Proceedings of the 2012 9th International Pipeline Conference IPC2012 September 24-28, 2012, Calgary, Alberta, Canada

IPC2012-90502
TIME-DEPENDENT CORROSION GROWTH MODELING USING MULTIPLE ILI DATA
S. Zhang, W. Zhou and M. Al-Amin Western University London, Ontario, Canada S. Kariyawasam and H. Wang TransCanada Pipelines Calgary, Alberta, Canada

ABSTRACT This paper describes a non-homogeneous gamma processbased model to characterize the growth of the depth of corrosion defect on oil and gas pipelines. All the parameters in the growth model are assumed to be uncertain; the probabilistic characteristics of these parameters are evaluated using the hierarchical Bayesian methodology by incorporating the defect information reported by the multiple in-line inspections (ILIs) as well as the prior knowledge about these parameters. The bias and random measurement error associated with the ILI tools as well as the correlation between the measurement errors associated with different ILI tools are taken into account in the analysis. The application of the model is illustrated using an example involving real ILI data on a pipeline that is currently in service. The results suggest that the model in general can predict the growth of corrosion defects reasonably well. The proposed model can be used to facilitate the development and application of reliability-based pipeline corrosion management. INTRODUCTION Historical pipeline incident data indicate that metal-loss corrosion is one of the most common threats to the structural integrity of pipeline networks worldwide [1]. Over the past decade, reliability-based corrosion management programs have been increasingly adopted by pipeline operators [2-5]. Such programs typically consist of three components, namely periodic high-resolution inline inspections to detect and size corrosion defects in a given pipeline, failure probability evaluation of the pipeline based on the inspection results, and defect repair, if necessary, to ensure that the failure probability of the pipeline is below an allowable failure probability level for a reference pipeline segment length (e.g. one km) and a reference period of time (e.g. one year). The corrosion growth modeling plays a critical role in various tasks involved in the pipeline corrosion management, such as prediction of the pipeline failure probability as a function of time, determination of the location and timing of defect mitigations, and development of re-inspection intervals

[2,6]. Overly conservative corrosion growth models will lead to unnecessary mitigation actions, which can result in significant cost penalties for the pipeline operators. On the other hand, non-conservative corrosion growth models may lead to critical defects being missed by mitigation actions; failures of these defects have serious consequences. The corrosion growth process is an inherently complex process that involves both the temporal and spatial variability. The temporal variability means that the growth path of a given defect varies with time, and the spatial variability means that the growth paths of different defects are different but may potentially be correlated. The probabilistic corrosion growth models reported in the literature can be classified as random variable (RV)-based models and stochastic process-based models. For the RV-based models, the corrosion growth (e.g. the growth of the depth of a corrosion defect) is generally assumed to follow a linear or power-law function of time (e.g. [7-13]). The parameters in the growth models are assumed to be time-independent random variables. Such models do not take into account the temporal variability of the corrosion growth process. The stochastic process-based models, on the other hand, consider that the growth path of a given corrosion defect varies with time and therefore are able to account for the temporal variability of the corrosion growth process. The objective of work reported in this paper was to develop a time-dependent growth model for the depth of corrosion defects on pipelines using the inline inspection (ILI) data and accounting for various uncertainties included in the ILI data. To this end, a non-homogeneous gamma processbased corrosion growth model was proposed in a hierarchical Bayesian framework. The Markov Chain Monte Carlo (MCMC) simulation techniques were employed to evaluate the probabilistic characteristics of the parameters involved in the model. The application and validity of the proposed growth model is illustrated using one example. ILI DATA AND UNCERTAINTIES The periodic ILI data provide valuable information for the corrosion growth modeling and structural integrity management for energy pipelines. The ILI data are subjected to

Copyright 2012 by ASME

measurement errors as a result of the uncertainties associated with the ILI tool and sizing algorithm [14]. It is commonly assumed in the literature that the measured defect depth follows a normal distribution with a mean value equal to the actual depth and a standard deviation characterizing the random measurement error [15-16]. This assumption however ignores the potential bias in the ILI data. Inspection tools with similar technologies and/or sizing algorithms are usually employed in consecutive ILIs on a given pipeline; as a result, certain degree of correlation (hereinafter referred to as temporal correlation) is likely to exist between the measurement errors associated with the data from multiple ILIs for the same pipeline. In this study, we assumed that the measured depth for defect i (i = 1, 2, , m) at inspection j (j = 1, 2, , n), yij, is related to the corresponding actual depth, xij, by = + + (1)

where aj and bj characterize the bias associated with inspection j, and denote the measurement error associated with defect i at inspection j, and is assumed to follow a multivariate normal distribution with zero means and variance-covariance matrix , that is ~(, ) (i=1, 2, , m) (2)

where i = (i1, i2, , ij, , in); ~ denotes the assignment of a probability distribution to a random variable; denotes transposition; MVN((, C) denotes a multivariate normal distribution with mean values of and variance-covariance matrix C, and =[klkl]nn, (k = 1, 2, , n; l = 1, 2, , n). kl represents the temporal correlation coefficient between inspection k and inspection l; k and l denote the standard deviations of the measurement errors associated with inspections k and l, respectively. To quantify the measurement error and bias of the ILI data, the depths of recoated defects obtained from multiple ILIs and the field measurement were compared. It is assumed that the defect depths obtained from the field measurement are errorfree and equal the actual defect depths. This assumption facilitates the calibration of measurement errors of ILI tools and quantification of the temporal correlation between the measurement errors associated with different ILI tools using the Bayesian method. Detailed discussions associated with the calibration model and the parameters estimations are given in Al-Amin et al. [17]. LITERATURE REVIEW OF STOCHASTIC PROCESSBASED CORROSION GROWTH MODELS In the literature, the stochastic process-based models to characterize the growth of corrosion defect typically employ the gamma process (GP) or the Markov process (MP). The gamma process commonly used in the corrosion growth modeling is a stochastic process with independent non-negative increments that are gamma distributed with a time-variant

shape parameter and a time-invariant scale parameter [18-21]. Pandey et al. applied the gamma process to model a generic deterioration process [18]. Maes et al employed the gamma process to characterize the growths of multiple corrosion defects and accounted for the spatial correlation between the defects [16, 22-23]. Provan and Rodriguez [24] used a non-homogeneous Markov process to characterize the growth of pitting corrosion based on the experimental data for aluminum, where the transition probability function was calculated numerically due to the difficulty in obtaining the analytical solutions. Hong [25] developed a non-homogeneous Markov process to characterize the corrosion growth process and a nonhomogeneous Poisson process to model the corrosion initiation. The advantage of Hong's model (see also [26]) lies in that the transition probability function was obtained analytically through the time transformation-condensation method. Valor et al. [27] and Caleyo et al. [28] also used a non-homogeneous Markov process to characterize the growth of pitting corrosion. The corrosion initiation time is assumed to follow a Weibull distribution; the parameters in the transition probability function are assumed to be a power-law function of time. The maximum depth of a population of defects was then characterized by an extreme value distribution. In these MPbased models, the values of the model parameters as well as the predicted defect depth are sensitive to the total number of discretized damage or deterioration states assumed [25]. Timashev et al. [29] developed a non-homogeneous Markov process model for the growth of corrosion defects on pipelines and applied the model to evaluate the failure probability of pipelines due to corrosion. To solve the transition probability function associated with the Markov process, the timedependent transition intensities were evaluated by employing the defect information obtained from repeated inline inspections. It can be inferred that the ILI data were assumed to be error-free and the defects on the pipelines were assumed to be spatially independent. Although the MP-based models provide a viable option for modeling the corrosion growth process, there are several difficulties associated with the MPbased models, namely 1) selecting an appropriate number of damage states and quantifying the corresponding transition probability function; 2) using the ILI data to evaluate the transition probability while accounting for the various uncertainties associated with the data, and 3) accounting for the spatial correlation between individual defects. The Bayesian-based methodologies provide a powerful tool to incorporate the ILI data into corrosion growth models. For example, the hierarchical Bayesian methodology and the Bayesian dynamic linear model have been used to update the parameters in the deterioration models based on inspection results [15-16, 30-31]. A two-stage Bayesian model was applied to analyze the reliability data for nuclear power facilities [32]. Based on inspection results, a three-level Bayesian model was used to derive the distributions for the corrosion pit depth and number of corrosion pits in a pressurized water reactor vessel [33] and steam generator tubes

Copyright 2012 by ASME

in nuclear facilities [16]. These Bayesian models can account for the probability of detection and measurement errors associated with the inspection tools. The Markov Chain Monte Carlo (MCMC) simulation was used to evaluate the posterior distributions of the distribution parameters for the pit depth. Maes et al. [15] proposed a hierarchical Bayesian approach to model the time-dependent growth of corrosion defects on pipelines based on the ILI data. They considered the temporal and spatial variability associated with the corrosion defect as well as the uncertainties (i.e. probability of detection and measurement errors) associated with the ILI tool. The growth of a given corrosion defect was characterized by a non-homogeneous gamma process with a time-dependent shape parameter and a time-independent scale parameter. The shape parameters for different defects were assumed to be independent of each other, whereas the scale parameters for different defects were assumed to be spatially correlated. They indicated that the MCMC simulation technique becomes time-consuming and may encounter convergence issues if the number of defects is large (say, greater than 100). Therefore, an equivalent Log-likelihood method was developed to estimate the mean and standard deviation of the distribution parameters. In the analysis, the initiation time of the corrosion defect is ignored; in other words, corrosion defects were assumed to start growing as soon as the pipeline is installed, which may not be always the case. The model characterizes the spatial correlation between individual defects using a set of so-called local-covariates (i.e. regression variables or explanatory variables), which may not be readily available in practice. The multivariate dynamic linear model (DLM) [34], which is a very versatile Bayesian tool commonly used in the time series analysis, has also been used to characterize the deterioration process with spatiotemporal characteristics. DLM is similar to the linear regression model but allows the model parameters to change, systematically, with time, capturing the temporal variability in the series. Little et al. [30-31] adopted a spatial-temporal DLM to characterize the corrosion process of an industry furnace and large industrial storage tanks using the Bayesian linear updating methodology. Empirical distance-based exponential covariance function was used to quantify the covariance matrix of multiple corrosion defects. Randell et al. [35] utilized DLM to characterize the growth of corrosion in complex industrial systems, and developed a methodology for inspection design and decision making based on the utility criteria. A full-scale offshore platform was used to illustrate the application of the proposed model. DLM was also used to predict the deterioration of various members in reinforced concrete bridges based on the inspection data [36]. The main advantage of DLM is that it can relatively easily account for the spatial and temporal variability of deterioration processes and incorporate the uncertainties involved in the inspection data. Furthermore, it is straightforward to update the probability distributions of the parameters in the deterioration models based on the inspection data. However, there are difficulties associated with the implementation of DLM for the monotonically increasing

corrosion process. Furthermore, the use of DLM to predict the corrosion growth in future times at which no inspection data are available requires linear extrapolation of the present growth rate into the future, which may be unrealistic. Based on above-summarized literature review, the nonhomogeneous gamma process-based hierarchical Bayesian model [37] was selected in this study to model the depth growth of corrosion defects on pipelines. The defect information reported by multiple inspections was used to derive the probability distributions for the parameters in the nonhomogeneous gamma process-based growth model based on the Bayesian updating. Details of the model are described in the following section. BAYESIAN UPDATING PARAMETERS OF GROWTH MODEL

Non-homogeneous gamma process-based hierarchical Bayesian model Consider that m active corrosion defects on a given pipeline have been inspected and sized by n inspections over a period of time. The measured depth of defect i (i = 1, 2, , m) at inspection j (j = 1, 2, , n), yij, is a function of the corresponding actual depth xij as given by Eq. (1). Further define yi = (yi1, yi2, , yij, , yin) and xi = (xi1, xi2, , xij, , xin). It follows from Eqs. (1) and (2) that yi can be characterized by a multivariate normal distribution given by Eq. (3): ~( + , ) (i = 1, 2, , m) (3)

where a = (a1, a2, , aj, , an) and b is an n-by-n diagonal matrix with diagonal elements equal to bj (j = 1, 2, , n). The parameters aj, bj (j = 1, 2, , n) and are assumed to be deterministic quantities and can be obtained from the calibration analysis of the inspection tools as described in Section 2 (e.g. set equal to the mean values of the corresponding posterior distributions derived from the Bayesian analyses). The actual defect depth of a given corrosion defect at time t (years) (t = 0 representing the time of installation of the pipeline), x(t), was assumed to be characterized by a nonhomogeneous gamma process that is parameterized as: G(x(t)|(t),) = (t)x(t)(t)-1e-x(t)/((t)) ( 4)

where G() denotes the probability density function (PDF) of the gamma distribution; () denotes the gamma function; is the so-called rate parameter (i.e. inverse of the scale parameter) [38-39], and (t) is a time-dependent shape parameter that was assumed to follow a power-law function of time given by (t) = 1(t-t0)2, with t0 denoting the initiation time of the defect growth (i.e. time elapsed from the installation of the pipeline up to the point when the defect starts growing). It follows that the mean and variance of x(t) at time t, denoted by m(t) and ((t))2, equal (t)/ and (t)/2, respectively. In the adopted

Copyright 2012 by ASME

model, 1 in conjunction with (i.e. 1/) represents the mean of the actual depth at the first unit increment of time; 2 characterizes the mean growth path over time, with 2 > 1 representing an accelerating growth, and 0 < 2 < 1 representing a decelerating growth. If 2 equals unity, the mean growth path will become linear and x(t) will be simplified to a homogeneous gamma process that has been extensively employed in the literature to characterize degradation processes, e.g. [21, 40-41]. An implicit assumption in adopting Eq. (4) to model the corrosion growth is that the rate of growth is non-stationary and x(t) has independent gamma distributed increments. By assuming the defect depth at t = t0 to be zero, the actual depth of a particular defect i at the time of the jth inspection, tj, can be obtained from the actual depth at the time of the (j-1)th inspection, tj-1, and the gamma distributed growth, xij, from tj-1 to tj; that is ~( | , ) = 1 0
2

= 1 +

( 5) ( 6)

Fig. 1 Structure of the hierarchical Bayesian corrosion growth model

where i is a defect-specific rate parameter associated with defect i, and ij is the time-dependent shape parameter associated with defect i given by = 1 0
2

The parameter t0i in Eq. (7) denotes the initiation time of defect i. In this study, we assumed that 1, 2, t0i and i are uncertain parameters and employed the Bayesian updating to derive the probability distributions for these parameters. The four-level hierarchical structure of the Bayesian time-dependent corrosion growth model is depicted in Fig. 1.

1 0 (j = 2, 3, , n)
2

(j = 1)

(7)

The parameters symbolized by circles in Fig. 1 are uncertain and need to be estimated whereas the parameters symbolized by squares are known quantities in the analysis. The first level in Fig. 1 includes the inspection data, i.e. the defect depths reported by inspections, which are associated with measurement uncertainties. The second level represents the auxiliary variables that include the actual depths at the times of inspections and increments of the actual depths between two consecutive inspections. This level makes the likelihood function for the measured depth mathematically tractable and facilitates the Bayesian updating. This is the so-called data augmentation technique first proposed by Tanner and Wong [42] in the context of missing data problems [43]. The second level also captures the growth path of the defect, where the growth of the actual depth over a particular time interval is assumed to follow a gamma distribution given by Eq. (6), which allows us to characterize the temporal variability and model the growth of the actual depth as an independent and positive increment. The third level includes the parameters of the gamma process (i.e. ij and i). Furthermore, ij is a function of 1, 2, and t0i, where 1 and 2 are assumed to be common for all the defects whereas the initiation time, t0i, is a defect-specific parameter. Finally, 1, 2, and t0i are defined in the fourth level. Figure 1 also includes three levels of known information for deriving the posterior distribution, namely 1) the hyper-parameters for the prior distributions of 1, 2, t0i, and i, 2) the background information, tj, denoting the time of the jth inspection since the installation of the pipeline, and 3) the variance-covariance matrix of the measurement error, , and the bias, a and b, obtained from the calibration of the inspection tools. Prior distribution For m active corrosion defects, the stochastic corrosion growth model described in the previous section involves two

Copyright 2012 by ASME

common parameters (i.e. 1 and 2) and 2m defect-specific parameters, namely the rate parameter i and the initiation time t0i (i = 1, 2, , m). To apply the Bayesian updating, prior distributions for the uncertain parameters need to be specified. In this study, the gamma distribution was selected as the prior distributions for 1, 2, and i. The selection of the gamma distribution for 1 and 2 is mainly based on the consideration that the distribution ensures 1 and 2 to be positive quantities and can be conveniently made as a non-informative distribution. The assignment of the gamma distribution as the prior distribution for i can lead to a conjugate posterior distribution for i, which improves the computational efficiency. The prior distribution for time t0i was assigned as a uniform distribution with a lower bound (denoted by lb) of zero and an upper bound (denoted by ub) equal to the time interval between the installation of the pipeline and the first detection of defect i. We further assumed that the prior distributions for i (t0i) associated with different defects are identical and mutually independent (iid). The prior distributions for 1, 2, i and t0i are summarized as follows: 2 ~(2 |, ) 0 1 ~(1 |, )

Bayes rule, an updated distribution, p(|D), for can be obtained as: (|) =
(|)() (|)()

where represents proportionality. The distribution p(|D) is the joint posterior distribution for . Equation (9) represents a typical one-layer Bayesian estimate, where the distribution parameters for the prior distribution for are deterministic quantities. However, if the uncertainties in the distribution parameters for the prior distribution for are considered, the joint prior distribution, f(), for the distribution parameters for (i.e., ) has to be introduced. Then Eq. (9) becomes (|, ) =
(|)(|)() (|)(|)()

(|)()

(9)

( |, ) ( = 1, 2, , ) ~

where U(|lb, ub) denotes the probability density function (PDF) of uniform distribution with a lower bound of lb and an upper bound of ub, and p, q, r, s, u, v, lb and ub denote the distribution parameters for the prior distributions and are assumed to be known quantities. Posterior distribution Given the specified prior distributions, the likelihood functions, the inspection data yi and calibration parameters a, b and associated with the inspection data, the posterior distributions for xi, 1, 2, i and t0i (i = 1, 2, , m) can be obtained from the Bayesian theorem. For brevity, let denote a vector consisting of all the uncertain parameters, i.e., = (x1, x2, , xi, , xm, 1, 2, , t0), where = (1, 2, , m) and t0 = (t01, t02, , t0m); let D denote all the inspection data, and denote the set of known and deterministic parameters, i.e., D = (y1, y2, , ym) and = {a, b, , t}, where t = (t1, t2, , tm). For clarity, is not included in the associated equations and descriptions in the following sections. The Bayesian method for estimating a set of unknown model parameters, denoted by , is an updating process by combining the information contained in the prior distribution and likelihood function of the observed data D. Let f() and L(D|) denote the joint prior distribution for and the likelihood function for D given , respectively. Based on the

(0 |, )( = 1, 2, , ) ~

(8)

The distribution, p(|D, ), represents the joint posterior distribution for the parameters in a hierarchical Bayesian framework. It is extremely difficult, if not impossible, to obtain an analytical expression for the joint posterior distribution given by Eqs. (9) and (10); furthermore, it is highly computationally intensive to directly evaluate the highdimensional integrations involved in these equations. To overcome these difficulties, the Markov Chain Monte Carlo (MCMC) simulation has been widely used over the last decade to evaluate the posterior distributions [44]. The key to the MCMC technique is sequentially drawing random samples of the parameters considered to create a Markov process whose stationary or target distribution is the joint posterior distribution being sought. MCMC is an iterative process in that at each step random samples are drawn from distributions that depend on the random samples drawn in the previous step. After an initial sequence of iterations (i.e. the so-called burn-in period), the random samples drawn from the subsequent iterations converge to the stationary distribution, i.e. the joint posterior distribution. If the number of iterations is large enough, the samples drawn after the burn-in period can then be used to evaluate the probabilistic characteristics (e.g. mean and standard deviation) of the posterior distributions, which is the same as the analysis involved in a conventional Monte Carlo simulation. The commonly used sampling algorithms in MCMC are Metropolis (or Metropolis-Hasting) algorithm and Gibbs sampler. A so-called slice sampling has also been increasingly employed in MCMC. Details about the aforementioned algorithms are available in the literature, e.g. [44-47]. In this study, the MCMC simulation was carried out using the software OpenBUGS [48]. NUMERICAL EXAMPLES In this section, an example involving real ILI data on a gas pipeline that is currently in service is used to illustrate the application of the corrosion growth model proposed in this study. The pipeline of interest was constructed in 1972 and

(|)(|)()(10)

Copyright 2012 by ASME

has been subjected to multiple inline inspections over the last decade. The time-dependent growth model was developed for 62 external corrosion defects that were excavated, measured in the ditch, and recoated in 2010. The apparent growth paths of these defects indicated by the ILI data in 2000, 2004 and 2007 as well as the field-measured depths in 2010 are shown in Figure 2, where wt denotes the pipe wall thickness. The field measured depths of the defects were assumed to equal the actual depths of the defects [17]; therefore, the actual depths of the 62 defects in 2010 are known. The probabilistic characteristics of the parameters of the growth models for each of the 62 defects were evaluated using the Bayesian updating based on the ILI data obtained in 2000, 2004 and 2007. The growth model was then validated by comparing the actual depths of the defects in 2010 with the corresponding depths predicted by the growth model. The calibrated biases, the measurement errors associated with individual ILI tools as well as the temporal correlations among the ILI data reported by the ILI tools in 2000, 2004 and 2007 are as follows: a1 = a2 = 2.04 (%wt), a3 = -15.28 (%wt), b1 = b2 = 0.97, b3 =1.4, 1 = 2= 6.62 (%wt), 3 = 9.51(%wt), 12 = 0.82, and 13 = 23 = 0.7 [17], where the subscript 1, 2 and 3 denote the parameters associated with the ILI data obtained in 2000, 2004 and 2007, respectively. The analysis results are shown in Fig. 3 through Fig.6.
70 ILI-reported data 60 50 Field-measured data

80 70 60

Predicted depth dp (%wt)

50 40 30

d p=d a+10%wt
20

d p=d a
10 0 0 10 20 30 40 50 60 Field measured depth da (%wt) 70 80

d p=d a-10%wt

Fig. 3 Comparison of the predicted depth with the field measure depth in 2010

40 30 20 10 0 2000

2001

2002

2003

2004 2005 2006 Inspection time (year)

2007

2008

2009

2010

Fig. 2 Apparent growth paths of individual defects characterized by the ILI-reported and field-measured depths

Figure 3 shows the overall comparison of the predicted depth, dp, of 2010 with the actual depth, da, obtained from the field measurement in 2010 for the 62 defects considered. Three lines are shown in this figure, namely dp = da, dp = da + 10%wt and dp = da - 10%wt. The results indicate that the predictions are reasonably good in that the majority of the predictions (about 90% of the 62 defects) fall in the region bounded by da 10%wt. The predicted depths deviate significantly (say, the absolute difference between the predicted and actual depths greater than 10%wt) from the corresponding actual depths for only six defects, with the maximum absolute deviation being approximately 20%wt. Further investigation revealed that the defects for which the growth model provides poor predictions are either pinhole or circumferential grooving type defects [49]. It has been reported that the measurement error of the ILI data for these types of defects is in general large and can even exceed 40%wt in some cases [23]. We also observed relatively large measurement errors involved in the ILI data for these types of defects in this study, which leads to a poor prediction from the growth model. The predicted growth paths associated with four defects are shown in Figs. 4(a) through 4(d); for comparison, the field measured depth in 2010 and the ILI-reported depths in 2000, 2004 and 2007 are plotted in the figures as well.

Defect depth (%wt)

Copyright 2012 by ASME

55 50 45 40 35

70 65 60 55 50 45 Upper bound(90 percentile) Mean of predicted depth Lower bound(10 percentile) ILI depth Field measured depth

Depth (%wt)

Depth (%wt)
Upper bound(90 percentile) Mean of predicted depth Lower bound(10 percentile) ILI depth Field measured depth

30 25 20 15 10 5

40 35 30 25 20 15 10 5 0 1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 Time (year)

0 1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 Time (year)

(a) Defect #1
70 65 60 55 50 45

(d) Defect #60

Fig. 4 Predicted defect depth growth from the year of defect initiation up to 2012

40 35 30 25 20 15 10 5 Upper bound(90 percentile) Mean of predicted depth Lower bound(10 percentile) ILI depth Field measured depth

0 1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 Time (year)

(b) Defect #2
45 Upper bound(90 percentile) 40 35 30 Mean of predicted depth Lower bound(10 percentile) ILI depth Field measured depth

25 20 15 10 5 0 1978 1980 1982 1984 1986 1988 1990 1992 1994 1996 1998 2000 2002 2004 2006 2008 2010 Time (year)

(c) Defect #7

Figures 4(a) to 4(d) depict the mean growth paths from the years of defect initiation up to 2010, for four selected defects (Defects #1, #2, #7, and #60) respectively; in addition, the lower bound and upper bound corresponding to an 80% confidence interval for the prediction are also shown in these figures. For a given defect, the mean, d(t), and standard deviation, d(t), of the defect depth at time t were calculated using Eqs. (4) through (7), i.e. d(t) = 1(t-t0)2/, d(t) = (1(tt0)2/2)0.5, where 1, 2, t0 and were treated as deterministic quantities and set equal to their corresponding mean values of the marginal posterior distributions evaluated using MCMC. The lower bound and upper bound of the predicted depth at given time t were quantified assuming that the depth at this particular time follows a normal distribution with a mean of d(t) and a standard deviation of d(t) according to the central limit theorem. Figure 4 indicates that the predictions obtained from the model for Defects #1, #2, and #60 are reasonably good, but the predicted depth for Defect #7 is about 16%wt lower than the actual depth. The poor prediction can be attributed to that Defect #7 is a pinhole defect, for which the ILI data tend to involve relatively large errors. This is also reflected in Fig. 4(c), which shows that the defect depths reported by the three ILI tools in 2000, 2004 and 2007 are almost the same but much lower than the actual depth in 2010. Note that the initiation time in the model plays an important role in characterizing the growth of defect. If the ILI data for a particular defect indicate a fast growing trend for the defect, it is likely to be identified by the Bayesian inference as a relatively new defect with a large initiation time. This observation is consistent with the experimental results reported in the literature [24, 50] indicating that metal-loss corrosion tends to have a higher growth rate at the early stage of the corrosion process. In the present study, for example, the ILI data for Defect #60 reveal a higher rate of growth than those for Defects #1 and #2; therefore, the mean of the initiation time obtained from MCMC for Defect #60 is 12 years larger than that for Defect #1 and 13 years larger than that for Defect #2.

Depth (%wt)

Depth (%wt)

Copyright 2012 by ASME

0.05 year=2000 year=2001 year=2002 year=2003 year=2004 year=2005 year=2006 year=2007 year=2008 year=2009

0.04

pdf (1/%wt)

0.03

0.02

0.01

0 20 30 40 50 60 Depth (%wt) 70 80 90

Fig.5 Time-dependent PDF for the depth of Defect #2

Figure 5 depicts the probability density function (PDF) for the depth of Defect #2 from 2000 to 2009. The figure indicates that the PDF curve moves rightwards along the horizontal axis as time increases from 2000 to 2009, and the curves also becomes wider as time increases. These observations are consistent with the fact that both the mean value and standard deviation of the defect depth in a given year increase with time. Figure 6 illustrates the PDF curves for the depths of Defects #1 through #10 in 2009. As expected, the PDF curves for the depths of different defects are markedly different because the parameters and t0 in the growth model are specific to individual defect.
0.09 0.08 0.07 0.06 Defect #1 Defect #2 Defect #3 Defect #4 Defect #5 Defect #6 Defect #7 Defect #8 Defect #9 Defect #10

pdf (1/%wt)

0.05 0.04 0.03 0.02 0.01 0 10 20 30 40 50 60 Depth (%wt) 70

the parameters involved in the model were assumed to be uncertain and evaluated using the hierarchical Bayesian methodology based on the inspection data obtained from multiple ILI runs. The biases, measurement errors as well as the correlations associated with the ILI tools were also taken into account in the Bayesian inference. The Markov Chain Monte Carlo simulation was employed to carry out the Bayesian updating and derive the posterior distributions of the parameters in the growth model. An example involving real ILI data for a gas pipeline was used to illustrate the proposed model. The parameters of the growth models for 62 external corrosion defects that were field measured and recoated were evaluated based on the defect depths reported by multiple ILI runs prior to the field measurement. The parameters were then used to predict the depths of the defects at the time of the field measurements. The predicted defect depths were compared with the corresponding field-measured depths to validate the growth model. The analysis results suggest that the model proposed in this study in general characterizes the growth of the defect depth reasonably well: the absolute differences between the predicted depths and the field measured depths are less than or equal to 10%wt for 90% of the 62 defects. The model predictions for pinhole and circumferential grooving type defects are relatively poor mainly due to the large measurement errors involved in the ILI data for those types of defects. Further investigations are needed to refine and improve the growth model by incorporating the bias and measurement error associated with the ILI data specific for the pinhole and circumferential grooving type defects and by accounting for the potential spatial correlations among multiple defects. The proposed model provides a powerful framework to deal with various uncertainties involved in the corrosion growth modeling based on the ILI data and will facilitate the corrosion management of oil and gas pipelines. ACKNOWLEDGEMENTS The authors from Western University gratefully acknowledge the financial support provided by the Natural Science and Engineering Research Council of Canada (NSERC) and TransCanada through the Collaborative Research and Development (CRD) program, and by the Faculty of Engineering at Western. The constructive comments on the paper provided by Prof. H.P. Hong at Western University as well as the two anonymous reviewers, and the assistances provided by Ms. P. Kwong and Mr. B. Ong from TransCanada during this study are greatly appreciated. REFERENCES [1] Cosham A., Hopkins P. and Macdonald, K.A., 2007, Best Practice for the Assessment of Defects in Pipelines Corrosion, Engineering Failure Analysis, 14, pp. 12451265. [2] Kariyawasam, S. and Peterson, W., 2008, Revised Corrosion Management with Reliability Based Excavation

80

90

Fig. 6 PDF curves for the depth of Defect #1 to #10 in 2009

CONCLUSION This paper describes a non-homogeneous gamma processbased model to characterize the growth of metal-loss corrosion defects on oil and gas pipelines. The shape parameter of the gamma process was assumed to be time-dependent and follow a power-law function of time, whereas the scale parameter of the gamma process was assumed to be time-independent and defect-specific. Furthermore, the corrosion initiation time for individual defect is accounted for in the growth model. All

Copyright 2012 by ASME

Criteria, Proceedings of IPC 2008, IPC2008-64536, ASME, Calgary. [3] Nessim, M. A., Zhou, W., Zhou, J., Rothwell, B. and McLamb, M., 2009, Target Reliability Levels for Design and Assessment of Onshore Natural Gas Pipelines, ASME Journal of Pressure Vessel Technology, 131(6), DOI:10.1115/1.3110017. [4] Nessim, M. A., Zhou, W., Zhou, J. and Rothwell, B., 2009, Reliability Based Design and Assessment for Locationspecific Failure Threats with Application to Natural Gas Pipelines, ASME Journal of Pressure Vessel Technology, 131(4), DOI:10.1115/1.3110019. [5] Zhou, J., Rothwell, B., Nessim, M. A. and Zhou, W., 2009, Reliability-based Design and Assessment Standards for Onshore Natural Gas Transmission Pipelines, ASME Journal of Pressure Vessel Technology, 131(3), DOI:10.1115/1.2902281. [6] Nessim, M. A., Dawson, J., Mora, R. and Hassanein, S., 2008, Obtaining Corrosion Growth Rates from Repeat Inline Inspection Runs and Dealing with the Measurement Uncertainties, Proceedings of IPC 2008, IPC2008-64378, ASME, Calgary. [7] Ahammed M. and Melchers R. E., 1996, Reliability Estimation of Pressurised Pipelines Subject to Localised Corrosion Defects, International Journal of Pressure Vessels and Piping, 69, pp. 261-272. [8] Ahammed M., 1998, Probabilistic Estimation of Remaining Life of A Pipeline in the Presence of Active Corrosion Defects, International Journal of Pressure Vessels and Piping, 75, pp. 321-329.

[14] Kariyawasam, S. and Peterson, W., 2010, Effective Improvements to Reliability Based Corrosion Management, Proceedings of IPC 2010, IPC2010-31425, ASME, Calgary. [15] Maes, M. A. Faber, M. H., and Dann, M. R., 2009, Hierarchical Modeling of Pipeline Defect Growth Subject to ILI Uncertainty, Proceedings of the ASME 28th international conference on Ocean, Offshore and Arctic Engineering, OMAE2009-79470, Honolulu, Hawaii, USA. [16] Yuan, X. X., Mao, D. and Pandy, M. D., 2009, A Bayesian Approach to Modeling and Predicting Pitting Flaws in Stream Generator Tubes, Reliability Engineering and System Safety, 94, pp. 1838-1847. [17] Al-Amin, M., Zhou, W., Zhang, S., Kariyawasam, S. and Wang, H., 2012, Bayesian Model for the Calibration of ILI Tools, Proceedings of IPC 2012, IPC2012-90491, ASME, Calgary. [18] Pandey, M. D., Yuan, X. and van Noortwijk, J. M., 2005, Gamma Process Model for Reliability Analysis and Replacement of Aging Structural Components, Proceedings ICOSSAR, paper No. 311, Rome, Italy. [19] Pandey, M. D., Yuan, X. X. and van Noortwijk, J. M., 2009, The Influence of Temporal Uncertainty of Deterioration on Life-cycle Management of Structures, Structure and Infrastructure Engineering, 5(2), pp. 145156. [20] van Noortwijik, J. M, van der Weide, J. A. M, Kallen, M. J. and Pandy, M. D., 2007, Gamma Process and Peaks-overThreshold Distribution for Time-dependent Reliability, Reliability Engineering and System Safety, 92, pp. 16511658. [21] van Noortwijik, J. M., 2009, A Survey of the Application of Gamma Process in Maintenance, Reliability Engineering and System Safety, 94, pp. 2-21. [22] Maes, M. A. and Dann, M. R., 2008, Hierarchical Bayes Methods for Systems with Spatially Varying Condition States, Canadian Journal of Civil Engineering, 34, pp. 1289-1298. [23] Maes, M. A., Dann, M. R., Breitung, K. W. and Brehm, E., 2008, Hierarchical Modeling of Stochastic Deterioration, Proceeding of the 6th International Probabilistic Workshop, Darmstadt. [24] Provan, J. W. and Rodriguez III, E. S., 1989, Part I: Development of Markov Description of Pitting Corrosion, Corrosion, 45, pp. 178-192. [25] Hong, H. P., 1999, Application of Stochastic Process to Pitting Corrosion, Corrosion, 55(1), pp. 10-16, NACE, Houston. [26] Hong, H. P., 1999, Inspection and Maintenance Planning of Pipeline under External Corrosion Considering

[9] Pandey M. D., 1998, Probabilistic Models for Condition Assessment of Oil and Gas Pipelines, NDT&E International, 31(5), pp. 349-358. [10] Amirat A., Mohamed-Chateauneuf A. and Chaoui K., 2006, Reliability Assessment of Underground Pipelines under the Combined Effect of Active Corrosion and Residual Stress, International Journal of Pressure Vessels and Piping, 83, pp. 107-117. [11] Stephens, M. and Nessim, M.A., 2006, A Comprehensive Approach to Corrosion Management Based on Structural Reliability Method, Proceedings of IPC 2006, IPC0610458, ASME, Calgary. [12] Teixeira A.P., Guedes Soares C., Netto T.A. and Estefen, S.F., 2008, Reliability of Pipelines with Corrosion Defects, International Journal of Pressure Vessels and Piping, 85, pp. 228-237. [13] Zhou W., 2010, System Reliability of Corroding Pipelines, International Journal of Pressure Vessels and Piping, 87, pp. 587-595.

Copyright 2012 by ASME

Generation of New Defects, Structural Safety, 21, pp. 203-222. [27] Valor, A., Caleyo, F., Alfonso, L., Rivas, D. and Hallen, J. M., 2007, Stochastic Modeling of Pitting Corrosion: A New Model for Initiation and Growth of Multiple Corrosion Pits, Corrosion Science, 49, pp. 559-579. [28] Caleyo, F., Velzquez, J. C., Valor, A. and Hallen, J. M., 2009, Markov Chain Modelling of Pitting Corrosion in Underground Pipelines, Corrosion Science, 51(9), pp. 2197-2207. [29] Timashev, S. A., Malyukova, M. G., Poluian, L. V. and Bushinskaya, A. V., 2008, Markov Description of Corrosion Defects Growth and Its Application to Reliability Based Inspection and Maintenance of Pipelines, Proceedings of IPC 2008, IPC2008-64546, ASME, Calgary. [30] Little, J., Goldstein, M. and Jonathan, P., 2004, Spatiotemporal Modelling of Corrosion in An Industrial Furnace, Applied Stochastic Models in Business and Industry, 20, pp. 219-238. [31] Little, J., Goldstein, M., Jonathan, P. and den Heijer, K., 2004, Efficient Bayesian Sampling Inspection Processes Based on Transformed Spatio-temporal Data, Statistical Modeling, 4, pp. 299-313. [32] Bunea, C., Charitos, T., Cooke, R. M. and Becker, G., 2005, Two-stage Bayesian Models: Application to ZEDB Project, Reliability Engineering and System Safety, 90, pp. 123-130. [33] Celeux, G., Persoz, M., Wandji, J. N. and Perrot, F., 1999, Using Markov Chain Monte Carlo Method to Solve Full Bayesian Modeling of PWR Vessel Flaw Distributions, Reliability Engineering and System Safety, 66, pp. 243252. [34] West, M. and Harrison, J., 1989, Bayesian Forecasting and Dynamic Models, Springer-Verlag, NY. [35] Randell, D., Goldstein, M., Hardman, G. and Jonathan, P., 2010, Bayesian Linear Inspection Planning for Largescale Physical Systems, Proceedings of the Institution of Mechanical Engineers, Part O: Journal of Risk and Reliability. DOI 10.1243/1748006XJRR322. [36] Wang, J. and Liu, X., 2010, Evaluation and Bayesian Dynamic Prediction of Deterioration of Structural Performance, Structure and Infrastructure Engineering, 6(6), pp. 663-647. [37] Banerjee, S., Carlin, B. P. and Gelfand A. E., 2004, Hierarchical Modeling and Analysis for Spatial Data, Chapman and Hall/CRC. [38] Ang, A. H. S. and Tang, W. H., 1975, Probability Concepts in Engineering Planning and Design, Volume I: Basic Principles, John and Wiley & sons, NY.

[39] Jonhson, R. A., 2000, Probability and Statistics for Engineers (6th edition), Prentice-Hall Inc, US. [40] Yuan, X. X., Pandey, M. D. and Bickelb, G. A., 2008, A Probabilistic Model of Wall Thinning in CANDU Feeders due to Flow-accelerated Corrosion, Nuclear Engineering and Design, 238(1), pp. 16-24. [41] Cheng, T. and Pandey, M. D., 2012, An Accurate Analysis of Maintenance Cost of Structures Experiencing Stochastic Degradation, Structure and Infrastructure Engineering, 8(4), pp. 329-339. [42] Tanner, M. A. and Wong, W. H., 1987, The Calculation of Posterior Distributions by Data Augmentation, Journal of the American Statistical Association, 82(398), pp. 528-540. [43] Press, S. J., 2003, Subjective and Objective Bayesian Statistics: Principles, Models, and Applications (2nd edition), Wiley-Interscience, NJ. [44] Gelman, A., Carlin, J. B., Stern, H. S. and Rubin, D. B., 2003, Bayesian Data Analysis (2nd edition), Chapman & Hall/CRC. [45] Congdon, P., 2003, Applied Bayesian Modeling, John Wiley and Sons Ltd, NJ. [46] Neal, R. M., 2003, Slice Sampling, The Annals of Statistics, 31(3), pp. 705-767. [47] Jasa, T. and Xiang, N., 2009, Efficient Estimation of Decay Parameters in Acoustically Coupled-spaces Using Slice Sampling, Journal of Acoustical Society of America, 126(3), pp. 1269-1279. [48] Lunn, D., Spiegelhalter, D., Thomas, A. and Best, N., 2009, The BUGS Project: Evolution, Critique and Future Directions (with discussion), Statistics in Medicine, 28, pp. 3049-3082. [49] Pipeline Operators Forum (POF), 2009, Specifications and Requirements for Intelligent Pig Inspection of Pipelines, Version 2009. [50] Aziz, P. M., 1956, Application of the Statistical Theory of Extreme Values to the Analysis of Maximum Pit Depth Data for Aluminum, Corrosion, 12(10), pp. 495-506.

10

Copyright 2012 by ASME