Engineering Failure Analysis 82 (2017) 1–15

Contents lists available at ScienceDirect

Engineering Failure Analysis

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

Statistical model and structural reliability analysis for onshore gas MARK
transmission pipelines
Konstantinos Pesinis, Kong Fah Tee⁎
Department of Engineering Science, University of Greenwich, UK

AR TI CLE I NF O AB S T R A CT

Keywords: The reliability of gas pipelines against rupture can be estimated based on historical failure data
Statistical analysis and theory of structural reliability. In this study, an integrated model for reliability analysis of
Gas pipelines failure data is presented jointly with a robust structural reliability model for the purpose of
Reliability comparison and cross-verification. The statistical model adopts a parametric hybrid empirical
External corrosion
hazard model complemented with a robust data processing technique known as the nonlinear
Rupture
quantile regression for reliability analysis and prediction. This model provides inferences on the
complete lifecycle reliability of the average pipe segment in the region under study. The struc-
tural reliability model is segment based and estimates rupture probabilities due to external metal
loss corrosion. The non-homogeneous Poisson process is used to model the generation of new
defects and the Poisson square wave process is used to model the growth of defects. The internal
pressure load is modelled as a discrete Ferry-Borges stochastic process. Then, an inspection and
maintenance plan is incorporated based on ASME B31.8S code of practice, for the service life
considered. The probability of detection and measurement error of the inspection tools is also
incorporated in the model. A numerical example of these models in an industrial context is
presented. Onshore gas transmission pipeline rupture data for the period from 2002 to 2014, was
obtained from the United States Department of Transportation Pipeline and Hazardous Materials
Safety Administration (PHMSA) database and analysed. The comparative study of the proposed
methodologies can assist gas pipeline operators in the informed implementation of optimal
maintenance strategies based on risk prioritization.

1. Introduction

Collection of incident data and database development for statistical analyses and probability prediction of future events is
considered a well-established practice when it comes to energy pipelines [1–5]. Structural reliability analysis (SRA) on the other
hand, offers an efficient and comprehensive way of directly assessing the loads and resistance associated with failure modes, ac-
counting for the separate effect of each random variable under consideration [6–9]. The uncertainties involved in the manufacturing
and operation processes are all incorporated in the SRA, by means of probabilistic approaches [10–12]. The comparison between SRA
and statistical approaches has always been challenging, for practical reasons. Pipeline historical failure data are scarce and generic, in
that they do not account for the wide variety of parameters and conditions associated with different pipelines [13,14]. However, the
validation of failure probability calculations (SRA) based on historical failure data is key to gaining confidence on the results,
provided the uncertainties associated with both methods are properly taken into account.
Previous studies that have utilized historical failure data for risk analysis of energy pipelines, evaluate the failure probabilities in

⁎
Corresponding author.
E-mail address: K.F.Tee@gre.ac.uk (K.F. Tee).

http://dx.doi.org/10.1016/j.engfailanal.2017.08.008
Received 16 January 2017; Accepted 22 August 2017
Available online 26 August 2017
1350-6307/ © 2017 Elsevier Ltd. All rights reserved.
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

the rate of occurrence of failures (ROCOF) sense for repairable systems [14,15]. This means that upon failure, the system is restored
to operation by repairing or even replacing some parts of the system, instead of replacing the entire system [16]. The ROCOF failure
rate considers a sequence of failure times within a time interval, as opposed to a single time to failure distribution [17,18]. However,
the ROCOF failure rate can be characterised only for a limited period of time, i.e. for as long as incident data are available. In practice,
most available incident databases that are fairly consistent, usually do not cover operation periods of > 30 years, while the actual
pipeline lifecycle can be significantly more than that. In this study, a reliability analysis of historical rupture data of gas transmission
pipelines is conducted, based on a non-repairable approach. The analysis is segment based. The time to rupture of a pipe segment is
defined as the time elapsing from the segment's start of operation until it ruptures. Upon rupture, the pipe segment is assumed to be
discarded and replaced by a new one, due to the detrimental effect of this type of failure. Therefore, pipes are examined as non-
repairable segments functioning within a repairable system, which is the entire pipeline network. This study proposes a model based
on the above described approach, which can provide reliability inferences on the average pipe segment's entire service life, even
when the data available are scarce, incomplete and both left and right censored. The average pipe segment exhibits the average
conditions in the region under study.
To illustrate the proposed model, statistical data was gathered from the Pipeline and Hazardous Materials Safety Administration
(PHMSA) database for the period 2002–2014, with regard to rupture incidents of onshore gas transmission pipelines. This database
provides information on the installation dates of pipes for which incidents have been reported as opposed to other popular databases,
where raw data is not reported. The incident data for the period 2002–2014 is considered fairly representative of the current state and
rehabilitation applications in regards to onshore gas transmission pipelines. Indubitably, PHMSA database covers thousands of miles
of onshore gas transmission pipelines and thus differences exist in materials, diameters, soil conditions and many other attributes.
Furthermore, there is discrepancy among the annual report database (that contains the pipeline details and mileage information) and
the incident database, which further hinders credible statistical analyses.
These uncertainties, associated with the statistical analysis of the PHMSA database, are taken into consideration by adopting a
hybrid empirical hazard model in liaison with a robust data processing technique, known as the nonlinear quantile regression. This
proposed combined model is thought to produce a credible expectation range of reliability performance of the infrastructure under
consideration for a complete service life of 100 years, as opposed to the ROCOF approach that can produce average annual failure
rates only for the time period under study, i.e. 13 years herein for the period 2002–2014. The resulting reliability forecasts are
compared with the results from a thorough SRA against external corrosion, on a reference pipe segment. The SRA models the ‘physics-
of-failure’ of the reference segment by adopting up-to date stochastic processes, by means of the Monte Carlo simulation. The non-
homogeneous Poisson process is used to model the generation of new defects and the Poisson square wave process is used to model
the growth of the defects. The internal pressure load is modelled as a discrete Ferry-Borges stochastic process. Then, an inspection and
repair program is applied to the reference segment based on ASME B31.8S [19], for a considered service life of 100 years. In addition,
a realistic characterisation of the probability of detection (PoD) and measurement error, associated with the inspection data is
incorporated in the model.
The main contributions of this work can be summarized in the following three aspects. First, the presentation of the robust
combination of an empirical hazard model with the nonlinear quantile regression technique, which can provide a credible range of
expected lifecycle performance for gas transmission pipelines on the segment level, even in the case of scarce and censored data.
Second, the validation of the statistical method based on a robust integration of stochastic process within a structural reliability
framework, which incorporates multiple defects, facilitates the identification of the critical defects and provides expected reliability
forecasts for the whole lifecycle of the examined pipe segment. Finally, the numerical application of the above-described metho-
dology on the PHMSA database for onshore gas transmission pipelines for the period 2002–2014, provides insights on the reliability
against rupture of pipes in this network, by considering the time to rupture of pipe segments, instead of the ROCOF failure rate which
is typically examined in literature.
The content of this paper is structured as follows. The hybrid empirical hazard model and the nonlinear quantile regression
methodology are presented in Section 2. The SRA model is described in Section 3. The numerical application of the above meth-
odologies is described in Section 4. Finally, some concluding remarks are presented in Section 5 on the basis of the outcomes from this
study.

2. Statistical analysis

2.1. Hazard functions and corresponding reliability

Gas pipelines, like most of the typical large-scale engineering structures, are considered to normally follow some common patterns
when it comes to their failure or hazard rate. One of the most commonly adopted hazard patterns in literature for large-scale
engineering structures, is the bath-tub pattern [20]. This is adopted in this study and is assumed to cover the whole lifecycle of gas
transmission pipelines, consisting of three phases as shown in Fig. 1. Phase 1 represents the incipient failure phase or ‘infant mor-
tality’ and the probability of failure decreases with time. Phase 2 represents random failures and the probability of failure is constant.
Phase 3 is called the wear-out or degraded phase, where the number of failures and thus the probability of failure increases with time.
Many models can be successfully employed to describe these mixed distributions. In this study, the exponential failure distribution
with a constant failure rate λ is employed for phase 2 and the Weibull distribution for phase 3. These are presented in Eqs. (1) and (2).
Phase 1 is neglected for gas transmission pipelines and is not taken into consideration herein. Phase 1 is thought to regard failures
at the initial operation years that are mainly caused by construction and/or manufacturing defects and human errors. However, since

2
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

Fig. 1. Bath-tub failure pattern.

gas pipelines are generally perceived as large cross-country infrastructure with long lifetimes, this first phase can be regarded trivial
in terms of both duration and magnitude. As a result, the hazard pattern considered, includes only phases 2 and 3 as presented in
Fig. 2.

h (t ) = λ for k1 ≤ t ≤ k2 (1)

β2 − 1
β t − k2 ⎞
h (t ) = λ + ⎛ 2 ⎞ ⎛⎜ ⎟⎜ ⎟ for t ≥ k2, a2 > 0, β2 > 1
⎝ α2 ⎠ ⎝ a2 ⎠ (2)

where h(t) is the hazard function and t is the age of the pipeline under consideration. Parameter λ is a constant failure rate and α2, β2
are the scale and shape parameter of the Weibull distribution. Parameters k1 and k2 refer to the start and the end of phase 2
respectively.
Once the hazard functions are obtained, the reliability functions can be derived. Eqs. (3) and (4) give reliability curves that
correspond to continuous hazard curves. The reliability functions corresponding to Eqs. (1) and (2) are:

R1 (t ) = e−λt for k1 ≤ t ≤ k2 (3)

…
t−k β2
−⎛ a 2 ⎞
R2 (t ) = e−λt + e ⎝ 2 ⎠ for t ≥ k2, a2 > 0, β2 > 1 (4)

Eqs. (1)–(4) entail the assumption that random failures occur throughout both phase 2 and phase 3, but in phase 3 an accelerating
over time degradation of the structure takes place and is described by a Weibull distribution.
β
Setting α = β22 and β = β2 − 1, Εq. (2) becomes:
α2

h (t ) = λ + a (t − k2 ) β (5)

According to Cox and Oakes [21] and Leemis [18], a discrete hazard function can be defined as:

fi
hi = , i = 1, 2,
R (ti − 1 ) (6)

where hi is the hazard function, R(ti − 1) is a reliability function, which is also referred to as the survival function and fi is a discrete
failure probability function that is equal to:

fi = R (ti − 1) − R (ti ), i = 1, 2, … (7)

In the limit of small time intervals Δt, the discrete hazard rate measures the rate of failure in the next instant on time for those
segments of the population (conditioned on) surviving to time t:

Fig. 2. Bath-tub failure pattern adopted for gas transmission pipelines.

3
K. Pesinis, K.F. Tee

hi =
N (ti − 1) − N (ti )
Δt ∗ N (ti − 1 )
, i = 1, 2,
… ,
Engineering Failure Analysis 82 (2017) 1–15

where N(ti − 1) and N(ti) are the numbers of functional segments of a population at time ti − 1 and ti, respectively.
Eq. (8) is also called the force of mortality and is a property of a time to failure distribution [22]. This approach is considered
adequate for the purpose of deriving the hazard function of gas transmission pipelines, since these can be divided in segments of the
same or similar characteristics, as described in the example of section 4. Also, this approach can efficiently deal with scarce, left or
right-censored data, by estimating the discrete hazards at those times that failure data are available [20].
Once the hazard function has been derived, parameters of Eq. (5) should then be estimated. The adopted approach for parameter
estimation is comprised of two parts. The first part is a heuristic method proposed by Sun et al. [20]. In specific, parameter k2 can be
(8)

determined by the discrete hazard bar chart that is produced from Eq. (8). Parameter k2 is thought to be easily obtainable in the
presence of complete and consistent data, but due to statistical uncertainties and fluctuations, this is not always the case and expert
opinion might be necessitated. Furthermore, as described above, λ is a constant rate of the exponential failure distribution and
represents the random failure phase. The value of λ can be obtained as the average of discrete hazard values that belong to this phase
(phase 2). It can be defined as:
1
λ=
Nk
∑ hi
1 < i ≤ Nk (9)

where k = k2 − k1.
Further, parameters α and β need to be estimated from the wear-out phase (phase 3), by means of the robust graphical method
described in the next section.

2.2. Nonlinear quantile regression

The most commonly adopted methods in literature for parameter estimation of probabilistic models are the maximum likelihood
method and the linear (or nonlinear) least-squares regression method [15,23–26]. Maximum likelihood method has the limitation
that it requires sufficient failure data to produce credible results. Least-squares regression has commonly been applied to obtain the
mean response of the dependent variable (hazard function in this study) as a function of the independent variable (time in this study).
However, both of them are thought to have difficulty in describing scarce and censored data and fail to yield a complete picture of the
variables' relationships. This is especially the case for heteroskedastic data, data with non-Gaussian error distributions and data
derived from sample selections. Also, they are sensitive to extreme outliers that can distort the results significantly. Furthermore,
many researchers aim to obtain an upper bound of the relationship such that the final result is conservative. So, it is especially useful
in applications where extremes are important, such as gas pipelines where upper quantiles of rupture incident levels are critical from
a public health perspective. This can be achieved by the quantile regression, since it estimates multiple rates of change (slopes), from
the minimum to maximum response [27,28].
Nonlinear quantile regression is considered to be able to deal with the above described challenges and is employed for the data
analysis and estimation of the parameters α and β, contained in Eq. (5). It is a quantifiable, repeatable and non-biased statistical
methodology that provides an estimate for a given quantile or, equivalently a given probability of exceedance. The greatest quantile
that can be calculated by the method depends on the number of data points available in the dataset. For datasets with a small number
of data, for instance < 20 data points, it is likely that the quantile regression lines for quantiles τ = {0.95, 0.98, and 0.99} might be
represented by the same line. The existence of various slope estimates for different quantiles, constitutes an indirect confirmation of
the heteroscedasticity of the dataset [24]. Errors are assumed to be independent and identically distributed (iid) in this study.
Typically a response variable y is a function of predictor variables x, so that y = f(x). To this end, Eq. (5) is rearranged as follows:
Q y = h (t ) − λ

and
x = t − k2

so that the following nonlinear relationship is defined:

Q y = ax β (10)

The conditional quantiles denoted by Qy(τ| x) are the inverse of the conditional cumulative distribution function of the response
variable Fy−1 (τ |x ) , where τ ϵ [0,1] denotes the quantiles [29]. For example, for τ = 0.90, Qy(0.90 | x) is the 90th percentile of the
distribution of y conditional on the values of x, i.e. 90% of the values of y are less than or equal to the specified function of x. In
contrast to the mean regression technique, which employs the least-squares procedure, parameters from the quantile regression are
obtained by minimizing Eq. (11) [30]:

S= ∑Q y ≥Qy y ] +
τ [Q yi − Q ∑Q y y ]
(1 − τ )[Q yi − Q
i  i
(11)
i i i < Q yi

The nonlinear iid quantile regression problem is solved using an interior point algorithm. This algorithm methodology is ana-
lytically described in Koenker and Park [31]. The nonlinear quantile regression methodology is part of the quantile regression

4
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

package, quantreg, in R software, which is used for the respective estimations in this study [32]. The method directly generates
quantiles at a specified exceedance level, removing the requirement for subjectively processing the data and therefore results in a
method that produces a unique answer [24,33].

3. Structural reliability approach

3.1. Generation of new external corrosion defects

The generation of new defects due to external corrosion, on a reference pipe segment, is characterised by means of the non-
homogeneous Poisson process (NHPP) [34–36]. The defect initiation times are assumed to be produced in a non-uniform way. The
total number of defects, N(t), generated over the whole pipeline lifecycle, (where t = 0 is the time of installation of the pipe), follows
a Poisson distribution with a probability mass function:

(S (t )) N (t ) e−S (t )
fp (N (t )| S (t )) = (t > 0)
N (t )! (12)
t
where N(t) are the total number of defects generated within time interval [0,t], S (t ) = ∫ λdτ d − 1dτ represents the expected number of
0
defects generated over the same time interval and λ(τ) is the assumed intensity factor corresponding to the pipe segment. It is
t
assumed that S (t ) = ∫ λdτ d − 1dτ = λτ d where λ and d are positive quantities that can be determined based on inspection data and/or
0
expert judgement. Considering n generated defects on the pipe component up to time T, the initiation times of the n defects are
denoted by to1, to2, …,ton (to1 ≤ to2 ≤ … ≤ ton ≤ T). The joint probability density function (PDF) of (to1, to2, …,ton) condition on N
(t) = n can be expressed as:
n
n! ∏i = 1 λ (ti )
fto1, … , ton | N (t ) (t1,…, tn |n) = (0 < t1 < t2 <… < tn ≤ T )
[S (t )]n (13)

The joint PDF of the initiation times for HPP, conditional on N(t) = n, is the same as the joint PDF of the order statistics of samples
of (Y1, Y2, …,Yn) where Y1, Y2, …,Yn are n independent and identically distributed (iid) random variables that are uniformly dis-
tributed over [0,T]. This conclusion for HPP can be generalised to NHPP, i.e., Y1, Y2, …,Yn are iid random variables with the
distribution [37]:
S (t )
P (Yi ≤ t ) = , (0 ≤ t ≤ T )
S (T ) (14)

It must be noted that, in literature, external corrosion defects are not considered to initiate immediately after the pipeline's
commissioning time [38,57]. Instead, initiation time is considered to be directly linked to the total elapsed time from the installation
of the pipeline till coating damage, plus the time cathodic protection can provide some level of corrosion prevention, after coating
damage. In the study of Velázquez et al. [38], corrosion field data for buried steel pipelines, indicated different corrosion initiation
times depending on the soil conditions, within a range of 2.57 to 3.06 years. To account for this fact herein in the NHPP model, the
initiation time of the first defect to1 is chosen to be derived from a uniform distribution with a lower bound of 2.57 and an upper
bound equal to 3.06.

3.2. Growth of defects

Velázquez et al. [38] and Gomes et al. [39] employed a time-independent non-linear power law model to characterise the growth
of corrosion defect depth in buried steel pipelines. The power law parameters are assumed to be random variables, derived from
extensive empirical studies. For given values of the parameters ki, αi and t0i, the depth of the ith defect at time t is null if t ≤ toi;
otherwise, it is given by
di (t ) = ki (ti − toi )αi (15)

where toi is the corrosion initiation of the ith defect and ki and αi are the proportionality and exponent factors respectively. In order to
describe the temporal uncertainty of defect growth in time, Markov Chain (MC) and gamma processes have commonly been adopted
in literature, in regard to energy pipelines [40–42]. However, in this study the rate of corrosion defect growth is modelled as a
Poisson square wave process (PSWP), as presented in Bazán and Beck [43]. While gamma process represents cumulative temporal
variability with stationary increments, the sample paths of deterioration are discontinuous in time and thus so are the growth of the
defects, something not realistic at some extent. The PSWP is also more advantageous compared to MC models since it does not require
discretization of the damage states or evaluation of transition probabilities.
In the model of this study, the proportionality factors ki are characterised by a PSWP with pulse heights (Uj) and durations
(tbj = tj + 1 − tj), which are random variables, as illustrated in Fig. 3 [43]. Pulse durations are adopted as exponential random
variables with parameter λ, where λ denotes the mean occurrence rate per unit time (or Poisson rate). The number of pulses, Z, within
a given period of time of ΔT follows a Poisson distribution with a probability mass function:

5
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

Fig. 3. Example of Poisson square wave process model.

U (Z = z |λ ) = (λ ΔT ) z exp(−λ ΔT ) z! (16)

The magnitudes of different pulses, U, are independent and identically distributed random variables and assumed to follow a T
location-scale distribution with distribution parameters μui, σui and νui. Finally, the exponent factors αi is assumed to follow an Inverse
Gaussian distribution with parameters μi and λi. For each pulse of the proportionality factor ki of the defect size in Eq. (16), the
increment in defect size is as follows:

di (t j + 1) = di (t j ) + Uj [(t j + 1 − to)ai − (t j − to)ai] for j = 0, 1, …, Z (17)

where Z is the number of pulses within a given sample and α is a realization of the exponent factor for the ith defect. Fig. 4 presents
one random realization of defect size sample paths in comparison to one random realization of the time-independent empirical power
law of Eq. (15) for the same random defect. The parameters (μu, σu, νu) and (μ, λ) of the random variables k and α are assumed to have
the values in Table 1.
Consistent with the typical assumption in industry practice, the length of the defect is assumed to appear on the pipe, at each
defect's initiation time toi, as a patch with a length and width, due to coating damage. Then the different defects' lengths are assumed
to remain unchanged with time and to follow a predefined probability distribution [36,44].

3.3. Internal pressure loading

The internal pressure is modelled by a simple but realistic discrete stochastic Ferry-Borges process [45,58]. Internal pressure loads
on the pipe segment are described as an annual sequence of independent and identically distributed random variables. Next, in
probability calculations, the maximum annual value of the sequence is used, which is obtained from the relevant internal pressure
probability distribution. An illustration of this model is illustrated in Fig. 5.

Fig. 4. Illustration of stochastic Poisson square wave process and time-independent power law defect growth models.

6
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

Table 1
Input parameters for the models for defect depth.

Model Variable Probability distribution Parameters Source

Defect depth k Τ-location-scale Location: μ = 0.168 mm/y Gomes et al. [39]

Scale: σ = 0.063
Shape: ν = 4.780
a Inverse Gaussian Mean: μ = 0.762
Shape: λ = 27.016
Defect density η Deterministic η = 0.321 def/m for 17 years pipeline age Valor et al. [48]

Fig. 5. Stochastic Ferry-Borges model for internal pressure of gas pipelines.

3.4. Uncertainties related to ILI tools

The ability of an inspection equipment to locate and quantify an actual external corrosion defect is also taken into consideration
herein [39]. The probability of detection (PoD) of a defect by a high resolution in line inspection (ILI) tool, usually depends on the
defect size di and the inherent ability of the tools based on their specifications. The PoD adopted in this study is of the following
exponential form:
PoD(di ) = 1 − e−qdi (18)
where q is a constant that defines the inherent tool detection capability according to vendor specifications [44].
Once a defect is detected, then it is assumed that an imperfect measure of its depth and length is obtained. The measured depths
and lengths by the ILI tool, diM and LiM respectively, are estimated based on the following equations:
di M = c1d + c2d di + εd (19a)

Li M = c1l + c2l Li + εl (19b)

where c1d (c1l) and c2d (c2l) are the biases of the ILI tool and are assumed to be deterministic quantities, while εd and εl are random
scattering errors related to the measured depths and lengths. The latter are assumed to follow a predefined normal distribution,
typically with zero mean and known standard deviations quantified from tool specifications [36].

3.5. Probability of rupture due to external corrosion

A corroding gas pipeline that is under internal pressure and contains multiple defects can fail by three different failure modes,
namely small leak, large leak and rupture, [58]. Small leak is the consequence of a defect penetrating the pipe wall. Large leak and
rupture, are together referred to as pipe burst and differ only by the occurrence or not, of an unstable axial propagation of the through
wall defect, resulting from plastic collapse of the pipe wall due to internal pressure at the defect. A detailed description of the
probabilistic methodology for the estimation of the failure probability can be found in Zhou [58] and Valor et al. [46]. The reliability
assessments, in this study, are implemented using the PCORRC model to calculate the failure (rupture) pressure of defects [47].

⎡ ⎛ ⎛ −0.157 L ⎞⎞ ⎤
2UTS wt ⎢ d
1 − i ⎜1 − exp ⎜ ⎟⎟ ⎥
i
Pf = 1.33
D ⎢ wt ⎜ ⎜ D (wt − di ) ⎟⎟ ⎥
⎢
⎣ ⎝ ⎝ 2 ⎠⎠ ⎥
⎦ (20)
where D is the pipe's diameter, wt is the wall thickness, UTS is the pipe's material ultimate tensile strength and the number 1.33 is the
PCORRC model error factor ([56]; [48]).
The rupture pressure is calculated according to the model developed by Ref. [49], as follows:

7
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

2 UTS wt
PR =
DM (21a)

2 1 2
⎧ ⎡1 + 0.6275L2 − 0.003375
⎪
M= ⎣
Dwt ( ) ⎤⎦
L2
Dwt
L≤ 20Dwt
⎨ L2
⎪ 3.3 + 0.032 Dwt L> 20Dwt (21b)
⎩
where M is the Folias bulging factor.
Next, the Monte Carlo (MC) framework is developed [26]. In each MC step, a vector of the random variables, which are associated
with the ith defect and are entailed in Eqs. (20, 21a, 21b), are randomly generated from their corresponding distributions. Then Pf and
PR are compared with the operating pressure Pop, which is also randomly generated based on a Ferry-Borges process realization. If a
defect's depth is equal or larger than the pipe's wall thickness (di ≥ wt) then a new failure pressure Pft is calculated by substituting
defect's depth with a delimited value di = 0.0009 ∗ wt. Then, if Pft ≥ Pyop, a small leak is counted. Otherwise, it is examined if
Pft > PR and if it is true, a large leak is considered and if not, a rupture instead. When di < wt, the same procedure is followed.
The number of ruptures, divided by the total number Nstp of Monte Carlo steps, constitute an unbiased estimator of the probability
of rupture PoRi of each defect. For each of the n defects, the above method is repeated Nstp times. Therefore, the process is repeated
n ∗ Nstp in total. It is also repeated for several future times tj within a set of time intervals. By the end of computations, the probability
of rupture PoRi(tj) of each of the corrosion defects in the pipeline at each time tj has been computed. Considering the definition of
reliability [45]:
Ri (t j ) = 1 − PoRi (t j ) (22)

and under the assumption that corrosion defects are independent, an upper bound of the probability of rupture of the pipe segment
can be obtained:
PoRseg (t j ) = 1 − Πi [1 − PoRi (t j )] (23)

or equivalently the reliability of the pipe segment against rupture:

Rseg (t j ) = 1 − PoRseg (t j ) (24)

Based on a study of Zhou [58], the assumption of independent multiple corrosion defects leads to higher probabilities of rupture
and thus more conservative results, which is the desired for long term probability of rupture predictions, if one considers the
detrimental effect of ruptures of gas transmission pipelines, from a public health perspective.

3.6. Maintenance plan

An inspection and repair program is applied to the reference segment based on ASME B31.8S [19], for the service life under
analysis. It is considered that various inspection and maintenance actions take place on pipe segments, within a pipeline network,
during their lifetimes and therefore their overall reliability is consistently maintained within a certain acceptable range. This range
varies depending on operators' decision making and different maintenance plans applied [50,51]. This study assumes that the re-
ference segment's overall reliability, is consistent to the implicit reliability in ASME B31.8S [14]. For the reliability prediction of the
reference segment, inspections and subsequent repairs are assumed based on high resolution ILI tools. The repair criteria prescribed
in the code of practice in ASME B31.8S are based on the maximum defect depth and the estimated failure pressure at the defect. The
proposed inspection intervals are dependent of the pipelines' class location and the second repair criterion (burst pressure at the
defect). The above-described prescriptions are summarized in Table 3.
After every inspection, the two repair criteria are examined and the pipe segment will be excavated and repaired if one of the
following is true [36]:
davg M ≥ ξn wtn (25a)

Pf M ≤ ηn Popn (25b)

where ξn and ηn denote the parameters selected based on the above described repair criteria [19], while wtn and Popn are the nominal
values of pipe wall thickness and operating pressure, respectively. davgM refers to the measured average defect depth, obtained from
Eq. (19a) and PfM to the estimated failure pressure. The latter is estimated by Eq. (20), by setting the nominal values of wall thickness
and pipe diameter (wtn and Dn) and also the values from Eqs. (19a, 19b) for each defect's average depth and length. It should be noted
that, the average values from the total number Nstp of iterations in the MC simulation, are used to characterise the defect depths and
lengths, for each different defect. Finally, the model error value is set to unity.
The repair actions on an excavated pipe segment first involve completely removing the existing coating of the segment. Plain
recoating or recoating plus sleeving, will be applied on the segment, based on the severity of the defects. Regardless of the specific
repair action, the segment is considered to be fully restored to pristine condition. This is a realistic assumption, considering that
sleeving offers at least a new pipe's burst capacity, while recoating mitigates the existing defects. The likelihood of a repair being of
poor quality is considered very low in industry practice and therefore ignored in this study [36].

8
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

3.7. Effect of maintenance on pipe segment's reliability

The reliability of the pipe segment, derived by Eq. (24), is expected to decrease monotonically in time, due to the degradation of
the structure. Following a repair, the segment's reliability is updated, by being restored to its original value at the time of each repair.
Thus, it can be described as follows [7,52]:

RA (t j ) = Rseg (t j − Ti ) Ti ≤ t j < Ti + 1 (26)

under the condition that the segment survives (does not fail) up to time T of each repair.
The overall segment's reliability, which is equivalent to the segment's survival function, considers the survival probability of the
segment up to time Tm of regular maintenance intervals and is formulated:

RS (t j ) = Rseg (t j − i·Ti )·Rseg (Tm )i i·Tm ≤ t j < (i + 1)·Tm (27)

Implicit on the above, is that the overall reliability (survival function) RS cannot increase throughout the life service, as opposed
to the conditional reliability RA, which is also referred to as availability [53].

4. Numerical application

4.1. Statistical analysis

The statistical analysis methodology presented in Section 2 is applicable to any energy pipeline network which has a large
population of segments and available historical rupture data. In fact, it is efficient with scarce and/or censored historical data. In this
numerical example, rupture data from the Pipeline and Hazardous Materials Safety Administration (PHMSA) database for the period
2002–2014 is examined, concerning onshore gas transmission pipelines. The pre-2002 data is excluded because the data is much less
detailed, which makes it very difficult to combine the data in these two periods together for analysis. The installation dates of pipes
are used, in order to derive the respective hazard and reliability functions. Upon rupture, the pipe segment is assumed to be discarded
and replaced by a new one, considering the segments as non-repairable joints that function within a repairable system, which is the
entire pipeline network. PHMSA database covers thousands of onshore gas transmission pipelines and thus differences exist in
materials, diameters, soil conditions and many other attributes. Discrepancy also exists among the annual report database (that
contains the pipeline details and mileage information) and the incident database, which necessitates data elaboration to ensure
credible inferences [54].
The proposed statistical analysis methodology, assumes that the network under consideration, can be divided into a number of
similar pipe segments, based on an adequate segmentation length. The resulting population of segments is considered to operate
under similar conditions. This is also assumed for those pipe segments that have reportedly failed thus far. This assumption certainly
entails some level of approximation, but taking into consideration that only one category of pipelines, i.e. onshore gas transmission is
examined. It is considered reasonable to expect similar behaviour at a significant extent, from this specific type of pipelines. Besides,
there are certain attributes that are common for the majority of onshore gas transmission pipelines that operated during 2002–2014.
In brief, steel, as pipeline material, accounts for over 99% of the total length. 97–98% of them, steel pipelines are cathodically
protected and coated and 80% of onshore gas transmission pipelines belong to the so-called class 1 areas (low-population-density
areas). It should be noted that, only pipe-related incidents are analysed. Pipe-related incidents include those occurring on body of
pipe and pipe seam, whereas non-pipe related incidents include those occurring on compressors, valves, meters, hot tap equipment,
filters and so on. Also, the rupture data used in this report are associated with the onshore (as opposed to offshore) gas transmission
(as opposed to gathering) pipelines, which account for the vast majority of gas pipelines in the US. To sum up, only the time to
rupture is of interest in this method and based on the determined segmentation length, the population is examined as “onshore gas
transmission pipeline segments”.
The segmentation length is assumed to be 12.5 m, which is approximately the typical length of a newly-constructed segment in
industry. Also, rupture incidents are assumed to be confined to the 12.5 m pipe segment. The rupture lengths of the 189 incidents,
which are reported by pipeline operators in the database for the period 2002–2014, is found to be around this magnitude (on average
around 10 m). As a result, the assumption that a pipe segment of 12.5 m is discarded after rupture and replaced by a new one, is
considered fairly representative of real practice. Thus, the gas transmission pipelines are assumed to form a statistically-related
population of 12.5 m segments performing a similar function in a similar location. As mentioned, 189 pipe-related rupture incidents
are reported in PHMSA, in the period 2002–2014. The time to rupture is of interest and as a result the installation dates of the
ruptured pipes are gathered. However, 10 rupture incidents concerned pipes, whose installation dates are unknown (are not reported
in the PHMSA). These 10 rupture incidents are not taken into consideration, adding extra uncertainty when all rupture incidents are
examined. However, this uncertainty is thought to be counterbalanced at a significant extent, by the fact that a percentage of
approximately 2% of the total mileage on an annual basis, is of unknown age (installation date) as well and thus not taken into
consideration.
Furthermore, the lack of detailed information in the PHMSA mileage database in respect to the age of the operating network,
necessitates to estimate average values. For every year between 2002 and 2014, the installation dates are available per decade. As a
result, different age groups of operating pipelines can be inferred for every year between 2002 and 2014. A careful examination of the
yearly changes of pipeline lengths for each installation date group for the period 2002–2014, indicates that each year 4 to 5 km of

9
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

Fig. 6. Rupture hazard bar chart of the onshore gas transmission pipeline network in 2002–2014.

new pipelines are added to the network on average. Some pipelines are replaced throughout the period 2002–2014 and thus the total
network mileage remains almost unchanged. In this study, the installation dates for years 2004 and 2014 are selected and the mean
time of their values are used to derive the respective pipeline age groups as presented in Table 4. It should be noted that, pipeline ages
only up to 64 years are considered, since above that value, it is not possible to derive plausible inferences about the respective
mileage per year.
The use of the average values of 2004 and 2014 returns a value of 18,477,255 km for 0–4 years old pipelines. Compared to the
initial observation that 4 to 5 km of new pipelines are installed every year for the period 2002–2014, it is deducted that the mean
values of 2004 and 2014 represent sufficiently the entire period 2002–2014. The same level of accuracy is expected for the rest of age
groups, amplified by the fact that 2004 is a very early year (3rd year) and 2014 is the final year of the studied period. Therefore, it can
be inferred that the totality of the pipe segments added and discarded from this network are well taken into account in this ap-
proximate approach. Furthermore, for every age group listed in Table 4, the lengths are first converted to segments (with seg-
mentation length of 12.5 m) and then divided by the number of years of each group, in order to derive the average number of pipe
segments based on their age per year. Finally, for every year the number of incidents is counted, based on the time to rupture of each
reported incident and the respective discrete hazards are then derived from Eq. (8). The resulting discrete hazard rate chart is
illustrated in Fig. 6. Within PHMSA, there are 20 rupture incidents, of pipe segments aged 65 years or more, with the oldest one being
a 99 years old segment that ruptured due to external corrosion in 2009. These are not part of the hazard bar chart of Fig. 6, since the
hazard dataset is right-censored at 64 years and only the remaining 159 reported incidents are considered. Resulting from the above,
a service life of 100 years is considered for reliability prediction.

4.2. Structural reliability analysis

For the purpose of comparison with the above-described statistical method, the approach presented in Section 3 is used. The
analysis is segment based and is applied on a specifically chosen reference segment. As mentioned, the SRA model estimates failure
probabilities due to external corrosion. Therefore, the reported rupture incidents due to external corrosion are gathered for the period
2002–2014 from PHMSA database and a reference pipeline is chosen based on their characteristics. In specific, the reference seg-
ment's random variables D, wt, UTS, Pop are defined with mean values equal to the average of the corresponding actual values of the
52 segments that ruptured due to external corrosion and with probability distributions and standard deviations, as proposed in
literature [36,55]. It should be noted that only metal loss external corrosion incidents, as opposed to stress corrosion cracking, are
counted. Moreover, average data from literature, which is based on extensive field studies on buried onshore energy pipelines, are
adopted to characterise the defect density and defect growth rate (Velazquez et al., 2009; [39,48]). All the above described values are
summarized in Tables 1 and 2.
The service life considered for reliability prediction is 100 years. The number of defects on the reference segment is characterised
by Eq. (12), where λ and d are assumed to be 0.0278 and unity, respectively, i.e. Λ(t) = 0.0139 t2, so that the expected number of
defects over the first 17 years of the period under examination is 0.321 per meter, according to the assumed average defect density
herein (Valor, 2014). The initiation time of the first defect to1 is obtained from a uniform distribution with a lower bound of 2.57 and
an upper bound equal to 3.06. The growth of each defect is modelled in Eq. (17) by adopting the probabilistic characteristics
presented in Table 1, as described in Section 3.2. Average values from literature, are adopted for the characterisation of the defect

10
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

Table 2
Probabilistic characteristics of the random variables.

Random variable Nominal value Unit Mean/nominal COV Distribution type Source of probabilistic characteristics

Annual maximum internal pressure, p 5.51 MPa 1.0 3% Gumbel Zhou and Zhang [55]
Diameter, D 493 mm 1.0 0 Deterministic Zhou and Zhang [55]
Wall thickness, t 7.09 mm 1.0 1.5% Normal Zhou and Zhang [55]
Tensile strength, σy 434 MPa 1.09 3% Normal Zhou and Zhang [55]
Defect length N/A mm 80 130% Lognormal Zhou and Zhang [55]; Zhang and Zhou [36]

length. It is assumed to have a mean value of 80 mm and a COV of 130% (Nessim and Stephens, 2006; [36,55]). Finally, the segments
diameter (D), wall thickness (wt) and ultimate tensile strength (UTS) are adopted to be invariant within the pipe segment.
The measurement by the ILI tools, defect depths and lengths, are assumed to be unbiased and therefore the parameters from Eqs.
(19a) and (19b) are set as c1d = c1l = 0 and c2d = c2l = 1. The random scattering errors εd and εl are defined as 7.8% ∗ wt and
7.8 mm respectively, according to common tool specifications [55]. The PoD associated with ILI tool is estimated from Eq. (18),
where q = 3.262 (mm− 1), i.e. the PoD is equal to 90% for a defect depth of 10% ∗ wt. Out of the 52 external corrosion rupture
incidents found in the PHMSA, 43 cases belong to Class 1 and the rest to Class 2; thus the safety parameters ξn and ηn in Eqs. (25a) and
(25b) are set as 0.5 and 1.39 respectively, based on Table 3. Since the majority belong to Class 1 pipelines, the initial inspection time
for the reference pipe segment, is assumed to take place at 10 years. Furthermore, an interval of 10 years is also selected for sub-
sequent inspections, based on the repair criteria in Table 3. Finally, the reliability of the reference pipe segment is evaluated through
106 Monte Carlo simulation in MATLAB software.

4.3. Results and discussions

Following the formation of the discrete hazard bar chart presented in Fig. 6, the methodology presented in Sections 2.1 and 2.2 for
hazard and reliability estimation, is applied next. First, the hazard bar chart needs to be distinguished in random failure and wear-out
phases, by deterministically defining parameter k2, which is the point that separates them. In the example herein, it is easily obtained
directly from observation of the chart and is evaluated as k2 = 27 years. This is thought to be the year after which wear out initiates
and thereupon increases in a rather consistent manner. This is verified not only by the fact that the majority of discrete hazards after
k2 have higher values, but also because there are no years with zero hazards. Furthermore, for the random failure period from year
k1 = 0 to k2 = 27 years, the hazard rate is defined as the average of the discrete hazards within that period. The obtained average
value is considered representative of the hazard trend for this period, since the respective values hardly fluctuate while there are
many zero values. In a special case of minimal fluctuations, along with non-zero values, the hazard function for this phase could be
more adequately characterised by the upper bound, as a more conservative evaluation from the safety viewpoint.
Next the non-linear quantile regression method is adopted, as per Section 2.2, to evaluate the parameters α and β of Eq. (5). As
mentioned, the non-linear iid quantile regression problem is solved using an interior point algorithm, which is an iterative procedure
that begins with initial estimates of the parameters. Different starting values of α and β of Eq. (5), could return slightly different final
estimates of α and β. The results verify what is implied by the hazard bar chart format; the heteroscedasticity of the data. This is
verified in an indirect way, since for different quantiles, different slope estimates are obtained. However, after 0.99 quantile, all slope
estimates are the same and therefore are represented by the same line. Thus, the 0.99 is chosen as the adequate quantile for parameter
evaluation in this example. The reason for this is that an upper bound estimation is desired from the safety viewpoint. Rupture
incidents of gas pipelines are very critical from a public health perspective and thus a conservative estimation of the hazard function
is of interest, so that all potential future rupture data will be below the upper bound with a 1% probability of exceedance. Non-linear
quantile regression is also proven to be robust with outliers in this application, as they are efficiently taken into consideration in the
0.99 quantile estimation. Therefore, this method is thought to produce a credible characterisation of parameters α and β. Finally, in
liaison with the discrete hazard estimation model, a reliability prediction for 100 years can be obtained, even though the dataset is
incomplete and right-censored at 64 years. In other words, the hazard curve is extended after 64 years until 100 years are reached.
The resulting parameters of Eq. (5) are presented in Table 5. The derived hazard function curve for all failure causes is presented in
Fig. 7. The respective reliability function curve for a prediction period of 100 years is presented in Fig. 8.
For the purpose of validation, the above statistical methodology is applied for rupture incidents due to external corrosion reported
to PHMSA in 2002–2014. Again, only those incidents for concerned pipelines of up to 64 years of age are considered but stress

Table 3
Inspection and repair criteria based on ASME B31.8S.

Repair criteria

Class 1st inspection year Remaining wall thickness (% of nominal) Failure pressure/MAOP
(%)

1 10 50 1.39
2 13 50 1.39

11
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

Table 4
Evaluation (approximate) of onshore gas transmission pipeline network ages in 2002–2014.

Pipeline age Length (km) Total number of segments Number of segments per year
(years)

0–4 18.477,255 1.478.180 369.545

5–14 49.254, 3525 3.940.348 394.034
15–24 46.207,805 3.696.624 369.662
25–34 45.487,671 3.639.013 363.901
35–44 83.936,3135 6.714.905 671.490
45–54 116.157,758 9.292.620 929.262
55–64 73.381,536 5.870.522 587.052

Table 5
Parameter estimation results of Eq. (5) for all failure causes.

Nonlinear (iid) quantile regression (starting estimates: α = 4 ∗ 10− 6 β = 10− 2)

Including standard errors

Quantile Power law fit Including standard errors

−6 −6
0.99 1.373 · 10 + 7.655 · 10 · (t-27) 0.206
1.373 · 10− 6 + 7.655 · 10− 6 ( ± 6.685 · 10− 7) ∗ (t-27)0.206

Table 6
Parameter estimation results of Eq. (5) for external corrosion incidents.

Nonlinear (iid) quantile regression (starting estimates: α = 10− 6 β = 10− 1)

Including standard errors

Quantile Power law fit Including standard errors

−7 −6
0.99 2.228 · 10 + 1.958 · 10 ·(t-37) 0.418
2.228 · 10− 7 + 1.958·10− 6 ( ± 1.302 · 10− 7) ∗ (t-37)0.418

Fig. 7. Hazard function curve for τ = 0.99 quantile (all failure causes).

corrosion cracking incidents are not counted. Even though external corrosion is considered a gradual deterioration failure me-
chanism, a random failure phase is again considered and the same methodology described in Section 2.1 is implemented. The
rationale behind this is the very nature of external corrosion mechanism since according to practical experience, substantial time is
normally required for metal loss to reach an extent that leads to failure. As a result, external corrosion failures in < 10 years for
example, can be thought random in nature, in that they are caused by extreme conditions.
Following this, the end point of the random failure phase for external corrosion ruptures is evaluated to be k2 = 37 years by the
respective hazard bar chart and then the upper bound based on 0.99 quantile for the wear out phase is estimated. It should be noted
that the end point of the random failure phase is not considered to be 30 years, even if the hazard rate has a peak value there. As

12
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

Fig. 8. Reliability function against rupture of the onshore gas transmission pipeline network in 2002–2014.

mentioned, the wear out phase is expected, except for the highest rate values, to also entail a consistent number of non-zero rate
values. As it can be observed from Fig. 9, after 30 years there are 7 years of zero incidents while upon year 38, incidents take place far
more consistently. Therefore, it is deducted that the ruptures of year 30, belong to the random failure phase, which ends at 37 years.
The resulting parameters of Eq. (5) for the external corrosion incidents are presented in Table 6 and the respective hazard curve in
Fig. 9. Finally, the resulting reliability function for 100 years is presented in Fig. 9. The aim is to compare and validate this result,
with the reliability function curve derived from the structural reliability analysis, on the considered reference pipe segment.
The results of the structural reliability method are illustrated in Fig. 10. In specific, the reliability of the reference pipe segment
without the consideration of maintenance activities derived from Eq. (24) is presented, along with the conditional reliability and the
overall segment's reliability from Eqs. (26) and (27) respectively, both of which take into account maintenance actions. The MC
simulation returns one maintenance action at 50 years, in which the segment is fully repaired and restored to pristine condition. As a
result, the conditional reliability (availability) is equal to 1 on year 50, while the overall reliability (survival function) as expected
does not increase, but is nonetheless positively affected by the maintenance action. The latter overall reliability curve is compared
with the reliability function obtained from the statistical method.
The comparative Fig. 11 presents the two reliability curves, obtained from the two different methods. These are considered to be
in sufficiently good agreement, taking into consideration the generic and approximate nature of the statistical method, as well as the
fact that certain relative values, have to be adopted in the structural reliability analysis. Therefore, the validity of both, as complete
integrated methods, each one in own field, is considered to be cross-verified. Moreover, a better knowledge of the state of onshore gas
transmission pipelines, as reported in PHMSA from 2002 to 2014 is also thought to be obtained, by means of explicit reliability curves
that describe the changes in reliability throughout a complete service life of 100 years, as opposed to average annual failure rates for
the 13 years among 2002 and 2014 that can be obtained by the usually adopted ROCOF methods. What is more, the relative values

Fig. 9. Hazard bar chart for rupture due to external metal loss corrosion.

13
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

Fig. 10. Results from the structural reliability approach.

Fig. 11. Comparison of the reliability function curves of the two different methods.

adopted to characterise defect density, growth rate and defect length can be investigated as part of a parametric study, in order to
increase the proximity among the two curves obtained from the two different methods. This can be of great pertinence to operators
who wish to determine values of specific defect parameters or other attributes that define risk in existing or new pipelines, based on
failure probabilities gathered from significant empirical evidence.

5. Conclusions

This study presents two distinct models for reliability prediction of onshore gas transmission pipelines against rupture and
compares them, in an effort to derive inferences on their applicability in relevant decision making. The first methodology is based on
historical incident data, while the other focuses on the analysis of a predominant failure mechanism of gas pipelines. In respect to the
statistical methodology, an empirical discrete hazard model is adopted in conjunction with a non-subjective parameter estimation
technique, namely non-linear quantile regression, to evaluate the hazard and reliability functions. The model can provide inferences
on a region's average pipe segment's reliability for its complete lifecycle, even in case of scarce, incomplete and censored data, as
opposed the commonly adopted ROCOF methods in literature that only account for the limited time period under study. Non-linear
quantile regression is used to yield a complete picture of the hazard dataset. Thus, it is considered that the generic nature of the
historical data is sufficiently accounted for and the resulting reliability lies within a credible expectation range. Furthermore, the
second model provides estimates of rupture probabilities against external corrosion on a reference pipe segment. The uncertainties
are modelled in an integrated manner through stochastic processes, associated with the segment-based loads and resistances. An
implicit inspection and maintenance plan is incorporated based on standardised codes of practice along with the corresponding
uncertainties of the inspection procedures. A numerical application of the two methods on the PHMSA database for the period
2002–2014 is presented with regard to rupture incidents on onshore gas transmission pipelines. The results demonstrate reasonable
proximity, which constitutes an inherent validation of the soundness of both methodologies and their estimations. Both methodol-
ogies can help pipeline operators, to evaluate plausible expectations for the performance of either new or existing pipelines. Those
evaluations can also correspond separately to specific parameters, like defect growth rates or defect density or other pipeline at-
tributes that define risk, by means of parametric studies. The result of the numerical application considered herein, provides addi-
tional knowledge on the state of the PHMSA onshore gas transmission network among 2002–2014 with respect to rupture, based on
the statistical model presented, which is more broad compared to the ROCOF based approaches employed usually in literature. This
period is considered representative of the up-to-date adopted techniques and strategies for operation and rehabilitation in industry,

14
K. Pesinis, K.F. Tee Engineering Failure Analysis 82 (2017) 1–15

amplifying the relevance of the results to reliability analyses for new or existing pipelines. The methodologies of this study can
therefore assist operators in making informed maintenance decisions based on reliability and risk.

References

[1] AER, Report 2013-B: Pipeline Performance in Alberta, 1990–2012, Alberta Energy Regulator, Calgary, Alberta, 2013.
[2] CONCAWE, Performance of European Cross-country oil Pipelines: Statistical Summary of Reported Spillages in 2013 and Since 1971, (2015) Brussels, Belgium: Report no
4/15.
[3] EGIG, 9th EGIG Report 1970–2013: Gas Pipeline Incidents, European Gas Pipeline Incident Data Group, Groningen, MA, Netherlands, 2015.
[4] PHMSA, phmsa.dot.gov, [Online] Available at http://phmsa.dot.gov, (2015) [Accessed September 2015].
[5] UKOPA, Pipeline Product Loss Incidents and Faults Report (1962–2013), UKOPA, Ambergate, UK, 2014.
[6] A.M. Alani, A. Faramarzi, M. Mahmoodian, K.F. Tee, Prediction of sulphide build-up in filled sewer pipes, Environ. Technol. 35 (14) (2014) 1721–1728.
[7] G. Barone, D.M. Frangopol, Reliability, risk and lifetime distributions as performance indicators for life-cycle maintenance of deteriorating structures, Reliab. Eng. Syst. Saf.
123 (2014) 21–37.
[8] Y. Fang, L. Wen, K.F. Tee, Reliability analysis of repairable k-out-n system from time response under several times stochastic shocks, Smart Structures and Systems 14 (4)
(2014) 559–567.
[9] K.F. Tee, L.R. Khan, H.P. Chen, A.M. Alani, Reliability based life cycle cost optimization for underground pipeline networks, Tunn. Undergr. Space Technol. 43 (2014)
32–40.
[10] Y. Fang, J. Xiong, K.F. Tee, An iterative hybrid random-interval structural reliability analysis, Earthquakes and Structures 7 (6) (2014) 1061–1070.
[11] M. Mahmoodian, A.M. Alani, K.F. Tee, Stochastic failure analysis of the gusset plates in the Mississippi river bridge, Int. J. For. Eng. 1 (2) (2012) 153–166.
[12] K.F. Tee, L.R. Khan, Reliability analysis of underground pipelines with correlation between failure modes and random variables, Journal of Risk and Reliability 228 (4)
(2014) 362–370 IMechE, Part O.
[13] J.F. Kiefner, R.E. Mesloh, B.A. Kiefner, Analysis of DOT reportable incidents for gas transmission and gathering system pipelines, 1985 through 1997, Pipeline Research
Council International, Inc., 2001 Catalogue, (L51830e).
[14] M. Nessim, W. Zhou, J. Zhou, B. Rothwell, Target reliability levels for design and assessment of onshore natural gas pipelines, J. Press. Vessel. Technol. 131 (6) (2009)
61701.
[15] F. Caleyo, L. Alfonso, J. Alcántara, J.M. Hallen, On the estimation of failure rates of multiple pipeline systems, J. Press. Vessel. Technol. 130 (2) (2008) 21704.
[16] L.R. Khan, K.F. Tee, Quantification and comparison of carbon emissions for flexible underground pipelines, Can. J. Civ. Eng. 42 (10) (2015) 728–736.
[17] H. Harold, H. Feingold, M. Dekker, Repairable Systems Reliability: Modeling, Inference, Misconceptions and their Causes, (1984).
[18] L.M. Leemis, Reliability: Probabilistic Models and Statistical Methods, Prentice-Hall, Inc., 1995.
[19] ASME, ASME B31.8S-2004, Managing System Integrity of Gas Pipelines: Supplemental to ASME B31.8, 2012 New York.
[20] Y. Sun, C. Fidge, L. Ma, Reliability prediction of long-lived linear assets with incomplete failure data, Quality, Reliability, Risk, Maintenance, and Safety Engineering
(ICQR2MSE), 2011 International Conference on. IEEE, 2011, pp. 143–147.
[21] D.R. Cox, D. Oakes, Analysis of Survival Data, CRC Press, 1984.
[22] C. Ebeling, An Introduction to Reliability and Maintainability Engineering, The McGraw-Hill Company, Inc., New York, 1997.
[23] O. Basile, P. Dehombreux, F. Riane, Identification of reliability models for non repairable and repairable systems with small samples, Proceedings of IMS2004 Conference on
Advances in Maintenance and Modeling, Simulation and Intelligent Monitoring of Degradation, 2004, pp. 1–8. Arles.
[24] T. Morrison, P. Spencer, Quantile regression–a statistician's approach to the local ice pressure-area relationship, OTC Arctic Technology Conference. Offshore Technology
Conference, 2014.
[25] S.E. Rigdon, A.P. Basu, Statistical Methods for the Reliability of Repairable Systems, Wiley New York, 2000.
[26] K.F. Tee, Y. Cai, H.-P. Chen, Structural damage detection using quantile regression, J. Civ. Struct. Heal. Monit. 3 (1) (2013) 19–31.
[27] B.S. Cade, B.R. Noon, A gentle introduction to quantile regression for ecologists, Front. Ecol. Environ. 1 (8) (2003) 412–420.
[28] C. Chen, An introduction to quantile regression and the QUANTREG procedure, Proceedings of the Thirtieth Annual SAS Users Group International Conference, SAS
Institute Inc., Cary, NC, 2005.
[29] R. Koenker, J.A.F. Machado, Goodness of fit and related inference processes for quantile regression, J. Am. Stat. Assoc. 94 (448) (1999) 1296–1310.
[30] Q.V. Cao, T.J. Dean, Using Nonlinear Quantile Regression to Estimate the Self-Thinning Boundary Curve, (2015).
[31] R. Koenker, B.J. Park, An interior point algorithm for nonlinear quantile regression, J. Econ. 71 (1) (1996) 265–283.
[32] R. Koenker, quantreg: Quantile Regression. R Package Version 4.94. http://CRAN.R-project.org/package=quantreg, (2012).
[33] Q.V. Cao, J. Wang, Evaluation of methods for calibrating a tree taper equation, For. Sci. 61 (2) (2015) 213–219.
[34] S.P. Kuniewski, J.A.M. van der Weide, J.M. van Noortwijk, Sampling inspection for the evaluation of time-dependent reliability of deteriorating systems under imperfect
defect detection, Reliab. Eng. Syst. Saf. 94 (9) (2009) 1480–1490.
[35] A. Valor, F. Caleyo, L. Alfonso, D. Rivas, J.M. Hallen, Stochastic modeling of pitting corrosion: a new model for initiation and growth of multiple corrosion pits, Corros. Sci.
49 (2) (2007) 559–579.
[36] S. Zhang, W. Zhou, Cost-based optimal maintenance decisions for corroding natural gas pipelines based on stochastic degradation models, Eng. Struct. 74 (2014) 74–85.
[37] V.G. Kulkarni, Modeling and Analysis of Stochastic Systems, CRC Press, 2009.
[38] J.C. Velázquez, F. Caleyo, A. Valor, J.M. Hallen, Predictive model for pitting corrosion in buried oil and gas pipelines, Corrosion 65 (5) (2009) 332–342.
[39] W.J.S. Gomes, A.T. Beck, T. Haukaas, Optimal inspection planning for onshore pipelines subject to external corrosion, Reliab. Eng. Syst. Saf. 118 (2013) 18–27.
[40] F. Caleyo, J.C. Velázquez, A. Valor, J.M. Hallen, Markov chain modelling of pitting corrosion in underground pipelines, Corros. Sci. 51 (9) (2009) 2197–2207.
[41] H.P. Hong, Inspection and maintenance planning of pipeline under external corrosion considering generation of new defects, Struct. Saf. 21 (3) (1999) 203–222.
[42] W. Zhou, H.P. Hong, S. Zhang, Impact of dependent stochastic defect growth on system reliability of corroding pipelines, Int. J. Press. Vessel. Pip. 96–97 (2012) 68–77.
[43] F.A.V. Bazán, A.T. Beck, Stochastic process corrosion growth models for pipeline reliability, Corros. Sci. 74 (2013) 50–58.
[44] M. Stephens, M. Nessim, A comprehensive approach to corrosion management based on structural reliability methods, 2006 International Pipeline Conference, American
Society of Mechanical Engineers, 2006, pp. 695–704.
[45] R. Melchers, Structural Reliability Analysis and Prediction, 2nd Edition, Wiley, 2004.
[46] A. Valor, F. Caleyo, J.M. Hallen, J.C. Velázquez, Reliability assessment of buried pipelines based on different corrosion rate models, Corros. Sci. 66 (2013) 78–87.
[47] B.N. Leis, D.R. Stephens, An alternative approach to assess the integrity of corroded line pipe-part I: current status, The Seventh International Offshore and Polar
Engineering Conference. International Society of Offshore and Polar Engineers, 1997.
[48] A. Valor, F. Caleyo, L. Alfonso, J. Vidal, J.M. Hallen, Statistical analysis of pitting corrosion field data and their use for realistic reliability estimations in non-piggable
pipeline systems, Corrosion 70 (11) (2014) 1090–1100.
[49] J.F. Kiefner, P.H. Vieth, A Modified Criterion for Evaluating the Remaining Strength of Corroded Pipe, Battelle Columbus Div, OH (USA), 1989.
[50] L.R. Khan, K.F. Tee, Risk-cost optimization of buried pipelines using subset simulation, Journal of Infrastructure Systems, ASCE 22 (2) (2016) 04016001.
[51] K.F. Tee, L.R. Khan, H.S. Li, Application of subset simulation in reliability estimation for underground pipelines, Reliab. Eng. Syst. Saf. 130 (2014) 125–131.
[52] N.M. Okasha, D.M. Frangopol, Redundancy of structural systems with and without maintenance: An approach based on lifetime functions, Reliab. Eng. Syst. Saf. 95 (5)
(2010) 520–533.
[53] K.B. Klaassen, J.C.L. Van Peppen, System Reliability, (1989) (Edward Arnold).
[54] NTSB, Integrity Management of gas Transmission Pipelines in High Consequence Areas, National Transportation Safety Board, 2015 SS-15/01.
[55] W. Zhou, S. Zhang, Impact of model errors of burst capacity models on the reliability evaluation of corroding pipelines, J. Pipeline Syst. Eng. Pract. 7 (1) (2015) 4015011.
[56] B. Fu, D. Stephens, D. Ritchie, C.L. Jones, Methods for Assessing Corroded Pipeline–Review, Validation and Recommendations, PRCI Report, 2000 Catalog, (L51878).
[57] A.W. Peabody, R. Bianchetti (Ed.), Control of Pipeline Corrosion, second ed., NACE International, Houston, Texas, 2001.
[58] W. Zhou, System reliability of corroding pipelines, Int. J. Press. Vessel. Pip. 87 (10) (2010) 587–595.

15