Documente Academic
Documente Profesional
Documente Cultură
www.elsevier.com/locate/compstruc
Solid Mechanics Laboratory, Department of Mechanical Engineering, Federal University of Rio de Janeiro, Ilha do Fundao,
P.O. Box 68503, Rio de Janeiro, RJ 21945-970, Brazil
Received 1 July 2001; accepted 9 January 2002
Abstract
This paper introduces a structural identication technique built on nite element (FE) model updating. The FE
model is parameterized by a structural parameter that continuously describes the damage in the structure, and besides,
an evolution equation of this damage parameter is presented. The model updating is accomplished by determining the
subset of this damage parameters that minimizes a global error derived from the dynamic residue vectors, which is
obtained by introducing the experimental modal properties into the original model eigenproblem. A mode-shape
projection technique is used in order to achieve compatibility between the dimension of the experimental and analytical
models. The adjusted model maintains basic properties of the analytical model as the sparsity and the symmetry, which
plays an important role in model updating-based damage identication. The verication and assessment of the current
structural defect identication is performed on a analytically derived bidimensional truss structure and on a cantilever
bidimensional EulerBernouilli beam through a virtual test simulator. This simulator is used to realistically simulate the
corrupting eects of noise, ltering, digital sampling and truncation of the modal spectrum. The eigensystem realization
algorithm along with the common-based normalized system identication were utilized to obtain the required natural
frequencies and mode shapes. 2002 Elsevier Science Ltd. All rights reserved.
Keywords: Continuum damage model; Damage identication; Finite element model updating
1. Introduction
It is needless to emphasize the importance of damage
detection techniques and health monitoring in aerospace, civil and mechanical engineering. It is essential to
determine the safety and reliability of their systems and
structures. Based on experimental modal analysis and
signal processing techniques, monitoring and interpreting changes on structural dynamic measurements can be
considered as a quite promising approach for damage
0045-7949/02/$ - see front matter 2002 Elsevier Science Ltd. All rights reserved.
PII: S 0 0 4 5 - 7 9 4 9 ( 0 2 ) 0 0 0 1 5 - 9
418
2. Theoretical foundations
The present section is devoted to the construction of
the modelling for the dynamics of damaged structures,
which will provide the theoretical basis for the defect
identication method introduced in this work.
2.1. A continuum damage model
In the present modelling, besides the classical variables that characterize the kinematics of a continuum
medium (displacements and velocities of material
points), an additional scalar variable b 2 0; 1 is introduced. This variable is related with the links among
material points and can be interpreted as a measure of
the local cohesion state of the material. This eld of
cohesion states describes the current state of damage
within the mechanical system as it evolves in time and
possesses as extreme values b 0 and 1. Therefore, if at
a certain time t, after a period of evolution of the system,
b 1, all the links and the initial material properties
have been preserved. On the other hand, if b 0 a local
rupture is considered since all the links among material
points have been broken. The variable b is associated to
the damage variable D [32], by the following relation:
b 1 D. As the degradation is an irreversible phenomenon, the rate b_ must be negative or equal to zero. A
detailed presentation of the basic principles that govern
the evolution of such kind of continuum damage can be
found in [10,23,24]. Just a summary of those principles is
presented here. For the sake of simplicity, the hypotheses of isothermal processes are adopted for the damage
evolution. Besides, it is also assumed the hypothesis of
small deformation and, consequently, the conservation
of mass principle is automatically satised.
From now on the variables in bold will be used to
denote vectors and the variables between square brackets will be used to denote tensors.
Let a body B that occupies a region X R3 with a
suciently regular boundary C be subjected at each time
instant t 2 I to external forces g : C2 I ! R3 and
b : X I ! R3 , to external microscopic forces p : X
I ! R, q : C2 I ! R and T
to prescribed displacements
S
ux; t 8 x 2 C1 , where C1 C2 ; and C1 C2 C.
Therefore, the principle of virtual power can be expressed as:
pint pext 0
C2
C2
3
where p is dened as a microscopic distance force while q
is a microscopic contact force, both in duality with b.
The microscopic forces are related to non-mechanical
actions (chemical and electromagnetic, for instance) that
can cause an evolution of the damage. So, substituting
(2) and (3) in (1) yields
Z
Z
qu ^u r r^u b ^u dX
g ^u dA
X
C2
Z
Z
H rb^ F b^ pb^ dX
qb^ dA 0;
8 ^u 2 Vv ; 8b^ 2 Vb
r
m
E
tre2 e e
21 m 1 2m
w kb C b_ k _
b
1
ktre2 le e w kb C b_ kb_
2
H krb
6
7
C2
Z
1
krb rb^ dX
kdiv u2 le e w b^ dX
2
X
X
Z
9
C b_ b^ dX 0; 8 b^ 2 Vb
C2
419
b 2 0; 1
and b_ 6 0
420
It is worthwhile mentioning that the damage detection method, that will be introduced later on, can be
developed from many of the damage models that already exist [32]. Indeed, at a rst sight, the method
builds simply on the kinematic parameterization of the
FE model resulting of the use of the damage model. The
need of the evolution equations, like those just presented
above, is often due to either of the two typical issues: the
development of the damage scenario which can be used
as an initial guess for the iterative numerical solution of
the damage detection problem or the analysis of the
residual life of the structure upon the damage distribution obtained by means of the proposed technique. This
complete set of equations describing the non-linear
evolution of the damage can also be explored in situations where no control can be applied to the input
signals and, as a consequence, the usual laboratorial
conditions are not accomplished. This is typical in several industrial applications. In those cases a non-linear
identication scheme in the time domain, like the one
proposed in [13], based on a realist modelling like the
one summarized in Eqs. (8) and (9) could be used.
Another important aspect that must be remarked is
highlighted in Eqs. (8) and (9), where one can see that
only the elastic terms are aected by the damage eld.
This will imply, in the nite element context, that the
mass matrix remains unchanged during the damage
evolution.
2.2. Basic equations
One can summarize the present non-destructive
damage detection procedure by four main steps: modal
testing [7], location of potentially damaged regions, assessment of severity and the real extension of the damage, and assessment of the validity of the provided result
by means of some error measures.
Generally speaking, a linear n-DOF nite element
model of an undamaged structure is represented by the
following equation of motion
M
u Ku 0
10
j 1; . . . ; nd
13
11
where the eigenvalue ki is the square of the natural frequency corresponding to vibration mode /i . Now, replacing the analytical modal parameters by experimental
ones into the eigenproblem (11), a dynamic residue
vector is dened over each identied mode:
i
Ri K x2E;i M /E i 1; . . . ; nm
12
nm
X
Ri Ri
14
i1
where nm is the number of modes to be used in the updating procedure and nd is the number of nodes used to
discretize the damage eld.
T
i
i
2x2E;i M/E K/E
15
where xE;i and /E are the experimental modal properties. The set of parameters fb1 ; . . . ; bnd gT that minimizes
the problem (13) is given by the solution of the optimality condition presented below
F
oGR
0 such that bp 2 0; 1
oB
p 1; . . . ; nd
16
421
18
20
where the operator S is, essentially, the projection dened over each component of B and reads as
Sv fv if v 2 0; 1; 1 if v > 1; 0 if v < 0g
21
422
min Ri Ri
24
i
/o
the residue R
Ri Zi
m
25
26
/o
T
iT i i
/i
o Zo Zo / o
27
Z
Z
Z
Z
28
/i
o
m /m Poi /m
o
o
o
where meansTthe Moore Penroses inverse, inasmuch
i
as the matrix Zi
o Zo can be singular and does not have
a classic inverse. Therefore, recalling Eq. (22), one can
rewrite the full ith mode shape in terms of its matrix
projection Pi and its measured partition /i
m as follows
(
)
i
/m
I
i
/i
29
f/i g
m Pi /m
i
P
oi
/o
After the projection operator Pi for the ith mode shape
is determined, the mode shape may be expanded and the
dynamic residue Ri with respect to the model DOF can
be computed.
K KR /E Ri
i 1; . . . ; nm
31
Once the present approach has been properly introduced, and assuming that the experimental modal
properties have already been obtained, one may formulate this damage identication algorithm as follows:
1. expand the measured mode shapes;
2. calculate the initial dynamic residue vectors; select
the set of parameters to be updated;
3. start the updating process assuming that all the selected parameters have initial value equal to one;
4. expand the measured modes (except for the rst iteration);
5. calculate the vector F and the tangent matrix A;
6. calculate the increment DB;
7. apply the projection operator S;
8. update the vector B and the stiness matrix Kd ;
9. kDBk
< 1 ? kGRk < 2 ?
kBk
Yes ) end
No ) go to 4.
At each iteration, the experimental mode shapes need
to be expanded (step 4) as the stiness matrix has to be
updated. Each such expansion is obtained by means of
the ill-posed problem (24), which can lead to several
solutions. Here, the one selected is given by (28), what
can represent troubles to the convergence of the algorithm.
3.5. Error measures
After using the proposed method to obtain an updated analytical model, one has to assess the reliability
of the provided results. In this work, four error measures
are used for this purpose. The relative error between the
experimental and the analytical natural frequencies Ef ,
the weighted normalized H2 norm [12] of the dierence
between the experimental frequency response functions
(FRFs) and the analytical ones, the amplitude correlation coecient (ACC) [8] and a normalized time-error
(TE) measure.
The weighted normalized H2 norm of the dierence
between the experimental FRF matrix GExp and models
FRF matrix Gm is dened as follows:
W
kDGxkH2
W x GExp x Gm x
H2
GExp x
32
H2
0;
1;
jxj < XB
jxj P XB
33
423
34
YExp t Ym t
YExp t
35
424
1200 Hz
4096
0450 Hz
10
sine chirp
425
r2s
r2n
36
Table 2
Simulation parameters for the rst six cases
Case
SNR
(dB)
Measured DOF
(NODES)
Number
of samples
Number of
averages
4.C
90
4096
10
4.D
4.E
4.F
4.G
4.H
90
50
50
50
20
2,4,6,8,10,12,
14,16,18,20
2,6,10,14,18
2,6,10,14,18
2,6,10,14,18
2,5,8,14,21
2,6,10,14,18
4096
4096
16,384
4096
8192
10
10
10
10
100
426
other vertical DOF starting at node 2. It has been considered that the signal was contaminated with noise such
that the SNR was 90 dB (see Eq. (36)), which corresponds to a very accurate measure.
The selected parameters to be updated were based on
the components of the mean initial dynamic residue
associated to the vertical DOF, which is depicted in Fig.
6. Based on a statistical procedure [9], the parameters 4,
5, 6 and 8, which are associated with the greatest components of the mean residue vector, were selected to be
updated. It should be remarked that, in this case, this
remarkable selection of parameters is mainly due to the
great number of measured DOF. In a more realistic
situation, where one has to deal with a small set of experimental data, this selection is not so ecient.
The result provided by CDIA is depicted in Fig. 7,
where the thin line represents the imposed damage and
the thick line represents the result obtained by CDIA. It
is clear from Fig. 7, that CDIA was able to capture the
essence of the damage eld. However, it should be remarked that the validity of the provided result must be
based only on the error measures computed from the
available data. Hence, Fig. 8 shows the error indicators
associated to the case (4.C).
The relative frequency error indicator Ef shows that
all the model natural frequencies, except the second one,
have got closer to the experimental natural frequencies.
The ACC measure clearly shows an improvement of the
FRFs from the second natural frequency on, and that
the frequency band around the rst natural frequency has reached worst approximation. Nevertheless,
it should be remarked that, within this band, the poor
FRFs curve tting, due to precision errors in frequency
associated to the discrete fourier transform may be responsible to make the ACC error measure worse within
this band. The normalized TE also shows a better
agreement between the response of the damaged structure and that of the updated model than the response of
the original one. As a last check, one clearly sees that the
normalized dynamic residue matrix has been minimized.
Hence, all the error measures are positive with respect to
the response provided by the CDIA, i.e., all of them
shows that the damage identication performed by
CDIA have led to a better agreement between the updated model and the actual damaged structure, which
implies that the result can be considered satisfactory,
fact that agrees with Fig. 7.
The second situation (4.D) is exactly equal the previous one (4.C), except that here one has half the sensors
such that the measured DOF are the vertical ones of the
following nodes: 2, 6, 10, 14 and 18. The result provided
by CDIA is shown in Fig. 9 and the error measures for
this case are shown in Fig. 10.
It is clear from Fig. 9 that the present result does not
describe the damage eld as well as it has been described
in the previous example, for the same algorithm stopcriterium. The weighted normalized H2 before and after
the updating process are respectively: 0.90889 and
0.76580, what means a positive result. Despite the fact
that one can clearly see that the result presented in Fig. 7
is more accurate than the one presented in Fig. 9 it is
hardly noticed by comparing Figs. 8 and 10. Therefore,
it is really dicult to assure that, in this situation, one
solution is more consistent than the other one. This
fact may be considered as predictable as long as it is
well known that, in general, the dynamic behavior of
a structure is slightly changed due to damage. According to the set of error measures depicted in Fig. 10, the
result shown in Fig. 9 would be perfectly accepted as
a satisfactory solution for the damage identication
problem.
The third case (4.E) to be analyzed is equal to the case
(4.D) except that the SNR has been set equal to 50 dB.
The result provided by CDIA is depicted in Fig. 11 and
its respective error measures are shown in Fig. 12. After
427
428
than the one shown in Fig. 11. Nevertheless, the comparison of the error measures for the situations (4.C)
and (4.E) slightly indicates that the solution in Fig. 7 is
better than the one presented in Fig. 11.
The fourth case (4.F) entirely corresponds to the
previous one (4.E), except that the FRFs were obtained
with 16,384 samples in time (see Table 1), fact that led to
a higher denition of the FRFs for the same band of
interest (0450 Hz). Since the normal mode shapes and
natural frequencies required for the present approach
come from the ERA algorithm, whose the required data
are the impulse response functions obtained out of the
measured FRFs, the mode shapes and natural frequencies can be dierent from the ones obtained in case (4.E).
The result obtained by CDIA is graphed in Fig. 13 and
its error measures are depicted in Fig. 14.
The weighted normalized H2 norm before and after
the updating process are respectively: 0.76766 and
0.29434. As it was expected, since changes in the dynamic behavior of the structure due to damage are slight
and the set of available experimental data (FRFs) is
more precise in this case than in the foregoing one, the
result provided in Fig. 13 is a better approximation of
the damage eld than the solution depicted in Fig. 11,
which is noted when comparing the Ef error measure in
Figs. 12 and 14 and it is even more clear when comparing the ACC measure or the weighted normalized H2
norm.
It is well known that after performing an experiment
if one decides to do a new one, such as keeping everything the same but the position of sensors, can be really
simple or almost not practical. However, despite the
diculties inherent to this process, a last assessment of
the identied damage eld can be obtained performing a
new experiment, if it is possible. Therefore, the damage
identication result depicted in Fig. 11 will be assessed
by a new experiment (4.G), where the sensors are relocated at more suitable positions, chosen such as to en-
x 2 x0 L; x0 3:25L
37
429
430
431
5. Concluding remarks
range encompassing the second frequency may contribute to this result. The same conclusion may be drawn
with respect to the rst frequency in the case (4.I).
An approach for detecting damage based on a continuum damage model and using partial experimental
432
433
434
Fig. 21. Result obtained by CDIA for the case 4.J with 20 elements.
Fig. 22. Result obtained by CDIA for the case 4.J with 40 elements.
Acknowledgements
Support for this research was partially provided by
Brazilian Spacial Agency (AEB) under contract 011/98
and National Council of Scientic and Technological
Development (CNPq).
Appendix A. Elemental stiness matrix for a bi-dimensional EulerBernouilli beam based on the continuum
damage model
Fig. 24 depicts a well known bi-dimensional Euler
Bernouilli beam element where the DOF are ordered as
follows
ue fvi hi vj hj gT
Fig. 23. Result obtained by CDIA for the case 4.J with 10 elements.
The method has been shown to be ecient for estimating damage parameters for the presented situations.
A:1
12I
bi bj
L2
B 4IL 2bi bj
B 12I
@ 2 b b
i
j
L
4I
bi 2bj
L
435
[2] Alvin KF, Park KC. A second order structural identication procedure via state-space-based system identication.
AIAA J 1994;32:397406.
[3] Baruch M, Bar Itzhack IY. Optimum weighted orthogonalization of measured modes. AIAA J 1978;16(4):34651.
[4] Brinker R, Kirkegaard PH, Anderson P. Damage detection
in an oshore structure. In: Proceedings of the 13th IMAC,
1995. p. 6617.
[5] Castello DA. Structural damage identication via stiness
matrix updating based on a continuum damage model.
MSc Dissertation, Federal University of Rio de Janeiro,
Brasil, 1999 [in Portuguese].
[6] Doebling SW, Farrar CR, Prime MB, Sheritz DW.
Damage identication and health monitoring of structural
and mechanical systems from changes in their vibration
characteristics: a literature review. Los Alamos National
Lab., Report LA-13070-MS, Los Alamos, NM, May 1996.
[7] Ewins DJ. Modal testing: theory and practice. Taunton,
Somerset, England: Research Studies Press Ltd; 1989.
[8] Grafe H. Model updating of large structural dynamics
models using measured response functions. PhD Dissertation, Imperial College of Science, Technology and Medicine, University of London. 1998.
[9] Farhat C, Hemez FM. Updating nite element models
using an element-by-element sensitivity methodology.
AIAA J 1993;31(9):170211.
[10] Fremond M, Nedjar B. Damage, gradient of damage and
principle of virtual power. Int J Solids Struct 1996;33(8):
1083103.
[11] Frank PM. Fault diagnosis in dynamic systems using
analytical and knowledge-based redundancea survey and
some new results. Automatica 1990;26:45974.
1
4I
2bi bj
12I
bi bj
bi 2bj
L
L2
2I3bi bj 4IL 2bi bj
2Ibi bj C
C
12I
4I
4IL 2bi bj
b
bi 2bj A
i
j
L
L2
2Ibi bj
4IL bi 2bj 2Ibi 3bj
4I
L
where bi and bj are the nodal cohesion parameters associated to the nodes i and j of the element depicted in
Fig. 24. One should notice that this result is a particular
one due to the fact that one could have chosen any kind
of shape functions such as quadratic, trigonometric and
etc. Therefore, the user has enough freedom to choose
the most suitable discretization of the damage eld as
well as the shape functions for some specic problem,
what makes the proposed approach very attractive inasmuch as this choice, in principle, does not have any
connection with the discretization of the FE model of
the original system.
References
[1] Alvin KF. Method for determining minimum-order mass
and stiness matrices from modal test data. AIAA J
1995;33:12835.
436
[30]
[31]
[32]
[33]
[34]
[35]
[36]
[37]
toral dissertation. Department of Aerospace Engineering Sciences, University of Colorado, Boulder, February
2000.
Sampaio RP, Maia NM, Silva JM. Damage detection
using the frequency response function curvature method.
J Sound Vib 1999;226(5):102942.
Santos JV, Soares CM, Soares CA, Pina HL. Development
of a numerical model for the damage identication on
composite plates. Compos Struct 2000;48:5965.
Skrzypek J, Ganczarski A. Modeling of material damage
and failure of structures: theory and applications. Berlin:
Springer-Verlag; 1999.
Smith S. Iterative use of direct matrix updates: connectivity
and convergence. In: Proceedings of the AIAA 33rd
Structures, Structural Dynamics, and Materials Conference (Dallas, TX). Washington, DC: AIAA; 1992. p. 1797
806.
Smith S, Beatie C. Simultaneous expansion and orthogonalization of measured modes for structural identication.
In: Proceedings of the AIAA Dynamic Specialist Conference (Long Beach CA). Washington, DC: AIAA; 1993.
p. 26170.
Stutz LT. Damage identication approach via exibility
matrix adjustment based on a continuum damage model.
MSc Dissertation, Federal University of Rio de Janeiro,
Brasil, 1999 [in Portuguese].
Tsou P, Shen M-HH. Structural damage detection and
identication using neural networks. AIAA J 1994;32(1):
17683.
Zou Z, Link M. Identication of delamination in sandwhich plates using vibration test data. Proceedings of the
Identication in Engineering Systems, 1996. p. 5866.