Engineering Structures 33 (2011) 12671276

Contents lists available at ScienceDirect

Engineering Structures
journal homepage: www.elsevier.com/locate/engstruct

Load rating of concrete-deck-on-steel-stringer bridges using field-calibrated

2D-grid models
Ahmet Turer a, , Bahram M. Shahrooz b
a
Civil Engineering Department, Structural Mechanics Lab., Middle East Technical University, 06531 Ankara, Turkey
b
Department of Civil and Environmental Engineering, University of Cincinnati, 765C Baldwin Hall, OH 45221, USA

article info abstract

Article history: This paper presents and discusses issues related to structural identification, calibrated model-based load
Received 18 August 2009 rating, and sensitivity of rating to the analytical model, along with experimental studies conducted on an
Received in revised form existing concrete-deck-on-steel-stringer bridge. The proposed model-updating procedure uses collected
5 January 2011
dynamic data (mode shapes, modal frequencies, and order of modes) as well as static deformed shape
Accepted 6 January 2011
Available online 7 February 2011
information. Two-dimensional (2D) grid models were developed to successfully simulate the transverse
load transfer mechanisms between girders, torsional flexibility, and effects of skewed bridge architecture.
Keywords:
The rating results obtained from the 2D-grid models were close to 3D-FEM-based evaluation, while
Bridge simplified 1D bar models had serious shortcomings. Grouping the parameters of the analytical model
Rating at different stages of model calibration enhanced the speed and convergence success of the objective
Modeling function. Although cross-braces are considered as non-structural members, they have been found to be
Calibration the most critical members of the selected bridge during rating studies. Failure of cross-braces deemed
to alter the load transfer mechanism between girders and possibly resulting in the premature failures of
interior girders.
2011 Elsevier Ltd. All rights reserved.

1. Introduction
Bridges are potentially the weakest links of a healthy trans-
portation system. According to 2005 bridge inventory statistics
in USA, about one quarter of all bridges (156,177 bridges out of
594,616 nationwide) [1] are structurally deficient and/or function-
ally absolute, while USA can be considered as one of the relatively
good examples regarding bridge safety considering other countries
around the globe. As recent unfortunate bridge collapses have bit-
terly reminded the society, evaluation and categorization of the
bridge conditions is a major task and should be given high impor-
tance (e.g. I-35W Mississippi River bridge in Minnesota, August 1,
2007 [2] and Harp Road bridge Oakville, Washington, 15 August
2007 [3,4]). Large numbers of bridges often times make the eval-
uation and categorization process an unmanageable task. Visual
inspection methods have known shortcomings such as visual lim-
itations, indirect and often uncertain correlation between appear-
ance and actual structural condition, subjectivity of the evaluator
due to his/her level of experience, variations in the moodpersonal
emotional state of the evaluator affecting the evaluation, and even
weather conditions during bridge evaluation. Therefore, a robust
and objective evaluation method is needed to assess the existing
condition of bridges, which would preferably be based on field
measured objective data and adequate level of analytical modeling.
The field-calibrated analytical model-based objective load rating
of bridges would not only eliminate false negative and false pos-
itive errors such as identifying a healthy bridge as deficient, or
a deficient bridge as healthy but also allows bridge engineers to
conduct correct ordering and prioritization of bridges that need re-
pair work and strengthening. Different aspects of bridge load rat-
ing have been studied in the past [5], including time dependency,
reliability, and probabilistic approaches [6,7]. This paper discusses
general concepts on field-calibrated analytical model-based load
rating, putting emphasis on different levels of analytical modeling
(1D, 2D, 3D models) on the load rating sensitivity. An overview of
load rating methods is presented and implementation is shown us-
ing an actual sample bridge.

1.1. Overview of bridge load rating

Corresponding author. Tel.: +90 312 210 5419; fax: +90 312 210 7991; GSM:
Bridge load rating utilizes a normalized unitless value which
+90 533 655 7665.
E-mail addresses: aturer@metu.edu.tr (A. Turer), shahrobm@ucmail.uc.edu
quantifies the remaining capacity of a bridge, after its self-weight,
(B.M. Shahrooz). in terms of standard live loads. Dynamic impact factors and
URL: http://www.ce.metu.edu.tr/aturer/ (A. Turer). different load factors are used during rating. The standard truck
0141-0296/$ see front matter 2011 Elsevier Ltd. All rights reserved.
doi:10.1016/j.engstruct.2011.01.004
1268 A. Turer, B.M. Shahrooz / Engineering Structures 33 (2011) 12671276
Fig. 1. Standard truck and lane loads defined by AASHTO.
Fig. 2. AASHTO rating equations and factors for allowable stress and load factor
methods.
and lane loads used for bridges are shown in Fig. 1 [8], although
different states and countries have additional standardized vehicle
or military loads (e.g., Ohio: 2F1, 3F1, 4F1, 5C1 trucks). The demand
on bridge members generated by the dead load is subtracted
from the total capacity to obtain the remaining capacity, which
is then divided by the live load demand to acquire load rating.
Fig. 2 summarizes the load rating equations which can be used
for both allowable stress and load factor methods by modifying
the coefficients A1 and A2. Bridge load rating is obtained for
inventory and operating levels using different allowable stresses or
A2 coefficients as shown in Fig. 2. The scope of this paper is limited
to allowable stress and load factor rating methods; however, Load
and Resistance Factor Rating (LRFR) is also available in AASHTO
Manual for condition evaluation of bridges [8]. Calculation of load
ratings for both operating and inventory levels can be performed
in LRFR, which uses limit states for strength, service, and fatigue.
The strength limit state rating considers the strength capacity of
a bridge under permanent and live loading. The service limit state
accounts for stress, deformation, and crack width. The fatigue limit
state accounts for cyclic stresses. Strength is the principal limit
state for all ratings and commonly determines the bridge posting,
closing, and repair decisions [9].
1.2. Levels of analytical modeling for bridge load rating
Load rating procedure requires computation of the dead and
live load demands on bridge members. Available computation
methods have different levels of complexity and detail for calcu-
lating the demands on members. The most widely known and gen-
erally used analysis type is 1D beam model analysis. Generally, a
single girder line is modeled using beam elements and load trans-
fers to the modeled girder from adjacent girders are considered by
using distribution factors. The drawbacks of the 1D beam analysis
can be mainly grouped as: oversimplified geometry, weakness in
modeling the transverse direction properties (diaphragms, cross-
braces) and irregularities (especially for skewed bridges), difficulty
in modeling composite action, and not being capable of simulating
geometric or material nonlinear behavior. Bridge Analysis and Rat-
ing Systems (BARS) program, probably being the most widely used
rating software by DOTs in USA, utilizes 1D approach, although,
currently BARS is being replaced by an improved program called
Virtis [10].
Alternatively, the limitations of 1D model can be mitigated
by using 2D grid or 3D Finite Element (FE) models. For concrete
deck on steel stringer bridges, for example, the girder lines and
cross-braces can be modeled as beam members in the longitudinal
and transverse directions, forming a grid pattern. The grid model
has advantage of modeling the width of a bridge with multiple
girder lines and transverse beams as opposed to modeling only
a girder line or considering the whole bridge as a single beam.
Furthermore, grid models are easier to model, analyze, and fully
automate for calibration compared to 3D-FE models. The 2D-grid
model referred in this study has a flat grid structure in the xy plane
selecting the z-axis for the vertical direction. Each node has three
degrees of freedom: one vertical translation along the z-axis and
two horizontal rotations in the x and y axes. Shell members may
also be used for the 2D-based modeling but current study utilizes
only lumped frame members representing composite structure of
the concrete deck and steel girders (Fig. 3).
Different levels of modeling for girder bridges are shown in
Fig. 4. 3D-FE models are superior to the 1D- and 2D-based mod-
eling in defining the geometric dimensions, continuity or dis-
continuity, material properties, and support conditions. Common
element libraries available for 3D-FE modeling (e.g., frame, truss,
shell, solid, anisotropic elements; laminated, heterogeneous non-
linear materials) allow users to better model the structural prop-
erties (i.e., material and geometric) of a bridge. Adjusted meshing
density at various section of a bridge allows important sections to
be more precisely modeled and analyzed. However, 3D-FE mod-
els generally have thousands of elements, very large number of
degrees of freedom, and unmanageable number of optimization
variables for calibration, which require relatively long analysis du-
rations and impose major challenges for automated calibration.
Time- or displacement-dependent geometric or material
nonlinearities may be added to FE elements allowing time- and
loading level-dependent nonlinear modeling and analysis (e.g.,
nonlinear time-history and push-over analysis). Modeling behav-
ior in the nonlinear range is especially important for LRFR studies,
since the ultimate section capacities may not be clearly defined in
models as in the case of composite bridges. 2D-grid models are be-
lieved to be the most efficient modeling level compared to the un-
realistic simplicity of 1D-models and unmanageable complexity of
the 3D and nonlinear models. 2D-grid models can be best utilized
for a large number of existing bridges since they are simple to con-
struct and calibrate if field measured data are available.
2. Grid modeling for concrete deck on steel stringer bridges
2.1. Bridge inventory
From about 600,000 existing 6 m (20 ft) and longer bridges
in USA, about 29,000 bridges are located in State of Ohio. The
concrete-deck-on-steel-stringer bridges constitute about 41% of
the total bridge population in Ohio [11]. If the 2D-grid model-
based calibration and load rating method can be streamlined, then
the model generation, field calibration, and load rating can be
 A. Turer, B.M. Shahrooz / Engineering Structures 33 (2011) 12671276 1269

Fig. 3. 2D-grid model approximation.

Fig. 4. Different levels of modeling for bridges with girders.

conducted for a large number of bridges in a fairly short amount of
time. In this way, a good percentage of bridges can be objectively
evaluated and better rated with respect to the visual inspection and
1D-beam model-based load rating.
2.2. General assumptions
The grid model calibration is carried out using experimentally
obtained mode shapes, modal frequencies, and static deforma-
tions (BGCIBridge Girder Condition Index [11]).
The calibrated analytical model that replicates measured
dynamic properties and static deflections is assumed to closely
represent the overall structural properties of an existing bridge.

A computer program was developed in MATLAB computing lan-
guage which would conduct structural analysis, model calibration,
and load rating of bridges. The advantage of using MATLAB over
other programming languages is that the software has its own
built-in functions and tool boxes for matrix manipulation and op-
timization routines for minimizing objective function in calibra-
tion. The programmable and customizable graphical interface is
also suitable for easy input and display of measured and generated
data. The basic assumptions inherent in the developed load rating
software are as follows.
Linear stiffness modeling and analysis are considered.
Discrete members in the form of a grid are used to model
composite behavior of girder and deck. Although longitudinal
and transverse girder lines are in the form of a grid (Fig. 3), the
model is assumed to adequately simulate the actual behavior of
deck and girders.
Support conditions are assumed to be properly simulated
by using linear springs in vertical, horizontal, and rotational
directions.
2.3. Automated calibration process of 2D-grid models
Structural model updating/analytical model calibration is well-
known and documented techniques to modify selected parameters
of an analytical model until the response of the computer analyses
and field measured structural properties closely match each
other [1214]. The most commonly used parameters for model
updating are mode shapes and modal frequencies [15]. Some
studies rarely use static load testing-based deflection and rotation
measurements [11,16], which are much easier to match compared
to strain measurement-based model updating since strains are
local responses of a structure which can be influenced by
discontinuities, material imperfections, and stress concentrations.
The uniqueness of this study is to incorporate static and dynamic
parameters at the same time for model updating.
An automated calibration program was written to optimize
the error between measurements and simulations. The selected
parameters for grid model calibration can be mainly grouped as
(a) element rigidity (EI) of longitudinal members, (b) element
rigidity (EI) of transverse members, (c) nodal lumped mass defined
1270 A. Turer, B.M. Shahrooz / Engineering Structures 33 (2011) 12671276
at joints, and (d) support spring coefficients. An improvement was
achieved for faster convergence of calibration process by using
staged level of parameter definition; otherwise, defining the mass
and sectional properties of all members as variables in the first
stage would extremely slow down the convergence and would also
risk the convergence to reach at the global minimum.
The staged variable process used in this study can be ex-
plained using a simple example. In the case of a simply supported
beam, the span can be modeled using one hundred members with
equal number of mass and cross-sectional parameters; however,
calibration of the analytical model by changing hundreds of pa-
rameters one by one would not be successful especially if the ini-
tial (nominal) stiffness values are too different from their actual
(existing) values. Instead, assuming that the structural parameters
of similar sections are close to each other, all of the one hundred
member stiffness parameters can be changed simultaneously as a
single variable. In this way, the stiffness parameter of all members
that would best fit to the measured responses can be obtained. At
this point, the second stage of optimization starts as the initial vari-
able group is divided into two or more sub groups, each sub group
representing the properties of members that belong to a portion of
the beam span. The calibration results of the first stage are assigned
to each sub group, but second stage optimization treats each sub-
group parameter as an independent optimization parameter. The
second stage convergence is much faster and successful than the
case if the calibration was started directly from the second stage
using nominal values. As each staged optimization reaches its own
global minimum, the parameter groups are further divided into
additional sub groups, increasing the number of parameters in a
controlled manner and further improving the calibration results.
It should be noted that the stage variable approach is not guar-
anteed to converge to the solution that will be obtained if the pa-
rameterization of the last step was used in one step to estimate the
parameters.
The calibration of the grid model requires definition of an
objective function (OF) and related minimization (optimization)
process. The selected structural analysis results consist of mode
shapes, modal frequencies, and deformed shapes for girders which
can also be obtained experimentally. The objective function is
basically defined as the summation of absolute errors between the
analytical and experimental results [11] and can be defined using
formal mathematical definitions [17] as
y = F (x) = (FBGCI (x), FMAC (x), Ffreq (x), Forder (x))
where x = (x1 , x2 , . . . , xm ) X is the parameter vector, X is the
parameter space, y = (y1 , . . . , yn ) Y is the objective vector, and
Y is the objective space. OF that will be used for calibration studies
is the scaled summation of objective space to match the analytical
and field responses in all domains. Therefore, the OF can be defined
as follows:
n
4
OF = ci yi = ci |Fi |
i=1 i =1
= c1 FBGCI (x) + c2 FMAC (x) + c3 Ffreq (x) + c4 Forder (x) (2)
where ci are adjusting factors of the objective vector. Parameter
vector variables can be constrained during optimization due to
logical or physical limitations such as the expected range of elastic
modulus or moment of inertia due to physical dimensions of the
girders. The OF has four major components:
1. Bridge Girder Condition Indicator (BGCI) is the static deformed
shape of girders when all experimental nodes are loaded
with a vertical unit load. The absolute difference between the
experimental nodal deformations (xexperimental ) and analytical
nodal deformations (xanalytical ) at each node is summed for all
nodes and calculated as
n_nodes
FBGCI (x) = (xexperimental ) (xanalytical ) .
(3)
i=1
2. Modal Assurance Criteria (MAC) defines a correlation between
the two mode shapes [15] and is defined as
{e }H {a }
2
MAC(e , a ) =
{e }H {e } {a }H {a }
2
{e }T {a }
= (4)
{e }T {e } {a }T {a }
where e and a are mode shapes referring to experimental
and analytical results, respectively, both being n 1 sized; n
being the total number of nodes in each mode. Hermitian
transpose of an array ( H ) refers to taking the transpose of
an array with their complex conjugate values for complex
modes in the form of a + bi, which is modified as a bi
during the transpose. Complex modes may be obtained from
field testing and nonlinear models with high damping can result
complex analytical modes. However, the MAC calculation of real
modes can be simplified as the dot product of modal vectors as
indicated in Eq. (4).
MAC values change between zero and one (1.0 shows perfect
match between two modes while zero indicates no correlation).
The MAC error term is defined as unity minus MAC coefficient
which is then modified by an importance factor. The modes at
lower frequencies are considered to be more important than
the modes at higher frequencies in determining the global
structural properties of a bridge. Therefore, the MAC errors are
scaled up using an importance factor of (n_modes + 1 i),
where n_modes is the total number of matching modes, i is the
current matching mode number, and MACi is the MAC result
of ith matching modes. The summation of all MAC error terms
would yield MAC component of the OF. Alternatively, mass
participation ratios may also be used to scale the MAC errors:
n_modes
FMAC (e , a ) = (1 MAC(e , a )i )
i=1
(1)
(n_modes + 1 i) . (5)
3. Frequency error is the difference between the experimentally
obtained modal frequencies (fexperimental,i ) and their analytically
obtained counterparts (fanalytical,i ), which are automatically
matched for mode number i by using the MAC matrix. The
absolute frequency errors are summed up for all matched
modes, again amplified by the modal order importance term
(n_modes + 1 i), where i refers to the mode number of each
matched frequency pair. Alternatively, mass participation ratios
may also be used to scale the frequency error:
n_modes
Ffreq (f ) =
fanalytical,i fexperimental,i
i=1
(n_modes + 1 i) . (6)
4. The order error term is introduced to prevent mode order
skipping or switches. Torsional and bending modes of a bridge
can often times switch places in the analytical model compared
to the experimentally obtained modes. The errors of the mode
 A. Turer, B.M. Shahrooz / Engineering Structures 33 (2011) 12671276
shape order in the analytical model are penalized using the
Forder term as
Forder (order)
n_modes
= (|orderi orderi1 1| (n_modes + 1 i)) (7)
i=2
where orderi refers to the analytically obtained mode shapes
mode order number that matches with the experimentally
obtained mode number i. For example, if analytically obtained
modes 3, 4, 7, and 6 are matching with the experimentally
obtained modes 1, 2, 3, and 4, the order error is computed as
(|4 3 1|) 4 + (|7 4 1|) 3 + (|6 7 1|) 2 =
0 + 6 + 4 = 10. If the experimental modes were ordered as 3,
4, 6, and 7 preventing mode switch as frequently seen between
torsional and bending mode shapes; then, the Forder term would
be equal to 0 + 3 + 0 = 3. Forder term would converge to zero
if the experimental modes 3, 4, 5, and 6 would match analytical
modes 1, 2, 3, and 4; eliminating any mode skipping.
It is important to have static and dynamic terms coexist in the
OF since consideration of only dynamic data measured from a
bridge (i.e., mode shapes and frequencies) may lead the calibration
process into an unrealistic stiffness state. For example, identical
mode shapes and modal frequencies can be obtained for a simply
supported beam for different rigidity (EI) and mass per length ( )
terms (i.e., changing both with the same proportions). Therefore,
static deformed shape information is crucial (when modal scaling
is missing) to calibrate a bridge analytical model so that structural
stiffness terms are also identified correctly.
The four error terms defined in OF in Eq. (1) have differ-
ent units and magnitudes. Directly summing up the error terms
is incorrect and various error terms should be modified us-
ing cj coefficients, which may be considered as adjusting fac-
tors. BGCI deflection differences have displacement units and
can be normalized by the average of experimentally measured
deflections. The number of terms in the BGCI error term is
equal to the number of nodes. Coefficient c1 can be defined as
I1 /[(number of nodes) (average of static deformations)], vari-
able I1 is the relative importance factor for static deformation
(BGCI) correlations. Similarly, coefficient c2 can be defined as
I2 /[n_modes (n_modes + 1)], where variable I2 is the rela-
tive importance factor for mode shape correlation and n_modes is
the number of modes used in the study. Coefficient c3 can be de-
fined as I3 /[(average of experimental frequencies) (n_modes
(n_modes + 1))]. Order error in Eq. (1) can be scaled by coefficient
c4 which is equal to I4 /(n_modes). The default value of I factors
can be taken as 1.0 or increased if a certain error term is more im-
portant (e.g., I1 = 1.25 I2 = I3 = 1.5 I4 = 0.6). Selection of
importance factors is a personal preference affected by engineer-
ing judgment, which may also be based on the level of reliability
and coherence of the collected field measurements. Detailed dis-
cussions about the choice of the weighting factors and its effects
on model updating and structural identification can be found in
the literature [18].
Simple step search algorithm was used for the OF optimization,
which is similar to simplex algorithm [19,20] but simpler in the
sense that it utilizes one point per variable instead of moving n + 1
points in an n-dimensional space. The method was modified to use
logarithmic scale (e.g., x (1 1x), 1x 0.1) instead of using
linear scale (e.g., x 1x). All of the parameters are updated one
by one and the OF is re-evaluated each time. If the OF is reduced,
the current value of the selected parameter is accepted and next
parameter is modified. If the OF is increased, the value of 1x is
reduced by a multiplier (e.g., 0.9) and therefore the direction
is reversed (Fig. 5(a)). Sensitivity analysis would be beneficiary
to determine the range of the calibration parameters. Defining
1271
constraints and reasonable ranges for some of the parameters may
not always be possible such as in the case of spring stiffness at the
support of a girder. Contact surface and support conditions make a
big impact on the boundary stiffness. Modified simplex algorithm
with logarithmic scale (Fig. 5(b)) is more suitable for automated
calibration of not well-defined variables such as spring coefficients
at the supports. At later stages of the optimization, the scaling
factor (i ) is modified to converge to unity (1) which would enable
fine tuning of variables around the optimal solution.
2.4. Load rating using 2D-grid models
As the grid model of a bridge is calibrated using dynamic
and static measurements, the load rating is performed on the
calibrated model. Standard HS20-44 (or combination of other
trucks) and lane loading are used for rating. The truck load is moved
over the analytical model by small displacement increments over
each lane. Sensitivity analyses conducted using 3D-FEM for the
incremental truck step length indicates that minimum rating
coefficients converge to a constant if step lengths are selected
smaller than 1 m on a 58.5 m long three-span continuous bridge
(1/20th of the shortest span). The member forces (moment and
shear) are computed for all members of a bridge corresponding
to all locations of the rating truck as it moves over all the
lanes. The same procedure is repeated for a moving unit load to
obtain influence lines of member forces. The three-dimensional
(3D) member forces (and/or stresses) matrices are constructed
utilizing each dimension for (a) the member number, (b) the truck
load or unit load location, and (c) lane number. The force or stress
matrices are used to obtain the minimum and maximum member
forcestress values for the most critical loading combinations.
On any slice taken from the 3D matrix cube a 2D matrix exists;
each column in the matrix belongs to a members internal force
influence line and each row belongs to the location of a moving
unit load or truck. For any given member, the worst locations of
moving loads on each lane that would generate the worst possible
combination (minimum and maximum) of live load demands are
detected; scaled for respective (a) lane, (b) concentrated, and (c)
truck loads; and superimposed using the theory of superposition.
Both distributed lane and truck load locations are considered for
all members separately and the largest (negative and positive) live
load demands for each member are found individually. The largest
live load demands are combined with the corresponding dead load
demands and load rating factors (including impact) to separately
calculate the load rating values of all members of a bridge. Different
rating methods, levels, and coefficients are used for the member
load rating calculations as shown in Fig. 2. The member with the
minimum rating would govern the bridge load rating.
2.5. Example application
The analysis, calibration, and rating features of the written
software are demonstrated using PRE-725-0803 bridge, which
was constructed in 1968, has three spans (17 m, 24 m, 17 m),
four lanes with 11.6 m width, 10 skew angle, and a general
inspection appraisal value of 6 (0:bad, 7:good) [21]. The bridge
geometry was simply defined using span lengths, bridge width,
number of girders, skew angle, and cross-brace spacing. The
model was automatically generated by the developed software.
The grid models nominal section properties were manually
defined using available bridge plans/drawings. The nodes that
were used during the field experiment to measure mode shapes
and deformed pattern (BGCI) were also defined through the
graphical user interface of the developed software by clicking
on each intersection, in the correct sequence. The data from
multiple dynamic tests using different nodes can also be defined
1272 A. Turer, B.M. Shahrooz / Engineering Structures 33 (2011) 12671276
Fig. 5. Step method search algorithm (a) incremental; (b) logarithmic.
and simultaneously used during the calibration process. The main
calibration parameters are automatically generated by the script
developed in MATLAB language (i.e., element rigidity (EI) of
longitudinal and transverse members, nodal mass assigned to all
joints, and support spring coefficients) and grouped under three
calibration stages (Fig. 6). Initially in stage #1, all the longitudinal
and transverse member sectional properties are assigned to two
independent variables. All nodal mass and spring coefficients are
also initially grouped under single variables. After reaching a
global minimum for stage #1, the parameters are sub-divided into
smaller groups and calibration process is continued. In stage two,
the exterior members for the longitudinal girders are defined as
additional calibration parameters (Fig. 6(b)); mass variables are
also divided into subcategories for the nodes at the supports,
exterior nodes, and interior nodes (Fig. 6(b)). Upon reaching a
global minimum for the OF for the second stage, the software
automatically goes through the third stage of the optimization
by further dividing and fine tuning the calibration parameters
(Fig. 6(c)). A three-staged calibration routine is defined by default;
however, additional sub-grouping after the third stage can easily
be programmed.
Total of 38 reference nodal points on the third and fourth girder
of PRE-725-0803 bridge were selected to obtain the mode shapes
and modal flexibility to obtained the deformed bridge shape
(BGCI). Total of 16 PCB 393C seismic ICP (IEPE) accelerometers
were used in a rowing scheme for multi inputmulti output
dynamic analysis using complex mode indicator function (CMIF)
algorithm [22,23]. The selected nodes for system identification
are shown in Fig. 7. The deformed shape (BGCI) is obtained
by loading experimental flexibility matrix with unit loads at all
measurement nodes and compared against analytical simulations
in Fig. 7. Total of 11 experimental modes were identified and
used in the model updating. From Fig. 8(a)(d), it is evident that
MAC and modal frequency correlations between the experimental
and analytical results significantly improve after calibration. The
calibration procedure effectively improves the MAC coefficients
as well as modal frequency matches. Furthermore, the resulting
grid model is capable of accurately simulating vibration modes and
frequencies of a skewed bridge.
Following the calibration process, the bridge load rating
procedure was performed by calculating bridge load rating factors
of all the members. The worst truck/lane loading locations and
combinations were determined for each member in the grid
model. The nominal (according to plans) member level bridge
rating coefficients for inventory and operating levels are shown
in Fig. 8(e) and (f), respectively. The member with the minimum
rating factor among all the members of the bridge model yielded
the load rating of the PRE-725-0803 Bridge. The calculated member
rating coefficients were grouped for transverse and longitudinal
members; cross-brace component (of the deck and cross-brace
assembly that forms the transverse members) is commonly not
accepted as structural members although transverse members are
found to have lower rating factors compared to the longitudinal
(deck + girders) members in this study.
Bridge load rating can be computed for the nominal (as-built)
and the calibrated (as-is) conditions of a bridge. Fig. 9 compares
the nominal and calibrated 2D-grid model-based rating factors,
where a tendency of reduced rating values is observed when the
calibrated grid model is used for rating. The reduction percentage
in rating coefficients of the sample bridge differs for different rating
methods, levels, and members; however, rating reduction after
model calibration is in the range of 15%.
 A. Turer, B.M. Shahrooz / Engineering Structures 33 (2011) 12671276 1273

1000 1000 1000

Stage 1: two groups for frames Stage 2: four groups for frames Stage 3: 12 groups for frames
800 800 800
600 600 600
400 400 400
200 200 200
0 0 0
200 200 200
400 400 400
600 600 600
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000
Grouping nodal variables

1000 1000 1000

Stage 1: one group for nodal mass Stage 2: four groups for nodal mass Stage 3: 11 groups for nodal mass
800 800 800
600 600 600
400 400 400
200 200 200
0 0 0
200 200 200
400 400 400
600 600 600
0 500 1000 1500 2000 0 500 1000 1500 2000 0 500 1000 1500 2000

(a) Stage 1 (b) Stage 2 (c) Stage 3

(* each symbol or line type represents an optimization variable.)

Fig. 6. Three stages of frame and nodal mass grouping.

The 3D finite element model of the bridge (Fig. 10) was also
constructed and calibrated using available field measurements in
order to compare against 2D-grid model results. The 2D-grid and
3D-FEM-based bridge load rating approach results are compared
in Fig. 9. The longitudinal members, which have combined effect
of deck and girder, yielded similar rating coefficients when 2D-
grid or 3D-FE models are used. When the rating factors of girders
and cross-braces are compared in Fig. 9, the cross-braces are found
to be in more critical condition compared to the girders for both
grid- and FE-based approaches; therefore, transverse members
are expected to reach their capacities earlier than the girders.
The cross-braces are commonly not evaluated during the rating
process, and standard details are used during design without much
attention to their capacities. Although cross-braces may not be
considered as structural members, they play an important role
for the load transfer between adjacent girders. The failure of a
cross-brace should still be considered as a structural damage since
load transfer between adjacent girders would be affected, making
the bridge more vulnerable to loads, and in turn reducing the
overall bridge rating. An additional rating was conducted using the
calibrated 3D-FE model of PRE-725-0803 Bridge to simulate the
loss of cross-braces in an attempt to understand their significance.
The analyses yielded a reduction of 23% in the load rating as the
damaged cross-braces are excluded from the analytical model. The
design of cross-braces in the codes as well as standard drawings
should be revisited, based on the effect of cross-braces on the
overall bridge rating.
2D-grid-based rating of transverse members considers a
smeared beam element for the deck, cross-braces, and horizontal
member at the bottom (Fig. 3) and has inherent assumptions
for calculating demand stresses. The curvatures of the transverse
members are calculated by dividing moment on the transverse
members by corresponding calibrated EI values. Then, bottom
brace strain is calculated by multiplying curvature by the distance
between the bottom brace and calculated transverse members
neutral axis location, which is a function of the effective deck
width (btransverse
eff Fig. 3). Effective deck width calculation of AASHTO
10.38.3.1 [24] was implemented for the transverse members and
used as the minimum of (1) one-fourth of the span length of the
transverse member (i.e., width of the bridge), (2) the distance
center to center of cross-braces, (3) 12 times the least thickness
of the slab. Cross-brace rating calculations based on 3D-FE models
have better discritization with distinct elements modeled for
each member of cross-braces; therefore, have a better control on
calculating the member demands and capacities. Calculated cross-
brace rating factors using 2D-grid model are consistently 1.5 times
lower than the 3D-FEM-based model results, which indicates that
the neutral axis (NA) is actually at a lower position (i.e., farther
away than the deck) than the value calculated for transverse
members of the grid model. The NA shifting is most probably
due to the semi-composite action caused by diagonally placed
weak cross-braces and/or narrower deck width in the transverse
direction. Therefore, neutral axis location should be shifted down
and/or effective deck width should be taken smaller than the
AASHTO approach for the transverse members. The 2D-grid and
3D-FEM rating results of the cross-braces consistently remained
lower than the rating results of girders for the selected PRE-725-
0803 Bridge, which is an interesting observation showing that the
cross-braces might be more critical than girders of some bridges.
2.6. General discussions
2D-grid modeling is merely a step behind 3D-FE modeling and
may be considered as the next generation of analytical modeling
for bridge rating. Such models are a significant improvement
over simplistic frame-based (1D) line models. Grid models can
be automatically generated and calibrated by using measured
field data. Moreover, grid models have the advantage of modeling
the width of a bridge which is especially critical for skewed
bridges and transverse load transfer calculations. Fast analyses
with small number of optimization parameters (compared to 3D-
FEM), suitability for iteration and automated optimization, and
powerful numerical capabilities for bridge load rating at element
level are additional advantages of grid model-based bridge rating.
The rating results of calibrated 2D-grid model are lower than
the rating values of nominal 2D-grid model, which may be an
1274 A. Turer, B.M. Shahrooz / Engineering Structures 33 (2011) 12671276
Fig. 7. Correlation of analytical and experimental BGCI functions for the sample bridge.
indication of the deterioration over the 40 years of bridge life.
Although the visual rating is 6 out of 7, the deck might have
suffered from deterioration such as water penetration and salt for
deicing. Some of the rebars in the concrete deck could have been
rusted, and unintended composite action between the deck and
girders might have been partially lost. Section loss on the girders
and braces is usually not significant due to regular maintenance
and painting, but supports at the abutments often times suffer
extensive rusting due to rain and snow waters. The comprehensive
testing on the bridge is expected to capture global changes on
the structural properties and therefore reflect in to the analytical
model during calibration studies.
The staged optimization routine is an improvement over
defining a large number of calibration variables in the beginning of
the optimization process and trying to manipulate all the variables
at the same time. The use of logarithmic scale for ill-defined
variable sensitivity ranges (such as support spring stiffness values)
has also proved to be very useful.
The same mode shapes and modal frequencies can be obtained
for the same geometry with different scales of stiffness and
mass values. Hence, the objective function (OF) should contain
both dynamic and static error terms at the same time to avoid
convergence to wrong global mass and stiffness values.
Finally, the accuracy of the load rating results and reliability
of the proposed method is naturally dependent on the quality
of the measured data, accepted assumptions, and preferences
made for the objective function: the time invariance, observability,
reciprocity, level of dynamic excitation, linearity of a bridge,
changing and uncertain boundary conditions, etc. are important
aspects in field testing which might affect the experimental results.
Geometry simplification by reducing in to 2D and a grid structure,
linear behavior, damping ratio, and similar assumptions causes
only an approximate approach to analytical modeling. Uniqueness
of the calibrated analytical models can also be questioned since
success of the results can be dependent on factors such as
importance and weighing constants assigned during calibration,
grouping order and level, constraints defined for variables, terms in
the objective function definition, etc. Nevertheless, the automated
field-calibrated 2D-grid modeling is a clear improvement over
nominal analysis or 1D-beam modeling.
3. Conclusions
Two-dimensional (2D) grid modeling-based calibration and
load rating was proposed and illustrated on a pilot study. The
advantages of grid modeling and automated calibration are
 A. Turer, B.M. Shahrooz / Engineering Structures 33 (2011) 12671276 1275

a b

c d

e f

Fig. 8. Sample bridge 2D-grid model results (a) MAC before calibration; (b) frequencies before calibration; (c) MAC after calibration; (d) frequencies after calibration;
(e) inventory level ASD rating; and (f) operating level ASD rating.

discussed. The proposed modeling, calibration, and load rating was
illustrated using an actual bridge.
Modeling and rating of transverse members with 2D-grid
modeling helps determining the critical members with the lowest
rating factor. When PRE-725-0803 Bridge (constructed in 1968)
is taken as an example, the ASD inventory-based girder rating is
found to be around 2.0 while minimum cross-brace rating factor
is found in the range of 1.0, which is quite low compared to the
general inspection value of 6 given to the bridge. Rating studies
using the grid models draw attention to modeling and evaluation
of cross-braces. Even though loss of x-braces is commonly not
considered to be a structural damage, simulation of x-brace loss
yielded a further reduction of about 23% in minimum bridge load
rating.
The example application using a sample bridge shows that
the bridge rating factors have a tendency to decrease when the
calibrated grid model is used for rating, although visual rating was
6 out of 7. The lower rating factor from the calibrated model may
be an indication of bridge deterioration during its 40-year long life
span or improper assignment of the initial grid member structural
properties due to unreasonable assumptions such as the effective
deck width and deck-girder interaction in the grid model.
1276 A. Turer, B.M. Shahrooz / Engineering Structures 33 (2011) 12671276
Fig. 9. Nominal versus calibrated grid model rating results.
Fig. 10. General view of 3D-FE model.
Acknowledgements
The authors acknowledge contribution received for dynamic
testing and post-processing of dynamic data by Dr. Mike Lenett and
Dr. Necati Catbas. Other team members and who took part during
structural testing and/or contributed intellectually to the research
are: Drs. Emin Aktan, Alper Levi, Victor Hunt, Arthur Helmicki,
Raymond Barrish, and Kirk Grimmelsman. Their assistance is
gratefully acknowledged. This study was a part of a series of
research projects funded by Ohio-DOT and FHWA.
References
[1] US Department of Transportation. Federal highway administration. National
Bridge Inventory. NBI. 2006. Cited on January 2007.
http://www.fhwa.dot.gov/bridge/defbr05.htm.
[2] I-35W Mississippi river bridge. 2008. Cited on April 2008.
http://en.wikipedia.org/wiki/I-35W_Mississippi_River_Bridge.
[3] Kline J. Weight of excavator too much for rural bridge. The daily world. August
2007. Cited on May 2008.
http://www.thedailyworld.com/articles/2007/08/15/local_news/03news.txt.
[4] List of bridge disasters. Cited on May 2008.
http://en.wikipedia.org/wiki/List_of_bridge_disasters.
[5] Galambos CF. Highway bridge loadings. Eng Struct 1979;1:2305.
[6] Akgul F, Frangopol DM. Time-dependent interaction between load rating and
reliabilityof deteriorating bridges. Eng Struct 2004;26:175165.
[7] Faber MH, Val DV, Stewart MG. Proof load testing for bridge assessment and
upgrading. Eng Struct 2000;22:167789.
[8] American Association of State Highway and Transportation Officials. AASHTO.
Manual for condition evaluation of bridges. 1994.
[9] Jaramilla B, Huo S. Looking to load and resistance factor rating.
US FHWA public roads magazine. vol. 69. No. 1. JulyAugust 2005.
http://www.tfhrc.gov/pubrds/05jul/09.htm.
[10] AASHTOW are transportation software solutions. September 2006. Cited on
January 2007. http://aashtoware.org/?siteid=28&pageid=117.
[11] Turer A. Condition evaluation and load rating of steel stringer highway bridges
using field calibrated 2D-grid and 3D-FE models. Ph.D. thesis. University of
Cincinnati. Civil and Env. Engr. Dept. 2000.
[12] Janter T, Sas P. Uniqueness aspect of model-updating procedures. AIAA J 1989;
28(3):53843.
[13] Katafygiotis LS, Papadimitriou C, Lam H. A probabilistic approach to structural
model updating. Soil Dyn Earthq Eng 1998;17:495507.
[14] Christodoulou K, Ntotsios E, Papadimitriou C, Panetsos P. Structural model
updating and prediction variability using Pareto optimal models. Comput
Methods Appl Mech Engrg 2008;198:13849.
[15] Allemang RJ. The modal assurance criteriontwenty years of use and abuse.
Sound Vib Mag 2003.
[16] Nicoud YR, Raphael B, Burdet O, Smith IFC. Model identification of bridges
using measurement data. Comput-Aided Civ Infrastruct Eng 2005;20:11831.
[17] Haralampidis Y, Papadimitriou C, Pavlidou M. Multi-objective framework for
structural model identication. Earthq Eng Struct Dyn 2005;34:66585.
[18] Christodoulou K, Papadimitriou C. Structural identification based on optimally
weightedmodal residuals. Mech Syst Signal Process 2007;21:423.
[19] Spendley W, Hext GR, Himsworth FR. Sequential application of simplex
designs in optimization and evolutionary operation. Technometrics 1962;4:
44161.
[20] Nelder JA, Mead R. A simplex method for function minimization. Comput J
1965;7:30813.
[21] Turer A. Condition evaluation and load rating of steel stringer highway bridges
using field calibrated 2D-grid and 3D-FE models. Ph.D. thesis. University of
Cincinnati; 2000.
[22] Phillips AW, Allemang RJ. The complex mode indicator function (CMIF) as a
parameter estimation method. In: International modal analysis conference
proceedings. 1998. p. 10.
[23] Allemang RJ, Brown DL. A complete review of the complex mode indicator
function (CMIF) with applications. In: International conference on noise and
vibration engineering proceedings. 2006 p. 38.
[24] Standard specifications for highway bridges. 17th ed. Adopted by the
American Association of State Highway and Transportation Officials, AASHTO;
2002.