You are on page 1of 16

Computers and Structures 80 (2002) 117132

www.elsevier.com/locate/compstruc

A frequency response function-based structural


damage identication method
Usik Lee *, Jinho Shin
Department of Mechanical Engineering, Inha University, 253 Yonghyun-Dong, Nam-Ku, Incheon 402-751, South Korea
Received 9 March 2001; accepted 9 October 2001

Abstract
This paper introduces an frequency response function (FRF)-based structural damage identication method (SDIM)
for beam structures. The damages within a beam structure are characterized by introducing a damage distribution
function. It is shown that damages may induce the coupling between vibration modes. The eects of the damageinduced coupling of vibration modes and the higher vibration modes omitted in the analysis on the accuracy of the
predicted vibration characteristics of damaged beams are numerically investigated. In the present SDIM, two feasible
strategies are introduced to setup a well-posed damage identication problem. The rst strategy is to obtain as many
equations as possible from measured FRFs by varying excitation frequency as well as response measurement point. The
second strategy is to reduce the domain of problem, which can be realized by the use of reduced-domain method introduced in this study. The feasibility of the present SDIM is veried through some numerically simulated damage
identication tests. 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Structural damage; Damage identication; Beams; Frequency response function; Damage-induced modal coupling;
Reduced-domain method

1. Introduction
Existence of structural damages within a structure
leads to the changes in dynamic characteristics of the
structure such as the vibration responses, natural frequencies, mode shapes, and the modal dampings.
Therefore, the changes in dynamic characteristics of
a structure can be used in turn to detect, locate and
quantify the structural damages generated within the
structure. In the literature, there have been appeared a
variety of structural damage identication methods
(SDIM), and the extensive reviews on the subject can be
found in Refs. [13].
The nite element model (FEM) update techniques
have been proposed in the literature [49]. As a draw-

Corresponding author. Tel.: +82-32-860-7318; fax: +82-32866-1434.


E-mail address: ulee@inha.ac.kr (U. Lee).

back of FEM-update techniques, the requirement of


reducing FEM degrees of freedom or extending the
measured modal parameters may result in the loss of
physical interpretability and the errors due to the stiness diusion that smears the damage-induced localized
changes in stiness matrix into the entire stiness matrix.
Thus, various experimental-data-based SDIM have been
proposed in the literature as the alternatives to the
FEM-update techniques.
The experimental-data-based SDIM depends on the
type of data used to detect, locate, and/or quantify
structural damages. They include the changes in modal
data [1018], the strain energy [19,20], the transfer
function parameters [21], the exibility matrix [22,23],
the residual forces [24,25], the wave characteristics [26],
the mechanical impedances [27,28], and the frequency
response functions (FRFs) [2931]. Most of existing
modal-data-based SDIM have been derived from FEM
model-based eigenvalue problems.
As discussed by Banks et al. [32], the modal-databased SDIM have some shortcomings. First, the modal

0045-7949/02/$ - see front matter 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 4 5 - 7 9 4 9 ( 0 1 ) 0 0 1 7 0 - 5

118

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

data can be contaminated by measurement errors as well


as modal extraction errors because they are indirectly
measured test data. Second, the completeness of modal
data cannot be met in most practical cases because they
often require a large number of sensors. On the other
hand, using measured FRFs may have certain advantages over using modal data. First, the FRFs are less
contaminated because they are directly measured from
structures. Second, the FRFs can provide much more
information on damage in a desired frequency range
than modal data are extracted from a very limited
number of FRF data around resonance [30]. Thus, the
use of FRFs seems to be very promising for structural
damage identication.
How to minimize the experimental measurement
errors, structure model errors, and the damage identication analysis errors has been an important issue in
most structural damage identication researches. To
develop or to choose a proper reliable SDIM, one needs
to well understand the degree of damage eects on the
dynamics of a structure as well as the aforementioned
errors. Some researchers [13,16,3237] have investigated
the damage-induced changes in natural frequencies,
mode shapes, and curvature mode shapes with varying
the location and severity of a damage. However, very
few attentions have been given to the eects of the
damage-induced coupling of vibration modes (simply,
damage-induced modal coupling) and the higher vibration modes omitted in the analysis on the accuracy of
predicted vibration characteristics of the damaged beam,
from a damage identication viewpoint.
The purposes of the present paper are: to develop an
FRF-based SDIM, in which an ecient reduced-domain
method of damage identication can be used, to investigate the eects of the damage-induced modal coupling
and the omitted higher vibration modes on the accuracy
of predicted vibration characteristics of the damaged
beam, and nally to verify the feasibility of the present
SDIM through some numerically simulated damage
identication tests.



o2
o2 w
qA
w f x; t
EI
ox2
ox2

where wx; t is the exural deection, f x; t the external


force, and EI is the bending stiness for the intact beam.
In Eq. (1), dot () indicates the partial derivative with
respect to time t.
For most practical vibration monitoring problems, it
might be dicult to assign a denitive representation for
the stiness of damaged area because the location, sizes,
and geometry of the damage are not known in prior.
Thus, one of the simplest approaches is to represent the
damage-induced change in stiness at damage location
by the degradation of elastic modulus as follows [16,
32,35]:
Ed x E1  dx

2.1. Dynamic equation of motion for damaged beams


Though the FRF-based SDIM developed in this
paper can be readily extended to the higher order
structures including Timoshenko beams and plate
structures, the BernoulliEuler beam is considered in
this paper as an example structure, for simplicity. The
beam has the length L, the mass density per length qA,
and the intact Youngs modulus E. For small amplitude
vibrations, the dynamic equation of motion for the
beams in an intact state is given by [38]

where Ed is the eective Youngs modulus in the damaged state, and dx is the damage distribution function
which may characterize the state of damage. The case
dx 0 indicates the intact state, while dx 1 indicates the complete rupture of material due to damage. It
seems to be reasonable to assume that the damage-induced changes in mass distribution are negligible because the damage does not result in complete breakage
with a loss of mass [13,17,18,35].
Assume that the damages in a beam are uniform
through the thickness of beam (i.e., thickness-through
damages). Then, the intact Youngs modulus E in Eq.
(1) can be replaced with the eective Youngs modulus
Ed to derive the dynamic equation of motion for the
beams in the damaged state as follows:




o2
o2 w
o2
o2 w

qA
w f x; t
EI
EI
D
ox2
ox2
ox2
ox2

where EID is the eective reduction of bending stiness


due to the presence of damages:
EID x

2. Dynamics of damaged beams

Edxy 2 dA

The second term in the left side of Eq. (3) should


vanish for the intact state. In this study, it is assumed
that there are no damages on the boundaries of beam.
Thus, the boundary conditions applied to a beam in the
intact state can be equally applied to the beam in the
damaged state.
2.2. Dynamic response of the intact beam
The dynamic equation of motion for uniform intact
beams (i.e., EI constant) is reduced from Eq. (1) as

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

EI

o4 w
qA
w f x; t
ox4

Forced vibration response can be obtained by superposing M normal modes as


wx; t

M
X
Wm xqm t

EI



o4 w o2
o2 w
qA
w f x; t

EI
D
ox4 ox2
ox2

where qm t are the modal (or generalized) coordinates


and Wm x are the normal modes satisfying the eigenvalue problem
0000

15

By using the normal modes of the intact beam, the


general solution of Eq. (15) can be assumed as
M
X
wx; t
Wm x
qm t
16

EIWm  qAX2m Wm 0 m 1; 2; . . . ; M

119

Substituting Eq. (16) into Eq. (15) and then applying


Eqs. (8) and (9) yields the modal equations for the
damaged beam as follows:
M
X


qm X2m 
qm 
qn fm t m 1; 2; . . . ; M 17
kmn 
n

and the orthogonality property


Z L
qAWm Wn dx dmn

EIWm00 Wn00 dx X2m dmn

where Xm are the natural frequencies for the intact beam


and dmn is the Kronecker symbol.
Substituting Eq. (6) into Eq. (5) and then applying
Eqs. (8) and (9) yields the modal equations as

qm X2m qm fm t m 1; 2; . . . ; M

The third term in the left side of Eq. (17) reects the
inuence of damage, which is characterized by the
symmetric matrix kmn dened by
Z L
dxWm00 Wn00 dx
DIM
18
kmn EI

10

where fm t are the modal (or generalized) forces dened


by
Z L
f x; tWm dx
11
fm t
0

The matrix kmn , which is called damage inuence matrix


(DIM) herein, depends on the mode curvatures as well
as the damage distribution function. Eq. (18) shows that
the o-diagonal terms of DIM induce the coupling between modal coordinates, which is called herein damage-induced modal coupling (DIMC). To the authors
knowledge, the DIMC has not been discussed in the
existing literatures on SDIM.
The natural frequencies of the damaged beam (Xm )
can be obtained from
j

k
2
det X2m  Xm dmn  kmn 0 no sum
19

Assume that a harmonic point force is applied at


x xF as

For the harmonic point force acting at x xF , the


general solutions of Eq. (17) can be assumed as

f x; t F0 d x  xF eixt


qm t qm t Dqm t

12

where F0 is the amplitude of the harmonic point force


and x is the excitation (circular) frequency. Substituting
Eq. (12) into Eq. (11) gives
fm t Wm xF F0 eixt

qm t

Wm xF
F0 eixt
Qm eixt
X2m  x2

where qm t are the modal coordinates for the intact


beam satisfying Eq. (10), and Dqm t are the damageinduced small perturbed solutions. Substituting Eq. (20)
into Eq. (17) gives

13
D
qm X2m Dqm 

Solving Eq. (10) for qm yields

20

M
M
X
X
kmn Dqn
kmn qn
n

m 1; 2; . . . ; M

21

14

The vibration response of the intact beam can be readily


obtained by substituting Eq. (14) into Eq. (6).
2.3. Dynamic response of the damaged beam
The dynamic equation of motion for damaged uniform beams can be reduced from Eq. (3) as

On applying Eq. (14) into the right side of Eq. (21)


and solving for Dqm t gives
Dqm t

M X
M
X

2

1
Xm  x2 dml  kml kmn Qn eixt
n

22

The third term in the left side of Eq. (21) is so small that
it can be neglected. Then, Eq. (22) can be approximated
in a simplied form as

120

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

Dqm t

M
X
kmn Qn ixt
e
2
2
X
n
m x

23

kmn

N
X

Z
EI

xDj xj

j1

On substituting Eqs. (14) and (23) into Eq. (20) and


substituting the result into Eq. (16) may yield the forced
vibration response of the damaged beam as follows:
"
M
X
Wm xWm xF
wx; t
X2m  x2
m
#
M X
M
X
Wm x Wn xF
F eixt
kmn 2

2 X2  x2

x
X
m
n
m
n

W xeixt

24

where M indicates the number of normal modes superposed in the analysis. The structural damping can be
taken into account in Eq. (24) by simply replacing the
natural frequencies Xm in Eq. (24) with Xm (1 igm )1=2 ,
where gm is the mth modal loss factor.
2.4. Damage inuence matrix

!
Wm00 Wn00 dx

Dj

N
X
j
kmn
Dj

27

j1

where N is the number of damage detection zones


(DDZs), and Dj , xDj , and 2xj represent the magnitude,
location, and size of the piecewise uniform damage over
the jth DDZ, respectively. Here, the DDZs indicate the
nite beam segments that are suspected of damages.
It can be observed from Eq. (27) that the damage-free
zones in which Dj 0 can be removed from the domain
of integration without degrading the accuracy of DIM.
This may drastically reduce the domain of problem or
the number of DDZs for which damage identication
analysis should be conducted. Based on this observation, the reduced-domain method of damage identication is introduced in Section 3.

3. Development of damage identication method


If the DIMC is negligible, Eq. (27) can be approximated as

The DIM depends on how the structural damage is


distributed along the beam. Once the damage distribution function dx is given, the DIM can be readily
computed from Eq. (18). As shown in Fig. 1, consider a
thickness-through damage of magnitude 0 6 D 6 1,
which is uniformly distributed over the small span 2x,
with its midpoint at x xD . The piecewise uniform
thickness-through damage can be represented by
dx DfH x  xD  x  H x  xD xg

xDj xj

25

where H x is the Hevisides unit function. Substituting


Eq. (25) into Eq. (18) yields the DIM as follows:
!
Z

kmn Km dmn

28

where
Km

N
X
j1

xDj xj

EI
xDj xj

Wm00 2 dx

Dj

N
X

kmj Dj

29

j1

Applying Eq. (28) into Eq. (19) may yield a set of


linear algebraic equations for unknown Dj as
n
o
2
kmj fDj g X2m  Xm
m 1; 2; . . . ; M and j 1; 2; . . . ; N

30

xD x

kmn

EI
xD x

Wm00 Wn00 dx

D
kmn D

26

If there exist many damages, say N local damages, Eq.


(26) can be further generalized as follows:

Fig. 1. A uniform beam with a piecewise uniform damage of


magnitude D.

Once the modal data (i.e., natural frequencies and mode


shapes) for a beam in both intact and damaged states
are provided by modal testing or theoretical vibration
analysis, Eq. (30) can be solved for unknown Dj to locate and quantify many local damages at a time, which
implies the structural damage identication. Thus, Eq.
(30) can be used as a means of structural damage identication. The SDIM derived from Eq. (30) is found to
be the same as the modal-data-based SDIM introduced
by Luo and Hanagud [16]. However, as discussed in
Section 1, the modal-data-based SDIM may have some
important limitations. Thus, this study aims to develop
an FRF-based SDIM as an alternative to the modaldata-based SDIM derived Eq. (30).
It might be relatively cheap and easy to use accelerometers to measure vibration responses of a structure.
The vibration signals measured by accelerometers can be
readily processed to obtain FRFs. There are several
dierent denitions of FRF [39]. Though any of them

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

can be used to develop an FRF-based SDIM, the inertance FRF is adopted in this paper.
The inertance FRF generated by the harmonic point
force applied at a point xF can be measured at a point x
as follows:
Ax; x

x; t
w
W x
x2
f xF ; t
F0

31

Substituting Eqs. (12) and (24) into Eq. (31) and


applying Eq. (27) may yield
x2

N X
M X
M
X
Wm x j Wn xF
k
Dj
2
2 mn X2  x2
j
m
n Xm  x
n

Ax; x x2

M
X
Wm xWm xF
m

32

X2m  x2

Because Eq. (32) provides the relationship between unknown damage information (i.e., damage location and
magnitude) and known vibration data, it can be used to
develop an algorithm for structural damage identication. In Eq. (32), the mode shapes (Wm ) and natural
frequencies (Xm ) of the intact beam are considered as
known quantities because they are provided in advance
by the modal testing or theoretical vibration analysis.
The inertance FRF, Ax; x, is also considered as known
quantity because it is measured directly from the damaged beam. However, the damage magnitudes Dj are the
unknown quantities to be determined.
In Eq. (32), the (response, FRF) measurement point
x and the excitation frequency x can be chosen arbitrary. For a specic set of x and x, Eq. (32) may yield
a linear algebraic equation for N unknown Dj . Thus,
choosing N dierent sets of excitation frequency and
measurement point may yield N linear algebraic equations for N unknown Dj in the form of
bXij cfDj g fYi g

i; j 1; 2; . . . ; N

33

where
(
Xij

x2q

Wm xp
X2m  x2q

)T

j
kmn

Wn xF
X2n  x2q

X Wm xp Wm xF


Yi A xp ; xq x2q
X2m  x2q
m
j
kmn
EI

34

35

xDj xj
xDj xj

Wm00 Wn00 dx

i p q  1P
p 1; 2; . . . ; P ; q 1; 2; . . . ; Q; PQ P N

36

37

where xp p 1; 2; . . . ; P denote the measurement


points and xq q 1; 2; . . . ; Q denote the excitation
frequencies. Solving Eq. (33) for N unknown Dj simply

121

implies the location and quantication of damages at a


time. Thus, Eq. (33) provides a new algorithm for FRFbased SDIM. The present FRF-based SDIM requires
the following data only:
1. natural frequencies of intact beam, i.e., Xm ;
2. modes shapes of intact beam, i.e., Wm ;
3. FRF of damaged beam, i.e., Axp ; xq .
The damage identication problem is a sort of inverse problem. Thus, if the number of useful data (or
equations) is not equal to the number of unknown
quantities to be determined, a proper optimization
solution technique is required. One of traditional
approaches is to minimize a suitable norm of the discrepancy between measured and computed quantities,
which is usually a quadratic form associated to the inverse of the covariance matrix. The minimization procedure may smear the damage over intact zones, which
results in the incorrect damage identication. Thus, to
avoid this sort of problem, how to setup a well-posed
damage identication problem has been an important
research issue in the subject of damage identication. To
cope with this issue, two feasible strategies are introduced in the following.
The rst strategy is to obtain a sucient number of
equations from Eq. (32) by choosing as many sets of
excitation frequency and (response) measurement point
as needed. The use of FRFs may help realize this
strategy. Because it is not always easy or practical to
increase the number of measurement points over a certain number, it seems to be much simple and easy rst to
x the measurement points and then to vary the excitation frequency until a sucient number of equations
are derived.
The second strategy is to reduce the (spatial) domain
of problem. From Eqs. (33) and (36), one may nd that
the number of unknown quantities is equal to that of
j
DDZs and the matrix kmn
requires denite integrals only
over the zones with damages. Thus, instead of examining whole domain of problem to search out damages
(i.e., full-domain method), one can reduce the domain of
problem in advance by removing the zones that are
found out to be damage-free to examine only the reduced
domain of problem (i.e., reduced-domain method). The
reduced-domain method will not degrade the accuracy
of damage identication results at all. To realize the
reduced-domain method, however one should know the
locations and sizes of damage-free zones in advance.
Unfortunately, this is impracticable for most cases.
Thus, one needs a method to search out damage-free
zones in the process of damage identication analysis.
In this paper, a three-steps method is introduced and
its feasibility is numerically veried in Section 4.
The rst step: Divide the domain of problem into N
DDZs and use Eq. (33) to predict N unknown damages

122

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

Dj for N DDZs. The rst prediction results are represented by Dj (rst step) j 1; 2; . . . ; N.
The second step: Divide each DDZ at the rst step
into M sub-DDZs to have total (M  N ) sub-DDZs and
use Eq. (33) to re-predict (M  N) unknown damages for
(M  N) sub-DDZs. The second prediction results are
represented by Dij (second step) (i 1; 2; . . . ; M and
j 1; 2; . . . ; N).
The third step: If Dij second step < Dj first step,
conclude that the ith sub-DDZ within the jth DDZ is
damage-free. Otherwise, the sub-DDZ is suspected of
damage.
Once damage-free zones are searched out and removed from the domain of problem by using the present
three-steps method, it is possible to put D 0 for all
removed damage-free zones and to conduct damage
identication only for the reduced domain, which is the
reduced-domain method of damage identication introduced in the present study. By iteratively using the
reduced-domain method, all damage-free zones can be
removed from the original domain of problem to leave
damaged zones only, which simply implies the location
of damages. The damage magnitudes are quantied
from Eq. (33) every iteration.
In summary, an FRF-based SDIM is introduced
based on the damage identication algorithm of Eq.
(33). In the present SDIM, the reduced-domain method
can be iteratively used to reduce the domain of problem.
The present SDIM can locate and quantify many local
damages at a time by using the FRFs experimentally
measured from the damaged beam. The appealing features of the present SDIM may include the followings:
(1) the modal data of damaged beam are not required in
the analysis; (2) as many equations as required to setup a
well-posed damage identication problem can be generated from the measured FRFs by varying the excitation frequency as well as the measurement point; (3) the
reduced-domain method based on the three-steps process of domain reduction can be iteratively used to eciently reduce the domain of problem and nally to
identify many local damages just within a few iterations.

4. Vibration characteristics of damaged beams


Many researchers [13,16,3237] have investigated the
damage-induced changes in natural frequencies, mode
shapes, and curvature mode shapes varying the location
and severity of damage. However, there have been very
few investigations, from a damage identication viewpoint, on the eects of the DIMC as well as the higher
vibration modes omitted in the analysis (simply the
omitted higher modes) on the accuracy of predicted vibration characteristics of the damaged beam. Thus, in
this section, some numerical investigations are given to
the DIMC and the omitted higher modes. As a repre-

sentative problem, a uniform beam of length L 1:2 m


is considered herein. The beam has the intact bending
stiness EI 11:2 N m2 and the mass density per length
qA 0:324 kg/m.
4.1. Eects of damage-induced model coupling
The DIM for the cantilevered beam with a piecewise
uniform damage at the midpoint of beam, i.e., xD 0:6
m, is shown in Table 1. Similarly, the DIM for the
cantilevered beam with three identical piecewise uniform
damages at xD 0:3, 0.6, and 0.9 m is given in Table 2.
The piecewise uniform damages considered for Tables 1

Table 1
Damage inuence matrix kmn =kref for the cantilevered beam
with one piecewise uniform damage: D 0:5; xD 0:6 m;
2x 0:133 m; kref 3:87

Table 2
Damage inuence matrix kmn =kref for the cantilevered beam
with three piecewise uniform damages: D1 D2 D3 0:5;
xD1 0:3 m, xD2 0:6 m, xD3 0:9 m; 2x1 2x2 2x3 0:133
m; kref 3:87

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

123

and 2 have the same magnitude D 0:5 and the same


size 2x 0:133 m.
Tables 1 and 2 show that, as a general rule, the diagonal terms of DIM (i.e., the direct eects of damage)
increase in magnitude as the mode number increases.
However, they decrease momentary at certain vibration
modes if a node of the modes is located in damaged
zones. For instance, k33 and k55 in Table 1 are smaller
than k22 and k44 , respectively, because a node of the third
and fth modes is located in the damaged zone. Eq. (27)
shows that, in general, DIM becomes larger as the
damage magnitudes increase. The o-diagonal terms of
DIM (i.e., the indirect eects of damage or the DIMC)
are relatively smaller than the diagonal terms. The odiagonal terms vanish completely when the damage is
uniformly distributed over the whole beam, regardless of
its magnitude, which can be readily proved from Eq.
(27) by using the orthogonality property for normal
modes.
Fig. 2 shows the eects of DIMC on the damageinduced changes in natural frequencies of the cantile-

Fig. 3. Eects of DIMC on the changes in natural frequencies


of the simply supported beam: (a) when D 0:5, (b) when
D 0:05.

Fig. 2. Eects of DIMC on the changes in natural frequencies


of the cantilevered beam: (a) when D 0:5, (b) when D 0:05.

vered beam depending on the magnitude of a piecewise


uniform damage. Fig. 3 is for the simply supported
beam. Neglecting the DIMC tends to underestimate
the damage-induced changes in natural frequencies. In
general, the eects of DIMC on the changes in natural
frequencies are found to be negligible, especially when
the damage is very weak. However, it will be desirable to
include the DIMC in the damage identication analysis
because damages are not known in prior for most
practical cases.
From Figs. 2 and 3, one may observe the followings.
First, in general, the percent changes in natural frequencies at the lower modes are larger than those at the
higher modes, and vice versa for the absolute changes in
natural frequencies. Second, the percent changes in
natural frequencies highly depend on mode number and
damage location. If damages are located at or very near
the nodes of a mode, the percent change in the natural
frequency of the corresponding mode is very small. For
instance, the percent changes in natural frequencies are
very small for the odd (e.g., third and fth) modes of

124

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

Fig. 4. Eects of DIMC on the change in inertance FRF of the


cantilevered beam with a damage of D 0:5.

cantilevered beam and for the even (e.g., second and


fourth) modes of simply supported beam. Very similar
results have been experimentally observed by Capecchi
and Vestroni [40]. Third, the percent changes in natural
frequencies converge to a certain steady state value as
the mode number increases. For instance, about 1%
when D 0:5 and about 0.1% when D 0:05 for the
cantilevered beam. Similarly, about 0.5% when D 0:5
and about 0.05% when D 0:05 for the simply supported beam.
Fig. 4 compares the inertance FRFs of damaged
beam, calculated with and without including the DIMC,
with that of intact beam. In general, the eects of DIMC
on the changes in inertance FRFs are found to be negligible. One notes that the third and fth resonance
peaks are not appeared in Fig. 4 because the FRF
measurement point (x 0:6) coincides with a node of
the third and fth modes.
4.2. Eects of the omitted higher modes
A suciently large number of normal modes and
natural frequencies of the intact beam are required for
accurate damage identication. However, in practice,
only a limited number of the lower normal modes and
natural frequencies can be provided by modal testing or
theoretical modal analysis. Thus, the errors due to the
omission of the higher normal modes are inevitable.
Fig. 5 shows the ratios between the omitted higher
modes-induced errors in natural frequencies and the
damage-induced changes in natural frequencies for the
cantilevered beam with a piecewise uniform damage.
Similarly, Fig. 6 shows the results for the simply supported beam. The omitted higher modes-induced error
in natural frequency, denoted by DX (omitted modes) in
Figs. 5 and 6, is dened by the dierence between the

Fig. 5. Comparison of the errors in natural frequencies due to


the omitted higher normal modes with the damage-induced
changes in natural frequencies for the cantilevered beam: (a)
when D 0:5, (b) when D 0:05.

exact and approximate natural frequencies of the damaged beam. The approximate natural frequencies are
calculated by using a nite number of normal modes. On
the other hand, the damage-induced change in natural
frequency, denoted by DX (damage) in Figs. 5 and 6, is
dened by the dierence between the exact natural frequency of the intact beam and that of the damaged
beam. The important thing here is that the omitted
higher modes-induced errors should be much smaller
than the damage-induced changes for very reliable
damage identication. From Figs. 5 and 6, one may
observe the followings.
First, if damages are located at or very near the nodes
of a normal mode, the omitted higher modes-induced
errors become very signicant for the natural frequency
corresponding to the normal mode. For example, the
damage considered herein is located at a node of the
third and fth modes of the cantilevered beam. Thus,
when total ve normal modes are used to calculate
natural frequencies, for instance, the omitted higher

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

Fig. 6. Comparison of the errors in natural frequencies due to


the omitted higher normal modes with the damage-induced
changes in natural frequencies for the simply supported beam:
(a) when D 0:5, (b) when D 0:05.

modes-induced errors in the third and fth natural frequencies are larger than 16% of the damage-induced
changes while those in the rst, second and fourth natural frequencies are about 10%. A very similar observation can be made for the simply supported beam. This
means that, from a damage identication viewpoint, it is
desirable to use only the natural frequencies of the intact
beam of which modes do not have nodes at or very near
the damage locations. However, this is almost impracticable because the damage locations are not known in
advance. This limitation is certainly one of the shortcomings for the modal-data-based SDIMs in which only
modal parameters are used for damage identication.
Second, the omitted higher modes-induced errors become increasingly signicant for weak damages. For
example, if total ve normal modes are used to calculate
the rst natural frequency of the cantilevered beam, the
omitted higher modes-induced error is about 10% of the
pure damage-induced changes when D 0:5, whereas
about 110% when D 0:05. A similar observation can

125

Fig. 7. Comparison of the errors in inertance FRFs due to the


omitted higher normal modes with the damage-induced changes
in inertance FRFs, for two dierent damage magnitudes: (a)
cantilevered beam, (b) simply supported beam.

be made for the simply supported beam. Because the


omitted higher modes-induced errors become very signicant especially for weak damages, a suciently large
number of normal modes should be considered to
compute suciently accurate modal data for the damaged beam.
Fig. 7 shows the ratios between the omitted higher
modes-induced errors in inertance FRFs and the damage-induced changes in inertance FRFs, depending on
damage magnitude. A harmonic point force of x 30
rad/s is applied at the midpoint of beam and the inertance FRFs are measured also at the midpoint of beam.
When D 0:5, about four normal modes are required
for the cantilevered beam to lower the omitted higher
modes-induced errors 10% below the pure damageinduced changes, and about three normal modes for
the simply supported beam. As the damage becomes
weaker, a much larger number of normal modes are
required to compute the FRFs to a required accuracy.
Fig. 7 show that, when D 0:05, about eight and ve

126

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

Fig. 8. Eects of damage magnitude on the changes in natural


frequencies of the cantilevered beam with a piecewise uniform
damage at the midpoint of beam.

Fig. 10. Eects of damage location on the changes in natural


frequencies of the simply supported beam with a piecewise
uniform damage of D 0:5.

normal modes are required for the cantilevered beam


and simply supported beam, respectively.

tends to converge to a certain steady state value as the


mode number increases. It can be also observed from
Figs. 9 and 10 that the eects of damages on the change
in a natural frequency become very weak when the
damages are located at the nodes of the corresponding
vibration mode. Because the fth mode of cantilevered
beam has nodes at x 0:33, 0.66, and 0.87 m, for instance, Fig. 9 shows that the percent change in the fth
natural frequency is found to be very small for all
damage locations considered in the gure. Similarly, as
the fourth mode of the simply supported beam has
nodes at xD 0:33 and 0.60 m, the percent change in the
fourth natural frequency is also found to be very small
when damage is located at xD 0:33 and 0.60 m.

4.3. Eects of damage magnitude and location


Fig. 8 shows the percent changes in natural frequencies depending on damage magnitude. In general,
as can be quickly observed from Eq. (30), the percent
changes in natural frequencies increase in proportion to
the magnitudes of damages.
Figs. 9 and 10 show the percent changes in natural
frequencies depending on damage location for the cantilevered and simply supported beams, respectively. The
sensitivity of the natural frequency to the change of
damage location at the lower modes is found to be
higher than that at the higher modes. And, as discussed
in Section 4.1, the percent change in natural frequency

Fig. 9. Eects of damage location on the changes in natural


frequencies of the cantilevered beam with a piecewise uniform
damage of D 0:5.

5. Numerical damage identication tests and discussions


In this section, the feasibility of the present FRFbased SDIM is tested through some numerically simulated damage identications. The feasibility tests are
conducted by rst placing some pre-dened piecewise
uniform damages in a uniform beam and then inversely
identifying them by use of the present SDIM.
Eqs. (33)(37) show that the present SDIM requires
the normal modes and natural frequencies for the intact
beam and the experimentally measured FRFs for the
damaged beam. For the intact beam, the normal modes
and natural frequencies are analytically computed from
Eqs. (7)(9) in which the material and structural properties are for a rened intact structure model. By rened, we mean that the measured and analytical modal
parameters are in good agreement.
As the FRFs are measured experimentally, it is liable
to be contaminated by various measurement noises.
Thus, following the approach used by Thyagarajan et al.

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

[31], an e% random noise is added to the FRFs analytically simulated from Eq. (31) to represent the measurement noises in measured FRFs:





e
A xp ; x q A xp ; x q 1
randn
38
100
where A is the FRFs simulated to include the measurement noises, and randn is the random noise generator
function in M A T L A B . In this study, it is assumed that
the random noise is uniformly distributed, with the
mean 0 and variance 1.
To measure the accuracy of the predicted damage
state with respect to the true one, a root mean squared
damage identication error (DIE) is dened as
s
Z
1 L Pred
d x  d True x2 dx
DIE
L 0
v
u N
u1 X

t
2xj DPred
 DTrue
39
j
j
L j
where L is the length of beam and the superscripts True
and Pred indicate the true and predicted damage states,
respectively. The subscript j indicates the quantities for
the jth DDZ. As the value of DIE is getting smaller,
the predicted damage state is getting closer to the true
one.
As an illustrative example, a simply supported uniform beam is considered. The beam has the length
L 1:2 m, the intact bending stiness EI 11:2 N m2 ,
and the mass density per length qA 0:324 kg/m. As
shown in Fig. 11, two damage problems are considered:
the one piecewise uniform damage problem (i.e.,
D 0:6, xD 1:0 m, 2x 0:044 m) and the three
piecewise uniform damages problem (i.e., D1 0:4,

127

xD1 0:2 m, 2x1 0:015 m; D2 0:5, xD2 0:6 m,


2x2 0:015 m; D3 0:6, xD3 1:0 m, 2x3 0:015 m).

5.1. Comparison between full-domain method and reduced-domain method


The full-domain and reduced-domain methods of
damage identication are compared. In the full-domain
method, damage identication is iteratively conducted
always over the whole span of beam with increasing the
number of DDZs, for instance, by three times every iteration, until the predicted damage converges to the true
one. In the reduced-domain method, the domain of
problem is reduced by searching out and removing
damage-free zones by use of the three-steps process
introduced in Section 3. The same procedure is repeated
in the next iteration, but now only for the reduced
domain.
Fig. 12 shows the detailed damage identication
processes for the full-domain and reduced-domain
methods. No random noise in FRFs is considered for
Fig. 12. The reduced-domain method is found to provide more accurate damage identication when compared with the full-domain method. Fig. 12(b) clearly
shows that the damage magnitudes in the truly damaged
zones are always predicted larger at the second step,
whereas the damage magnitudes in the truly damagefree zones are always predicted smaller at the second
step. Comparing the damage magnitudes predicted at
the rst and second steps of iteration, the zones in which
the predicted damage magnitudes become smaller at the
second step are considered to be damage-free and they
are all indicated by the minus () sign at the last step
(i.e., third step) of iteration, as shown in Fig. 12(b). On
the other hand, the other zones are kept suspecting of
damage and indicated by the plus () sign. In Section
5.3, the three-steps process will be further tested with
taking into account the noise in FRFs.

5.2. How to choose excitation frequencies

Fig. 11. Two example problems for numerically simulated


damage identication tests.

It might be important to well understand which excitation frequencies should be chosen in order to obtain
reliable damage identication from the measured FRFs
which may depend on excitation frequency. Fig. 13
shows the DIEs obtained by varying the excitation frequency x of the harmonic point force applied at the
midpoint of beam. It is found that the DIEs are strongly
dependent of the excitation frequency. It is interesting
to observe from Fig. 13 that the most reliable damage
identication (i.e., small value of DIE) can be obtained
when the excitation frequencies are chosen very near the
natural frequencies. In theory, if structural damping
is taken into account in the analysis, the excitation

128

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

5.3. Feasibility of the three-steps process for reduceddomain method


In Section 5.1, the feasibility of the three-steps process used in conjunction with the reduced-domain
method was tested for the beam with a single damage,
shown in Fig. 11(a), without taking into account the
noise in FRFs. In this section, the feasibility tests are
rerun with taking into account the random noise in
FRFs. Fig. 14 shows the damage predictions at the rst
two steps of the rst iteration of the reduced-domain
method of damage identication, depending on the level
of random noise in FRFs. Fig. 14 shows that in general
the damage magnitudes in the truly damaged zones are
predicted larger at the second step, whereas those in the
truly damage-free zones are predicted smaller. However,
this rule is broken if the random noise in FRFs becomes
larger than a certain value: for instance, about 10% for
the problem considered herein. Thus, for the successful
applications of the present three-steps process to damage identication, the true noise in FRFs should be
minimized as low as possible, probably less than about
5%, which can be met by most well-prepared vibration
tests. Otherwise, the full-domain method would be
rather recommended.

5.4. Multi-excitation frequency/multi-measurement point


approach

Fig. 12. Two methods of damage identication: (a) full-domain


method, (b) reduced-domain method.

frequencies can be chosen to be equal to the natural


frequencies. However, exciting the beam at even natural
frequencies (e.g., second and fourth natural frequencies)
does not provide good damage identication because the
pre-specied location of damage exactly coincides with a
node of the even normal modes. In general, it is recommended to choose the excitation frequencies very
near to the natural frequencies of which normal modes
do not have any node within the most candidate-damaged zones.

Eq. (23) shows that the number of DDZs is equal to


the number of unknown damage magnitudes, Dj . Once
the number of DDZs is xed, one can choose a proper
number of excitation frequencies and measurement
points to derive as many linear algebraic equations
as the number of unknown Dj . Thus, depending on the
number of excitation frequencies and measurement
points, various approaches can be considered: the singleexcitation frequency multi-measurement point (SFMP)
approach, the multi-excitation frequency single-measurement point (MFSP) approach, and the multi-excitation frequency multi-measurement point (MFMP)
approach. For a damaged beam divided into total
27 DDZs, for instance, one may choose either one excitation frequency twenty-seven measurement points
(1F27P), or three excitation frequencies nine measurement point (3F9P), or nine excitation frequencies three
measurement point (9F3P), or 27 excitation frequencies
one measurement points (27F1P). The SFMP approach
seems to be impractical for most problems because it
is not so easy to increase the number of measurement
points over a certain limit number. Because the FRFs
are dependent of measurement points, the MFSP approach can provide poor damage identication if the
measurement points are not properly chosen. Thus, to
cope with these problems met when SFMP or MFSP

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

129

Fig. 13. Excitation frequency dependence of DIE when a 5% random noise in FRFs is considered.

Fig. 14. Illustration of the three-steps process used in the reduced-domain method of damage identication: the simply supported
beam with one piecewise uniform damage at the middle of beam.

approach is applied, it is recommended in general to use


the MFMP approach for most reliable damage identication.

Table 3 compares the damage identications by


SFMP and MFMP approaches, in terms of DIE with
varying the random noise in FRFs up to 9%. The

130

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

Table 3
Comparison of the DIEs depending on the application method of the present SDIM and the random noise in FRFs
Random noise in
FRFs (%)

SFMP approach

MFMP approach

Full-domain method

Reduced-domain
method

Full-domain method

Reduced-domain
method

0
1
3
5
7
9

9:56  1012
7:99  103
2:33  102
3:61  102
4:24  102
5:01  102

8:21  1012
3:25  103
1:47  102
2:16  102
2:61  102
3:62  102

5:32  1012
3:16  103
1:12  102
2:04  102
2:52  102
3:33  102

3:76  1012
2:57  103
9:56  103
1:16  102
2:18  102
2:84  102

Fig. 15. Damage identication results by the full-domain method, for dierent levels of random noise in FRFs.

MFMP approach is shown to provide more accurate


results for both the full-domain and reduced-domain
methods. This is probably because the abundant damage
information contained in the measured FRFs can be
reected in the damage identication analysis more efciently by the use of MFMP approach. Table 3 shows
that in general the reduced-domain method used in
combination with MFMP approach provides the most
reliable damage identication.

5.5. Application to the three damages problem


Both the full-domain and reduced-domain methods
of damage identication (all in combination with
MFMP approach) are applied to the beam with three
damages (Fig. 11b) and their results are given in Figs. 15
and 16, respectively, for dierent levels of random noise
in FRFs. It is shown that both methods fairly well locate

and quantify the pre-dened three damages up to about


9% random noise in FRFs, but putting a small level of
incorrect damages at damage-free zones.
When compared with the full-domain method, the
reduced-domain method seems to provide more reliable results as far as the noise in FRFs is so small
that the three-steps process works correctly. Fig. 17
shows the details of a reduced-domain method based
damage identication, in which a 5% random noise in
FRFs is considered so that the three-steps process works
correctly. However, it is numerically investigated that
the three-steps process fails to discriminate damage-free
zones from damaged zones if the random noise in FRFs
becomes larger than about 9% for the present example
problem. Thus, when a 10% random noise in FRFs is
considered, the three-steps process fails to work and Fig.
16(d) certainly shows that the reduced-domain method
provides a misleading damage identication by putting
non-zero damages at damage-free zones.

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

131

Fig. 16. Damage identication results by the reduced-domain method, for dierent levels of random noise in FRFs.

6. Conclusions

Fig. 17. Details of the reduced-domain method of damage identication for the simply supported beam with three piecewise uniform damages, when a 5% random noise in FRFs is considered.

In this paper, an FRF-based SDIM is derived from


dynamic equation of motion for damaged beams. The
appealing features of the present SDIM include the
followings. First, the modal data of damaged structure
are not required in the analysis. Second, a large number
of equations can be readily derived by varying the excitation frequency as well as the response measurement
point. Lastly, the domain or size of problem can be
drastically reduced by iteratively using the reduceddomain method introduced in this paper.
Numerical investigations on the dynamics of damaged beam may conclude that: (1) neglecting the DIMC
may underestimate the damage-induced changes in
natural frequencies, (2) the eects of DIMC on the
changes in natural frequencies and FRFs are in general
negligible, and (3) the damage-induced changes in natural frequencies are relatively large at the lower modes,
but highly dependent of modes.
The feasibility of the present SDIM is veried
through some numerically simulated damage identication tests. It is shown that the three-steps process introduced for the reduced-domain method of damage
identication is valid as far as the noise in FRFs is
smaller than a certain limit value: for instance, 9% for
the example problems considered in the present study. In
general, the reduced-domain method is found to provide
the most reliable damage identication when it is used in
combination with the MFMP approach.

132

U. Lee, J. Shin / Computers and Structures 80 (2002) 117132

Acknowledgements
This work was supported by Korean Research
Foundation Grant (KRF-2001-041-E00034).
References
[1] Doebling SW, Farrar CR, Prime MB. A summary review
of vibration-based damage identication method. Shock
Vibr Dig 1998;30(2):91105.
[2] Salawu OS. Detection of structural damage through changes
in frequency: a review. Engng Struct 1997;19(9): 71823.
[3] Zou Y, Tong L, Steven GP. Vibration-based modeldependent damage (delamination) identication and health
monitoring for composites structuresa review. J Sound
Vibr 2000;230(2):35778.
[4] Collins JD, Hart GC, Hasselman TK, Kennedy B.
Statistical identication of structures. AIAA J 1974;12(2):
18590.
[5] Kabe AM. Stiness matrix adjustment using mode data.
AIAA J 1985;28(9):14316.
[6] Smith SW, Hendricks SL. Damage detection and location
in large space trusses. AIAA SDM Issues of the International Space Station A Collection of Technical Papers,
AIAA, Washington, DC, 1988. p. 5663.
[7] Lin RM. In: Analytical model improvement using modied
IEM Proc International Conference on Structural Dynamic Modeling. Glasgow, Scotland: National Agency for
Finite Element Methods and Standards; 1993. p. 18194.
[8] Zimmermann DC, Kaouk M. Structural damage detection
using a minimum rank update theory. J Vibr Acoustics
1994;116:22231.
[9] Lim TW. Structural damage detection using constrained
eigenstructure assignment. J Guidance, Cont Dyn
1995;18(3):4118.
[10] Adams RD, Cawley P, Pye CJ, Stone BJ. A vibration
technique for non-destructively assessing the integrity of
structures. J Mech Engng Sci 1978;20(2):93100.
[11] Hearn G, Testa RB. Modal analysis for damage detection
in structures. J Engng Mech 1991;117(10):304263.
[12] Grin SF, Sun CT. Health monitoring of dumb and smart
structures. 28th Annual Technical Meeting of SES, 1991.
[13] Pandey AK, Biswas M, Samman MM. Damage detection
from changes in curvature mode shapes. J Sound Vibr
1991;145(2):32132.
[14] Cempel C, Matke HG, Ziolkowski A. Application of
transformed normal modes for damage location in structures. In: Stanley P, editor. Structural integrity assessment.
Oxford: Elsevier; 1992. p. 24655.
[15] Davini C, Gatti F, Morassi A. A damage analysis of steel
beams. Meccanica 1993;22:2737.
[16] Luo H, Hanagud S. An integral equation for changes in the
structural dynamics characteristics of damaged structures.
Int J Solids Struct 1997;34(35/36):455779.
[17] Bicanic N, Chen HP. Damage identication in framed
structures using natural frequencies. Int J Numer Meth
Engng 1997;40:445168.
[18] Hassiotis S. Identication of damage using natural frequencies and Markov parameters. Comput Struct 2000;74:
36573.

[19] Stubbs N, Kim JT, Topole K. An ecient and robust


algorithm for damage location in oshore platforms.
Proceedings of the ASCE 10th Structures Congress, 1992.
p. 5436.
[20] Cornwell P, Doebling SW, Farrar CR. Application of the
strain energy damage detection method to plate-like
structures. J Sound Vibr 1999;224(2):35974.
[21] Lew JS. Using transfer function parameter changes for
damage detection of structures. AIAA J 1995;33(11):2189
93.
[22] Lin CS. Location of modeling errors using modal test data.
AIAA J 1990;28(9):16504.
[23] Pandey AK, Biswas M. Damage diagnosis of truss
structures by estimation of exibility change. Int J Anal
Exp Modal Anal 1995;10(2):10417.
[24] Ricles JM, Kosmatka JB. Damage detection in elastic
structures using vibratory residual forces and weighted
sensitivity. AIAA J 1992;30(9):23106.
[25] Sheinman I. Damage detection and updating of stiness
and mass matrices using mode data. Comput Struct
1996;59(1):14956.
[26] Feroz KT, Oyadiji SO. Damage detection in rods using
wave propagation and regression analysis. Proc 1999
ASME Design Engineering Technical Conferences, Las
Vegas, NV, DETC99/VIB-8372, 1999.
[27] Cawley P. The impedance method of non-destructive
inspection. J NDT Int 1984;17:5965.
[28] Wong BS, Guan TC, King LM. Mechanical impedance
inspection of composite structures. Br J NDT 1993;35:39.
[29] Choudhury AR, He J. Structural damage location using
expanded measured frequency response function data.
Proc IMAC XIV, 1996. p. 93442.
[30] Wang Z, Lin RM, Lim MK. Structural damage detection
using measured FRF data. Comput Meth Appl Mech
Engng 1997;147:18797.
[31] Thyagarajan SK, Schulz MJ, Pai PF. Detecting structural
damage using frequency response functions. J Sound Vibr
1998;210(1):16270.
[32] Banks HT, Inman DJ, Leo DJ, Wang Y. An experimentally validated damage detection theory in smart structures. J Sound Vibr 1996;191(5):85980.
[33] Weissenburger JT. Eect of local modications on the
vibration characteristics of linear systems. J Appl Mech
1968;35:32735.
[34] Sato H. Free vibration of beams with abrupt changes of
cross-section. J Sound Vibr 1983;89(1):5964.
[35] Yuen MMF. A numerical study of the eigenparameters of
a damaged cantilever. J Sound Vibr 1985;103:30110.
[36] Davini C, Morassi A, Rovere N. Modal analysis of
notched bars: tests and comments on the sensitivity of
an identication technique. J Sound Vibr 1995;179(3):
51327.
[37] Lee U, Chang J, Kim N. Structural micro-damage identication. AIAA Paper 2000-1503, 2000.
[38] Meirovitch L. Analytical methods in vibration. New York:
Macmillan; 1967.
[39] Ewin DJ. Modal testing theory and practice. Hertfordshire: Research Studies Press; 1984.
[40] Capecchi D, Vestroni F. Monitoring of structural systems
by using frequency data. Earthquake Engng Struct Dyn
1999;28:44761.

You might also like