Proceedings of the XXVI Iberian Latin-American Congress on Computational Methods in Engineering CILAMCE 2005

Brazilian Assoc. for Comp. Mechanics & Latin American Assoc. of Comp. Methods in Engineering
Guarapari, Espı́rito Santo, Brazil, 19th –21st October 2005

Paper CIL25-0445

EVALUATION OF NUMERICAL AND EXPERIMENTAL TECHNIQUES FOR

DAMAGE DETECTION

Carlos Cristiano Hasenclever Borges

cchb@lncc.br
Laboratório Nacional de Computação Cientı́fica
Petrópolis - RJ - Brasil

Flávio de Souza Barbosa

Alexandre Abrahão Cury
Anderson Rodrigues Vilela
flavio@numec.ufjf.br
acury@numec.ufjf.br
vilela@numec.ufjf.br
NUMEC - Núcleo de Pesquisa em Métodos Computacionais em Engenharia
Universidade Federal de Juiz de Fora
Juiz de Fora - MG - Brasil

Christian Cremona
cremona@lcpc.fr
Laboratoire Central des Ponts et Chaussées
Paris - France
Abstract. Structural systems are usually subjected to degradation processes due to a
combination of causes, such as: design or constructive problems, unexpected loading or
a natural process of degradation due to the structures ageing. The effects of that dete-
rioration results in damaged regions whose main characteristics are localized stiffness
decreases of the structure.
A common procedure presented in several papers on damage detection to evaluate
those localized damages is done using measured vibrational data of the structure as nat-
ural frequencies and vibration modes in damaged and/or undamaged states of the struc-
ture.
Firstly, the present study analyzes the sensibility of the structural vibration charac-
teristics as a function of geometrical and quantitative damage variations. Such analysis is
done through experimental tests of controlled damages in a cantilever beam. The effects
of the localization and magnitude of damages on the structural natural frequencies vari-
ation is then estimated and a profile of the sensibility of the vibration data is developed.
Secondly, the efficiency of a damage detection technique based on the strain energy
deviation between damaged and undamaged structural vibration modes is evaluated by
means of the experimental tests.
Finally, to circumvent the case of reduced set of experimental data, the applied dam-
age detection technique is evaluated through a proposed methodology of damage identi-
fication with incomplete modal data.

Keywords: Damage Identification, Dynamic Analysis, Vibrational Data, Interpolation

 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

1. INTRODUCTION
The assessment of structural damage via measured vibration data is the subject of
several recent studies due to its wide practical applications (Vilela et al, 2003) and (Al-
vandi et al, 2002). The problem consists in detecting and localizing the affected region as
well as determining its level of damage.
An useful and robust damage detection method must detect structural damages taking
into account the difficulties inherent to the experimental data acquisition. For example,
the obtainment of a representative and sufficient values of vibration data of the structure,
i.e., the mode shape in all the extension or volume of the structure is one of the main
difficulties in dynamic data acquisition. Such information is necessary to the knowledge
of the geometric characteristics of the vibrational modes, and the accuracy of the damage
detection model. Another inconvenient of measured vibration data is the presence of
noise, which must be treated by the damage identification model with an adequate level
of robustness.
An experimental cantilever beam is constructed and vibrational data are obtained al-
lowing tests involving a real structure and all the process of experimental data acquisition
and manipulation and its implication. Besides, a comparison with a computational model
that represents the cantilever beam using numerical techniques will analyze the accuracy
that one can expect in this type of simulation.
In the sequence, this work analyzes the sensibility of the structural vibration char-
acteristics due to quantitative damage variations. This analysis is performed through ex-
perimental tests of controlled damages in the experimental cantilever beam and by means
of computational simulations. After that, the efficiency of a damage detection technique
based on the strain energy deviation between damaged and undamaged structural vibra-
tion modes is evaluated by means of the experimental tests.
Finally, numerical experiments are performed using interpolation strategies as an at-
tempt to increase the numbers of experimental data available in order to augment the
robustness of the proposed method for several levels of incomplete vibrational measured
data.

2. DAMAGE ASSESSMENT METHODS

There are several methods of damage assessment available in the literature (Doebling
et al, 1996). Variations in vibrational data has been a consensus in terms of its potential
to indicate that some process of damage is occurring in a structural system. However,
the direct analysis of the variations in natural frequencies and mode shapes do not only
usually brings the necessary and confident information to asseverate the real damage state.
Efficient manipulation of those data is mandatory to enhance the quality and robustness
of the results obtained in damage assessment.
Here, one shows the direct variation of vibrational data in experimental and computa-
tional cases to analyze the viability of using those changes as damage indicator. Moreover,
two damage identification models that seek a better utilization of the vibrational data are
introduced and tested if they can provide more efficiency in damage detection problem.
Thus, one shortly describes or referrers these tools as following:
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

2.1 Changes of dynamic characteristics

When one uses computational structural models as part or complement of the damage
identification process one needs to obtain numerically the vibrational data values. The
dynamic properties inherent to a structural system without damping (natural frequency
and mode shapes) are defined by the eigenvalue problem as described by the equation:

(K − wi2 M)φi = 0, (i = 1, 2, ..., n) (1)

where K and M are, respectively, the stiffness and the mass matrices of the structural
system; wi are the natural frequencies in (rad/s) associated to the vibration modes φi and
n is the number of modes considered in the analysis. To obtain the numerical results of
Eq. (1) one usually uses the Finite Element Method (FEM) discretization (Zienkiewicz et
al, 2000).
Assuming that the introduction of damages in modelling affects directly the stiffness
matrix without modifying the mass matrix, the structural vibrational data variations may
be used as a damage indicator.
That is one of the simplest methods for damage detection that is based on the analysis
of the dynamic characteristics changes due to a stiffness alteration (or damage) (Barbosa
et al, 2004).

2.2 Modal Assurance Criterion - MAC

The Modal Assurance Criterion is often applied to damage evaluation (Alvandi, 2004)
and its efficiency is directly connected with the number of sensored points in the structure.
The MAC’s coefficients express the existing correlation between two measured mode
shapes. In Eq. (2), [φA ] and [φB ] are matrices made up of mA and mB modes, respec-
tively, measured in n points of the structure, and the MAC’s coefficients are defined by:
¯ Pn ¯
j ¯2
¯ i
k=1 [φA ]k [φB ]k
M AC ij = Pn i 2
Pn j 2 (2)
k=1 ([φA ]k ) k=1 ([φB ]k )

where [φA ]ik and [φB ]jk are, respectively, the k th component of the modes [φA ]i and [φB ]j .
The MAC’s coefficient varies from 0 to 1, where 1 means a perfect correlation (Alle-
mang et al. 1982).

2.3 Strain Energy Method - SEM

This method has already been used in engineering (Stubbs et al, 1992) and (Cornwell
et al, 1999) and biomechanics applications (Vilela et al, 2003) and is based on the analysis
of the mode shape strain energy variation of the structural flexural elements before and
after the damage. The strain energy of a Bernoulli-Euler beam is given by:
Z L µ ¶2
1 ∂ 2v
U= EI(x) dx (3)
2 0 ∂x2

where U is the deformation energy, EI(x) is the flexural rigidity of the beam, L is the
beam’s length and v is the beam’s vertical deflection. For a particular mode shape, φi (x),
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

the energy associated with that mode shape is:

Z µ 2 ¶2
1 L ∂ φi
Ui = EI(x) dx (4)
2 0 ∂x2
If the beam is subdivided into Nd regions, then the energy associated with each region j
due to the ith mode is given by:
Z µ 2 ¶2
1 aj+1 ∂ φi
Uij = EI(x) dx (5)
2 aj ∂x2

where aj and aj+1 are the limits of the region j. The energy fraction is defined as the
ration between the deformation energy of the beam and the energy of the element, given
by Eq. (6) and Eq. (7):
Uij
Fij = (6)
Ui
and
Nd
X
Fij = 1 (7)
i=1

Similar quantities can be defined for a damaged structure and are given as:
Uij∗
Fij∗ = (8)
Ui∗
Z µ 2 ∗ ¶2
∗ 1 L ∗ ∂ φi
Ui = EI (x) dx (9)
2 0 ∂x2
Z µ 2 ∗ ¶2
∗ 1 aj+1 ∗ ∂ φi
Uij = EI (x) dx (10)
2 aj ∂x2
where (*) indicates a quantity calculated using the damaged mode shapes, φi .
In the case of small damages, the first order approximation gives:

Fij∗ = Fij +superior order terms (11)

One obtains than:

R aj+1 ³ ´
³ 2 ∗´ RL
∂ 2 φ∗i ∂ φi
Uij∗ /Ui∗
aj
EIj∗ (x) 0
EI ∗
∂x2
(x) ∂x2
dx/
dx
1= = Ra ³ 2 ´ RL ³ ´ (12)
Uij /Ui j+1 2
EIj (x) ∂∂xφ2i dx/ 0 EI(x) ∂∂xφ2i dx
aj

By considering that the damage is small and localized, Eq. (12) may be simplified as
follows:
R aj+1 ³ ∂ 2 φ∗i ´ R L ³ ∂ 2 φ∗i ´
ˆ j
EI aj ∂x2
dx/ 0 ∂x2 dx
= ³
R aj+1 ∂ 2 φi ´ R L ³ ∂ 2 φi ´ (13)
ÊIj∗ 2 dx/ 2 dx
aj ∂x 0 ∂x
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

where ÊIj and ÊIj∗ can be obtained using mean value theorem (Stubbs et al, 1992).
By calculating the average value for the number of modes n used one obtains:
R aj+1 ³ ∂ 2 φ∗i ´ R L ³ ∂ 2 φ∗i ´
ˆ
EI j 1 Xn
aj ∂x 2 dx/ 0 ∂x2
dx
βj = = ³ ´ ³ ´ (14)
ÊIj∗ n i=1 R aj+1 ∂ 2 φi dx/ R L ∂ 2 φi dx
aj ∂x2 0 ∂x2

where βj is the damage indicator of the region j. In order to have more representative
values and to be able to compare them in various examples, one standardizes the damage
indicator like:
βj − β
Zj = (15)
σβ

where β and σβ represent the mean and standard deviation of β, respectively.

3. ON THE CONSIDERATION OF INCOMPLETE VIBRATIONAL MODE SHAPES

As defined by Eq. (14), the SEM is based on the relation between the mode shapes of
damaged and undamaged structure. Those data are integrated in the structure spam. The
obtained scalar for each element of the discretized structure, by means of FEM, indicates
if the element presents or not some damage level.
For that, it is necessary the integration in both structure states with all translation
degrees of the discretized structure. Such data are obtained experimentally and in the
greater number of the cases, it is impossible to obtain all of them. Limitations in the
equipment, difficulties in accessing some places of the structure are problems that need to
be tackled in practice to obtain the translational mode shapes components.
One possible solution for that is to expand the experimental vibrational modes in or-
der to complete the missing components. In this way, the integration defined by equation
Eq. (14) becomes viable to be calculated.
Expansion or contraction techniques for the degrees of freedom of the discretized
structure are well-known in damage identification problems (Doebling et al, 1996). The
main inconvenience of those approach is the error introduced by the approximation strat-
egy. Depending on the damage identification model used, the introduced error can become
impracticable an accurate damage identification and in some cases, the method becomes
incapable to assess the structural damage.
In this paper two strategies for mode expansion are implemented in order to obtain
the missing translational degrees of freedom of the mode shapes.
Firstly, one uses cubic spline interpolation. This method is applied by parts allowing
the use of a low degree of the polynomial and is very useful when one has a lot of points
to be interpolated. The low order of the polynomial used keeps the computational costs
reasonable and avoids oscillations that usually appears in high order polynomial degrees.
In cubic splines the first and second derivatives are continue in internal points. The second
technique used is the hermitian cubic polynomial interpolation. As the same way, this
method use a third degree polynomial to adjust the missing data. Both the interpolation
methods applied are defined in the INTERP1 routine for unidimensional interpolation of
MATLAB software.
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

4. MODAL IDENTIFICATION
As previously mentioned this work has an experimental part to obtain measurements
for a real structure what allows to compare with results obtained by computational simu-
lations. In practice, the equipments used in data acquisition measure longitudinal strains
which one can transform in vertical deflections. After that, it is necessary to apply effi-
cient methods that will manipulate these data to obtain the vibrational data. One uses for
that task the algorithms described as following.

4.1 Random decrement method

The technique of the random decrement is a temporal nonparametric technique, very
simple and its completely described in references (Asmussen, 1997) and (Cole, 1971).
The method is very fast in certain cases, even 100 times faster than the algorithm of
the Fourier Fast Transformation (FFT) (Brincker et al, 1994). In this work the random
decrement method is employed in order to filter the noise from dynamic measurements
and also to approach the transient response.

4.2 Ibrahim time domain method

The Ibrahim method is a temporal technique of modal identification based on a tran-
sient response. Using two time-shifted matrices deduced from the transient response,
Ibrahim presents an identification technique of modal parameters of the structure. By
solving an eigenvalues problem, the modal parameters of the structure can be estimated.
Details about this technique may be found in (Ewins et al, 2000).

5. EXPERIMENTAL TESTS
5.1 Description of the tested beam
The present study analyzes the dynamic behavior of the cantilever beam shown in
Fig. 1. This beam was instrumented with strain-gages and its dimensions are presented in
Fig. 2(a) and Fig. 2(b). The mechanical properties of the beam are:
• Material: aluminium.
• Density (ρ): 2700 kg/m3 .
• Young’s modulus (E): 70 GPa.
Seven equidistant strain-gages were bounded to the beam as shown in Fig. 1. The
strain-gages were placed in the measurement points B to H, defining eight regions (1 to 8).
Damages were imposed on region 4 (see Fig. 2(a)). The dynamic tests were performed
using an impulsive excitation applied to point I. The modal components in points A and
I were calculated assuming no vertical deflection in point A and no longitudinal strain
in point H. Firstly, the undamaged beam was tested and then, using a drilling machine,
four successive controlled damage levels were imposed on region 4, and the structure was
re-tested for each damage level. Figures 3(a) to 3(d) show the controlled damages (holes).
The equipment employed in the tests is listed in the following:
1
• Seven 120Ω strain-gages connected in 4
Wheatstone bridge.
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

Figure 1: Tested beam.

Strain-gages

A B C D E F G H I

19 mm
1 2 3 4 5 6 7 8

80 mm 75 mm 75 mm 75 mm 75 mm 75 mm 75 mm 75 mm
600 mm

Damage Localization

(a) Tested beam - Superior view.

3,2 mm

19 mm

(b) Tested beam - Cross-

section.

Figure 2: Tested beam characteristics.

• PCI A/D converter model 6013, Low Cost series, manufactured by National Instru-
ments.

• A Pentium IV computer.

Using the differential equation of a beam, which relates vertical deflections v and
longitudinal strains ²x (Eq. (16)) and appropriate boundary conditions the modal compo-
nents, may be calculated in terms of the strain measurements, being h the beam depth.

d2 v 2²x
2
= (16)
dx h
Due to the equipment limitations, only the dynamic properties of the first vibration
mode shape were measured and analyzed.
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

(a) Damage Level 1: one 1mm-hole. (b) Damage Level 2: one 2mm-hole.

(c) Damage Level 3: one 1mm-hole and one (d) Damage Level 4: two 2mm-hole.
2mm-hole.

Figure 3: Damage levels 1, 2, 3 and 4.

6. NUMERICAL TESTS
There are some goals that one intends to achieve by means of the numerical tests per-
formed. One of them is to analyze the equivalence between the damage state of a struc-
ture obtained by two different ways: experimental measurements where one achieves the
vibrational data in experimental tests as explained in the previous section and a compu-
tational correspondent model where the vibrational data are obtained by means of FEM
discretization (Zienkiewicz et al, 2000). The real and theoretical model adjustment is al-
ways difficult and tests involving both models are usually useful to provide new insights
for better system identification techniques. Besides, the real model has the additional
inconvenience of tackling noised data.
Another important analysis is concerned to the comparison between two damage
identification strategies namely as MAC and SEM. Both are based on the structural modal
shapes but developed with different approaches. Quantitative values of vibrational data
variation before and after the damage process are included in the tests expecting to show
the level of influence of the damage over this type of data and its reliability to use in
damage assessment.
Finally some expansion techniques applied on the vibrational data obtained experi-
mentally and simulated computationally are carried out. Two types of interpolation are
tested and the main question to be analyzed is if it worths to introduce new points by
means of interpolation despite of the approximation error that is also introduced.

Methodology
As previously described, the experiments are executed in a cantilever beam built in
laboratory and computationally reproduced by means of FEM discretization using two-
dimensional frame elements. Based on the number of the strain-gages used to monitoring
the real beam, a natural option adopted for the discretized cantilever beam is shown in
Fig. 4. One notices it is composed by 8 elements.
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

NODES ELEMENTS

1 2 3 4 5 6 7 8

80 mm 75 mm 75 mm 75 mm 75 mm 75 mm 75 mm 75 mm
600 mm

Figure 4: Discretized cantilever beam.

The strategy to introduced the damage in the computational model is to reduce not
only the inertia but also reducing the cross-section area of the damaged element of the
computational modeled beam. The reduction of these parameters is obtained by an equiv-
alence element method shown in Fig. 5.
For each of the two first damage levels considered in the experimental tests (defined
by the hole size) represented by level 1 and level 2, i.e, only one hole is performed, a
different diameter (D) was adopted for a posterior definition of the correspondent variable
in FEM element, (L) as defined in Fig. 5. For the case when one has two holes (level 3
and level 4) the same procedure is adopted but taking into account the cumulative effect
referent to the reduction of each L value relative to each diameter (D1 and D2 ), as shown
in Fig. 6.
19 mm 19 mm
3,2 mm

3,2 mm
D
m

m
m

m
75

75

Figure 5: Equivalence element method - Levels 1 and 2.

19 mm 19 mm
3,2 mm

3,2 mm

D2
D1
m

m
m

m
75

75

Figure 6: Equivalence element method - Levels 3 and 4.

Considering that:
Vhole = Vprism (17)
For levels 1 and 2:
πD2 πD2
× 3, 2 = L × 3, 2 × 75 → L = mm (18)
4 300
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

For levels 3 and 4:

π(D12 + D22 ) π((D12 + D22 )
× 3, 2 = L × 3, 2 × 75 → L = mm (19)
4 300
which leads to:

Reduced Cross-Section Area= [(19 − L) × 3, 2] mm2 (20)

Reduced Inertia= [(19 − L) × 3, 23 ]/12 mm4 (21)

The quantitative effect of the damage over the vibrational data is realized in two ways:
by the direct analysis of natural frequencies variation and by performing MAC in the
experimental and computational models vibrational data obtained. The level of variation
in both evaluations indicates if these measurements are useful to show the presence of
damage in structures.
The SEM is used over the same data acquired from the experimental tests and com-
putational simulation in a attempt of not only showing the presence of a damage process
but also of localizing its position and give some insight about the damage value.
After that, the vibrational data are submitted to a interpolation using cubic splines
and hermitian cubic polynomial in a attempt of increase the precision of the damage
identification for both experimental and computational vibrational data.
As described above, all those tests are performed for the experimental case and the
correspondent computational model and have the goal of investigates the coherence be-
tween the real and the theoretical computational approach adopted.

7. RESULTS
First results are related to the analysis of the natural frequencies variation of the can-
tilever beam subject to damage. Figures 7(a) and 7(b) show, respectively, the variation
of the first natural frequency for experimental data and for numerical data obtained com-
putationally. In these figures, damage level ‘0’ (zero) represents the undamaged tested
structure.
The total frequency variation from the undamaged case to the fourth damage level is
inferior to 0,10%. Although this variation is almost insignificant it should allow to estab-
lish any correlation between the damage level and the natural frequency. This correlation
was clearly detected elsewhere in numerical (Barbosa et al, 2004) and experimental works
(Alvandi, 2004). The computational simulations show that increasing the damage level,
the natural frequency decreases. The same result is not achieved for the experimental data.
This fact is due to the equipment limitations and shows the importance of the precision in
the measurements which contributes for a noise reduction.
The use of MAC is carried out in advance. As one uses only one mode shape, MAC’s
coefficients reduce to only one number for each analysis. Using the undamaged mode
shape as [φA ] and the damaged mode shapes as [φB ] (see Eq. (2)) the graphics presented
in Fig. 7(c) and Fig. 7(d) are obtained for the experimental and computational cases,
respectively.
There are practically no variations with the MAC’s coefficients. These characteristics
were also detected in other works in the literature (Alvandi, 2004). But, as in the previous
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

test involving natural frequencies the computational case presents decreasing MAC values
for increasing damages scenarios in contrast with experimental case where no correlation
is detected.

7.024 7.312

7.022
7.3115

FIRST NATURAL FREQUENCY (Hz)

FIRST NATURAL FREQUENCY (Hz)

7.02

7.018 7.311

7.016
7.3105
7.014

7.012 7.31

7.01
7.3095
7.008

7.006 7.309
0 1 2 3 4 0 1 2 3 4
DAMAGE LEVEL DAMAGE LEVEL

(a) First natural frequency variation - Experimental (b) First natural frequency variation - Computational
case. case.

1.00000005 1.00000000

1.00000000 0.99999995

0.99999995 0.99999990

0.99999990 0.99999985
MAC

MAC

0.99999985 0.99999980

0.99999980 0.99999975

0.99999975 0.99999970

0.99999970
0.99999965

0.99999965
0 1 2 3 4 0.99999960
0 1 2 3 4
DAMAGE LEVEL DAMAGE LEVEL

(c) MAC’s coefficients - Experimental case. (d) MAC’s coefficients - Computational case.

Figure 7: Damage levels 1, 2, 3 and 4.

In the attempt to localize the damage region one performs tests using SEM for the
numerical and computational mode shapes. According to reference (Alvandi, 2004), the
SEM coefficients are damage indicators for Z values greater than 0,5.
Figure 8(a) and Fig. 8(b) show the results for the experimental and computational
case for the four damage levels used. Observing Fig. 8(a), one can notice that the region 4
has SEM coefficients (Z) greater than 0,5 for all damage levels, showing that this damage
detection method is capable to localize damage regions in this type of problem. On the
other hand, the quantitative value of damage is not identified. Similar results are obtained
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

in (Alvandi, 2004) for SEM. One can also observe in Fig. 8(a) and Fig. 8(b) that the
accuracy to concentrate the value of Z different from zero only in region 4 is greater for
the computational case where the numerical case spread with more intensity high Z values
in the neighborhood of region 4, probably due to imprecise measures and consequently
mode shapes.

LEVEL 1 LEVEL 1
LEVEL 2 LEVEL 2
1 LEVEL 3 2 LEVEL 3
LEVEL 4 LEVEL 4
0.8
1.5
0.6
Z

0.4 1

Z
0.2
0.5

0
0
1
2 1
3 2
4 3
5 4
6 5
REGIONS 6
7
8 4 7 4
3
ELEMENTS 8 3
2 2
1 1
DAMAGE LEVEL DAMAGE LEVEL

(a) SEM’s coefficients - Experimental case. (b) SEM’s coefficients - Computational case.

Figure 8: Damage levels 1, 2, 3 and 4.

The SEM showed a reasonable capacity to indicate where the damage is located in
the structure if one considers that only values of Z greater than 0.5 represents structural
damage. However, mainly in the experimental case other undamaged regions present
values close from 0.5 for the Z values possibly due to noise in the mode shapes. This
fact can bring difficulties and confusion for the damage area definition of a structure.
To circumvent that diminishing the imprecision in the damaged regions identification one
expands the number of experimental obtained points in an attempt to augment the samples
in the real damage region to concentrate high values of Z for those new points and avoid
that the undamage points have significants Z values. The inconvenience is that the new
sample points present some approximation error due to interpolation process that can
misguide the search for the damage parameters.
In a first test one duplicates the number of initial experimental points producing a
16-element discretized cantilever beam. One can observe the results for hermitian cubic
interpolation in Fig. 9(a) and 9(b) for experimental and computational case, respectively.
Figures 10(a) and 10(b) show the same when one uses the cubic spline for interpolation.
The damage now is not only in element 4 but in elements 7 and 8 of the new discretization.
The graphics indicate a better and correct concentration of the damage region in compar-
ison to the 8-element beam as expected. The improvement of the damage assessment
quality is noticed for the experimental case and mainly for the computational case that
has reliable vibrational data, i.e., without noise. That shows the importance of strategies
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

to reduce the noise level in experimental measurements for efficient damage identification
(Barbosa et al, 2001).

LEVEL 1 LEVEL 1
1.5 LEVEL 2 2.5 LEVEL 2
LEVEL 3 LEVEL 3
LEVEL 4 LEVEL 4
2
1
1.5
Z

Z
1
0.5

0.5

0 0
0 0

5 5

10 10

ELEMENTS 15 ELEMENTS 15
4 4
3 3
2 2
20 1 20 1
DAMAGE LEVEL DAMAGE LEVEL

(a) SEM’s coefficients - Experimental case. (b) SEM’s coefficients - Computational case.

Figure 9: Cubic - Double of experimental points

LEVEL 1
LEVEL 1 2.5 LEVEL 2
1.5 LEVEL 2 LEVEL 3
LEVEL 3 LEVEL 4
LEVEL 4 2

1 1.5
Z

1
0.5
0.5

0
0
0
0

5
5

10
10

ELEMENTS ELEMENTS 15
15
4
4 3
3 2
2 20 1
20 1
DAMAGE LEVEL DAMAGE LEVEL

(a) SEM’s coefficients - Experimental case. (b) SEM’s coefficients - Computational case.

Figure 10: Spline - Double of experimental points

Finally, an interpolation is applied to generate the triple of original points, i.e., a

beam with 24 elements. Thus, the damage is located in elements 10, 11 and 12 of the new
discretization. The results can be observed in Fig. 11(a) and 11(b) for hermitian cubic
interpolation and in Fig. 12(a) and 12(b) for cubic spline interpolation. One notice that
artificially increases in the numbers of points that will perform the damage identification
produce an enhance in the quality of the response when SEM is used. Moreover, one
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

observes that both types of interpolation produce results in similar damage identification
level.

2
LEVEL 1 2.5 LEVEL 1
LEVEL 2 LEVEL 2
LEVEL 3 LEVEL 3
1.5 2 LEVEL 4
LEVEL 4

1.5
1
Z

Z
1
0.5
0.5

0
0
0 0

5 5

10 10

15 15
ELEMENTS ELEMENTS
20 20
34 34
25 12 25 12
DAMAGE LEVEL DAMAGE LEVEL

(a) SEM’s coefficients - Experimental case. (b) SEM’s coefficients - Computational case.

Figure 11: Cubic - Triple of experimental points

1.5 LEVEL 1 2.5 LEVEL 1

LEVEL 2 LEVEL 2
LEVEL 3 LEVEL 3
LEVEL 4 2 LEVEL 4

1
1.5
Z
Z

1
0.5

0.5

0 0
0 0

5 5

10 10

15 15
ELEMENTS
20 20
ELEMENTS 34 34
25 12 25 12
DAMAGE LEVEL DAMAGE LEVEL

(a) SEM’s coefficients - Experimental case. (b) SEM’s coefficients - Computational case.

Figure 12: Spline - Triple of experimental points

 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

8. CONCLUSIONS
Damage identification techniques are evaluated in this paper using experimental data
measured in a cantilever beam and providing subsidy for comparison with theoretical
computational model. This type of analysis is very useful to furnish insights to improve
the process of system identification and to diagnose the main influences of the errors
that an experimental experience is subjected to. One notices that the level of accuracy in
damage assessment is highly influenced by data errors/noise inherent to equipment and
measurements in experimental vibrational data acquisition.
Only the variation of structural dynamic properties, shown that is not an efficient
damage indicator even if one uses some processing tool as MAC. Despite of that, vi-
brational data variation has enormous potential and are considered the best indicator of
changes in the structural properties. An algorithm that tackles these variations in a effi-
cient way can achieve good performance in the damage identification.
An algorithm named SEM is used to obtain a certain robustness level in the damage
assessment. The SEM showed the ability to locate the damage using the translational
degrees of mode shapes included with very small damage scenarios.
In order to circumvent the lack of experimental data that is generally obtained due to
difficulties in real simulations and to augment the capability of SEM to localize the dam-
aged regions one introduces an interpolation process using cubic splines and hermitian
cubic polynomial. Both present similar behavior and increased the level precision in the
damage assessment.
The difficult of SEM to quantify the damage level can be treated by a two-step
process, for example where, after well-defined the damage location by SEM, another
technique, like genetic algorithm (Barbosa et al, 2001), will focus only in the determina-
tion of the damage intensity.

Acknowledgements
The authors would like to thank UFJF (Universidade Federal de Juiz de Fora), FAPEMIG
(Fundação de Amparo à Pesquisa do Estado de Minas Gerais) and CNPq (Conselho Na-
cional de Desenvolvimento Cientı́fico e tecnológico) for the financial support.

REFERENCES
Allemang R.J., Brown D.L., 1982, “A Correlation Coefficient for Modal Vector Analysis.”
Proceedings of IMAC 1.

Alvandi, A., Cremona, C., 2002, “Reliability of bridge integrity assessment by dynamical
testing”, First European Workshop on Structural Health Monitoring, Ecole Normale
Supérieure de Cachan.

Alvandi, A., 2004, “Contribution a lutilization pratique de levaluation dynamique pour la

detection dendommagements dans les ponts.” Ph.D. thesis, cole National des Ponts et
Chausses, Marne la Vale, France.

Asmussen, J.C., 1997. “Modal Analysis Based on the Random Decrement Technique -
 CILAMCE 2005 – ABMEC & AMC, Guarapari, Espı́rito Santo, Brazil, 19th – 21st October 2005

Applications to Civil Engineering Structures”. Ph.D. thesis, Aalborg University, Dane-

mark.

Barbosa, F. S.; Borges, C.C.H and Cury, A. A., 2004. “Modeling of structural damage
identification based on variation of modal characteristics of structures.” XXV CILAMCE
- Iberian Latin-American Congress on Computational Method in Engineering, Recife,
Brazil. (in portuguese)

Barbosa, H.J.C, Borges, C.C.H., 2001, “A Genetic Algorithm for Damage Identification
in Framed Structures Using Vibration Data”, IX International Symposium on Dynamic
Problems of Mechanics, pp 47-52, Florianópolis, SC.

Brincker, R.; Krenk S. and Jensen J.L., 1994. “Estimation of correlation functions by the
random decrement technique.” Proceedings of IMAC 10, 610-615.

Cole, H.A.; 1971. “Failure Detection of a Space Shuttle Wing by Random Decrement.”
NASA TMX - 62,04.

Clough, R.W., Penzien, J., 1993, “Dynamic of Structures”, McGraw-Hill.

Cornwell, P.J., Doebling, S.W., Farrar, C.R., 1999, “Application of the Strain Energy
Damage Detection Method to Plate-Like Structures”, Journal of Sound and Vibration,
vol.224, n.2, pp.359-374.

Doebling, S.W., Farrar, C.R., Prime, M.B., Shevitz, D.W., 1996, “Damage identification
and health monitoring of structural and mechanical systems from changes in their vibra-
tion characteristics: a literature review”, Technical Report LA-1307-MS, Los Alamos
National Laboratory, New Mexico.

Ewins, D. J.; 2000. “Modal Testing: Theory, Practice and Application.” Research Studies
Press.

Vilela J., Ribeiro H. C., Guevara-Jr. N.O., Barbosa F. S., Barra L. P. S., 2003,
“Investigação do processo de ósseo-integração de implantes dentários através de re-
spostas dinâmicas”, XXIV CILAMCE Iberian Latin American Congress on Computa-
tional Methods, Ouro Preto, MG Brasil.

Stubbs, N.; Kim J.-T. and Topole K.; 1992. “An efficient and robust algorithm for dam-
age localization in offshore platforms.” Proceedings of the ASCE Tenth Structures
Congress, 543-546.

Zienkiewicz, O.C., Taylor, R.L., 2000, “The Finite Element Method”, Butterworth-
Hienemann.