中国科技论文在线 http://www.paper.edu.

cn
ARTICLE IN PRESS

JOURNAL OF
SOUND AND
VIBRATION
Journal of Sound and Vibration 305 (2007) 298–307
www.elsevier.com/locate/jsvi

Short Communication

Structural damage identification based on residual force vector

Q.W. Yang, J.K. Liu
Department of Mechanics, Zhongshan University, Guangzhou 510275, PR China
Received 17 June 2005; received in revised form 26 October 2005; accepted 17 March 2007

Abstract

Structural damage identification methods based on the residual force vector are studied in this paper. Using the residual
force vector, the node residual force vector is defined to locate the suspected damaged elements preliminarily. Then three
damage quantification techniques are studied to identify damages more precisely. The first is the algebraic solution of the
residual force equation, the second is the MREU technique and the third is the natural frequency sensitivity method. A
mode shape expansion technique based on the best achievable eigenvector concept is presented to solve the incomplete
measurement problem. These damage detection methods are demonstrated on a numerical example and the measurement
noises are discussed.
r 2007 Elsevier Ltd. All rights reserved.

1. Introduction

Techniques based on dynamic parameters for detecting damages in a structure have attracted much
attention in recent years. Modal frequencies and mode shapes are the most popular parameters used in the
identification. The basic idea of these techniques is that modal parameters are functions of the physical
properties of the structure (mass, damping and stiffness). Therefore, changes in the physical properties will
cause changes in the modal properties. Many methods were developed recently using modal parameters as
damage indicators.
An important class of damage identification methods is based on the updating or modification of structural
matrices. The residual force vector is widely used in many damage detection methods using optimal matrix
modification [1–5]. Chen and Garba [1] put forward a theory for assessing the occurrence, location and extent
of potential damage using on-orbit response measurements. This method detects damages by using the
minimum norm solution of the residual force equation. Zimmerman and Kaouk [2] made use of a minimum
rank update theory to detect structural damages. The damage sites are located firstly by the residual force
vector and the damage extents are assessed by the minimum rank update theory. Doebling improved this
method and presented a new technique termed the minimum-rank elemental update (MREU) by computing
the minimum rank updates directly to the elemental stiffness parameters [3]. Chiang and Lai [4] presented a
two-stage structural damage detection method. The residual force vector is used to localize damages

Corresponding author.
E-mail address: [email protected] (J.K. Liu).

0022-460X/$ - see front matter r 2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jsv.2007.03.033
转载
中国科技论文在线 http://www.paper.edu.cn
ARTICLE IN PRESS
Q.W. Yang, J.K. Liu / Journal of Sound and Vibration 305 (2007) 298–307 299

preliminarily and the simulated evolution method is employed to determine damage extents. Mares and Surace
proposed a genetic algorithm to identify damage in elastic structures. The location and quantification of the
extent of the damage is performed with genetic techniques implemented by using the residual force method,
which is based on conventional modal analysis theory [5]. In short, these above methods all begin with the
residual force vector but use different techniques to obtain damage extents, so the accuracy of the residual
force vector is very important to those methods. The minimum norm method is shown to be unfeasible in
damage identification in practice because the residual force equation is ill conditioned with the measurement
noises, while the minimum rank update techniques can obtain better results only when the number of modes
used in calculation equals the rank of perturbed matrix [2,3]. However, the rank of the perturbed matrix is
unknown after damages occurrence; in other words, the number of damaged elements is unknown, which
limits the use of the minimum rank update method. The genetic algorithms [4,5] need much expense of the
computation and the results of damage quantification sometimes are not ideal because the mode shapes used
in calculation often have large measurement noises.
In this study, the residual force method is developed in order to detect structural damages successfully when
the measured modal parameters are incomplete and have noises. A mode shape expansion algorithm based on
the best achievable technique is proposed to solve the problem of the test/analysis degree of freedom (dof)
mismatch. Damage locations are determined firstly by the node residual force vector, which is defined by the
residual force vector. Three techniques for damage quantification are studied to obtain damage extents after
the suspected damaged elements are determined, the first is the algebraic solution of the residual force
equation, the second is the MREU technique and the third is the natural frequency sensitivity method. These
damage detection methods are demonstrated on a numerical example with the measurement noises.

2. The residual force equation

Without loss of generality, by assuming that the mass matrix is unchanged as damage occurs, the eigenvalue
equation for an n dofs finite element model of a damaged structure is
ðK d  ldj MÞfdj ¼ 0, (1)

K d ¼ K u  DK, (2)
where M is the mass matrix, Ku and Kd are the stiffness matrices associated with the undamaged and damaged
structural models, respectively, DK is the corresponding changes, ldj and fdj are the jth eigenvalue and
eigenvector of the damaged structure, respectively.
Substituting Eq. (2) into Eq. (1) yields
ðK u  ldj MÞfdj ¼ DKfdj . (3)
Letting bj ¼ ðK u  ldj MÞfdj , Eq. (3) can be rewritten as
DKfdj ¼ bj , (4)
where bj is the jth residual force vector.

3. Damage identification

Eq. (4) can be expressed as

2 3 8 9
DkT1 >
> bj1 >
>
6 T7 >
> >
>
6 Dk2 7 < bj2 >
> =
6 7
6 . 7fdj ¼ . , (5)
6 .. 7 >
> .. >
>
4 5 >
> >
>
: bjn >
> ;
DkTn
where Dki (i ¼ 1n) is the stiffness change vector of the ith dof. Dki is nonzero only if the ith dof is damaged.
By inspecting Eq. (5) it can be seen that the residual force vector bj will have nonzero elements only in the
中国科技论文在线 http://www.paper.edu.cn
ARTICLE IN PRESS
300 Q.W. Yang, J.K. Liu / Journal of Sound and Vibration 305 (2007) 298–307

damaged dofs. So we can find damaged elements according to the relation between the element number and
the dof number. For the vibration data with measurement noises, the damaged dofs can be determined by the
larger absolute elements in bj.
For the convenience of actual engineering practices, the node residual force vector nbj for the jth mode
shape is defined as
8 9
>
> nbj1 >
>
>
> >
>
< nbj2 >
> =
nbj ¼ .. , (6)
>
> . >
>
>
> >
>
: nbjq >
> ;

where q is the number of nodes in the structural finite element model. Each element in nbj is defined as the sum
of the corresponding absolute values of elements in bj associated with the dofs of this node. For example,
assuming the 1st node is corresponding to the 1st and 2nd dofs, then the 1st node residual force is defined as
   
nbj1 ¼ bj1  þ bj2 . (7)
Those nodes having larger values in nbj are the most suspected damaged nodes and the damaged elements
can be determined according to the relation between the element number and the node number.

3.1. The algebraic solution of the residual force equation

While the suspected damaged elements are determined by using the above described localization algorithm,
a simple approach to obtain damage extents termed the algebraic solution is by directly solving the residual
force equation. The perturbed global stiffness matrix DK can be expressed as
X
N
DK ¼ al K l , (8)
l¼1

where al (alA[0, 1]) and Kl are the lth elemental damage parameter and stiffness matrix, respectively, N is the
number of elements. The value of al is 0 if the lth element is undamaged and al is 1 if the lth element is
completely damaged. From Eq. (8) we can obtain
X
N
Dki ¼ al kli , (9)
l¼1

where kli is the ith stiffness vector in the lth element stiffness matrix. Both single damage and multiple damages
are discussed in the following derivation.
For the case of single damage occurring in the structure, without loss of generality, assume that the qth
element including the ith dof is damaged and the damage extent is aq. Eq. (9) reduces to
Dki ¼ aq kqi . (10)
From Eq. (5) it is obvious that
DkTi fdj ¼ bji . (11)
Substituting Eq. (10) into Eq. (11) the damage extent can be obtained as
bji
aq ¼ q T . (12)
ðki Þ fdj
For the case of multiple damages occurring in the structure, assume that the number of the damaged
elements is r and the damage extents are a1, a2, y, ar, respectively, then Eq. (9) becomes
X
r
Dki ¼ al kli . (13)
l¼1
中国科技论文在线 http://www.paper.edu.cn
ARTICLE IN PRESS
Q.W. Yang, J.K. Liu / Journal of Sound and Vibration 305 (2007) 298–307 301

Correspondingly, r equations associated with the damaged dof are selected from Eq. (5) to calculate the
damage extent, i.e.,
2 T3 8 9
Dk1 >
> bj1 >
>
6 T7 >
> >
>
6 Dk2 7 < bj2 >
> =
6 7
6 . 7fdj ¼ . . (14)
6 .. 7
4 5 > .. >
>
> >
>
>
> >
DkT : bjr >
;
r

Substituting Eq. (13) into Eq. (14) yields

2P 3
r T 8 9
al kl1 fdj
6 l¼1 7 > > bj1 >
>
6 Pr 7 > >b > >
6 l¼1 al k2 fdj 7 >
l T
< j2 >=
6 7
6 7 ¼ . . (15)
6 .. 7 > .. >
6 . 7 > > >
>
4 Pr 5 > : bjr >
> >
;
lT
l¼1 a l k r fdj

Eq. (15) can be rewritten as

2 T T 3 8 9
k11 fdj k21 fdj  kr1 T fdj 8 a1 9 > bj1 >
6 T 7>
> > >
> > > >
6 k1 f T
kr2 T fdj 7> >
> > > > > >
6 2 dj k22 fdj  7< a2 = < bj2 =
6 7 . ¼ .. . (16)
6 . .. .. .. 7 > .. > >
6 .. . . . 5>7>
> >
> >
> . >>
>
4
T T
: > ; > >
:
>
>
;
k1r fdj k2r fdj  rT
kr fdj a r b jr

The damage extents can be easily obtained by solving the linear Eq. (16) using the Gauss elimination method.

3.2. The natural frequency sensitivity method

The algebraic solution of the residual force equation needs less computation efforts than that of the
minimum norm solution and the minimum rank solution. However, as described in the introduction, the
algebraic solution also depends on the accuracy of the residual force vector; in other words, mode shape noises
have quite large effects on its damage detection result. Those damage quantification methods using mode
shape information are all inevitably affected by the large measurement noises of mode shapes. So a better
choice to obtain the damage extents is only using the natural frequency information because the lower natural
frequencies can be measured very precisely. In this paper, the natural frequency sensitivity method [6,7] is also
employed to identify damages after damages are located using the node residual force vector. Assuming that
the number of measured natural frequencies is m, the damage quantification formulation from measured
modal frequencies with the first-order approximation is as follows:
8 9 8 9
> Dl1 > > a1 >
> > > >
> Dl2 >
>
< >
= > a2 >
>
< >
=
.. ¼ Sf .. , (17)
>
>
> . >>
>
>
>
> . >
>
>
>
: >
; : >
> ;
Dlm ar
where Sf denotes the first-order sensitivity matrix of natural frequencies and Dlj (j ¼ 1m) is the change of the
natural frequency before and after damage. From Eq. (17), the damage extents can be obtained as
8 9 8 9
> a1 > > Dl1 >
>
> > >
> >
> >
>
>
< a2 = < Dl2 >
> =
þ
.. ¼ Sf .. . (18)
>
>
> . >
>
>
>
>
> . >
>
>
> ;
: > >
: >
;
ar Dlm
中国科技论文在线 http://www.paper.edu.cn
ARTICLE IN PRESS
302 Q.W. Yang, J.K. Liu / Journal of Sound and Vibration 305 (2007) 298–307

When the damage is not small, a second-order approximation on the damage [8] can be performed or an
iteration scheme [9] can be used to estimate damage extents more precisely, which are not further discussed in
this paper.

4. Mode shape expansion

Lim and Kashangaki [10] used the best achievable technique to locate damages. This paper makes use of it
to expand the incomplete measured mode shape. From Eq. (3), one obtains
ðK u  ldj MÞ1 DKfdj ¼ fdj . (19)
Substituting Eq. (8) into Eq. (19) yields
Ljl gjl ¼ fdj , (20)
where
Ljl ¼ ðK u  ldj MÞ1 K l ; l ¼ 1  N, (21)

gjl ¼ al fdj ; l ¼ 1  N. (22)

Eq. (20) shows that the mode shape of the damaged structure is the linear combination of the nonzero
columns of the matrix Ljl associated with the damaged elements. So Eq. (20) can be used to expand mode
shapes after the probably damaged elements are identified preliminary. The elements used to mode shape
expansion can be down selected by the best achievable technique. Eq. (20) can be partitioned into measured
dofs, represented by m, and unmeasured dofs, represented by s, to obtain
" m# ( m)
Ljl fdj
s gjl ¼ . (23)
Ljl fsdj

The best achievable eigenvector for the measured dofs fma

dj can be obtained as

fma m m þ m
dj ¼ Ljl ðLjl Þ fdj , (24)
where Lmis the matrix
jl Lm
jlwhere the zero columns have been removed to enhance computational efficiency,
and the superscript ‘+’ indicates the pseudo inverse of a matrix. The distance between the vector fm ma
dj and fdj
can be computed using the Frobenious norm
 
 ma 
d jl ¼ fm
dj  fdj  , (25)
F

where kkF represents the Frobenious norm. If the number of measured modes is m, an N  m matrix of d’s can
be constructed as
2 3
d 11 d 21    d m1
6 d 12 d 22    d m2 7
6 7
D¼6 6 .. .. .. .. 7
7. (26)
4 . . . . 5
d 1N d 2N    d mN
Searching for values that are considerably smaller than others in the matrix D and the columns of their
corresponding matrices, Ljl are used as the basis vectors to expand the mode shape. Without loss of generality,
assuming that the matrices associated with the smaller elements in D are Lj1 ; Lj2 ; . . . ; Lje , the matrix used to
expand the jth mode shape can be obtained as
h i
E j ¼ Lj1 Lj2    Lje . (27)

The mode shape expansion formulation is as follows:

fsdj ¼ E sj ðE m þ m
j Þ fdj . (28)
中国科技论文在线 http://www.paper.edu.cn
ARTICLE IN PRESS
Q.W. Yang, J.K. Liu / Journal of Sound and Vibration 305 (2007) 298–307 303

5. Numerical example

A plane truss structure (shown in Fig. 1) is taken as an example to verify the proposed methods. The basic
parameters of the structure are as follows: E ¼ 200 GPa, r ¼ 7.8  103 kg/m3, L ¼ 1 m and A ¼ 0.004 m2. The
relation between the element number and the node number is listed in Table 1. Assume that only the 1st, 2nd,
3rd, 4th and 5th nodes are measured locations and the other nodes are unmeasured sites. Natural frequencies
and mode shapes are contaminated with 0.05% and 5% random noises, respectively. Three damage cases are
studied in this work. The first damage case assumes a single damage occurs in the 10th element with a stiffness
loss of 10%. The second damage case assumes element 16 is damaged with a stiffness loss of 10%. The third
damage case assumes that elements 9 and 10 are damaged at the same time with stiffness losses both of 10%.
The ratios of frequency changes before and after damage for these three damage cases are shown in Table 2.
For the first damage case, mode 4 is used to calculate the node residual force vector because it undergoes a
significant shift in frequency according to Table 2. The node residual force vector of mode 4 is shown in Fig. 2.
Nodes 2 and 3 are the most suspected damaged nodes from Fig. 2 since their residual force values are much
larger than others. According to Table 1, element 10 is the corresponding element associated with nodes 2 and
3. By using Eq. (12), the algebraic solution of the damage extent can be obtained as a10 ¼ 0.1035 (3.5%). The
value in bracket denotes the comparative error between the calculated value and the assumed value. The
minimum rank solution by the MREU technique is shown in Fig. 3 and the damage extent of element 10 is
a10 ¼ 0.0743 (25.7%). Using the natural frequency sensitivity method, damage extent of element 10 can be
obtained as a10 ¼ 0.1025 (2.5%). According to the above results of the three damage quantification methods,
conclusion can be easily drawn that the natural frequency sensitivity method has the best precision and the
MREU technique has the worst.
For the second damage case, mode 1 is used to obtain the node residual force vector since it undergoes a
significant shift in frequency from Table 2. The node residual force vector is shown in Fig. 4. Nodes 5, 7 and 8
are most probably the damaged nodes for their node residual force values are considerably larger than others.
According to Table 1, elements 16 and 19 are the suspected damaged elements. The algebraic solution can be
calculated using Eq. (17) to be a16 ¼ 0.0903 (9.7%) and a19 ¼ 0.0494. The result shows that the algebraic
solution is fallible because element 19 is falsely identified. Results obtained by MREU using mode 1 and using
modes 1 and 2 are shown in Figs. 5 and 6. According to Figs. 5 and 6, the better result is obtained when only
mode 1 is used because the number of modes equals the number of nonzero perturbations in this case. Damage
extent of element 16 in Fig. 5 is a16 ¼ 0.0823 (17.7%). Using the natural frequency sensitivity method, one
obtains a16 ¼ 0.1076 (7.6%) and a19 ¼ 0.0032. Element 16 is identified from the damage quantification results
to be the genuine damaged element and element 19 is excluded. From the above results, one can see that only
the natural frequency sensitivity method is the most feasible and has the best accuracy for damage
identification.
In the third damage case, mode 4 is used to calculate the node residual force vector and the result is shown
in Fig. 7. Nodes 2, 3 and 9 are damaged according to Fig. 7 and their corresponding elements are 9, 10 and 11.
The algebraic solutions from Eq. (17) are a9 ¼ 0.0965 (3.5%), a10 ¼ 0.1054 (5.4%) and a11 ¼ 0.0409. Again,
element 11 is falsely identified. Results by MREU using modes 4 and 5 and using modes 4, 5 and 6 are shown
in Figs. 8 and 9. It is obvious that the MREU technique is valid only when the number of mode shapes used to
calculate equals the rank of the perturbed matrix. This is a potential drawback to any minimum rank optimal
matrix update technique, because in practice the expected order of the damage will typically be unknown.

11 10 9 8 7 6
4 8 12 16 20

1 3 5 7 9 11 13 15 17 19 21 23

60˚ 2 6 10 14 18 22
L L L L L L
1 2 3 4 5

Fig. 1. A plane truss structure.

中国科技论文在线 http://www.paper.edu.cn
ARTICLE IN PRESS
304 Q.W. Yang, J.K. Liu / Journal of Sound and Vibration 305 (2007) 298–307

Table 1
Nodes that are related to each element in the truss

Element number Corresponding nodes

1 11
2 1
3 1, 11
4 10, 11
5 1, 10
6 1, 2
7 2, 10
8 9, 10
9 2, 9
10 2, 3
11 3, 9
12 8, 9
13 3, 8
14 3, 4
15 4, 8
16 7, 8
17 4, 7
18 4, 5
19 5, 7
20 6, 7
21 5, 6
22 5
23 6

Table 2
The ratios of frequency changes before and after damage

Mode Damage in element 10 (%) Damage in element 16 (%) Damages in element 9 and 10 (%)

1 0.17 1.67 0.34

2 0.32 0.42 0.69
3 0.01 0.60 0.02
4 0.73 0.06 1.72
5 0.52 0.25 0.54
6 0.55 0.76 0.58

600
Node residual force

500
400
300
200
100
0
1 2 3 4 5 6 7 8 9 10 11
Node number

Fig. 2. The node residual force vector for mode 4 when element 10 is damaged.

Damage extents of elements 9 and 10 from Fig. 8 are a9 ¼ 0.079 (21%) and a10 ¼ 0.0735 (26.5%), which have
large errors. Results by using the natural frequency sensitivity method are a9 ¼ 0.1154 (15.4%), a10 ¼ 0.0952
(4.8%) and a11 ¼ 0.0091. Elements 9 and 10 are identified from the damage quantification results to be the
中国科技论文在线 http://www.paper.edu.cn
ARTICLE IN PRESS
Q.W. Yang, J.K. Liu / Journal of Sound and Vibration 305 (2007) 298–307 305

0.08
0.07

Damage extent
0.06
0.05
0.04
0.03
0.02
0.01
0
1 4 7 10 13 16 19 22
Element number

Fig. 3. Result of MREU using mode 4 when element 10 is damaged.

180
160
Node residual force

140
120
100
80
60
40
20
0
1 2 3 4 5 6 7 8 9 10 11
Node number

Fig. 4. The node residual force vector for mode 1 when element 16 is damaged.

0.09
0.08
0.07
Damage extent

0.06
0.05
0.04
0.03
0.02
0.01
0
1 4 7 10 13 16 19 22
Element number

Fig. 5. Result of MREU using mode 1 when element 16 is damaged.

0.1
0.08
Damage extent

0.06
0.04
0.02
0
-0.02
-0.04
-0.06 1 4 7 10 13 16 19 22
Element number

Fig. 6. Result of MREU using mode 1 and 2 when element 16 is damaged.

中国科技论文在线 http://www.paper.edu.cn
ARTICLE IN PRESS
306 Q.W. Yang, J.K. Liu / Journal of Sound and Vibration 305 (2007) 298–307

900

Node residual force

800
700
600
500
400
300
200
100
0
1 2 3 4 5 6 7 8 9 10 11
Node number

Fig. 7. The node residual force vector for mode 4 when element 9 and 10 are damaged.

0.09
0.08
Damage extent

0.07
0.06
0.05
0.04
0.03
0.02
0.01
0
1 4 7 10 13 16 19 22
Element number

Fig. 8. Result of MREU using mode 4 and 5 when element 9 and 10 are damaged.

0.3

0.2
Damage extent

0.1

-0.1

-0.2

-0.3
1 4 7 10 13 16 19 22
Element number

Fig. 9. Result of MREU using mode 4, 5 and 6 when element 9 and 10 are damaged.

genuine damaged elements and element 11 is excluded. On condition that the suspected damaged elements are
determined preliminarily by the node residual force vector, the above results show that the natural frequency
sensitivity method is the most feasible technique in engineering practice.

6. Conclusion

Structural damage identification methods based on the residual force vector are studied in this paper. The
node residual force vector is defined according to the residual force vector to localize damages preliminarily. In
order to solve the incomplete measurement problem, a mode shape expansion algorithm based on the best
achievable technique is presented. After the suspected damaged elements are determined using the node
中国科技论文在线 http://www.paper.edu.cn
ARTICLE IN PRESS
Q.W. Yang, J.K. Liu / Journal of Sound and Vibration 305 (2007) 298–307 307

residual force vector, three different damage quantification methods are used to identify damages more
precisely. A plane truss structure is analyzed as a numerical example to demonstrate these damage detection
methods with measurement noises.
Results of numerical example show that the node residual force vector can preliminarily locate the probable
damages in measured locations or unmeasured locations, but the absolutely accurate damage localization is
impossible for the measured mode shapes have large measurement noises.
The algebraic solution of the residual force vector is the simplest damage quantification technique among
these damage detection methods. This advantage is especially obvious when the structural damages are
separate. But this method often identifies false damages when the measurements have noises.
The MREU technique can identify structural damages only when the number of mode shapes used in the
calculation equals the rank of the perturbed matrix. Since in practice the expected rank of the perturbed
matrix is unknown, the MREU technique is limited to be used in actual engineering practice. Even if this
precondition is satisfied, the damage extents obtained by MREU often have large errors with measurement
noises.
On condition that the suspected damaged elements are determined preliminary by the node residual force
vector, the natural frequency sensitivity method is the most practical technique because it can not only obtain
damage extents of the genuine damaged elements accurately, but also exclude the false damaged elements.
This method is robust because the lower natural frequencies of structures can be measured very precisely. The
drawback of this method is a second-order approximation or an iteration scheme must be performed with
increasing damages, which adds to the computation efforts. This coupled damage detection method would
have broad foreground in the actual engineering practices.

Acknowledgments

This work is supported in part by the TRAPOYT, NSFC, GDSF and the Foundation of Zhongshan
University Advanced Research Center.

References

[1] J.-C. Chen, J.A. Garba, On-orbit damage assessment for large space structures, AIAA Journal 26 (9) (1988) 1119–1126.
[2] D.C. Zimmerman, M. Kaouk, Structural damage detection using a minimum rank update theory, Journal of Vibration and Acoustics
116 (2) (1994) 222–231.
[3] S.W. Doebling, Minimum-rank optimal update of elemental stiffness parameters for structural damage identification, AIAA Journal
34 (12) (1996) 2615–2621.
[4] D.-Y. Chiang, W.-Y. Lai, Structural damage detection using the simulated evolution method, AIAA Journal 37 (10) (1999)
1331–1333.
[5] C. Mares, C. Surace, An application of genetic algorithms to identify damage in elastic structures, Journal of Sound and Vibration 195
(2) (1996) 195–215.
[6] R.M. Lin, H. Du, J.H. Ong, Sensitivity based method for structural dynamic model improvement, Computer & Structures 47 (3)
(1993) 349–369.
[7] A. Messina, A.I. Jones, J.E. Williams, Damage detection and localization using natural frequency changes, in: Proceedings of
Conference on Identification in Engineering Systems, Swansea, UK (1996) 67–76.
[8] A. Messina, J. E Williams, T. Contursi, Structural damage detection by a sensitivity and statistical-based method, Journal of Sound
and Vibration 216 (5) (1996) 791–808.
[9] Z.Y. Shi, S.S. Law, L.M. Zhang, Structural damage detection from modal strain energy change, Journal of Engineering Mechanics,
ASCE 126 (12) (2000) 1216–1223.
[10] T.W. Lim, T.A.L. Kashangaki, Structural damage detection of space truss structures using best achievable eigenvectors, AIAA
Journal 32 (5) (1994) 1049–1057.