You are on page 1of 9

DYNAMIC ASSESSMENT OF CURVED CABLE-STAYED BRIDGE BY

MODEL UPDATING
By James M. W. Brownjohn1 and Pin-Qi Xia2

ABSTRACT: Bridges are indispensable components of the infrastructure of modern society, and their assessment
via techniques of structural dynamics is assuming greater importance. This assessment concerns performance of
the as-built structure compared to the design and can also extend to the assessment of structural deterioration
or damage. Simple validation of numerical results by dynamic testing has met some success; however, feedback
from testing into analysis is usually crude, and only recently have systematic techniques been developed that
can be supplied to such structures. This paper investigates the application of sensitivity-based model updating
technology to the dynamic assessment of the Safti Link Bridge, a curved cable-stayed bridge in Singapore.
Downloaded from ascelibrary.org by New York University on 05/10/15. Copyright ASCE. For personal use only; all rights reserved.

Based on the measured modal data from prototype testing, the simulated dynamic properties obtained via finite-
element analysis have been significantly improved by modification of uncertain structural parameters such as
Youngs modulus of concrete and structural geometry.

INTRODUCTION only from the modeling errors resulting from simplifying as-
sumptions made in modeling the complicated structures but
Cable-stayed bridges with modern distinctive styles are in- also from parameter errors due to the uncertainties in material
creasing in number worldwide. These bridges are now built in and geometric properties and boundary conditions. The sim-
more unusual styles for structural and aesthetic reasons (Menn plifying assumptions depend on the intended application of the
1996; Rito 1996). Examples include the Lerez Bridge (Troy- model and on the experience of the engineer. However, given
ano et al. 1998), a single inclined tower bridge; the Katsushika an appropriate structural model, there are various methods
Harp Bridge (Takenouchi 1998), with a single pylon and S- (Mottershead and Friswell 1993) for forcing a match be-
shaped deck; the Marian Bridge (Kominek 1998), with a single tween analytical predictions and measured responses, some of
L-shaped pylon; the Alamillo Bridge (Casa 1995), with a sin- which involve direct updates to structural stiffness or mass
gle inclined pylon; and the Safti Link Bridge (Tan 1996), matrices. Modification of structural parameters having clear
which has a curved deck and single offset pylon. The unique physical significance is the preferred route.
structural styles of these bridges beautify the environment but This technique of model updating for correcting uncertain-
also add to the difficulties in accurate structural analysis. The ties from modeling, geometry, material, and analysis to im-
accurate assessment of these and other types of bridges using prove the analytical results based on the experimental modal
dynamics-based methods has become of increasing concern results has emerged in the 1990s as a subject of great impor-
due to their infrastructural role. tance for mechanical and aerospace structures. It has been de-
Dynamics-based assessment (Severn et al. 1989; Felber 1995; veloped to the point of becoming a standard structural analysis
Law and Ko 1995; Felber and Cantieni 1996) of these unusual and assessment tool for these structures. However, this updat-
bridges is based on a comparison of the experimental modal ing technology is still difficult to apply as a standard engi-
analysis (EMA) data obtained during full-scale tests with the neering tool for civil engineering structures, because of the
finite-element analysis (FEA) predictions. One purpose of the difficulties in prototype testing and experimental data analysis
comparison is so that the finite-element (FE) models can be resulting from the nature, size, location, and usage of these
used to predict performance during unusual loads such as earth- structures and the greater uncertainties in material physical
quakes (Dumanoglu et al. 1991; Brownjohn et al. 1992). Even properties. In particular, for field testing of a bridge, the tester
if the resulting response is large enough that material and ge- is at the mercy of varying environmental conditions not limited
ometric nonlinearities become significant, the starting point for to temperature and traffic. Without the benefit of the relatively
nonlinear analysis would be a realistic linear model. For ex- controlled and repeatable conditions of laboratory testing, the
ample, in the case of a suspension bridge, the large cable os- response parameters are likely to change. Moreover, when re-
cillations will vary the geometric stiffness, which can be accom- lying on ambient excitation, the requirements of nonstation-
modated in nonlinear analyses if the operating tension is known. arity for reliable spectral analysis are inevitably violated, and
Material nonlinearities can also be incorporated based on the a greater skill is required to recognize when a response is not
low level characteristics. Other motivations for dynamics-based a feature of the structure but of the excitation.
assessment include the validation of design assumptions em-
The FE model updating process has several functions in
bodied in the FE model and health assessment (i.e., the iden-
civil engineering structural dynamics. First, if the results of
tification of structural deterioration or damage).
updating processes are disseminated to the engineering com-
Confidence in using FE models for performance predictions
munity via case studies, they can serve as a guide for further
may be lacking due to relatively large differences between
modeling of similar structures. For example consistent over-
experimental and analytical modes. The differences come not
estimation of foundation fixity could lead to more realistic
1
design assumptions. Unless there is feedback from full-scale
Assoc. Prof., School of Civ. and Struct. Engrg., Nanyang Technol.
Univ., Nanyang Ave., Singapore 639798. E-mail: cjames@ntu.edu.sg
tests into the FE model, such an effect will not be unearthed.
2
PhD Student, School of Civ. and Struct. Engrg., Nanyang Technol. Because all civil engineering structures are prototypes, the
Univ., Nanyang Ave., Singapore 639798. E-mail: pp2906893@ntu.edu.sg benefit is not so immediate as is the case of mechanical and
Note. Associate Editor: David McCallen. Discussion open until July aerospace engineering where developed models can be applied
1, 2000. To extend the closing date one month, a written request must immediately to the study of production variants. Second, the
be filed with the ASCE Manager of Journals. The manuscript for this new improved model can be used for further appraisals of
paper was submitted for review and possible publication on April 5, 1999.
This paper is part of the Journal of Structural Engineering, Vol. 126, the actual structure tested (e.g., for predicting effects of un-
No. 2, February, 2000. ASCE, ISSN 0733-9445/00/0002-02520260/ usual loads or of making structural modifications). In addition,
$8.00 $.50 per page. Paper No. 20619. increased confidence in structural modeling and the conse-
252 / JOURNAL OF STRUCTURAL ENGINEERING / FEBRUARY 2000

J. Struct. Eng. 2000.126:252-260.


quent reductions in uncertainties may lead to more economic
designs. A final and increasingly important benefit is the po-
tential for use as a damage detection technique (Doebling et
al. 1998) and as a database for the bridge management systems
(Hawk and Small 1998; Hearn 1998).
It is only recently that the civil engineering community has
begun to adopt this advanced technology. Cantieni (1996) in-
vestigated model updating of a concrete arch bridge, and Pavic
et al. (1998) applied the technique to footbridges and concrete
floors. These investigations marked the beginning of the suc-
cessful application of the model updating technology to civil
engineering structures. The application of the model updating
technology to even more complex structures such as cable-
stayed bridges is still a challenge to the investigator.
Downloaded from ascelibrary.org by New York University on 05/10/15. Copyright ASCE. For personal use only; all rights reserved.

Model updating procedures can be classified as being one-


step procedures (global methods) or iterative (local methods).
Global methods directly reconstruct the updated global mass
and stiffness matrices from the reference data (measured fre- FIG. 1. Safti Link Bridge
quencies and mode shapes). Local methods are based on cor-
rections applied to local physical parameters of the FE model. ments, and the total width of the traffic lanes is 8 m with 2-
They have a good physical interpretation of the obtained mod- m walkways. The deck has a 180-m radius of curvature to the
ifications and preserve the symmetry, positive-definiteness, and centerline where nine bridge stay cables are attached at 10-
sparseness of the stiffness matrix. Global methods simply de- m centers, supporting the deck from the pylon, which is held
liver an updated matrix that regenerates the response data with in position by three pairs of back stay cables.
often no guaranteed preservation of the attributes of the ma- The concrete deck girder comprises a central 2-m-deep box
trix. The effective and most popular local methods for model with wing-tip girder plus 3 m of cantilever on each side.
updating are generally based on the sensitivity analysis. Fig. 3 shows the detail of the deck girder. Prestressed trans-
This paper investigates the dynamic assessment of the Safti verse diaphragms are provided at 5-m intervals to improve
Link Bridge, a curved deck, single offset pylon, cable-stayed torsional rigidity, to resist transverse bending, and to transfer
bridge that is a landmark in Singapore, via FE model updating load between the stay cable anchorage and the bridge deck.
based on sensitivity analysis. The analysis was done by using Loads are also transferred from the deck to the bridge stays
the ANSYS code (ANSYS 1997) for FEA in conjunction with by two 0.35-m-thick webs along the center of the girder.
the FEMtools (FEMtools 1998), a set of software tools used For greater stability, the deck is fixed at the abutments,
mainly for model updating in mechanical engineering/aero- which rest on a row of 1.4-m-diameter bored piles. A system
space applications. Using even imperfect prototype testing of 1.25 by 1.25 m ground beams links the back stay anchorage
data, the FE model has been modified to reflect the true dy- to each other and to the base of the octagonal tapered pylon.
namic behavior of the structure with acceptable accuracy. Five 1.55-m-diameter bored piles support the pylon footing,
while back stay anchorage combines an arrangement of small
BRIDGE DESCRIPTION bored piles and ground anchors. The pylon, ground beams,
The Safti Link Bridge shown in Fig. 1 was completed in and ground anchors together with the back stays, form a sep-
1995. World-renowned structural engineers, Professors T. Y. arated support structure connected to the deck via the bridge
Lin and Philip Chow initiated its conceptual design, which was stays. The fourth bridge stay (counting from the west side) is
executed by T. Y. Lin SEAsia, Singapore. vertical, locating its termination in the pylon approximately 30
The bridge comprises a curved concrete box deck and a m above its anchorage in the deck, with the pylon continuing
single independent offset pylon. Fig. 2 shows the schematic to a height 42.55 m above its base.
plan, angled, and elevation views of the bridge. The arc of the All bridge stays were stressed at anchor blocks located in
curved deck along the centerline spans 100 m between abut- the bridge deck after construction of the deck to progressively

FIG. 2. Schematic Views of Safti Link Bridge

JOURNAL OF STRUCTURAL ENGINEERING / FEBRUARY 2000 / 253

J. Struct. Eng. 2000.126:252-260.


geometric stiffness could be set up in a way that mimics the
actual in situ construction process of the bridge by introducing
axial strains in the cables and then running a static analysis to
determine an equilibrium state in which the cables are pre-
stressed. This state could be used in the free vibration anal-
ysis, but the strategy could not be incorporated into the au-
tomated iterative cycle of FEMtools; thus for tension structures
with significant geometric stiffness, a different approach would
be required. The principle of sensitivity-based model updating
can be applied directly in a semiautomatic loop involving it-
FIG. 3. Detail of Deck Girder erated free-vibration solutions of any FE code that provides
for geometric stiffness. This has been done for one suspension
lift it clear of the construction supports, while the back stays bridge (Brownjohn 1997).
were stressed from the ground anchor caps. The stay cables Gravity also affects stay cable axial stiffness, and for the
Downloaded from ascelibrary.org by New York University on 05/10/15. Copyright ASCE. For personal use only; all rights reserved.

are VSL System SSI 200. Each stay cable comprises a bundle single element used the reduced effective modulus Ei for each
of parallel monostrands enclosed in a polyethylene sheath. cable with horizontal span L, weight density , and operating
Each monostrand is a helicoid arrangement of seven 5-mm stress was found via Ernsts formula
steel wires, for which the effective elastic modulus (according Ei = E/{1 (2L2E/123)} (1)
to design procedures) is taken as 195 GPa.
Although the pylon is gently tapered, the geometrical prop-
DYNAMIC PROPERTIES erties were taken as the mean values of those at the nodes of
the element for each beam element in the pylon, introducing
Dynamic properties of the bridge were obtained from FEA a minor error into the analysis that could be accounted for in
and EMA. For the FEA, the model was set up and run using the updating.
ANSYS. Additionally, the model was exported in ASCII for- As for the boundary conditions, the foundation restraints
mat to run in FEMtools for correlations and model updating. and abutment restraints due to the piles are essentially fixed
The EMA was based on prototype testing. in translation at the support points, whereas the deck ends were
assumed as pinned.
FE Modeling
FEA has for some decades been an accepted tool for sim- Prototype Testing
ulating structural behavior, but creating a good model is not
an easy task. Many different modeling strategies are available A dynamic site test of the bridge was conducted to deter-
(i.e., which element types, how many degrees of freedom, mine the character of the 3D vibration mode shapes and fre-
etc.). These strategies depend on the skill and experience of quencies up to approximately 10 Hz. The pylon was inacces-
the analyst and on the intended application of the model (e.g., sible, and so measurements were restricted to the deck and
for predicting behavior due to static, dynamic, and/or thermal made in vertical and radial directions at locations 0E to 10E,
loading). However, preparation of an FE model that will be a 0W to 10W, and 4C as shown in Fig. 4. Two techniques were
candidate for updating requires the consideration of additional employed: ambient vibration testing (AVT) and forced vibra-
factors not normally taken into account in conventional FE tion testing (FVT). A set of eight force balance accelerometers
model construction. Of these, the choice of updating param- was used with low-pass filtered signals digitally recorded on
eters is most important. The inaccuracies or uncertainties in a portable computer via an analog-to-digital converter. A 5.5-
the structure must be expressed as parameters such that these kg instrumented sledgehammer was used for FVT; breezes,
uncertainties can be assessed quantitatively. In addition, when vehicles, or pedestrians were relied on for the AVT. The testing
constructing FE models for updating, they have to be ex- is fully described elsewhere (Brownjohn et al. 1999).
changeable between the FEA and updating software. In this
case study, only element types supported both in the ANSYS Correlation Analysis
and FEMtools were used in the generation of the FE model. To correlate the results between initial FEA and EMA, the
Based on the above considerations, the structural compo- FE model data generated in ANSYS and the measured data
nents of the box-type deck were modeled by shell elements. were imported into the updating software FEMtools through
The pylon and ground beams were modeled as conventional interface programs. The pairing of FE model nodes and mea-
3D beam elements. The shell elements and 3D beam elements surement points (circle points) is shown in Fig. 5. The corre-
are supported both in ANSYS and FEMtools; thus they can lation of dynamic properties is listed in Table 1. The modal
be exchanged between both codes, but restrictions to certain assurance criterion (MAC) value is a coefficient analogous to
element types presented some difficulties. the correlation coefficient in statistics or coherence in signal
The tensions in the stay cables should have the effect of processing. It compares ordinates of mode shapes from FEA
increasing the stiffness of the deck against transverse (vertical and EMA and gives a value of unity for perfect correlation,
and lateral) vibrations via geometric stiffness effects. Unfor- while returning a value of zero for uncorrelated orthogonal
tunately, FEMtools cannot process the effects of cable ten- modes. It is defined as follows:
sions, and so the cables were each modeled with a single con-
ventional unloaded 3D beam with low flexural rigidity. Prior Ta Te 2
analyses had been conducted (Brownjohn et al. 1999) using MAC(a , e) = (2)
(aTa)(eTe)
SAPIV in which it was found that neglecting cable tension
resulted in reductions of 2 and 4% in first (symmetric) and where a and e = analytical and experimental mode shape
second (antisymmetric) vertical bending modes with the re- vectors, respectively. The superscript T denotes the vector
ductions of <0.5% for other modes. For the case of a suspen- transpose.
sion bridge deriving its stiffness from cable tensions, the ap- Table 1 shows that the correlation was not too good between
proximation of ignoring the tensions would be unacceptable; FEA frequency fFEA and EMA frequency fEMA (except for first
however, in this case it is not unreasonable. In ANSYS the bending mode) with differences Df exceeding 10% and even
254 / JOURNAL OF STRUCTURAL ENGINEERING / FEBRUARY 2000

J. Struct. Eng. 2000.126:252-260.


shown in Fig. 6 in which the EMA ordinates are shown as
dots on the FEA modes.
Two other ways of comparing FEA and EMA modes are
shown in Figs. 7 and 8. Fig. 7 shows pairing of frequencies
between initial FEA and EMA emphasizing errors as depar-
tures from a diagonal line with unit slope. The MAC matrix,
which illustrates orthogonality conditions between all combi-
nations of analytical and experimental mode shapes, is shown
in Fig. 8. For comparable modes, the MAC values are high,
whereas off the diagonal dissimilar modes are indicated by
values close to zero.

FE MODEL UPDATING
The initial FEA for the bridge was not 100% successful in
Downloaded from ascelibrary.org by New York University on 05/10/15. Copyright ASCE. For personal use only; all rights reserved.

the assessment of dynamic properties. Model updating was


used to improve the FEA predictions.

Updating Procedures
In applying FE model updating technology (Brownjohn and
Xia 1999), the procedure for model updating contains three
aspects: selection of responses as reference data, selection of
parameters to update, and model tuning. Prior experience
shows that a successful updating is strongly dependent on cor-
rect selection of responses and on the choice of uncertain
model parameters.
The responses include resonant frequencies and mode shape
ordinates that are strongly dissimilar between FEA and EMA.
In this case all of the EMA mode shape ordinates and fre-
quencies were selected as responses. The choice of parameters
is a crucial step in model updating. The important issues are,
FIG. 4. Detail of Measurement Points first, how many parameters should be selected, and, second,
which parameters from the many candidates are preferred? It
generally requires the application of physical and mathemati-
cal insight. Physically, the selected parameters must be uncer-
tain in the model. Otherwise, the blindly updated structural
components may lose their originally certain properties and
produce meaningless results in the updated FEA. Mathemati-
cally, if the estimation of too many parameters is attempted,
then the problem may appear ill-conditioned or arbitrary be-
cause the observations are limited in vibration testing. To have
a well-conditioned updating problem, and bearing in mind the
limitation of the measurements, it is necessary to select those
updating parameters that will be most effective in producing
FIG. 5. Pair of FE Model Nodes and Measurement Points a genuine improvement in the modeling of the structure.
Therefore, the number of updating parameters should be kept
small, and such parameters should be chosen with the aims of
TABLE 1. Correlation between Initial FEA and EMA
correcting recognized uncertainty in the model and ensuring
fFEA fEMA Df MAC that the data should be sensitive to them. One good way to
Number (Hz) (Hz) (%) (%) Mode shape assess this is to conduct a sensitivity analysis that computes
(1) (2) (3) (4) (5) (6) the sensitivity coefficient defined as the rate of change of a
1 1.26 1.18 6.39 96.5 First bending particular response quantity with respect to a change in a struc-
2 1.61 2.76 41.58 63.8 Second bending tural parameter. Structural parameters with consistent large
3 2.62 3.59 27.09 96.4 First torsion values are chosen. For all selected responses and parameters,
4 3.71 4.61 19.50 87.6 Third bending the sensitivity matrix [S ] is obtained as follows:
5 4.88 6.10 19.96 97.8 Second torsion
6 6.16 7.00 11.94 86.5 Fourth bending Ri
7 7.00 9.10 23.11 94.6 Third torsion [S ]ij = (3)
Pj
where Ri and Pj represent a structural response and parameter,
reaching 40%. The correlation of mode shapes expressed by respectively. The subscripts are i = 1, . . . , N for N responses
MAC values seems good except for the second bending mode and j = 1, . . . , M for M parameters. Eq. (3) is a differential
(MAC value of 63.8%), although MAC statistics are imperfect analysis for the sensitivity coefficient. The sensitivity matrix
indicators of modal correlation that cannot compete with vi- can be computed for all physical element properties (material,
sual comparison. The descriptions given in column 6 in Table geometrical, boundary, etc.) by using direct derivation or per-
1 are the mode shapes with predominant vertical bending or turbation techniques depending on whether or not mass and
torsion of the deck girder. Being a complex asymmetric struc- stiffness are proportional to the property.
ture, the vertical, lateral, and torsional responses occurred to- When sensitivity analysis is used to help selection of param-
gether in each mode to varying degrees. The mode shapes are eters for model tuning, one should start with all possible pa-
JOURNAL OF STRUCTURAL ENGINEERING / FEBRUARY 2000 / 255

J. Struct. Eng. 2000.126:252-260.


Downloaded from ascelibrary.org by New York University on 05/10/15. Copyright ASCE. For personal use only; all rights reserved.

FIG. 6. Mode Shapes of Dominant Vibration of Deck Girder

rameters, then identify sensitive and insensitive areas, and then Youngs modulus E of the walkway deck (parameter No. 1)
eliminate ineffective (low sensitivity) parameters. The effec- or mass density of web and lower central deck (parameter
tive parameters (high sensitivity) can be further examined by No. 12), these two parameters were still selected because they
arriving at a selection suitable for model tuning. The uncertain apply to the major structural components. The parameters of
parameters in the bridge structure may include Youngs mod- the pylon were included even though pylon response was not
ulus E and mass density of the reinforced concrete compo- measured, because the pylon participates in most of the deck
nents, cross-sectional area Ax and inertia moment I of the beam vibration modes.
structures such as pylon, the thickness H of shell components Although the fixities of the piled foundations and deck end
of the deck girder, and the boundary fixities. Out of the pos- bearings were apparently very uncertain, it turned out that the
sible total of 138 structural parameters, only 21 of these, sensitivities of the responses to the boundary parameters were
shown to have a stronger influence on responses, were chosen close to zero. Pinned deck bearings and fully fixed piles were
and are listed in Table 2. The envelope of normalized sensi- used, and boundary conditions were not chosen as structural
tivities of responses to the selected parameters is shown in parameters.
Fig. 9. After selection of responses and parameters, an iterative
Although the responses are not very sensitive to either procedure for model tuning was conducted. The selected pa-
256 / JOURNAL OF STRUCTURAL ENGINEERING / FEBRUARY 2000

J. Struct. Eng. 2000.126:252-260.


the sensitivity coefficient matrix in terms of a truncated Taylor
series expansion limited to the linear term as follows:
{Re} = {Ra} [S ]({Pu} {P0}) (4)
or
{R} = [S ]{P} (5)
where {R} = difference between the vector of experimental
response {Re} values and the vector of analytical response
{Ra} values; {P} is the difference between the vector of up-
dated parameter {Pu} values and the vector of current param-
eter {P0} values; and [S ] is the sensitivity matrix that can be
obtained from (3). The sensitivity matrix [S ] is usually a rec-
tangular matrix (N M). Eq. (5) may be determined, over-
Downloaded from ascelibrary.org by New York University on 05/10/15. Copyright ASCE. For personal use only; all rights reserved.

FIG. 7. Pair of Frequencies between Initial FEA and EMA


determined, or underdetermined depending on which number
of responses is equal to, larger than, or smaller than the num-
ber of parameters, respectively. In any case, the estimation of
{P} in (5) can be solved using the pseudoinverse technique
or Bayesian estimation technique. Because the Taylor series
expansion is truncated after the first term, the neglected higher-
order terms necessitate several iterations, particularly when
{R} contains large values. It should be noted that when ex-
tremely large discrepancies exist between the experimental and
analytical models, the validity of the Taylor series truncations
in (4) is undermined, and the iterative process is prone to di-
vergence. Usually, it is required that a reasonable approxi-
mation is obtained by manual methods (i.e., engineering judg-
ment before updating automatically), but Lin et al. (1995)
proposed a modification to the modal sensitivity method that
enables convergence for larger magnitudes of FE modeling
errors.
Although the tuning procedure is an automatic process, it is
still not a black box and requires correct choice of tuning
parameters and reference responses to obtain the best results.
When running an exercise, it is impossible to obtain a desir-
FIG. 8. MAC Matrix between Initial FEA and EMA able result after only one or two tuning procedures, and it is
necessary to frequently adjust the tuning strategies for a suc-
rameters were estimated during an iterative process in the tun- cessful updating. For example, if a satisfactory result can not
ing procedure. A tuning procedure was finished when a given be obtained after one sequence of iterations, it can be repeated
difference margin (tolerance) was achieved or an allowed after deleting or replacing responses or parameters, or after
number of iterations was completed. In the formulation of pa- changing the convergence margin or maximum number of it-
rameter estimation, based on the penalty function method, the erations.
true responses or experimental responses were expressed as To avoid physically impossible updated parameter values,
functions of analytical responses, structural parameters, and the lower and upper bounds for the parameter values should

TABLE 2. Changes in Selected Parameters


p
Number Type VI VU (%) Structural component
(1) (2) (3) (4) (5) (6)
1 E 31 GPA 28.0 GPA 9.68 Walkway deck
2 E 31 GPA 25.5 GPA 17.74 Upper central deck
3 E 31 GPA 37.0 GPA 19.35 Traffic deck
4 E 35 GPA 36.1 GPA 3.14 Pylon
5 E 31 GPA 27.5 GPA 11.29 Web and lower central deck
6 E 31 GPA 32.4 GPA 4.52 Lower deck
7 E 31 GPA 33.4 GPA 7.74 Parapet wall
8 2,400 kg/m3 1,740 kg/m3 27.50 Parts of pylon linking cables
9 2,400 kg/m3 1,960 kg/m3 18.33 Walkway deck
10 2,400 kg/m3 2,180 kg/m3 9.17 Traffic deck
11 2,400 kg/m3 2,210 kg/m3 7.92 Pylon
12 2,400 kg/m3 3,150 kg/m3 31.25 Web and lower central deck
13 2,400 kg/m3 3,110 kg/m3 29.58 Lower deck
14 2,400 kg/m3 1,840 kg/m3 23.33 Parapet wall
15 Ax 4.85 m2 3.5 m2 27.84 Parts of pylon linking cables
16 H 0.2 m 0.15 m 25.0 Walkway deck
17 H 0.2 m 0.19 m 5.0 Upper central deck
18 H 0.2 m 0.25 m 25.0 Traffic deck
19 H 0.35 m 0.28 m 20.0 Web and lower central dock
20 H 0.15 m 0.18 m 20.0 Lower deck
21 H 0.4 m 0.55 m 37.5 Parapet wall

JOURNAL OF STRUCTURAL ENGINEERING / FEBRUARY 2000 / 257

J. Struct. Eng. 2000.126:252-260.


Downloaded from ascelibrary.org by New York University on 05/10/15. Copyright ASCE. For personal use only; all rights reserved.

FIG. 10. Pair of Frequencies between Updated FEA and EMA

FIG. 9. Envelope of Sensitivity of Responses to Parameters

be applied. If a parameter value reaches its allowable extreme


during iterative modal tuning, then the parameter becomes in-
effective during the rest of the procedure. The bound levels of
parameters should be set lower in every tuning procedure to
guarantee the accumulated bound levels not to exceed the al-
lowable extreme after finishing the final tuning procedure. It
is possible that convergence cannot be obtained to a satisfac-
tory degree when parameter bounds are defined. A trade-off
between physically acceptable parameter values and the con-
vergence level is then required.

Updated Results
FIG. 11. MAC Matrix between Update FEA and EMA
Table 3 lists the correlation values after updating. It can be
seen that the differences Df between FEA frequency fFEA and
ical properties. Although it is possible to place limits on the
EMA frequency fEMA were all reduced to below 10%. The cor-
allowable ranges of the chosen structural parameters, the final
relation of mode shapes was also improved, as MAC values
results should be checked against normal practice. In this case,
all exceeded 90%, with the exception to the fourth bending
the updated values of Youngs modulus E and the mass density
mode shape. The worst MAC value is 84.7%. Even though
of the concrete components were within the ranges given by
this sounds high, the MAC values should be used with caution
the British Code of Practice for the Structural Use of Concrete
because they can mask some significant variations in the mode
and the American Concrete Institute Building Code according
shape that could, for example, be identified and localized
to Neville (1982).
using coordinate modal assurance criterion. For the second
The mode shapes of dominant vibration from updated FEA
bending mode, the MAC value has increased from 63.8 to
are compared with EMA in Fig. 6. To visualize clearly, the FE
96.3%. The good pairing of frequencies between updated FEA
model is shown only as a wire frame; the lower components
and EMA is shown in Fig. 10; all paired points are close to
such as the web, lower central deck, and lower deck of the
the diagonal. Inspection of the MAC matrix (Fig. 11) shows
deck girder were selected out. The circle points in Fig. 6 rep-
an improvement on the prior FE model (Fig. 8).
resented the EMA mode shapes. Visual checking shows sat-
Table 2 lists the changes p in value of the selected param-
isfactory agreement.
eters with initial value VI and updated value VU. The initial
To check the reliability of the updated model, an exercise
values of parameters in the table were the design values of
was done by rerunning the updated numerical model with per-
parameters. For the reinforced concrete, the design values were
turbed parameters about values of the updated parameters to
quite different from actual updated values. Theoretically, it is
generate a perturbed model, taking the frequencies and mode
very difficult to determine accurately the values of these phys-
shapes as reference data, and performing an updating proce-
dure for the perturbed model. It was found that the perturbed
TABLE 3. Correlation between Updated FEA and EMA
model converged very fast to the updated mode only after five
fFEA fEMA Df MAC iterations. The maximum difference of the frequency was only
Number (Hz) (Hz) (%) (%) Mode shape 0.15%, and the minimum MAC value arrived at 99.8% after
(1) (2) (3) (4) (5) (6) the updating procedure. The exercise verified the success of
1 1.14 1.18 3.67 97.8 First bending the updating procedure with experimental data.
2 2.62 2.76 4.89 96.3 Second bending
3 3.29 3.59 8.36 97.1 First torsion DISCUSSION
4 4.18 4.61 9.36 90.8 Third bending
5 5.86 6.10 4.00 91.9 Second torsion The FE model updating technology has been applied suc-
6 6.60 7.00 5.70 84.7 Fourth bending
7 8.70 9.10 4.38 91.4 Third torsion
cessfully in the dynamic assessment of the cable-stayed bridge.
It is important to emphasize here what is learned from this
258 / JOURNAL OF STRUCTURAL ENGINEERING / FEBRUARY 2000

J. Struct. Eng. 2000.126:252-260.


exercise as it is likely that the process will increase in popu- crease. However, close inspection shows that the changes in
larity. the parameters are in accord with the changes in dynamic be-
havior of the updated FE model. Except for Mode 1, all the
FE Modeling for Updating frequencies increased, consistent with a global increase in stiff-
ness; Table 2 shows increases in Youngs modulus E and thick-
There are some particular considerations when modeling a ness H of the traffic deck, lower deck, and parapet wall,
structure as a candidate for FE model updating. The initial FE whereas mass density of the traffic deck and parapet wall
model used in this study was not the first attempt at modeling were reduced. The traffic deck and lower deck are thin-walled
the bridge. A previous FE model adapted from one used by closed boxes dominating the bending and rotational stiffness
the designer for checking static load combinations was initially properties of the wing-tip deck.
chosen and adapted as a candidate for updating. This model Of course, the changes in selected parameters also must
featured a spine beam by using conventional 3D beam el- meet the requirement of mode shapes closing to EMA values.
ements to represent the deck girder (Brownjohn et al. 1999; From this point of view, it should be difficult to interpret the
Xia 1999). These beam elements incorporated all the bending, changes in the parameters. However, it could be realized that
torsional, and inertial properties of the box-deck, whereas low
Downloaded from ascelibrary.org by New York University on 05/10/15. Copyright ASCE. For personal use only; all rights reserved.

all updated values of selected parameters represented their ac-


density elements capable of transferring static loads but not tual values for closing the model to reality.
representing inertia properties were used for the deck.
Updating was applied to this model, but the results were not Effect of Quality of Testing Data on Updating
ideal, with a maximum frequency error of 15%. Even to
achieve this level of agreement, six selected parameters Because updating is based on the assumption that test data
changed by 100%, losing physical relevance for the structure. are correct, the reliable updating depends on the accuracy of
By condensing deck properties into a spine beam model the the test data. It also depends on the quality of the test data
original geometrical properties of the wing-tip box deck were (Mottershead and Friswell 1993), and this case study illustrates
lost. The lesson here is that the level of detail has to be high this. The MAC value for the fourth bending mode (before
if the physical properties are to remain in reasonable ranges. updating) was 84.7%, which is low compared to values for
other modes, due to the less reliable data for the higher mode.
Model of Cables Modal data for higher modes are not so accurate both exper-
imentally and analytically, partly due to spatial resolution of
As mentioned earlier, the stressing in the cables was not mode shapes, and updating does not work as well for these
modeled, and single conventional 3D beam elements with re- modes. The effect is also visible for third bending, third tor-
duced effective axial stiffness were used for each stay cable, sion, and second torsion modes having lower MAC values
simplifying the effects of cable tension and sag on lateral and compared to first bending, second bending, and first torsion
axial stiffness, respectively. For all but the lowest two vibra- (around 91% against 97%). Apparently quality of test data is
tion modes, this simplification has a negligible effect on the more critical for identifying and updating the higher modes.
results. Furthermore, trial studies showed that unless the cable As well as specific effects on higher modes, test data quality
dynamics themselves are an issue it is best to avoid modeling will also be affected by the ambient effects (e.g., wind, traffic,
the cables with multiple elements; otherwise, the free vibration and temperature). The vibration sources used in AVT also con-
solution will be dominated with cable modes. These extra taminate the signal with noise. This represents a major prob-
modes, introduced using multiple elements, appear at frequen- lem with the largest civil engineering structures, which require
cies close to those predicted by simple cable vibration for- impractical force inputs to match signal/noise ratios that are
mulas, but involve weak participation of the deck. Otherwise, acceptable in laboratory testing. When using AVT methods,
the main structural modes are unaffected. the response peak in the frequency domain is reduced and
The axial stiffnesses of the cables have a small uncertainty, slightly flattened by noise, distorting the mode shape. Prob-
because even though the tensions and mass density are known, lems of bias, resolution, noise, and averaging are acute for
the elastic modulus used in (1) is a nominal design value. The large structures with low frequencies, nonstationary inputs,
sensitivities with respect to axial stiffnesses are small, and and variable ambient effects. Moreover, for AVT, the measured
their inclusion as updating parameters suggests impossible in- mode shape is usually an operating deflected shape con-
creases in their values. Hence they are taken as certain param- taining contributions from other modes that will only be neg-
eters. ligible if modes are well separated.
Despite the simplifications, the updating produced good re- Little can be done to improve noisy data even if the noise
sults. There remains a consistent underestimation of modal fre- level can be measured; the best that can be done is to define
quencies, and it is tempting to reconcile this with the under- confidence bands relating to noise levels. To protect the in-
estimation due to neglecting geometric stiffness effects, except vestment of time and money in testing, the best equipment and
that the actual distribution of frequency errors is not biased in greatest care in selection of transducer locations and signal
a manner consistent with the effects of geometric stiffness. ranges are required. To that end a prior dynamic analysis is
The lesson here is that if the dynamics of the stay cables usually required.
themselves are not relevant, a reasonable approximation can
be obtained with equivalent springs. CONCLUSIONS

Interpretation of Parameter Changes The following are concluded from this paper:

The selected parameters included the physical properties The dynamic properties obtained by the FEA for the com-
such as Youngs modulus E and mass density and the geo- plex structures such as the Stafi Link Bridge are not al-
metrical properties such as cross-sectional area Ax and thick- ways consistent with the measured results due to the mod-
ness H of the structural components. The changes in these eling errors and the uncertainties in the structure. It is
parameters represented the global changes of stiffness and necessary to improve the FE model for successful dy-
mass, leading to global changes of dynamic properties of the namic assessment of the structure.
bridge structure. The changes in selected parameters have a The model updating is a feasible and effective technology
random appearance, because some increase whereas others de- for improvement of the FE model by modification of the
JOURNAL OF STRUCTURAL ENGINEERING / FEBRUARY 2000 / 259

J. Struct. Eng. 2000.126:252-260.


parameters with uncertainties existing in the structure, Stayed Bridge in Sevilla (Spain). Earthquake Engrg. and Struct. Dyn.,
based on the prototype testing data. The successful appli- 24(1), 3551.
Doebling, S. W., Farrar, C. R., and Prime, M. B. (1998). A summary
cation of the model updating technology to the dynamic review of vibration-based damage identification methods. The Shock
assessment of the Safti Link Bridge represented a crucial and Vibration Digest, 30(2), 9105.
step toward using the technology for dynamic assessment Dumanoglu, A. A., Brownjohn, J. M. W., and Severn, R. T. (1991). Seis-
of complex structures in civil structural engineering. mic analysis of the Faith Sultan Mehmet (Second Bosporus) suspension
The appropriate FE modeling of a structure is the key to bridge. Earthquake Engrg. and Struct. Dyn., 21, 881906.
successful updating. The structure should be modeled Felber, A. J. (1995). Seismic assessment of existing bridges: one role
for dynamic testing. Proc., Bridges into the 21st Century, Hong Kong
with as much detail as possible to represent geometric and Institution of Engineers, 765772.
structural form. Felber, A., and Cantieni R. (1996). Advances in ambient vibration test-
Using a conventional unloaded 3D beam element to ing: Ganter Bridge, Switzerland. Struct. Engrg. Int., Zurich, Switzer-
model a tensioned cable in the cable-stayed bridge can land, 6(3), 187190.
still produce good results in analysis of dynamic proper- FEMtools users manual version 1.4. (1998). Dynamic Design Solutions,
ties, but the accuracy of the updating is limited unless Leuven, Belgium.
Hawk, H., and Small, E. P. (1998). The BRIDGEIT bridge management
Downloaded from ascelibrary.org by New York University on 05/10/15. Copyright ASCE. For personal use only; all rights reserved.

geometric stiffness effects can be included. system. Struct. Engrg. Int., Zurich, Switzerland, 8(4), 309314.
There is still much to learn about FE model updating. For Hearn, G. (1998). Condition data and bridge management system.
example, the methods for arriving at a starting point close Struct. Engrg. Int., Zurich, Switzerland, 8(3), 221225.
enough to the solution to allow convergence are not for- Kominek, M. (1998). The Marian Bridge, Czech Republic. Struct.
malized. Also, updating places special requirements on Engrg. Int., Zurich, Switzerland, 8(4), 283284.
modal test strategy, and it is recognized that some form Law, S. S., and Ko, J. M. (1995). Ambient vibration measurements of
the Tsing Ma Bridge towers. Proc., Bridges into the 21st Century,
of dynamic analysis should precede a modal test to aid in Hong Kong Institution of Engineers, 585592.
selecting the correct transducer locations and frequency Lin, R. M., Lim, M. K., and Du, H. (1995). Improved inverse eigen-
range for capturing the modes that contribute the most to sensitivity method for structural analytical model updating. J. Vibra-
the updating. The optimization also extends to selection tion and Acoustics, 117, 192198.
of the appropriate type and number of structural param- Menn, C. (1996). The place of aesthetics in bridge design. Struct.
eters in relation to the responses. Guidelines for these Engrg. Int., Zurich, Switzerland, 6(2), 9395.
Mottershead, J. E., and Friswell, M. I. (1993). Model updating in struc-
steps will evolve through further studies on civil struc- tural dynamics: A survey. J. Sound and Vibration, 167(2), 347375.
tures. Neville, A. M. (1982). Properties of concrete, 3rd Ed., Pitman Publishing,
Ltd., London.
APPENDIX. REFERENCES Pavic, A., Hartley, M. J., and Waldron, P. (1998). Updating of the an-
alytical models of two footbridges based on modal testing of full scale
ANSYS user manual, release 5.4. (1997). Swanson Analysis System, Inc., structures. Proc., 23rd Int. Seminar on Modal Analysis, Society for
Houston. Experimental Mechanics, Inc., Bethel, 11111118.
Brownjohn, J. M. W. (1997). Vibration characteristics of a suspension Rito, A. (1996). Aesthetics and bridge design. Struct. Engrg. Int., Zu-
footbridge. J. Sound and Vibration, 202(1), 2946. rich, Switzerland, 6(2), 9192.
Brownjohn, J. M. W., Dumanoglu, A. A., and Severn, R. T. (1992). Am- Severn, R. T., Brownjohn, J. M. W., Dumanoglu, A. A., and Taylor, C.
bient vibration survey of the Faith Sultan Mehmet (Second Bosporus) A. (1989). A review of dynamic testing methods for civil engineering
suspension bridge. Earthquake Engrg. and Struct. Dyn., 21, 907924. structures. Civ. Engrg. Dyn., London, U.K., 124.
Brownjohn, J. M. W., Lee, J., and Cheong, B. (1999). Dynamic perfor- Takenouchi, K. (1998). Single pylons for curved cable-stayed bridges.
mance of a curved cable-stayed bridge. Engrg. Struct., 21(11), 1015 Struct. Engrg. Int., Zurich, Switzerland, 8(4), 269272.
1027. Tan, S. C. (1996). Safti Link Bridge, design and construction. Defence
Brownjohn, J. M. W., and Xia, P. Q. (1999). Finite element model up- Technology Group Seminars Series, Ministry of Defense, Singapore,
dating of a damaged structure. Proc., 17th Int. Modal Analysis Conf., 14.
Society for Experimental Mehcanics, Inc., Bethel, Vol. 1, 457462. Troyano, L. F., Armisen, J. M., and Suarez, M. A. A. (1998). The in-
Cantieni, R. (1996). Updating of analytical models of existing large clined towers of the Ebro and Lerez Bridge. Struct. Engrg. Int., Zu-
structures based on modal testing. Proc., Workshop US-Europe on rich, Switzerland, 8(4), 258260.
Bridge Engrg.: Evaluation Mgmt. and Repair, ASCE, Reston, Va., 15 Xia, P. Q. (1999). Condition assessment of bridges by dynamic mea-
17. surements. Tech. Rep., School of Civ. and Struct. Engrg., Nanyang
Casa, J. R. (1995). Full scale dynamic testing of the Alamillo Cable- Technological University, Singapore.

260 / JOURNAL OF STRUCTURAL ENGINEERING / FEBRUARY 2000

J. Struct. Eng. 2000.126:252-260.

You might also like