You are on page 1of 12

Front. Struct. Civ. Eng.

2013, 7(3): 276–287


DOI 10.1007/s11709-013-0207-9

RESEARCH ARTICLE

Inverse Gaussian process-based corrosion growth


modeling and its application in the reliability analysis for
energy pipelines
Hao QIN, Shenwei ZHANG, Wenxing ZHOU*

Department of Civil and Environmental Engineering, Western University, London N6A 5B9, Canada
*
Corresponding author. E-mail: wzhou@eng.uwo.ca

© Higher Education Press and Springer-Verlag Berlin Heidelberg 2013


ABSTRACT This paper describes an inverse Gaussian process-based model to characterize the growth of metal-loss
corrosion defects on energy pipelines. The model parameters are evaluated using the Bayesian methodology by
combining the inspection data obtained from multiple inspections with the prior distributions. The Markov Chain Monte
Carlo (MCMC) simulation techniques are employed to numerically evaluate the posterior marginal distribution of each
individual parameter. The measurement errors associated with the ILI tools are considered in the Bayesian inference. The
application of the growth model is illustrated using an example involving real inspection data collected from an in-service
pipeline in Alberta, Canada. The results indicate that the model in general can predict the growth of corrosion defects
reasonably well. Parametric analyses associated with the growth model as well as reliability assessment of the pipeline
based on the growth model are also included in the example. The proposed model can be used to facilitate the
development and application of reliability-based pipeline corrosion management.

KEYWORDS pipeline, metal-loss corrosion, inverse Gaussian process, measurement error, hierarchical Bayesian, Markov
Chain Monte Carlo (MCMC)

1 Introduction corrosion growth modeling [1–3]. For example, Hong [1]


proposed a non-homogenous Markov process to model the
Metal-loss corrosion is a major threat to the structural corrosion growth based on data obtained from corrosion
integrity of steel energy pipelines world-wide, the majority experiments. Maes et al. [2] and Zhang et al. [3] reported
of which are highly pressurized underground oil and non-homogenous Gamma process- (NHGP-) based models
natural gas pipelines. The reliability-based corrosion formulated in the hierarchical Bayesian framework to
management programs have been increasingly employed characterize the growth of defects on pipelines based on
by the pipeline operators. The corrosion growth modeling data obtained from multiple inline inspections (ILI). The
plays a key role in the various tasks involved in the study in Ref. [3] differs from that in Ref. [2] as the former
pipeline corrosion management program such as determi- considers the initiation time of each individual defects and
nation of the inspection interval, evaluation of the failure incorporates a general form of the measurement errors
probability and development of phased mitigation actions associated with ILI data, which includes the biases and
to satisfy the safety constraint and optimize the allocation random scattering error as well as the correlations between
of limited resources. the random scattering errors associated with the ILI tools
The corrosion process is subjected to significant used in different inspections.
uncertainties of both temporal and spatial variability. The A study reported in the literature [4] suggests that the
Markov process and Gamma process are two stochastic inverse Gaussian process (IGP) is a viable alternative to
process models that have been commonly applied in characterize the corrosion process. The mathematical
tractability of IGP [4] also facilitates incorporating the
Article history: Received Feb. 21, 2013; Accepted Jun. 8, 2013 IGP-based corrosion growth model in a Bayesian frame-
Hao QIN et al. Inverse Gaussian process-based corrosion growth modeling 277

work to evaluate and update the model parameters based Based on Eq. (1), the mean and variance of X(t), denoted
on the corrosion inspection data. However, reports on the by E[X(t)] and Var[X(t)], respectively, are given by
IGP-based models to characterize the growth of corrosion
defects on pipelines are scarce in the literature. E½X ðtÞ ¼ ΘðtÞ, (3a)
In this study, we aim to develop an IGP-based growth
model for the depths of corrosion defects on energy Var½X ðtÞ ¼ ΘðtÞ=: (3b)
pipelines by incorporating the ILI-reported depths collected
from multiple ILI runs. The growth model is formulated in 3 Inspection data and growth models for
a hierarchical Bayesian framework to deal with uncertain- multiple defects
ties in the growth model and inspection data. The Markov
Chain Monte Carlo (MCMC) simulation techniques are The inspection data reported by the ILI tools are always
employed to carry out the Bayesian inference and associated with certain sizing errors including biases and
numerically evaluate the probabilistic characteristics of random scattering error that will impact the estimation of
the model parameters. The proposed growth model is model parameters. Suppose m active corrosion defects on a
illustrated and validated through an example involving a set particular pipe segment have been inspected and sized n
of real ILI data collected from a natural gas pipeline that is times over a certain period of time. The measured depth of
currently in service in Alberta, Canada. The use of the the ith defect at the jth inspection, yij, (i = 1, 2, …, m; j = 1, 2,
developed growth model for evaluating the time-dependent
…, n) can be related to the corresponding actual depth, xij,
system reliability of a joint of the above mentioned pipeline
through the following equation [6,7]:
is also included in the example.
yij ¼ aj þ bj xij þ εij , (4)

2 Inverse Gaussian process where aj and bj denote the constant and non-constant bias
associated with the ILI tool used in the jth inspection, and εij
Consider that X is a random variable and follows an inverse denotes the random scattering error associated with the
Gaussian distribution with the probability density function ILI-reported depth of the ith defect at the jth inspection, and
(PDF), fX(x|μ,θ), parameterized as [5] is assumed to be normally distributed with a zero mean and
rffiffiffiffiffiffi known standard deviation [8]. It is further assumed that for
 
 – 2=3 ðx – Þ2 a given inspection j, εij and εlj (i≠l) are independent,
fX ðxj,Þ ¼ x exp – Ið0,1Þ ðxÞ, (1) whereas for a given defect i, εij and εik (j≠k) are correlated
2π 22 x
with a correlation coefficient of ρjk (i.e., the correlation
with μ (μ > 0) and θ (θ > 0) denoting the mean and shape coefficient between the random scattering errors associated
parameters of X, respectively, and I(0,1)(x) denoting an with the jth and kth inspections) [8]. Denote Ei = (Ei1, Ei2 …
indicator function that equals unity for x > 0 and zero Ein)′ as the vector of random scattering errors associated
otherwise. The mean and variance of X are μ and μ3/θ, with the n inspections for defect i, with “′” representing
respectively. transposition. It follows from the above assumption
Let {X(t); t≥0} denote an inverse Gaussian process that Ei is multivariate normal-distributed and has a PDF
(IGP) with a PDF given by [4] given by
   2   
 1 X–1
fEi ðεi Þ ¼ pffiffiffiffiffi X 1=2 exp – εí
1
fX ðtÞ xðtÞΘðtÞ, ΘðtÞ ε , (5)
n=2   2 i Ei i
ð 2πÞ  E 
0  2 1 i

 – ΘðtÞ P
pffiffiffiffiffiffiffiffiffiffi B xðtÞ C where Ei denotes the n  n variance-covariance matrix
¼ =2πΘðtÞxðtÞ – 3=2 exp @ – A
2xðtÞ of Ei with each of elements equal to ρjkσjσk (j = 1, 2, …, n; k
= 1, 2, …, n), and σj denotes the standard deviation of the
  random scattering error associated with the tool used at the
 Ið0,1Þ xðtÞ , (2) jth inspection. A Bayesian methodology has been reported
in Ref. [8] to evaluate aj, bj, σj and ρjk by comparing the
where Θ(t) and (Θ(t))2 denote the mean and shape ILI-reported and corresponding field-measured depths for
functions, respectively, with  denoting a scale parameter. a set of static (i.e., repaired) corrosion defects.
The IGP has three properties: 1) X(t) equals zero at t = 0 The growth of the depth associated with each individual
with probability one; 2) X(t) has independent increments; defect is assumed to follow an inverse Gaussian process
and 3) for all t2 > t1≥0, X(t2) – X(t1) follows an inverse given by Eq. (2), where Θ(t) is assumed to be a power-law
Gaussian distribution with a PDF given by Eq. (1), where function of time, i.e., Θ(t) = α(t – t0)δ with t0 denoting the
the mean and shape parameters equal Θ(t2) – Θ(t1) and initiation time of the corrosion defect, α denoting the
(Θ(t2) – Θ(t1))2, respectively. average growth over the first unit time interval since t0, and
278 Front. Struct. Civ. Eng. 2013, 7(3): 276–287

δ indicating the trend of the mean growth path. The mean where xi = (xi1, xi2, …, xij, …, xin)′ with the jth element, xij =
Pj
growth (i.e., Θ(t)) is an accelerating, decelerating and k¼1 Δxik ; a = (a1, a2, …, aj, …, an)′, and b is an n  n
linear growth path for δ > 1, δ < 1 and δ = 1, respectively. diagonal matrix with the jth element equal to bj.
In this study, the four parameters of the IGP-based
growth model, namely α, t0, δ and , were assumed to be
4.2.2 Likelihood functions of the growth of depth
uncertain parameters. Furthermore, we assumed α and t0 to
be defect-specific, and δ and  to be common for all
Assume that Δxij and Δxlj (i ≠ l) are conditionally
defects. It then follows from Eq. (2) that the growth of the
independent of each other for inspection j; in other words,
ith defect between the (j – 1)th and jth inspections, ΔXij, is
the exchangeability condition [9] is applicable to Δxij (i =
inverse Gaussian distributed and has a PDF of
1, 2,…, m) conditional on αi, δ,  and ti0 for inspection j.
fΔXij(Δxij|ΔΘij, (ΔΘij)2). The mean value of ΔXij, ΔΘij,
The joint PDF of Δxi can be written as
equals αi(tj – ti0)δ for j = 1 and αi(tj – ti0)δ – αi(tj-1 – ti0)δ for j 
= 2, 3, …, n, where αi denotes the average growth of the ith 
fΔXi ðΔxi ai ,δ,,t0i Þ ¼ Πj¼1 fΔXij ðΔxij jai ,δ,,t0i Þ
n
defect over the first unit time interval since ti0; tj denotes
the time of the jth inspection, and ti0 denotes the initiation pffiffiffiffiffiffiffiffiffiffi
¼ Πj¼1
time of the ith defect. The actual depth of the ith defect at the n – 3=2
=2πΔΘij Δxij
time of the jth inspection, xij, is then obtained from xij = xi,j-1
!
+ Δxij, with xi0 assumed to be zero. ðΔxij – ΔΘij Þ2
 exp – : (7)
2Δxij
4 Bayesian updating of the growth model
4.3 Prior distribution
4.1 Overview the Bayesian updating
For m active corrosion defects, the growth model described
The Bayesian updating is an approach to evaluate the in Section 3 involves a total of 2m + 2 basic parameters,
probability distributions of a set of unknown parameters of namely m defect-specific parameters αi and initiation times
a given model by combining the existing knowledge of ti0 (i = 1, 2,…, m) and two common parameters δ and . In
these parameters with the new information contained in the this study, the gamma distribution was selected as the prior
data [9]. The mechanism for combining the information is distributions of αi, δ and  (i = 1, 2,…, m) based on the
Bayes’ theorem. The existing knowledge of model consideration that the gamma distribution ensures αi and 
parameters is reflected by the prior distribution. The new to be positive quantities and can be conveniently made as a
information in the data is incorporated in the Bayesian non-informative distribution. The prior distribution of the
updating through the so-called likelihood function, and the initiation time ti0was assigned a uniform distribution with a
probability distribution obtained from the updating is lower bound of zero and an upper bound equal to the time
known as the posterior distribution. The likelihood interval between the installation of the pipeline and the first
function, prior and posterior distributions are described detection of the ith defect. The prior distributions of αi and
in the following sections. ti0 associated with all defects were further assumed to be
identical and mutually independent (iid). Details of the
4.2 Likelihood function prior distributions of αi, ti0 (i = 1, 2,…, m),  and δ are
given as follows:
4.2.1 Likelihood function of the inspection data αi e πG ðp,qÞ, (8a)

Consider defect i and denote yi = (yi1, yi2,…, yij, …, yin)′ ti0 e πU ð0,uÞ, (8b)
and Δxi = (Δxi1, Δxi2, …, Δxin)′. It follows from Eqs. (4)
and (5) as well as the assumptions described in Section 3
that the likelihood of the inspection data, yi, conditional on  e πG ðr,sÞ, (8c)
the actual depth increments, Δxi, is given by
X  – 1=2 δ e πG ðv,wÞ, (8d)
 
Lðyi jΔxi Þ ¼ ð2πÞ – n=2  ε 
where πG(∙) and πU(∙) denote the PDF of the gamma and
 X–1  uniform distributions, respectively, and p, q, u, r, s, v, and
1
 exp – ðyi – a – bxi Þ# Ei
ðyi – a – bxi Þ , w are known distribution parameters of the prior distribu-
2 tions. The selection of the values of these parameters is
(6) discussed in Section 5.
Hao QIN et al. Inverse Gaussian process-based corrosion growth modeling 279

4.4 Posterior distribution pipe segment with a length of about 82 km, were
excavated, field measured and mitigated in 2010. The
Let θ denote the unknown model parameters and D denote field-measured depths were assumed to be free of
the inspection data. The joint posterior distribution of measurement errors [8]; that is, the actual depths of the
θ, p(θ|D) can be obtained by combining the joint prior 62 defects in 2010 are known. A summary of the ILI-
distribution of θ, π(θ), with the likelihood, L(D|θ), of the reported and field-measured depths for the defects is given
inspection data according to Bayes theorem [9,10]: in Table 1, where the symbol wt denotes pipe wall
thickness and the symbol %wt denotes the unit of defect
LðDjθÞπðθÞ
pðθjDÞ ¼ depth measured by MFL tools. The apparent growth paths
!LðDjθÞπðθÞdθ indicated by the ILI-reported and field-measured depths for
the 62 defects are illustrated in Fig. 1. Note that the non-
monotonic trend of the growth paths for some defects is
/ LðDjθÞπðθÞ, (9a) attributed to the measurement errors involved in the
where “/” represents proportionality. If the distribution inspection data. The three sets of ILI data obtained in 2000,
parameters (denoted by λ) of the prior distribution of θ are 2004 and 2007 were used to carry out the Bayesian
also considered to be uncertain, the joint posterior updating and evaluate the probabilistic characteristics of
distribution of the parameters θ and λ, denoted by p (θ, the parameters of the growth models for each of the 62
λ|D), can be formulated as the same manner as Eq. (9a) defects. The growth model was then validated by
based on the hierarchical Bayesian theorem, i.e., comparing the predicted depths in 2010 with the
corresponding actual depths.
LðDjθÞπðθjlÞπðlÞ The calibrated biases and random scattering errors
pðθ,ljDÞ ¼
!:::!LðDjθÞπðθjlÞπðlÞdθdl associated with individual ILI tools as well as correlations
between the random scattering errors of different ILI tools
are as follows [8]: a1 = a2 = 2.04 (%wt) and a3 = – 15.28
/ LðDjθÞπðθjlÞπðlÞ: (9b) (%wt); b1 = b2 = 0.97 and b3 = 1.4; σ1 = σ2 = 5.97 (%wt)
and σ3 = 9.05 (%wt); ρ12 = 0.82 and ρ13 = ρ23 = 0.7, with
It is not possible to analytically derive the joint posterior the subscripts “1,” “2” and “3” denoting the parameters
distributions of the parameters of the growth model; associated with the ILI-reported data in 2000, 2004 and
therefore, the MCMC simulation techniques were used to 2007, respectively.
numerically evaluate the posterior distribution. In this The distribution parameters of the prior distributions
study, the software OpenBUGS [11] was employed to carry given by Eq. (8) were selected as follows: p = 1, q = 1, u =
out the MCMC simulation and Bayesian inference of the 28 (year), r = 1, s = 1, v = 1 and w = 1. These values were
probabilistic characteristics of the model parameters. The observed to lead to a good convergence of the MCMC
generic distribution option was used to define the inverse simulation. The impact of the distribution parameters of
Gaussian distribution in OpenBUGS [11]. the prior distributions on the posterior distributions as well
as the predictive accuracy of the growth model is discussed
in the following section.
5 Numerical example A total of 20000 sequences of samples were generated
from the MCMC simulation and the first 3000 sequences
5.1 Model validation of samples were treated as the burn-in period and therefore
discarded. The samples in the remaining sequences were
This section presents the growth models developed for 62 used to make inference of the statistics of the model
external corrosion defects identified on a natural gas parameters. The PDFs of the posterior marginal distribu-
pipeline that is currently in service in Alberta, Canada. tions of  and δ, which are common for all defects, are
This pipeline was constructed in 1972 and has been plotted in Figs. 2(a) and (b), respectively. For comparison,
inspected by the magnetic flux leakage (MFL) tools in the prior distributions are also plotted in the same figure.
2000, 2004 and 2007. The 62 defects, which spread over a For brevity, the posterior distributions of αi and ti0

Table 1 Summary of the ILI-reported and field-measured depths for the 62 defects
ILI-reported depth
field-measured depth in 2010
2000 2004 2007

min (%wt) 3 3 6 12
max (%wt) 56 55 60 65
mean (%wt) 27 28 27 32
280 Front. Struct. Civ. Eng. 2013, 7(3): 276–287

box indicate the 2.5- and 97.5-percentile values of the


corresponding marginal distribution. The bracketed number
at the right end of the arm represents the defect number. The
vertical line indicates the overall mean of this parameter for
all the defects. The overall mean of α and t0 for the 62
defects are 1.2%wt/year and 10.4 years, respectively.
Figures 2 and 3 indicate that most of the marginal
distributions of model parameters are not symmetric and
even highly skewed for certain defects. Therefore, the
median values of the marginal distributions were selected
as the point estimates of model parameters and used to
predict the corrosion growth based on Eq. (2).
The predictive quality of the growth model was
demonstrated by a comparison of the predicted depths in
Fig. 1 Apparent growth paths indicated by the ILI-reported and 2010 with the corresponding field-measured depths for the
field-measured depths 62 defects (see Fig. 4). The predicted depths were obtained
by substituting the median values of αi, ti0 and δ associated
with each of the 62 defects evaluated from the MCMC
simulation into Eq. (3a). Figure 4 suggests that the
proposed model can predict the growth of corrosion
defects reasonably well for majority of the defects
considered in that the predicted depths for 90% of the 62
defects fall between the two bounding lines representing
actual depth 10%wt. The range of 10%wt is commonly
used in the pipeline industry as a confidence interval for the
calibration of the ILI tools and was used in this study as a
metric of the predictive accuracy of the growth model.
The mean, 10- and 90-percentile values of the growth
paths for four arbitrarily selected defects, Defects #13, #15,
#18 and #19, are plotted in Fig. 5(a) through 5(d),
respectively. The mean and standard deviation of the defect
depth at a given time were evaluated using Eqs. (3a) and
(3b), respectively, based on the median values of the model
parameters associated with this particular defect. The 10-
and 90-percentile values were then approximately eval-
uated assuming that the depth at a given time is Gaussian-
distributed according to the central limit theorem. For
comparison, the corresponding ILI-reported depths in
2000, 2004 and 2007 as well as the field-measured depth
in 2010 are also plotted in the same figure. The results in
Fig. 5 indicate that the predicted average growth rate
differs from defect to defect, which is expected because the
parameter α is defect-specific. For example, the average
growth rate of Defect #19 is the highest among the four
defects plotted and equals 1.9%wt/year, followed by
Defects #15, #13 and #18 with average growth rates
Fig. 2 Illustration of the prior and posterior distributions of  and equal to 1.6, 1.3 and 1.1%wt/year, respectively.
δ. (a) The prior and posterior distributions of ; (b) the prior and
posterior distributions of δ 5.2 Parametric analysis

associated with each of the 62 defects are depicted by the To investigate the impact of the assumptions associated
box-plots instead of PDF curves and shown in Figs. 3(a) with the model parameters on the predictions, two
and 3(b), respectively. In Fig. 3, the mid-line, left and right scenarios, designated by Scenarios I and II respectively,
sides of a given box indicate the median, 25- and 75- were further considered. For brevity, the case correspond-
percentile values of the posterior marginal distribution of ing to the results shown in Section 5.1 is referred to as the
each parameter. The left and right ends of the arms of each baseline case, i.e., α and t0 are both defect-specific and δ
Hao QIN et al. Inverse Gaussian process-based corrosion growth modeling 281
282 Front. Struct. Civ. Eng. 2013, 7(3): 276–287

Fig. 3 Box-plots of the posterior distributions of α and t0. (a) Box plot of the posterior distribution α; (b) box plot of the posterior
distribution t0
Hao QIN et al. Inverse Gaussian process-based corrosion growth modeling 283

Fig. 4 Comparison of the predicted depths in 2010 with the corresponding field-measured depths

Fig. 5 Predicted growth path of a given defect. (a) Defect #13; (b) (a) Defect #15; (a) Defect #18; (a) Defect #19
284 Front. Struct. Civ. Eng. 2013, 7(3): 276–287

and  are common for all defects. Scenario I assumes that 5.3 Application in the system reliability assessment
only t0 is defect-specific, whereas scenario II assumes that
only  is common for all the defects. The predicted depths In this section, we apply the developed growth model to
corresponding to Scenarios I and II were obtained in the the system reliability assessment of the corroding gas
same manner as that for the baseline case. The results pipeline considering three distinctive failure modes,
indicate that the percentage of the predictions for the 62 namely small leak, large leak and rupture. Details of the
defects falling within the region of actual depth 10%wt is mechanisms of small leak, large leak and rupture can be
72% for Scenario I and 48% for Scenario II. The fact that found in [12]. The failure of a pipeline segment due to
Scenario I results in a poorer prediction than the baseline small leak, large leak and rupture can be formulated based
case suggests that it is more reasonable to assume α to be on the following three limit state functions [13].
defect-specific (as in the baseline case) than common for The limit state function for a corrosion defect penetrat-
all the defects (as in Scenario I). This is because α ing the pipe wall at a given time t, g1(t), is
represents the mean of the growth within the first unit time g1 ðtÞ ¼ 0:8wt – dðtÞ, (10a)
interval since initiation and is expected to differ from
defect to defect due to the spatial variability of the with wt denoting the wall thickness of the pipeline, and d(t)
corrosion defects. Furthermore, the fact that the predictions denoting the depth of the corrosion defect at time t, i.e., x(t)
corresponding to Scenario II are markedly worse than used in the growth model. The use of 0.8wt as opposed to
those corresponding to the baseline case implies that δ does wt in the above equation is consistent with typical industry
not vary markedly from defect to defect. Therefore, practice, where a remaining ligament thinner than 0.2wt is
assuming δ to be common for all the defects (as in the considered prone to developing cracks that could lead to
baseline case) allows the Bayesian updating to assimilate leaks.
the growth trends associated with all the defects and The limit state function for plastic collapse under
therefore make a reasonable inference of the statistics of δ. internal pressure at the defect at time t, g2(t), is given by
Our investigation indicates that assuming  to be either
defect-specific or common for all defects, in general, has a g2 ðtÞ ¼ rb ðtÞ – pðtÞ, (10b)
negligible impact on the predictions. where rb(t) denotes the burst pressure resistance of the pipe
To investigate the influence of the prior distributions of at the defect at time t and p(t) is the internal pressure of the
model parameters on the predictive quality, we considered pipeline at time t.
two additional scenarios, designated by Scenarios III and The limit state function for the unstable axial extension
IV respectively, both of which are the same as the baseline of the through-wall defect that results from the burst, g3(t),
case except that the shape parameter (i.e., p) of the prior is given by
distribution of αi equals 10 in Scenarios III and 0.1 in
Scenarios IV (as opposed to 1.0 in the baseline case). The g3 ðtÞ ¼ rrp ðtÞ – pðtÞ, (10c)
analysis results indicate that about 21% (61%) of the
where rrp(t) is the pressure resistance of the pipeline at the
predictions of the 62 defects fall within the region of actual
location of the through-wall defect resulting from the burst
depth 10%wt for Scenario III (IV). This is mainly
at time t. Note that we use g1(t), g2(t) and g3(t) to
because p = 10 and 0.1 in conjunction with q = 1 in Eq. (8a)
emphasize that g1, g2 and g3 are all time-dependent as a
implies that the means of the growth in the first unit time
result of the deterioration of pipe resistance over time and
interval since t0 equal 10%wt/year and 0.1%wt/year,
time-dependency of the internal pressure.
respectively. This prior knowledge about the growth rate
Given the above, the probabilities of small leak, large
might be too distant from the realistic scenario to lead to a
leak and rupture within a time interval [0, t], denoted by
good convergence of the MCMC simulation as well as
Psl(t), Pll(t) and Prp(t) respectively, are defined as follows
good estimate of the posterior distributions. The prior
[14]:
distribution of αi in the baseline case (i.e., p = q = 1)
implies that the mean and coefficient of variation (COV) of Psl ðtÞ ¼ Prob½ts <tb \ ts £t, (11a)
the growth in the first unit time interval since t0 equal 1%
wt/year and 100%, respectively, which are considered Pll ðtÞ ¼ Prob½tb <ts \ tb £t \ g3 ðtbÞ > 0, (11b)
reasonably representative of the reality. From this
perspective, the prior distribution of αi specified in the Prp ðtÞ ¼ Prob½tb <ts \ tb £t \ g3 ðtb Þ£0, (11c)
baseline case can be regarded, to certain extent, as an
informative distribution. The comparison between the where Prob[$] denotes the probability of an event; ts and tb
baseline case and Scenarios III and IV highlights the denote the times at which the defect depth just reaches
importance of properly selecting the prior distributions for 0.8wt and p(t) outcrosses rb (t) for the first time,
the Bayesian updating. respectively, and “∩” represents the intersection (i.e.,
Hao QIN et al. Inverse Gaussian process-based corrosion growth modeling 285

joint event). Note that the probability of small leak L(t) is the defect length (i.e., in the longitudinal direction of
occurring first then followed by a burst is ignored in the the pipeline) at time t and ζ is the model error.
analysis. The probabilities given by Eq. (11) were The flow stress-dependent failure criterion for a through-
evaluated through the Monte Carlo (MC) simulation wall flaw developed by Kiefner et al. [16] was employed in
technique following the procedures described in [13]. this study to calculate rrp(t) as follows:
The PCORRC model [15] as given by Eq. (12) was
selected to evaluate rb(t) 2wtf
" !!# rrp ðtÞ ¼ , (13)
2u dðtÞ 0:571LðtÞ M ðtÞD
rb ðt Þ ¼  1– 1 – exp – pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ,
D wt Dðwt – dðtÞÞ where σf is the flow stress and equals 0.9σu [12] and M(t) is
the Folias factor given by (the notation M(t) is used to
(12)
emphasize that M is in general a function of time because
where σu is the pipe tensile strength; D is the pipe diameter; the defect length can grow over time)
8 sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
>
> LðtÞ2 LðtÞ4 LðtÞ2
>
< 1 þ 0:6275 – 0:003375 , £50;
M ðt Þ ¼
Dwt ðDwtÞ2 Dwt
(14)
>
> LðtÞ2 LðtÞ2
>
: 3:293 þ 0:032 , > 50:
Dwt Dwt
To carry out the reliability analysis, we selected nine assumed to be fully correlated. The statistics associated
defects located on one particular pipe joint (approximately with the parameters involved in the reliability analysis are
17.5 m long) from the 62 defects considered in Section 5.1. summarized in Table 2. The spatial variability of diameter,
Although the defects have been repaired and ceased wall thickness, tensile strength, defect depth, defect length
growing since 2010, they were hypothetically assumed to and model error associated with different defects was
be still active and follow the growth path predicted from ignored and all these parameters were assumed to be
the growth model. This assumption allows evaluation of mutually independent.
the time-dependent failure probability of the pipeline under A total of 100000 simulation trials were used to evaluate
the threat of the nine active defects for the sole purpose of the probabilities of small leak, large leak and rupture over a
illustrating the application of the developed growth model 10-year forecasting period since the most recent inspection
in the reliability analysis. The pipe joint selected was made (i.e., from years 2008 to 2017). For each simulation trial,
from API 5L X52 steel with a specified minimum yield the growth paths of the nine defects over the period from
strength (SMYS) of 359 MPa and a specified minimum years 2008 to 2017 (i.e., corresponding to t = 36 to 45 in
tensile strength (SMTS) of 456 MPa, and has an outside Eq. (2)) were generated from Eq. (2) with the values of αi,
diameter of 508 mm, an operating pressure of 5.66 MPa ti0, δ and  equal to the corresponding medians of their
and a nominal wall thickness of 5.56 mm. Due to lack of posterior distributions evaluated from the MCMC simula-
information of the defect lengths associated with the nine tion. The probabilities of small leak, large leak and rupture
defects in the ILI report of 2007, we assumed that the over a ten-year forecasting period are shown in Fig. 6.
defect lengths are static and follow an identical and Figure 6 indicates that the probability of small leak at a
independent lognormal distribution with a mean of 30 mm given time is the highest over the entire forecasting period,
and a COV of 50% based on the information provided in for this example, followed by the probabilities of large leak
[12]. The internal pressures of different years were and rupture in a descending order. The probabilities of
assumed to be identical and independent Gumbel- small leak, large leak and rupture gradually increase over
distributed random variables [17]; at a given year, the time with a similar trend. For example, the failure
internal pressures associated with all the defects were probability at the end of the forecasting period is about
Table 2 Probabilistic characteristics of the basic random variables used in the reliability analysis
parameter unit distribution type nominal value mean standard deviation source
D mm deterministic 508 508 – [17]
wt mm normal 5.56 5.56 0.25 [12]
L mm lognormal – 30 15 [12]
σu MPa normal 456 492 14.78 [18]
p MPa Gumbel 5.66 5.93 0.12 [12]
ζ – Gumbel 1.00 1.08 0.29 [19]
286 Front. Struct. Civ. Eng. 2013, 7(3): 276–287

Fig. 6 Failure probabilities of small leak, large leak and rupture

twice as high as that at the beginning of the forecasting comparison of the predicted and field-measured depths
period for each of the three failure modes considered. indicates that the proposed model can predict the corrosion
growth reasonably well: the absolute differences between
the predicted depths and the actual depths are less than and
6 Conclusions equal to 10%wt for 90% of the 62 defects.
To illustrate the application of the growth model in the
We propose an inverse Gaussian process- (IGP-) based reliability analysis, nine defects located on the same pipe
model to characterize the growth of depths of metal-loss joint were selected from the 62 defects. The failure
corrosion defects on oil and gas steel pipelines. The model probability of the pipe joint due to corrosion and internal
parameters (i.e., α, t0, δ and ) were assumed to be pressure was then evaluated for three distinctive failure
uncertain and evaluated using the hierarchical Bayesian modes, i.e., small leak, large leak and rupture over a
methodology based on the imperfect inspection data forecasting period of 10 years since the most recent
obtained from multiple ILI runs. The parameters α and t0 inspection. The simple Monte Carlo technique was
were assumed to be defect-specific, whereas δ and  were employed in the reliability analysis by taking into
assumed to be common for all defects. The biases, random consideration uncertainties in the corrosion growth,
scattering errors as well as correlations between the material and pipe geometric properties as well as internal
random scattering errors associated with the ILI tools pressure. The growth model and procedure for the
used in different inspections were accounted for in the reliability analysis described in this study can be relatively
Bayesian inference. The Markov Chain Monte Carlo easily implemented in a pipeline corrosion management
simulation techniques were employed to carry out the framework to facilitate the structural integrity management
Bayesian updating and numerically evaluate the probabil- of pipelines. It is noted that the growth model developed in
istic characteristics of the parameters in the growth model. this study ignores the spatial correlation between the
An example involving real ILI data collected from a growths of different defects, and is a phenomenological
natural gas pipeline that is currently in service in Alberta, model that does not attempt to address the mechanism of
Canada was used to illustrate the proposed growth model metal-loss corrosion.
and its application in evaluating the time-dependent failure
Acknowledgements The authors gratefully acknowledge the financial
probability of the pipeline due to corrosion. The growth
support provided by the Natural Sciences and Engineering Research Council
models were developed for 62 external corrosion defects (NSERC) of Canada and TransCanada Corporation through the Collaborative
that have been subjected to multiple ILI runs, and were Research and Development (CRD) program. The helpful comments provided
subsequently excavated, field measured and repaired. The by the anonymous reviewer are gratefully appreciated.
ILI data were used to carry out the Bayesian updating and
evaluate the parameters in the growth models correspond-
ing to each of the 62 corrosion defects considered. The References
median values of the posterior distributions of the model
parameters were then used to predict the depths of the 1. Hong H P. Application of stochastic process to pitting corrosion.
defects at the time of the field measurements. A Corrosion, 1999, 55(1): 10–16
Hao QIN et al. Inverse Gaussian process-based corrosion growth modeling 287

2. Maes M A, Faber M H, Dann M R. Hierarchical modeling of 2009, 28(25): 3049–3082


pipeline defect growth subject to ILI uncertainty. In: Proceedings of 12. CSA-Z662 Oil and Gas Pipeline Systems. Canadian Standards
the ASME 28th International Conference on Ocean. Offshore and Association, Rexdale, Ontario, 2007
Arctic Engineering, Honolulu, Hawaii, USA, 2009, OMAE2009– 13. Zhou W, Hong H P, Zhang S. Impact of dependent stochastic defect
79470 growth on system reliability of corroding pipelines. International
3. Zhang S, Zhou W, Al-Amin M, Kariyawasam S, Wang H. Time- Journal of Pressure Vessels and Piping, 2012, 96–97: 68–77
dependent corrosion growth modeling using multiple ILI data. In: 14. Zhou W. Reliability evaluation of corroding pipelines considering
Proceedings of IPC 2012, IPC2012–90502. ASME, Calgary, multiple failure modes and time-dependent internal pressure.
2012 Journal of Infrastructure Systems, 2011, 17: 216–224
4. Wang X, Xu D. An inverse Gaussian process model for degradation 15. Leis B N, Stephens D R. An alternative approach to assess the
data. Technometrics, 2010, 52(2): 188–197 integrity of corroded line pipe- part II: Alternative criterion. In:
5. Chikkara R S, Folks J L. The Inverse Gaussian Distribution. New Proceedings of the 7th International Offshore and Polar Engineering
York: Marcell Dekker, 1989 Conference. Honolulu, 1997, 635–640
6. Al-Amin M, Zhou W, Zhang S, Kariyawasam S, Wang H. Bayesian 16. Kiefner J F, Maxey W A, Eiber R J, Duffy A R. Failure stress levels
model for the calibration of ILI tools. In: Proceedings of IPC 2012, of flaws in pressurized cylinders, Progress in flaw growth and
IPC2012–90491. ASME, Calgary, 2012 fracture toughness testing, ASTM STP 536. American Society of
7. Fuller W A. Measurement Error Models. New York: John Wiley & Testing and Materials, 1973, 461–481
Sons, Inc, 1987 17. Zhou W. System reliability of corroding pipelines. International
8. Jaech J L. Statistical Analysis of Measurement Errors. New York: Journal of Pressure Vessels and Piping, 2010, 87(10): 587–595
John Wiley & Sons, Inc, 1985 18. Jiao G, Sotberg T, Igland. RSUPERB. 2M- statistical data: Basic
9. Bernardo J, Smith A F M. Bayesian Theory. New York: John Wiley uncertainty measures for reliability analysis of offshore pipelines.
& Sons Inc, 2007 SUPERB JIP report no. STF70 F952112, SUPERB project 700411,
10. Gelman A, Carlin J B, Stern H S, Rubin D B. Bayesian Data 1995
Analysis. 2nd ed. New York: Chapman & Hall/CRC, 2004 19. Zhou W, Huang G. Model error assessments of burst capacity
11. Lunn D, Spiegelhalter D, Thomas A, Best N. The BUGS project: models for corroded pipelines. International Journal of Pressure
Evolution, critique and future directions. Statistics in Medicine, Vessels and Piping, 2012, 99–100: 1–8

You might also like