You are on page 1of 8

Probabilistic Fatigue Evaluation of Riveted Railway Bridges

Boulent M. Imam1; Timothy D. Righiniotis2; and Marios K. Chryssanthopoulos3

Abstract: A probabilistic fatigue assessment methodology for riveted railway bridges is presented. The methodology is applied to a
typical, short-span, riveted U.K. railway bridge under historical and present day train loading. On the loading side, the problem is
randomized through dynamic amplification and traffic volume; on the resistance side, the S-N curves and the cumulative damage model
Downloaded from ascelibrary.org by ARIZONA,UNIVERSITY OF on 05/29/14. Copyright ASCE. For personal use only; all rights reserved.

are treated probabilistically. Model uncertainty is represented by the ratio between actual and calculated stresses, the latter obtained
through finite element analysis. Annual response spectra for a fatigue-critical connection are developed through Monte Carlo simulation,
which show that there is a continual and accelerating increase in the mean stress range experienced by the connection with time. S-N
curves proposed in United States and United Kingdom codes are used in combination with Miner’s rule, to estimate the remaining fatigue
life of the connection for different target failure probabilities. Parametric studies revealed that fatigue life estimates exhibit the highest
sensitivity to detail classification, to S-N predictions in the region of high endurances, and to model uncertainty. This highlights the
importance of field monitoring for old bridges approaching the end of their useful life.
DOI: 10.1061/共ASCE兲1084-0702共2008兲13:3共237兲
CE Database subject headings: Bridges, railroad; Steel; Probability; Fatigue life; United Kingdom.

Introduction probabilistic treatments have presented their results in terms of


either a reliability index ␤ 共Brühwiler and Kunz 1993; Kunz and
A significant part of the railway bridge stock around Europe and Hirt 1993兲 or a failure probability P f 共Tobias and Foutch 1997兲.
North America consists of metallic railway bridges. Riveted con- Both ␤ and P f are evolving with time, capturing the reduction in
struction was employed for the majority of these bridges, which the available safety margin against fatigue as the number of load
were built during the second half of the 19th and the beginning of cycles experienced by the structure increases. These probabilistic
the 20th centuries. This implies that many of these bridges have performance measures can be associated with the widely accepted
exceeded 100 years of age. Train loads and speeds, as well as safety factor format 共AASHTO 1998兲.
traffic volumes, have been increasing continually over the years Following previous work 共Imam et al. 2005, 2006兲, the authors
and, today, these bridges sustain loads that are substantially dif- have extended the deterministic fatigue assessment of a United
ferent to those experienced 100 or even 50 years ago. This, Kingdom typical, simply supported, short-span, riveted railway
bridge 共Fig. 1兲, in order to account for both material and loading
coupled with the fact that old riveted bridges were never designed
uncertainties. By focusing on the most critical connection 共S7,
against fatigue, renders their fatigue assessment a task of consid-
Fig. 1兲, as identified in a deterministic manner by Imam et al.
erable importance in railway infrastructure management.
共2005, 2006兲, its associated loading is randomized by introducing
So far, the majority of the fatigue assessment procedures
uncertainties in the train frequencies as well as its dynamic am-
employed for old riveted bridges have been based on determinis-
plification, the latter captured via the dynamic amplification factor
tic methods. However, in the past decade or so, probabilistic
共DAF兲. Further uncertainty in loading is incorporated by consid-
methodologies have emerged as a viable alternative incorporating ering the epistemic or modeling uncertainty, represented by the
uncertainty in both resistance and loading 共Brühwiler and Kunz ratio of actual-to-calculated stresses. Thus, through the use of
1993; Kunz and Hirt 1993; Tobias and Foutch 1997兲. As is ap- Monte Carlo simulation, annual response spectra, in the form of
propriate for time-dependent reliability problems, fatigue-related stress ranges, are generated for four distinct time periods span-
ning the bridge’s life since 1900.
1
Postdoctoral Researcher, School of Engineering 共C5兲, Univ. of On the resistance side, in addition to the Miner sum, S-N
Surrey, Guildford, Surrey GU2 7XH, U.K. 共corresponding author兲. curves used in the United Kingdom 共BS 1980; Railtrack 2001兲
E-mail: b.imam@surrey.ac.uk and the United States 共AREA 1996兲, for the fatigue detail under
2
Lecturer in Structural Engineering, School of Engineering 共C5兲,
consideration, are modeled probabilistically. By combining load-
Univ. of Surrey, Guildford, Surrey GU2 7XH, U.K. E-mail:
t.righiniotis@surrey.ac.uk
ing and resistance uncertainties in a reliability-based formulation,
3
Professor of Structural Systems, School of Engineering 共C5兲, Univ. failure probability time profiles and remaining fatigue life esti-
of Surrey, Guildford, Surrey GU2 7XH, U.K. E-mail: mkchry@ mates are generated for the critical detail of the particular bridge
surrey.ac.uk analyzed in this paper.
Note. Discussion open until October 1, 2008. Separate discussions
must be submitted for individual papers. To extend the closing date by
one month, a written request must be filed with the ASCE Managing
Editor. The manuscript for this paper was submitted for review and pos-
Fatigue Detail Classification
sible publication on December 18, 2006; approved on September 5, 2007.
This paper is part of the Journal of Bridge Engineering, Vol. 13, No. 3, Several fatigue detail classifications are suggested in United
May 1, 2008. ©ASCE, ISSN 1084-0702/2008/3-237–244/$25.00. Kingdom 共BS 1980; Railtrack 2001兲 and American 共AREA 1996兲

JOURNAL OF BRIDGE ENGINEERING © ASCE / MAY/JUNE 2008 / 237

J. Bridge Eng. 2008.13:237-244.


Downloaded from ascelibrary.org by ARIZONA,UNIVERSITY OF on 05/29/14. Copyright ASCE. For personal use only; all rights reserved.

Fig. 1. Finite element model of a typical riveted railway bridge Fig. 2. Results of full-scale fatigue tests on old steel and
wrought-iron riveted members and connections compared with S-N
curves suggested by U.K. codes 关adopted from BS 共1980兲; Railtrack
codes to represent the fatigue behavior of riveted connections. 共2001兲兴
BS5400 共BS 1980兲 suggests the use of its Class D for riveted
details. Class B can also be used, following some tests on riveted
girders 共Xie et al. 2001兲. Since the latter class relates to the stress versus drilled兲, the possible presence of corrosion on test speci-
at the edge of a hole, its S-N curve has to be divided by a stress mens from actual girders, the different termination criteria used in
concentration factor of 2.4 共suggested in BS5400兲 in order to the tests 共first observed cracking, fracture or excessive deforma-
express it in terms of commonly used nominal stress. This class tion兲 and, finally, the variation in damage accumulated before
will be termed as “modified Class B” hereafter. The U.K. railway removing the specimens for testing 共Imam 2006兲. The data pre-
assessment code 共Railtrack 2001兲 proposes a specific class for sented in Figs. 2 and 3 imply a higher standard deviation than that
wrought-iron riveted details 共Class WI兲. assumed in either U.K. or U.S. classifications.
On the other hand, in the American Railway Engineering As- By studying Figs. 2 and 3, it can be seen that the modified
sociation 共AREA 1996兲 code, a distinction is made between steel Class B S-N curves 共mean and design兲 and the AREA curves
riveted connections with low or normal clamping force, the effect 共mean and design兲 for punched holes seem to provide a reason-
of which was investigated extensively by Zhou et al. 共1995兲. For able fit to the experimental results. The Class D mean S-N curve
the former, Category D is suggested, whereas for the latter, a 共Fig. 2兲 lies much too close to the upper bound of the experi-
combination of Category D and C curves is proposed. For normal mental data, while the Category D curve 共Fig. 3兲 becomes too
clamping, a further distinction is made between punched and conservative for higher endurances. The AREA S-N curve for
drilled holes. In the same code, a classification for wrought iron drilled holes has not been considered further, since the method of
riveted connections is also given 共Category E兲. hole preparation used in the construction of riveted bridges at the
When all stress ranges in a stress history are below the con- beginning of the previous century was likely to be by punching.
stant amplitude fatigue limit 共CAFL兲, ⌬␴0, the assumption of an The fatigue data on wrought-iron riveted members and con-
infinite life may be adopted. However, long life 共number of cycles nections presented in Figs. 2 and 3 are replotted in Fig. 4 in order
⬎108兲 variable amplitude fatigue tests carried out on full-scale to make direct comparisons with the mean and design S-N curves
welded details 共Fisher et al. 1983, 1993兲 have resulted in fatigue for wrought-iron riveted details 共Railtrack Class WI and AREA
cracking even when only 0.1% of the applied stress ranges was Category E兲. Due to the limited number of full-scale test results
above the CAFL. For this reason, U.S. codes 共AREA 1996; on wrought-iron elements, small-scale test results on wrought-
AASHTO 1998兲 treat variable amplitude loading by extending the
S-N curves downwards without a change of slope at the CAFL. In
the case of U.K. codes 共BS 1980; Railtrack 2001兲, variable am-
plitude loading is treated by changing the slope of the S-N curve
from m to m + 2 for stress ranges below the CAFL, which is
defined at N = 107 cycles.
The mean and design 共defined as “mean minus two standard
deviations”兲 S-N curves of the above-mentioned U.K. detail clas-
sifications for steel riveted details are shown in Fig. 2, whereas
Fig. 3 presents their American counterparts. In addition to the S-N
curves, Figs. 2 and 3 also present full-scale experimental results
on old steel and wrought-iron riveted members and connections
retrieved from the literature. The majority of these were retrieved
from Kulak 共2000兲. A considerable degree of scatter can be ob-
served in these test results, which is greater than that evidenced
with welded details. Several reasons contributing to this scatter
may be identified: The wide range of rivet clamping force values
in the specimens, the different stress ratios that were applied Fig. 3. Results of full-scale fatigue tests on old steel and
during different tests, the different material properties 共old steel wrought-iron riveted members and connections compared with S-N
versus wrought iron兲, the method of hole preparation 共punched curves suggested by the AREA 共1996兲 code

238 / JOURNAL OF BRIDGE ENGINEERING © ASCE / MAY/JUNE 2008

J. Bridge Eng. 2008.13:237-244.


Probabilistic Framework

Resistance Random Variables


The fatigue life of a particular structural detail under constant
amplitude loading may be estimated from an S-N curve, which is
described by

N共⌬␴兲m = C 共1兲
where N = number of cycles to failure; ⌬␴ = applied constant
amplitude stress range; and C and m = constants relevant to the
fatigue detail in question. In a probabilistic formulation, N is
Downloaded from ascelibrary.org by ARIZONA,UNIVERSITY OF on 05/29/14. Copyright ASCE. For personal use only; all rights reserved.

generally considered a lognormal random variable whose distri-


bution has a constant standard deviation and varying mean
Fig. 4. Results of full-scale and small-scale fatigue tests on
共Chryssanthopoulos and Righiniotis 2006兲. If it is further assumed
wrought-iron riveted members and connections compared with U.K.
that m and ⌬␴ are deterministic, it follows from Eq. 共1兲 that C
and American S-N curves
will also be lognormal with the same coefficient of variation
共CoV兲 as N.
In this work, mean values for C are obtained from BS 共1980兲
iron members and connections are also presented in Fig. 4. It can for modified Class B and the U.K. railway assessment code 共Rail-
be seen that the Class WI design curve captures the scatter for track 2001兲 for Class WI. In both cases, the CoV of C is assumed
wrought-iron riveted connections reasonably well. In contrast, the equal to 0.30 as proposed by BS 共1980兲. For the AREA S-N
AREA Category E mean and design curves seem to be conserva- curves, the mean values and CoV for C proposed by Moses et al.
tive with respect to the experimental results. The reason for the 共1987兲 are used. Accordingly, a CoV of 0.14 is assumed for the
observed short endurances for the two data points lying below the AREA Categories C and D, and a CoV of 0.10 is assumed for
S-N curves 共indicated by circles in Fig. 4兲, which were tested at a Category E.
stress range of 90 MPa, may be due to the fact that these speci- The probabilistic description of the S-N curves also requires
mens had been pretested at lower stress ranges and may have the calculation of a reference stress range associated with C. In
contained nonidentified cracks hidden by the rivet head 共Fisher the case of the British S-N curves, this reference stress range is
et al. 1987兲. However, given the limited number of tests, these the CAFL, ⌬␴0, defined at a life N = 107. Therefore, for each
two outliers should be treated with considerable caution. sample value of C, the corresponding CAFL is calculated through
Figs. 2–4 suggest that the BS5400 modified Class B, the Eq. 共1兲, by taking logarithms on both sides 共for N = 107兲 and re-
Railtrack Class WI S-N curve, and the AREA curve for punched arranging as
holes present the most reasonable options with respect to the
experimental results. This suggestion is largely based on com- log C − 7
log共⌬␴0兲 = 共2兲
parison of the S-N curves with the test results lying in the region m
N ⬍ 107 cycles, which comprise the majority of the experimental Eq. 共2兲 implies that ⌬␴0 is a derived random variable, fully cor-
data. There are very few test data in the region beyond N related with C. The same approach is adopted for the AREA S-N
⬎ 107 cycles, which corresponds to stress ranges below curves, the only difference being that the reference stress range is
50– 60 MPa. As will be seen later in the paper, this is precisely defined at N = 2 ⫻ 106. In graphical terms, the effect of random
the region of interest for riveted railway bridges, since the major- variability is to produce a parallel shifting of the S-N curves.
ity of the stress cycles are close to or below the CAFL of the In the case of variable amplitude loading, fatigue damage is
detail classifications considered. Small changes in the S-N curves calculated using the Miner sum which is given as 共Miner 1945兲
in this high endurance region 共N ⬎ 107兲 can have a significant
effect on fatigue life predictions. More experimental results under n
low stress ranges 共less than 50– 60 MPa兲 are obviously required D= 兺i Nii 共3兲
in order to ascertain the validity of the proposed S-N curves and,
thus, increase confidence in estimating remaining fatigue lives. where D = total damage; ni = applied number of cycles at a stress
It has to be emphasized that there is very limited published range ⌬␴i; and Ni = fatigue life corresponding to the same stress
work on full-scale tests of riveted bridge connections such as range. Fatigue failure is assumed to occur when D is greater than
stringer-to-cross-girder connections 共Imam 2006兲. Due to the or equal to a damage limit ⌬, which is typically taken equal to
complex behavior of such connections, which involves out-of- unity in deterministic analysis. However, ⌬ is susceptible to con-
plane deformations and the development of secondary stresses siderable uncertainty, especially in the case of variable amplitude
共Al-Emrani 2005兲, it is often difficult to interpret test results on an loading. A lognormal distribution with a mean value of 0.90 and a
S-N basis using nominal stresses. As is well known, the S-N CoV of 0.30 is considered here in accordance with earlier studies
curves found in fatigue codes are generally developed for uniaxial 共Wirsching 1995兲.
stress conditions and it has yet to be verified whether they can be
used confidently for estimating the fatigue life of such riveted
Loading and Modeling Random Variables
connections. Notwithstanding these comments, and given the ab-
sence of specific test results on stringer-to-cross-girder connec- Live loading on the bridge is represented by a rail traffic model,
tions, the BS5400 modified Class B, the Railtrack Class WI, and which was developed by the authors in previous studies 共Imam
the AREA punched hole curves are selected in this paper for et al. 2006兲. It is divided into four distinct periods 共1900–1920,
fatigue damage calculation. 1920–1940, 1940–1970, and 1970–2005兲, each associated with

JOURNAL OF BRIDGE ENGINEERING © ASCE / MAY/JUNE 2008 / 239

J. Bridge Eng. 2008.13:237-244.


Table 1. Random Variables for Probabilistic Analysis
Random
variable Mean CoV Distribution
a
C 共BS5400 modified Class B兲 2.34⫻ 10 15
0.30 Lognormal
共Railtrack Class WI兲 9.33⫻ 1013 0.30 Lognormal
共AREA Category C兲 3.04⫻ 1012 0.14 Lognormal
共AREA Category D兲 1.44⫻ 1012 0.14 Lognormal
共AREA Category E兲 5.60⫻ 1011 0.10 Lognormal
⌬ 0.90 0.30 Lognormal
DAF 1.10 0.14 Normal
␣ 0.80 0.14 Normal
Downloaded from ascelibrary.org by ARIZONA,UNIVERSITY OF on 05/29/14. Copyright ASCE. For personal use only; all rights reserved.

Annual train frequency f ti From 共Imam et al. 2006兲 and BS 共1980兲 0.14 Lognormal
a
Units: MPa3.

particular characteristics in rail traffic. Historical trains are pro- A summary of all the random variables and their distribution
posed for the first three periods, whereas the fourth period is characteristics considered in the probabilistic analyses is shown in
represented by the BS 共1980兲 medium traffic model, as suggested Table 1.
by the code for current rail traffic. Although load evolution and
its impact on the fatigue life of welded details has been consid- Damage and Remaining Life Formulation
ered in highway bridges 共Righiniotis 2006兲, this study adopts
nonevolving loads from present day onwards, which is likely to For the purposes of the probabilistic analyses, annual response
be a nonconservative assumption given the current outlook on the spectra are developed for each period of the rail traffic model
development of rail freight in the United Kingdom and other 共1900–1920, 1920–1940, 1940–1970, 1970–2005兲. These spectra
countries. Further details regarding the proposed rail traffic model are developed by multiplying the static stress ranges obtained
and the train characteristics 共axle weights, spacings, and train from previous deterministic analyses of the bridge under the pas-
frequencies兲 can be found in Imam et al. 共2006兲. sage of individual train types over one track of the bridge 共Imam
The annual frequency for each train type crossing the bridge et al. 2006兲 with the DAF and the ␣ factor. For each individual
is assumed to be lognormally distributed, as proposed by train crossing, a different value of DAF and ␣ is sampled from the
Ebrahimpour et al. 共1992兲, with a CoV of 0.14. Randomness assumed probability distributions in order to account for random-
in train frequencies is, of course, directly linked to the annual ness in these variables. This process of calculating the determin-
number of applied stress cycles. istic stress ranges is carried out f ti times, where f ti is the annual
One of the major sources of loading uncertainty is associated frequency of train type i in any given time period. Effectively a
with dynamic amplification. Following the field measurements Monte Carlo simulation is carried out, with random variables
presented by Byers 共1970兲 and Tobias and Foutch 共1997兲 on DAF and ␣ sampled from assumed distributions. Since the train
short-span steel railway bridges, a normal distribution with a frequencies f ti are also taken to be random, the above process is
mean value of 1.10 and a standard deviation of 0.15 is assumed repeated 103 times, in order to capture the uncertainty in the train
for the DAF. frequencies.
Modeling or epistemic uncertainty is typically captured by the The annual damage Da,l in each period l can be calculated
from its annual response spectrum and will be given, for the case
differences between the analytically/numerically obtained stress
of the two-slope British S-N curves, as
ranges ⌬␴a and the actual stress ranges ⌬␴ f . In general, actual
stresses obtained through field measurements are lower than their
calculated counterparts 共Sweeney et al. 1997兲. This can be attrib- Da,l = D⬘a,l + Da,l
⬙ 共4兲
uted to several reasons, such as the beneficial effects of the lon-
gitudinal and transverse distribution of train axle loads through where
the rails, sleepers and ballast, which are typically ignored in the

冉 冊
k1
analysis, the higher design axle loads given in codes as compared n ⌬␴i m
to the true axle loads, the partial end fixity of the various mem- ⬘ =
Da,l 兺
i=1 10
i
7
⌬␴0
if ⌬␴i 艌 ⌬␴0 共5兲
bers, the participation of the bracing and the floor system on the
response of the structure and, finally, the composite action of

冉 冊
the ballast and track with the bridge superstructure. In order to k2
n ⌬␴i m+2
account for this class of uncertainties, a factor ␣ = ⌬␴ f / ⌬␴a is
introduced, which is assumed to follow a normal distribution with
⬙ =
Da,l 兺
i=1 10
i
7
⌬⌬␴0
if ⌬␴i 艋 ⌬␴0 共6兲

a mean value of 0.80. This value was obtained by comparing


static stresses obtained analytically or numerically with static val- where k1 and k2 = number of stress range blocks ⌬␴i in the annual
ues obtained through field measurements 共Byers 1976; Adamson response spectrum above and below the CAFL, respectively; and
and Kulak 1995; Sweeney et al. 1997兲. By comparison, the ni = number of cycles in each ⌬␴i, which as mentioned previously
AREA 共1996兲 code suggests a value of 0.85 for span lengths up to is randomized through f ti.
23 m. This is likely to be higher than an average value. Following By contrast, for the single slope AREA S-N curves 共Category
the proposal of Byers 共1976兲, a CoV of 0.14 is adopted for ␣. E兲, the annual damage will be given as

240 / JOURNAL OF BRIDGE ENGINEERING © ASCE / MAY/JUNE 2008

J. Bridge Eng. 2008.13:237-244.


Downloaded from ascelibrary.org by ARIZONA,UNIVERSITY OF on 05/29/14. Copyright ASCE. For personal use only; all rights reserved.

Fig. 5. Annual spectrum of calculated stress ranges for the period Fig. 7. Annual spectrum of calculated stress ranges for the period
1900–1920 1940–1970

k1
1
Da,l = 兺
C i=1
ni共⌬␴i兲m 共7兲 Based on the limit state function of Eq. 共8兲, the probability of
fatigue failure can be defined as the probability of g 艋 0. In other
words
In the case of the combined AREA S-N curve for punched holes
共combination of Categories C and D兲, Eq. 共7兲 is modified accord- P f = P关g共X兲 艋 0兴 = P关Tr 艋 t兴 = ⌽共− ␤兲 共10兲
ingly.
The probability of failure is calculated by considering a limit where g 艋 0 represents the “failure” domain; ␤ = reliability index,
state function given as and ⌽共 兲 = standard normal distribution function. This time-
varying failure probability is estimated using Monte Carlo simu-
lp Tl
lation with 106 samples. The sample size was determined through
g=⌬−D=⌬− 兺 兺 Da,l
l=1 j=1
共8兲 a convergence study as discussed by Imam 共2006兲.

where l p = number of periods 共or annual response spectra兲 of


interest; and Tl = number of years in each period. For the traf- Results and Discussion
fic model adopted here 共historical load model plus BS5400
medium traffic兲, l p = 4, T1 = 20, T2 = 20, T3 = 30, and T4 艌 35.
Annual Response Spectra
The last inequality indicates that the sum in Eq. 共8兲 is calculated
by extrapolating the BS5400 medium traffic model into the fu- Figs. 5–8 show the annual response spectra for critical inner
ture, up to the point of fatigue failure of the connection 共g 艋 0兲, stringer-to-cross-girder connections for each distinct period of
without accounting for any future load evolution. Under this as- the rail traffic model. As mentioned previously, these spectra in-
sumption, the remaining fatigue life Tr of the connection, assum- corporate the random effect of DAF, ␣, and f ti. In all the figures,
ing that all failures occur in the period from 1970 onwards, is the CAFL for the modified Class B and the Class WI considered
given as in the U.K. codes, as well as the mean values ␮ and the CoV of
l p Tl the stress range are indicated. As can be seen, the majority of the
⌬ − 兺 兺 Da,l stress cycles experienced by the connection are below the fatigue
l=1 j=1 limits. It is also apparent that there is a continuous increase in the
Tr = 共9兲
Da,4 level of the mean stress range, in effect doubling in value between
the beginning and the end of the 20th century. On average, the
For failures occurring in preceding periods, the denominator of
increase is much more significant in the latter half of this period.
Eq. 共9兲 is modified accordingly.

Fig. 6. Annual spectrum of calculated stress ranges for the period Fig. 8. Annual spectrum of calculated stress ranges for the period
1920–1940 1970 onwards

JOURNAL OF BRIDGE ENGINEERING © ASCE / MAY/JUNE 2008 / 241

J. Bridge Eng. 2008.13:237-244.


Downloaded from ascelibrary.org by ARIZONA,UNIVERSITY OF on 05/29/14. Copyright ASCE. For personal use only; all rights reserved.

Fig. 9. Probability density function of predicted fatigue life for Fig. 10. Probability of fatigue failure versus time for U.K. detail
various detail classifications classification

In essence, the distributions corresponding to successive periods the remaining fatigue life of the connection is found to be
may be seen to shift to the right, thus pushing progressively more 110 years. For the same probability of failure and assuming its
stress ranges above the CAFL. American counterpart 共AREA punched holes兲, the remaining
Fitted probability distributions to each response spectrum are fatigue life drops to 40 years. However, given the large number of
also shown in Figs. 5–8. Different distribution types were tested bridge connections present in the U.K. railway network, it is pru-
and the best fit was found to be provided by the Weibull law, dent to consider a target lower than 2.3%. For example, it is
whose probability density function 共PDF兲 is given as estimated that roughly 6,000 bridges similar to the type presented

冋 冉 冊册
in Fig. 1 are in service today with an average of about 16 fatigue-

f共x兲 = 冉冊
␭ x
␩ ␩
␭−1
exp −
x


共11兲
critical connections each, which corresponds to a total of almost
100,000 共Imam 2006兲. Thus, a 1% failure probability target 共or a
corresponding ␤ = 2.3兲 would imply 1,000 connection failures
where ␭ and ␩ are the distribution parameters. The estimated within 85 years assuming modified Class B or within 25 years
values of these parameters are given in the respective figures. assuming the AREA punched holes classification. Of course, once
Since the stress ranges close to and above the fatigue limit make a connection fails, load redistribution is likely to occur within
the largest contribution to fatigue damage, the distribution param- such bridges, which is not taken into account here. The differ-
eters were determined considering the best fit with respect to ences between the results associated with the wrought-iron S-N
the stress range blocks on the right tail of the histograms. The curves 共Railtrack Class WI versus AREA Category E兲 are found
first period spectrum is represented by an exponential distribution to be significantly higher, with the AREA curve leading to very
关Eq. 共11兲 with ␭ = 1兴, whereas the following three are better de- high probabilities of fatigue failure.
scribed by the two-parameter Weibull. The great degree of variation observed in remaining fatigue
life estimates 共Figs. 10 and 11兲 can be attributed to the differences
Remaining Fatigue Life between the S-N curves that were considered 共Figs. 2–4兲.
Comparison of the BS5400 modified Class B mean S-N
Fig. 9 depicts the PDF of the predicted fatigue life of the connec- curve with the AREA mean curve for punched holes shows
tion according to two U.K. and one U.S. detail classification. The that the former results in more conservative life estimates up to
mean life ␮ and its standard deviation ␴ are also presented in the N ⬇ 5 ⫻ 107 cycles. Thereafter, the latter becomes more onerous.
figures. The PDF are very well described by lognormal distribu- Since the majority of the stress ranges experienced by the con-
tions, as might be expected given the input distributions shown in nection are close to and below the CAFL, as was demonstrated in
Table 1. From an engineering point of view, it is important to
emphasize that the significance of these distributions lies in the
very high standard deviations rather than in the mean values. The
estimated values for the standard deviations offer a quantitative
assessment of the overall uncertainty associated with fatigue
evaluation procedures and underline the importance of adopting
intensified inspection and management plans for this class of
bridges.
Failure probabilities from year 2004 onwards are calculated
according to Eq. 共10兲 and based on the spectra shown in
Figs. 5–8. Fig. 10 shows the probability of fatigue failure P f
of the connection against time for two U.K. detail classifications
共modified Class B and Class WI兲, whereas Fig. 11 presents results
obtained by considering the AREA S-N curves 共punched holes
and Category E兲. It is evident from Figs. 10 and 11 that detail
classification has a significant effect on the fatigue failure prob-
ability of the connection. For a 2.3% probability of failure 共or a Fig. 11. Probability of fatigue failure versus time for the AREA
corresponding ␤ = 2.0兲 and assuming a modified Class B detail, detail classification

242 / JOURNAL OF BRIDGE ENGINEERING © ASCE / MAY/JUNE 2008

J. Bridge Eng. 2008.13:237-244.


emphasize that accurate estimation of stress ranges is essential, a
process that would greatly benefit from field measurements.

Conclusions

A methodology for probabilistic fatigue evaluation of riveted rail-


way bridges was presented, which is an extension of earlier de-
terministic studies carried out by the authors. Response spectra
for a fatigue-critical connection were developed and Weibull dis-
tributions were fitted to the response histograms. These spectra
correspond to four time periods, the first dating back to 1900,
Downloaded from ascelibrary.org by ARIZONA,UNIVERSITY OF on 05/29/14. Copyright ASCE. For personal use only; all rights reserved.

reflecting changes in rail traffic. It was found that the majority of


Fig. 12. Effect of different variables on the remaining fatigue the stress cycles experienced by the connection are close to and
life corresponding to a target 2.3% probability of failure 共AREA below the CAFL, but also that there is a continual and accelerat-
classification兲 ing increase in the mean stress range with time, ranging from
5.8 MPa to 12.6 MPa from the beginning to the end of the 20th
century.
Comparison of experimental data retrieved from the literature
Figs. 5–8, and correspond to the lower part of the S-N curves with available S-N curves revealed that the BS5400 modified
共⌬␴ ⬍ 50 MPa兲, this explains the lower remaining fatigue life es- class B, the Railtrack Class WI, and the AREA curve for punched
timates obtained by considering the AREA classification for holes present the most reasonable options for the fatigue assess-
punched holes. In the case of the wrought-iron detail classifica- ment of riveted connection details. However, substantial differ-
tions, it can be seen in Fig. 4 that the AREA Category E S-N ences in fatigue life estimates were obtained using these different
curves are more onerous than the Railtrack Class WI curves along classifications for riveted details. These differences were attrib-
the whole spectrum. Due to the single slope of the AREA curves uted to the deviations between the S-N curves in the low stress/
共m = 3兲, the deviations between the two sets of curves can be seen high endurance region, typically beyond 10 million cycles. Since
to increase for lower stress ranges and higher lives. For stress the S-N curves in this region are supported by only a few test
ranges ⌬␴ ⬍ 50 MPa, which is the region of interest, there are results, any remaining life estimates for this class of bridges
very large differences in fatigue life predictions between the should be viewed with caution. Nonetheless, the results suggest
AREA and Railtrack curves, as can be seen in Figs. 10 and 11. that a number of critical bridge connections are approaching the
Given the very few test results available in the low stress/high end of their service life and, as a result, a timely management of
endurance region and the significant differences in the proposed repairs and replacement needs to be developed, especially since it
S-N curves therein, the numerically derived fatigue life estimates is unlikely that fatigue evaluation procedures will be sufficiently
presented herein should be viewed with caution. On the other improved within a relatively short time frame relevant to this
hand, given the implications of fatigue damage on a large number class of bridges. Further parametric studies revealed that fatigue
of connections in a busy network, these estimates suggest that life exhibits the highest sensitivity to detail classification and the
measures to improve our ability to predict 共e.g., inspection, field ␣ factor, which represents model uncertainty. This highlights the
monitoring兲 and prolong 共e.g., repair兲 fatigue life are essential importance of field monitoring for old bridges approaching the
for effective infrastructure management. A timely management of end of their useful life. The damage limit and the DAF were
repairs and replacement needs to be in place well before riveted found to be of less importance.
bridges reach the end of their already considerably long service The results presented in this paper are based on the assumption
life. of nonevolving loads from present day onwards. It would be
sensible to consider future increases in train axle loads and fre-
Parametric Studies quencies, which would obviously reduce the remaining life esti-
mates further. In particular, possible introduction of higher axle
Parametric studies in order to investigate the effect of different
loads 共from 25 t to 30 t per axle兲, which is currently being dis-
variables on the failure probability of the connection over time
cussed across Europe, could have significant effect on remaining
have also been carried out. Each time a single variable 共in the
fatigue life estimates, given the evidence presented herein regard-
case of random variables, their mean value兲 was varied, while
ing the implications of pushing stress ranges beyond the currently
keeping the others fixed. Fig. 12 shows the effect of the damage
limit ⌬, ␣ factor, and DAF on the characteristic 共2.3% probability accepted fatigue limit levels.
of failure兲 fatigue life of the connection. The base model is asso-
ciated with the characteristics presented in Table 1 combined with
the AREA S-N curve for punched holes. Fig. 12 shows the per-
centage increases 共towards the right兲 and the percentage decreases Acknowledgments
共towards the left兲 brought about by changing, one at a time, the
mean values of the random variables. Notwithstanding the domi- The work described in this paper forms part of a project supported
nant influence of the detail classification, the most substantial by Network Rail. The opinions expressed are those of the writers
change in fatigue life is affected by the ␣ factor, which represents and do not necessarily represent those of the sponsoring organi-
the model uncertainty. In comparison, the effect of the DAF and zations. We would like to thank Mr. Brian Bell, Network Rail
the damage limit ⌬ can be seen to be less profound. These results project officer for valuable discussions and feedback.

JOURNAL OF BRIDGE ENGINEERING © ASCE / MAY/JUNE 2008 / 243

J. Bridge Eng. 2008.13:237-244.


Notation Proc., National Structural Engineering Conf., Vol. I, ASCE,
New York, 153–170.
The following symbols are used in this paper: Chryssanthopoulos, M. K., and Righiniotis, T. D. 共2006兲. “Fatigue
C ⫽ constant in an S-N curve depending on material reliability of welded steel structures.” J. Constr. Steel Res., 62共11兲,
1199–1209.
properties and loading conditions;
Ebrahimpour, A., Maragakis, E. A., and Ismail, S. 共1992兲. “A fatigue
D ⫽ total fatigue damage;
reliability model for railway bridges.” Proc., 6th Specialty Conf. on
Da,l ⫽ annual damage in period l;
Probabilistic Mechanics and Structural Engineering, and Geotechni-
DAF ⫽ dynamic amplification factor;
cal Reliability, ASCE, New York, 320–323.
f共 兲 ⫽ probability density function; Fisher, J. W., Mertz, D. R., and Zhong, A. 共1983兲. “Steel bridge members
f t ⫽ annual train frequency; under variable amplitude long life fatigue loading.” Transportation
g ⫽ limit state function; Research Board, NCHRP Rep. No. 267, National Research Council,
k ⫽ number of different stress range blocks in response Washington, D.C.
Downloaded from ascelibrary.org by ARIZONA,UNIVERSITY OF on 05/29/14. Copyright ASCE. For personal use only; all rights reserved.

spectrum; Fisher, J. W., Nussbaumer, A., and Keating, P. B. 共1993兲. “Resistance of


l p ⫽ number of periods; welded details under variable amplitude long-life fatigue loading.”
m ⫽ inverse of the slope of the S-N curve; Transportation Research Board, NCHRP Rep. No. 354, National Re-
N ⫽ number of cycles to failure; search Council, Washington, D.C.
n ⫽ number of applied cycles; Fisher, J. W., Yen, B. T., Wang, D., and Mann, J. E. 共1987兲. “Fatigue
P f ⫽ probability of failure; and fracture evaluation for rating riveted bridges.” Transportation
Tl ⫽ number of years in period l; Research Board, NCHRP Rep. No. 302, National Research Council,
Tr ⫽ remaining fatigue life; Washington, D.C.
Imam, B. 共2006兲. “Fatigue analysis of riveted railway bridges.” Ph.D.
␣ ⫽ ratio of measured to calculated stress, model
thesis, Univ. of Surrey, U.K.
uncertainty; Imam, B., Righiniotis, T. D., and Chryssanthopoulos, M. K. 共2005兲.
␤ ⫽ reliability index; “Fatigue assessment of riveted railway bridges.” Int. J. Steel Struct.,
⌬ ⫽ damage limit; 5共5兲, 485–494.
⌬␴ ⫽ stress range; Imam, B., Righiniotis, T. D., Chryssanthopoulos, M. K., and Bell, B.
⌬␴0 ⫽ constant amplitude fatigue limit; 共2006兲. “Analytical fatigue assessment of riveted rail bridges.” Proc.,
␭, ␩ ⫽ parameters of Weibull distribution; Inst. Civil Eng. (ICE), Bridge Eng., Vol. 159共BE3兲, London, 105–116.
␮ ⫽ mean; and Kulak, G. L. 共2000兲. “Fatigue strength of riveted shear splices.” Prog.
␴ ⫽ standard deviation. Struct. Eng. Mater., 2, 110–119.
Kunz, P., and Hirt, M. A. 共1993兲. “Reliability analysis of steel railway
Subscripts
bridges under fatigue loading.” Proc., IABSE Colloquium Copen-
i ⫽ train type; and hagen 1993, Remaining Structural Capacity, International Association
i ⫽ stress range block in response spectrum. for Bridge and Structural Engineering 共IABSE兲, Zurich, Switzerland,
53–60.
Miner, M. A. 共1945兲. “Cumulative damage in fatigue.” ASME J. Appl.
References Mech., 12共3兲, 159–164.
Moses, F., Schilling, C. G., and Raju, K. S. 共1987兲. “Fatigue evaluation
procedures for steel bridges.” Transportation Research Board,
Adamson, D. E., and Kulak, G. L. 共1995兲. “Fatigue tests of riveted bridge
NCHRP Rep. No. 299, National Research Council, Washington, D.C.
girders.” Structural Engineering Rep. No. 210, Univ. of Alberta,
Railtrack. 共2001兲. “Railtrack line code of practice: The structural assess-
Canada.
ment of underbridges.” RT/CE/C/025, Railtrack, U.K.
Al-Emrani, M. 共2005兲. “Fatigue performance of stringer-to-floor-
Righiniotis, T. D. 共2006兲. “Effects of increasing traffic loads on the fa-
beam connections in riveted railway bridges.” J. Bridge Eng., 10共2兲,
tigue reliability of a typical welded bridge detail.” Int. J. Fatigue,
179–185.
28共8兲, 873–880.
American Association of State Highway Transportation Officials
Sweeney, R. A. P., Oomen, G., and Le, H. 共1997兲. “Impact of site mea-
共AASHTO兲. 共1998兲. LRFD bridge design specifications, 2nd Ed.,
surements on the evaluation of steel railway bridges.” Proc., IABSE
Washington, D.C.
American Railway Engineering Association 共AREA兲. 共1996兲. “Chapter Workshop Lausanne 1997, Evaluation of Existing Steel and Compos-
15: Steel structures.” Manual for railway engineering, Washington, ite Bridges, International Association for Bridge and Structural Engi-
neering 共IABSE兲, Zurich, Switzerland, 139–147.
D.C.
British Standards 共BS兲. 共1980兲. “Steel, concrete and composite bridges: Tobias, D. H., and Foutch, D. A. 共1997兲. “Reliability-based method for
Part 10: Code of practice for fatigue.” BS5400, British Standards fatigue evaluation of railway bridges.” J. Bridge Eng., 2共2兲, 53–60.
Institute, London. Wirsching, P. H. 共1995兲. “Probabilistic fatigue analysis.” Probabilistic
Brühwiler, E., and Kunz, P. 共1993兲. “Remaining fatigue life of a riveted structural mechanics handbook, C. Sundararajan, ed., Chapman and
railway bridge.” Proc., IABSE Colloquium Copenhagen 1993, Hall, New York.
Remaining Structural Capacity, International Association for Bridge Xie, M., Bessant, G. T., Chapman, J. C., and Hobbs, R. E. 共2001兲.
and Structuctural Engineering 共IABSE兲, Zurich, Switzerland, “Fatigue of riveted bridge girders.” Struct. Eng., 79共9兲, 27–36.
375–383. Zhou, Y. E., Yen, B. T., Fisher, J. W., and Sweeney, R. A. P. 共1995兲.
Byers, W. G. 共1970兲. “Impact from railway loading on steel girder spans.” “Examination of fatigue strength 共Sr-N兲 curves for riveted bridge
J. Struct. Div., 96, 1093–1103. members.” Proc., 12th Annual Bridge Conf., Engineer’s Society of
Byers, W. G. 共1976兲. “Rating and reliability of railway bridges.” Western Pennsylvania, Pittsburgh, Pa., 102–109.

244 / JOURNAL OF BRIDGE ENGINEERING © ASCE / MAY/JUNE 2008

J. Bridge Eng. 2008.13:237-244.

You might also like