A Bayesian Approach For Estimating Characteristic Bridge Traffic Load Effects

Enright, B.1, Leahy, C.2 and OBrien, E.J.2&3

1
School of Civil Engineering, Dublin Institute of Technology, Bolton Street, Dublin 1, Ireland
2
Roughan & O'Donovan Innovative Solutions, Arena Road, Sandyford, Dublin 18, Ireland
3
School of Civil, Structural & Environmental Engineering, University College Dublin, Newstead, Dublin 4, Ireland
email: [email protected], [email protected], [email protected]

ABSTRACT: This paper investigates the use of Bayesian updating to improve estimates of characteristic bridge traffic loading.
Over recent years the use Weigh-In-Motion technologies has increased hugely. Large Weigh-In-Motion databases are now
available for multiple sites on many road networks. The objective of this work is to use data gathered throughout a road network
to improve site-specific estimates of bridge loading at a specific Weigh-In-Motion site on the network. Bayesian updating is a
mathematical framework for combining prior knowledge with new sample data. The approach is applied here to bridge loading
using a database of 81.6 million truck records, gathered at 19 sites in the US. The database represents the prior knowledge of
loading throughout the road network and a new site on the network is simulated. The Bayesian approach is compared with a
non-Bayesian approach, which uses only the site-specific data, and the results compared. It is found that the Bayesian approach
significantly improves the accuracy of estimates of 75-year loading and, in particular, considerably reduces the standard
deviation of the error. With the proposed approach less site-specific WIM data is required to obtain an accurate estimate of
loading. This is particularly useful where there is concern over an existing bridge and accurate estimates of loading are required
as a matter of urgency.

KEY WORDS: Bayesian; Updating; Bridge; Characteristic; Traffic; Load; Modelling.

1 INTRODUCTION of the measured traffic and a new stream of traffic is

Accurate modelling of bridge traffic loading is critically simulated [10–12]. Hundreds or thousands of years of traffic
important in bridge engineering. At the design stage it allows can be simulated in this way and the characteristic load effects
the design of bridges which are fit for purpose, while reducing can be obtained directly from the simulated traffic, without
the waste associated with overdesign. In the assessment of the need for extrapolation. These long run simulations are
existing structures, estimates of traffic loading are possibly time consuming but they allow for loading scenarios, such as
more important. Where there is concern over an existing truck meeting or overtaking events, which did not occur
bridge, this information will determine whether a bridge needs during the WIM measuring period.
to be repaired or replaced. If a bridge is saved as a result, WIM data is becoming increasingly more common and this
significant cost savings can be made. is reflected in recent bridge loading literature. Enright et al.
The most accurate method for modelling bridge traffic [13] used 21 million trucks from 11 sites to examine long span
loading is to use Weigh-In-Motion (WIM) data. WIM systems bridge loading in Alabama. Nowak & Rakoczy [14] simulated
measure truck weights and axle configurations as they pass bridge loading using 35 million truck records from 32 sites in
along a road at normal highway speeds [1]. Statistical the US and Fu & You [15] examine multi-truck loading
methods are generally used to extrapolate from the relatively events using 63 million truck records gathered in New York.
short WIM measurement period to the return period used for This paper develops a method for using the large amounts
bridge design/assessment. Return periods of 75 years [2] or of data which are now available for many road networks to
1000 years [3] are commonly used. help estimate loading at a new site on the same network.
A common extrapolation approach is to fit a statistical Bayesian updating is used as it provides a mathematical
distribution to the measured data and to extrapolate to the framework for combining prior knowledge about a certain
required return period. The generalized extreme value statistical property with new sample data [16]. This is useful
distribution is a popular choice in the literature. This in engineering where data from various sources must be
distribution can be separated into a family of three extreme combined in the decision-making process and has been used
value distributions, comprising of the Gumbel (type I), for many civil engineering applications. For example, Hong
Fréchet (type II) and Weibull (type III) distributions. The type and Prozzi [17] use this approach to update the parameter
I Gumbel distribution is used by some authors [4], [5] but the distributions for a pavement performance model. Enright and
type III Weibull distribution is perhaps the most common [6– Frangopol [18] use it to predict the deterioration of concrete
9]. These extreme value distributions must be fitted to block bridges as it allows the results of bridge inspections to be
maxima, with maximum daily or weekly values often used. combined with prior information to obtain improved
As an alternative to direct extrapolation from WIM data, predictions. Other studies [19], [20] also apply a Bayesian
Monte Carlo simulations can also be used. With these approach to improve the accuracy of structural engineering
simulations, statistical distributions are fitted to the properties models.
 In this work 81.6 million trucks from 19 sites in the US are underlying random variable X (in this case the maximum
used to test the approach. These truck records provide the weekly GVW), and this cdf can be used to calculate
prior knowledge about loading in the US. This prior data is characteristic values:
then combined with new site-specific data, using the Bayesian
approach, in order to obtain improved estimates of
characteristic loading at the new site. The Bayesian approach
(4)
is compared with a Weibull extrapolation approach, which
uses only site-specific data, and the improvements in accuracy
examined. The proposed approach reduces the amount of
WIM data required to make accurate estimates of loading. The prior distribution of the parameter vector is obtained by
This is very useful where there is concern over the safety of fitting a Weibull distribution to each site-year of prior data.
an existing bridge and an estimate of loading is required in The data from the Colorado site is used to simulate the new
order to make a prompt decision about the future of the site-specific data and is not included in the prior data. This
bridge. new site-specific data is modelled by randomly sampling 25
non-consecutive weeks of data from the Colorado site.
2 METHODOLOGY Bayesian updating is then used obtain the posterior
2.1 Bayesian updating distribution of the parameter vector and hence the cdf for
loading at that site. For the Bayesian approach, the Weibull
A random variable (maximum weekly GVW) is described by distribution must be fitted to the entire distribution, rather than
a parameter vector θ (e.g. for a Weibull distribution with three the tail of the maximum daily values, which is commonly
parameters, ). However, these parameters are used [8], [21], [22]. With the aim of examining the trend in
themselves random variables due to the uncertainty associated this tail region, the Weibull distribution is fitted to maximum
with parameter estimation. These random variables have an weekly GVWs, rather than maximum daily. GVW is used
assumed ‘prior’ distribution '(θ). When a new set of sample here to assess the proposed method but the same approach
data, x, is acquired this prior distribution can be updated to could easily be applied to bridge load effects required for
obtain the ‘posterior’ distribution ''(θ). The posterior assessment.
distribution is calculated as:
The Generalized Extreme Value (GEV) distribution is given
by:
(1)
(5)

where ) = conditional probability, or likelihood , defined on {x : 1 + ξ(x - μ)/σ > 0}, where:
of the new data x, given θ.
μ = location parameter, -∞ < μ < ∞
The denominator is a normalising constant which ensures that σ = scale parameter, σ > 0
is a complete probability density function (PDF). As a ξ = shape parameter
result Eq. (1) can be rewritten as:
The Weibull distribution is a subset of the GEV distribution,
where the shape parameter ξ is negative. The location and
(2) scale parameters are similar to mean and standard deviation
and the shape parameter describes the behaviour in the tail of
the distribution.
This is the approach used here as it avoids the
computationally intensive calculation of the denominator
. The likelihood function is calculated for 2.2 WIM Data
the new set of data as: The prior data used here is WIM data from the United States
Federal Highway Administration’s Long-Term Pavement
Performance (LTPP) Program. Initially the LTPP collected
WIM data with inconsistent quality control measures [23]. In
(3) 1999 a plan was developed which, among other things,
improved and centralised quality control. This led to a
significant improvement in WIM data reliability and since
The posterior is often used to obtain the expected value 2003 ‘research quality’ WIM data is being collected at 28 of
of θ, which gives a point estimator of the parameter vector. the Specific Pavement Studies LTPP sites. The 81.6 million
However the approach used in this work includes the truck records were collected over the period 2005-2012 and
uncertainty associated with estimating the value of θ. The are detailed in Table 1.
posterior distribution of the parameter vector can be used to
calculate the cumulative distribution function (cdf) for the
 Table 1. WIM Data
Years of Trucks/ MMW 75yr 8
Site State Data Weekday GVW2
1 Arizona 5.7 575 63
2 Arizona 5.7 4,988 93 6

-ln ( -ln (p) )

3 Arkansas 6.0 5,526 100
4 California 5.0 5,939 68 4
51 Colorado 6.5 1,473 73
6 Delaware 5.5 930 66
2
7 Illinois 7.4 3,139 98
8 Indiana 4.5 1,489 69
9 Kansas 6.6 1,851 81 0
Prior Data
10 Louisiana 4.9 506 77
Test Site
11 Maine 4.4 835 67 -2
12 Maryland 6.8 1,030 51 50 100 150 200
13 Minnesota 6.2 316 67 GVW (t)
14 New Mexico 4.7 716 61
15 New Mexico 4.7 2,934 89 Figure 1. Weibull distributions of the prior data alongside the
16 Pennsylvania 5.6 5,315 90 Illinois site used to test the proposed approach.
17 Tennessee 5.6 5,474 89
18 Virginia 6.0 1,082 63
19 Wisconsin 5.3 987 73
1
Site used to test the Bayesian approach
2
MMW GVW: mean maximum weekly gross vehicle weight
(a good measure of intensity of loading at a site [24])

Bridges in the US are designed and assessed using the HL-93

load model [2], [25]. This model applies to “normal vehicular
use” (all legal trucks, illegal overloads and un-analysed
routine permits) and does not include special permit vehicles.
In order to describe a methodology which can be used to
examine “normal vehicular use”, special permit vehicles are
removed from the database. These vehicles are identified in
the WIM data using filtering rules, as described by Leahy
[26].
2.3 Estimating the prior distribution using kernel density
estimators (KDE)
The Colorado site is used as the test data for the Bayesian
approach. The other 18 sites are used as the prior data. Four
years of data are used from each site as this amount is
available for each site and gives each site an equal weighting.
A Weibull distribution is fitted to the maximum weekly Figure 2. Prior data used with kernel density estimated
GVWs for each site-year of prior data – see Figure 1. This distribution for different ranges of shape parameter ξ.
gives 72 sets of Weibull parameters ( ). These parameter
values are shown in Figure 2. In this figure, values of σ are A continuous distribution, rather than discrete points, is
plotted against μ for different ranges of the shape factor, ξ. needed for the prior distribution. In order to fit a distribution
Figure 1 also shows the Colorado site used to test the to the 72 data points, kerned density estimators (KDEs) are
Bayesian approach. It can be seen that the loading at the used. With the KDE method, each data point is replaced with
Colorado site agrees well with the general trend of the prior a component density. These densities are known as kernel
data. As a result it is expected that the proposed method will functions and are added to obtain an estimate of the overall
perform well with this test site. distribution. Any distribution can be used for the kernel
functions but the normal distribution is a popular choice and is
used here. The product kernel method, as described by Scott
[27], is used. The kernel function is essentially a trivariate
normal distribution where all variables are independent. The
estimated distribution is shown in Figure 2. For visualisation
purposes, the distribution is plotted for different ranges of the
shape parameter ξ.
In the Bayesian calculations, the weighting of the prior data,
with respect to the new data, is determined by the bandwidth
of the kernel function. The bandwidth, in this case, refers to
the variance of the trivariate normal distribution, where each
of the three variables has its own independent standard
deviation. If the bandwidth is too large, oversmoothing will
occur and the trends in the prior data will be lost. On the other
hand, if the bandwidth is too small the posterior distribution
will be restricted to match very closely to the prior data.
Optimal bandwidths exist for sample data which matches a
certain theoretical distribution, but this is not the case here. In
this work a trial and error approach, as recommended by
Silverman [28], is used and an optimal value selected by
visual inspection of the marginal distributions – see Figure 3. (c) Shape parameter
Bandwidth values of three tonnes, one tonne and 0.04 are used Figure 3. Histograms of marginal distributions of prior data
for μ, σ and ξ, respectively. The selection of these values is with the kernel density estimated distribution.
somewhat subjective but when applying this approach 2.4 Assessing accuracy of proposed Bayesian approach
bandwidth selection should be carefully considered as it does
influence the results. The Bayesian approach is assessed to determine its accuracy
in estimating the 75-year load. 75 years is the design life used
in US bridge design [2]. The benchmark 75-year load is first
calculated by fitting a Weibull distribution to all the data
available at the Colorado site (6.5 years) and extrapolating to
the 75-year load – see Figure 4. This approach assumes that
this extrapolation from 6.5 year of data represents the true 75-
year loading.
The accuracy of the method is assessed by randomly
selecting 25 non-consecutive maximum weekly values from
the Colorado data. The Bayesian approach, using the prior
data, is then applied to this sample dataset and the 75-year
load estimated. The accuracy of the estimate is then
determined by comparison with the benchmark.
A non-Bayesian approach, where no prior data is used, is
then applied to the same sample data. A Weibull distribution
(a) Location parameter is fitted directly to the 25-week sample. The accuracy in
estimating the 75-year value can then be compared with the
Bayesian approach – see Figure 4. This process is then
repeated for 100 different random datasets in order to assess
the standard deviation of the error.

(b) Scale parameter

Figure 4. Example of the Bayesian and Weibull approach

applied to a 25-week sample from the Colorado site (“Site
data” includes the weekly maxima from 6.5 years of
measurements at the site).
3 RESULTS loading are required to make a prompt decision about the
The errors for both the Bayesian and non-Bayesian future of the bridge.
approaches are shown in Figure 5 and Table 2. A mean error The method performs well with the Colorado site used to
of zero with a small standard deviation is preferable. It can be test it. However, this site has similar loading to the general
seen for the non-Bayesian approach, where no prior data is trend in the prior data. It is possible that the Bayesian
used, that there is large variation in the errors. This variation approach may not perform as well at a site with unusually low
is not desirable when estimating characteristic loading on or high levels of loading as the method may tend to push
bridges as it may result in an estimate of loading which is estimates of loading closer to that of an average site.
significantly greater or less than the true value. This variation The proposed Bayesian approach examined here used a
is significantly reduced with the Bayesian approach, with the large WIM database as the prior data. The WIM
standard deviation reducing from 19.3% to 4.4%. The mean measurements were gathered at 18 WIM sites which were
error is also improved from 7.8% to 2.3%. subject to strict quality control measures. The results of the
method depend on the extent of good quality prior data which
is available. It is acknowledged such a large database of good
0.08 quality prior data may often not be available when assessing
Non-Bayesian bridge traffic loading and that the improvement in accuracy
0.07
Bayesian may not be as significant if less prior data is available.
0.06
ACKNOWLEDGMENTS
Frequency

0.05
The authors gratefully acknowledge the Federal Highway
0.04 Administration’s Long-Term Pavement Performance Program
0.03 for access to an extensive WIM database and the Irish
National Roads Authority for their financial support.
0.02

0.01 REFERENCES
[1] COST323, Weigh-in-Motion of Road Vehicles - Final Report. PARIS:
0 LCPC, 2002.
-60 -40 -20 0 20 40 60
[2] AASHTO, AASHTO LRFD Bridge Design Specifications, 5th ed.
Error (%) Washington D.C.: American Association of State Highway and
Transportation Officials, 2010.
Figure 5. Errors in estimating 75-year load for the Bayesian [3] EC1, Eurocode 1: Actions on structures, Part 2: Traffic loads on
and non-Bayesian approaches, for 100 sample datasets. bridges. European Standard EN 1991-2:2003. Brussels: European
Committee for Standardization, 2003.
[4] D. I. Cooper, “Development of short span bridge-specific assessment
live loading,” in Safety of Bridges, P. C. Das, Ed. London: Thomas
Table 2. Error with Bayesian and non-Bayesian approaches Telford, 1997, 64-89.
[5] A. O’Connor and E. Eichinger, “Site-specific traffic load modelling for
Mean Standard Deviation bridge assessment,” Proceedings of The Institution of Civil Engineers.
Bayesian 2.3% 4.4% Bridge Engineering, vol. 160, no. 4, 185-194, 2007.
Non-Bayesian 7.8% 19.3% [6] S. F. Bailey and R. Bez, “Site specific probability distribution of
extreme traffic action effects,” Probabilistic engineering mechanics,
vol. 14, no. 1, 19-26, 1999.
[7] T. J. Miao and T. H. T. Chan, “Bridge live load models from WIM
4 CONCLUSION data,” Engineering Structures, vol. 24, no. 8, 1071-1084, 2002.
[8] E. J. OBrien, B. Enright, and A. Getachew, “Importance of the tail in
A Bayesian methodology is proposed for obtaining additional truck weight modelling for bridge assessment,” ASCE Journal of
value from the large amount of WIM data which are available Bridge Engineering, vol. 15, no. 2, 210-213, 2010.
for multiple sites on many road networks. The approach uses [9] S. A. Graves, E. J. OBrien, and A. O’Connor, “The Determination of
Site-Specific Imposed Traffic Loadings on Existing Bridges,” in M. J.
this existing, or ‘prior’, network data to help predict site- Ryal, G. A. R. Parke, & J. E. Harding. (Eds.), The Fourth International
specific loading at a new site on the network. The prior data Conference on Bridge Management, London, 2000.
used here is from 18 WIM sites in the US. This WIM data is [10] B. Enright and E. J. OBrien, “Monte Carlo simulation of extreme
used to determine the prior distribution of the parameters of a traffic loading on short and medium span bridges,” Structure and
Infrastructure Engineering, vol. 9, no. 12, 1267-1282, 2013.
Weibull distribution and this prior distribution is updated [11] A. O’Connor and E. J. OBrien, “Traffic load modelling and factors
using the new site-specific data. The Bayesian approach is influencing the accuracy of predicted extremes,” Canadian Journal of
compared with the alternative of fitting a Weibull distribution Civil Engineering, vol. 32, no. 1, 270-278, 2005.
directly to the site-specific data without using any prior data. [12] B. Sivakumar, M. Ghosn, and F. Moses, Protocols for Collecting and
Using Traffic Data in Bridge Design - NCHRP Report 683.
The errors in estimating characteristic loading are compared
Washington D.C.: Transport Research Board, 2008.
for the proposed Bayesian approach and the non-Bayesian [13] B. Enright, C. Carey, and C. C. Caprani, “Microsimulation Evaluation
approach. The Bayesian approach achieves significant of Eurocode Load Model for American Long-Span Bridges,” Journal
reductions in the variation of the error while also improving of Bridge Engineering, vol. 18, no. 12, 1252-1260, 2013.
[14] A. Nowak and P. Rakoczy, “WIM Based Simulation Model Of Site
the mean of the error. The proposed method allows accurate
Specific Live Load Effect On The Bridges,” in B. Jacob, A. H.
estimates of loading to be achieved with less site-specific Mcdonnell, F. Schmidt, & W. Cunagin (Eds.), 6th International
WIM data. This is particularly useful where there is concern Conference on Weigh-In-Motion, Dallas, 2012, 352-358.
over the safety of an existing bridge and accurate estimates of
[15] G. Fu and J. You, “Truck load modeling and bridge code calibration,”
in Faber, Kohler, & Nishijima (Eds.), Applications of Statistics and
Probability in Civil Engineering, London, 2011, 406-413.
[16] A. H.-S. Ang and W. H. Tang, Probability concepts in engineering :
emphasis on applications in civil & environmental engineering, 2nd ed.
Wiley, New York, 2007.
[17] F. Hong and J. A. Prozzi, “Estimation of pavement performance
deterioration using Bayesian approach,” Journal of infrastructure
systems, vol. 12, no. 2, 77-86, 2006.
[18] M. P. Enright and D. M. Frangopol, “Condition prediction of
deteriorating concrete bridges using Bayesian updating,” Journal of
Structural Engineering, vol. 125, no. 10, 1118-1125, 1999.
[19] T. Igusa, S. G. Buonopane, and B. R. Ellingwood, “Bayesian analysis
of uncertainty for structural engineering applications,” Structural
Safety, vol. 24, no. 2, 165-186, 2002.
[20] J. L. Beck and S.-K. Au, “Bayesian updating of structural models and
reliability using Markov chain Monte Carlo simulation,” Journal of
Engineering Mechanics, vol. 128, no. 4, 380-391, 2002.
[21] B. Enright, “Simulation of traffic loading on highway bridges,” PhD
Thesis, University College Dublin, 2010.
[22] OBrien, E. J., Schmidt, F., Hajializadeh, D., Zhou, X.-Y., Enright, B.,
Caprani, C. C., … Sheils, E., “A Review of Probabilistic Methods of
Assessment of Load Effects in Bridges,” Submitted for publication,
2013.
[23] D. Walker and D. Cebon, “The Metamorphosis of LTPP Traffic Data,”
in B. Jacob, A. H. Mcdonnell, F. Schmidt, & W. Cunagin (Eds.), 6th
International Conference on Weigh-In-Motion, Dallas, 2012, 242 -
249.
[24] E. J. OBrien and B. Enright, “Using Weigh-In-Motion Data to
Determine Aggressiveness of Traffic for Bridge Loading,” Journal of
Bridge Engineering, ASCE, vol. 18, no. 3, 232-239, 2012.
[25] AASHTO, The Manual for Bridge Evaluation, 2nd ed. Washington
D.C.: American Association of State Highway and Transportation
Officials, 2011.
[26] C. Leahy, “Predicting Extreme Traffic Loading on Bridges Using
Weigh-In-Motion Data,” PhD Thesis, University College Dublin, 2014.
[27] D. W. Scott, Multivariate density estimation: theory, practice, and
visualization, vol. 383. John Wiley & Sons, 1992.
[28] B. W. Silverman, Density estimation for statistics and data analysis,
vol. 26. London: Chapman & Hall, 1986.