Ingenieur-Archiv 60 (1989) 117-- 123 Ingenieur-Archiv

© Springer-Verlag 1989

An experimental modal analysis technique for large-scale structures

J. Wallasehek, Darmstadt

Summary: Using the theory of stationary random vibrations an experimental modal analysis technique,
previously suggested and applied by Luz [2--4], is derived. This method does not require an artificial
excitation of the structure under investigation and is based on measurements of the system's velocity
only. It is well suited for applications to large-scale structures such as high-rise buildings, towers or bridges.

Ein Verfahren zur experimentellen Modalanalyse bei groBen Strukturen

[~bersieht: Mit I-Iilfe der Theorie stationi~rer Zufallsschwingungen wird die yon Luz [2--4] vorgeschlagene
Methode zur experimentellen Modalanalyse erkl/£rt. Dabei wird keine k/instliche Erregung der zu unter-
suchenden Strnktur ben6tigt und es genfigt, die Schwinggeschwindigkeit des Systems zu messen. I)as Verfah-
ren ist speziell ffir grote Banwerke, wie z. B. Hoehhiiuser, Tfirme oder Talbrficken geeignet.

1 Introduction

Today experimental modal analysis has become an accepted method for determining the dynamic
characteristics of linear structural systems. The most important information about linear struc-
tures are essentially the eigenfrequencies and the associated mode-shapes. Since these properties
belong to a mathematical model of the structure rather than to the structure itself, they cannot
be measured directly. Usually the structure is excited at certain points and its dynamical response
to this excitation is observed. Then frequency response functions arc calculated by means of fast
Fourier transform techniques and the parameters of an associated mathematical model (eigen-
frequencies and mode-shapes) are estimated using curve-fitting: Although this method has
been strongly advanced over the years, it features a serious drawback which becomes obvious
when largeseale structures such as high-rise buildings or bridges are to be analysed: I t is the
artificial excitation of the structure by means of shaking or impact which is often not appropriate
for large structures. The huge amount of energy which is necessary to induce structural vibrations
may cause local damage, and the measurement of the exciting forces usually is not a simple task.
Because of this problem especially civil engineers became interested in experimental modal-
analysis techniques which require no artificial excitation. These methods use excitations by
natural microtremors, wind or traffic loads which are available everywhere and at any time. Of
course, no quantitative information about the exciting forces can be used in this ease. Never-
theless the determination of the eigenfrequeneies and mode-shapes of the structure is possible
in most situations, and how this can be done will be described in the following.

2 Formulation of the problem

For illustration we discuss the ease of a one-dimensional continuous system whose deflection is
given by the scalar function w(x, t) of the spatial variable x and time t. The definition of the
system's transfer function G(x, ~; D) is based on the steady-state response to a single concentrated
118 Ingenieur-Archiv 60 (1989)

Fig. 1. Definition of the transfer function

harmonic excitation F~(x -- ~) e j~t (see Fig. 1, 8: Dirac distribution). The transfer function is
defined as the ratio
G(x, ~; ~) = W ( x ) / F (1)
(generally complex-valued) between the amplitude W(x) of the response w(x, t) = W ( x ) e ~gt
and the force amplitude f . The transfer function depends on the location ~ where the force is
applied and on the frequency tg.
We will assume t h a t the systems to be investigated are subject to so-called proportional
damping. Then G(x, ~; D) can be written as

G(x, ~; O) = ~ H~(O)Ck(x) ~k(~), (2)

k--1

a bilinear form of the (real-valued) mode-shapes ~k(x) where

1 1
Hk(D) -- (3)
mk w~ -- ~ + 2jDkD~o k
is called the modal transfer function, mk is the modal mass, D~. the modal damping and wk the
eigenfrequency associated with mode #.
Our aim is to estimate ~k(x) and cok using measurements of the system's response w(x, t) only.
The result can then easily be generalized to the general case of two- and three-dimensional
structures allowing for displacements in any direction.

3 Spectral theory of random vibrations

We are interested in the structure's response to random excitations like wind-pressures or micro-
tremor induced base-motions. These loadings can be considered to be the realisations of stochastic
(field-) processes and in this case also the deflection of the system has random character. I n the
mathematical sense stochastic (field-) processes are described b y probability distributions defined
on probability spaces. However, the most important information for engineering applications can
be obtained from just the first- and second-order moments.
Let {/(x, t; l~)l denote the stochastic field-process of the excitation in Fig. 2, where ,u is
used to distinguish between the various realisations/(x, t) which form the field-process. The first-
and second-order moments of {/(x, t; #)} are called mean
m1(x, t) = E[/(x, t)] (4)
and correlation function
K::(x, y; t, u) ~- E[/(x, t)/(y, u)] (5)
where E [ ] is the expectation operator. These moments can be expressed in terms of the proba-
bility distributions of {/(x, t ; #)} and can equivalently be defined by time- and space-averages if the
process is ergodic.
I~ --L i
f(x,t)

w(x,t) Fig. 2. Beam under stochastic excitation

J. Wallaschek: An experimental modal analysis technique for large-scale structures 119

Let {w(x, t ;/~)} be the beam's deflection. The joint statistics of {/(x, t ; #)} and {w(x, t ; #} can be
described by the cross-correlation function

K1u(x, y; t, u) = E[/(x, t) w(y, u)]. (6)

The stochastic field-processes {/(x, t; #)} and {w(x, t;#)} are related to each other because to
every rcalisation of {/(x, t;/~)} -- specified by fixing # -- there is exactly one realisation of
{w(x, t; #)} with the same #.
If for example the beam of Fig. 2 is described b y the model of Euler and Bernoulli and if
viscous external and internal damping are taken into account, then the relation between the
loading and the deflection is given by the equation of motion

1 @ d., -fit EI --8x~ @ d~ -~ @ oA -fffi w(x, t) =/(x, t) (7)

(El: bending stiffness; A: cross-sectional area; o: mass density; d , d.,: external and internal
damping) and the boundary conditions

w( x, t ) l~=o = O, w ( x , t) x=L = o,

(8)
~2w( x, t) I = 0
~w(X,~x2 t) x=0 = 0,
~X2 Ix= L

which hold for every pair of realisations. I t can be assumed that similar mathematical relations
between excitation and the system's response exist for most mechanical systems even if the
equation of motion or the boundary conditions are not known.
The dynamics of linear systems can be described by means of the impulse response function
g(x, e; t) which describes the motion at location x due to a unit impulse applied at location e at
time t = 0 assuming that the structure was at rest for t < O. I t can be shown that the impulse
response function is the inverse Fourier transform
+co

g(x, e; t)
'f -~
-oo
a(x, e; ~9) e j~t dr9 (9)

of the transfer function. By virtue of the superposition principle the response to the excitation
/(x, t) can then be expressed as a convolution integral
t

w(x, t) = f f g(x, e;t - r)/(e, r) de dr (10)

0 D

where the integration over e extends on the whole domain D and vanishing initial conditions have
been assumed.
This deterministic relation holds for every pair of realisations and after taking the expectation
on both sides of (10) we find
t

t) = f f e; t - r) r) de dr (11)
0 D

for the mean and

Kww(x, y; t, u ) - - f f / f .q(x, e; t -- v)g(y,~;u -- o) K#(e, ~]; v, a, d~ d~] d~ dr (12)

OD OD
for the correlation function of the system's response.
Most stochastic processes which are encountered in technical applications can be considered
to be stationary, which means that their first- and second-order moments do not depend on the
choice of the time-origin. As a consequence the mean depends only on the spatial coordinate,
and the correlation function depends only on the spatial coordinates and the time-lag t -- u
120 Ingenieur-Arehiv 60 (1989)

between the two instants which are compared. Without loss of generality only centered processes,
whose mean vanishes identically, need to be considered, since this case can always be obtained
by a simple coordinate transformation. The only important information about ~ stochastic
field-process is then contained in the correlation function.
The correlation function of a stationary stochastic field-process can be written as the inverse
Fourier transform

Kf1(x, t; y, u) ~- f ~oz(x, y; 17) e j9¢t-~) d/2 (13)

of the non-negative function ~z(x, y; 17) which is called the spectral density and contains in-
formation about the power localization of the stochastic field-process {](x, t;/~)}, see [1].
If the excitation process is stationary, then also the system's response will be stationary and
y,~w(X, y; 17) = f f G(x, $; D) ~os](~,~; D) G*(y, rt; D) d~ d~ (14)
D D

can be derived from (12) for the spectral density of the system's response. Once the spectral
density ~o~(x, y ; D) is known, all other relevant informations about the system's response can be
calculated, e.g. the variance of the deflection is
+ec

a2,~(x) = E[w2(x, t)] = f ~%~(x, x; 17) d/2. (15)

Note tha~ all these functions can be expressed in terms of ~OzI(~, # ; 1?) and the transfer function of
the system.
So far we have only been concerned with the displaeemen~ of the structure. We can however
easily calculate the correlation functions and spectral densities of other system responses, like
e.g. bending-strain or velocity; in the latter ease we find
~o~(z, y; 17) = 1?~w(X, y; 17), (16)
For a physical system the spectral density of the velocity can easily be determined with the help
of a digital signal analysator if the velocity of the structure is measured.

4 Determination of eigenfrequeneies and mode-shapes

If (14) and (2) are inserted into (16), we find

v~(x, y; ~) = E E ~%~(~) H~(~)//~(~) ~(x) ~(y) (17)

k=l 1=1

for the spectral density of the velocity where

Ikt(D) = f f ~0z(x, y; D) ~E(X) ~l(Y) dx dy (18)
D D

can be interpreted as the cross spectral density of generalized forces. Rearranging (17) we obtain

~ ( x , y; £2) = ~ ~2Ikk(~2 ) lHk(1?)l 2 ~%(X) %(y)

k=l

+ E E 17%~(~)H~(~)H~(1?)~(~) ~(y) (19)

k~l /=1
k=~l

where the contribution of the terms has been separated into two parts. Setting x ~ y and inte-
grating (19) over D yields

~o~(D) = f yJ~(x, x; D) dx = ~ 1?eIkk(D) IHk(X2)i2 Jke

D k=l

+ ~ s~%~(17)//k(~)H~(1?)Jk~ (20)
k=l
k~-i
J. Wallasehek: An experimental modal analysis technique for large-scale structures 121

ReHk Im H k
I
1 1

\J
(d k Wk

Fig. 8. Modal transfer Iunetion He(D)

with
Jel = f We(x) ~q(x) dx. (21)
D

For structures with constant mass-distribution Jet = 0 holds for k 4= l because of the orthogona-
lity of the eigenfunctions. Although Jk, will not vanish in general, the contribution of the mixed
terms in (20) is considerably smaller than that of the first sum as long as the eigenfrequencies of
the system are well separated. This is due to the fact that for not too large damping the function
He(D) has a pronounced peak near the eigenfrequency ok whereas it is small for other frequencies
(see Fig. 3).
For broad band excitation Iez(tg) is a slowly varying function of D, so that (20) attains its
relative maxima where the functions jHe(.c2)l~ are maximal. These frequencies are

De = toe V1 -- 2D~ (22)

and lie close to the eigenfrequencies of the structure if the damping is small. Neglecting the
damping in (22), which seems justified for most civil engineering structures, we can locate the
eigenfrequencies we at the frequencies where v-p~(~) attains its relative maxima.
In order to find the mode-shapes ~k(x) we take into account that due to the characteristic
behavior of Hk(f2) the sum (19) is locally, for D ~ ~oe, dominated by a single term
w~,~(x, y; P- ~ we) ~ ~[ee(~e) IHe(~e)l ~ ~e(x) ~e(Y). (23)
Dividing then ~t~(x,y;coe) by T ~ ( y , y; o~e) we estimate the mode-shape ~k(x) as being propor-
tional to
~(x, y; o~e) (24)
~]~(y, y; o~e) "
The method proposed to determine ~oe and ~e(x) can be applied if the velocity of the structure
under investigation is measured. It is restricted to the ease of slightly damped structures with
separated eigenfrequencies and stationary stochastic excitations which have non-vanishing
spectral content in the frequency-range investigated. However, these limiting conditions are met
by most large-scale civil engineering structures.

5 Applications

The method as described in this paper has been successfully used in the past to determine eigen-
frequencies and eigenmodes of civil engineering structures [2--4]. Among the systems which have
been investigated were the freeway bridge over the Koeher-valley in southern Germany (span
1100 m, height 180 m), a HDI~ nuclear power plant and a highrise building at the University of
Stuttgart. In all eases the results showed good agreement with data obtained by other methods.
Typically about 100 measurement points have been selected and the velocity of the structure
has been measured using Willmore-type seismometers of high sensitivity.
122 Ingenieur-Archiv 60 (1989)

6 Some open problems

Besides the principal limitations discussed at the end of Sect. 4 another problem related to the
method described above may arise if the main source of excitation is micro-tremor induced base
motion. Especially for stiff structures this may be the main energy source. Then the boundary
conditions of the system are no longer autonomous. Since the boundary conditions arc essential
for the eigenfrequeneies and eigenmodes, we must clearly specify which boundary conditions
apply in the theoretical investigation.
In order to fix ideas let us consider the beam of Fig. 4 which is excited by the (micro-tremor
induced) motion of its end-points as well as by transverse (wind) loading. Assuming the model of
Euler-Bernoulli to be appropriate to describe the system's dynamics the equation of motion
is (7) and the boundary conditions are

w(x, t) x=o = --yl(t)' w(x, t) x=L = --y.~(t),

(25)
~2w(x,~x~ t) x=o = O, ~w(X,~x2 t) ~=L = O.

A reference model which is appropriate for this problem is the simply supported beam, and
applying experimental modal analysis to this problem we would like to identify the eigenfre-
queneies and eigenmodes of a simply supported beam.
It is however not possible to obtain the solution of (7), (25) directly in terms of the eigenmodes
of the simply supported beam because of the inhomogeneous boundary conditions, [5]. We will
therefore introduce a variable
L--x x
v(x, t) = w(x, t) + ~ y~(t) + --/y2(t) (26)

which describes the elastic deformation of the beam with respect to the straight line passing
through the beam's end-points. Upon insertion of (26) in (7) and (25) we obtain

1 + d2 ~ E I - - + dl + eA v(x,t)

=/(x,t) + L-- x [~A~)l(t) + dl~)i(t)] @ -~

x [~A~)~(t) + diy~(t)] (27)

and homogeneous boundary conditions. The transformation (26) thus reduced the equation of
motion (7) with time-varying, inhomogeneous boundary-conditions (26) to an equation of motion
with constant, homogeneous boundary conditions.
In order to find the eigenfrequeneies and eigenmodes of the simply supported beam by the
method described in this paper measurements of the (relative) velocity ~(x, t) have to be made
because the boundary conditions are only homogeneous in the variable v(x, t). With the help of
seismometers, however, only the absolute velocity,b(x, t)can be registrated. I f the transformation
(26) is not taken into account and eigenfrequeneies and eigenmodes are calculated from (22)
and (24), the eigenfrequencies will still be correct, but the eigenmodes will be erraneous.

L H
f(x,t)

y~ (t) , Y2(t)

f w(x,t) Fig'. 4. Beam with moving supports

J. Wallaschek: An experimental modal analysis technique for large-scale structures 123

I f we assume for simplicity, t h a t the processes yl(t) and y2(t) are stationary, mutually un-
correlated and t h a t they are uncorrelated with/(x, t) we find

2 L - x L - s.(+)
~(x, y; o)e) ~ ~%Iee(coe) lHe(¢oe)l 2 q)e(x) q~e(Y) ~- ~ - - L (2S)

where Sn(D ) and $22(~2) are the power-spectral densities of yl(t) and y2(t), respectively and it is
easily seen that ~ m a y not be used to determine the mode-shapes.
Since it is relatively easy to take the transformation (26) into account, it is suggested that the
motion of the structure's boundary is measured simultaneously. }'or two- and three-dimensional
structures, however, this will increase the necessary measurement effort.

7 Conclusions

I n this paper an experimental modal analysis technique for large-scale structures which was
proposed by Luz [2--4] has been explained using the theory of stationary random vibrations.
I t has been shown that it is possible in m a n y cases to identify the eigenfrequeneies and eigen-
modes of a structure by measurements of the velocity only. The problem of timevarying
boundary conditions, which arises in base-motion excited structures, has also been adres-
sed. The proposed method might become i m p o r t a n t in connection with vibration-monitoring and
crack identification in civil engineering structures.

References

1. Lin, Y. K. : Probabilistic theory of structural dynamics. Malabar: Krieger-Publ. House 1967

2. Luz, E. : Experimental modal analysis of ]argescale structures. Paper presented at the Conf. on vibration
problems in engineering, China, June 1986
3. Luz, E.: Bestimmung yon Bauwerksparametern and -zust~nden mit Hilfe yon Schwingungsmessungen
unter stochastischer Anregung. Materialpr/i~ung 28 (1986) 173--177
4. Luz, E.: Zur experimentellen Modalanalyse yon Bauwerken. Materialprfifung 28 (1986) 301--306
5. Meirovitch, L. : Elements of vibration analysis. Tokyo: McGraw Hill 1975.

Received November 10, 1988

Dr.-Ing. J. Wallaschek
Institut ffir Mechanik
Teehnische Hoehschule Darmstadt
Hochschulstrai~e 1
D- 6100 Darmstadt
Federal Republic of Germany

9*