Download as pdf or txt
Download as pdf or txt
You are on page 1of 9

)$&8/7(,772(*(3$67(:(7(16&+$33(1

'(3$57(0(17%85*(5/,-.(%28:.81'(
$)'(/,1*%28:0(&+$1,&$
:'(&52</$$1
%+(9(5/((

.$7+2/,(.(
81,9(56,7(,7
/(89(1

6WDWXV

- 0$(&. 0 $%'(/ :$+$% * '( 52(&.  'DPDJH GHWHFWLRQ LQ UHLQIRUFHG
FRQFUHWH VWUXFWXUHV E\ G\QDPLF V\VWHP LGHQWLILFDWLRQ 3URFHHGLQJV ,60$ 
1RLVHDQG9LEUDWLRQ(QJLQHHULQJ/HXYHQ%HOJLXPSS6HSWHPEHU

,5-0$(&.

7(/   )$;  


(PDLOMRKDQPDHFN#EZNNXOHXYHQDFEH
Damage detection in reinforced concrete structures by
dynamic system identification

J. Maeck, M. Abdel Wahab, G. De Roeck


Department of Civil Engineering, division Structural Mechanics, K.U.Leuven, Belgium
e-mail : johan.maeck@bwk.kuleuven.ac.be

Abstract
Service loads, environmental and accidental actions may cause damage to constructions. Regular inspection
and condition assessment of engineering structures are necessary so that early detection of any defect can be
made and structure’s remaining safety and reliability can be determined. When the structural damage is
small or it is in the interior of the system, its detection cannot be done visually. A useful more elaborate non-
destructive evaluation tool is vibration monitoring. It relies on the fact that occurrence of damage or loss of
integrity in a structural system leads to changes in the dynamic properties of the structure. In this paper,
different techniques will be presented and compared to derive from experimentally determined modal
characteristics of a reinforced concrete beam its dynamic bending stiffness. The degradation of stiffness, due
to the cracking of the reinforced concrete, gives information on the position and intensity of the occurred
damage.

information is needed to localize uniquely the


1. Introduction damage.
Besides this updating technique, also a direct
In the framework of developing a nondestructive stiffness calculation of the dynamic stiffness is
vibration testing method for monitoring the presented based on the quotient of the modal
structural integrity of constructions in civil bending moment to the corresponding modal
engineering, it is important to be able to determine curvature. Herein, besides experimental
the dynamic stiffness in each section of the eigenfrequencies, also modeshapes are needed to
structures from measured modal characteristics. The calculate curvatures from.
damaged structure results in a dynamic stiffness
reduction of the cracked sections. 2. Test program & data processing
From the dynamic stiffnesses, one obtains
directly an idea of the extension of the cracked An experimental program is set up to establish the
zones in the structure. The dynamic stiffness relation between damage and changes of the
reduction can also be associated with a degree of dynamic system characteristics. In the test program,
cracking in a particular zone. Determining the a concrete beam of 6 meter length is subjected to an
damage parameters from modal parameter shifts increasing static load to introduce cracks.
belongs to the group of inverse problems [4]. The test beam is subjected to two identical point
After a short description of the test set-up, the loads (symmetric at 1m from the middle). After
results of a sensitivity based updating technique is each load step (6 in total, point load being
presented. The proposed updating technique uses respectively 4, 6, 12, 18, 24 and 25.3 kN), an
only measured eigenfrequencies since these are experimental modal analysis is performed on the
sensitive indicators of structural integrity [9]. beam. A free-free set-up for the beam is established
However methods using only eigenfrequencies by using very flexible springs to support the beam
cannot distinguish damage at symmetrical locations (rigid body frequencies << first eigenfrequency).
in a symmetric structure [3]. For a damage pattern Accelerometers were placed each 20cm at both
in a symmetric structure, also modeshape longitudinal edges of the upper side of the beam (62
measurement points in total).
m3
m4
The sensitivity matrix, S, is defined as the first
K2
0.85m
derivative of the modal parameters (z) with respect
m7

0.5m
1.0m
to the unknown parameters (θ), i.e.;
m6 m9 hinge support
0.5m
K1 0.5m

 ∂z1 ∂z1 ∂z1 ∂z1 


0.5m

L
m2

 ∂θ ∂θ l 
1.0m

∂θ 2 ∂θ 3
m1 0.85m m8

 1 
∂z 2 ∂z 2 ∂z 2
L
∂z 2 
 ∂θ 1 ∂θ 2 ∂θ 3 ∂θ l 
hydraulic jack
m5

displacement
[ S ] nxl = ∂z 3 ∂z 3 ∂z 3 ∂z 3  (2)
 L 
roller support transducer

 ∂θ 1 ∂θ 2 ∂θ 3 ∂θ l 
L L L L L
∂z n ∂z n ∂z n ∂z n 
 ∂θ L
 1 ∂θ 2 ∂θ 3 ∂θ l 

A finite difference approximation is used to


calculate the elements of the sensitivity matrix.
Each column of the sensitivity matrix is computed
Fdyn from two finite element analyses for two θ values
differing by an incremental amount δθ. Therefore,
l+1 analyses are required to compute the sensitivity
matrix. The use of the finite difference
Figure 1 Static & dynamic set-up
approximation provides usually gradients with
acceptable accuracy especially when the gradient is
The stochastic subspace identification technique is not small. The perturbation value used in this
applied to the dynamic response of the beam in approximation should be carefully chosen in order
order to extract the modal parameters [8]. to minimize the truncation, condition and round-off
errors [6].
3. Updating technique Rewriting equation (1) in more detail:

[ zm - z j ] nx1 = [ S j ] nxl [θ j +1 - θ j ] lx1 (3)


The main goal of updating techniques is to achieve
a good agreement between experimental and
calculated numerical modal parameters. The used zm represents the measured modal parameters used
method minimizes in an iterative manner the in the updating algorithm. The subscript j indicates
differences between numerical and experimental the iteration number at which the sensitivity matrix
modal parameters. The resulting least-square is computed. The number of modal parameters (e.g.
problem is solved by the Gauss-Newton method. eigenfrequencies, mode shapes and/or modal
Practical implementation of the Gauss-Newton curvatures) is equal to n, while l is the number of
method relies upon the application of the singular unknown parameters to be updated.
value decomposition. The Gauss-Newton equation To be able to compute θj+1 from equation (3), the
is similar to the truncated Taylor series used in the inverse of the matrix Sj should be calculated. If S is
penalty function method [5]. The penalty function not square, the inverse of S does not exist and the
equation can be written in the following form: pseudo inverse S✝ is used, which can be calculated
by solving the singular value decomposition of S,
S δθ - δ z = 0 (1) i.e.;
where δz is the discrepancy between the measured
modal data and the finite element solution. δθ is the y 0
[ S ] nxl = [U ] nxn  kxk 
T
nxl [V ] lxl (4)
perturbation in the unknown parameters to be  0 0
updated. S is the Jacobian or sensitivity matrix
containing the first derivative of the calculated where k ≤ min (n,l), U and V are unitary matrices,
modal parameters (z) with respect to the unknown i.e., UT U = I and VT V = I.
parameters (θ).
The pseudo inverse of S is then computed as: parameter β characterizes the relative length of the
damaged zone. It ranges between 0 to 1. If β
 y −1 0
[ S + ] lxn = [V ] lxl  kxk T
 lxn [U ] nxn (5) becomes small, a very local damage at the middle of
 0 0 the beam is obtained, while if β is equal to 1, the
Now, the solution of equation (3) is: beam is damaged over its whole length. α
characterises the relative magnitude of damage. If α
[θ j +1 ] lx1 = [θ j ] lx1 + [ S j+ ] lxn [ zm - z j ] nx1 (6) is equal to 1, no damage is present, whereas if α
drops to zero, the bending stiffness vanishes at the
In equation (6), it is assumed that the measured middle of the beam. The third parameter is the
data are equally weighted. The lower frequencies power n which characterizes the variation of the E-
are in general more accurately determined than the modulus from the centre of the beam (x=0) to the
higher ones. Therefore, in the updating algorithm end of the damaged zone (x=βL/2). If n is larger
the lower natural frequencies should be given more than 1, a flat damage pattern is produced, otherwise
weight than the higher. To account for this fact, a steep pattern is obtained.
equation (3) is pre-multiplied by the weighting By using this proposed function, not only the
matrix Wz , i.e., updating parameters are reduced to three but also a
realistic damage pattern is always guaranteed. It
[Wz ] nxn [ z m - z j ] nx1 = [Wz ] nxn [S j ] lxn [θ j +1 - θ j ] lx1 (7) should be noted that a symmetric damage pattern is
assumed since in the present application the beam is
The matrix Wz is a diagonal matrix containing loaded symmetrically. It is of course easy to account
the weight factors for each modal component. for a non-symmetric damage pattern by assuming
The updating algorithm described in this section two different stiffness variations at the left and the
is implemented in the finite element program right hand side of the centre line of the beam. In
ANSYS [2] and the matrix laboratory software such a case, the number of updating parameters will
MATLAB [7]. be increased from 3 to 5 parameters (α, (β1, n1) for
In order to detect damage along the beam length the left-hand side and (β2, n2) for the right hand
using model updating, one possibility is to update
side).
the bending stiffness of each element in the finite
It should be noted that in transferring the damage
element model. This means that a large number of
function to the finite element model, β is used to
updating parameters are required to properly
calculate the number of damaged beam elements.
describe the variation of the bending stiffness along
the beam length. Consequently, a considerable
amount of computational time is needed to calculate
the sensitivity matrix (2). Besides, due to
measurement and/or discretization errors a realistic
damage pattern is not always guaranteed if the
bending stiffness of each element can vary
independently. Therefore it would be advantageous
to build a function that can describe a damage
pattern by only few representative parameters and
has the flexibility to represent small as well as large
damage zones. Assuming that the reduction in the
bending stiffness can be simulated by a reduction of
the E-modulus, the following function is proposed: Figure 2 Damage function

[ ] π x 
n Before applying any static load to the beam, a
E = E0 1 − (1 − )cos 2
t with t =   (8)
2  β L 2 
dynamic test is performed. This test serves as a
reference for later comparison at the different
where β, α and n are the damage parameters. L is damage stages. A finite element model containing
the beam length and x is the distance along the 30 beam elements is constructed. The results of this
beam measured from the centre line. A sketch of the initial model are fitted in a global way to the
proposed function is shown in Figure 2. The reference test results. This can be done by either
updating the E-modulus or the density of the whole
beam. In Table 1, the measured natural frequencies 4. Direct stiffness calculation
for the first four bending modes are given for the
reference test of the beam and compared to the The direct stiffness calculation uses the
finite element results. The E-modulus is adapted to experimental modeshapes in deriving the dynamic
give the best agreement with the test results. It can stiffness. The big advantage is that no numerical
be seen that the difference between the finite model is needed to obtain the dynamic stiffness
element results and the measured natural distribution for statically determined structures. For
frequencies is in general less than 1%. This means hyperstatic structures, the reaction forces and
that the beam model represents the measurement consequently the internal forces are dependent on
quite well and can be used with confidence for the stiffness of the structure. Therefore an iterative
damage detection after each load step. procedure is applied to a numerical model to find
the EI distribution of a hyperstatic structure.
Mode 1 2 3 4
Measured 21.9 60.3 117.0 192.0 The method makes use of the basic relation that the
F.E. 22.16 60.66 117.74 192.13 dynamic bending stiffness in each section is equal
% difference 1.1 0.6 0.6 0.06 to the bending moment in that section divided by
Table 1 Natural frequencies – reference test – the corresponding curvature. The eigenvalue
E=38.0 109 N/m2, ρ=2500 kg/m3 problem for the undamped system can be written as:
K ϕ =ω2 M ϕ (9)
The updating algorithm is now applied to the beam
in which K is the stiffness matrix, M the (analytical)
at the different load steps. The natural frequencies
mass matrix, ϕ the measured modeshape and ω the
of the first four bending modes are used as modal
measured eigenpulsation. This can be seen as a
parameters (vector z) in the updating algorithm.
pseudo static system: for each mode internal
The three damage parameters β, α and n are
(section) forces are due to inertial forces which can
considered as updating parameters (vector θ) be calculated as the product of local mass and local
making a 4×3 sensitivity matrix ([S]4×3). β is acceleration (= ω2.ϕ). A lumped mass matrix is used
converged to values less than 1.0 for the first three in (9). As the measurement mesh is rather dense,
load steps indicating that the whole beam is not yet this is acceptable. Alternative is using the consistent
damaged. At load step 4, β reaches 1.0 already. α mass matrix, which is not done in the present paper.
varies from 0.85 (15% damage) at step 1 up to 0.49 The contribution of rotational inertia is proven to be
(51% damage) at step 6. n converged to values negligible in the present case.
between 1.4 to 2.2. Figure 3 shows the evolution of Because of the free-free set-up in the dynamic
damage along the beam length for the different load tests, the internal forces should be in static
steps. equilibrium. Due to measurement errors, this is not
7
exactly the case. Therefore a Gram-Schmidt
beam 3
6.5
x 10
orthogonalization is applied to the experimental
6
modeshapes. The corrected ϕ' can be calculated
from:
5.5

x1T M ϕ’ = 0
5 (10)
x2T M ϕ’ = 0
EI

4.5

4
step 1 with ϕ' = ϕ + a.x1 + b.x2 and x1, x2 respectively the
step 2
step
step
3
4
translational and rotational rigid body mode of the
3.5 step
step
5
6
free beam. Equation (10) is in fact nothing else than
3
imposing the total vertical and moment equilibrium
0 1 2 3 4 5 6
length (m) of the inertial forces of the beam.
From (9), using the orthogonalized modeshape,
Figure 3 Stiffness degradation for all loadsteps
the bending moment in each section is now
determined.
The next step in deriving the dynamic bending
stiffness consists of the calculation of the curvatures
along the beam for each modeshape. Direct respectively the first derivatives of displacements
calculation of curvatures from measured and rotations are minimized. The weight of these
modeshapes, e.g. by using the central difference extra conditions is set by the penalty factors α and
approximation, results in oscillating and inaccurate β.
values. Advantages of this Mindlin approach is that
A smoothing procedure accounting for the directly curvatures are available, boundary
inherent inaccuracies of the measured modeshapes conditions can be imposed easily (in this
should be applied. Therefore a weighted residual experimental set-up curvatures at the free beam
penalty-based technique is adopted which closely ends have to be zero) and the approximated modal
resembles the finite element approach. deflections have not to go through all measurement
points. A drawback is the difficulty to choose
The beam is divided in a number of elements appropriate penalty factors. Too high values cause
separated by nodes corresponding to the locking of the system. Therefore alternatives with
measurement points. Each node has 3 degrees of quadratic displacement, linear rotation and constant
freedom: the modal displacement va, the rotation ψ curvature interpolations have been studied.
and the curvature κ, which are approximated In the following, α and β are chosen in such way
independently (Figure 4). Linear shape functions that the median of the relative error on the modal
are used. deflections is 2-3%, which is a reasonable
estimation of the anticipated measurement
v =v N +v N inaccuracy. Hence, every modeshape has its own
a a,1 1 a,2 2
penalty factors.
= N + N
1 1 2 2 Figure 5 shows for the first modeshape the
= N + N different finite element variables along the beam
1 1 2 2
axis in the reference (undamaged) state. Figure 5a
Figure 4 Finite element variables shows the approximated vs. the experimental
modeshape (+), figure 5b the relative error, figure
This is analogous to the Mindlin plate element, for 5c and 5d the modal rotations and curvatures.
which the rotations are approximated independently
from the bending deflection. median 0.021
The objective function, which has to be 0.3 0.1

minimized, contains the difference between 0.2 0.08


relative error
modeshape

0.1 0.06
approximative and measured modeshapes. Two
0 0.04
penalty terms are added to enforce continuity of -0.1 0.02
rotations and curvatures in a mean, smeared way. -0.2 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6
-a- -b-

(v a − ϕ m ) 2 /e 2 dv
∫ ∫
0.3 0.14

= dx + ( − a ) 2 dx 0.2 0.12

2 2 dx (11)
0.1 0.1
curvature
rotation

0 0.08

/ e4
d 2
+ ∫ ( −
-0.1 0.06
) dx minimum -0.2 0.04
2 dx -0.3 0.02

-0.4 0

ϕm denotes the measured modeshape, and L is the e 0 1 2 3


-c-
4 5 6 0 1 2 3
-d-
4 5 6

length of a finite element. Elements are chosen in


such way that nodes coincide with measurement Figure 5 va vs. ϕm , relative error, ψ and κ for
points. 1st bending mode of beam.
The first term expresses that the average
difference between approximation and measurement To obtain the dynamic bending stiffness for the
has to be minimized. Without any other terms in the reference state, one should calculate the bending
potential function one will find a (piecewise linear) moments. Figure 6 shows the results for the first
approximation through all measurement points. In bending mode. Figure 6a shows the measured and
order to obtain a filtering of experimental errors and the (almost identical) orthogonalized modeshape,
so a smoothing of the deflection, two extra terms according to (10), plotted upon each other. Figure
are added. Differences between the (independent) 6b shows the distributions of inertia forces
approximations of rotations and curvatures with according to (9) from which the modal shear forces
and bending moments can be determined (Figure 6c 7.E+06

& 6d). Dividing now the sectional bending moments 6.E+06

by the modal curvatures (Figure 6e, from Figure 5),


5.E+06
one obtains the dynamic bending stiffness (Figure
6f). 4.E+06

EI (Nm2)
3.E+06
5
x 10
0.6 2

0.4 2.E+06
reference step1
modeshape

inertia force 1
0.2 step2 step3
0 1.E+06
0 step4 step5
-0.2 -1 step6
0 1 2 3 4 5 6 0 1 2 3 4 5 6 0.E+00
5 5
x 10 -a- x 10 -b-
5 10
2 2.5 3 3.5 4
bending moment
shear force

7.E+06
0 5

6.E+06
-5 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6
-c- x 10
6
-d- 5.E+06
0.2 8

EI (Nm2)
0.15 6 4.E+06
dynamic EI
curvature

0.1 4

0.05 2 3.E+06
0 0
0 1 2 3 4 5 6 0 1 2 3 4 5 6 2.E+06
-e- -f- reference step1
step2 step3
1.E+06
step4 step5
Figure 6 EI for 1st bending modes of beam. 0.E+00
step6

2 2.5 3 3.5 4
7.E+06
From Figure 6 it can be noticed that at the sections
6.E+06
of almost-zero moments (or almost-zero curvatures),
5.E+06
the approximation for EI is not accurate anymore.
Higher modes have even more sections with zero
EI (Nm2)

4.E+06

curvatures (Figure 7). Combining information of 3.E+06


different modes can give a solution.
2.E+06
reference step1
step2 step3
8.E+06 1.E+06
step4 step5
step6
0.E+00
7.E+06 2 2.5 3 3.5 4
7.E+06

6.E+06 6.E+06
EI (Nm2)

5.E+06
5.E+06
4.E+06
EI (Nm2)

4.E+06 Mode 1
3.E+06
Mode 2
Mode 3
Mode 4 2.E+06
3.E+06 reference step1
0 1 2 3 4 5 6 1.E+06 step2 step3
step4 step5
step6
Figure 7 EI for first 4 bending modes (reference 0.E+00
2 2.5 3 3.5 4
state)
Figure 8 Dynamic stiffness degradation for first
The Figures 5, 6 & 7 dealt with the reference four bending modes
undamaged state without any static preload. The
same can be done after each load step, to examine The beam zone from 2m to 4m is a zone of almost
the dynamic (bending) stiffness degradation. The constant static bending moment, which should result
evolution of the dynamic bending stiffness through in the same cracking and consequently the same
different loadsteps is shown in Figure 8. In each dynamic stiffness. Due to the own weight of the
plot information of one mode is used. beam the bending moments in the middle are
slightly higher and so the reduction of the dynamic
stiffness. It can be noticed that the stiffness
degradation derived from the third and certainly effects due to damage from those due to
from the fourth bending mode is less accurate. environmental changes. As mentioned earlier, to
Results from the first two bending modes are detect asymmetric damage in a symmetric structure
comparable, except from the central zone for the in a unique manner, also modeshape information is
second mode, which is inaccurate due to zero by necessary.
zero division. On the contrary, the direct stiffness calculation
In Figure 9 the degraded dynamic bending makes use of the experimental modeshapes to
stiffness is plotted versus the bending moment of derive the dynamic stiffness from modal curvature
the preceeding static test, using updating as well as calculation. The advantage of the latter method is
direct stiffness calculation results. The own weight that no numerical model is needed to obtain the
is also accounted for in the moment calculation. dynamic stiffness distribution. However a rather
Figure 9 shows that after the first loadstep, the dense measurement grid is necessary to be able to
bending stiffness already decreases by about 20 %. identify accurately the curvatures of the higher
modes.
When results of the updating are compared to
6.5E+06
those from the direct stiffness method, one can
6.0E+06 notice that both methods estimate a dynamic
5.5E+06
stiffness decrease of around 50% after the ultimate
loadstep.
EI (Nm2)

5.0E+06

4.5E+06

4.0E+06 Acknowledgements
direct stiffness calculation using 1st mode
3.5E+06 direct stiffness calculation using 2nd mode
sensitivity based updating This work was carried out in the framework of
3.0E+06
FKFO-project No. G.0243.96, supported by the
0 10 20 30 40 50 Belgium National Fund for Scientific Research.
M static (kNm)

Figure 9 Dynamic stiffness degradation vs. static References


load history
1. Abdel Wahab M., De Roeck G., Effect of
temperature on dynamic system parameters of a
The stiffness degradation calculated from the direct highway bridge, Structural Engineering
stiffness method corresponds well with updating International, No.4, pp.266-270, 1997.
results. 2. ANSYS revision 5.3, Swanson Analysis
System, 1994.
5. Conclusion 3. Choy, F.K., R. Liang, P. Xu, Fault
Identification of Beams on Elastic Foundation,
The present paper describes two techniques to Computers and Geotechnics, Vol.17, pp. 57–
calculate the stiffness degradation of a damaged 176, 1995.
reinforced concrete beam. The starting point of the 4. Doebling S.W., Farrar C.F., Prime M.B.,
techniques is completely different. Shevitz D.W., Damage identification and health
Since the updating techniques are using only monitoring of structural and mechanical
experimental eigenfrequencies, which can be systems from changes in their vibration
cheaply acquired, this approach provides an characteristics: a literature review, Research
inexpensive structural assessment technique. One report LA-13070-MS, ESA-EA, Los Alamos
accelerometer at a well-chosen location can provide National Laboratory, Los Alamos, New
the eigenfrequency information. Mexico, 1996.
A drawback in that case is that no information is 5. Friswell, M.I. and Mottershead, J.E., Finite
available on correspondence of the numerical to the element model updating in structural dynamics,
experimental modeshapes. Another disadvantage of KLUWER Academic Publisher, Dordrechts,
using only eigenfrequencies is that these are also The Netherlands, 1995.
sensitive to environmental influences [1]. This
means that a procedure is needed to separate the
6. Gill, P., Murray, W. and Wright, M., Practical
optimisation, Academic Press Limited, San
Diego, 1981.
7. MATLAB revision 5.2, The Mathworks Inc.,
1998.
8. Peeters B., Abdel Wahab M., De Roeck G., De
Visscher J., De Wilde W.P., Ndambi J.-M.,
Vantomme J., Evaluation of Structural Damage
by Dynamic System Identification, Proceedings
of ISMA21, Vol.3, Leuven, Belgium, 1349-
1361, 1996.
9. Salawu O.S., Detection of structural damage
through changes in frequency: a review, Eng.
Structures, Vol.19, No. 9, pp.718-723, 1997.

You might also like