ARTICLE IN PRESS

JOURNAL OF
SOUND AND
VIBRATION
Journal of Sound and Vibration 270 (2004) 1–14
www.elsevier.com/locate/jsvi

Damage detection in structures using a few frequency response

measurements
H.Y. Hwanga,*, C. Kimb
a
Department of Aerospace Engineering, Sejong University, 98 Kwangjin-Gu, Kunja-dong, Seoul 143-747, South Korea
b
School of Mechanical Engineering, Kyungpook National University, Daegu 702-701, South Korea
Received 22 January 2002; accepted 4 February 2003

Abstract

This paper presents methods to identify the locations and severity of damage in structures using
frequency response function (FRF) data. Basic methods detect the location and severity of structural
damage by minimizing the difference between test and analytic FRFs, which is a type of model updating or
optimization method; however, the preferred method proposed in this paper uses only a subset of vectors
from the full set of FRFs for a few frequencies and calculates the stiffness matrix and reductions in explicit
form. To verify the proposed method, examples for a simple cantilever and a helicopter rotor blade are
numerically demonstrated. The proposed method identified the location of damage in these objects, and
characterized the damage to a satisfactory level of precision.
r 2003 Elsevier Ltd. All rights reserved.

1. Introduction

The problem of monitoring the structural health of an object consists of obtaining information
about the existence, location, and extent of damage in the structure using non-destructive
methods. One approach is to monitor and interpret changes in structural dynamic measurements
based on experimental modal analyses and signal-processing techniques. The extraction of the
natural frequency and mode shape of a vibrating structure can be accomplished using modern
vibration testing equipment and instrumentation. The modal and structural dynamic data can
then be utilized for cost-effective health monitoring and operational life assessment without
requiring the structure to be dismantled [1].

*Corresponding author. Tel.: +82-2-3408-3773; fax: +82-2-3408-3333.

E-mail address: [email protected] (H.Y. Hwang).

0022-460X/03/$ - see front matter r 2003 Elsevier Ltd. All rights reserved.
doi:10.1016/S0022-460X(03)00190-1
 ARTICLE IN PRESS

2 H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14

A general approach for detecting damage, proposed by several researchers [2–15], is based on
comparing the frequency changes obtained using experimental data collected from the structure
with the sensitivity of the modal parameters obtained from an analytical model of the structure.
The sensitivity of the natural frequencies of the structure to changes in the stiffness, mass, and
damping are calculated using a finite element analysis. However, methods based on a sensitivity
analysis require an accurate analytical model of the structure being investigated, which can be
difficult to obtain. Also, most of the damage detection algorithms require a significant amount of
modal test data. These requirements make the damage detection procedure expensive, time-
consuming, and impractical for real-life structures in service.
Abdalla et al. [1] formulated the damage detection problem as a convex optimization
problem involving linear matrix inequality constraints. Pandey et al. [2] presented an evaluation of
changes in the structure flexibility matrix as a candidate method for identifying both the
presence and the location of the damage. Farhat et al. [3] developed a sensitivity-based
methodology for improving the finite element model of a given structure using test modal data
and a few sensors. Li et al. [5] used modal characteristics extracted from vibration tests with an
original finite element model in an identification approach developed to combine the advantages
of two classes of techniques: eigensensitivity and multiple-constraint matrix adjustment. Smith [7]
applied alternating projection algorithms for approximating a matrix in structural model
updating. The desired matrix properties, such as sparsity, definiteness, and the satisfaction of
eigenconstraints, were imposed as side constraints for a minimization problem formulated to
produce an updated matrix model that better matched the measured data. Ahmadian et al. [8]
addressed the problem of selecting a side constraint and determining the regularization parameter
when updating the model; the constraint weight was determined by the regularization
parameter. They considered methods based on singular value decomposition, cross-validation,
and L-curves.
In this paper, structural damage was detected using techniques based on the physical meaning
of the stiffness matrices and on a few frequency response functions (FRFs) that can be obtained
from a test. The basis for this method is that damage produces a decrease in the dynamic stiffness
EI: The symmetry and physical properties of the stiffness matrices were maintained by
considering their sparsity.
Despite the presence of experimental errors in the FRF test data, it is generally assumed that
this data is a better representation of how the structure behaves than the initial predictions from
the FE model. Consequently, in this paper, it is assumed that to the FE model was adjusted, or
updated, so that the results it produces are in some sense closer to the experimental results.
There has been a significant amount of work on generating and testing different updating
methods over the past 20 years. The resulting algorithms may be split into several categories based
on whether they work in the frequency or modal domains and whether they adjust the mass and
stiffness matrices directly (direct methods) or make parametric changes to the model (indirect or
parametric methods) [7].
Although it is fairly easy to detect the presence of damage in a structure from changes in the
natural frequencies, it is difficult to determine the location of the damage. This is because damage
at two different locations may produce the same amount of frequency change. Previous research
[2–12] based on modal methods that use the natural frequency and modal shape changes should
not be used in damage assessment methodologies because the modal shapes of a damaged and
 ARTICLE IN PRESS

H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14 3

undamaged system are nearly the same. Hence, in this paper, FRFs are used to detect damage,
rather that the natural frequencies and modal shapes.
In this paper, to reduce the computational time required to analyze a structure that has been
damaged at more than one location, two FRFs were used iteratively for different frequency
values. This created a large number of equations to accurately identify the damage, and
eliminated the need for a large amount of FRF data. The method worked for each discrete
frequency and could be optimized using the least-squares method. The damage detection equation
was rearranged to avoid ill-conditioned problems while the matrices were being inverted. The EI
stiffness values were separated from the element stiffness matrix when using the least-squares
method; this minimized the number of unknowns in the stiffness matrix and preserved the
intrinsic FEM connectivity properties of the model. The derived damage detection method was
tested on a simple cantilever and a helicopter rotor blade.

2. Formulation of the equations of motion

Damage will alter the dynamic characteristics of a structure. This is characterized by changes in
the modal parameters, i.e., the modal frequencies, damping values, and mode shapes associated
with each modal frequency. Changes also occur in some of the structural parameters, such as the
mass, damping, stiffness, and flexibility matrices of the structure.
The equations of motion of a structure with N degrees of freedom and viscous damping
coefficients can be expressed as [16]
.
½MfxðtÞg ’
þ ½DfxðtÞg þ ½KfxðtÞg ¼ ff ðtÞg; ð1Þ
where ½M; ½D; and ½K represent the n n mass, damping, and stiffness matrices. If we assume a
harmonic input, the external force and displacement can be expressed as ff ðtÞg ¼ fFðoÞgejot and
fxðtÞg ¼ fX ðoÞgejot : Substituting into Eq. (1) yields
ð o2 ½M þ jo½D þ ½KÞfX ðoÞgeiot ¼ fFðoÞgeiot : ð2Þ
From the above equation, the FRF matrix, ½HðoÞ; is defined as
½HðoÞ ¼ ð o2 ½M þ jo½D þ ½KÞ 1 : ð3Þ
Then Eq. (2) can be expressed as
fX ðoÞg ¼ ½HðoÞfF ðoÞg: ð4Þ

3. Damage detection methods using FRF data

The objective of this paper is to identify the location and amount of structural damage, which is
related to the stiffness matrices. Therefore, the damping matrix ½D in Eq. (1) is neglected even
though damping reduces the natural frequency slightly. Small local changes in the mass of a beam
can cause significant changes in the frequency. Hence, we also assume that the mass of the
structure does not change after it has been damaged. We assign the damaged part of the structure
to the connection between two substructures, illustrated in Fig. 1.
 ARTICLE IN PRESS

4 H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14

Substructure Substructure
I II
Unknown

Fig. 1. Damaged structure divided into two substructures.

The stiffness matrix of the damaged part can then be obtained from
½Kunknown ¼ ð½K o2 ½MÞ ð½Knew o2 ½MÞ; ð5Þ
where ½K is the stiffness matrix of the structure without damage, and ½Knew ¼ ½KI þ ½KII is
composed of the stiffness matrices of substructure I and II without connections, i.e.,
" # " #
½KI ½0 ½0 ½0
½Knew ¼ þ : ð6Þ
½0 ½0 ½0 ½KII
The subscript ‘unknown’ indicates the stiffness value of the damaged parts in the structure.
Using the relation ½HðoÞ 1 ¼ ð½K o2 ½MÞ and Eq. (5), we can theoretically obtain the
stiffness matrix for the damaged portion of the structure by subtracting the inverse of the FRF
with damage from the inverse of the FRF without damage. However, this method requires a large
number of calculations because many sub-stiffness matrices must be calculated continuously due
to the damaged points. Also, one needs to measure all the modes of the structure, especially the
high-frequency modes, to obtain a good estimate of the stiffness matrix. Because of obvious
limitations in experimental instrumentation, it becomes increasingly difficult to measure the
higher frequency response data. Hence, to find the location and amount of damage using just a
few test data, it is necessary to iterate the calculations using both the theoretical FRFs from the
finite element model and the experimental FRFs. However, only a few lower frequency modes
need to be measured to generate the necessary stiffness matrices.
In Eq. (3), if damping is neglected, the FRFs matrix can be expressed as
½HðoÞ ¼ ð o2 ½M þ ½KÞ 1 : ð7Þ
Let the FRFs from the test be ½HðoÞT : Using the mass matrix ð½MA Þ and the stiffness matrix
ð½KA Þ from the finite element model, the simulated FRFs are
½HðoÞA ¼ ð o2 ½MA þ ½KA Þ 1 : ð8Þ
When a structure contains damage, ½HðoÞT differs from ½HðoÞA due to the stiffness reduction.
Hence, we can find the location and amount of damage through iterations by matching ½HðoÞT to
½HðoÞA : This basic computational scheme is illustrated in Fig. 2. However, in practice, these
calculations are extremely difficult to perform because of the matrix inverting process, which may
introduce highly ill-posed problems.
This method
pffiffiffi was tested using a simple discrete model with six degrees of freedom. The most
reliable 1= 2 of the peak FRF values were used to compare ½HðoÞT and ½HðoÞA : In practical
applications, the measurements will include errors arising from many potential sources; therefore,
errors in the amplitude measurement of the FRFs were considered. An intentional 10% random
noise error was added to simulate noise effects. During the iterative process, according to the
scheme shown in Fig. 2, 1% of the stiffness values were reduced in the undamaged ½K:
 ARTICLE IN PRESS

H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14 5

Start

Find [H(ω)]T

Change Ki Find [H(ω)]A while

i=2,3~n reducing Ki values

Reduce K value
Compare error of 1/ 2 max.,
min. of [H(ω)]A and [H(ω)]T
No

Yes No
Iteration= 100? error<=0.5%

Yes

Plot
[ H(ω)]A and [H(ω)]T

End

Fig. 2. Computational scheme for the basic method.

The FRFs at different iteration times during the reduction of one spring constant value are
shown in Fig. 3. As the iteration number increases, ½HðoÞA approaches ½HðoÞT and the spring
constant is reduced to the value that was assumed in the damaged structure.
The above method can be used to find the position and the amount of damage at one element
using just one FRF. However, for structures that have been damaged at more than one element,
we have to compare the test FRF to analytical FRFs obtained using different combinations of the
sub-stiffness matrices. This makes the method difficult to apply due to the sharp increase in the
amount of calculations that are required. Therefore, a new approach is introduced that uses a
specific FRF and iterates for different frequencies.
From Eq. (7), the test FRF is
½HðoÞT ¼ ð o2 ½M þ ½KT Þ 1 : ð9Þ
Rearranging the above equation yields ½KT ¼ ½HT 1 þ o2 ½M: However, it is extremely difficult to
obtain ½HT 1 because we cannot measure all the FRFs, and the inversion process creates ill-
conditioned problems. Therefore, a new method is required that locates the damaged position
using just a few test sets. From Eq. (9), ½HðoÞT ð o2 ½M þ ½KT Þ ¼ ½I; which can be expressed in
detail as
2 3 2 3 2 3
H11 ðoÞ H12 ðoÞ H1n ðoÞ T11 ðoÞ T12 ðoÞ T1n ðoÞ 1 0 0
6 7 6 7 6 7
6 H21 ðoÞ H22 ðoÞ H2n ðoÞ 7 6 T21 ðoÞ T22 ðoÞ T2n ðoÞ 7 6 0 1 07
6 7 6 7 6 7
6 7 6 7¼6 7 ;
6 7 6 7 6 7
6 7 6 7 6 7
4 5 4 5 4 5
Hn1 ðoÞ Hn2 ðoÞ Hnn ðoÞ T Tn1 ðoÞ Tn2 ðoÞ Tnn ðoÞ 0 0 1 n n
ð10Þ
 ARTICLE IN PRESS

6 H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14

FRF H(w) FRF H(w)

-3.5 -3

-4
-4
-4.5

-5
-5
-5.5
log (FRF)

log (FRF)
-6 -6

-6.5
-7
-7

-7.5
-8
-8

-8.5 -9
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
Frequency (rad/s) Frequency (rad/s)

20th iteration 30th iteration

FRF H(w) FRF H(w)

-3.5 -3

-4
-4
-4.5

-5
-5
-5.5
log (FRF)

log (FRF)

-6 -6

-6.5
-7
-7

-7.5
-8
-8

-8.5 -9
0 500 1000 1500 2000 2500 3000 0 500 1000 1500 2000 2500 3000
Frequency (rad/s) Frequency (rad/s)

40th iteration 50th iteration

Fig. 3. FRFs at different iteration times.

where ½T ¼ ð½KT o2 ½MÞ: This expression is frequency dependent rather than a constant. For a
specific xth row of ½HT ; we can obtain the 1 n matrix

½ Hx1 Hx2 Hxn T ð½KT Þn n ¼ ½0 y 010 y 0  þ o2 ½ Hx1 Hx2 Hxn T ½Mn n

by rearranging the above equation. However, its inverse matrix cannot be determined uniquely to
give ½KT : Thus, the above equation is used iteratively for different frequencies, oi : This provides a
sufficient number of equations using the least-squares method.
 ARTICLE IN PRESS

H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14 7

If we consider the first row, the different n frequencies yield to the following equation:
2 3 2 3 2 3
H11 ðo1 Þ H12 ðo1 Þ H1n ðo1 Þ K11 K12 K1n 1 0 0
6 7 6 7 6 7
6 H11 ðo2 Þ H12 ðo2 Þ H1n ðo2 Þ 7 6 K21 K22 K2n 7 61 0 07
6 7 6 7 6 7
6 7 6 7 ¼6 7
6 7 6 7 6 7
6 7 6 7 6 7
4 5 4 5 4 5
H11 ðon Þ H12 ðon Þ H1n ðon Þ T Kn1 Kn2 Knn 1 0 0
2 3 2 T 3
H11 ðo1 Þ H12 ðo1 Þ H1n ðo1 Þ M11 M12 M1n
2 36 7 6 7
& 6 H11 ðo2 Þ H12 ðo2 Þ H1n ðo2 Þ 7 6 M21 M22 M2n 7
6 76 6 7 6 7
þ4 o2i 56 7 6 7: ð11Þ
7 6 7
6 7 6 7
& 4 5 4 5
H11 ðon Þ H12 ðon Þ H1n ðon Þ T Mn1 Mn2 Mnn
If the left-side term ½HT is ½A and the right-side term is ½B
½A½KT ¼ ½B: ð12Þ
The solution of Eq. (12), ½KT ¼ ð½AT ½AÞ 1 ½AT ½B using the least-squares method, will give the
unknown value ½KT : However, in our case, the determinant of ½AT ½A is usually close to zero;
thus, a different approach is required.
When we take the inverse matrix and consider the properties of the stiffness matrix elements
that are distributed near the diagonal, we can reduce the unknown values and required FRFs for a
general structure. For example, if we use only two elements that exist in the first column of the ½K
matrix, and make use of the fact that the mass ½M is a diagonal matrix for static structures, from
Eq. (11) we obtain
2 3 8 2
9
H11 ðo1 Þ H12 ðo1 Þ >
> 1 þ o 1 H 11 ðo 1 ÞM11 >
>
6 7" >
> >
>
6 H11 ðo2 Þ H12 ðo2 Þ 7 # >
> 1 þ o 2
H ðo ÞM >
>
6 7 K11 < 2 11 1 11 =
6 H11 ðo3 Þ H12 ðo3 Þ 7 ¼ 1 þ o 2
H ðo ÞM : ð13Þ
6 7 K > 3 11 1 11
>
6 7 21 >
> >
>
4 ^ ^ 5 >
> ^ >
>
>
: >
;
2
H11 ðon Þ H12 ðon Þ 1 þ on H11 ðo1 ÞM11
Thus, we can identify all the non-zero stiffness values using just a few FRFs in one row of the
FRF matrix.
The discrete system model was used again to verify this method. The structure was assumed to
be damaged at several locations. In the FRF matrix, the first row was used repeatedly for various
frequency values. The largest error was determined by comparing the FRF before and after the
damage. The least-squares method of Eq. (13) was used to calculate the reduction of the stiffness
values. The identified stiffness reductions were the same as the assumed damage values in the
discrete model.
Although this method gave accurate estimations of the damage for this case, it does not work well
for some other cases. For example, when the columns in matrix ½A have the same values, the damage
estimations are incorrect. Although the orders of the elements in the ith and jth columns of the
stiffness matrix are different from each other, they have a similar constitution that has same element
 ARTICLE IN PRESS

8 H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14

values such as ½ 0 0 0 106 2 106 106 T and ½ 0 0 0 106 106 2 106 T in the
simple discrete model with six degrees of freedom. This gives the same FRF values using either Eq. (8)
or Eq. (9).
The above method cannot be applied to general structures that have the same element material
property, type, and size throughout, as is the case in some analysis models. For example, in a
uniform beam model, the stiffness matrices of each element are equal. The same element matrices
are superimposed to obtain the global stiffness matrix. Since we use only one specific row of
FRFs, the values that are obtained from the matrix multiplication are equal even though the
element values are different from each other. To overcome this problem, we concentrated on
changes in the stiffness matrix caused by damage.
In Eqs. (8) and (9), if we assume that the mass of the structure remains constant before and
after the damage, the change in the stiffness as a result of the damage is ½DK ¼ ½KA ½KT ¼
½HA 1 ½HT 1 : When multiplied by ½HT ; this yields

½HT ½DK ¼ ½HT ð½KA o2 ½MÞ ½I; ð14Þ

where ½HA can be calculated with using ½KA and ½M: If we assume that we can measure the first
row of the FRF matrix to obtain the specific frequency, the above equation can be rearranged as
follows:

2 3
DK11 DK12 ? DK1n
6 DK DK22 DK2n 7
6 21 ? 7
½ HT11 HT12 ? HT1n T 6 7
4 ^ ^ & ^ 5
DKn1 DKn2 ? DKnn
2 3 1
HA11 HA12 ? HA1n
6H HA22 ? HA2n 7  
6 A 7
¼ ½ HT11 HT12 ? HT1n T 6 21 7 1 0 ? 0 ; ð15Þ
4 ^ ^ & ^ 5
HAn1 HAn2 ? HAnn A

where ½DK is the change in the stiffness value of the damaged elements.
The previous methods detected damage in structures using modal test data and element-by-
element adjustments to the FE model. In this method, the following element equations of the mass
and stiffness matrices are used in a beam model to minimize the number of unknowns in the
stiffness matrix and to preserve the intrinsic FEM connectivity properties of the model [13]

2 3
12 6L 12 6L
EI 6
6 6L 4L2 6L 2L2 7
7
½Ke ¼ 6 7;
L 4 12
3
6L 12 6L 5
6L 2L2 6L 4L2
 ARTICLE IN PRESS

H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14 9

2 3
156 22L 54 13L
rAL66 22L 4L2 13L 3L2 7
7
½Me ¼ 6 7; ð16Þ
420 4 54 13L 156 22L 5
13L 3L2 22L 4L2
where E; I; L; r; and A are Young’s modulus, area moment of inertia, length of the beam element,
density, and cross-sectional area, respectively. Separating the EI values from Eq. (16) to reduce
the number of variables and to maintain the physical meaning of the stiffness matrix yields
0 2 31
12 6L 12 6L
B 1 6 6L 4L2 6L 2L2 7 C
B 6 7C
½Ke ¼ EI B 3 6 7C ¼ EI½K 0 e ; ð17Þ
@L 4 12 6L 12 6L 5A
6L 2L2 6L 4L2
where ½K 0 e can be calculated using a finite element model. From the 1 n row on the right side of
Eq. (15), if a value exists in the ith column then the stiffness value corresponding to the ith column
element of ½DK will be changed. The ½K 0 e value in Eq. (17) is known; only the EI values are
unknown. Thus, Eq. (15) becomes
ð½HT Þxn ½K 0 ð:;iÞ DðEIÞ ¼ ð½HT Þxn ð½HA Þð:;iÞ1 ½Iðx;iÞ ; ð18Þ
where ð½HT Þxn is the xth row in the FRF matrix that was obtained from the test, ½K 0 ð:;iÞ DðEIÞ is
used to apply Eq. (17) to the ith column of ½DK in Eq. (15), and ð½HA Þð:;iÞ1 ¼ ð½KA o2 ½MÞð:;iÞ1 : If
we let the known values of the left side of Eq. (18) be aðx; iÞ and the known values on the right side
be bðx; iÞ; we have
DðEIÞ ¼ bðx; iÞ=aðx; iÞ: ð19Þ
Thus, we can calculate the change in stiffness due to structural damage by using calculated DðEIÞ
values.

4. Numerical examples

4.1. Example 1: cantilever with and without cracks

In this section, the damage detection method was applied to a simple cantilever beam system.
For the system model shown in Fig. 4, the FRFs can be obtained from an experiment with the
physical model. However, representative FRFs were synthesized numerically using the continuous
system model instead. It was assumed that the structure can be modelled using the finite element
method with basic structural elements (i.e., bars, beams, plates, membranes, and shells). It was
also assumed that the damage size is the same as the element size.
In the cantilever, E ¼ 3 107 psi, L ¼ 30 in, I ¼ 0:0833 in4, A ¼ 1 in2, and r ¼ 0:00073 lb s2/in.
The model had 10 elements. The stiffness and mass matrices were obtained by superimposing
Eq. (16). Since the cantilever had a fixed boundary condition at the left end, the first and the
second rows and columns were eliminated from the 22 22 matrix, as we are concerned only with
 ARTICLE IN PRESS

10 H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14

30 in

1in
1 2 3 4 5 6 7 8 9 10

1in

Fig. 4. Cantilever beam model.

transverse displacements and rotations. Hence, the following 20 20 stiffness matrix ½KA and
mass matrix ½M were created
2 3
2:2213 0 1:1107 1:6660 ? 0 0 0 0 0
6 0 6:6640 1:6660 1:6660 0 0 0 0 0 7
6 ? 7
6 7
6 1:1107 1:6660 2:2213 0 ? 0 0 0 0 0 7
6 7
6 1:6660 1:6660 0 6:6640 0 0 0 0 0 7
6 ? 7
6 7
6 0 0 1:1107 1:6660 ? 0 0 0 0 0 7
6 7
6 0 0 1:6660 1:6660 0 0 0 0 0 7
6 ? 7
6 7
6 ^ ^ ^ ^ & ^ ^ ^ ^ ^ 7
6 7
66 7
½KA ¼ 1 10 6 0 0 0 0 ? 1:6660 0 0 0 0 7 ;
6 7
6 0 0 0 0 ? 1:6660 0 0 0 0 7
6 7
6 7
6 0 0 0 0 ? 0 1:1107 1:6660 0 0 7
6 7
6 0 0 0 0 6:6640 1:6660 1:6660 0 0 7
6 ? 7
6 7
6 0 0 0 0 ? 1:6660 2:2213 0 1:1107 1:6660 7
6 7
6 0 0 0 0 1:6660 0 6:6640 1:6660 1:6660 7
6 ? 7
6 7
4 0 0 0 0 ? 0 1:1107 1:6660 1:1107 1:6660 5
0 0 0 0 ? 0 1:6660 1:6660 1:6660 3:3320 20 20

2 3
16:269 0 2:8157 2:0336 ? 0 0 0 0 0
6 0 3:7543 2:0336 1:4079 0 0 0 0 0 7
6 ? 7
6 7
6 2:8157 2:0336 16:269 0 ? 0 0 0 0 0 7
6 7
6 2:0336 1:4079 0 3:7543 0 0 0 0 0 7
6 ? 7
6 7
6 0 0 2:8157 2:0336 ? 0 0 0 0 0 7
6 7
6 0 0 2:0336 1:4079 ? 0 0 0 0 0 7
6 7
6 7
6 ^ ^ ^ ^ & ^ ^ ^ ^ ^ 7
6 7
6 7
½M ¼ 1 10 4 6 0 0 0 0 ? 2:0336 0 0 0 0 7 :
6 7
6 0 0 0 0 ? 1:4079 0 0 0 0 7
6 7
6 7
6 0 0 0 0 ? 0 2:8157 2:0336 0 0 7
6 7
6 0 0 0 0 3:7543 2:0336 1:4079 0 0 7
6 ? 7
6 7
6 0 0 0 0 ? 2:0336 16:269 0 2:8157 2:0336 7
6 7
6 0 0 0 0 1:4079 0 3:7543 2:0336 1:4079 7
6 ? 7
6 7
4 0 0 0 0 ? 0 2:8157 2:0336 8:1343 3:4414 5
0 0 0 0 ? 0 2:0336 1:4079 3:4414 1:8771 20 20
 ARTICLE IN PRESS

H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14 11

We assumed that stiffness value was altered from E ¼ 3 107 to 2 107 psi due to damage at the
3rd, 4th, and 9th elements of the model. The stiffness matrix ½KT that was necessary to simulate
the test FRF was obtained from Eq. (16). The mass matrix was assumed to be equal to the
analytical values. Using these matrices, the FRFs were calculated from Eqs. (8) and (9). Note that
the FRF in Eq. (9) should be obtained from an actual test rather than from the simulated values
used in this example.
Two FRF values in one row of the global FRF matrix were used, corresponding to two
elements with stiffness values. The FRF was assumed to have been measured at a test frequency of
3000 rad/s. Using the first row of the global matrix, the value bð1; iÞ in Eq. (19) was
bð1; iÞ ¼ ½ 0:0 0:0 0:0279 0:3813 0:0469 0:0044 0:0748
0:0687 0:0 0:0 0:0 0:0 0:0 0:0
0:0654 0:3018 0:0654 0:1056 0:0 0:0 1 20 :

Here, the 3rd to the 8th and the 15th to the 18th elements had non-zero values. If the cantilever
beam were undamaged, all values of bð1; iÞ would be zero. If we consider the properties of
Eq. (16), we can determine that the 3rd, 4th, and 9th components are damaged. Also, we can
calculate the decrease in EI values using Eqs. (18) and (19). By taking independent values that are
not superimposed, we can obtain the reduction in the stiffness value of EI ¼ 8:33 105 lb m2,
which corresponds to bð1; 3Þ=0.0279 in the 3rd element, bð1; 8Þ= 0.0687 in the 4th element, and
bð1; 15Þ= 0.0654 in the 9th element. If we assume that only the E values were affected by the
damage (the I values of each element remain unchanged), we obtain DE ¼ 1 107 psi, which was
assumed initially.
Next, suppose that each element has a different amount of damage, i.e., the 3rd and 9th element
stiffness values were changed from E ¼ 3 107 to 2 107 psi, and the 4th element stiffness value
was changed to E ¼ 2:5 107 psi. Using the same procedure, and assuming the same 3000-rad/s
frequency, the calculated b values using the first row of the global FRF matrix are
bð1; iÞ ¼ ½ 0:0 0:0 0:0288 0:3815 0:0004 0:1788 0:0292
0:0286 0:0 0:0 0:0 0:0 0:0 0:0
0:0621 0:2867 0:0621 0:1003 0:0 0:0 1 20 :

From this, bð1; 3Þ=0.0288, and the reduced stiffness value of the 3rd element is
EI ¼ 8:33 105 lb m2. Also, from bð1; 8Þ= 0.0286 and bð1; 15Þ= 0.0621, the reduced stiffness
values of the 4th and 9th elements are EI ¼ 4:165 105 and 8.33 105 lb m2, respectively. For a
fixed value of I ¼ 0:0833 in4, the change in Young’s modulus is DE ¼ 1 107 psi for the 3rd and
9th elements, and DE ¼ 0:5 107 psi for the 4th element. Therefore, using this proposed method,
we can accurately find the location and the amount of damage in a structure.

4.2. Example 2: the helicopter rotor blade

The proposed method was also applied to the helicopter rotor blade shown in Fig. 5. The rotor
blade had a length of 1100 mm. A VR-12 airfoil was used between 220 and 880 mm, and a VR-15
airfoil between 1036 and 1100 mm. A linear airfoil shape transition was assumed between 880 and
1036 mm.
 ARTICLE IN PRESS

12 H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14

200

220
880 1100

Fig. 5. Configuration of the rotor blade.

Table 1
Change in the natural frequency of the rotor blade due to damage (rad/s)
Mode Before damage After damage
1st 2.7592e+001 2.7012e+001
2nd 1.9023e+002 1.8660e+002
3rd 5.3105e+002 5.1870e+002
4th 1.0306e+003 1.0164e+003
5th 1.7194e+003 1.6850e+003
6th 2.5841e+003 2.5052e+003
7th 3.5890e+003 3.5424e+003
8th 4.7469e+003 4.6331e+003
9th 6.0690e+003 5.9033e+003
10th 7.5217e+003 7.4423e+003

The rotor blade was modelled as a cantilever beam since a simplified analysis is generally
performed in the aircraft conceptual design phase. A total of 16 stiffness values were used, and the
element masses were assumed to be concentrated at each node. A total of 42 beam elements were
used so that 86 86 mass and stiffness matrices were constructed. Since fixed boundary
conditions were applied at the root, the first and the second rows and columns were removed so
that 84 84 mass and stiffness matrices remained.
The 3rd, 19th, and 24th elements were assumed to be damaged. The stiffness EI of the 3rd
element was reduced by 50% from 107.929 to 53.960 N m2, the 9th element from 14.271 to
7.136 N m2, and the 24th element from 14.271 to 10.000 N m2. The changes in the natural
frequencies are shown in Table 1. From the many FRFs, we chose to compare Hð1; 4Þ of the
damaged rotor blade to Hð1; 4Þ of the undamaged rotor blade in Fig. 6.
Non-zero values of the 1 84 bð1; iÞ vector obtained using this proposed method are listed in
Table 2. These values indicate that the 3rd, 19th, and 24th elements are damaged. Using the same
method as in the previous example, the identified change in stiffness of the 3rd, 19th, and 24th
elements are DEI3 ¼ 53:9690 N m2 ; DEI19 ¼ 7:1353 N m2 ; DEI24 ¼ 4:2713 N m2 : These values are
exactly the same as the assumed values. Therefore, we conclude that this method of damage
detection can be used for structures with damage at more than one location.
 ARTICLE IN PRESS

H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14 13

FRF H(w)
-4.5

-5 H(1,4) before
damage

-5.5
H(1,4) after
damage
-6
l og (FRF)

-6.5

-7

-7.5

-8
0 500 1000 1500 2000 2500 3000
Frequency (rad)

Fig. 6. Comparison of the H(1,4) FRF before and after damage.

Table 2
Non-zero bð1; iÞ values of the bð1; iÞ vector
Columns 3–6 0.0023 0.0001 0.0023 0.0001
Columns 35–38 0.0015 0.0001 0.0015 0.0001
Columns 45–48 0.0009 0.0000 0.0009 0.0000

5. Conclusions

In this paper, we developed damage detection methods that use FRFs to find the location and
severity of damage in structures.
We started with a basic method that required a sharp increase in computational effort when the
structure was damaged at more than one location. To overcome this problem, the least-squares
method was used for different frequencies of a specific FRF; however, this was not satisfactory
since the method encountered problems calculating the inverse of the matrices when the stiffness
matrices had repeated elements. The best method identified structural damage using just a few
FRFs. To minimize the number of unknowns, common stiffness variables were extracted from the
finite elements. A few FRFs were sufficient to accurately identify the extent and location of the
damage. The repeated use of a few FRFs for different frequency values, considering only a vector
subset of the full set of FRFs, is the original work of this paper.
The required number of measured FRFs could be further reduced by using more constraint
equations. Future research in this field should be focused on discovering the limitations of these
methods using actual test data instead of simulated data. Also, geometric effects of damage
directions must be investigated when modelling and testing for structural damage.
 ARTICLE IN PRESS

14 H.Y. Hwang, C. Kim / Journal of Sound and Vibration 270 (2004) 1–14

Acknowledgements

This work was supported by a Korea Research Foundation Grant (KRF-99-003-

E00104E1302).

References

[1] M. Abdalla, K.M. Grigoriadis, D.C. Zimmerman, Enhanced damaged detection using linear matrix inequalities,
Proceedings of the 16th International Modal Analysis Conference, Santa Barbara, CA, 1998, pp. 144–150.
[2] A.K. Pandey, M. Biswas, Damage detection in structures using changes in flexibility, Journal of Sound and
Vibration 169 (1994) 3–17.
[3] C. Farhat, F.M. Hemez, Updating finite element dynamics models using an element-by-element sensitivity
methodology, American Institute of Aeronautics and Astronautics Journal 31 (9) (1993) 1702–1711.
[4] P. Hajela, F.J. Soeiro, Structural damage detection based on static and modal analysis, American Institute of
Aeronautics and Astronautics Journal 28 (6) (1990) 1110–1115.
[5] C. Li, S.W. Smith, Hybrid approach for damage detection in flexible structures, Journal of Guidance, Control, and
Dynamics 18 (13) (1995) 419–425.
[6] H.T. Banks, D.J. Inman, D.J. Leo, Y. Wang, An experimentally validated damage detection theory in smart
structures, Journal of Sound and Vibration 191 (1996) 859–880.
[7] S.W. Smith, Iterative matrix approximation for model updating, Mechanical Systems and Signal Processing 12 (1)
(1998) 187–201.
[8] H. Ahmadian, J.E. Mottershead, M.I. Friswell, Regularisation methods for finite element model updating,
Mechanical Systems and Signal Processing 12 (1) (1998) 47–64.
[9] S.W. Doebling, Rank optimal update of elemental stiffness parameters for structural damage identification,
American Institute of Aeronautics and Astronautics Journal 34 (12) (1996) 2615–2621.
[10] J.S. Lew, Using transfer function parameter changes for damage detection of structures, American Institute of
Aeronautics and Astronautics Journal 33 (11) (1995) 2189–2193.
[11] I. Sheinman, Damage detection and updating of stiffness and mass matrices using mode data, Computers and
Structures 59 (1) (1996) 149–156.
[12] D. Li, Z. Zheng, K. He, B. Wang, Damage detection in offshore structures by the FRF method, Proceedings of the
International Offshore Mechanics and Arctic Engineering, Calgary, Vol. 1, Part B, 1992, pp. 601–604.
[13] D.L. Logan, A First Course in the Finite Element Method, 2nd Edition, PWS, Boston, 1992.
[14] H.Y. Hwang, Identification techniques of structure connection parameters using frequency response functions,
Journal of Sound and Vibration 212 (1998) 469–769.
[15] H.Y. Hwang, Identification Methods for Physical Parameters in Structure, Doctoral Thesis, Georgia Institute of
Technology, Atlanta, GA, 1993.
[16] D.J. Ewins, Modal Testing: Theory and Practice, Research Studies Press, Letchworth, Hertfordshire, England,
1984.