cylinder:
2 1
single: PaW --:--k
I z
1 + e- i2k ,L
parallel: p w 2 _ _ _ _,...,---_
a ikz(l _ e- i2k,L)
2 J(kaR)
cylinder:
PaW kaJ'(kaR)
In each case k z = Jk~ - ~~, ~m = m2rr / W. The variation of these pressure contri-
butions against frequency are shown in Figure 7.22 for a few values of wavenumber
~m'
7.6 Cylindrical Cavity 273
- infinite
...... parallel
-cylinder
~~~~F~~~~~~
Freq [kHz]
o. 5. 10. 15. 20. 30.
Figure 7.22: Comparison of pressure contributions from infinite plate, two parallel plates,
and the circular cylinder.
To further elaborate on the plots, look just at the portion of the pressure term
that is specifically different. We have
single:
parallel:
cylinder:
The second form in each case corresponds to the appropriate expansion of k z • For
each m, a frequency spectrum of k z goes through zero once. For the parallel plates
this occurs where
When m = 0, this corresponds to the characteristic equation for the free vibrations
of the fluid with free boundaries. Similarly, the zeros occur for the cylinder when
When m = 0, this corresponds to the characteristic equation for the free vibrations
of the fluid with free boundaries, and as m -+ 00 it corresponds to the free
274 Chapter 1. Structure-Fluid Interaction
vibrations of the fluid with fixed boundaries. Note that there are nO spectral peaks
for frequencies such that
ka <; or w < 2rrmca/W
where W is the space window. At these frequencies. the pressure waves are evanes-
cent.
Although we only focused on the pressure contribution of the fluid loading.
this example serves to demonstrate the rather complicated interaction between the
wave phenomena and the structural dynamics that can occur in coupled structural
systems.
Problems
7.1 Show that the far -field pressure radiated from a piston of radius a is
• 2 iiJJ 1(kaa sin 0) e- ikaT
P = PaW .
kaa smO r
- Reference [66]. p. 97
7.2 Show that the far-field pressure radiated from a point loaded infinite plate is
• 2 iiJ cosO
p(r, 0) = PaW - - - - - - - - - - - - - - , - -
1 - ikah(P/Pa)cosO[1- (W/Wc)2sin4 0)
P = Iwu[1l'a]
2[ -iPaca
2 k L +
i(wZM - K)]
(2)2
1l'a tan a W 1l'a
The thin-walled structures of interest in this chapter are generally enclosed regions,
as typified by an aircraft fuselage. They are modeled as combinations of folded
plates and cylindrical shell segments as shown in Figure 8.1. Other examples of
structures which fall into this category include plates with stringers, corrugations,
channels, ducts, troughs, open and closed thin-walled tubes with one or more cells.
As pointed out in Chapter 5, the only way to efficiently handle problems with
complicated boundaries and discontinuities is to develop a matrix methodology
for use on a computer. The approach we use here is similar to that of the spec-
tral element method developed in Chapter 5 for the waveguide analysis; how-
ever, for the present extended structures, the stiffness is established in the fre-
quency/wavenumber domain.
The method to be developed has much in common with the finite strip methods
of Cheung [14] and Eterovic and Godoy [49]. There are significant differences,
however. Our elements use the exact dynamic solution and hence have the advan-
tage that there is no restriction on size or frequency range - they are not strip
elements.
Figure S.l: Complex multicelled thin-walled structure modeled as a collection of flat and
curved extended elements.
275
276 Chapter 8. Thin-Walled Structures
Shape Functions
The Helmholtz potentials for both symmetric and anti symmetric loadings are given
as
<l>nm(X, y)
Hznm(x, y)
m
(8.1)
where L is the length of the element as shown in Figure 8.2, and k p and ks are the
wavenumbers as given in Chapter 4. The displacements u and v, therefore, have
spectral representations with the kernels
will drop the ± and it need only be necessary to make the change ~ -+ -~ as
appropriate. The subscripts n and m are cumbersome to carry along. therefore in
the subsequent developments these will be dropped; their presence will be inferred
from the tilde.
To get the shape functions. we replace the above coefficients with the frequency
and wavenumber dependent generalized nodal degrees of freedom U I. VI. U2. and
V2. Although we have only four coefficients. we have two functions and this results
in two sets of shape functions. These can be arranged as
The frequency and wavenumber dependent shape functions for the two-noded
element are significantly more complicated than those we have looked at before
and are given as
8r(x) = ikJ"z;] (x) + ~h;;(x). (8.4)
and
II = (klk2 - 52)(el - e2). !z = (klk2 + 52)(1 - ele2)
Ll == I? - If. el = e- ik , L. e2 = e- ik2L
Figure 8.3 illustrates the wavenumber dependence of the shape functions where
5 = 2n m / W and W = 4 m. The shape functions occur in pairs where 83 and 84 are
the mirror images of 81 and 82. respectively. Note that as m increases each shape
function exhibits behavior localized to the node. This occurs because both kl and
k2 are complex for large m as illustrated in Figure 6.5.
Working in the analogous way. we get for the semi-infinite plate
(8.6)
where
[k l k 2e- ik ,x + 52e-ik2X1/ Ll • 8~(X) = ik l 5[e- ik ,x - e- ik2X 1/ Ll
[klk2e-ik2X + 52e-ik,x]/ Ll , 8r(x) = ik25[e-ik2X - e-ik,x]/Ll
_ klk2 +5 2 (8.7)
278 Chapter 8. Thin-Walled Structures
1.
". -m=O
\ •... •.... m=100
....•.•.
-m=200
\
". \
\
'.'.'.'.
'.'. ·····m=300
\
g1
.
~\ g2
.............................. '.'
~~• .lJ
o. ••••
.------:c~=---=...........- - r~....................................
". .~ . . . . . . ". g1x2
~,............·-;:~-....pos~ion (xlL]
I, , , , I , , , , I , , , , I , , , , I !, , , .! ' , , , I " ' , !, , " I , , , , !, , , , I , , , , I " , , I , , , , I " ' , I
o. 2. 4. 6. 8. O. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
Figure 8.3: Membrane shape functions at 5 kHz. At left is g"(x) and at right g"(x),
We reiterate the significance of the shape functions is that they allow the complete
dynamic description of the element to be captured entirely in terms of the nodal
degrees of freedom - the functions themselves are used to interpolate (exactly)
between the nodes.
(8.8)
where now the dynamic stiffness matrix [knm ] is of size [4 x 4]. is complex and
symmetric, and can be written as
8.1 Membrane Spectral Elements 279
(8.9)
k13 :hnm
2ik2k~ [klk2e-ik2Lz21 + ~;e-ik2Lzl1]
k24 :hnm
nm
2iklk~ [klk2e-ik2LzlI + ~;e-ikILz2d
The determinant quantity il nm is
A very interesting aspect of this matrix is that it has a form (in terms of repetition
of entries) similar to that of the spectrally formulated beam element of Chapter 5.
This comes about essentially because both the beam and the 2-D membrane have
two degrees of freedom per node.
Figure 8.4 shows the frequency and wavenumber dependence of the diagonal
stiffnesses. Note that the higher m give significantly larger stiffnesses and that
significant "resonant" behavior is only seen at the lower m values. This is consistent
with the behavior of the shape functions.
The one noded or throw-off element is treated in the same way. The unit normals
at Node 1 are nx = -1, ny = 0 thereby giving the relation between the nodal loads
(per unit length) and stresses as
(8.10)
Substituting for the stresses (using the expressions of Section 4.1), we eventually
get the relationship between the nodal loads and nodal displacements as
10.[1~
1............•............•..
Figure 8.4: Dynamic stiffnesses normalized as kll /[Eh/ L(1 - v 2 )] and k 22 /[Gh/ L], re-
spectively.
Nt.' I 0 (8.12)
8.1 Membrane Spectral Elements 281
The last condition corresponds to an arbitrary normal loading with S(y) represent-
ing a spatial variation of the load having the Fourier coefficients am, bm, and P(t)
the time variation of the load with the Fourier coefficients fin.
Simple Example
As a simple example. we will consider the response of an infinite sheet subjected
to a concentrated load. This is a problem whose solution can be represented in
many forms - it is instructive to see how it relates to the element approach.
We conceive of the infinite sheet as made of two semi-infinite elements sharing
common uland VI displacements along the axis x = O. To help in understanding
the assemblage process, consider a Node 2 at x = +00 and a Node 3 at x = -00.
The stiffness relations for each element is
{
-
Nxxl
N xyl
}12 ~m[2~~ + 2klk2 - k~J
ik)k~ ]{~: }
{ fo!.xxl
-
N xy1
} 13 -~m[2~~ + 2klk2 - k~]
iklk~ ]{~: }
The negative off-diagonal term comes about because the second element is rotated
1800 about the y axis. (The effect of orientation is developed more formally later
in this chapter.) The assemblage process is simply a matter of imposing nodal
eqUilibrium. In this case, of course, we are dealing with a generalized node. but
the process is completely equivalent to what was done in Chapter 5. We need
balance of both the x and y loads, that is.
~ [2!~~~ 2i~k~ ] { ~: } = { ~ }
This gives the solution
_ .6.-
Ul = P VI = 0
2J.Lhi k2k~ ,
This is a result obtained by Lamb [72]. It is surprising that the V displacement
along x = 0 is zero.
The actual responses are reconstructed by using the double summation. Rather
than do that, consider the response at an arbitrary position. Using the shape func-
tions and nodal solution. we get
If we now replace the functions of y with their exponential fonn, then we can write
the radial displacement for the first mode as
We can put this in a fonn suitable for approximating the far field by the method of
stationary phase (this approximation was developed in Chapter 7) as
"r(X, z) =
roo \{I(~)e-i<l>(~) d~
Loo
Specializing the method to our case, we have
\II =
~o = k p sin 9 ,
",(r, 9) ~ -i Pcos 9
2(2J.t + J..)h
J
2n e- i [k,.r+lTj 41
kpr
This shows that the displacement diminishes as 1/,.;r in the far-field - a conclu-
sion we also drew for the flexural behavior of an infinite sheet subjected to a point
load.
Displacement Representation
The general spectral solution for the dynamic response of the plate in vacuum can
be written as
8.2 Spectral Elements for Flexure 283
W (x ,
A
~
y) = L..-m [Amne-ik,x + Bmne-ik2X (8.13)
The subscripts m and n are understood on the wavenumbers kJ and k2, and L
is introduced to associate the reflections coming from a boundary at x = L. For
convenience, most of the subsequent expressions will omit the summation and
omit the m and n subscripts; both of these will be implied.
Since the acoustic problem is also of interest, we will just briefly describe here
how fluid loading can be incorporated into the response of plates. Developing
a full matrix formulation for the acoustic problem in finite enclosed structures
is beyond the intended scope of this book. The primary difficulty is associated
with the multiple radiation loadings and we would need to develop an appropriate
"bookkeeping" scheme to keep track of the various loadings. The problem is com-
pounded because the various radiating surfaces are at different angles. What we
will show here, however, is that the fluid loading effect itself can be incorporated
by simply modifying the spectrum relation. The solutions are then valid for the
exterior problem where the fluid is on the "outside" of the structure and acts to
conduct energy away from the structure.
By utilizing the general equation of motion for the plate element presented in
Equation (6.14) and replacing the loading in terms of the pressure, a characteristic
equation similar to that of Equation (7.20) can be determined as
kZ =Jp-p_C2
a Sm
(8.14)
These are very easy to incorporate into the structural response. In the subsequent
developments, we take that kJ and k2 are general enough to handle these cases.
Shape Functions
First consider the throw-off element with C = D = O. The end conditions require
that
284 Chapter 8. Thin-Walled Structures
The constants A and B can be rewritten in terms of the nodal displacement WI and
rotation {J x I. Doing this, we can therefore rewrite the displacements in the form
(8.15)
The functions 8i(X) are the frequency dependent plate shape functions and are
given by
(8.16)
The complete description of this element has now been captured in the two nodal
degrees of freedom W1 and {JI'
The development of the two-noded element, shown in Figure 8.2, follows along
the same lines. The end conditions on the element are
dw(O) - dw(L) -
w(O) = WI, ~ =1/F\, w(L) = W2, -- = 1/F2
dx
Solving for the coefficients in terms of the nodal degrees of freedom allows the
transverse displacement to be written in the form
(8.17)
The frequency and wavenumber dependent shape functions for the two-noded
element are significantly more complicated than those for the throw-off element
and are given as
8\(X) [r\h\(x) + r2h2(x)]/ ~
82(X) [r\ h 3(x)+ r2h4(x)]/ ~
(8.18)
83(X) [r\h 2(x) + r2 h l(x)]/ ~
- -
84(X) = -[r1h4(x) + r2h3(x)]/ ~, ~ = r? - r}
where
h\ (x) ik 2 [e- ik ,X - e2e-ik,(L-x)] - ik1 [e-ik,X - e\e-ik,(L-x)]
h 2(x) -ik2 [e2 e - ik ,x - e-ik,(L-x)] + ikl [e\e-ik,x _ e-ik,(L-X)]
h 3(x) [e-ik,x + e2 e - ik ,(L-x)] _ [e-ik,x + e\e-ik,(L-x)]
h 4(x) [e2 e - ik ,x + e-ik,(L-x)] - [e\e-ik,x + e-ik,(L-x)]
and
r1 =i(k2-kl)[1-e1e2L r2=i(k2+k\)[el-e2]
e1 = e-ik,x, e2 = e-ik,x
Functionally, these are the same shape functions used in describing the beam
waveguide behavior; the difference is that the wavenumbers k1 and k2 have a
dependence on l; .
Figure 8.5 illustrates the wavenumber dependence of the shape functions. The
shape functions occur in pairs where 83 and 84 are the mirror images of 81 and
82, respectively. Note that the higher m tend to give distributions localized to the
node.
8.2 Spectral Elements for Flexure 285
1
-m=O
.... -m=100
-m=200
·····m=300
g1
g2
position [X/L]
I. • I , I I. . I I. I.
"
o. 2. 4. 6. 8. o. 2. 4. 6. 8. 10.
Figure 8.5: Shape functions gi(X) (left) and their derivatives (right) for a plate in flexure at
1 kHz.
This is a dynamic stiffness relationship between the loads applied at the edge of a
semi-infinite plate, and the deflections along this edge.
The process to obtain the stiffness relation for the double noded element is the
same as for the throw-off element and leads to
(8.20)
The matrix [k] is the dynamic plate element stiffness matrix; its individual terms
are
and
ZII
1_ e-ik,Le-ik2L, ZI2 = e- iklL _ e- ik2L
Notice that [k] is a symmetric complex [4 x 4] matrix. This matrix relates applied
forces and moments along the edges to transverse displacement and rotation there.
Figure 8.6 shows the variation of the two main diagonal terms as a function of
frequency. The main point to note is that the number of spectral peaks diminish as
m is increased. This is in keeping with the idea that the plate stiffens as m increases.
-m=O
.... ·m=50
-m=100
m=150
10
k11 k22
Frequency [kHz]
I I I I I I I I I I I I I I I I I I I I I I I I
Figure 8.7: Assemblage of 3-D folded plate elements. The complete connection edge is a
generalized node.
[R] =
[
cos8
Si~ 8 o
o
o
- sin8
o
cos8
o
0
0
0
1
l
Chapter 8. Thin-Walled Structures
The resulting general plate element models both in-plane and out-of-plane waves.
Note that the global rotation about the y axis, ;Py, is equal to the negative local
v,x. To illustrate the procedure for constructing the structural stiffness matrix,
assume that the structure is in a state of (dynamic) equilibrium, thus each node
Goint) must also be in a state of (dynamic) equilibrium. Consider a free body
diagram containing a node under the action of externally applied loads internal
nodal loads (Nx_'" Nxy , ... , St.,). Replacing the member nodal forces in terms of
the (augmented) stiffness matrices gives equilibrium as
(8.21 )
where [K] is called the dynamic structural stiffness matrix. A more detailed
description of this assembling procedure for static problems is given in Refer-
ence [40].
The applied loads are taken in the form
8.3 Folded Plate Structures 289
where S(y) is the distribution shape along a node. We then have the structural
stiffness relation
where {S P} is a vector indicating the degrees of freedom which are loaded. For a
single impact, say, {S} has zeros at all indices excluding the term representing the
load point and becomes {S}P.
16 x 512 x 512 x 8 x 2 = 67 MB
The spectral method, like other techniques for dynamic analysis of plate struc-
tures, is computationally intensive; it requires several hundred thousand, or even
millions, of solutions to a small system of linear equations. Unlike other tech-
niques however, the spectral method gives exact results over large regions and
automatically produces the frequency response function.
Another benefit of the spectral element method of finding dynamic behavior
of structures is that it is ideally suited for multiprocessor computers of the single
instruction multiple data (SIMD) type. The many solutions to the small dynamic
stiffness matrix equations cited above are computations which are completely in-
dependent of each other, and can therefore be performed simultaneously. Although
the efficiency of most parallel algorithms usually decreases as the number of pro-
cessors increases, the spectral element method maintains maximum efficiency on
the largest of multiprocessor computers [26,27,28]. As computer hardware ar-
chitectures become increasingly parallel the methods underlying the approaches
shown here will gain tremendous computational advantages.
Box Responses
Our test structure is a thin-walled cylindrical box with a square cross section having
sides oflength 102 mm as shown in Figure 8.8. We model the box as a collection of
folded plates with the folds located at the nodal dots. Later, to make the example
more interesting, we increase the thickness of the plates so as to make the box
have properties (in the context of our applied force history) of both a folded plate
structure and a simple beam; the relationship between these two structures gives
an interesting insight into the behavior of complex structures.
The top face of the box is impacted at its center with the smoothed force history
of Figure 1.7. Velocity responses at three locations on an x - z plane through the
impact site are shown in Figure 8.9. The sets of curves labeled "w" are flexural
responses. The "u" curve is an in-plane velocity response in the side panel which
is induced by the original flexural wave interacting with the folds. In-plane waves
travel more rapidly than flexural waves and their effects will therefore be seen
much earlier.
wtop
Time [f..LS]
!
Apart from the arrival times, there is very little insight immediately obvious from
these plots. We now show the effect of the various folds on the reflections and in
doing so illustrate some of the unique features of the spectral element method.
Evolution of Resonance
An interesting study is to look at the effect of adding folds to the plate. Figure 8.10
shows the sequence of models of the box beam where an additional comer is
added at each step. For clarity, only the right half of the symmetric sections are
shown. The four models in this figure have the same properties as the complete
102 mm x 102 mm structure. Throw-off elements (denoted by elements ending
with a squiggle, I) absorb all waves transmitting through the last comer in each
292 Chapter 8. Thin-Walled Structures
.- ...~
o 1 2 3
Figure 8.10: Folded plate sequence geometries. The structures are symmetric about the line
of action of the force.
j folds
=c=~O
2
full
Frequency [kHz)
I~,~~~~~~~~~~,ITime[~)~~'~'~I~,~,~'~I~"~'~I~~I~"~,~,~I~,~,,.,~I~,~,~,I
o. 1000. 2000. 3000. 4000. O. 2. 4. 6. 8. 10. 12. 14.
Figure 8.11: The evolution of structural resonance at the impact site. (a) Time responses,
w(t), (b) system frequency response 161.
model. The sequence helps to explain the responses in the lowest curves of Fig-
ure 8.9.
Each fold adds a new set of reflections. These reflections are discernible in Fig-
ure 8.11 (a). A deeper insight is gained by looking at the system transfer functions
for the same cases where in the frequency domain we have
Ii = OP
What we see is the formation of spectral peaks. This is essentially the vibration
resonance behavior. It is interesting to note that while the time domain responses
have established themselves by the third fold, the frequency plots show that the
spectral peaks will get sharper. This is because the spectral element method is
showing the system response not modified by the spectrum of the input loading.
Figure 8.12 shows the same information but at a location 200mm down the
cylinder. Again we see the evolution of the resonances, but we also see a very
8.4 Structural Applications 293
folds: 0 ~-------------------------~
full
Frequency [kHz]
L...o.................JL-<-..............J..................I....L.........'-'-.-...J'I Ti me [J.I.S]1....,....,....' ....
' ,-,I'...'....' ....' ....
1 ....
' ,...........
1....
' ....
' ,....,....1........
, ....
' ,...1....,....' ....
' I....,....,....' ......
....J, 'I
o. 1000. 2000. 3000. 4000. O. 2. 4. 6. 8. 10. 12. 14.
Figure 8.12: The evolution of structural resonance at y = 200 nun. (a) Time responses,
w(t), (b) system frequency response IGI.
--top
--middle
------ bottom
------ beam
x=700mm
Time [/:\s]
O. 1000. 2000_ 3000_ 4000. 5000.
Figure 8.13: Box displacements.
the box beam side panels and Timoshenko results show a pronounced attenuation
while the top and bottom faces show enhanced response. The transition frequency
occurs at the first natural frequency of the box beam's cross section; to illustrate this,
we compare in Figure 8.14 the resonance frequencies of a square frame having the
same properties as a unit length of the box beam cross section (acting under plane
strain conditions). The multiple resonances of the frame coincide, approximately,
with the beginning of the peaks in the box beam response.
top
side
00 00 0
I , , , , I , , , , I , , t , I , , , I ,
The pertinent question arises: can the box beam be approximated by a Timo-
shenko beam model? The significant difference between the two models is that
the folded plate model, being three dimensional, has many more modes than the
simpler beam theory. The issue is not so much whether or not the modes are ex-
cited but whether they appear in the monitored degrees of freedom. For example,
the vertical (global z direction) behavior of the centerline of the box beam side
8.5 Segmented Cylindrical Shells 295
panel is similar to that of the Timoshenko beam; however, its out -of-plane behavior
bears no resemblance. An understanding of when the Timoshenko theory is not
applicable can be obtained by considering the vibrations of the cross section. The
first resonance occurs at about 2.8 kHz (which is inside the input range of 0 to
12 kHz) and therefore is significant. If, however, the cross section were reduced
to 25 mm x 51 mm, say, while keeping the same amount of material the first reso-
nance would increase to about 15 kHz and the cross sectional behavior would not
be significant; a Timoshenko theory would then suffice.
We saw that the membrane behavior of a flat plate required four generalized
degrees of freedom, while the out-of-plane flexural behavior also required four.
A folded plate structure, which exhibits both membrane and flexural behavior,
then required eight degrees of freedom. A segment of a curved cylindrical shell,
as shown in Figure 8.15, exhibits behavior similar to that of the folded plate and
therefore has eight degrees of freedom. For the folded plate, we assembled the
[8 x 8] stiffness as a combination of two [4 x 4] elements plus a rotation matrix.
But the curved element has all displacements coupled and hence we have to tackle
the stiffness matrix directly as an [8 x 8] system. This is too cumbersome to do
explicitly; consequently, we layout a computer based method for establishing the
shape functions and subsequently the stiffness matrix. This, of course, adds to
296 Chapter 8. Thin-Walled Structures
This is homogeneous and therefore, at best, we can only get amplitude ratios.
For example, we can solve for the remaining two terms as a function of U o . In
anticipation of a later need, we augment the system with
ow
--1/1=0 or
os
and write the solutions at a particular wavenumber k j as
where the symbol eI> indicates an amplitude ratio. Although the vector {eI>} shown is
normalized with respect to u o , it is possible for other modal vectors to be normalized
differently. This must be done for each mode k j and hence there are eight vectors.
We choose to represent these as [eI>] = [[ eI> A], [eI> B)] where
and [eI> B) is the same but evaluated at the wavenumbers -k j • That is, the [4 x 4]
partitions [eI>A] and [eI>B] are evaluated at +k j and -k j , respectively; they are
fully populated and typically are not symmetric. The normalizations are arranged
so that the uncoupled flat plate solutions are easily recovered.
For each mode, the corresponding amplitude Uoj is undetermined; to make the
notation resemble what we have already used, we will label each of the these
8.S Segmented Cylindrical Shells 297
A, B, .... The solution for the displacements can then be expressed as Consider
a segment of shell of length L in the hoop direction. We begin by expressing the
displacements as
u(s) = A<I>lI e- ik ,s + B<I>\2e- ik2S + ... + G<I>l7 e- ik3 (L-s) + H<I>lse- ik4 (L-s)
with similar expressions for V, and v" but involving the amplitude ratios <l>2j'
w,
<l>3j, and <l>4j, respectively. The length is introduced to include reflections coming
from a boundary located at s = L. This displacement solution can be rewritten as
where the diagonal matrix of exponentials and the vector of amplitudes are give
by
j IA} = {n IB} ={ ~I
We wish to replace the vectors {A} and {B} in terms of the nodal displacements at
s = 0 and s = L. That is, we introduce
{{A}}
{B}
= [G) {{Uh} = [[G
(uh [G 11
lI ]
2
where each partition of [G) is of size [4 x 4]. We are now in a position to write
the displacements in terms of the shape functions. They are
There are a total of 4 x 4 x 2 = 32 shape functions. While not obvious from the
above, it turns out that, even in this general case, the collection of shape functions
associated with the degrees of freedom at the second node are the mirror image
of those associated with the degrees of freedom at the first node. The results for
the throw-off element are simply those associated with [G I d. These formulas can
be used to recover all the shape functions already derived in this chapter and in
Chapter 5.
By way of example, Figure 8.16 shows the g33(S) cos(~mY) shape function of a
270° shell segment; this shape function is associated with the WI degree of freedom
and thus can be plotted as a radial displacement off the original shape. The other
shape functions behave in a similar manner. It is clear from this figure and from
the above developments, that once the nodal degrees of freedom are determined
then the shape functions can be used to compute the responses at any intermediate
locations. This is a crucial attribute since the segments can be very large.
8.5 Segmented Cylindrical Shells 299
r
these resultants in matrix form as
(8.26)
The stiffness matrix in this case is of size [4 x 4]. These stiffness relations can be
used to recover the results already given for flat elements.
For illustrative purposes, Figure 8.17 shows the normalized k 11 , k22 , k33 , and
k44 diagonal terms for the same shell used to illustrate the shape functions. As
is typical with spectral elements, they exhibit a very large dynamic range. The
normalizations are with respect to the stiffnesses for straight thin beams [40] but
modified for plates. That is, they are presented as
- 3
k11/(Eh/L), kzZ/(Gh/ L) , k33/(12D / L ),
Note that the kll stiffness is substantially less than the static values, and it is only
after the cut-off frequency (~ 3 kHz for m = 0) does it become greater than unity.
On the other hand, k33 is always significantly larger than the static straight value;
this is because the L 3 in the denominator predicts an inordinately small static value.
300 Chapter 8. Thin-Walled Structures
K11
K22
K33
- - unity reference
--m=O
---------m=40
Figure 8.17: Normalized stiffnesses kll , k22 , k33 , and k44 , for a 270 shell segment.
0
n
Figure 8.18.
- p
2R
R = 250mm
-:u·® CD·
h =25mm
Figure 8.18: Geometry for the point loaded circular cylinder.
We form the complete cylinder by combining two curved elements of the same
material properties each of which is a half-circle. Because of symmetry, we could
model the shell with a single half-circle element or as a single whole-circle element.
In the latter case, half of the applied load must be placed at each node and the extra
conditions of" 1 = 0, ~1 = 0; "2 = 0, ~2 = 0 be imposed. The two element model
8.5 Segmented Cylindrical Shells 301
was chosen so as to verify the assemblage process. The elements are joined at
the common nodes by merely summing together the appropriate dynamic stiffness
matrix components to form a global stiffness matrix [K]. We must therefore first
rotate each element stiffness to this global system. The transformation requires the
use of a simple [6 x 6] rotation matrix [ T ] that is of the form
[ T ] = [[R(oa)] 0 ]
[R(a + 2,8)]
where R(O) is the [3 x 3] rotation matrix [40]. This takes into account that the ends
of the curved element are oriented differently to each other. In our case, if the nodal
coordinates are (Xl, Zl) and (X2, Z2), then with 2p = J(X2 - ZI)2 + (Y2 - ZI)2 and
J
q = R2 - p2, the angles are given by
Since only two elements are present, we will assign the coordinate system of one to
be the global coordinate system and rotate the other. The global stiffness matrix is
then simply formed by adding the two global element stiffness matrices together.
Typical velocity reconstructions for three points, 0 = 0°, 90°, 180°, along the
circumference are shown in Figure 8.19.
- - curved elements
o 64 flat elements
Time [us]
I I I I I I I I I I I I I ! t I I I I I I ! ! ! I I
o. 2000. 4000. 6000. 8000. 10000.
Figure 8.19: Comparison of the nonnal velocity reconstructions for the point excitation of
a closed cylinder. The lengthwise position is y - O.
For comparison purposes, the same shell was modeled using 64 flat spectral
elements. The two sets of results are indistinguishable even though there are very
many reflections. It is worth pointing out that when fewer flat elements were
302 Chapter 8. Thin-Walled Structures
used the results deteriorated as the time increased. This shows the significant
computational savings in using the curved element.
There are obvious multiple reflections occurring. The waves travel around the
circumference and interact with each other. We can get an alternative insight by
looking at the system response function {; where u = {; P and P is the input load.
Note that this facility is an integral attribute of the spectral element formulation.
From Figure 8.20, we can see that many resonance frequencies are being estab-
lished. These would be the frequencies of interest if an impulse/modal analysis
experiment vibration experiment were being performed. It is interesting to note that
the w response at 0° and 180° exhibit similar resonant behavior, but the response
at 90° is missing many of the intermediate resonances.
zero reference
90
180
Freq [kHz]
II I I I I I !!! I !! I I II ! I I I I I! 1 I! I I ! I I!! I I I I I 1 I! ! ! ! I! ! ! ! I
o. 1. 2. 3. 4. 5. 6. 7. 8. 9. 10.
Figure 8.20: System response functions for the point excitation of a closed cylinder.
A significant peak appears at about 3.3 kHz. An analysis of the amplitude ratios
shows that this peak is coming entirely from the first mode even though the response
overall is dominated by the third mode. This is the same peak that is observed in
Figure 6.20 where the curvature acts effectively as a continuous boundary and sets
up a standing wave.
also be treated. But it will have the added advantage of being formulated in the
frequency domain.
A more challenging difficulty is to make the super-element part of the current
collection of extended elements. The essence of the challenge is the delimiting
of the boundaries. The current plate elements have a very large lateral dimension
and somehow the small super-element must be attached to it. There are a number
of possibilities two of which are shown in Figure 8.21. We now try to elaborate
further on these possibilities.
Figure 8.21: Some schemes for utilizing the spectral super-element concept and attaching
them to the extended elements. Left is a super-element with large parallel edges, right is a
super-element inside an extended spectral element that has parallel and circular boundaries.
The first is to mesh the complete region between the two parallel sides and
reduce it to just the edge degrees of freedom. The size of the super-element is
too large to do this directly. therefore. the key difficulty of the problem is how to
combine extenders with the meshed region before performing the condensation.
Suppose this is done. We now need to transform the frequency domain element to
the wavenumber domain. Let us represent the stiffness relation as
where [ \11 ] represents the coefficients for the Fourier transform. If we choose the
nodal spacing for {u} to coincide with ll. y in the transform. then we have
{i} == [ k Hu} .
Because of the special nature of [ \11] the [ k ] matrix will also have a special
form. In fact. we expect it to be block diagonal. Note that the relation between
[k] and [k] is essentially that of a double Fourier series. This will require a modal
representation that has a great number of modes (because the slab is long); if.
however. we can accelerate this representation (by utilizing the extenders. for
example) then again we can handle any local discontinuity.
Problems 305
u(y) = LU m COS~mY,
m
Problems
8.1 Show that for a plane wave, the in-plane element displays uncoupled the u
and v degrees of freedom and the stiffness is given by
kllnm 0 kl3nm 0 ]
[ knm] = [ 0 k22nm 0 k24nm
kl3nm 0 kllnm 0
o k24nm 0 k22nm
8.3 Let the distributed displacement fields (U(x, Y, z)} be represented by an ex-
pression of the form u(x, Y, z) - LN(x, y, z)J{u} with similar expressions
for the other displacements v(x, y, z) and w(x, y, z). Show that the potential
energy for a general elastic body may be written as
which are general expressions for the stiffness matrix and loading.
- Reference [40], p. 169
8.4 Consider the transverse deflection of a plate and let the deflections be repre-
sented by
M 4
w(x, Y) = LYm(Y) L Nr(x)ur
m-l r-l
This expression forms the basis for the finite strip method.
- Reference [14], p. 29
8.5 Continuing the previous problem, let
[m ]m. = [m ] l phYmYn dy
One moves forward from this mystic landing to Fourier analysis, pe-
riodogram analysis, Fourier analysis over groups, and one comes
rapidly to great technological applications, and always a sense of the
actual and potential unities that lurk in the corners of the universe.
P.J. DAVIS and R. HERSH [32]
It is opportune now to assess the accomplishments of this book and (perhaps more
important) to point out what is missing.
This book has emphasized the analysis of structures in situations that facilitate
their treatment as connected waveguides. Thus, the primary effort is devoted to
developing the spectral analysis method for solving wave propagation problems in
rods, beams, and plates. It is shown that within the spectral framework, the incorpo-
ration of damping, coupling effects and higher-order theories, is a straightforward
and simple matter. The introduction of the spectral element approach formalized
the connection procedure for use on a computer and accordingly extended the
range of application of these methods to more general structures. Furthermore,
the inherent parallelism of the spectral methods make them eminently suitable for
hosting on the new multi-processor computer architectures. Additionally, the in-
corporation of experimental data in many of the examples demonstrates the affinity
of the spectral methods for experimental methods.
What emerges is an approach that exhibits a striking unity between the analytical,
computational, and experimental methodologies.
A topic introduced but not developed is that of a global approach to waves in
structures. The spectral element approach demonstrated in Chapters 5 and 8 is very
exciting because it offers the possibility of performing wave analysis in compli-
cated structures without having to resort to expensive, time-consuming cascading
methods. The efficiency of the approach resides in the fact that the only unknowns
are those associated with the connectivities at the joints and therefore are few in
number. The beauty of such an approach is that the structure can be analyzed as
a whole without worry of the detailed history of each wave in each member -
only if necessary is the detailed wave behavior reconstructed, and this can be done
quite efficiently as part of the post-processing. A further side benefit is that it brings
wave propagation and vibration analyses closer together and it makes such global
307
308 Afterword
methods as modal analysis possible, but all the time it still has the possibility of
recovering the detailed behavior. It seems that what is really needed in dynamic
structural analyses is some global or higher level evaluation of the results. The
spectral analysis approach has the potential to offer this - the frequency domain
energy measures introduced in Chapter 5 could playa role - but it is not devel-
oped as yet. The development of a global approach to wave propagation in complex
structures must be considered the next challenge.
Only barely treated in this book is the whole series of problems dealing with
waves in extended media interacting with discontinuities (such as holes or cracks)
causing dynamic stress concentrations, diffraction, and scattering. These problems
can be treated by spectral analysis but they require the development of some new
tools invariably involving multiple series (or integral) representations. These topics
are really part of the more general one of the exact formulation of wave problems
and their subsequent asymptotic approximation. While many such solutions can
be found in the literature, they generally are too cumbersome to be of direct value
within our connected structures context. An exciting possibility is to extend the
spectral super-element concept to encompass these problems. Some conjectures
for achieving this were given at the end of Chapter 8, but very little work in the
area has been done as yet. The compelling aspect of the spectral super-element
approach is its suitability for hosting on multi-processor computers.
These two problems represent different departures from the contents of this
book. The first is to a level of integration, the second to a level of more local detail.
The spectral methodologies presented in this book provide a useful middle ground
for beginning both studies.
Appendix:
Bessel Functions
Bessel functions, and the other special functions, arise as solutions to many bound-
ary value problems and their use can greatly extend the range of the spectral meth-
ods. The major properties of the Bessel functions are summarized here so as to
facilitate their use. Additional sources of information are References [1,105,124],
the book by Press et al. [102] has Fortran coding of these functions.
d2w
Z2_- + zdw
- + (Z2 - n2 )w = 0, n~O
dz 2 dz
is called Bessel's equation and has the two independent solutions
where in, Yn are Bessel functions of the first and second kind, respec-
tively. These are shown plotted in Figure A.I. Note that Y(z) is singular as z --+ 0
and that both functions tend to zero as z --+ 00. Actually, in and Yn exhibit a
damped oscillation and for this reason, the solutions are often written in terms of
the Hankel functions of the first and second kind given by
where In, Kn are called modified Bessel functions of the first and second kind,
respectively. These are shown plotted in Figure A.2 as e- z I(z) and e+ z K(z),
respectively.
309
310 Appendix: Bessel Functions
1.0 ............
.8 ..... ....... ............ ., ......• J
.6
.4
.2
.0 -2
-.2 .~--~~~~. -10
-.4
-.6 .....: ........ ;. ........:- ......... .
-.8 ... ., .. .. .. ...
-1.0 z
O. 2. 4. 6. 8. 1 O. 12. 14. 16. 18. 20.
1.0
.8 ...• v
.6
.4 .:_0
.2 /'""--.;,?--.;..........:..:
....... : _ 1
.0 ·.-2
-.2 ~.,.,c,.,;._.: - 1 0
-.4
-.6 ......... ,: ........:.
-.8 . . .. ,.,,':;,
~
. ........: ...... .
-1.0 Z
O. 2. 4. 6. 8. 10. 12. 14. 16. 18. 20.
Many equations are transformable into Bessel equations and a very useful gen-
eral form is
d2w dw
Z 2 _ + (1 + 2a)z- + (/32 Z2Y + <5 2 )w = 0
dz 2 dz
This has the solution [l05]
w = z: [C lv(/3;Y) + C2YV(/3;Y)] ,
1
Limiting Behavior
When manipulating Bessel functions, the following recurrence relations are very
useful:
2n
-Mn - M n- 1 ,
z
_2 dMn
dz
2n
--Ln +Ln- 1 ,
z
dL n
2-
dz
Appendix: Bessel Functions 311
K
.8
_0
.6 -1
..
-2
.4
:=r=~~~~~'''~'''~'.~~.. -10
.2 .-' . . . . . . . . . . . . . . '.' . . .
~ , . . , . ... . . . .:
.0 Z
O. 2. 4. 6. 8. 10. 12. 14. 16. 18. 20.
1.0 ......... " .....
.8 ......
_0
.6 _1
-2
.4 -10
.2
.0 Z
O. 2. 4. 6. 8. 10. 12. 14. 16. 18. 20.
where Mn represents any of the functions in, Yn , H;" H;, and Ln represents either
In or einTf Kn.
Bessel functions have the small argument expansions
~ ( Z2)m
1 (
In
)
Z =
(1
2Z
)n
:::0 -"4
1I
m!r(n + m + I)
1 (n - m - I)! 2
__ (lz)-n"
7r 2 ~
n-I
m!
(lz2)m
4
+ -In(lZ)in(z)
7r 2
m=O
1 1
+ 1) + 1/!(n + m + 1)}(_~z2)m _ _ __
00
__ (~z)n L{1/!(m
7r m=O m!(n + m)!
(1 )n ~ (I 2)m 1
2Z :::o"4Z m!r(n + m + 1)
~(~z)-n L (n - mm -
n-1
l
1)1
. (_~z2)m + (-I)n+lln(~z)In(z)
m=O •
00 1
+(-1)" ~(~z)n L{1/!(m + 1) + 1/!(n + m + 1)}(~z2)m _ _ __
m=O m!(n + m)!
where r is the gamma function and 1/!(n) = L lin - .5772. For very small
arguments, in particular,
312 Appendix: Bessel Functions
10 ~ 1 - ~Z2,
10 ~ 1 + ~Z2,
J
n
=::} fT cos (z _ ::.4 _ mr)
'1/;-;' 2 '
1 •
In=::} ;;::c::: e- ,
v2rrz
It is more apparent here that 1n, Yn exhibit a damped oscillation and In, Kn behave
as exponentials. The Hankel functions have the limiting behaviors
That is, they have the exponential form of the circular functions with the difference
that they decrease in amplitude for large argument. Consequently, the use of the
Bessel functions in spectral analysis of waves is similar to that of the circular
functions.
References
[19] Craig, R.R., Structural Dynamics, Wiley and Sons, New York, 1981.
[20] Cranch, E.T. and Alder, A.A., "Bending Vibrations of Variable Section Beams,"
Journal of Applied Mechanics, 78, pp. 103-108, 1956.
[21] Crighton, D.G., "The Free and Forced Waves on a Fluid-loaded Elastic Plate,"
Journal of Sound and Vibration, 63(2), pp. 225-235, 1979.
[22] Crighton, D.G., "The 1988 Rayleigh Medal Lecture: Fluid Loading - The Inter-
action between Sound and Vibration," Journal of Sound and Vibration, 133(1), pp.
1-27,1989.
[23] Crighton, D.G. and Innes, D., "Low Frequency Acoustic Radiation and Vibration
Response of Locally Excited Fluid-Loaded Structures," Journal of Sound and Vibra-
tion, 91(2), pp. 293-314,1983.
[24] Crighton, D.G., Dowling, A.P., Ffowcs Williams, J.E., Heck], M. and Lepping-
ton, F.G., Modern Methods in Analytical Acoustics: Lecture Notes, Springer-Verlag,
Berlin, 1992.
[25] Crowley, F.B., Phillips, J.W. and Taylor, C.E., "Pulse Propagation in Straight and
Curved Beams - Theory and Experiment," Journal of Applied Mechanics, 39, pp.
1-6,1972.
[26] Danial, A.N., Inverse Solutions in Folded Plate Dynamics, Ph.D. Thesis, Purdue
University, 1994.
[27] Danial, A.N. and Doyle, J.F., "A Massively Parallel Implementation of the Spectral
Element Method for Impact Problems in Plate Structures," Computing Systems in
Engineering, 5, pp. 375-388,1994.
[28] Dania!, A.N. and Doyle, J.F., "Dynamic Response of Folded Plate Structures on a
Massively Parallel Computer," Journal of Computers and Structures, 54, pp. 521-
529,1995.
[29] Danial, A.N. and Doyle, J.F., "Transverse Impact of Damped Plates near a Straight
Edge," Journal of Vibration and Acoustics, 117, pp. 103-108, 1995.
[30] Danial, A.N., Rizzi, S.A. and Doyle, J.F., "Dynamic Analysis of Folded Plate Struc-
tures," Journal of Vibration and Acoustics, 118, pp. 591-598, 1996.
[31] Davies, R.M., "A Critical Study of the Hopkinson Pressure Bar," Philosophical
Transactions of the Royal Society, 240, pp. 375-457.,1948.
[32] Davis, PJ. and Hersh, R., The Mathematical Experience, Birkhiiuser, Boston, 1981.
[33] Desmond, T.P., "Theoretical and Experimental Investigation of Stress Waves at a
Junction of Three Bars," Journal of Applied Mechanics, 48, pp. 148-154, 1981.
[34] Dockerty, G.J. and Leech, C.M., "Stress Propagation through Jointed Beam Sys-
tems using Finite Element Theory," International Journal of Mechanical Sciences,
23,pp.457-471,1981.
[35] Doyle, J.F., Farris, T.N., and Martin, M.T., Crack Identification in Frame Struc-
tures, In Dynamic Fracture Mechanics, M.H. Aliabadi, editor, pp. 237-283, Com-
putational Mechanics Publications, Southampton, 1995.
[36] Doyle, J.F., "An Experimental Method for Determining the Dynamic Contact Law,"
Experimental Mechanics, 24, pp. 10-16, 1984.
[37] Doyle, J.F., "Further Developments in Determining the Dynamic Contact Law,"
Experimental Mechanics, 24, pp. 265-270, 1984.
[38] Doyle, J.F., "Determining the Contact Force during the Transverse Impact of Plates,"
Experimental Mechanics, 27, pp. 68-72, 1987.
[39] Doyle, J.F., "A Spectrally Formulated Finite Element for Longitudinal Wave Prop-
agation," International Journal of Analytical and Experimental Modal Analysis, 3,
pp. 1-5, 1988.
[40] Doyle, J.F., Static and Dynamic Analysis of Structures, Kluwer, The Netherlands,
1991.
References 315
[41] Doyle, J.F., "A Genetic Algorithm for Determining the Location of Structural Im-
pacts," Experimental Meclwnics, 34, pp. 37-44,1994.
[42] Doyle, J.F., "Determining the Location and Size of a Transverse Crack in a Beam,"
Experimental Meclwnics, 35, pp. 272-280, 1995.
[43] Doyle, J.F. and Farris, T.N., "A Spectrally Formulated Finite Element for Wave
Propagation in 3-D Frame Structures," International Journal of Analytical and Ex-
perimental Modal Analysis, pp. 223-237,1990.
[44] Doyle, J.F. and Kamle, S., "An Experimental Study of the Reflection and Transmis-
sion of Flexural Waves at Discontinuities," Journal of Applied Meclwnics, 52, pp.
669-673,1985.
[45] Doyle, J.F. and Kamle, S., "An Experimental Study of the Reflection and Transmis-
sion of Flexural Waves at an Arbitrary T-Joint," Journal of Applied Meclwnics, 54,
pp. 136-140, 1987.
[46] Dyer, I., "Moment Impedance of Plates," The Journal of the Acoustical Society of
America, 32, pp. 1290--1297, 1960.
[47] M. Eisenberger, "Exact Static and Dynamic Stiffness Matrices for General Variable
Cross Section Members," AIAA Journal, 6, pp. 1105-1109, 1990.
(48) Elmore, W.C. and Heald, M.A., Physics of Waves, Dover, New York, 1985.
(49) Eterovie, A.L. and Godoy, L.A., "An Exact Strip Method for Folded Plate Struc-
tures," Computers and Structures, 32, pp. 263-276, 1989.
[50) Ewing, W.M. and Jardetzky, W.S., Elastic Waves in Layered Media, McGraw-Hill,
New York, 1957.
[51) Fahy, F.J., Sound and Structural Vibration: Radiation, Transmission and Response,
Academic Press, New York, 1985.
[52) Finnveden, S., Exact Spectral Finite Element Analysis of Stationay Vibrations in a
Railway Car Structure, Royal Institute of Technology, Stockholm, 1994.
[53) Gopalakrishnan, S., Spectral Analysis of Wave Propagation in Connected Wave-
guides, Ph.D. Thesis, Purdue University, 1992.
[54) Gopalakrishnan, S. and Doyle, J.F., "Wave Propagation in Connected Waveguides
of Varying Cross Section," Journal of Sound and Vibration, 175(3), pp. 347-363,
1994.
[55] Gopalakrishnan, S. and Doyle, J.F., "Spectral Super-Element for Wave Propaga-
tion in Structures with Local Non-Uniformities," Journal of Computer Methods and
Applied Meclwnics in Engineering, 121, pp. 77-90,1995.
[56] Graff, K.F., Wave Motion in Elastic Solids, Ohio State University Press, 1975.
[57) Guyan, R.J., "Reduction of Stiffness and Mass Matrices," AIAA Journal, 3, pp. 380,
1965.
[58] Haines, D.W., "Approximate Theories for Wave Propagation and Vibrations in Elas-
tic Rings and Helical Coils of Small Pitch," International Journal of Solids and
Structures, to, pp. 1405-1416, 1974.
[59) Hamilton, W.R., The Mathematical Papers of Sir W.R. Hamilton, Cambridge Uni-
versity Press, Cambridge, 1940.
[60) Hoff, N.J., The Analysis of Structures, Wiley and Sons, New York, 1956.
(61) Hsieh, D.Y. and Kolsky, H., "An Experimental Study of Pulse Propagation in Elastic
Cylinders," Proceedings of Physical Society, 71, pp. 608-612,1958.
[62] Hsieh, D.Y. and Lee, J.P., "Experimental Study of Pulse Propagation in Curved
Elastic Rods," The Journal of the Acoustical Society of America, 54(4), pp. 1052-
1055,1973.
[63) Inee, E.L., Ordinary Differential Equations, Dover, New York, 1956.
[64) Jones, R.M., Meclwnics of Composite Materials, McGraw-Hill, New York, 1975.
[65] Jordan, R.W. and Whiston, G.S., "Remote Impact Analysis by Use of Propagated
Acceleration Signals IT: Comparison between Theory and Experiment," Journal of
Sound and Vibration, 97, pp. 53--63,1984.
316 References
[66) Junger, M.e. and Feit, D., Sound, Structures, and their Interaction, MIT Press,
Cambridge, 1986.
[67) Kalnins, A., "On Fundamental Solutions and Green's Functions in the Theory of
Elastic Plates," Journal of Applied Mechanics, 33, pp. 31-38,1966.
[68) Kane, T.R., "Reflection of Flexural Waves at the Edge of a Plate," Journal ofApplied
Mechanics, Vol. 21-22, pp. 213-220,1954.
[69) Knauss, W.G., "Uniaxial Wave Propagation in a Viscoelastic Material Using Mea-
sured Materials Properties;' Journal ofApplied Mechanics, 35, pp. 449-453, 1968.
[70) Kolsky, B., Stress Waves in Solids, Dover, New York, 1963.
[71) Konenkov, Y.K., "Diffraction of a Flexural Wave by a Circular Obstacle in a Plate,"
Soviet Physics-Acoustics, Vol. to,No.2, pp. 153-156, 1964.
[72) Lamb, B., "On the Propagation of Tremors over the Surfaces of an Elastic Solid,"
Philosophical Transactions of the Royal Society of London, A 203, pp. 1-42, 1904.
[73) Lanczos, C., The Variational Principles of Mechanics, University of Toronto Press,
Toronto, 1966.
[74) Langhaar, B.L., Energy Methods in Applied Mechanics, Wiley and Sons, New York,
1962.
[75) Langley, R.S., "Analysis of Power Flow in Beams and Frameworks using the Direct
Stiffness Method"," Journal of Sound and Vibration, 136,3, pp. 439-452,1990.
[76) Lee, J.P. and Kolsky, B., "The Generation of Stress Pulses at the Junction of Two
Noncollinear Rods," Journal ofApplied Mechanics, VoLl5, pp. 809-8l3, 1972.
[77) Leissa, A.W., Vibration of Sheils, NASA SP-288, 1973.
[78) Leung, A.Y.T., "Accelerated Convergence of Dynamic Flexibility in Series Form,"
Engineering Structures, 1, pp. 203-206,1979.
[79) Ljunggren, S., "Generation of Waves in an Elastic Plate by a Vertical Force and by
a Moment in the Vertical Plane," Journal of Sound and Vibration, 90, pp. 559-584,
1983.
[80) Main, I.G., Vibrations and Waves in Physics, 3rd ed., Cambridge University Press,
Cambridge, 1993.
[81) Markus, S., Mechanics ofVibrations ofCylindrical Sheils, Elsevier, New York, 1988.
[82) Martin, M.T., Gopalakrishnan, S. and Doyle, J.F., "Wave Propagation in Multiply
Connected Deep Waveguides," Journal ofSound and Vibration, 174(4), pp. 521-538,
1994.
[83) Martin, M.T., Power Flow Analysis of Wave Propagation in Trusses, M.S. Thesis,
Purdue University, 1990.
[84) Martin, M.T., Inverse Problems in Structural Dynamics, Ph.D. Thesis, Purdue Uni-
versity, 1994.
[85) Martin, M.T. and Doyle, J.F., "Impact Force Identification from Wave Propagation
Responses," International Journal of Impact Engineering, 18, pp. 65-77,1996.
[86) Martin, M.T. and Doyle, J.F., "Impact Force Location in Frame Structures," Inter-
national Journal of Impact Engineering, 18, pp. 79-97, 1996.
[87) Maynard, J.D., Williams, E.G. and Lee, Y., "Nearfield Acoustic Holography: I.
Theory of Generalized Holography and the Development of NAH," Journal of the
Acoustical Society of America, 78(4), pp. 1395-1413, 1985.
[88) McConnell, K.G. and Chang, C.N., "A Study of the Axial-Torsional Coupling
Effect on a Sagged Transmission Line," Experimental Mechanics, 26, pp. 324-329,
1986.
[89) Mead, DJ., "A Comparison of Some Equations for the Flexural Vibration of Damped
Sandwich Beams," Journal of Sound and Vibration, 83, pp. 363-377, 1982.
[90) Medick, M.A., "On Classical Plate Theory and Wave Propagation," Journal of Ap-
plied Mechanics, 28, pp. 223-228,1961.
References 317
[91] Mendelsohn, D.A., Achenbach, J.D. and Keer, L.M., "Scattering of Elastic Waves
by a Surface-Breaking Crack," Wave Motion, 2, pp. 277-292,1980.
[92] Menke, W., Geophysical Data Analysis: Discrete Inverse Theory, Academic Press,
Orlando, 1984.
[93] Mindlin, R.D. and Herrmann G., "A One-Dimensional Theory of Compressional
Wave in an Elastic Rod," Proceedings of First U.S. National Congress of Applied
Mechanics, pp. 187-191, 1950.
[94] Morely, S.D., "Elastic Waves in a Naturally Curved Rod," Quarterly Journal of
Mechanical and Applied Mathematics, 14, pp. 155-172, 1961.
[95] Naghdi, P.M. and Berry, J.G., "On the Equations of Motion of Cylindrical Shells,"
Journal of Applied Mechanics, 21(2), pp. 160-166, 1964.
[96] Nakao, T., Okano, T. and Asano, I., "Theoretical and Experimental Analysis of
Flexural Vibrations of the Viscoelastic Timoshenko Beam," Journal ofApplied Me-
chanics, 52, pp. 728-731,1985.
[97] Norton, M.P., Fundamentals of Noise and Vibration Analysis for Engineers, Cam-
bridge University Press, Cambridge, 1989.
[98] Oi, K., "Transient Response of Bonded Strain Gages," Experimental Mechanics, 6,
pp. 463-469,1966.
[99] Pao, Y.H. and Mow, C.C., The Diffraction of Elastic Waves and Dynamic Stress
Concentrations, Crane, Russak and Company Inc., New York, 1973.
[100] Percival, C.M. and Cheney, J.A., "Thermally Generated Stress Waves in a Disper-
sive Elastic Rod," Experimental Mechanics, 9, pp. 49-57,1969.
[101] Philipson, L.L., "On the Role of Extension in the Flexural Vibrations of Rings,"
Journal of Applied Mechanics, 23, pp. 364-366,1956.
[102] Press, W.H., Flannery, B.P., Teukolsky, S.A. and Vetterling, W.T., Numerical
Recipes, 2nd ed., Cambridge University Press, Cambridge, 1992.
[103] Redwood, M., Mechanical Waveguides, Pergamon Press, New York, 1960.
[104] Reissner, E., "Stress and Displacement of Shallow Spherical Shells," Journal of
Mathematical Physics, 25(1), pp. 80-85, 1946.
[105] Reiton, F.E., Applied Bessel Functions, Dover, New York, 1965.
[106] Richards, T.H. and Leung, Y.T., "An Accurate Method in Structural Vibration
Analysis,," Journal of Sound and Vibration, 55, pp. 363-376,1977.
[107] Ripperger, E.A. and Abramson, B.N., "Reflection and Transmission of Elastic
Pulses in a Bar at a Discontinuity in Cross Section;' Third Midwestern Conference
on Solid Mechanics, pp. 29-39, 1957.
[108] Rizzi, S.A., A Spectral Analysis Approach to Wave Propagation in Layered Solids,
Ph.D. Thesis, Purdue University, 1989.
[109] Rizzi, S.A. and Doyle, J.F., "Spectral Analysis of Wave Motion in Plane Solids with
Boundaries," Journal of Vibration and Acoustics, 114, pp. 133-140, 1992.
[110] Rizzi, S.A. and Doyle, J.F., "A Spectral Element Approach to Wave Motion in
Layered Solids," Journal of Vibration and Acoustics, 114, pp. 569-577, 1992.
[111] Rybicki, E.F. and Kanninen, M.F., "A Finite Element Calculation of Stress Intensity
Factors by a Modified Crack-Closure Integral," Engineering Fracture Mechanics, 9,
pp.931-938,1977.
[112] Sanjuan, M.A.F., "Comments on the Hamiltonian Formulation for Linear and Non-
linear Oscillators including Disssipation," Journal of Sound and Vibration, 94, pp.
734-736, 1995.
[113] Schueller, J.K. and Wall, T.M.P., "Impact of Fruit on Flexible Beams to Sense
Modulus of Elasticity," Experimental Mechanics, 31, pp. 118-121, 1991.
[114] Schwieger, H., "Central Deflection of a Transversely Struck Beam," Experimental
Mechanics, 10, pp. 166-169, 1970.
318 References
[115] Seelig, J.M. and Hoppmann, W.H.II, "Impact on an Elastically Connected Double-
Beam System," Journal ofApplied Mechanics, 31, pp. 621-626,1964.
[116] Shames, I.H. and Dym, C.L., Energy and Finite Element Methods in Structural
Analysis, Hemisphere, Washington, 1985.
[117] Shoua, E.D., "The Composite Damping Capacity of Sandwich Cantilever Beams,"
Experimental Mechanics, 8, pp. 300-308,1968.
[118] Sneddon, LN., Fourier Transforms, McGraw-Hill, New York, 1951.
[119] Sotiropoulos, G.H., "Comment on Substructure Synthesis Methods," Journal of
Sound and Vibration, 94, pp. 150-153, 1984.
[120] Speigel, M.R., Fourier Analysis with Applications to Boundary Value Problems,
McGraw-Hili, New York., 1974.
[121] Stevens, K.K., "Force Identification Problems: An Overview," Proceedings of SEM
Spring Meeting, Houston, pp. 838-844,1987.
[122] Suh, N.P, "Helical Coils as Impact Load Dispersers," Journal of Engineering for
Industry, 92(1), pp. 197-207, 1970.
[123] Timoshenko, S.P. and Woinowsky-Krieger, S., Theory of Plates and Shells,
McGraw-Hill, New York, 1968.
(124] Tranter, C.J., Bessel Functions, Hart, New York, 1968.
[125] Viktorov, I.A., Rayleigh and Lamb Waves Physical Theory and Applications, Plenum
Press, New York, 1967.
(126] Weaver, W. and Gere, J.M., Matrix Analysis of Framed Structures, Van Nostrand,
New York, 1980.
(127] Wittrick, W.H., "On Elastic Wave Propagation in Helical Springs," International
Journal of Mechanical Science, 8, pp. 25-47,1966.
[128] Yeats, W.B., Collected Poems ofW.B. Yeats, Macmillan and Co., London, 1960.
[129] Zachary, L.W. and Burger, C.P., "Dynamic Wave Propagation in a Single Lap
Joint," Experimental Mechanics, 20, pp. 162-166, 1980.
Index