Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Engineering Structures 126 (2016) 457–468

Contents lists available at ScienceDirect

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

Reliability of a corroded RC beam based on Bayesian updating of the


corrosion model
Sharvil Alex Faroz (Doctoral candidate), Nikil N. Pujari (Doctoral candidate), Siddhartha Ghosh ⇑ (Professor)
Structural Safety, Risk & Reliability Lab., Department of Civil Engineering, Indian Institute of Technology Bombay, Mumbai 400076, India

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

Article history: RC structures degrade primarily due to corrosion-induced damages, mostly by the loss of steel rebar vol-
Received 28 September 2015 ume. The prediction of time-varying damage due to corrosion is important for assessing the residual life
Revised 2 August 2016 of a structure and making decisions on maintenance/repair. Existing models of prediction fail to provide
Accepted 3 August 2016
realistic estimates of steel loss over time. This paper presents a new methodology for the corrosion reli-
ability analysis of reinforced concrete components, where the corrosion loss is estimated employing
Bayesian updating using field measurements, which also incorporates instrument biases and errors in
Keywords:
the NDT evaluations. Based on compiled experimental results, the updated model provides significantly
Corrosion
Pitting
improved estimation of steel loss over time. Sample reliability analyses over the design life of a corroded
Bayesian updating RC beam show the importance of updating over the existing model, for both pitting and uniform corro-
NDT sion considerations.
Measurement error Ó 2016 Elsevier Ltd. All rights reserved.
Reliability

1. Corrosion in reinforced concrete primary interest of a structural/maintenance engineer in such


cases is to find the loss of steel in corrosion affected reinforcements
Reinforced concrete (RC) structures occupy a major share of the over time which is related to a moment (or shear) failure at the
infrastructural stock of many countries across the world. The rein- ultimate limit state (ULS). Although a beam may also be governed
forcing steel within concrete is protected from corrosion due to the by the serviceability limit state (SLS) in terms of cracking of cover
formation of a passive film around the steel rebar, along with the concrete or loss of stiffness [3,4], in the present study we focus
low porosity, permeability and high electrical resistivity of good only on the ultimate limit state of flexure [7,8]. Though models
quality concrete [1]. If these aspects are compromised, then corro- exist to predict loss of steel over time, significant deviations from
sion of steel in concrete commences. Generally, aggravating agents actual measurements are typically reported [9]. Deviation of the
like chloride (Cl), carbon dioxide (CO2) and to some extent sul- model from in situ situations may be due to a lack of knowledge
phur dioxide (SO2) penetrate through the cover concrete and about the composition of corrosion products and/or modelling
depassivate the steel to initiate corrosion [2]. Rust produced due the complexity of corrosion process [10]. In addition, these factors
to corrosion occupies more volume than the volume of steel con- also depend on (ambient) conditions which are inherently random.
sumed to produce that rust. This rust accumulates in the steel- The prediction model, which is developed based on limited test
concrete interface exerting pressure on concrete which leads to specimens, may not represent the actual structure either.
cracking, spalling and delamination of cover concrete [3]. Overall, A precise estimation of the residual steel area of rebars is possi-
corrosion leads to (i) loss of steel section, (ii) cover cracking and ble through the destruction of cover concrete, however such oper-
(iii) loss of bond, ultimately leading to a reduction in the capacity ations may not always be feasible. As an alternative to invasive
of a structural member over time [4]. An estimate of the time- techniques, continuous or routine monitoring of the structure
varying capacity of RC structures/components is of prime impor- using non-destructive techniques (NDT) may be more appropriate.
tance from the perspective of life-cycle performance. These techniques include but are not limited to: infra-red ther-
Experimental evidences suggest that one parameter which can mography [11–13] and X-ray imaging [14–16] for determining
be associated with various damage levels of an RC structure is the steel loss; and linear polarisation resistance technique (LPR)
the amount of steel lost due to corrosion [5,6]. Therefore, the [17,18] for measuring the rate of corrosion. Although these
schemes can provide a reasonable estimate of the time-varying
condition of a structure, perpetual inspection is seldom possible
⇑ Corresponding author. and measurements are restricted to a few optimised numbers in
E-mail address: sghosh@civil.iitb.ac.in (S. Ghosh).

http://dx.doi.org/10.1016/j.engstruct.2016.08.003
0141-0296/Ó 2016 Elsevier Ltd. All rights reserved.
458 S.A. Faroz et al. / Engineering Structures 126 (2016) 457–468

order to minimise the life-cycle cost [19]. Therefore, a planned For the work presented in this paper, the likelihood function
scheme of non-destructive data acquisition through intermittent used for the Bayesian updating incorporates a model for the ran-
monitoring and using it – via Bayesian inference – for future pre- dom measurement error (along with the model/prediction error).
diction of corrosion is deemed suitable for the life-cycle manage- The effects of pitting are included using the pitting factor varying
ment of an RC structure. over space and/or time, over and above the updated prediction of
The effectiveness of a probabilistic approach to estimate corro- uniform corrosion. In a major extension of the work by Faroz
sion is underlined by many researchers over the last two decades. et al. [9], the time-varying reliability of a corroded singly rein-
The classical statistical methods, however, rely on large quantities forced beam is estimated based on its flexural failure; and finally,
of data to be available to make accurate predictions [20]. Thus, a comparison between results from the updated and the original
when limited with small amount of data, Bayesian updating can corrosion models is presented.
be adopted by combining empirical or physics-based prediction
models with NDT measurements. 2. Steel loss due to corrosion
Corrosion in an RC structure involves two principal phases, (i)
initiation and (ii) propagation [2]. Initiation denotes the period Faraday’s law has been traditionally adopted to model the loss
from construction of new structures to corrosion commencement, of steel due to corrosion. This law does not consider the effect of
while propagation primarily involves loss of steel (and cracking of thickening of rust layer over the corroding rebar surface and over-
cover) once corrosion has initiated. In the past, there have been estimates the loss of steel [28]. As corrosion progresses, thickness
studies addressing the corrosion related problem in RC structures of the rust goes on increasing and the rate of rust formation
pertaining to these phases. Some of these studies used monitored depends on the thickness of rust already produced [29]. Due to a
data in conjunction with Bayesian inference, for example Rafiq progressive increase in the diffusion distance, the rate of rust for-
et al. [21], Keßler et al. [22], and Samarakoon and Sælensminde mation decreases [30]. For uniform corrosion, thickness of the rust
[23]. Of these, while Rafiq et al. [21] developed a methodology to layer is proportional to the total mass of rust per unit length of the
predict corrosion initiation by making use of embedded sensors corroding bar ðW r Þ, and the rate of rust formation may be
in concrete cover, the other two works relied primarily on the expressed as
half-cell potential test data to obtain an updated estimate of the
corrosion initiation time. The research in the past decade has dW r kw
¼ c1 ð1Þ
shifted to the propagation phase, without overlooking the initia- dt Wr
tion period [24]. In the recent past, there have been studies apply-
where t is the time from corrosion initiation, kw is a constant of pro-
ing Bayesian inference using monitored data, in the propagation
portionality and the exponent c depends on the rust layer formed
phase as well. Suo and Stewart [25] used visual inspection to
over the corroding surface [30]. When the rust layer does not
update the prediction of proportion of surface cracking due to cor-
adhere to the corroding surface, c approaches 1 and for an adherent
rosion related volume expansion. The most significant work in the
rust layer it approaches 2. For diffusion based processes, it takes a
use of Bayesian updating for the estimation of steel loss due to cor-
value between 2 and 3 depending on the densification of the corro-
rosion was that by Akiyama et al. [26], wherein the authors consid-
sion products [31], 2 being for uniform density. The amount of rust
ered both the phases of corrosion. The corrosion initiation time
produced by the time t can be obtained by integrating Eq. (1):
was updated using chloride concentration results through partially
destructive coring. Further, they adopted visual inspection on crack Z t 1c
width to update the steel loss model, through an empirical relation W r ðtÞ ¼ ckw dt ð2Þ
0
between the two. The updated corroded profile of rebars was then
used to obtain the failure probability of the structure. However, Liu and Weyers [29] adopted this model with c ¼ 2 to estimate
they adopted a steel loss model linearly varying with time, that the rust formation. The amount of steel consumed ðW s Þ can be
did not consider the effect of thickening of the rust layer over related to the amount of rust generated as [32]
the corroding rebar surface. Also, the corrosion loss was considered
W s ¼ asr W r ð3Þ
to be uniform, and pitting was disregarded. This study, based on
crack width inspection, implied that inspections need to be per- and combining Eqs. (2) and (3) with c ¼ 2, the steel loss can be
formed at an advanced stage of degradation, which may not be obtained as
the best strategy, as such reactive inspection is likely to be more Z 12
t
expensive than a proactive one [21]. W s ðtÞ ¼ 2a2
sr kw dt ð4Þ
Considering the random nature of the corrosion process, limited 0
amount of data and lack of confidence in prediction, Faroz et al. [9]
where kw ¼ pAw D0 icorr [32], D0 is the original diameter of the rebar
developed a similar methodology for updating the corrosion model
and icorr is the mean annual rate of corrosion, yielding
using the available steel loss data. Their corrosion growth model qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
was parabolic in time, reflecting the effect of the thickening of W s ðtÞ ¼ 2pD0 icorr Aw a2sr t ð5Þ
the rust layer. The present work improves this method to incorpo-
rate two major aspects: In these equations, the parameters Aw and asr are related to the
corrosion process and the amount of rust generated, respectively.
 The output of an NDT depends on random factors – like the con- Therefore, these parameters cannot be measured directly and are
dition and working environment of the structure, sensitivity of treated in the present work as un-observable parameters. Bhargava
the NDT equipment, material imperfection and operator skills et al. [32] gave estimates for these quantities based on the exper-
[27], etc. – leading to random measurement errors. imental evaluations by Liu [10]. Performing a regression analysis
 The corrosion of rebars due to chloride ingress is not uniform on these laboratory-based data for D0 measured in mm, icorr in
because of the heterogeneous ambience of a rebar embedded lA/cm2, W s in mg/mm and t in years, they found: Aw ¼ 2:48614
in concrete. This localised or ‘pitting’ corrosion – a spatial vari- and asr ¼ 0:61309. Their time-varying steel loss model for uniform
ability – if ignored, may lead to an underestimation of the prob- corrosion, thus, can be expressed as
ability of failure [7,8]. pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
W s ðtÞ ¼ 2:423 D0 icorr t ð6Þ
S.A. Faroz et al. / Engineering Structures 126 (2016) 457–468 459

2.1. Models for estimating pitting corrosion Table 1


Time-varying Gumbel parameters of R.

The uniform corrosion model of Eq. (5) is an idealistic assump- D0 (mm) L0 = 50 mm L0 = 100 mm
tion, whereas a realistic estimation of the failure of a structure due m j m j
to corrosion needs to account for pitting effects. A brief review of
12 0.151gt þ 1 0.308gt 0.306gt þ 1 0.312gt
available pitting models for RC structures is presented in this 14 0.120gt þ 1 0.274gt 0.261gt þ 1 0.275gt
section. 20 0.093gt þ 1 0.329gt 0.277gt þ 1 0.360gt
González et al. [33] defined the pitting depth factor Rp as the
ratio of the maximum pit depth ðpmax Þ and the average penetration
of corrosion based on uniform corrosion ðpavg Þ (Fig. 1). They found W s ðtÞ
Rp to vary between 4 and 8 under simulated natural corrosion test-
gt ¼ ð8Þ
W0
ing and between 5 and 13 for impressed current corrosion testing.
where W 0 is the mass per unit length of an un-corroded bar and
The DuraCrete model [34] used a statistical description of the pit-
W s ðtÞ is obtained from Eq. (5). The average residual area of steel
ting factor, with Rp  N ð9:28; 4:04Þ. The statistics of Rp was further
can be expressed in terms of gt :
developed by Stewart [7] and Stewart and Alharthy [8], where the
pitting factor was assumed to be time-invariant and statistically Aavg ¼ A0 ð1  gt Þ ð9Þ
independent for each discrete element of a rebar. Further, it was
where A0 is the original cross-sectional area of a rebar.
assumed that only one critical pit formed inside the considered
Kashani et al. [38] defined an area pitting coefficient as a func-
element, and the minimum area ðAmin Þ was calculated using the
tion of the location along the length of a rebar: Ra ¼ AðxÞ=Aavg ,
formulation suggested by Val and Melchers [35]. Later, Darmawan
where AðxÞ is the cross-sectional area of the corroded bar at loca-
[36] modified their probabilistic model to obtain a time-varying
tion x. They found that Ra follows a lognormal distribution with
model for Rp , by making the location and scale parameters of a
time-varying parameters. Tang et al. [39] suggested a mixed nor-
Gumbel distribution to be functions of icorr and t.
mal distribution for AðxÞ. However, it is unclear which and how
Zhang et al. [37] commented that using Rp may result in an
many normal modes need to be adopted in the time-varying model
inaccurate calculation of the residual cross-sectional area, as the
of AðxÞ.
form and number of pits formed may be complex, to determine
Based on this critical review of existing models, to incorporate
the actual reduced cross-section, as opposed to the simplified
the effect of pitting in corrosion, the model suggested by Zhang
assumption of a single critical pit [7,8,36]. Instead, they suggested
et al. [37] is found to be the most appropriate choice in modelling
a new pitting factor (R) relating the minimum area to the area
steel loss, since it directly characterises the minimum residual
based on uniform corrosion. R expresses the area spatial hetero-
cross-sectional area in a statistical sense.
geneity in pitting:
Aavg 3. Bayesian inference
R¼ ð7Þ
Amin
where Aavg is the average residual cross-sectional area of a corroded Bayesian inference provides a mathematical approach to com-
rebar estimated from the mass loss over an element length based on bine previous notions or beliefs about any phenomenon with
uniform corrosion (Fig. 1). Based on 3D scanning of corroded rebars, newly available information [40,41]. Rebar corrosion in RC is a
Zhang et al. [37] created a time-varying probabilistic model for R phenomenon for which we have scientific or empirical models,
(Table 1), where the scale (m) and location (j) parameters of the but the estimations based on these models are usually different
Gumbel distribution are expressed as functions of the mass loss from the field values. These models rely significantly on some
ratio at time t, which is defined as (un-observable) parameters which we are not very confident about
and their values come from limited statistical data available to the
researchers who calibrated these parameters. Considering that the

Fig. 1. Rebar cross-section for uniform and pitting corrosion.


460 S.A. Faroz et al. / Engineering Structures 126 (2016) 457–468

Table 2
Statistics of Aw .

D0 (mm) icorr (lA/cm2) W s (mg/mm) asr W r (mg/mm) t (years) kw Aw

16 2.41 19.75 0.778 25.39 1.84 175.6 1.446


0.723 27.32 202.8 1.674
0.700 28.22 216.4 1.786
0.629 31.40 268.0 2.212
0.622 31.76 274.1 2.262
0.523 37.77 387.7 3.200
0.348 56.76 875.6 7.228
16 1.79 30.21 0.778 38.83 3.54 212.9 2.366
0.723 41.78 246.6 2.741
0.700 43.15 263.1 2.923
0.629 48.03 325.8 3.621
0.622 48.57 333.2 3.703
0.523 57.76 471.3 5.237
0.348 86.81 1064 11.83
16 3.75 14.98 0.778 19.25 0.72 257.4 1.366
0.723 20.72 298.1 1.581
0.700 21.40 317.9 1.687
0.629 23.81 393.8 2.089
0.622 24.08 402.7 2.136
0.523 28.64 569.6 3.022
0.348 43.04 1286 6.826
12.7 1.8 15.64 0.778 20.10 2.38 84.90 1.182
0.723 21.63 98.31 1.369
0.700 22.34 104.9 1.460
0.629 24.86 129.9 1.808
0.622 25.14 132.8 1.849
0.523 29.90 187.9 2.616
0.348 44.94 424.3 5.909

monitoring of corrosion through NDT is a fairly routine event for asr  Uð0:348; 0:778Þ ð11Þ
important structures these days, a Bayesian updating of the avail-
Bhargava et al. [32] proposed a procedure to compute Aw from
able corrosion model is proposed here based on limited but actual
available data. Following this procedure, 28 values of Aw are gener-
evaluations of the in situ condition. Bayesian inference is adopted
ated (Table 2) using the data reported by Liu [10] (this is the same
to update the un-observable random variables (H) based on the
data set used by Bhargava et al.). The prior probability model for Aw
observed/measured data (X). According to Bayes’ theorem, the
is obtained by fitting this statistic using the Kolmogorov-Smirnov
prior information described by the joint probability density func-
0 goodness-of-fit test. A lognormal distribution is found suitable to
tion, f ðHÞ, when combined with the monitored data – quantified
model Aw :
by a likelihood function pðX j HÞ, results in an updated/posterior
00
distribution, f ðH j XÞ: ln Aw  N ð0:9073; 0:6752Þ ð12Þ
00 0 Fig. 2(a) shows the empirical and the fitted CDFs of Aw . The prior
f ðH j XÞ / pðX j HÞf ðHÞ ð10Þ
PDF used in the updating is presented in Fig. 2(b). The use of a log-
In general, H ¼ ½ H1 H2 . . . Hl T and X ¼ ½ X1 X2 . . . Xn T , normal distribution is also justified based on Eq. (5), which requires
where Hl are the model variables to be updated subjected to the Aw to be non-negative.
observations Xn . The joint prior distribution is taken as the product of the densi-
In the steel loss model of Eq. (5), Aw and asr are the un- ties of individual parameters, by assuming (for the sake of simplic-
observable parameters that are usually obtained by a regression ity) each parameter to be independent of the other:
analysis of data obtained under a specific set of conditions. There- 0 0 0
fore, there exists significant uncertainty associated with these
f ðHÞ ¼ f ðAw Þf ðasr Þ ð13Þ
parameters due to the difference between the test and field condi- Prior joint densities based on correlated parameters can also be
tions. As a consequence, these parameters need to be treated as updated, if the necessary information regarding correlation is avail-
random variables in the corrosion model. The Bayesian updating able [43].
in this work considers H ¼ ½ Aw asr T , whose probability models
require to be updated based on steel loss ðW s Þ measurements. 3.2. Errors in modelling and measurement

3.1. Modelling prior probabilities The likelihood function used in Eq. (10) serves as a very useful
means of combining the observed/measured data with the prior
Through a compilation of data reported in literature, Jamali information. However, in using the measured data, one must not
et al. [42] gave the values of asr in the range of 0.348–0.778, corre- rule out the very likely possibility of discrepancy in the adopted
sponding to various possible corrosion products. They pointed out physics-based or empirical model and the presence of error in
that the rust generated at various sites – such as marine, urban and the specific measurement techniques used. The likelihood function
industrial environments – comprises of various products. In the used in Bayesian updating incorporates these aspects.
absence of any further information on the composition of rust NDT tools for corrosion induced damage monitoring are prone
products, we conservatively adopt a uniform distribution for the to measurement errors and have inherent deficiencies that may
prior probability model of asr : lead to uncertain observations [44]. For the present study, we
S.A. Faroz et al. / Engineering Structures 126 (2016) 457–468 461

Fig. 2. (a) Empirical and fitted (prior) CDFs of Aw ; (b) prior PDF of Aw .

assume that there exists an NDT tool capable of measuring the sion, an NDT instrument can be calibrated against ‘‘true” measure-
average (that is, uniform) loss of steel in concrete. In the context ments obtained after taking the corroded bar out and using precise
of the corrosion model of Eq. (5), W m denotes the quantity mea- techniques such as 3D scanning or a simple Archimedean principle
sured by the NDT and W p denotes the steel loss predicted by the of finding the volume of the un-corroded bar. This work does not
existing model: focus on obtaining biases for a specific NDT instrument, rather it
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi emphasises on the importance of b0 and b1 in the prediction of
Wp ¼ 2pD0 icorr Aw a2sr t ð14Þ steel loss over time. The methodology presented here shows how
the biases for a calibrated instrument can be incorporated in the
Relating the readings from an NDT to the ‘‘true value” of the updated prediction. Fitting a straight line on the squared steel loss
observed parameter is possible through measurement error mod- data reported by Torres-Acosta and Sagues [47], b0 is obtained as
els, such as lognormal or logit [45]. However, in the present study 1.42% of W 0 (as presented later in Fig. 3(b)). The non-
a simpler linear model [27] is adopted: dimensional scaling bias b1 is assumed to vary in the range 0.75–
1.25, to which the sensitivity of the results is discussed.
W f ¼ b0 þ b1 W m þ em ð15Þ
The difference between the predicted value and the true value
ðW p  W f Þ has to be zero for a correct prediction by the model.
where W f denotes the actual steel loss, b0 2 R is the constant bias
For a realistic model, however, a global random error
(typically expressed in % of W 0 ), b1 2 Rþ is the scaling bias of the
NDT instrument, and em  N ð0; rem Þ denotes the random error in
e  N ð0; re Þ is required to incorporate the deficiencies of the
model and the measurement techniques. Here,
measurements. b0 ¼ 0 and b1 ¼ 1 signifies an unbiased NDT tool.
½W p  ðb0 þ b1 W m Þ represents each realisation of the zero mean
Determining the value of these biases may be termed as the
Gaussian random variable e. Typically, re is estimated by the user
‘‘calibration” of an NDT tool. Al-Amin et al. [46] demonstrated
based on the beliefs about the model uncertainty and the presence
how these biases (and the random error) can be evaluated using
of noise and bias in the measurements [48]. The global error con-
Bayesian inference. Obtaining the bias statistics needs the actual
sists of the measurement error (em ) and the model error (eM ), with
‘field’ values of the measured quantity. For steel loss due to corro-

Fig. 3. (a) Prediction following the existing model; (b) calculation of the constant bias b0 .
462 S.A. Faroz et al. / Engineering Structures 126 (2016) 457–468

r2e ¼ r2eM þ r2em . Each of these errors can also be modelled sepa- Table 3
Steel loss data.
rately [49]. rem is typically obtained from a separate calibration
exercise of an NDT instrument. Thus, if rem is already available, Specimen t (days) W m (mg/mm)
reM can be included in the set of model variables H. 1 CPA1 13 22.89
The likelihood function is assumed to follow a normal distribution, 2 CPA2a 17 31.47
with re as the standard deviation of the quantity ½W p  ðb0 þ b1 W m Þ. 3 CPB1 24 40.15
4 CPB2 20 34.46
The likelihood for the ith measurement can be expressed as 5 CPC1 21 36.24
"  2 # 6 CPC2 28 53.44
1 W pi  ðb0 þ b1 W mi Þ 7 CPD1 18 28.64
pðW mi j HÞ ¼ p ffiffiffiffiffiffi
ffi exp  ð16Þ
2pre 2r2e 8 CPD2a 20 34.13
9 CPE1 39 69.88
10 CPE2 20 34.94
For n measured data, either a recursive scheme can be adopted
11 CPF1 18 31.32
(where the previously updated density will be used as the prior 12 CPF2a 17 30.79
for the next sequence of updating) or, if the data become available 13 CPG1 10 8.240
for all past measurements together, the likelihood function can be 14 CPG2 10 2.726
15 CP1 12 17.89
constructed as
16 CP2 30 33.73
Y
n
a
Data used for updating.
pðX j HÞ ¼ pðW mi j HÞ ð17Þ
i¼1

The choice of the likelihood function does not affect the updating investigation on a corroding 21 mm diameter rebar in concrete,
procedure. However, as suggested by Karandikar et al. [48], one over an anodic region, were presented. The experiment was con-
should consider the aspect that, since the likelihood is based on ducted using a constant current density with different measured
the user’s beliefs, the effect of the likelihood spread on the updated values of icorr for each specimen. For the present study, all observa-
estimations should be evaluated. tions are scaled to an icorr = 100.0 lA/cm2. This is done using the
pffiffiffiffiffiffiffiffiffiffi
fact that W / icorr t , and keeping t the same for a particular data,
pffiffiffiffiffiffiffiffiffi pffiffiffiffiffiffiffiffi
3.3. MCMC simulation for the posterior model we have W scaled =W original ¼ 100= icorr . Here, W original is the steel
loss measured in their test and W scaled is the steel loss used in this
With the availability of prior density functions and the likeli- paper. From the same work, additional data are used for the valida-
hood incorporating the measured data, the posterior or updated tion of the updated model. These data, including the ones used for
density of the desired parameter(s) can now be computed through updating, are presented in Table 3 and are shown in Fig. 3(a) along
Eq. (10). However, the evaluation of the posterior density is diffi- with the estimates based on the model of Eq. (6). Based on these
cult because the proportionality constant corresponding to Eq. data points, Fig. 3(b) illustrates the calculation of the constant bias
(10) is usually analytically intractable. Conventional Monte Carlo b0 in measurements.
(MC) simulations require the full expression of the posterior den- The updating is performed for a set of values of re (2.5%, 5.0%,
sity to draw samples from it. To circumvent this situation, a Mar- 10.0% and 20.0% of W f ). Fig. 4 shows the effect of the adopted re
kov chain Monte Carlo (MCMC) simulation technique using the value on W s ðtÞ for an instrument with b0 ¼ 1:42% and b1 ¼ 1:0.
Metropolis-Hastings (MH) algorithm is adopted here. MH algo- The enhancement in prediction of steel loss over time is clearly vis-
rithm allows samples to be drawn from the posterior by making ible in the plots of the updated model when compared to the pre-
use of a proposal density function. Considering this, the proposal diction based on the existing model (Eq. (6)), against the measured
density chosen in this paper is a bivariate normal distribution: data set. In addition to the expected value of the probabilistic steel
" #! loss model used here, these plots also show the 95% confidence
r21 0 interval on the updated prediction of W s ðtÞ. This confidence inter-
qðH00 j H0 Þ  N H0 ; ð18Þ
0 r22 val typically envelops the measured data points. As expected, an
increase in re widens the 95% confidence interval. Besides, this
where H0 and H00 are the subsequent samples on the Markov chain. increase changes the mean (updated) estimation as well. Further,
The values of r21 and r22 are selected such that the acceptance rate effects of the measurement bias are studied by considering differ-
lies between 0.23 and 0.45 to ensure sufficient mixing of a Markov ent b1 values as noted in Table 4. Table 4 presents the change in the
chain [50], and are kept constant throughout the simulation. To parameters for updated distributions of Aw and asr . Fig. 5 shows
reduce the auto-correlation between these ‘dependent’ samples, how b1 changes the results for an assumed re ¼ 2:5%. These plots
the so-called ‘thinning’ technique is employed in this work, where show that a change in the scaling bias changes the expected value
– post burn-in – samples are stored only after a ‘lag’ of 500 itera- of W s more than it changes its confidence interval.
tions. A total of 5000 samples are used after burn-in. The results of this Bayesian updating of the time-varying steel
00
Samples from the posterior density function, f ðH j XÞ are then loss model makes a definite case for using such an approach in cor-
used to estimate various functions of the variables Aw and asr , such rosion studies. Fig. 4 and Fig. 5 clearly show the improvements that
as W s ðtÞ. This probabilistic characterisation not only provides the can be achieved by using Bayesian updating over the existing cor-
expectation of W s over time, but also its variance, which can be rosion estimation models. The next section extends this study to
used to estimate confidence intervals on the steel loss prediction. demonstrate how this updating influences the prediction of time-
varying reliability of a simple RC beam exposed to corrosive
3.4. Bayesian updating for steel loss over time environments.

The proposed approach requires case-specific field measure-


ments for updating the steel loss model using a Bayesian frame- 4. Reliability analysis
work. For the lack of any such data, experimental data from the
work of Torres-Acosta and Sagues [47] are adopted for the purpose For the purpose of an illustrative reliability analysis, we con-
of demonstration. In their work, the results of an experimental sider a simply-supported RC beam reinforced with three bars on
S.A. Faroz et al. / Engineering Structures 126 (2016) 457–468 463

Fig. 4. Prediction using the updated corrosion model (b0 = 1.42%, b1 = 1.0) for (a) re ¼ 2:5%, (b) re ¼ 5:0%, (c) re ¼ 10:0%, and (d) re ¼ 20:0%.

Table 4 The beam is designed as an under-reinforced section so that flexural


Measurement biases considered and their effects on updated Aw and asr (re ¼ 2:5%). failure occurs through the yielding of steel at any time instant. Con-
b0 (% of b1 Updated Aw Updated asr sidering a slab thickness of 150 mm with a uniform floor finish of
W0) 1.0 kN/m2 [52] and a uniform live load of 2.5 kN/m2 [53], we arrive
Sample Sample Sample Sample at a 230 mm by 530 mm beam (d = 500 mm), for the load case of
mean SDa mean SD
1:5DL þ 1:5LL, DL being the dead load and LL the live load. It has
1.42 0.75 12.76 3.109 0.6991 0.06568 a single layer of three 20 mm bars, with f ck = 25 MPa and
1.42 0.875 14.41 3.581 0.7001 0.06636
1.42 1.0 15.94 3.698 0.7044 0.06259
f yk = 415 MPa.
1.42 1.125 17.65 3.919 0.7064 0.06137 Reliability analysis of this beam is performed considering both
1.42 1.25 19.19 3.917 0.7125 0.05711 uniform and pitting corrosion of the rebars. It is assumed that
a
SD = standard deviation. although shear reinforcements get corroded along with flexural
reinforcements, flexural failure always precedes shear failure. Cor-
rosion in reinforced concrete also results in the loss of bond
between steel and concrete, which is ignored in our analyses along
the tension side only, (Fig. 6) for the limit state of collapse due to with the effect of corrosion on the yield strength and ductility of
flexure. This beam is assumed to carry uniformly distributed dead reinforcing steel. For the purpose of reliability analysis, in the limit
and (arbitrary point-in-time) live loads from a floor of dimension state function, the time-varying moment capacity of the beam is
4.65 m  4.65 m. The beam is designed following the Indian Stan- calculated by removing the factors 1.15 for f yk and 1.5 for f ck in
dard IS:456 [51] for a design moment the design equation (Eq. (19)):
 
f yk f yk nb A0  
Md ¼ nb A0 d  1:005 ð19Þ f y Ast ðtÞ
1:15 bfck Mr ðtÞ ¼ f y Ast ðtÞ d  0:771 ð20Þ
bfc
where f yk is the characteristic yield strength of the rebar, f ck is the
characteristic compressive strength of the concrete, nb is the num- where f y and f c are the random variables denoting yield strength of
ber of tension rebars, A0 is the (un-corroded, at t ¼ 0) area of a ten- rebar and compressive strength of concrete, respectively. Ast ðtÞ is
sion rebar, b is the width and d is the effective depth of the beam. the residual cross-sectional area of rebars at time t. The material
464 S.A. Faroz et al. / Engineering Structures 126 (2016) 457–468

Fig. 5. Prediction using the updated corrosion model (re ¼ 2:5%) for b0 ¼ 1:42% and (a) b1 ¼ 0:75, (b) b1 ¼ 0:875, (c) b1 ¼ 1:125, and (d) b1 ¼ 1:25.

Fig. 6. Discretisation of a beam.

Table 5 at any given cross-section, steel loss in individual rebars is the same
Statistics of material and geometrical parameters. (that is, full correlation in W s for any two rebars). Also, modelling
Load Distribution Bias CoVa Ref.
error in the design equation is neglected for the lack of any informa-
tion thereof.
fy Lognormal 1.113 0.09 [40]
Traditionally, corrosion reliability analyses have been based on
fc Normal 1.264 0.1266 [60]
b Normal 1 0.02 [61]
the assumption of uniform corrosion [7]. This indirectly implies
d Normal 1 0.02 [61] that the critical section is that at which the highest external bend-
A0 Normal 1 0.04 [61] ing moment acts because of the spatial invariance in uniform cor-
a
CoV = coefficient of variation.
rosion. Pitting corrosion requires a spatial analysis as discussed in
Section 2.1. For this purpose, the beam of length L is divided into
k ¼ L=D elements of equal length D (Fig. 6). The statistics of the
(f y and f c ) and geometrical (b; d and A0 ) variations adopted in the spatial area heterogeneity factor (R) proposed by Zhang et al.
present work are given in Table 5. For simplicity, it is assumed that, [37], adopted for modelling the time-varying pitting corrosion in
S.A. Faroz et al. / Engineering Structures 126 (2016) 457–468 465

this study, remain the same for each of these elements and each re ¼ 2:5%; b0 ¼ 1:42%; b1 ¼ 1:0. These values are used for the pur-
bar. pose of illustration only, and any appropriate value may be used.
The factor Rj ðtÞ for the jth element at time t is considered to be Fig. 7(a) presents the histogram of posterior samples of asr ,
fully correlated for all rebars, and the minimum area is assumed to which shows that this random variable has a very skewed distribu-
occur at the centre of each element. Assuming Aavg to be the same tion. The Kolmogorov-Smirnov (K-S) goodness-of-fit test fails for
for all elements at any time t, the residual steel area for an element most standard probability models (normal, lognormal, gamma,
comprising of nb rebars is calculated using Eqs. (7)–(9): exponential, extreme types, etc.). A non-parametric kernel density
h i estimator [56] can be useful to obtain a smooth PDF for such vari-
nb A0 1  WWs ðtÞ ables. A kernel density usually consists of a weighing function xð:Þ
Astj ðtÞ ¼ ð21Þ
0

Rj ðtÞ with a certain bandwidth (h) which defines its smoothness. The
PDF is generally expressed as
The moment capacity of the jth element Mrj ðtÞ is calculated using
X
n
Eq. (20). As suggested by Zhang et al. [37], an element length of ^f ðHÞ ¼ 1 xðH  Hi ; hÞ ð25Þ
D ¼ 150 mm is used for all pitting corrosion analyses in our study, n i¼1
for which the statistical characteristics of Rj ðtÞ are scaled from the
where Hi ; i ¼ 1; . . . ; n are the sampled values. The weighing function
base element size L0 ¼ 50 mm to D, as m ¼ m0 þ j0 lnðD=L0 Þ and is taken in the form of a Gaussian kernel Kð:Þ for modelling asr :
j ¼ j0 . For uniform corrosion, the reliability analyses are per-
!
formed only at the mid-span of the beam with Rj ðtÞ ¼ 1:0. 1 h 1 h2
xðh; hÞ ¼ K ¼ pffiffiffiffiffiffiffi exp  2 ð26Þ
The applied moment due to uniformly distributed loads are cal- h h h 2p 2h
culated at the centre of each element
With a bandwidth of 100, the resulting CDF for asr is plotted along
wDL þ wLL  
M qj ¼ Lxj  x2j ð22Þ with the empirical one in Fig. 7(b), which shows that the kernel
2 density estimator provides a very good fit. For Aw , the best proba-
where wDL and wLL are the (unfactored) dead and live loads per unit bilistic model based on the K-S test comes out to be a generalised
length of the beam and xj is the location of the centre of the jth ele- extreme value (GEV) distribution with the following parameters:
ment. The loads are considered to be random variables. In the shape = 0.4035, scale = 1.622, and location = 14.13. The posterior
absence of any probabilistic model recommended for loads in the histogram of Aw along with the fitting of its empirical CDF using this
relevant Indian Standards, we adopt the models suggested by SriVi- GEV model is shown in Fig. 8. These kernel density and GEV models
dya and Ranganathan [54]. These load statistics are presented in for asr and Aw , respectively, are used in the reliability analysis. The
Table 6. posterior (paired) samples of asr and Aw also show a linear correla-
The limit state function for collapse against flexure for the jth tion of q ¼ 0:9589, which is included in the reliability analysis
element can now be expressed as along with the fitted CDFs of these two variables.

g j ðtÞ ¼ Mrj ðtÞ  Mqj ðtÞ ð23Þ


4.2. Results of the reliability analysis
with a time-varying probability of failure for the beam, treating it as
a series system: Based on a Rackwitz-Fiessler reliability analysis, the initial reli-
ability of the beam, before the initiation of any corrosion, is evalu-
Pf ðtÞ ¼ Pðg j ðtÞ min8j < 0Þ ð24Þ ated as bð0Þ ¼ 6:272, which is high for a civil structural component.
This high reliability can primarily be associated with the low value
The reliability of the beam bðtÞ is obtained using a first-order of bias (mean to nominal ratio) in the live load model adopted.
reliability analysis following the Rackwitz-Fiessler algorithm [55], Also, the standards used for designing the beam are not based on
for both pitting and uniform corrosion cases. For both cases, the rigorous reliability methods, which may result in very conservative
time-varying reliability analysis is conducted using the existing designs, in general (as per Ranganathan [57], a reasonable value is
(deterministic) model for steel loss (Eq. (6)) and the model based in the range of 4.3–5.5 for beams designed as per this equation,
on Bayesian updating. The reliability of the beam is calculated governed by flexure). However, the goal of this study is not to
every five years over a typical design life of 50 years. obtain the actual reliability of a general beam, but to see the effects
of using a Bayesian updated corrosion model instead of the existing
4.1. Updated models for asr and Aw empirical model on the time-varying reliability of the beam. This
high reliability also argues against the use of (Monte Carlo or
The Rackwitz-Fiessler analysis following the first-order reliabil- other) simulation-based analysis, which would require an inordi-
ity method (FORM) requires all random variables to be modelled nately large number of simulations to arrive at a reasonable esti-
with continuous distribution functions. As discussed in Section 3.3, mate of the probability of failure.
Bayesian updating for steel loss provides sufficient samples (5000, For a parametric study, the beam is subjected to two different
in this case) from the posterior distributions of asr and Aw . These corrosion rates, (a) icorr = 1.0 lA/cm2 signifying low-to-moderate
two parameters need to be modelled with continuous CDFs, which corrosion and (b) icorr = 10.0 lA/cm2 signifying severe corrosion.
can be used for a statistical characterisation of W s ðtÞ used in Eq. Results of the reliability analyses are presented in graphical form
(21). For these characterisations and for subsequent reliability in Figs. 9 and 10. These plots present how the reliability of the
analyses (using the Bayesian updated model of W s ) the following beam reduces over a design life of 50 years due to corrosion of
set of values are adopted for measurement errors and biases: rebars (both uniform and pitting). For each case, these plots pro-
vide a comparison between results based on the existing (deter-
ministic) model and the Bayesian updated (probabilistic) model
Table 6
Statistics of load parameters. for steel loss.
As it can be expected, the pitting corrosion model results in a
Load Distribution Bias CoV
lower reliability of the beam compared to the uniform corrosion
Dead load (DL) Normal 1.050 0.10 model at any instant of time. This remains true for both the exist-
Live load (LL) Lognormal 0.179 0.52
ing and updated models. The updated model (which is deemed
466 S.A. Faroz et al. / Engineering Structures 126 (2016) 457–468

Fig. 7. (a) Histogram of updated asr ; (b) empirical and fitted CDFs of updated asr .

Fig. 8. (a) Histogram of updated Aw ; (b) empirical and fitted CDFs of updated Aw .

Fig. 9. Time-varying reliability of the beam for icorr = 1.0 lA/cm2. Fig. 10. Time-varying reliability of the beam for icorr = 10.0 lA/cm2.

more realistic compared to the existing model, based on the work updated models becomes more prominent for the severe corrosion
presented earlier in Section 3) shows that the actual effect of cor- rate of 10.0 lA/cm2. The probability of failure is estimated as
rosion is significantly more than that predicted by the existing
Pf ðtÞ  U½bðtÞ ð27Þ
model of steel loss. The difference between the existing and the
S.A. Faroz et al. / Engineering Structures 126 (2016) 457–468 467

For the case of pitting corrosion, P f after 50 years is 8:652  109 [6] Almusallam AA, Al-Gahtani AS, Aziz AR, Rasheeduzzafar. Effect of
reinforcement corrosion on bond strength. Constr Build Mater 1996;10
based on the existing model and 1:100  103 based on the updated (2):123–9.
model (for icorr = 10.0 lA/cm2). This significant difference between [7] Stewart MG. Spatial variability of pitting corrosion and its influence on
structural fragility and reliability of RC beams in flexure. Struct Saf 2004;26
the two models clearly justifies the use of Bayesian updating in cor- (4):453–70.
rosion reliability analysis. Similarly, the results highlight the effect [8] Stewart MG, Al-Harthy A. Pitting corrosion and structural reliability of
of pitting, as opposed to uniform corrosion, which manifests in a corroding RC structures: experimental data and probabilistic analysis. Reliab
Eng Syst Saf 2008;93(3):373–82.
significant increase in P f (from 4:792  107 to 1:100  103 ) at [9] Faroz SA, Pujari NN, Ghosh S. A Bayesian Markov Chain Monte Carlo approach
the end of design life in case of a severe corrosion. for the estimation of corrosion in reinforced concrete structures. In: Topping
BHV, Iványi P, editors. Proceedings of the twelfth international conference on
computational structures technology. Stirlingshire, UK: Civil-Comp Press;
5. Concluding remarks 2014. http://dx.doi.org/10.4203/ccp.106.150 [paper 150].
[10] Liu Y. Modeling the time to corrosion cracking of the cover concrete in chloride
contaminated reinforced concrete structures PhD thesis. Blacksburg,
Corrosion of rebars is the primary durability concern in rein- USA: Virginia Polytechnic Institute and State University; 1996.
forced concrete structures. Considering that these structures con- [11] Chung L, Paik IK, Cho SH, Roh YS. Infrared thermographic technique to
stitute a large share of civil structures across all countries, and measure corrosion in reinforcing bar. Key Eng Mater 2006;321–323:821–4.
[12] Kobayashi K, Banthia N. Corrosion detection in reinforced concrete using
that many are exposed to polluted and corrosive environments, induction heating and infrared thermography. J Civil Struct Health Monitor
the issue of corrosion needs more attention from academic 2011;1(1–2):25–35.
research in conjunction with maintenance/management processes. [13] Baek S, Xue W, Feng MQ, Kwon S. Nondestructive corrosion detection in RC
through integrated heat induction and ir thermography. J Nondestruct Eval
From a structural engineering point of view, loss of rebar area is 2012;31(2):181–90.
the most significant aspect of corrosion because it may eventually [14] Michel A, Pease BJ, Geiker MR, Stang H, Olesen JF. Monitoring reinforcement
lead to the possible collapse of a structure. The present study pro- corrosion and corrosion-induced cracking using non-destructive X-ray
attenuation measurements. Cem Concr Res 2011;41(11):1085–94.
poses a novel approach of Bayesian updating of existing steel loss [15] Akiyama M, Frangopol DM. Estimation of steel weight loss due to corrosion in
model based on monitored data. This approach, while making RC members based on digital image processing of X-ray photogram. In:
use of typical structural health monitoring/NDT outcomes, pro- Proceedings of the 3rd international symposium on life-cycle civil engineering
(IALCCE 2012), Vienna, Austria. p. 1885–91.
vides significantly better estimates of steel loss due to corrosion,
[16] Itty PA, Serdar M, Meral C, Parkinson D, MacDowell AA, Bjegović D, et al. In situ
compared to the existing empirical model. It should also be noted 3D monitoring of corrosion on carbon steel and ferritic stainless steel
that the Bayesian updating presented here did not require an unre- embedded in cement paste. Corros Sci 2014;83:409–18.
alistically large number of measurements to improve the estima- [17] Ahmad S, Bhattacharjee B. A simple arrangement and procedure for in-situ
measurement of corrosion rate of rebar embedded in concrete. Corros Sci
tions significantly. These can be ‘‘regular” corrosion monitoring 1995;37(5):781–91.
based on a prescriptive scheme, as recommended for concrete [18] Alghamdi SA, Ahmad S. Service life prediction of RC structures based on
bridges by the Indian Roads Congress [58], or measurements based correlation between electrochemical and gravimetric reinforcement corrosion
rates. Cem Concr Compos 2014;47:64–8.
on an optimum maintenance scheme, as suggested by Kim et al. [19] Frangopol DM, Lin KY, Estes AC. Life-cycle cost design of deteriorating
[59]. A time-varying reliability analysis of a sample RC beam over structures. ASCE J Struct Eng 1997;123(10):1390–401.
its design life clearly shows the difference between predictions [20] Ang AHS, Tang WH. Probability concepts in engineering. 2nd ed. New York,
USA: John Wiley & Sons; 2007.
based on the existing and updated models, which highlights the [21] Rafiq MI, Chryssanthopoulos MK, Onoufriou T. Performance updating of
importance of using an updated model. The Bayesian approach concrete bridges using proactive health monitoring methods. Reliab Eng Syst
presented here also explains how biases in instruments and mea- Saf 2004;86(3):247–56.
[22] Keßler S, Fischer J, Straub D, Gehlen C. Updating of service-life prediction of
surement/modelling errors can be suitably incorporated in the reinforced concrete structures with potential mapping. Cem Concr Compos
updated estimates. This work shows how both pitting and uniform 2014;47:47–52.
corrosion models can utilise the Bayesian approach, and how the [23] Samarakoon SMSMK, Sælensminde J. Condition assessment of reinforced
concrete structures subject to chloride ingress: a case study of updating the
assumptions of uniform corrosion may significantly underestimate
model prediction considering inspection data. Cem Concr Compos
the probability of failure. 2015;60:92–8.
The proposed framework of corrosion reliability analysis can be [24] Otieno MB, Beushausen HD, Alexander MG. Modelling corrosion propagation
further extended to include more complex corrosion models, for in reinforced concrete structures – A critical review. Cem Concr Compos
2011;33(2):240–5.
example (a) different corrosion for individual rebars, (b) corrosion [25] Suo Q, Stewart MG. Corrosion cracking prediction updating of deteriorating RC
of shear reinforcements along with flexural rebars, (c) non-uniform structures using inspection information. Reliab Eng Syst Saf 2009;94
loading, etc. (8):1340–8.
[26] Akiyama M, Frangopol DM, Yoshida I. Time-dependent reliability analysis of
existing RC structures in a marine environment using hazard associated with
Acknowledgement airborne chlorides. Eng Struct 2010;32(11):3768–79.
[27] Zheng R, Ellingwood BR. Role of non-destructive evaluation in time-dependent
reliability analysis. Struct Saf 1998;20(4):325–39.
The authors would like to thank the three anonymous review- [28] Malumbela G, Moyo P, Alexander M. A step towards standardising accelerated
ers whose critical comments and useful suggestions have been corrosion tests on laboratory reinforced concrete specimens. J SA Inst Civil Eng
valuable in enhancing the quality of the paper. 2012;54(2):78–85.
[29] Liu Y, Weyers RE. Modeling the time-to-corrosion cracking in chloride
contaminated reinforced concrete structures. ACI Mater J 1998;95(6):
References 675–81.
[30] Benarie M, Lipfert FL. A general corrosion function in terms of atmospheric
pollutant concentrations and rain pH. Atmos Environ 1986;20(10):1947–58.
[1] Chen D, Mahadevan S. Chloride-induced reinforcement corrosion and concrete
[31] Melchers RE. Mathematical modelling of the diffusion controlled phase in
cracking simulation. Cem Concr Compos 2008;30(3):227–38.
marine immersion corrosion of mild steel. Corros Sci 2003;45(5):923–40.
[2] Tuutti K. Corrosion of steel in concrete. Tech rep 4-82. Stockholm, Sweden:
[32] Bhargava K, Ghosh AK, Mori Y, Ramanujam S. Analytical model for time to
Swedish Cement and Concrete Research Institute; 1982.
cover cracking in RC structures due to rebar corrosion. Nucl Eng Des 2006;236
[3] Li CQ, Zheng JJ, Lawanwisut W, Melchers RE. Concrete delamination caused by
(11):1123–39.
steel reinforcement corrosion. ASCE J Mater Civil Eng 2007;19(7):591–600.
[33] González JA, Andrade C, Alonso C, Feliu S. Comparison of rates of general
[4] Hariche L, Ballim Y, Bouhicha M, Kenai S. Effects of reinforcement
corrosion and maximum pitting penetration on concrete embedded steel
configuration and sustained load on the behaviour of reinforced concrete
reinforcement. Cem Concr Res 1995;25(2):257–64.
beams affected by reinforcing steel corrosion. Cem Concr Compos 2012;34
[34] DuraCrete. Statistical quantification of the variables in the limit state
(10):1202–9.
functions. The European Union – Brite EuRam III, Project No. BE95-1347; 2000.
[5] Alonso C, Andrade C, Rodriguez J, Diez JM. Factors controlling cracking of
[35] Val DV, Melchers RE. Reliability of deteriorating RC slab bridges. ASCE J Struct
concrete affected by reinforcement corrosion. Mater Struct 1996;31
Eng 1997;123(12):1638–44.
(211):435–41.
468 S.A. Faroz et al. / Engineering Structures 126 (2016) 457–468

[36] Darmawan MS. Pitting corrosion model for reinforced concrete structures in a [48] Karandikar JM, Kim NH, Schmitz TL. Prediction of remaining useful life for
chloride environment. Magaz Concr Res 2010;62(2):91–101. fatigue-damaged structures using Bayesian inference. Eng Fract Mech
[37] Zhang W, Zhou B, Gu X, Dai H. Probability distribution model for cross- 2012;96:588–605.
sectional area of corroded reinforcing steel bars. ASCE J Mater Civil Eng [49] Beck JL, Katafygiotis LS. Updating models and their uncertainties. I: Bayesian
2014;26(5):822–32. statistical framework. ASCE J Eng Mech 1998;124(4):455–61.
[38] Kashani MM, Crewe AJ, Alexander NA. Use of a 3D optical measurement [50] Chib S, Greenberg E. Understanding Metropolis-Hastings algorithm. Am Stat
technique for stochastic corrosion pattern analysis of reinforcing bars 1995;49(4):327–35.
subjected to accelerated corrosion. Corros Sci 2013;73:208–21. [51] BIS. IS 456 Indian standard for plain and reinforced concrete – Code of
[39] Tang F, Lin Z, Chen G, Yi W. Three-dimensional corrosion pit measurement and practice. New Delhi, India: Bureau of Indian Standards; 2000.
statistical mechanical degradation analysis of deformed steel bars subjected to [52] BIS. IS 875 Code of practice for design loads for buildings and structures, Part
accelerated corrosion. Constr Build Mater 2014;70:104–17. 1: dead loads. New Delhi, India: Bureau of Indian Standards; 1987.
[40] Pujari NN, Ghosh S, Lala S. Bayesian approach for the seismic fragility [53] BIS. IS 875 Code of practice for design loads for buildings and structures, Part
estimation of a containment shell based on the formation of through-wall 2: imposed loads. New Delhi, India: Bureau of Indian Standards; 1987.
cracks. ASCE-ASME J Risk Uncertain Eng Syst, Part A: Civil Eng 2015. http://dx. [54] SriVidya A, Ranganathan R. Reliability based optimal design of reinforced
doi.org/10.1016/AJRUA6.0000840. concrete frames. Comput Struct 1995;57(4):651–61.
[41] Rastogi R, Ghosh S, Ghosh AK, Vaze KK, Singh PK. Fatigue crack growth [55] Nowak AS, Collins KR. Reliability of structures. New York, USA: McGraw Hill;
prediction in nuclear piping using Markov chain Monte Carlo simulation. 2000.
Fatigue Fract Eng Mater Struct 2016. http://dx.doi.org/10.1111/ffe.12486. [56] Härdle W, Müller M, Sperlich S, Werwatz A. Nonparametric and
[42] Jamali A, Angst U, Adey B, Elsener B. Modeling of corrosion-induced concrete semiparametric models. Berlin, Germany: Springer; 2004.
cover cracking: a critical analysis. Constr Build Mater 2013;42:225–37. [57] Ranganathan R. Structural reliability analysis and design. Mumbai, India: Jaico
[43] Zhu B, Frangopol DM. Reliability assessment of ship structures using Bayesian Publishing House; 1999.
updating. Eng Struct 2013;56:1836–47. [58] IRC. IRC:SP:60-2002 An approach document for assessment of remaining life
[44] Song HW, Saraswathy V. Corrosion monitoring of reinforced concrete of concrete bridges. Indian Roads Congress, New Delhi, India; 2002.
structures – A review. Int J Electrochem Sci 2007;2(1):1–28. [59] Kim S, Frangopol DM, Zhu B. Probabilistic optimum inspection/repair planning
[45] Simola K, Pulkkinen U. Models for non-destructive inspection data. Reliab Eng to extend lifetime of deteriorating structures. ASCE J Perform Constr Facilit
Syst Saf 1998;60(1):1–12. 2011;25(6):534–44.
[46] Al-Amin M, Zhou W, Zhang S, Kariyawasam S, Wang H. Bayesian model for [60] BIS. IS 10262 Concrete mix proportioning – Guidelines. Bureau of Indian
calibration of ILI tools. Proceedings of the Biennial international pipeline Standards, New Delhi, India; 2009.
conference, IPC, vol. 2. p. 201–8. [61] Lu R, Luo Y, Conte JP. Reliability evaluation of reinforced concrete beams.
[47] Torres-Acosta AA, Sagues AA. Concrete cracking by localized steel corrosion – Struct Saf 1994;14(4):277–98.
Geometric effects. ACI Mater J 2004;101(6):501–7.

You might also like