Sie sind auf Seite 1von 25

Mechanical Systems and Signal Processing 52-53 (2015) 181205

Contents lists available at ScienceDirect

Mechanical Systems and Signal Processing


journal homepage: www.elsevier.com/locate/ymssp

Damage identification techniques via modal curvature


analysis: Overview and comparison
Daniele Dessi n, Gabriele Camerlengo
INSEAN-CNR, Marine Technology Research Institute, National Research Council, Via di Vallerano 139, 00128 Rome, Italy

a r t i c l e i n f o

abstract

Article history:
Received 2 February 2014
Received in revised form
12 May 2014
Accepted 23 May 2014
Available online 8 July 2014

This paper aims to compare several damage identification methods based on the analysis of
modal curvature and related quantities (natural frequencies and modal strain energy) by
evaluating their performances on the same test case, a damaged EulerBernoulli beam.
Damage is modelled as a localized and uniform reduction of stiffness so that closed-form
expressions of the mode-shape curvatures can be analytically computed and data accuracy,
which affects final results, can be controlled. The selected techniques belong to two
categories: one includes several methods that need reference data for detecting structural
modifications due to damage, the second group, including the modified Laplacian operator
and the fractal dimension, avoids the knowledge of the undamaged behavior for issuing a
damage diagnosis. To explain better the different performances of the methods, the
mathematical formulation has been revised in some cases so as to fit into a common
framework where the underlying hypotheses are clearly stated. Because the various damage
indexes are calculated on exact data, a sensitivity analysis has been carried out with
respect to the number of points where curvature information is available, to the position of
damage between adjacent points, to the modes involved in the index computation. In this
way, this analysis intends to point out comparatively the capability of locating and
estimating damage of each method along with some critical issues already present with
noiseless data.
& 2014 Elsevier Ltd. All rights reserved.

Keywords:
Damage detection
Mode-shape curvature
Modal strain energy
Structural health monitoring

1. Introduction
Damage identification has been one of the most challenging topics in theoretical and experimental structural mechanics
for several decades. This topic has addressed the following different approaches strictly dependent on the kind of expected
damage and on the information supposed to be available. Thus, techniques developed for finding cracks in a fuselage panel
after disassembling an airframe substantially differ from those focusing on data acquired during in-service monitoring of a
containership. In the latter case, the failure of a single structural element may be not so critical as for airplanes and
represents the lowest damage level one is interested to detect.
The analysis of the state-space of the dynamical response provides the shared background for different types of methods
applicable to structures vibrating under operational or ambient loads. In the following we intend to focus on techniques
based on the determination of the response invariants (e.g. basis functions) because, if system linearity is assumed, they can
benefit of operational modal analysis for the identification of vibration modes. The so-called ambient excitation, in a less
n

Corresponding author. Tel.: 39 0650299254; fax: 39 065070619.


E-mail address: daniele.dessi@cnr.it (D. Dessi).

http://dx.doi.org/10.1016/j.ymssp.2014.05.031
0888-3270/& 2014 Elsevier Ltd. All rights reserved.

182

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

rigorous interpretation, is characterized by an almost flat power spectral density in the frequency range of the concerned
vibration modes. If one successfully characterizes the load as ambient excitation, there is no need of measuring the input
force. Stochastic gust for airplanes, irregular sea for ships, random traffic loading for a bridge provide some examples (e.g.
see [1,2]). Therefore, techniques based on the detection of vibration modes in the damaged condition still deserve attention.
Among many reviews of the existing literature concerning damage identification methods, Doebling et al. [3] and Fan
and Qiao [4] deal in particular with the subset of methods related to variations in basic modal properties. Changes or lack of
smoothness in mode shape or mode shape curvature, analysis of dynamically measured flexibility, updating of structural
model parameters provide different examples of this family. In the present paper the focus is on techniques processing
information about mode shape curvature or strain modes with or without the knowledge of baseline data. Though the first
attempts in detecting structural damage through changes in mode frequency date back to late seventies [5], the main efforts
in using also information from mode shapes took place since nineties. Stubbs and Osegueda [6] pointed out the existence of
a clear relation between damage and the subsequent variation of modal parameters for an EulerBernoulli beam model. This
principle of comparing modal properties before and after damage has set the path for several following investigations.
Successively, Stubbs et al. [7,8] defined an algorithm based on the evaluation of modal curvature integrals and applied their
method to a real structure. In the same period, also Pandey et al. [9] had considered the variation of the curvature of the
mode shapes as an indicator of the presence of damage. Later on, following Pandey et al. [9], Wahab and De Roeck [10]
defined and applied a curvature damage factor to real data on a prestressed concrete bridge. In the same years, Cornwell
et al. [11,12] presented a damage index based on the strain energy for 1D and 2D structures. Kim and Stubbs [13] and Kim
et al. [14] considered and improved the index proposed by Stubbs et al. [8] defining a mode shape based detection method
(MBDD). Following a slightly different way, Choi et al. [15] developed a damage index (denoted as compliance index)
assuming the invariance of the bending moment distribution after the occurrence of damage (projection over modal basis is
again introduced). The work of Galvanetto and Violaris [16] deserves a particular mention because, though a quantitative
damage index is not actually defined, they used the proper orthogonal modes instead of vibration mode shapes, implicitly
assuming possible nonlinearity of the response. Hu and Wu [17] applied the index developed by Cornwell et al. [12] to
aluminum plates with surface cracks, introducing the idea of a local interrogation window moving over the surface. After
reviewing the existing methods Radzien ski et al. [18] proposed new algorithms based on processing and combining
information from frequency variations and modal strains (FDI and HDD methods). Fan and Qiao [19] introduced an
identification method based on a damage location factor (DLF) and a damage severity correction factor (DSCF) through a
rigorous mathematical theory, clearly explicating hypotheses upon which simplifications were made. In last decade, Wu and
Law [20], Catbas et al. [21] and more recently Sung et al. [22] worked on approaches based on the application of the uniform
load surface theory (ULS, see [23]) and its extensions to curvatures. Though this family of methods originates from the
analysis of changes in the flexibility matrix (see the paper of Pandey and Biswas [24]), the normalized uniform load surface
method (NULS) proposed by Sung et al. [22] has reached the capability of estimating the damage severity in multiple
locations. All the methods mentioned above require baseline data, i.e., the evaluation and the comparison of the modal
properties (or some functions depending on them) in the intact and damaged conditions. A different approach is based on
assuming a model of the damaged structure which depends in a straightforward way on the damage intensity. For instance,
Lestari et al. [25] used a perturbation solution for a beam with a uniform reduction of stiffness, transforming the damage
detection into a system identification problem. In a rather different perspective, avoiding any too specific hypothesis on the
damage model, the lack of smoothness in the response and then in the basis functions can be used as an indicator of
damage. Ratcliffe [26] presented an index (modified Laplace operator, MLO) based on the comparison between the modal
curvatures of the damaged mode shapes affected by irregularities and a third order polynomial representing the undamaged
local curvature (recently, Qiao et al. [27] proposed a fourth order polynomial). This concept, first named as (resonant)
gapped smoothing method (GSM), was applied to the detection of delamination in a composite beam [28] and, some years
later, inspired some extensions to cover broadband curvature data from the analysis of frequency response functions [29,30]
and also to deal with 2D structures [30]. Cao and Qiao [31] extended the idea of Ratcliffe [26] introducing a novel Laplacian
scheme. The proposed algorithm for damage detection defines an array of Laplace operators for computing multiresolution
modal curvatures, improves the robustness to noise with a Gaussian filter and introduces a nonlinear damage operator for
enhancing the damage footprint. This approach has been further expanded recently (see [32,33]) to improve the
performance on noisy data by reshaping the algorithm so that the use of wavelets assumes a fundamental role. A
completely different approach has been proposed by Hadjileontiadisa et al. [34] and Li et al. [35]. They exploited a technique
developed in fractal geometry based on the definition of fractal dimension (FD) for estimating local irregularities in
waveforms. Wang and Qiao [36] defined a generalized fractal dimension (GFD). It is worth noting that [27] made a
comparison among GSM, ULS, GFD and a non-baseline strain energy method (SEM).
Numerical and experimental applications which show the effectiveness of the developed techniques are provided in
almost all papers. However, it is difficult for the reader to make a direct comparison of these techniques. The first reason lies
in using different test cases intended as a combination of observed structure, occurred damage and measured response. The
second reason concerns the presence of numerical and experimental uncertainties which vary depending on the test-case.
In the present paper, we try to overcome these limitations so that a clearer comparison of the damage indexes becomes
feasible. The common test case is provided by a simply supported EulerBernoulli beam with uniform cross sections. The
techniques selected for comparison are those developed by Stubbs et al. [8], Cornwell et al. [12], Kim and Stubbs [13], Choi
et al. [15] and Fan and Qiao [19] among the methods that require baseline data, and Ratcliffe [26], Hadjileontiadisa et al. [34],

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

183

Li et al. [35] and Wang and Qiao [36] among the methods based on the analysis of the damaged behavior alone. All these
techniques are representative of different approaches which process eigenfrequencies and mode shape curvatures, directly
or in the form of modal strain energy, without any sophisticated post-processing to overcome the presence of noise. In fact,
in the present work, the mode shapes are provided analytically also in the presence of damage, which is modelled as a
uniform stiffness reduction, to avoid the presence of errors in input data. As a consequence, it has been possible to achieve
an analytical representation of mode-shape curvature: sampling of these functions has provided the shared input data for all
methods. A plain description of the selected techniques has afforded the possibility of comparing them theoretically
especially regarding their basic assumptions. A particular effort in this direction has been done for techniques linked to the
concept of modal strain energy where a unique derivation has been introduced. The index performances in terms of
identifying damage location and of estimating the given damage have been compared in the paper. In particular, a
sensitivity analysis with respect to the number of nodes where curvature is evaluated, to the number of considered vibration
modes, to the damage level and extension has been carried out.
2. Analytical modes of a notched beam
As mentioned before, the input data for the calculation of damage indexes is provided by the analytical expression of the
mode shapes and their curvatures. The theory is based on a generalization of the analysis presented by Naguleswaran [37]
for stepped EulerBernoulli beams. Mainly oriented to the evaluation of the response but still worth to be mentioned here
are also the papers of Koplowa et al. [38] and Jaworski and Dowell [39]. The main steps for achieving the final expressions
are briefly recalled in the following.
For the sake of simplicity, the simply supported planar beam has negligible shear deformability (EulerBernoulli beam).
As shown in Fig. 1, damage is modelled as a localized and uniform reduction of the bending stiffness distribution D(x) along
the dimensional coordinate x. Thus, the damaged beam of length L is considered as the union of three beam portions as
shown in Fig. 2, where the shortest one, with a reduced cross-section with respect to the others, represents the damaged
portion. For each beam portion, Lk, Ak and Dk denote the length, cross-section area and bending stiffness, respectively. For
the sake of generality, it is convenient to formulate identification methods using non-dimensional coordinates ; , with
x=L and y=L, and non-dimensional variables. Three coordinate systems O1 ; 1 ; 1 , O2 ; 2 ; 2 and O3 ; 3 ; 3 are
defined for the three beam portions so as to write easily the corresponding equation of motion (note that the first coordinate
system O1 ; 1 ; 1 coincides with the global reference system O; ; ). For each beam portion of non-dimensional length
k Lk =L k 1; 2; 3, conditions at the edges are defined in terms of the boundary conditions whereas at the interfaces
between the beam portions, continuity of displacement w, slope , bending moment M and shear force V is imposed (all
non-dimensional). The generic expression of mode-shapes in non-dimensional form is
k k C 1
sin k k C k2 cos k k C k3 sinh k k C k4 cosh k k
k

k 1; 2; 3

Fig. 1. Simply supported beam with the damaged portion in grey (dimensional quantities).

A2 , D2
1

A1, D1

O1

O2
2

A3, D3

O3

Fig. 2. Geometrical non-dimensional parameters of the notched beam.

184

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

where C kr and k are, respectively, the coefficients and frequencies to be determined in the following. The beam boundary
conditions in terms of non-dimensional mode-shapes are expressed as
(
1 0 0
2
applied in O1
1 0 0
(

3 3 0

applied in E

3 3 0

whereas at the interfaces the continuity of displacement, slope, bending moment and shear force is applied:
8
1 1 2 0
>
>
>
> 0
<
1 1 02 0
applied in O2
1 1 D21 2 0
>
>
>
>

:
1 1 D21 2 0
8
2 2 3 0
>
>
>
> 0
<
2 2 03 0
2 2 D32 3 0
>
>
>
>
: D32 0
2

applied in O3

with D21 D2 =D1 and D32 D3 =D2 the bending stiffness ratios. It is worth underlying that the sudden variation of stiffness
implies that curvatures lose continuity as well. After substituting the generic expression of mode shapes (Eq. 1) into the set
of 12 conditions (Eqs. (2)(5)), the 12 unknown constants C r
(k 1; ; 3; and r 1; ; 4) and the natural frequencies can be
k
calculated so that the expressions of mode shapes are finally determined.1 The final expressions for both the mode shapes
and the mode-shape curvatures are quite complicate but can be easily retrieved with symbolic algebra programs. The
relative frequency variation is defined as
r i

in  i
;
i

with in the natural frequency of the damaged system. The previous relation can be expressed in terms of non-dimensional
frequency i
(the dependence on k is artificial) as
k
2

n
i
 i
k
k
2

r i

i
k

1
, with a
In Fig. 3, r i is plotted for different values of the stiffness reduction, i.e., for different values of Dn =D D21 D32
damaged length 2 0:1 placed between 0.6 and 0.7 of the beam length. It appears that the trends are different for all the
vibration modes: the modes are most affected when stiffness discontinuity occurs close to peaks in the mode shape curves.
The dependence of r i on Dn =D is nonlinear also if it can be locally linearized, e.g. for small values of the stiffness
reduction. On the opposite, differences on the mode shapes are hardly distinguishable. For this reason, in Fig. 4 the
difference between the damaged mode (30% stiffness reduction) and the undamaged mode is plotted by multiplying this
quantity by 10 (dashed curve). As shown in Fig. 4 for the first bending mode, differences are detectable only as a thickening
of the curve superposition. Differences appear much more relevant if the mode shape curvature is considered as shown in
Figs. 57. The discontinuity in the bending stiffness is reflected sharply on the curvature of the mode shapes as a result of
Eqs. (4) and (5).

3. Damage estimation based on mode curvature analysis with a baseline


3.1. Basic concepts
This section gives a presentation of several damage indexes related to the analysis of curvature and frequencies of
vibration modes. The majority of these methods are based on comparing modal strain energy in the damaged and
undamaged conditions. Though these techniques have to share a common theoretical background, they were originally
introduced in different ways sometimes preventing an easy comparison. The objective of providing a common theoretical
background for all the methods is pursued addressing first the method proposed by Fan and Qiao [19] from which the other
methods can be derived under suitable hypotheses with few exceptions.
1
To obtain these coefficients, one has to enforce that the determinant of the homogeneous system, depending on the non-dimensional frequencies 1 ,
2 and 3 , is set to zero. It is worth noting that the resulting equation is dependent only on one variable, e.g. 41 A1 2 L4 =D1 , because 2 and 3 are linearly
related to 1 through the beam parameters ( is the material density and denotes the dimensional frequency expressed in rad/s).

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

185

Relative frequency variation

0.005
0
-0.005
st
1 Mode
nd
2 Mode
3 rd Mode
4 th Mode
th
5 Mode

-0.01
-0.015
-0.02
0.65

0.7

0.75

0.8

0.85

0.9

0.95

1.05

D* / D

1-st mode shape

Fig. 3. Relative variation of the natural frequencies with respect to reduced stiffness ratio in the damaged portion.

1
*
D / D = 0.9
*
D / D = 0.8
*
D / D = 0.7
Difference (values x 10)

0.5

0
0

0.2

0.4

0.6

0.8

Fig. 4. First mode shape for several stiffness reductions in the damaged beam section. The dashed curve representing the difference between the intact and
the damaged modes in the worst case has been multiplied by 10.

1-st mode shape cruvature

D */ D = 0.9
D */ D = 0.8
D */ D = 0.7

0
-5
-10
-15
-20
-25
-0.2

0.2

0.4

0.6

0.8

1.2

2-nd mode shape cruvature

Fig. 5. First mode shape curvature for several values of the stiffness reduction in the damaged portion.

D */ D = 0.9
D */ D = 0.8
*
D / D = 0.7

50

-50
0

0.2

0.4

0.6

0.8

Fig. 6. Second mode shape curvature for several values of the stiffness reduction in the damaged portion.

186

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

3-rd mode shape cruvature

D / D = 0.9
*
D / D = 0.8
D */ D = 0.7

100
50
0
-50
-100
0

0.2

0.4

0.6

0.8

Fig. 7. Third mode shape curvature for several values of the stiffness reduction in the damaged portion.

Fig. 8. Definition of nodes and elements along the beam.

It is worth recalling that these methods are based on or ultimately connected to the evaluation of strain energy, which is
R
defined for a solid as U 12 V r: dV, with r and the stress and strain tensors, respectively. Relative to slender structures like
beams, e.g. described with EulerBernoulli theory, the previous definition can be simplified and expressed in terms of the
beam curvature. For the successive developments, it is useful to decompose the strain energy with respect to the
contribution of vibration modes, which leads to define the modal strain energy associated with the i-th mode:
!2
Z
2
1 L
d i x
Dx
dx;
8
U i
2
2 0
dx
where the mode shapes in the dimensional problem are still indicated with i x for the sake of simplicity. The previous
equation can be specified for a single portion of the beam between the coordinates xj and xj 1 (xj j  1h with h the
constant element length, see Fig. 8):
!2
Z
2
1 xj 1
d i x
i
Dx
dx;
9
Uj
2
2 xj
dx
A common assumption in many investigations is supposing that stiffness variations are negligible inside each beam portion
(element). Thus, if Dj is the stiffness inside the element, it follows
!2
Z
Dj xj 1 d2 i x
i
dx Dj i
10
Uj
j ;
2
2 xj
dx
where i
j , the integral of the square of the modal curvature, represents a pure geometrical quantity. In the following, the
physical quantities when referred to the damaged structure will be indicated with n (e.g., the modal strain energy after
damage will be U in ).
It is also convenient to express the strain energy as commonly done in matrix structural analysis. The main reason for
this choice lies in simplifying the notation. The second reason for that is linked to the fact that this choice was made also by
Fan and Qiao who set
U i 12 ziT Kzi ;

11
i

where the discrete vibration mode z

(corresponding to xj ) is the eigenvectors of the generalized eigenvalue problem:

K i Mz 0

12

with K and M the stiffness and the mass matrix, respectively, and i the system eigenvalues, equal to the square of natural
2
frequencies i . Assuming that damage determines a perturbation (reduction) of the stiffness matrix, the perturbed
eigenvalue problem is
K K  i i Mzi zi 0:

13

It is worth recalling the following result for the perturbed eigenproblem:


i ziT Kzi ;

14

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

zi

zjT Kzi
;
i j
j 1;j a i
N

187

15

that shows how the relevance of eigenvector variation is related to the closeness of structural frequencies (modal density of
the intact structure).
3.2. Fan and Qiao
The method introduced by Fan and Qiao [19] is based essentially on combining the perturbed eigenvalue problem with
the variation of strain energy in presence of damage. This paper follows a previous attempt of relating the strain energy
change with stiffness reduction (using Eq. (14)) due to Shi et al. [40]. Pre-multiplying Eq. (13) by ziT , developing products
and recalling that ziT K i M 0 (see Fan and Qiao for further details) yields
ziT Kzi i ziT Mzi ziT Kzi  i ziT Mzi 0:

16

Using Eq. (12), it is possible to avoid any reference to mass matrix in the previous equation:
ziT Kzi 

i iT i

z Kz ziT Kzi  i ziT Kzi 0


i
i

17

This scalar equation contains terms related to the strain energy of the undamaged and damaged structures. The strain
energy for the i-th mode of the damaged structure is




18
U in 12 ziT ziT K K zi zi ;
which gives, expanding products and disregarding higher-order terms,
U in C 12 ziT Kzi ziT Kzi 12 ziT Kzi ziT Kzi ;

19

and similarly for j-th beam element,


n
1 iT
Kj zi ziT Kj zi 12 ziT Kj zi ziT Kj zi :
U i
j C2 z

20

To proceed further, it is necessary to make some assumptions about damage. Within the hypothesis of superposition of
damage effects, the global variation of the stiffness matrix of the structure K is assumed to be given by the sum of the
stiffness variations on each element j, i.e., K N
j 1 Kj . Moreover, the local damage is set to be linearly dependent on the
variation of the bending stiffness relative to the j-th element, that is Kj j Kj , with j Dnj  Dj =Dj . Thus, according to
Eq. (10),
n
n
n
Dnj i
1 j Dj i
U i
j
j
j
n

U in D in ;

21
22

with D the average stiffness value after damage (usually supposed equal to the undamaged stiffness D for a uniform beam).
The previous expressions allow simplifying Eq. (20) as
n
i
zi Kj zi CDj i
j  j :

23

In the same way, Eq. (17) can be expressed as (K N


j 1 Kj :
N
i iT
z Kj zi j ziT Kj zi

j1 i
j1

j ziT Kj zi 

j1

j1

i iT
z Kj zi 0:
i

24

Substituting Eq. (23) into the previous equation and taking into account Eq. (21) yield


N
N
i N
in
i
Dj i
2 j Dj i
j 2
j j Dj j  j
i j1
j1
j1


N

n
i
C0
 i Dj i
j  j
i j 1

25

Considering that the first and third terms represent (global) curvature integrals for the damaged and undamaged cases,
one finally obtains

 

N
n
i
i in i
j i
j j

i
j1

26

188

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

or
n
i
j

j1

n
i
i
j j
in i

i
j

j1

in

1
1

i
j
i

i
i

27

The previous system of equations requires that the number of considered vibration modes M is equal to the number of beam
elements N, i.e., the possible damage locations. This approach has two limitations: first, improving the accuracy in
determining the damage position requires identifying high-order modes, which is not a trivial task; as a second point,
uncertainties in system coefficients and r.h.s. frequencies have an increased impact on the solution as system dimension
grows. For this reason, Fan and Qiao [19] proposed their technique for an evaluation of damage severity after identifying the
damage position (and extension). To achieve this purpose, they defined a DSCF (Damage Severity Correction Factor) which
was evaluated after that the DLF (Damage Location Factor) had identified the most damage-sensitive modes for every beam
element. Thus, the damaged elements were identified processing these information. At the end, the product DSCF times DLF
gives the coefficients of Eq. (27) evaluated only in regions where damage is likely to be present.

3.3. Kim and Stubbs


The approach of Kim and Stubbs [13] was presented by the authors as a particular case of the relation introduced in the
paper of Stubbs and Osegueda [6] that expresses the change in system frequencies with the variation of strain energy. On
the other hand, in developing the damage index Kim and Stubbs (and later Kim et al. [14]) exploited the concept of modal
stiffness which is equivalent to the definition of modal strain energy. Thus, the method of Kim and Stubbs can be also
interpreted as a particular case of the theory developed by Fan and Qiao [19].
Supposing zi 0, Eq. (17) becomes
ziT Kzi 

i iT i
z Kz C 0;
i

28

where the relative change in the i-th eigenvalue can be isolated:


i ziT Kzi
C
:
i
ziT Kzi

29

Imposing again zi 0, the expression of damaged strain energy (Eq. (19)) becomes
U in C ziT Kzi ziT Kzi :

30

Assuming damage in the element j, that is, K Kj , the variation of strain energy in that element is
n
iT
iT
Kj zi ziT Kj zi U i
Kj zi :
U i
j z
j Cz

31

Obtaining from the previous equation the expression ziT Kzi and substituting into Eq. (29), one obtains
in
i
i U j  U j
C
;
i
i
U

32

which can be obtained also from Stubbs and Osegueda [6].2


Kim and Stubbs express the strain energies in terms of curvature integrals as
i
n in
i Dj j  Dj j
C
;
i
D i

33

with D the average stiffness value of the undamaged beam. At this point, an additional hypothesis needs to be made about
the expression of global strain energy. If Dj D, one can solve for the damage intensity as
Dnj
D

i i
i
j
i
;
n
i
j

34

n
2
In fact, disregarding the variation of mode shapes, i.e., zi C 0 and i
 i
j 0 and supposing that damage affects only a single element jd, the global
j
in
in
i
variation of strain energy is given by the variation of the strain energy in a single element, thus U in  U i N
 U i
j 1 U j
j C U jd  U jd , which after
substitution in the index formulated by Stubbs and Osegueda gives Eq. (32).

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

Furthermore, it is defined as


i i
i
M

i1
j
i
;
j
in
M
i 1 j

189

35

which is exactly the inverse of the index defined by Kim and Stubbs that assumed j C j 1 and used this expression also
for the estimation of damage severity. It is interesting to note that, passing from Eq. (31) to Eq. (33), it has been assumed that
n
n
U i
Dnj i
which is not consistent with the assumption zi C0. Thus, deriving the method of Kim and Stubbs from the
j
j
method of Fan and Qiao, the position zi 0 seems to be applied only in the perturbed eigenvalue equation. On the other
n
hand, if one assumes without distinction zi 0, the relation U i
Dnj i
j
j holds, and one obtains
i
j
i

i
;
i

36

which becomes a particular case of Eq. (27).


3.4. Stubbs and Osegueda
As anticipated before, Stubbs and Osegueda [6] started from the relationship:
i U in  U i
C
;
i
U i

37

which can be further expanded with respect the variation of the bending stiffness as

i U in  U i
1 U in 
C

Dj ODj 2

i
U i
U i Dj 

38

Dj 0

Assuming a linear dependence with respect to damage, it is set as



N
i
1 U in 
f ij Dj with f ij i
;

i
U Dj 
j1

39

Dj 0

or, re-writing Eq. (39) in the form of a linear system of equations,



N
Dj U in 

;
F ij j i with F ij i

i
U Dj 
j1

40

Dj 0

This equation has to be solved in the unknown damage indicators, 1, 2 and so on. One of the critical points is that, if the
intact structure is symmetric (also taking into account the boundary conditions), the sensitivity matrix F F ij  is singular. In
fact, Stubbs and Osegueda had to introduce the analysis of longitudinal vibrations for breaking this symmetry. To obtain the
i
coefficients Fij, one has to recall that these coefficients represent the energy fractions, expressed also as i
j = . Thus,
assigning j and calculating numerically i =i from a structural model, the energy fractions are the unknowns of the
following equation:
i
N
i
j
i j :
i
j1

41

As already remarked for the method of Fan and Qiao [19], solving Eq. (40) requires that the number of modes M is equal to
the number of structural elements. An alternative way also proposed by Stubbs and Osegueda is using an iterative procedure
with regridding over the experimental dofs.
3.5. Cornwell et al.
The method of Cornwell et al. [12] (also presented in [11]) is based on a simple statement which prescribes that the ratio
between local and global strain energy does not change before and after damage occurs:
n
U i
j

U i
j

U i

C
in

42

Recalling the relations between strain energy and curvature integrals (Eqs. (21) and (22)), one obtains
n
Dnj i
j
n

D i

Dnj i
j
D i

43

190

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205
n

Assuming D CD (acceptable if damage is small), one can solve for the stiffness reduction:
Dnj
Dj

i
i
j =
n
in
i
j =

44

To enhance the performance of the damage indicator, the contribution of all the modes can be retained:
j

i i
M
i 1 j =
in in
M
i 1 j =

45

which is the inverse of the index defined by Cornwell et al. (as for [13], this provides also an estimation of the damage
severity). To highlight the hypotheses behind relation (42), it is useful to expand the l.h.s. with respect to the stiffness:

n
in
U i
U i
U j 
j
j

Dj ;
46

U in U i Dj U in D 0
j

where it is assumed that damage occurs only in the j-th element. After evaluating and re-arranging the derivative
expression on the r.h.s., it follows
0
1


n
in
in 
U i
U i
U i
1 @U j 

j
j U
j
ADj


47


U in U i U i Dj D 0 U i Dj D 0
j

n
Dnj i
If the variation of mode curvature is neglected, i.e., U i
j
j , one obtains


in 
U j 
U in 

i


j ;

Dj
Dj 

Dj 0

48

Dj 0

and, substituting the above expressions into Eq. (47), one has
!
n
U i
U i
i
U i
j
j
j
j

1

Dj
U in U i U i
U i

49

Furthermore, dividing and multiplying by Dj, one has


!
n
U i
U i
U i
U i
j
j
j
j

1

j
U in U i U i
U i

50

The previous equations show that Cornwell's approximation is valid if jj j{1.


3.6. Stubbs et al.
Stubbs et al. [8] start from the consideration of the order of magnitude relative to terms in Eq. (42), i.e., for N sufficiently
large,
n
U i
j

in

U i
j

{1;

{1

51

1:

52

U i

which also implies


n
U i
j

in

1C

U i
j
U i

Substituting into the previous equation the local expression of modal strain energy in terms of curvature integrals (Eqs. (10)
n
and (21)) along with the global expressions making use of the average stiffness (U in D in and U i D i ) yields, after
some algebra,
!
D i
n in
i
D

j
Dj
Dnj
!
C
53
n
Dj
D
in
i
i
n
D
j n
Dj
n

Assuming a uniform bending stiffness (Dj D) and a small damage (D CD CDnj ), Eq. (53) is simplified as
i
in i
Dnj
j
C
i
n
Dj i in
j

54

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

191

Stubbs et al. [8] defined an index after summing over the considered modes at both numerator and denominator:
j C

in i
M
j i
i 1
i in
in
M
i 1 j

55

which is the inverse of the index originally defined by Stubbs et al. This index was normalized by Stubbs et al. for
highlighting damage zones where it assumes values different from unity.

3.7. Choi et al.


Choi et al. [15] proposed a damage index (compliance index) suitable to be applied to 1D and 2D structures using the
concept of compliance. Its derivation is a bit different from theories already considered in this section and, though in Choi
et al. [41] the compliance index was further generalized, it is here articulated with reference to the present case. The relation
between the bending moment M(x) and the curvature, wx according to EulerBernoulli theory is
2 wx
M Dx
:
x2

56

The above equation can be integrated over the length of each beam element for providing an average bending moment:
Mj

1
xj

xj 1

xj

M dx

Dj
xj

xj 1
xj

2 wx
dx
x2

57

with xj xj 1  xj , where the stiffness does not vary over the element j. The same result can be obtained for a damaged
element:
n

Mj

Dnj
xj

xj 1

xj

2 wn x
dx
x2

58

Choi et al. assumed that the load does not change after damage (this applies only to statically determinate structures) and,
as a consequence, the distribution of bending moment has to remain the same. This formally implies
R xj 1 2 wx
dx
xj
x2
C
2
n
R xj 1 w x
Dj
dx
xj
x2

Dnj

59

Applying the above formula to each identified mode shape yields


R xj 1 2 i x
dx
Dnj
xj
x2
:
C
Dj
R xj 1 2 in x
dx
xj
x2

60

It is worth mentioning that Choi et al. defined a damage index (again, the inverse of that actually defined here) for
identifying the damage position summing over all the modes and adding unity to both numerator and denominator, thus
obtaining
R xj 1 2 i x
dx
xj
x2
:
^j
i
n
2
R xj 1 x
dx
1 M
i 1 xj
x2
1 M
i1

61

This choice was done for avoiding that the denominator in Eq. (60) is close to zero as it may happen for the first modes.
However, summation over the modes may provide an overall index value that significantly oscillates requiring normalization. In the present case, also in the perspective of obtaining index values comparable with the other damage indexes, it
has been preferred to introduce an absolute value in the above equation as


R

i
 xj 1 2 x

1 M
dx


i 1  xj
2

x

:
j
62
R

2 in
 xj 1 x

M
dx
1 i 1  xj



x2

192

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

4. Damage estimation based on mode curvature analysis without a baseline


In this section, two methods based on the analysis of mode shape curvature in the damaged case alone have been
considered. Though these methods offer the advantage of avoiding any reference to the undamaged case, they suffer of noise
affecting input data more than methods based on calculation of global quantities. Calculation of curvatures from mode
shapes affects also baseline data, but many of them rely on subsequent evaluation of integral expressions that mitigate
propagation of errors. For this reason, optimal sampling for reducing differentiation errors [42] as well as sophisticate postprocessing (e.g. fuzzy logic, see [43]) has a deeper impact on techniques looking for lack of smoothness of curvature mode
shapes. The selected methods are the modified Laplacian operator [26] and the fractal dimension [34-36] which illustrates
two different approaches for addressing this problem without handling the presence of noise as it has been done
successively in several ways (see references in Section 1).
4.1. Modified Laplacian operator
Ratcliffe [26] proposed a damage index which is based on the analysis of discontinuities of curvatures in the damaged
beam. The tool for highlighting the lack of smoothness is the evaluation of the Laplacian operator at the nodes where the
curvature values are available. For a 1D structure, the result is proportional to the calculation of the second-order spatial
derivative with finite differences. Thus, the Laplacian operator applied to the i-th mode i is
Li xj in xj 1  2in xj in xj  1 ;

63

in

where denotes again the dimensional mode shape after normalization. Similar indicators of damage location have been
identified by several authors looking to discontinuities (jumps) in curvature distribution, which is basically provided by
Li xj . However, Ratcliffe noticed that if damage is small, the Laplacian operator fails in highlighting sharply the damage
location. For this reason, Ratcliffe also proposed a way for amplifying the sensitivity to damage: for each j-th beam element,
a third-order polynomial, say P i xj , is defined using the values of Li in the points xj  2 , xj  1 , xj 1 and xj 2 . The new
damage indicator i
j is defined as
i
i
i
j P xj L xj ;

j 3; ; N  1

64

Choosing a different i, it is possible to take into account the curvature relative to different modes.
In the present application, because of the availability of the analytical expression of the curvatures, the Laplacian is
evaluated at the mid-point of the damaged beam element and the index i
j has been properly normalized with respect to its
maximum value for putting in evidence the damaged zones:
i

L j in x j 1  2in x j in x j  1

65

where the index j denoting the Laplacian addresses the j-th element and the point x j is the mid-point of the j-th element.
4.2. Fractal dimension
Several authors have proposed index based on the concept of fractal dimension (FD), initially introduced by Katz [44], for
the analysis of generic waveforms. The FD allows estimating the intrinsic complexity of a geometrical object. If it is applied
to a smooth curve, FD gets the value of unity. If the curve is not smooth, the lack of regularity of the curve may make the
geometrical object more similar to a strip, i.e., it partially loses the appearance of a curve. In this case, FD will assume values
greater than unity as this irregular behavior grows. To define an index for locating damage running over the elements of the
beam, the definition of FD needs to be restricted to a curve portion. With reference to the grid defined over the beam (see
Fig. 8), a set of points falling in the neighborhood of P j  xj ; xj can be defined, where x represents the i-th curvature
mode shape in the present application. The generic element in this neighborhood is P j k  xj k ; xj k , where k assumes
integer values between  M=2; ; M=2, with M an even integer indicating the number of samples in the window. Thus, two
geometrical parameters can be defined with reference to the beam portion associated with the node j:
dj maxdistP j  M=2 ; P j  M=2 k ;
Lj

M=2  1

k  M=2

k 1; ; M

distP j k ; P j k 1 ;

66
67

with distA; B the function giving the distance between the points A and B. The FD for the node xj can be defined as
FDj
log10

log10 n
 
dj
log10 n
Lj

j M=2 1; ; N  M=2

68

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

193

The integer n is set to M (the window length) by Li et al. [35] whereas is set to N (the overall number of points) by
Hadjileontiadisa et al. [34] and Wang and Qiao [36]. Wang and Qiao [36] also introduced a stretching in the x dimension by
multiplying the node abscissa by a factor S.
5. Results
In the following, a comparison of all considered damage identification techniques is carried out taking into account their
capability of identifying the damage position (Section 5.1) and the damage intensity (Section 5.2). The unique exception is
relative to the method of Stubbs and Osegueda [6] because it cannot be applied to a symmetric structure (symmetry is
intended in terms of shape and boundary conditions).
If not further specified, all the first five beam modes contribute to the calculation of the damage index. In Table 1 the
correspondence between the number of beam elements N and the element jd affected by damage (uniform stiffness
reduction) is shown. In Section 5.3 partially damaged elements are also considered. Several stiffness reductions will be
considered throughout the damage analysis, ranging from 1% to 20%.
5.1. Damage position
In this section, results about the identification of the damage position are presented by showing the distribution of the
various damage indexes along the structure.
The damage indexes proposed by Cornwell et al. [12] (Fig. 9) and Choi et al. [15] (Fig. 10) are plotted varying the number
of beam elements relatively to a given damage equal to Dn =D 0:95. The method of Cornwell et al. behaves regularly for all
the grids and always permits the identification of the damaged element. The index remains close to 1 for the undamaged
elements whatever the number of elements is considered but performs better as long as N grows. The compliance index
proposed by Choi et al. (Eq. (62)) performs correctly if the number of elements is N 20 and N 100 (its value approximates
also the reduced stiffness ratio). However, the damage index locates ambiguously the damaged element in the case N 5:
both beam elements 3 and 4 seem candidate in the same way to be affected by damage.
At first glance results obtained with the method of Stubbs et al. [8] seem not so good as in the previous cases. In Fig. 11
the damage index is plotted separately for each case (5, 20 and 100 elements) for highlighting better the index variation at
the damaged element. The damage index assumes uniform values very close to unity in the undamaged elements. Thus,
even if the variation due to damage is small, the index may reveal the presence of damage choosing proper scaling of the
vertical axis.
The index proposed by Kim and Stubbs [13] presents relevant oscillations at the boundaries and in positions
symmetrically distributed with respect to the beam mid-point (see Fig. 12). Only if the beam is divided into few elements,

Table 1
Indication of the damaged element jd with respect to the number of
elements N in numerical tests.
N

jd

5
10
20
100
300

4
8
16
80
240

Damage index

1.05

0.95

0.9

0.2

0.4

0.6

0.8

Fig. 9. Distribution of the damage index j defined by Cornwell et al. [12] for several grids (N 5, 20, 100).

194

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

Damage index

1.05

0.95

0.9

0.2

0.4

0.6

0.8

Fig. 10. Distribution of the damage index j defined by Choi et al. [15] for several grids (N 5, 20, 100).

1.01

Damage index

1
0.99
1

0.995
1
0.999
0

0.2

0.4

0.6

0.8

Fig. 11. Distribution of the damage index proposed by Stubbs et al. [8] for several grids (stiffness reduction j 1 is plotted).

1.1

Damage index

1.05
1
0.95
0.9
0.85
0.8

0.2

0.4

0.6

0.8

Fig. 12. Distribution of the damage index j proposed by Kim and Stubbs [13] for several grids (N 5, 20, 100).

the index oscillations disappear. To explain this behavior, the damage index is evaluated with just one mode at a time.
For the case with N 20, the damage index is first evaluated using the first bending mode, then using the second mode and
so on up to the fifth mode (Fig. 13). The largest oscillations involve especially the first (at the edges), the second and the third
mode. On the other hand, these oscillations are negligible for the remaining two modes. The reason for that can be
explained considering Eq. (34). Close to the nodes of mode shape curvature (including the boundaries), curvature integrals
in
i
become small. Thus, while i
j and j
j can be disregarded in the numerator, the whole fraction is highly sensitive to small
changes in the denominator. On the contrary high-order modes are less affected due to high slopes in the neighborhood of
nodes. Thus, averaging over all the modes allows reducing the detrimental effect of the low-order modes in the inner points,
keeping still distinguished the damage location. Moreover, when damage occurs close to the nodal points the index
variation is absorbed by these undesired index oscillations. It is worth noting also that for all the modes the index assumes
nearly the same value at the damage position (mean 0.75).

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

195

Damage index

0.8

st

1 mode
nd
2 mode
3 rd mode
4 th mode
5 th mode

0.6

0.2

0.4

0.6

0.8

Fig. 13. Distribution of the damage index j proposed by Kim and Stubbs [13] computed separately with the first five bending modes (N 20).

Damage index

1.1

1.05

0.95

0.9

0.2

0.4

0.6

0.8

Fig. 14. Distribution of the damage index proposed by Fan and Qiao [19] for several grids (stiffness reduction j 1 is plotted).

The index proposed by Fan and Qiao [19], even if it is the most reliable with respect to the estimation of damage intensity
(see later), presents undesired oscillations at the beam edges as shown in Fig. 14. This observation agrees with the remarks
made by Fan and Qiao in their paper and reported at the end of Section 3.2 concerning to critical issues relative to Eq. (27).
Looking at Fig. 14, it appears that using less elements (N 5) leads to a clearer identification of the damaged portion with
respect to the cases N 10 and N 20. The estimation of damage severity can be refined with more dense grids but falsepositive cases can be excluded only on the basis of the analysis carried out with less nodes.
To highlight better the comparison between the methods with a baseline, several diagrams presenting together the
different index distributions are shown in Figs. 1517. The number of elements is the same with N 20 and three stiffness
ratios, Dn =D 99%; 95% and 90%, are considered for the damaged element. The scales of the y-axis are different for each
curve so as to highlight relative differences in the index values which address the presence of damage. All the methods do
not present false-negative cases whereas index oscillations may exist especially at the boundaries. In general, this numerical
behavior is likely to occur with methods more accurate in estimating the damage intensity than others less performing.
Next, the indexes not based on a comparison with a reference condition are taken into account. As mentioned in
Section 3, no estimation of damage intensity is addressed by these indexes. First, the results from the application of the
method of Ratcliffe [26] in the case Dn =D 0:95 are shown in Fig. 18. The position of damage is highlighted through
variations of the (normalized) index applied to the first mode shape and evaluated for different grids (N 20, 100, 300).
Oscillations due to the index definition are present but they are symmetrical with respect to the point indicating the exact
damage position. Indeed, this index is not much affected by the grid as long as the damage extension covers completely a
beam element between two nodes. In Figs. 192021, the investigation is performed changing the considered mode for two
grids made of 20, 100 and 300 elements. Though the index performs well independently to the considered modes, some
index oscillations occur for the grid with the smallest number of points. Thus, a degradation of the index behavior may occur
in combination with coarse grids and with its application to high-order modes.
The indexes based on fractal dimension (FD) are here applied to modal curvatures and not directly to mode shapes as
they were originally because of the availability of their analytical expressions. Curvature mode shapes have been previously
normalized with respect to unity and FD is assumed to be computed on a grid considering the mid-points of each element
for avoiding ambiguities with the value of D(x) at nodes of damaged elements. The FD according to Hadjileontiadisa et al.
[34] is applied to the first three modes' shapes considering M4 and a ratio N=M 75 which is the same as that adopted in

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

Damage index

1.01
1
0.99
1.025

Choi

Cornwell

0.975
1.01
1
0.99
1.025
1
0.975

Fan

Kim

1.002
1

Stubbs

0.998

0.4

0.2

0.8

0.6

Fig. 15. Index distribution for all the baseline methods (N 20 and Dn =D 0:99).

Choi

0.95

Damage index

Cornwell

0.9
1.1
Fan

1
0.9
1
0.9
0.8
1.01

Kim

Stubbs

1
0.99

0.2

0.4

0.6

0.8

Fig. 16. Index distributions for all the baseline methods (N 20 and Dn =D 0:95).
1.1
1

Damage index

Choi

0.9
1.2
1
0.8

Cornwell

1.125
1
0.875
1.2
1
0.8

Fan

Kim

1.01
1
0.99

Stubbs
0

0.2

0.4

0.6

0.8

Fig. 17. Index distribution for all the baseline methods (N 20 and Dn =D 0:90).
1
0

Damage index

196

-1
1
0
-1
1
0
-1
0

0.2

0.4

0.6

0.8

Fig. 18. Distribution of damage index for different grids according to Ratcliffe [26].

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

197

1st mode

Damage index

-1
1

2nd mode

0
-1
1

3rd mode

0
-1
0

0.2

0.4

0.6

0.8

Fig. 19. Distribution of damage index for different modes with N 20 according to Ratcliffe [26].

1st mode
mo

Damage index

-1
1

2nd mode

0
-1
1

3rd mode

0
-1
0

0.2

0.4

0.6

0.8

Fig. 20. Distribution of damage index for different modes with N 100 according to Ratcliffe [26].

mo
1st mode

Damage index

0
-1
1

2nd mode

0
-1
1

3rd mode

0
-1
0

0.2

0.4

0.6

0.8

Fig. 21. Distribution of damage index for different modes with N 300 according to Ratcliffe [26].

their paper (Fig. 22). Results relative to a given damage with Dn =D 0:95 are indeed quite satisfactory in this case and get to
be significantly bad only for N 20 as shown in Fig. 23 (M is the same as before). However, increasing the stiffness reduction
and using the first curvature mode shapes provide more chances of identifying the damage location (see Fig. 24). The
approach of Wang and Qiao [36] (GFD) is similar in the choice of parameters in calculating FD but a stretching is applied on
x-axis affecting the computation of distances. The achieved estimation of the damage location in Fig. 25 is clearer (top and
mid diagrams) but with the coarse grid (N 20) requires changing the y-scale to be satisfactorily precise in locating small
stiffness reduction (scales in the y-axis vary from figure to figure). The estimation of FD on the first curvature mode shape

198

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

Damage index

1.25

1st mode
2nd mode
3rd mode

1.2
1.15
1.1
1.05
1
0

0.2

0.4

0.6

0.8

Fig. 22. Distribution of the damage index (FDCD) for N 300 and M 4 (Hadjileontiadisa et al. [34]) calculated over different modes with given stiffness
reduction Dn =D 0:95.

1st mode
2nd mode
3rd mode

Damage index

1.4
1.3
1.2
1.1
1
0

0.2

0.4

0.6

0.8

Fig. 23. Distribution of the damage index (FDCD) for N 20 and M 4 (Hadjileontiadisa et al. [34]) calculated over different mode shapes with assumed
Dn =D 0:95.

1.12
D* / D = 0.95
D* / D = 0.90
D* / D = 0.80

Damage index

1.1
1.08
1.06
1.04
1.02
1
0.98

0.2

0.4

0.6

0.8

Fig. 24. Distribution of the damage index (FDCD) for N 20 and M 4 (Hadjileontiadisa et al. [34]) calculated on the first mode shape with different Dn =D.

obtained by Li et al. [35] is slightly better than those obtained by Hadjileontiadisa et al. [34] also with the coarser grid
(Fig. 26). Considering the second mode and the intermediate grid case N 100, it is interesting to compare in Fig. 27 the
techniques of Hadjileontiadisa et al. [34], Wang and Qiao [36] and Li et al. [35] for the same choice of the length of the
sliding window M, i.e., M2. In fact, there is not much difference between the techniques in highlighting the damaged
location apart from different y-scales which, fundamentally, depends on the value assigned to n.
5.2. Damage intensity
The evaluation of damage intensity is the second step in damage analysis. The methods avoiding any reference or
baseline data (like the Laplacian operator or the fractal dimension) are excluded because they do not directly offer this kind

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

199

1.01
1.005

Damage index

1
1.001
1

1.0001
1
0

0.2

0.4

0.6

0.8

Fig. 25. Distribution of the damage index (GFD, Wang and Qiao [36]) for N 20; 100; 300 and M 4 calculated on the first mode (solid line: Dn =D 0:95,
dashed: Dn =D 0:90, with points: Dn =D 0:80).

D* / D = 0.80

Damage index

2
0

D* / D = 0.90

1.05
1
1.05

D* / D = 0.95

1
0

0.2

0.4

0.6

0.8

Fig. 26. Distribution of the damage index (FD) for N 20 and M 2 (Li et al. [35]) calculated on the first mode shape.
5

Li (M = 2)

Damage index

0
-5
1.0004

GFD (M = 2)

1.0002
1
1.4

Hadjileontiadisa (M = 2)

1.2
1
0

0.2

0.4

0.6

0.8

Fig. 27. Distribution of the damage indexes for N 100 and M 2 from Li et al. [35], Wang and Qiao [36] and Hadjileontiadisa et al. [34] calculated on the
second mode shape.

of result. Provided that the selected identification technique has correctly identified the damage position, the same index
can be often exploited for estimating the damage if no normalization had been made on it. Though generally following this
rule for remaining adherent with examined theories, in the case of method of Choi et al. [15] Eq. (60) has been considered
for damage estimation. Thus, with reference to the index value at the damaged element jd 34 (N 100), a comparison with
the given stiffness reduction D Dnjd =D is given in terms of the estimated one, D  Djnd estim =D or jd in Fig. 28. The bisector
assumes a particular meaning: it represents the points of an ideal method which perfectly identifies the damage intensity,

200

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

0.7

|j| (estimated damage)

0.6
0.5
0.4
0.3
Cornwell
Fan
Kim
Stubbs
Choi

0.2
0.1
0

0.1

0.2

0.3

0.4

0.5

0.6

0.7

*
j

( D - D ) / D (given damage)
Fig. 28. Identified damage severity jjd j for all the methods (jd 34, N 100).

0.1

|j| (estimated damage)

0.08
Overestimating

0.06
Underestimating

0.04
Cornwell
Fan
Kim
Stubbs
Choi

0.02

0
0

0.02

0.04

0.06

0.08

0.1

( D - D j ) / D (given damage)
Fig. 29. Identified damage severity jjd j for small stiffness reduction (jd 34, N 100).

that is, given and predicted damage coincide. This line divides the plane into two distinct areas, the upper triangle where
damage overestimations fall and the lower triangle where damage underestimations lie. Looking at Fig. 28, it is evident that
all the techniques usually underestimate or overestimate independently to the amount of structural damage, at least in the
present application. According to this classification, the indexes of Kim and Stubbs, and Cornwell et al. overestimate, whilst
the modified Choi et al. index slightly underestimates damage and the Stubbs et al. index seems losing an acceptable
correlation with damage severity. Indeed, the index of Fan and Qiao succeeds in getting correctly the damage and
no degradation is present as stiffness reduction gets worse (this applies substantially to the Choi et al. index as well).
Recalling that usually the aim is identifying damage in the early stages of its development, a closer look for low values
of stiffness reduction can be interesting by looking at Fig. 29. A relevant premise in discussing these data is that under- or
over-estimation does not necessarily preclude a good identification of the position, that is a matter of distribution of index
values all over the analyzed region. Secondarily, from the point of view of providing conservative estimations, the techniques of Kim and Stubbs and Cornwell et al. can be still considered as reliable methods. As said before, in this perspective

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

201

the Stubbs et al. index indeed provides too low values with respect to the others. Fan and Qiao and Choi et al. slightly
underpredict damage but this is quite irrelevant and demonstrates to be the best indexes for obtaining the damage intensity.
5.3. Damage affecting more than one element
As anticipated before, in this section the hypothesis that damage coincides with a single element is lifted. The analysis is
carried out keeping constant the damage length but shifting the damage location gradually, as shown in Fig. 30. A shift equal
to 0% represents the initial damage location with the element jd entirely damaged and a shift equal to 100% denotes that the
damage domain has been shifted of an amount equal to the length of a single element, and coincides now with the element
jd 1. Taking the damage position as a parameter, even if only small shift will be allowed, this investigation appears again as
a sensitivity analysis. However, since all the methods explicitly or implicitly assume this coincidence in the theoretical
background, the loss of this property brings the studied techniques close or even beyond the limits of their applicability. The
authors seem aware of this when they handle experimental data with more complicate schemes not accounted for in the
original theory. Here, the focus is on understanding how much the mathematical core of the technique is affected by shifting
the damage with respect to the original position. The given reduction of stiffness is Dn =D 0:90 whereas the number of
elements for the index calculation is N 20. This investigation is limited to methods with a baseline because the second
group of methods generally requires a larger number of experimental dofs. Moreover, using the element mid-point for the
index calculation (in non-baseline methods) implies that shifts of the damaged portion less than half the element length are
not detectable.
The methods of Cornwell et al. [12] and Choi et al. [15] behave in a similar way (also) with respect to damage shift. It is
interesting to observe in Figs. 31 and 32 that when damage is shifted by 50% of its length, the damage indices on the two
equally damaged elements differ slightly. Similarly, if damage is totally shifted on the adjacent element (shift is 100%), the
estimated damage intensity is varied with respect to the unshifted case. This is an effect due to relating the damage index to
the estimation of curvature changes which are not the same for all the modes.
Concerning the identification of damage location, the method of Stubbs et al. [8] performs even better. As shown in
Fig. 33, the index assumes nearly the same values as long as damage is shifted from one element to the other. Unfortunately,
the estimated damage is far from its real value.

Fig. 30. Position of the damage portion (checkerboard pattern) with respect to the element grid for different shifts from the initial damage location (0% shift).

Damage index

1.1

0% shift
25% shift
50% shift
75% shift
100% shift

0.9

0.8

0.2

0.4

0.6

0.8

Fig. 31. Effect of damage shift for Cornwell et al. [12].

202

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

Damage index

1.05

0% shift
25% shift
50% shift
75% shift
100% shift

0.95

0.9

0.2

0.4

0.6

0.8

Fig. 32. Effect of damage shift for Choi et al. [15].

Damage index

1.005

0.995

0% shift
25% shift
50% shift
75% shift
100% shift

0.99
0

0.2

0.4

0.6

0.8

Fig. 33. Effect of damage shift for Stubbs et al. [8].

Damage index

0.9

0% shift
25% shift
50% shift
75% shift
100% shift

0.8

0.7

0.2

0.4

0.6

0.8

Fig. 34. Effect of damage shift for Kim and Stubbs [13].

The method of Kim and Stubbs [13] as shown in Section 5.1 presents relevant oscillations of the index not related to the
presence of damage. This undesired feature becomes even worse when damage is shared by two adjacent elements because
the damage index gets lower values of the same order of damage oscillations (see Fig. 34).
The analysis of damage shift for the method of Fan and Qiao [19] shows some unexpected features. The index of Fan and
Qiao [19] becomes less precise if elements are not uniformly damaged. If damage is partially shifted to the successive
element, index oscillations appear not only at the beam edges (see Fig. 35) compromising the clear capability of the method
of identifying both position and intensity shown in the unshifted case. As outlined at the end of Section 3.2, Fan and Qiao
[19] decomposed the overall index into the product of two quantities, DLF and DSCF. The DLF, used for locating damage, is
shown in Fig. 36, indicating that damage occurs between elements 16 and 17 (it depends on the assumed threshold).
The DSCF is then evaluated only on possible damage locations, i.e., for 0:70 r r 0:90, and the final estimation of damage
intensity is then reported in Fig. 37, showing a good agreement with target values.

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

0%
25%
50%
75%
100%

20

Damage index

203

10
0
-10
-20

0.2

0.4

0.6

0.8

Fig. 35. Effect of damage shift for Fan and Qiao [19].

DLF

4
3
0% shift
25% shift
50% shift
75% shift
100% shift

2
1
0

0.2

0.4

0.6

0.8

Fig. 36. Identification of damage position via DLF according to Fan and Qiao [19].

1.05

DLF* DSCF

1
0.95
0% shift
25% shift
50% shift
75% shift
100% shift

0.9

0.85

0.2

0.4

0.6

0.8

Fig. 37. Effect of damage shift based on DLF and corrected with DSCF according to Fan and Qiao [19].

6. Final remarks
In this paper, a comparison of the performances of several damage indices based on the analysis of quantities related
(more or less directly) to mode-shape curvatures and frequencies has been carried out. Both techniques requiring reference
data and techniques doing without have been considered. The use of noiseless data (analytical mode-shapes) has aimed to
evaluating and comparing the intrinsic capability of each mathematical formulation in addressing the basic inverse problem
(from vibration modes to damage). Therefore, additional post-processing of raw identification results or combinations of
different techniques have not been taken into account at this stage with only one exception regarding the method of Fan and
Qiao [19].

204

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

The main achievements from this analysis are listed in the following:

 Detection of damage location and estimation of damage severity are tasks that have to be considered separately, because







an index performing well in identifying the damaged beam element may be not so accurate in providing the damage
intensity and vice versa. Examples of this statement are the methods of Stubbs et al. [8] and Fan and Qiao [19]. It is worth
noting that the method of Kim and Stubbs [13] underperformed in both the cases whereas the method of Choi et al. [15]
appears as a good compromise between damage localization and damage estimation. Cornwell et al. [12] give better
results in damage localization and overestimate damage.
The use of these techniques requires to be carefully understood for avoiding unpleasant situations: this is the case of
methods which solve linear systems whose coefficients are determined in the undamaged condition (e.g. the method of
Stubbs et al. [6]). Rank deficiency or at least ill-conditioning is likely to occur in monitoring physical systems
characterized by symmetry. This applies in general because solving inverse problems brings linear systems with illconditioned coefficient matrices.
False-positive cases are likely to occur much more than false-negatives, even if the latter might occur for particular
damage locations.
Also in the present ideal condition because of the link between beam element and damage extension, it is not always
provided that convergence of the damage identification is obtained increasing the number N of beam elements.
In general, methods based on baseline data are less demanding in terms of spatial sampling than methods based on the
analysis of smoothness of the curvature alone.
The previous concept applies also to the number M of considered modes: some indices perform better with low-order
modes, other with high-order modes. Averaging over all the modes is a chance, but positive contributions in locating
damage may be balanced and hidden by negative ones.
Removing the condition that damage extension covers entirely one element has revealed to be not so critical with the
exception of the methods of Kim and Stubbs [13] and Fan and Qiao [19]. In the case of Fan and Qiao [19], an approach
based on relating mode shapes and frequency variations to stiffness reduction is highly ill-conditioned and requires
separating the two problems, the identification of the position and the estimation the severity of damage. With
dedicated post-processing, Fan and Qiao [19] can still provide a fine tool for detecting damage. Nonetheless, it is worth
underlining that the position of damage along the beam and the spread of damage on consecutive elements has an
impact on the accuracy of damage identification in all the methods.
The methods based on the analysis of the smoothness of the curvature mode shape offer good chances of identifying
damage even without any reference to the undamaged structure. The modified Laplacian operator (MLO) of Ratcliffe
[26] behaves reasonably well for the simply supported beam even with a low number of points. The methods based
on fractal dimension (FD) appear to be more demanding in terms of required samples on the curve than the MLO. It
is worth remarking that the FD can be applied also to displacement mode shapes which allows avoiding relevant
noise on the processed data. All the considered FD methods can be considered as particular cases of the GFD
introduced by Wang and Qiao [36] for particular choices of M, n and stretching of the x-axis. Indeed, the techniques
of Wang and Qiao [36] and Li et al. [35] appear as significative improvements of the original work of Hadjileontiadisa
et al. [34].

Indeed, this analysis can not yet be considered as an exhaustive guideline for selecting the most appropriate technique
for damage identification in practical cases but just a starting point. Index thresholds for discriminating if damage is
present have not been discussed because they often are problem dependent. Mixed formulations or multi-level approaches
are also conceived for reducing the possibilities of false-positive cases. Moreover, some assumptions might turn to be
critical. In practical applications, the assumption that damage determines a uniform reduction of stiffness is hard to meet.
Indeed, smooth functions describing the damage distribution have a negative impact on highlighting curvature
discontinuities. A second point concerns the fact that experimental uncertainties and numerical errors are not taken into
account. This assumption is the working hypothesis which this paper is based upon. Nonetheless, techniques which
perform similarly may suffer differently from the presence of noise on data likely to be introduced by the calculation of
curvatures. From this point of view, an exception is given by FD which is usually applied also to displacement data. This
problem is more critical in methods based on the analysis of pointwise properties than in those calculating integral
quantities like strain energy or curvature integrals. Optimal sampling for reducing errors in second-order spatial
derivatives [42], filtering techniques like those introduced by Cao and Qiao [31,32], or post-processing advanced
techniques like fuzzy logic [43] are examples of feasible approaches for dealing with noisy data collected in experimental
campaigns.

Acknowledgment
This paper has been funded by Progetto Bandiera RITMARE, granted by Italian Ministry of Education and Research.

D. Dessi, G. Camerlengo / Mechanical Systems and Signal Processing 52-53 (2015) 181205

205

References
[1] L. Hernans, H. Van der Auweraer, Modal testing and analysis of structures under operational conditions: industrial applications, Mech. Syst. Signal
Process. 13 (2) (1999) 193216.
[2] R. Mariani, D. Dessi, Analysis of the global bending modes of a floating structure using the proper orthogonal decomposition, J. Fluids Struct. 28 (2012)
115134.
[3] S.W. Doebling, C.R. Farrar, M.B. Prime, A summary review of vibration-based damage identification methods, Shock Vib. Dig. 30 (2) (1998) 91105.
[4] W. Fan, P. Qiao, Vibration-based damage identification methods: a review and comparative study, Struct. Health Monit. 10 (1) (2011) 83111.
[5] P. Cawley, R.D. Adam, The location of defects in structures from measurement of natural frequencies, J. Strain Anal. 14 (2) (1979) 4957.
[6] N. Stubbs, R. Osegueda, R. Global, Nondestructive damage evaluation in solids, Modal Anal. 5 (2) (1990) 6779.
[7] N. Stubbs, J.-T. Kim, K. Topole, An efficient and robust algorithm for damage localization in offshore platforms, in: Proceedings of the ASCE 10th
Structures Congress, San Antonio, TX, 1315 April 1992. ASCE, Reston, VA , pp. 543546.
[8] N. Stubbs, J.-T. Kim, R.F. Charles, Field verification of a nondestructive damage localization and severity estimation algorithm, in: Proceedings of the
13th International Modal Analysis Conference, Nashville, TN, 1316 February 1995. SEM, Bethel, CT, pp. 210218.
[9] A.K. Pandey, M. Biswas, M.M. Samman, Damage detection from changes in curvature mode shapes, J. Sound Vib. 145 (2) (1991) 321332.
[10] M.M. Abdel Wahab, G. De Roeck, Damage detection in bridges using modal curvatures: application to a real damage scenario, J. Sound Vib. 226 (2)
(1999) 217235.
[11] P. Cornwell, M. Kam, B. Carlson, Comparative study of vibration-based damage identification algorithms, in: Proceedings of the 16th International
Modal Analysis Conference, Santa Barbara, CA, 25 February 1998. SEM, Bethel, CT, pp. 17101716.
[12] P. Cornwell, S.W. Doebling, C.R. Farrar, Application of the strain energy damage detection method to plate-like structures, J. Sound Vib. 224 (2) (1999)
359374.
[13] J.T. Kim, N. Stubbs, Improved damage identification method based on modal information, J. Sound Vib. 252 (2) (2002) 223238.
[14] J.T. Kim, Y.S. Ryu, H.M. Cho, Damage identification in beam-type structures: frequency-based method vs mode-shape-based method, Eng. Struct. 25 (1)
(2003) 5767.
[15] S. Choi, S. Park, S. Yoon, Nondestructive damage identification in plate structures using changes in modal compliance, NDT&E Int. 38 (7) (2005)
529540.
[16] U. Galvanetto, G. Violaris, Numerical investigation of a new damage detection method based on proper orthogonal decomposition, Mech. Syst. Signal
Process. 21 (3) (2007) 13461361.
[17] H. Hu, C. Wu, Development of scanning damage index for the damage detection of plate structures using modal strain energy method, Mech. Syst.
Signal Process. 23 (2) (2009) 274287.
[18] M. Radzieski, M. Krawczuk, M. Palacz, Improvement of damage detection methods based on experimental modal parameters, Mech. Syst. Signal
Process. 25 (6) (2011) 21692190.
[19] W. Fan, P. Qiao, A strain energy-based damage severity correction factor method for damage identification in plate-type structures, Mech. Syst. Signal
Process. 28 (2012) 660678.
[20] D. Wu, S.S. Law, Damage localization in plate structures from uniform load surface curvature, J. Sound Vib. 276 (2004) 227244.
[21] F.N. Catbas, M. Gul, J.L. Burkett, Damage assessment using flexibility and flexibility-based curvature for structural health monitoring, Smart Mater.
Struct. 17 (2008) 112.
[22] S.H. Sung, H.J. Jung, H.Y. Jung, Damage detection for beam-like structures using the normalized curvature of a uniform load surface, J. Sound Vib. 332
(6) (2013) 15011519.
[23] Z. Zhang, A.E. Aktan, Application of modal flexibility and its derivatives in structural identification, J. Res. Nondestruct. Eval. 10 (1) (1998) 4361.
[24] A.K. Pandey, M. Biswas, Damage detection in structures using changes in flexibility, J. Sound Vib. 169 (1) (1994) 317.
[25] W. Lestari, P. Qiao, S. Hanagud, Curvature mode shape-based damage assessment of carbon/epoxy composite beams, J. Intell. Mater. Syst. Struct. 18 (3)
(2007) 189208.
[26] C.P. Ratcliffe, Damage detection using a modified Laplacian operator on mode shape data, J. Sound Vib. 204 (3) (1997) 505517.
[27] P. Qiao, W. Lestari, M. Shah, J. Wang, Dynamics-based damage detection of composite laminated beams using contact and noncontact measurement
systems, J. Comput. Math. 41 (10) (2007) 12171252.
[28] C.P. Ratcliffe, W.J. Bagaria, Vibration technique for locating delamination in a composite beam, AIAA J. 36 (6) (1998) 10741077.
[29] C.P. Ratcliffe, A frequency and curvature based experimental method for locating damage in structures, J. Vib. Acoust. 122 (3) (2000) 324329.
[30] M.K. Yoon, D. Heidera, J.W. Gillespie Jr., C.P. Ratcliffe, R.M. Crane, Local damage detection using the two-dimensional gapped smoothing method,
J. Sound Vib. 279 (2005) 119139.
[31] M. Cao, P. Qiao, Novel Laplacian scheme and multiresolution modal curvatures for structural damage identification, Mech. Syst. Signal Process. 23 (4)
(2009) 12231242.
[32] M. Cao, W. Xu, W. Ostachowicz, Z. Su, Damage identification for beams in noisy conditions based on Teager energy operator-wavelet transform modal
curvature, J. Sound Vib. 333 (6) (2014) 15431553.
[33] M. Cao, M. Radzienski, W. Xu, W. Ostachowicz, Identification of multiple damage in beams based on robust curvature mode shapes, Mech. Syst. Signal
Process 46 (2) (2014) 468480.
[34] L.J. Hadjileontiadis, E. Douka, A. Trochidis, Fractal dimension analysis for crack identification in beam structures, Mech. Syst. Signal Process. 19 (3)
(2005) 659674.
[35] H. Li, Y. Huang, J. Ou, Y. Bao, Fractal dimension-based damage detection method for beams with a uniform cross-section, Comput.-Aid. Civ. Infrastruct.
26 (3) (2011) 190206.
[36] J. Wang, P. Qiao, Improved damage detection of beam type structures using uniform load surface, Struct. Health Monit. 6 (2) (2007) 99110.
[37] S. Naguleswaran, Vibration of an EulerBernoulli beam on elastic end supports and with up to three step changes in cross-section, Int. J. Mech. Sci. 44
(12) (2002) 25412555.
[38] M.A. Koplow, A. Bhattacharyya, B.P. Mann, Closed form solutions for the dynamic response of Euler-Bernoulli beams with step changes in cross
section, J. Sound Vib. 295 (1) (2006) 214225.
[39] J.W. Jaworski, E.H. Dowell, Free vibration of a cantilevered beam with multiple steps: comparison of several theoretical methods with experiment,
J. Sound Vib. 312 (4) (2008) 713725.
[40] Z.Y. Shi, S.S. Law, L.M. Zhang, Structural damage localization from modal strain energy change, J. Sound Vib. 218 (5) (1998) 814833.
[41] S. Choi, S. Park, N. Stubbs, Nondestructive damage detection in structures using changes in compliance, Int. J. Solids Struct. 42 (2005) 44944513.
[42] E. Sazonov, K. Powsiri, Optimal spatial sampling interval for damage detection by curvature or strain energy mode shapes, J. Sound Vib. 285 (4) (2005)
783801.
[43] M. Chandrashekhar, R. Ganguli, Damage assessment of structures with uncertainty by using mode-shape curvatures and fuzzy logic, J. Sound Vib.
326 (3) (2009) 939957.
[44] Michael J. Katz, Fractals and the analysis of waveforms, Comput. Biol. Med. 18 (3) (1988) 145156.

Das könnte Ihnen auch gefallen