International Journal of Fatigue 95 (2017) 104–119

Contents lists available at ScienceDirect

International Journal of Fatigue

journal homepage: www.elsevier.com/locate/ijfatigue

Data-based models for fatigue reliability of orthotropic steel bridge

decks based on temperature, traffic and strain monitoring
Isaac Farreras-Alcover a,⇑, Marios K. Chryssanthopoulos b, Jacob E. Andersen a
a
COWI A/S, Kgs. Lyngby, Denmark
b
University of Surrey, UK

a r t i c l e i n f o a b s t r a c t

Article history: A novel methodology is presented for probabilistic fatigue life prediction of welded joints in orthotropic
Received 27 June 2016 bridge steel decks. Monitoring data were used to specify time-series model parameters for the main dri-
Received in revised form 22 September vers of fatigue damage in such structures, namely pavement temperatures and heavy traffic intensities,
2016
which influence the stress range distributions at critical locations. Polynomial regression models were
Accepted 26 September 2016
Available online 28 September 2016
developed to quantify the relationship between fatigue loading, derived using S-N principles from strain
measurements at welded joints, with pavement temperatures and heavy traffic counts. The different
models were integrated within a fatigue reliability framework, in which the uncertainties arising from
Keywords:
Fatigue
material properties and fatigue damage at failure were modelled via random variables. A Monte Carlo
Steel bridges scheme was then deployed to predict S-N fatigue damage using the fatigue loading regression models
Orthotropic steel decks and simulated time-series of heavy traffic and pavement temperatures. Thus, fatigue reliability profiles
Structural Health Monitoring were generated, which accounted for different scenarios in terms of future changes in traffic and pave-
Regression analysis ment temperature. The proposed methodology was illustrated considering actual monitoring outcomes
Time-series models from the Great Belt Bridge (Denmark) with reliability profiles developed for both ‘baseline’ and ‘adverse’
Structural reliability scenarios in the context of asset integrity management. The combined effect of higher temperature and
Asset management
heavy traffic levels was shown to result in considerable reductions in fatigue reliability, with a commonly
used threshold being reached up to 40 years earlier compared to the baseline ‘no change’ scenario.
However, this reduction was not uniform for all the fatigue details considered, emphasizing the impor-
tance of monitoring different locations, based on a thorough understanding of the fatigue behaviour of
the orthotropic steel deck.
Ó 2016 Elsevier Ltd. All rights reserved.

1. Introduction mental conditions, loadings and structural responses from civil

Existing bridge infrastructure deteriorates with time due to the
inherent aging of materials combined, in many cases, with increas-
ing operational demands and severe environmental conditions.
This leads to a situation where the service life of some structures
is reached earlier than anticipated, thus compromising their
functionality and, potentially, the safety of the users. ASCE [1]
has estimated, for instance, that approximately 11% of the more
than 600,000 bridges currently operational in USA can be rated
as structurally deficient.
As a result, there is an increasing need for accurate assessment
methods leading to a better allocation of the limited available
resources. Nowadays, reliable monitoring data concerning environ-
⇑ Corresponding author.
E-mail addresses: ifal@cowi.dk (I. Farreras-Alcover), mkchry@surrey.ac.uk
(M.K. Chryssanthopoulos), jca@cowi.dk (J.E. Andersen).
infrastructure assets can be acquired. This has put forward an
alternative paradigm for assessing and predicting structural per-
formance, promoting advances in the field of Structural Health
Monitoring (SHM). One of the underlying premises in SHM
approaches is that the data obtained through monitoring systems
can be used to reduce epistemic uncertainties associated with
the deterioration and loading processes leading, in turn, to more
accurate structural performance assessments and service life
predictions.
Focusing on the challenging case of fatigue assessment and
motivated by the above advancements, several studies deal with
the potential of utilising monitoring data to predict remaining fati-
gue lives [2]. In order to treat both aleatoric (e.g. material proper-
ties, stress range spectra, etc.) and epistemic (e.g. fatigue damage
model) uncertainties embodied in the fatigue assessment process,
monitoring outcomes can be integrated within a probabilistic
framework underpinned by fatigue reliability methods. For

http://dx.doi.org/10.1016/j.ijfatigue.2016.09.019
0142-1123/Ó 2016 Elsevier Ltd. All rights reserved.
 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119
example, combining field measurements and structural reliability
analysis, Frangopol et al. [3] assessed the fatigue reliability of an
existing highway bridge using a monitoring period of 11 days by
comparing the allowable effective stress range with a stress range
distribution fitted to SHM data. Ni et al. [4] developed a fatigue
reliability model in relation to monitoring strain data from the
Tsing Ma Bridge (China), whereas Guo et al. [5] used the data pro-
vided by the Weigh-in-Motion (WIM) system installed in the
Throgs Neck Bridge (USA) to characterize probability distributions
of axle weights, axle spacings and vehicle positions.
In the case of orthotropic steel decks subject to fatigue loading,
cracks can initiate in the welds between the deck plate and the lon-
gitudinal trough stiffener, in longitudinal stiffener splice joints, or
in cross-beam to longitudinal stiffener connections. De Jong [6]
has drawn attention to typical fatigue-induced cracks observed in
a number of OSDs employed in both fixed and movable bridges
in the Netherlands, whereas Kolstein [7] has classified OSD welded
joints according to their S-N fatigue performance. Xiao et al. [8]
studied fatigue cracks at trough splice joints and Ya et al. [9] at
trough-to-deck joints. For a comprehensive overview of crack
types in OSDs, the reader is referred to De Jong [10]. By way of
example, Fig. 1 shows fatigue cracks in two typical details of OSD
welded joints, i.e. longitudinal stiffener splices and stiffener-to-
deck welded joints.
In general, the fatigue life of an OSD is a function of the specific
configuration chosen, i.e. stiffener geometry and connection types,
the nature of the traffic load and the characteristics of the surfacing
layer(s). With regard to the role of the latter, early research has
demonstrated that it acts compositely with the steel deck plate,
see for example Cullimore et al. [11]. This leads to a reduction in
stresses experienced by the steel deck, attributed in part to the
composite action but also due to the load dispersal enacted
through the thickness of the surfacing layer(s). It is worth noting
that these materials, broadly divided into bituminous and
polymer-based, are characterised by visco-elastic behaviour,
which, for any given composition, is influenced by temperature
variations and loading frequency and leads to non-linear
through-thickness strain distributions. Thus, it is important to
develop appropriate means for taking into account the interaction
between surfacing and steel deck under actual conditions, in order
to support design choices (material and thickness specification), as
well as make realistic fatigue assessments. In this context,
researchers have, in the past decade or so, developed both finite
element and theoretical models, e.g. De Jong [10] and Kolstein
et al. [12] respectively, to quantify the impact of the pavement
layer on the stresses developed at the steel deck under wheel type
loading. In these studies, the temperature dependence of the sur-
Fig. 1a. Typical cracks in longitudinal stiffener (trough) splices, from De Jong [10].
105
facing stiffness and its concomitant effect on stress levels in OSD
welded joints was emphasized. In turn, this plays an important role
in determining the fatigue life of OSD welded joints [12]. With the
increasing utilisation of sensors, recent studies have processed
field measurements of stress ranges at OSD welded joints and
have confirmed the influence of temperature on these quantities
[13–15].
The complexity in determining strain cycles at welded joints of
orthotropic steel decks (OSD) has also prompted the use of SHM for
their fatigue assessment [2,13,14]. Guo et al. [16] used monthly-
averaged monitoring data to determine the remaining fatigue life
at OSD welded joints. The proposed method, however, did not
account for the effect of uncertainties relating to the monitored
strain levels, nor for the joint effect of future traffic and tempera-
ture levels. These effects were considered by Sugioka et al. [17],
where the remaining fatigue life of OSD welded joints was esti-
mated in probabilistic terms within an S-N approach, via a MCS
accounting for temperature variations and traffic intensities. How-
ever, the methodology therein did not utilise monitoring data to
derive the stress ranges at selected welded joints, but relied
instead on a simplified analytical model. In consequence, the
effects arising from the stochastic nature of the traffic load and
the complexities associated with visco-elastic behaviour of surfac-
ing layers were not captured.
To overcome the above limitations, the present paper develops
a novel methodology for probabilistic fatigue life prediction of
welded joints in orthotropic steel bridge decks. From a SHM per-
spective, it can be regarded as a local, data-based approach to fati-
gue life prognostics. Firstly, monitoring outcomes are used to
specify time-series model parameters for the main drivers of fati-
gue damage in orthotropic steel decks, i.e. pavement temperatures
and heavy traffic intensities, which influence the time-varying
stresses at a number of fatigue critical details. Polynomial regres-
sion models [15] are then developed to quantify the relationship
between a dependent variable, namely the relevant fatigue loading
for an S-N approach using strain measurements at welded joints,
and two independent variables, namely pavement temperatures
and heavy traffic counts. The different developed models are inte-
grated within a fatigue reliability framework, in which the uncer-
tainties arising from material properties and fatigue damage at
failure are modelled via random variables. A Monte Carlo scheme
is then deployed to predict S-N fatigue damage using the fatigue
loading regression models and simulated time-series of heavy traf-
fic and pavement temperatures. Thus, fatigue reliability profiles are
generated, which can account for different scenarios in terms of
future changes in traffic and pavement temperature. The proposed
methodology is illustrated considering actual monitoring out-
106 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119
Fig. 1b. Typical cracks in longitudinal stiffener-to-deck welds, from De Jong [10].
comes from the Great Belt Bridge (Denmark) with reliability pro-
files developed for both ‘baseline’ and ‘adverse’ scenarios. The main
novelty of this work lies in modelling independently the random
processes of pavement temperatures and heavy traffic counts,
and using these developed models to simulate the time-varying
S-N fatigue loading. This makes it possible to quantify the effect
of different scenarios on fatigue life predictions, thus facilitating
the development of inspection and maintenance strategies within
life-cycle asset integrity management.
2. The SHM system of the Great Belt Bridge
The Great Belt Bridge (see Figs. 2 and 3) is a suspension bridge
in Denmark that opened in 1998. It has a main span of 1624 m and
a maximum hanger length of 177 m. Its cross-section is formed by
a closed steel box girder with an orthotropic deck, formed by lon-
gitudinal troughs and cross beams spaced every 4 m. The deck
plate thickness is 12 mm and is covered by a surfacing layer which
consists of four main layers: a 300 g/m2 of primer applied to the
cleaned blasted steel surface to create adhesion between the bitu-
men surfacing and the steel deck, a 4 mm bituminous mastic, a
25 mm mastic asphalt with maximum aggregates size of 8 mm
(intermediate layer) and a 30 mm mastic asphalt with maximum
aggregates size of 12 mm (wearing course).
An extensive Structural Health Monitoring system was installed
on the bridge, recording simultaneously strains at selected welded
joints of the OSD, pavement temperatures and traffic intensities.
Given the earlier remarks regarding the complex interactions that
influence the behaviour of the steel deck, this monitoring scheme
paves the way for data-based models which can quantify the struc-
Fig. 2. Cross section of the suspended deck of the Great Belt Bridge.
tural response at pre-selected locations, in terms of strain, as a
function of time-varying environmental and mechanical loading,
typified by changes in temperature and traffic. The system has
been operational for some years, thus providing a suitable platform
for model development aimed at fatigue assessment and remain-
ing life prediction. The monitoring data could also be used to vali-
date numerical and theoretical models of the OSD system for
future design and assessment purposes, though this objective is
not pursued in the current paper.
2.1. Pavement temperature
The pavement temperature monitoring system is installed at
two different cross sections, see Fig. 3. Each cross section is instru-
mented with two sensors placed in the middle of the slow lane;
one in the northern lane (T1 and T3) and another in the southern
lane (sensors T2 and T4). The sensors are embedded 1 cm into
the pavement wearing course and record pavement temperatures
every five minutes.
2.2. Vehicle traffic
Since the bridge inauguration, the traffic on the Great Belt
Bridge has been continuously increasing, exceeding an annual bi-
directional traffic flow of 10 million vehicles since 2006. The cross-
ing vehicles are automatically classified at the toll locations into
different vehicle classes according to their dimensions, as summa-
rized in Table 1.
The outcome of the traffic monitoring system consists of
hourly-aggregated vehicle traffic counts corresponding to the dif-
 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119 107

W E
Strain section:
Temp. section: T1,T2 SG1 to SG15 Temp. section: T3,T4

1624 m
Fig. 3. Location of the pavement temperature and strain sensors.

Table 1
Characteristics of the different vehicle classes.

Vehicle class Length range [m] Height range [m] Approx. vehicle type
1 0–3 No limit Motorcycle
2 3–6 No limit Car
3 6–20 <2,8 Car with trailer
4 6–10 >2,8 Van
5 10–20 >2,8 Truck
6 >20 No limit Articulated truck

ferent vehicle classes described in Table 1. As such, no vehicle

weighing is performed. The installation of a weigh-in-motion sta-
tion installed in the vicinity of the section instrumented with strain
gauges would result in quantification of each vehicle crossing (e.g.
vehicle velocity, axle weights distribution, etc.). This would enable
more advanced analyses to be performed in relation to the charac-
terisation of traffic-induced fatigue damage.
Fig. 5. Instrumented welds.

2.3. Strain
Fig. 5 shows one particular detail of the strain gauge system,
together with the data acquisition units. The strain gauge data
The strain monitoring system of the bridge consists of an instru-
are sampled at a frequency of 100 Hz in order to capture the strain
mented cross section, spatially positioned between the Eastern
cycles induced by the different axle configurations of the vehicles,
tower and the anchorage location as shown in Fig. 3, intended at
as well as possible dynamic amplification effects arising from the
monitoring nominal stresses. The cross section is monitored by
vehicle-pavement interaction. Strain records are converted into
15 uniaxial strain gauges (SGs) placed under the traffic lanes head-
stresses by assuming linear elastic behaviour. The values of the
ing eastwards, see Fig. 4. Ten strain gauges, namely SG numbers 1,
monitored strains capture only the effect of live loads, since the
3, 4, 6, 7, 9, 10, 12, 13 and 15 are used to monitor the transverse
SGs were installed when the bridge was already in operation
nominal strains at the trough-to-deck weld and are placed at the
(and, hence, dead load effects were already present). However,
mid-point between cross-beams. A further five gauges, namely
due to the presence of high tensile residual stresses in welded
SG numbers 2, 5, 8, 11 and 14 are used to monitor the longitudinal
joints, typically of yield point magnitude, fatigue strength is gov-
nominal strains at trough splice welds. They are placed at a dis-
erned by the applied stress range regardless of the associated
tance of 0.5 m (measured following the longitudinal direction of
stress ratio [18]. The former can be accurately estimated using
the troughs, see Fig. 5) from the nearest cross-beam. The SGs num-
the monitored strain histories, even though the stress ratio, which
bered from 1 to 9 are placed under the slow traffic lane, whereas
also depends on the dead load stresses, is not determined.
the rest are placed under the fast traffic lane (shown as blue).

Fast lane Slow lane

Fig. 4. Location of strain gauges.

108 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119

Fig. 6a shows a realization of the strain time series monitored at × 10 -5

SG7 on 15/05/2011. An overall trend is observed in the strain val-
ues, which is caused by temperature variations experienced by the 8 2nd axle 3rd axle
orthotropic deck. A similar behaviour was observed in Ni et al. [4]. 4th axle
6 1st axle 5th axle
Fig. 6b shows 100 min of strain data from the same day at SG7

Strain
starting at 10:00 AM. The passage of several heavy vehicles can 4
be clearly inferred from the presence of sharp spikes. Finally,
Fig. 7 traces the strain profile during the passage of a single heavy 2
vehicle over SG7. It can be concluded that the 100 Hz sampling fre-
0
quency enables the identification of individual strain cycles due to
the passage of vehicle axles over a specific point. This is of utmost -2
importance in determining correctly the stress range histograms, 0 0.5 1.0 1.5
which are required for fatigue assessment, especially for the Time [seconds]
multi-axle class 5 and 6 heavy vehicles.
Fig. 7. Stress time-history during the passage of a heavy vehicle at SG7.
It is noted that a weigh-in-motion station installed in the vicin-
ity of the section instrumented with strain gauges would enable to
characterize quantitatively each vehicle crossing (e.g. vehicle Fig. 9 shows a scatter plot of the hourly maximum stress ranges
velocity, axle weights distribution, etc.). This could enable the per- against the corresponding hourly-averaged pavement temperature
formance of more advanced analyses. for a typical annual temperature range at the Great Belt Bridge site
(from
12 °C to 43 °C), confirming the temperature dependence of
2.4. Stress range computation stress ranges at welded joints highlighted in the introduction.

Stress ranges are computed from the monitored strains by
applying the rainflow counting algorithm to the monitored strains
duly converted to stresses according to ASTM [19]. Due to compu-
tational constraints, sequences of 500 data points at a time are con-
sidered for the application of the cycle counting algorithm. With a
sampling frequence of 100 Hz, this represents a sampling duration
of 5 s, and leads to stable results given the time required by heavy
vehicles to pass over the monitored welds (e.g. less than 1 s, as can
be seen in Fig. 7).
Fig. 8a shows a typical stress range histogram, with a bin size of
0.5 MPa, obtained using the available monitoring data for year
2012 at SG1. As can be seen, the vast majority of stress cycles cor-
respond to low amplitudes, induced by light vehicles. This has also
been observed in other studies [20–22]. The upper tail of the his-
togram (see Fig. 8b) is associated with the passage of heavily
loaded Class 5 and 6 vehicles.
3. Data-based modelling
3.1. Pavement temperature
Fig. 10 shows the daily-averaged pavement temperatures
obtained from 2008 until 2010 recorded by the sensor T2. This is
the temperature sensor closest to the section instrumented with
strain gauges and its data have therefore been considered in devel-
oping the models for pavement temperature prediction. The max-
imum daily-averaged temperature between 2008 and 2010 is
31.8 °C and the minimum is
8.4 °C. A yearly pattern is clearly
observed.
Fig. 11a presents the monthly-averaged values of the pavement
temperatures (starting from daily-averaged values) whereas
Fig. 11b shows the corresponding standard deviations, which
appear to oscillate around 3 °C.

(a) × 10 -5

10
strain [-]

-5
0 5 10 15 20
Time [h]

(b) × 10 -5

6
4
strain [-]

2
0
-2
-4
0 20 40 60 80 100
Time [min]

Fig. 6. (a) Measured time series of strains during 24 h at SG7 and (b) during 100 min at SG7 (15/05/2011).
 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119 109

10 2
8
10

10 6

Cycles [-]

Cycles [-]
10 1
10 4

10 2

10 0 10 0
0 10 20 30 40 30 35 40 45
∆ σ [MPa] ∆ σ [MPa]

Fig. 8. (a) Stress range histograms at SG1 and (b) associated upper tail, year 2012.

50 Table 2 presents the model parameters for year 2010 estimated

SG 7 via non-linear least squares, whereas the corresponding model fit
is shown in Fig. 12. It is noted that the model parameters presented
∆ σ max in 1h [MPa]

40 SG 8
in Table 2 apply strictly to the Great Belt Bridge, with predicted
30 daily-average temperatures spanning between
3 °C and 23 °C.
Nonetheless, the generic model in Eq. (1) could be used in other
20 bridges with different pavement types and under different climate
conditions, provided monitored pavement temperatures are avail-
10 able to determine the particular model parameters.
The original time series, i.e. daily-averaged pavement tempera-
0 tures recorded during 2010 at sensor T2, is then de-seasonalized by
-10 0 10 20 30 40 50 subtracting the daily mean values estimated via Eq. (1) and divid-
T [°C] ing by the monthly standard deviations of the time series.


Fig. 9. Maximum hourly stress range vs. hourly averaged pavement temperatures. T Dt ðtÞ
T Dt
T t ¼ ð2Þ
rT;t
40 where T t is the deseasonalized daily-averaged pavement tempera-
ture at time t, T Dt ðtÞ the daily-averaged pavement temperature from
30 the original time series and rT;t the monthly standard deviation of
the time series, as shown in Fig. 11b.
[deg C]

20 Fig. 13 presents the resulting de-seasonalized time series. The

analysis of its autocorrelation function (ACF), see Fig. 14a, confirms
10 that the series is stationary, since it decreases quickly to zero.
A joint analysis of the ACF and the partial ACF (PACF) in Fig. 14
0 indicates that T Dt can be adequately modelled by an AR(1), see Eq.
(3), with a positive coefficient, since the ACF decreases exponen-
-10 tially to zero as the lag increases and the PACF cuts off after lag
Jan 2008 Jan 2009 Jan 2010 Jan 2011 1. Thus,

Fig. 10. Daily averaged-pavement temperatures from 2008 to 2010. T t ¼ uT;1  T t
1 þ T;t ð3Þ

where uT;1 is the first AR coefficient and T;t is a random normal

Data from the year 2010 have been used as a training dataset for
error term at time t. Table 3 shows the estimated values of the
temperature model development, since the data availability during
model parameters corresponding to Eq. (3).
that year was almost complete. The approach followed consists of
Analysis of the AR(1) model residuals showed that they are nor-
fitting an autoregressive (AR) model to the de-seasonalized tem-
mally distributed, zero centred and with a standard deviation
perature time series.
equal to 0.764 °C. Furthermore, analysis of the ACF and PACF resid-
In order to model the fluctuations in the mean values of the
uals confirmed their independence, thus verifying the modelling
pavement temperatures, a generic sinusoidal function [15] has
hypothesis.
been considered, namely.


T Dt ¼ a1  sinða2  t þ a3 Þ þ mT ð1Þ 3.2. Vehicle traffic


where T Dt is the expected mean value of the daily-averaged pave- The data utilised in developing the model consist of Eastward
ment temperature at time t (expressed in days, i.e. 0 6 t 6 365), heading traffic, specifically vehicles with dimensions larger than
a1 , a2 and a3 are the parameters of the sinusoidal function and 10 m in length and 2.8 m in width, i.e. vehicle classes 5 and 6 as
mT is the overall annual average of the pavement temperatures. described previously, which are essentially the classes correspond-
110 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119

(a) 25 (b) 4.5

20 4

mean(T) [° C]

std(T) [° C]
15 3.5

10 3

5 2.5

0 2
0 Feb. Apr. Jun. Aug. Oct. Dec. 0 Feb. Apr. Jun. Aug. Oct. Dec.

Fig. 11. Mean value (a) and standard deviations (b) of daily-averaged pavement temperatures from 2008 to 2010.

traffic counts in Fig. 16. The number of heavy vehicles is shown

Table 2 to have increased continuously from bridge inauguration up until
Parameters of the sinusoidal model for daily-averaged pavement temperatures (year
the year 2007 but have decreased thereafter until 2010, this being
2010).
the last year for which traffic data was made available for the work
Parameter Value Units presented herein. These observations underline the non-
a1 13.4 °C stationarity of the traffic time series which is a direct consequence
a2 0.01709 rad/day of the changing economic environment over the eleven year period
a3
1.78 rad considered.
mT 10.1 °C
Apart from the abovementioned seasonal effects and trends due
to economic cycles, both daily and hourly patterns also exist on the
traffic time series. These can be seen in Fig. 17, where the traffic
40 data are shown following hourly and daily aggregation schemes
T (2010)
respectively. They reflect the working patterns of the freight indus-
30 Fit
try, which manifest in weekday vs. weekend effects, as well as peak
hour effects within each day.
20
Bearing all the above in mind, the heavy traffic time series can
°C

10 therefore be described by the superposition of several patterns

with different temporal scales and underlying drivers, as summa-
0 rized in Table 4.
The approach followed to model the traffic counts entails first
-10 the fitting of regression models to the de-seasonalized traffic time
Jan 10 Mar 10 May 10 Jul 10 Sep 10 Nov 10 Jan 11 series. Then, AR models are introduced to account for the regres-
Fig. 12. Original time series and sinusoidal fit.
sion error terms.
The training dataset for developing these models consists of
monitoring data from 2008 to 2009.
The traffic time-series is de-seasonalized as follows:

BDt ðtÞ
lB;t

Fig. 13. Time series of T Dt .
ing to heavy vehicles. Fig. 15 shows the data from 2008 until 2010
aggregated through a daily discretization step. Some seasonal
effects can be observed, namely during the periods corresponding
to summer and winter holidays, when the overall traffic intensities
decrease. In addition, a negative global trend seems to exist, which
can be clearly observed by analysing the annual evolution of the
Bt ¼ ð4Þ
rB;t
where Bt is the deseasonalized daily traffic at time t, BDt ðtÞ is the
daily-aggregated counts of heavy vehicles, lB;t and rB;t are the daily
average and the standard deviation considering a weekly time step,
see Fig. 18.
However, in contrast to the temperature case, the ACF of the de-
seasonalized time series (Fig. 19) alternates cyclically with no evi-
dence of decaying characteristics, indicating the existence of an
autocorrelation structure, which, in essence, is caused by the
day-of-the-week effect (Fig. 17b).
Therefore, regression models are used to capture the day-of-
the-week effect on the de-seasonalized time series as:
Bt ¼ k1 þ k2 X2;t þ    þ kk X k;t þ B;t ð5Þ
where Bt is the value of the de-seasonalized traffic series at time t, ki
is the ith regression coefficient, Xi;t is the ith dummy explanatory
variable and B;t a random error process term at time t.
 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119 111

(a) 1 (b) 1

0.5 0.5

PACF
ACF

0 0

-0.5 -0.5
0 5 10 15 20 0 5 10 15 20
Lag Lag

Fig. 14. ACF (a) and partial ACF (b) of T Dt .

Table 3
Model parameters of the AR(1). × 10 5
6
Parameter Value 95% C.I.
uT;1 0.614 [0.532–0.695] 5.5

[veh./year]
The set of dummy explanatory variables X ;t f and their coeffi- 5
cients estimated via a linear regression least squares regression
approach, are summarized in Table 5.
4.5
Normal weekdays (X1-X5) present, as expected, higher traffic
intensities than weekends (X6 and X7). Moreover, the highest val-
4
ues for the standardized heavy traffic counts occur from Monday
to Wednesday, whereas the lowest value is on Saturdays. The
remaining variables (X8-X15) account for special days within the 3.5
2000 2002 2004 2006 2008 2010
annual calendar and are included in order to minimize errors creep-
ing into the time series model despite their relatively low annual Fig. 16. Annual evolution of the heavy traffic volume.
frequency.
The analysis of the residuals of the regression model in Table 5
normal random error term. Table 6 presents the values of the
reveals significant autocorrelations. To account for this, the residu-
parameters for the AR(7) model which was finally selected between
als of the traffic regression model have been modelled by an AR
models of different order.
model:
Table 7 summarizes the parameters of the normal error term mt
X
p
for the AR(7) model, whose parameters have now been fully spec-
B;t ¼ uB;i  B;t
i þ mt ð6Þ ified and can, thus, be used for the prediction of traffic counts.
i¼1
By way of example, Fig. 20 shows a comparison between
where B;t is the regression error at time t, uB;i are the parameters of model-based predictions and traffic observations corresponding
the autoregressive model, p is the order of the AR model and mt is a to a subset of the validation dataset, i.e. data from the year 2010.

2500
Traffic intensity [veh./day]

2000

1500

1000

500

0
Jan 2008 Jan 2009 Jan 2010 Jan 2011

Fig. 15. Heavy vehicle counts, daily aggregation.

112 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119

200 2500
(a) (b)

2000

Traffic intensity [veh./d]

Traffic intensity [veh./h]
150

1500

100

1000

50
500

0 0
28 Jan 02 Feb 07 Feb 12 Feb 12 Feb 28 Jan 02 Feb 07 Feb 12 Feb 17 Feb

Fig. 17. (a) Hourly- and (b) daily- aggregated traffic counts (Friday Feb 1st to Friday Feb 15th 2008).

Table 4 A very good match between model predictions and observations is

Temporal patterns of the heavy traffic counts. found. It is noted that the model-based predictions of the daily-
Temporal scale of Driver Effect aggregated heavy traffic counts BDt ðtÞ are obtained by combining
the pattern Eqs. (4)–(6).
Bridge lifetime National economy Variations in annual It should be pointed out that the predictive performance of the
traffic counts developed models depends strongly on the values of the daily aver-
Year Annual working patterns Variations during ages ðlB;t Þ and associated standard deviations ðrB;t Þ of the traffic
specific days in year
Week Weekly working patterns Day-to-day variation levels, which were known from actual data in the case of the com-
within a week parison shown in Fig. 20. In the case of simulations of future traffic,
Day Bridge location in relation to Hour-to-hour variation such values will have to be estimated on the basis of scenarios
origin-destination within a day regarding possible economic cycles. This will decrease the predic-
tive performance of the developed models, due to the inherent
uncertainty related to future conditions and associated traffic
levels.

2000 μ B,t
3.3. Fatigue loading
σ B,t
Considering the fatigue process within a S-N framework, the
[veh./day]

1500
fatigue loading, during a time interval Dt, on a welded joint in
terms of applied stresses can be expressed via a performance indi-
1000 cator function DDt ðtÞ:

X
Nc

500 DDt ðtÞ ¼ Dr m

i ð7Þ
i¼1

Jan 2008 Jan 2009 Jan 2010 where Dt is an arbitrary time interval duration ½t; t þ DtÞ and Dri is
the ith stress range out of the total number of stress cycles (N c )
Fig. 18. Daily mean and corresponding standard deviation of traffic considering a
within the time interval Dt and m is the slope of the S-N endurance
weekly time step.
curve.
The indicator DDt ðtÞ provides an estimate of the fatigue loading
for a detail following a single-slope S-N fatigue endurance curve
with slope m. As is well known, for many typical welded joints
m = 3 [18], with several studies adopting this approach in address-
ing the fatigue assessment of metallic structures under variable
loading [21,23]. A degree of conservatism in estimating fatigue
loading, and in turn fatigue damage, may be introduced through
this approach if a particular detail is characterised by a multi-
slope endurance curve (e.g. m1 = 3 and m2 = 5).
Farreras-Alcover [15] proposed a weighted least squares
approach to parameter estimation in polynomial regression mod-
els to characterize the correlation pattern among daily-
aggregated heavy vehicle counts BDt , daily-averaged pavement
temperatures T Dt and daily-aggregated S-N fatigue loading DDt .
This was achieved by introducing the ratio DDt/BDt, as a normalized
fatigue loading indicator per heavy vehicle to represent the depen-
Fig. 19. ACF of the deseasonalized time series of heavy traffic counts. dent variable, and TDt as the independent variable.
 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119 113

Table 5
Parameter estimates for the traffic regression model TRM.

Variable Description Estimate Variable Description Estimate

X1 Monday 0.744 X8 Holiday
1.130
X2 Tuesday 0.758 X9 1st January
0.644
X3 Wednesday 0.756 X10 31st December
1.576
X4 Thursday 0.671 X11 Day previous to extra-long weekenda 1.273
X5 Friday
0.143 X12 Day previous to a long weekendb 0.533
X6 Saturday
1.539 X13 Sunday previous to a holiday 0.023
X7 Sunday
1.076 X14 24–25th December
0.522
X15 26th December 0.835
a
Weekend followed by 2 or more holiday days.
b
Weekend followed by one holiday day.

Table 6
Parameter estimates for the AR(7) model.

Autoregressive model AR(7) Parameters

uB;1 uB;2 uB;3 uB;4 uB;5 uB;6 uB;7
0.307
0.124 – – – 0.119 0.044

Table 7
Parameters for the normal model of the error term m t .

Model residuals mt truncated (ET)

s.t.d. mean s.t.d
AR(7) 0.132 1.24  10
2 0.069

The formulation, which allows the calculation of prediction

bands of the quantity DDt =BDt as a function of T Dt , is described by
Eq. (8). Such prediction bands (see Fig. 21 for an illustration) are
calculated following the methodology described in Farreras-
Alcover [15].
DDt
ðT 0 Þ ¼ T 0 hW  t n
p
1;1
a2  stot ð8Þ
BDT

T
where : T 0 ¼ ½ 1 T 10 . . . T p0  ð9Þ
Fig. 21. 95% confidence and prediction bands for daily DDt =BDt (SG1).
T 0 is a given pavement temperature for which the prediction band
is calculated, n is the number of available datapoints of a training
dataset used to calculate the vector of parameters hW of a polyno- Once the parameters are estimated using a training dataset, Eq.
mial regression model of order p, t is the variate of a t-probability (8) can simulate realizations of the daily-aggregated S-N fatigue
distribution of n
p
1 degrees of freedom, a is a parameter to loading as a function of the operational (i.e. daily-aggregated heavy
traffic counts BDT ) and environmental conditions (i.e. daily-
denote a ð1
a)% confidence interval and s2w is the weighted resid-
averaged pavement temperatures T DT ).
ual variance of the model.

2200
Observation
2000 Model prediction
1800

1600
B ∆t [veh./day]

1400

1200

1000

800

600

400

Jun 2010 Jul 2010 Aug 2010 Sep 2010

Fig. 20. Predictions and observations of daily-aggregated heavy traffic (subset of the validation dataset).
114 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119

Fig. 21 shows the confidence bands (distribution of the mean
value of a dependent variable given a value of the independent
variable) and prediction bands (distribution of the value of a
dependent variable given a value of the independent variable) for
a p ¼ 4 regression model at SG1 obtained following the weighted
least squares approach outlined above and considering as the
training dataset for parameter estimation the monitoring data
recorded during 2012. As might be expected, both prediction and
confidence bands become wider in regions of reduced or space
data, such as the upper part of the temperature range.
4. Probabilistic fatigue life prediction
4.1. Methodology
The time-varying Limit State Function gðx; tÞ for S-N probabilis-
tic fatigue life prediction is formulated as:
gðx; tÞ ¼ Dc
DðtÞBm
where Dc is the critical fatigue damage corresponding to failure, DðtÞ
the S-N fatigue damage up to time t, Bm is a model uncertainty vari-
able associated with the accuracy of stress range estimation and x is
a vector of random variables.
The critical fatigue damage at failure, Dc , is modelled as a log-
normal random variable with a median value of unity and a 30%
coefficient of variation, as suggested in Wirsching [24] and further
recommended by JCSS [25].
The statistical properties of Bm are determined by the type of
global and local stress analysis methods used to estimate the stress
ranges. In general, model uncertainties are involved both in the
method for calculating nominal stresses in the vicinity of a fatigue
sensitive detail and in the estimation of stress concentration fac-
tors due to gross and local discontinuities, or due to deviations
from an ideal shape. Typically, a log-normally distributed random
variable with unit mean and a coefficient of variation in the range
of 0.25–0.30 has been used in offshore related work [26]. Note that
in this study, where the stress ranges are estimated directly from
monitoring data obtained in the vicinity of critical detail locations,
the variability in Bm associated with global stress analysis is cap-
tured by the variance of the regression residuals via the distribu-
tion of the prediction bands, see Eq. (8). Any additional
variability, which may be related to local effects and the position
at which the nominal stress is obtained in relation to the detail,
was, for simplicity, neglected.
For a single-slope fatigue endurance curve and adopting the
Palmgren-Miner rule for damage summation under variable ampli-
tude loading, the fatigue damage DðtÞ can be rewritten as:
1 Xt
1 Xt
DðtÞ ¼  Drm
t ¼  DDt ðtÞ
A t¼0 A t¼0
where A and m are the parameters of the S-N curve for any partic-
ular detail, which depend on material, manufacturing and geomet-
ric factors, and Drt the stress ranges summed up to time t. The
parameters A and m can be considered as jointly distributed ran-
dom variables, though in many practical applications, S-N test
results have been interpreted through a simpler approach in which
m is treated deterministically and the observed variability is
lumped into a log-normally distributed variable for A [26].
Thus, linking Eq. (11) with the data-based regression and time-
series models presented in the preceding sections, the fatigue dam-
age can be expressed as follows:
!
1 X X
t pþ1
DðtÞ ¼  BDt ðtÞ  hW;i
1  T Dt ðtÞi
1 þ t n
p
1  stot ð12Þ
A t¼0 i¼1
with all quantities previously defined.
Finally, Eq. (12) can be inserted into Eq. (10) to obtain a fatigue
limit state function which integrates S-N theory with the data-
based models described in the previous sections:
!
1 X X
t pþ1
gðx; tÞ ¼ Dc
 BDt ðtÞ  hi
1  T Dt ðtÞi
1 þ tn
p
1  stot ð13Þ
A t¼0 i¼1
In order to perform a fatigue reliability analysis, the material
parameter A needs to be specified in probabilistic terms for the
ð10Þ particular detail considered. Using the statistical method described
in EN 1990. Eurocode 0 - Basis of structural design [27], the mean and
standard deviation of A are estimated for detail 71 (trough-splice
welds, welded against a backing strip) and detail 50 (trough-to-
deck weld, partial penetration). It is assumed that a lognormal dis-
tribution is appropriate with a CoV equal to 0.58 [25]. Moreover,
the characteristic values given in EN 1993-1-9:2005. Eurocode 3:
Design of steel structures - Part 1–9: Fatigue [28] which correspond
to a 95% of survival for log(N) at N C ¼ 2 million at a 75% confidence
level are introduced and, thus, the estimated statistics are summa-
rized in Table 8.
4.2. Monte Carlo Simulation scheme
A MCS scheme is implemented in order to determine the time-
varying fatigue reliability profile for the different instrumented
welds following Eq. (13).
For each simulation cycle, values of A and D are sampled from
their respective probability distributions. Pavement temperatures
T Dt ðtÞ and heavy vehicle counts BDt ðtÞ are then generated for a time
increment Dt with the regression model (Eq. (8)), used to deter-
mine the corresponding S-N fatigue loading DDt ðtÞ. This time-
stepping process is repeated until the limit state function is vio-
lated (i.e. until gðx; t f Þ < 0) and a realization of the time to failure
ðt f Þ is determined. The next simulation cycle then begins until a
total of NMC time-to-failure realizations are obtained, N MC being
the sample size of the MCS.
Finally, a lognormal distribution is fitted to the frequency dia-
gram of tf comprising a total of NMC realizations. This process is
repeated for all details under consideration. The reliability index
at time t, bðtÞ, is finally calculated as:
ð11Þ Z t 
bðtÞ ¼
U
1 f tf ðtÞdt ð14Þ
0
where U
1 is the inverse of the cumulative distribution function of
the standard normal distribution and f tf ðtÞ the lognormal probabil-
ity density function (PDF) of tf . Fig. 22 presents the proposed
methodology leading to the time-varying fatigue reliability profiles
in the form of a flow-chart.

Table 8
Estimated statistics for the S-N parameter A.

Assumed Estimated
Detail category CoV 5% characteristic value of A [MPa3] lA [MPa3] rA [MPa3]
50 (trough-to-deck) 0.58 2:50  10 11
7:30  10 11
4:23  1011
71 (trough-splice) 0.58 7:16  1011 2:09  1012 1:21  1012
 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119 115

Define temperature and traffic scenarios

Define sample size of the MCS, NMC

Set Sensor Index: k=1

Selection of PDF for the fatigue detail coefficient (A) and

Miner's critical damage accumulation index (Δ)

Set MCS index, n=1

Generate realizations A(n) and Δ(n)

Set temporal index t=1 day

Simulate and

Simulate

1 No
Δ ( )− ( ) ( ) <0 ? t=t+1
( )

Yes
( )

No
n=NMC? n=n+1

Fit lognormal PDF to (1) ( )} and compute the

fatigue reliability profile as ( ) = −Φ (

No
k=kmax? k=k+1

Yes
End

Fig. 22. Methodology for the time-varying fatigue reliability calculation.

5. Application
5.1. Dataset selection
The proposed methodology for data-based probabilistic fatigue
life prediction utilises a training dataset to specify the parameters
for the temperature, traffic and fatigue loading models. In the
present study, the training datasets comprise data from 2010 for
temperature modelling, data from 2008 to 2009 for traffic mod-
elling and data from 2012 for modelling of fatigue loading. These
time periods were chosen taking into account factors such as com-
pleteness and data reliability.
The time interval Dt used to characterize DDt has been chosen to
be 1 day (i.e. 24 h). In consequence, pavement temperatures have
been averaged over 24 h to determine the daily-averaged pave-
ment temperature T Dt . Similarly, heavy vehicle counts for classes
116
(a)
3 T0
Temperature increase [ °C]
2.5
2 B1
1.5
0.5
0
2020 2040 2060 2080 2100 2120 2140
Fig. 23. Scenarios for (a) pavement temperature and (b) heavy traffic counts.
Table 9
Definition of scenarios.
Identifier T0
B0 Scenario S0
B1 Scenario S2
5 and 6 have been aggregated over the same period to determine
the daily vehicle counts BDt . Further considerations regarding the
implications of the selection of the size of Dt are discussed in
Farreras-Alcover [29].
5.2. Traffic and temperature scenarios
Future traffic levels are subjected to non-stationary effects due
to, for example macro-economic conditions. Moreover, climate
change is predicted to result in an increase of future average tem-
peratures. In the context of long-term asset management, it is
therefore desirable to consider different traffic and temperature
scenarios and quantify their effect on the fatigue life of the welds
under consideration.
Two temperature trends (T0 and T1) are examined. Scenario T0
does not consider any systematic annual temperature increase,
whereas scenario T1 considers a mean air temperature increase
of 2.9 °C by the year 2100 [30]. This trend has been generated by
the Danish Meteorological Institute considering the scenario A1B
of future emissions of greenhouse gases [31]. It has further been
assumed herein that an increase in air temperatures leads to an
equal increase in pavement temperatures [10].
Two different traffic trends have also been defined. Traffic trend
B0 considers the future heavy traffic levels to be equal to those
obtained from the training dataset, while traffic trend B1 considers
an annual average increase of 1% until the year 2030. The heavy
traffic levels from 2030 and onwards are assumed to remain con-
stant. Fig. 23 depicts the scenarios considered by combining the
different temperature and traffic trends with year 2012 considered
as reference. The proposed methodology makes it possible to quan-
tify the influence of adverse future changes in temperature and
traffic trends both individually and in combination.
The individual traffic and temperature trends are paired accord-
ingly resulting in four combinations (scenarios), as summarized in
Table 9.
I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119
(b)
1.1
T1
Annual traffic increase [-]
B0
1.05
1
0.95
0.9
2020 2040 2060 2080 2100 2120 2140
Year Year
5.3. Definition of target reliability
T1 EN 1990. Eurocode 0 - Basis of structural design [27] states that
Scenario S1 the target reliability index for a fatigue limit state corresponding
Scenario S3 to a standard reference period of 50 years for class RC2 (medium
consequences) structural members can vary from 1.5 to 3.8,
depending on the degree of inspectability, reparability and damage
tolerance. Long-span bridges are typically designed with service
lives well above 50 years and, consequently, a lower target
reliability range might be appropriate. However, fatigue failures
in long-span bridges are normally associated with high failure con-
sequences and, for this reason, it was decided to illustrate the
methodology developed using the upper bound of 3.8 as a target
value.
5.4. Results
Fig. 24 shows the simulated profiles of T Dt and BDt during
730 days as well as the corresponding simulated profile of DDt ðtÞ
using the regression model for SG2 (trough splice weld).
Fig. 25 displays the lognormal probability plot for 1000 simu-
lated realizations of t f corresponding to the scenario S0 at SG1.
As can be seen, the lognormal distribution fits very well the simu-
lated fatigue lifetimes. The sample size of the different MCS has
been equal to 500, which has been shown to be sufficient for the
MCS to converge with regard to first and second order statistics,
as seen in Fig. 26.
Fig. 27 presents the time-varying fatigue reliability profiles
obtained at trough splice welds respectively for the scenario S0.
As expected, the reliability indices decrease non-linearly with time.
The estimated time to failure for a target reliability b ¼ 3:8 corre-
sponding to the scenario S0 for the critical welded joints consid-
ered in this study are summarized in Table 10.
SG8 (66 years) and SG1 (106 years) are respectively the trough
splice and trough-to-deck welds presenting the lowest time to
reach the target reliability.
Fig. 28 shows respectively the reliability profiles at SG1 corre-
sponding to the different scenarios considered. Table 11 summa-
rizes the corresponding times for reaching b ¼ 3:8 for both SG1
and SG8.
As expected, the increase of temperature levels (scenario S1),
heavy traffic levels (scenario S2) and temperature and traffic levels
(scenario S3) result all in lower reliability profiles. In both welds,
 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119 117

40

20

T∆ t [ °C]
0

-20
0 100 200 300 400 500 600 700 800
4000
B ∆ t [veh./h]

3000

2000

1000

0
0 100 200 300 400 500 600 700 800
× 10 6
15

10
D∆ t [MPa3 ]

0
0 100 200 300 400 500 600 700 800
Time [days]

Fig. 24. Simulated profiles T Dt , BDt and DDt at SG2 during 730 days, scenario S0.

1
0.9995
0.995 0.8

0.95
0.6
COV [-]
Probability

0.75
0.5 0.4

0.25
0.2
0.05
0
0.005 0 200 400 600 800 1000
0.0005 Sample size of MCS, N

10 2 10 3 10 4 10 5 Fig. 26. COVðt f Þ of the MCS as a function of NMC.

Fig. 25. Lognormal fit to simulated tf at SG1, scenario S0.

The obtained reliability profiles are conservative due to the

the most unfavourable scenario is S3, followed by S2, S1 and S0.
The decrease in the time to failure when b ¼ 3:8 due to increases
in the temperature (scenario S1) is more noticeable at SG1
(22.6% reduction) than at SG8 (3.0% reduction). This is explained
by the major sensitivity to pavement temperature of the regression
models corresponding to trough-to-deck welds, which could be
attributed to the local structural behaviour and the longitudinal
position of the monitored points within a span.
The joint increase of future traffic levels and temperatures cap-
tured by the scenario S3 decreases significantly the times to reach
failure for b ¼ 3:8, with corresponding reductions of 27.2% (SG8)
and 37.7% (SG1).
adoption of a single-slope S-N fatigue curve where all stress cycles
are assumed to be damaging. This represents a more severe model
than the standard Eurocode design model EN 1993-1-9:2005. Euro-
code 3: Design of steel structures - Part 1–9: Fatigue [28]. Moreover,
system effects (i.e. stress redistribution), which are generally
expected to be favourable in the case of an orthotropic welded
deck, are not accounted for. As a result, the calculated reliability
profiles can be regarded as a conservative lower bound of the
deck’s actual reliability level. In consequence, the estimated times
to failure may be regarded as appropriate warning points at which
to schedule detailed inspections [32,33], thus linking monitoring
outcomes with management strategies. Similarly, crossing these
118 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119

Fig. 27. Time-varying fatigue reliability profiles at trough splice welds, scenario S0.

Table 10
Times to reach b = 3.8 at different SGs, scenario S0.

SG1 SG2 SG3 SG4 SG5 SG6 SG7 SG8 SG9

t. such that bðtÞ < 3:8 [years] 106 87 157 343 205 137 128 66 142

12
Scenario S0 6. Conclusions
11 Scenario S1
Scenario S2
10 A novel methodology for probabilistic fatigue life prediction of
Scenario S3
9 monitored welded joints of orthotropic bridge steel decks has been
presented. The methodology consists of using monitoring data to
8
characterize the random processes of pavement temperatures
7
β

and heavy traffic counts by means, respectively, of autoregressive

6 models for the case of pavement temperatures and autoregressive
models combined with regression models for the case of heavy
5
traffic intensities. Those processes can be regarded as the main dri-
4
β * =3.8
vers leading to stress ranges variations, i.e. S-N fatigue loading, at
3 welded joints of orthotropic steel decks, due to the temperature-
2 dependent composite action between the pavement and the steel
0 50 100 150 deck of orthotropic decks. Polynomial regression models are used
Time [years] to characterize the normal correlation pattern between the S-N
fatigue loading derived from strain measurements, taken as the
Fig. 28. Time-varying fatigue reliability profiles at SG1, scenario S0 to S3.
dependent variable, and pavement temperatures and heavy traffic
counts. The different models are integrated within a fatigue relia-
bility framework, in which the uncertainties arising from material
Table 11 properties and fatigue damage at failure are modelled via random
Times to reach b ¼ 3:8, SG1 and SG8, different scenarios. variables. A Monte Carlo Simulation scheme is used to simulate the
t. s.t. bðtÞ < 3:8 [years] S0 S1 S2 S3
time-varying fatigue damage in an S-N context by inputting simu-
lations of time-series of heavy traffic counts and pavement tem-
SG1 106 82 76 66
SG8 66 64 56 48
perature into the developed regression models. The novelty in
this work consists of modelling independently the random pro-
cesses of pavement temperatures and heavy traffic counts (i.e.
the independent drivers of strain variations and fatigue damage)
thresholds could also be regarded as points at which less conserva- to simulate the fatigue process and to determine time-varying reli-
tive fatigue damage models should be developed that could also ability profiles. This makes it possible to quantify the effect of dif-
benefit from updating based on inspection outcomes. ferent scenarios in terms of changes in pavement temperatures
It should be noted that no fatigue cracks have so far been and heavy traffic counts in fatigue life prediction. From a Structural
detected at the orthotropic steel deck of the Great Belt Bridge, this Health Monitoring perspective, the method developed can be
being consistent with the predicted fatigue lives calculated follow- regarded as a novel local, data-based approach to structural perfor-
ing the proposed methodology. The methodology can be applied to mance prognosis, based on environmental and operational moni-
other locations to be instrumented with strain gauges in the future. toring data streams. It can be implemented to other long-span
 I. Farreras-Alcover et al. / International Journal of Fatigue 95 (2017) 104–119
bridges with orthotropic decks, several of which are known to be
sensitive to fatigue damage, with the model parameters estimated
from in situ SHM data.
The proposed methodology has been illustrated considering
actual monitoring outcomes from the Great Belt Bridge (Denmark).
The results have quantified the adverse effect on fatigue life from
higher future pavement temperatures, or from increased heavy
traffic levels, or both. For the case of the welded joint associated
with the lowest predicted fatigue life (i.e. SG8), the joint conse-
quence of increasing traffic and temperature levels via the consid-
ered scenarios results in a reduction of 27% in the time needed to
reach a nominal target reliability. However, this reduction is not
uniform across the range of fatigue details considered, emphasiz-
ing the importance of monitoring different locations, based on a
thorough understanding of the fatigue behaviour of the orthotropic
steel deck. This could be accomplished through a combination of
modelling and prior experience, building on which the specifica-
tion of an appropriate SHM strategy can be procured to aid asset
integrity management, as demonstrated by the work presented
in this paper.
Acknowledgements
The authors would like to acknowledge the European Commis-
sion for funding the SmartEN training network, under the Marie
Curie ITN FP7 program (grant agreement number 238726), within
which this research was carried out. The authors would also like to
thank Storebælt A/S for allowing access to the monitored data
from the Great Belt Bridge (Denmark) in the context of the pre-
sented research.
The research findings, views and conclusions contained herein
are those of the authors and should not be interpreted as repre-
senting the official policies, either expressed or implied, of the
authors’ employers or the sponsoring organisations.
References
[1] ASCE. Report Card for America’s Infrastructure; 2013 Mar.
[2] Farreras-Alcover I, Andersen JE, Chryssanthopoulos MK. Performance
assessment and prediction of welded joints in orthotropic decks considering
hourly monitoring data. Struct Eng Int 2013;23(4):436–42 (7) [November 1].
[3] Frangopol DM, Strauss A, Kim S. Bridge reliability assessment based on
monitoring. J. Bridge Eng. 2008;13(3):258.
[4] Ni YQ, Ye XW, Ko JM. Monitoring-based fatigue reliability assessment of steel
bridges: analytical model and application. J Struct Eng 2010;136(12):1563–73.
[5] Guo T, Frangopol DM, Chen Y. Fatigue reliability assessment of steel bridge
details integrating weigh-in-motion data and probabilistic finite element
analysis. Comput Struct 2012;112–113:245–57 [Elsevier Ltd].
[6] DeJong FBP. Overview fatigue phenomenon in orthotropic bridge decks in the
Netherlands. In: Orthotropic Bridg. Conf Calif Sacramento, CA, USA, 2004. p.
489–512.
[7] Kolstein MH. Fatigue classification of welded joints in orthotropic steel bridge
decks. Delft University of Technology; 2007.
119
[8] Xiao Z, Yamada K, Inoue J, Yamaguchi K. Fatigue cracks in longitudinal ribs of
steel orthotropic deck. Int J Fatigue 2006;28:409–16.
[9] Ya S, Yamada K, Ishikawa T. Fatigue evaluation of rib-to-deck welded joints of
orthotropic steel bridge deck. J Bridge Eng 2011;16(4):492–9.
[10] De Jong FBP. Renovation techniques for fatigue cracked orthotropic steel
bridge decks PhD Thesis. Technische Universiteit Delft; 2006.
[11] Cullimore MSG, Flett ID, Smith JW. Flexure of steel bridge deck with asphalt
surfacing. IABSE period. University of Bristol; 1983.
[12] Kolstein M. Stress reduction due to surfacing on orthotropic steel decks. In:
10th Nord Steel Constr Conf Copenhagen, 2004. p. 425–37.
[13] Song Y, Ding Y. Fatigue monitoring and analysis of orthotropic steel deck
considering traffic volume and ambient temperature. Sci China Technol Sci
2013;56(7):1758–66 [May 4].
[14] Laigaard J, Bitsch N, Gjelstrup H. Control of traffic loads on Great Belt Bridge.
Bridge Maintenance, Safety, Manage. Stresa, Italy: Taylor & Francis Group;
2012. p. 3360–7.
[15] Farreras-Alcover I, Chryssanthopoulos MK, Andersen JE. Regression models for
structural health monitoring of welded bridge joints based on temperature,
traffic and strain measurements. Struct Health Monit 2015;14(6):648–62.
[16] Guo T, Li A, Li J. Fatigue life prediction of welded joints in orthotropic steel
decks considering temperature effect and increasing traffic flow. Struct Health
Monit 2008;7(3):189–202 [July 21].
[17] Sugioka K, MacRae GA, Saleh MF, Beamish M. Lifetime evaluation of
orthotropic steel bridge decks. In: Int Orthotropic Bridge Conf Sacramento,
USA, 2008.
[18] Maddox SJ, Richards KG. Fatigue strength of welded structures. Cambridge,
UK: Abington Publishing; 1991.
[19] Standard Practices for Cycle Counting in Fatigue Analysis. Designation: E 1049-
85. West Conshohocken, PA; 1997.
[20] Fasl JD, Helwig TA, Samaras SLWVA, Yousef AA, Frank KH, Potter DL, et al.
Development of rapid, reliable, and economical methods for inspection and
monitoring of highway bridges. Bridge Maintenance, Safety, Manag. Life-Cycle
Optim. Philadelphia, Pennsylvania, USA: Taylor & Francis Group; 2010. p.
413–9.
[21] Guo T, Chen Y-W. Field stress/displacement monitoring and fatigue reliability
assessment of retrofitted steel bridge details. Eng Fail Anal 2011;18(1):354–63
[January Elsevier Ltd].
[22] Liu M, Frangopol DM, Kwon K. Fatigue reliability assessment of retrofitted
steel bridges integrating monitored data. Struct Saf 2010;32(1):77–89
[January Elsevier Ltd].
[23] Wang Y, Li ZX, Li AQ. Combined use of SHMS and finite element strain data for
assessing the fatigue reliability index of girder components in long-span cable-
stayed bridge. Theor Appl Fract Mech 2010;54(2):127–36 [October].
[24] Wirsching P. Fatigue reliability for offshore structures. J Struct Eng 1984;110
(10):2340–56.
[25] JCSS. JCSS Probabilistic Model Code; 2007.
[26] Chryssanthopoulos MK, Righiniotis T. Fatigue reliability of welded steel
structures. J Constr Steel Res 2006;62(11):1199–209 [November].
[27] EN 1990. Eurocode 0 - Basis of structural design. Brussels (Belgium); 2007.
[28] EN 1993-1-9:2005. Eurocode 3: Design of steel structures - Part 1–9: Fatigue.
Brussels (Belgium); 2005.
[29] Farreras-Alcover I. Data-based models for assessment and life prediction of
monitored civil infrastructure assets PhD Thesis. UK: University of Surrey;
2014.
[30] Danmarks Meteorologiske Institut. Fremtidige klimaforandringer i Danmark.
Copenhagen; 2012.
[31] Bernstein L, Bosch P, Canziani O, Chen Z, Christ R, Davidson O, et al. Climate
Change 2007: Summary for Policymakers; 2007.
[32] Chung H-Y. Fatigue reliability and optimal inspection strategies for steel
bridges. Philosophy: The University of Texas; 2004.
[33] Chen ZW, Xu YL, Wang XM. Fatigue reliability analysis of long span bridges
with structural health monitoring systems. In: 4th Int Conf Struct Health
Monit Intell Infrastruct Zurich, Switzerland, 2009.