You are on page 1of 13

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/248209444

Markov chain modelling of pitting corrosion in underground pipelines

Article  in  Corrosion Science · September 2009


DOI: 10.1016/j.corsci.2009.06.014

CITATIONS READS

69 436

4 authors, including:

F. Caleyo Julio César Velázquez


ESIQIE, Instituto Politécnico Nacional Instituto Politécnico Nacional
83 PUBLICATIONS   1,199 CITATIONS    38 PUBLICATIONS   388 CITATIONS   

SEE PROFILE SEE PROFILE

Alma Valor
Instituto Politécnico Nacional
44 PUBLICATIONS   676 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Magnetic anisotropy from crystallographic texture and Barkhausen noise measurements in pipeline steels - 425102817 View project

Analysis of the elastic-plastic behavior of pressurized internal cracks in welded joints of low carbon steel pipes View project

All content following this page was uploaded by Julio César Velázquez on 29 August 2018.

The user has requested enhancement of the downloaded file.


This article appeared in a journal published by Elsevier. The attached
copy is furnished to the author for internal non-commercial research
and education use, including for instruction at the authors institution
and sharing with colleagues.
Other uses, including reproduction and distribution, or selling or
licensing copies, or posting to personal, institutional or third party
websites are prohibited.
In most cases authors are permitted to post their version of the
article (e.g. in Word or Tex form) to their personal website or
institutional repository. Authors requiring further information
regarding Elsevier’s archiving and manuscript policies are
encouraged to visit:
http://www.elsevier.com/copyright
Author's personal copy

Corrosion Science 51 (2009) 2197–2207

Contents lists available at ScienceDirect

Corrosion Science
journal homepage: www.elsevier.com/locate/corsci

Markov chain modelling of pitting corrosion in underground pipelines


F. Caleyo a,*, J.C. Velázquez a, A. Valor b, J.M. Hallen a
a
Departamento de Ingenierı́a Metalúrgica, ESIQIE, IPN, UPALM Edif. 7, Zacatenco, México D. F. 07738, Mexico
b
Facultad de Física, Universidad de La Habana, San Lázaro y L, Vedado, 10400 La Habana, Cuba

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

Article history: A continuous-time, non-homogenous linear growth (pure birth) Markov process has been used to model
Received 14 April 2009 external pitting corrosion in underground pipelines. The closed form solution of Kolmogorov’s forward
Accepted 7 June 2009 equations for this type of Markov process is used to describe the transition probability function in a dis-
Available online 14 June 2009
crete pit depth space. The identification of the transition probability function can be achieved by corre-
lating the stochastic pit depth mean with the deterministic mean obtained experimentally. Monte-Carlo
Keywords: simulations previously reported have been used to predict the time evolution of the mean value of the pit
A: Steel
depth distribution for different soil textural classes. The simulated distributions have been used to create
B: Modelling studies
C: Pitting corrosion
an empirical Markov chain-based stochastic model for predicting the evolution of pitting corrosion depth
and rate distributions from the observed properties of the soil. The proposed model has also been applied
to pitting corrosion data from pipeline repeated in-line inspections and laboratory immersion
experiments.
Ó 2009 Elsevier Ltd. All rights reserved.

1. Introduction cess. He investigated the effect of corrosion defect size on the


load-resistance ratio and obtained the probability transition matrix
The stochastic nature of pitting corrosion has been recognised of the process to estimate the probability of failure of the pipeline.
since the mid-1930s [1,2]. One of the assumptions that can be In recent years, important advances have been made in model-
made about this damage is that it retains no memory of the past, ling pitting corrosion through Markov chains [5,8–10]. Bolzoni
so only the current state of the damage influences its future devel- et al. [8] used a continuous-time, three-state Markov process to
opment. This important characteristic allows pitting corrosion to model the first stages of localised corrosion considering three pos-
be categorised as a Markov process [3]. The discretisation of the sible states of the metal surface: passivity, metastability and local-
pit depth space in a finite or countable set of non-overlapping ised corrosion. Valor et al. [5] proposed a new stochastic model in
states makes pitting corrosion a good candidate for Markov chain which pit initiation is modelled as a Weibull process, while pit
modelling [3]. growth is modelled using a non-homogenous, linear growth Mar-
Provan and Rodriguez [4] used a non-homogenous Markov pro- kov process. The theory of extremes was used to combine both pro-
cess to model pit depth growth for the first time. The strengths and cesses and to adequately reproduce the experimental observations
limitations of the Provan and Rodriguez’s approach are discussed of pitting corrosion for a range of materials and environments. La-
in [5]. Later, Morrison and Worthingham [6] used a continuous ter, Zhang et al. [9] used this stochastic model to investigate the
time birth process with linear intensity k for determining the reli- pitting corrosion susceptibility of pure Mg and Mg alloys.
ability of high pressure corroding pipelines. These authors divided Particularly pertinent to the focus of the present study is the
the space of the load-resistance ratio into discrete states and model developed by Timashev et al. [10] to assess the conditional
solved numerically the Kolmogorov equations to find the intensi- probability of pipeline failure and to optimise the maintenance of
ties of transition between damage states. They showed that the operating pipelines. This model is based on the use of a continu-
estimated probability distribution function of the load-resistance ous-time, discrete state pure birth homogenous Markov process for
ratio closely reproduced the distribution obtained from field mea- stochastically describing the growth of corrosion-caused metal loss.
surements. Hong [7] refined the Morrison and Worthingham’s The initial conditions are determined by the (measured or assumed)
model by using the analytical solution of the Kolmogorov equa- corrosion depth distribution at t = 0. The probability pi(t) that the
tions for the same homogeneous continuous type of Markov pro- defect depth is in the i-th state at a given moment in time t is deter-
mined by the (measured or assumed) depth distribution at t. Using
standard methods, the intensity or rate (ki) of the unknown, time-
* Corresponding author. Tel.: +52 55 57296000x54205; fax: +52 55
57296000x55270. independent transition probabilities are calculated by iteratively
E-mail address: fcaleyo@gmail.com (F. Caleyo). solving the system of Kolmogorov’s forward equations:

0010-938X/$ - see front matter Ó 2009 Elsevier Ltd. All rights reserved.
doi:10.1016/j.corsci.2009.06.014
Author's personal copy

2198 F. Caleyo et al. / Corrosion Science 51 (2009) 2197–2207

8
< dp1 ¼ k1 p1 ðtÞ tion of soil properties. Thus, with a knowledge of the pit depth dis-
dt
ð1Þ tribution at a given point in time and the measured values of the
: dpi ¼ k p ðtÞ  k p ðtÞ
dt i1 i1 i i soil characteristics, the proposed Markov framework can be used
to predict the time evolution of the pitting depth and rate distribu-
The referenced authors illustrated that by estimating ki, the cor- tions. Real-life case studies, involving simulated and experimental
rosion depth and rate distributions can be estimated for any time pit depth distributions, are presented to illustrate the proposed
point. From that, the conditional probability of pipeline failure Markov chain modelling framework.
can be estimated, and repair and maintenance actions can be
scheduled. 2. Model development
The Markov chain approach to corrosion growth proposed by
Timashev and co-workers has many merits and strengths, but 2.1. Markov chain representation of pitting corrosion damage
some of its limitations are worthy of consideration. First, the rela-
tive complexity of the iterative method used to determine the tran- The reader is referred to [3] and [15] for a formalism of the the-
sition intensities ki requires that the pipe wall thickness be divided ory of Markov processes. In this section, only the definitions rele-
into a relatively small number of Markov states. Second, the use of vant to the focus of the present study are presented. As a matter
time-independent transition rates ki (time homogeneity condition) of general definition, it is assumed that the pipe wall thickness
implies that their estimated values actually represent the average has been divided in N discrete states and that the corrosion dam-
of the time-dependent intensities ki(t) over the selected period of age (pit) depth, at any point in time t, can be represented by a dis-
time. The time homogeneity condition entails that the staying time crete random variable D(t) with P{D(t) = i} = pi(t), i = 1, 2 . . . N.
of the corrosion penetration front in a given damage state is expo- Furthermore, it is assumed that the probability that the damage
nentially distributed. It also implies that the corrosion growth rate at the i-th state advances one state during a very short interval
is treated implicitly as a constant, while the corrosion depth is of time dt can be written as ki(t)d t + o(dt). For a continuous-time,
treated as a linear function of the exposure time. Due to the non- non-homogenous linear growth Markov process with intensities
linearity of the pit growth process, these assumptions hold true ki(t) = ik(t), the probability that the process presently in state i will
only if the estimations are made for long exposure times over rel- be in state j (j P i) at some later time obeys the following system of
atively short time spans. Finally, the solution to the Kolmogorov Kolmogorov’s forward equations (note the differences with respect
equations applies only for the particular conditions of a pipeline to Eq. (1)):
undergoing repeated in-line inspections. (
One important aspect that remains unsolved in the area of reli- dpi;j ðtÞ kj1 ðtÞpi;j1 ðtÞ  kj ðtÞpi;j ðtÞ; jPiþ1
ability-based pipeline integrity management is the accurate esti- ¼ ð2Þ
dt ki ðtÞpi;i ðtÞ
mation of future pit depth and growth rate distributions from a
(single) measured or assumed pit depth distribution. Obviously, For a Markov process defined by the system of equations (2),
such estimation can be carried out only if oversimplifications are the conditional probability of transition from the m-th state to
made (see, for example, Morrison and Desjardins’ discussion in the n-th state (n P m) in the interval (t0,t), that is,
Ref. [11]) or if additional information is available besides the pit pm,n(t0,t) = P{D(t) = njD(t0) = m}, can be obtained in closed form
depth distribution. For example, Alamilla and Sosa [12] recently (see page 304 in [15]):
developed a stochastic model to estimate the corrosion damage
 
velocity in operating pipelines. In their model, the results of the n1
in-line inspection are used not only to provide the corrosion depth pm;n ðt 0 ; tÞ ¼ efqðtÞqðt0 Þgm ð1  efqðtÞqðt0 Þg Þnm ð3Þ
nm
distribution at the time of the inspection but also to make infer-
ences about the intensity of the homogenous Poisson process used where:
to simulate the generation of corrosion defects. Z t
In the case of external pitting corrosion in underground pipe- qðtÞ ¼ kðt0 Þdt
0
ð4Þ
lines, it seems reasonable to postulate that the available predictive 0
models for pit growth as a function of the soil characteristics
[13,14] could provide the additional information necessary to pre- In words, Eq. (3) means that the pitting corrosion damage in-
dict, with a reasonable accuracy, the time evolution of the pitting crease in an interval of length t  t0 follows a negative binomial
depth and rate distributions. distribution NegBin(r,p) with parameters r=m and
In this work, a non-homogenous linear growth pure birth Mar- p ¼ ps ¼ efqðtÞqðt0 Þg . From pm,n(t0,t), it is possible to estimate the
kov process, with discrete states in continuous time, is used to probability distribution function f(t) of the damage rate t associ-
model external pitting corrosion in underground pipelines. A mod- ated with the damage process over the interval of length
el for pit growth recently developed by the authors [14] has been Dt = t  t0, when the damage depth is at the m-th state:
used to perform Monte Carlo simulations aimed at predicting the f ðt; m; t 0 ; tÞ ¼ pm ðt 0 Þpm;mþtDt ðt0 ; tÞDt ð5Þ
distribution of maximum pit depths as a function of the pipeline
age and the physicochemical characteristics of the soil. The bino- Now, from f(t;m,t0,t), it is straightforward to derive the pitting
mial closed form solution of the forward Kolmogorov equations rate probability distribution associated with the entire pit popula-
is adopted to express the probability of transition between pit tion using:
depth (Markov) states in a given interval of time. In so doing, the X
N
Markov-derived stochastic mean of the pit depth distribution is f ðt; t0 ; tÞ ¼ f ðt; m; t0 ; tÞ ð6Þ
supposed to be equal to the deterministic mean of the distribution m¼1

obtained through Monte-Carlo simulations. This supposition has Critical to the goal of this study is the fact that if the initial dam-
been made for different exposure times and different soil classes age state at t = ti is ni, so that D(ti) = ni, then the time-dependent,
defined according to soil physicochemical characteristics that are stochastic mean M(t) = E[D(t)] of the linear growth Markov process
easy to measure in the field. A multivariate regression analysis can be expressed as [3]:
has been used to obtain a predictive model for pit growth [14],
which was employed to specify the transition probability as a func- MðtÞ ¼ ni eqðtti Þ ð7Þ
Author's personal copy

F. Caleyo et al. / Corrosion Science 51 (2009) 2197–2207 2199

Under certain simple assumptions1, the stochastic mean of the pro- Let E[n  m]/Dt be the average damage rate in the time interval
cess can be assumed to be equal to the deterministic mean DðtÞ of a (t0, t0 + Dt) for a pit with depth in the m-th state at t0. In the Appen-
damage process for which the increment in the time interval Dt is: dix to this paper it is shown that the instantaneous rate of damage

 
 
 predicted by the stochastic model for this case is:
D DðtÞ ¼ kðtÞ DðtÞDt ð8Þ

 E½n  m m
where k can be interpreted as the deterministic intensity of the tðm; t0 Þ ¼ ! m ð15Þ
Dt Dt!0 t 0
damage
process.
If DðtÞ represents the deterministic mean pit depth at time t, Thus, the stochastically-predicted instantaneous damage rate
and if the power law model is considered to be an accurate deter- agrees with the rate predicted in Eq. (10) by the deterministic
ministic representation of the pit growth process, then one can model. The agreement of the stochastic and deterministic rates of
write [14]: damage supports the adequacy of the proposed in this work-Mar-

kov chain approach for pitting corrosion modelling.
DðtÞ ¼ jðt  t sd Þm ð9Þ Let us assume that the probability distribution of the corrosion
depth at t0 is known, that is P{D(t0)) = m} = pm(t0). For example, this
where j and m are the pitting proportionality and exponent param- can be obtained if the corrosion damage in the pipeline is moni-
eters, respectively; and tsd is the starting time of the pitting corro- tored using in-line inspection. In this case, t0 would be the time
sion damage. In systems where passivity breakdown and/or of the inspection and the value of the probabilities pm would be
inclusion dissolution are the prevalent mechanisms for pit initia- estimated from the ratio of the number of corrosion pits with
tion, tsd would represent the initiation time of stable pit growth. depths in the m-th state to the total number of pits observed. If
In the case of underground pipelines, this parameter would repre- the transition probability function pm,n(t0, t) is known, then the
sent the total elapsed time from pipeline commissioning to coating pit depth distribution at any future moment in time can be esti-
damage plus the time period in which the cathodic protection is mated using [3,15]:
effective in preventing or attenuating external pitting corrosion
after coating damage. X
n
pn ðtÞ ¼ pm ðt 0 Þpm;n ðt0 ; tÞ ð16Þ
It is easy to show

that the pitting rate, obtained by taking the m¼1
time


derivative of DðtÞ, obeys the functional form of Eq. (8) with
k ¼ m=s and s = t  tsd: Based on the foregoing discussion, it is postulated that the pit-


ting corrosion damage evolution in underground pipelines can be
d DðtÞ m  undertaken as follows. The measured or assumed pit depth proba-
¼ DðsÞ ð10Þ
dt s bility distribution at t0 is used as the initial corrosion damage dis-
tribution pm(t0). The transition probability function pm,n(t0, t),
Valor et al. [5] showed that, when Markov processes are used to
which is completely identified if the function q(t) is known, would
model pitting corrosion, a direct physical meaning can be given to
be obtainable if a predictive model were available for relating the
the functions k(t) and q(t), being related to the pit growth rate and
parameters tsd and m in Eq. (12) to the physicochemical conditions
pit depth, respectively. In the present work, k(t) and q(t) are also
of the soil [14]. From the results of the in-line inspection and the
related to the pitting rate and depth, respectively, though in a
knowledge of the local soil characteristics pertinent to such a pre-
slightly different manner than in [5]. It isassumed that the exper-
dictive model, it would be possible to estimate the pitting corro-
imentally observed deterministic mean DðtÞ of the pitting corro-
sion damage evolution. The next section presents the prediction
sion depth is equal to the Markov-derived stochastic mean pit
model developed to identify pm,n(t0, t) from in-field measured soil
depth M(t):
characteristics.


MðtÞ ¼ DðtÞ ð11Þ
2.2. Predictive model for the intensity of transition probabilities
Let dt1 represent the sojourn time of the corrosion damage in the
first state of the chain. In this case ni = 1 between tsd and tsd + dt1. The prediction of the function q(t) from the soil characteristics
If dt1 is significantly less than the simulation time span, it is easy relies upon the assumptions expressed by Eqs. (11) and (12). The
to show from Eqs. (7), (9) and (11) that the value of the function estimation of DðtÞ is carried out using a predictive model based
q(t) can be approximated as: on Eq. (9), which was developed by the authors for predicting max-
qðtÞ ¼ lnðjðt  tsd Þm Þ ð12Þ imum pitting damage in underground pipelines [14]. The experi-
mental details and corrosion data used to produce this model can
Furthermore, from Eq. (4), it follows that: be found elsewhere [14,16]. The soil classes considered in [14]

m are also used in this work. If DðtÞ represents the maximum pit
kðtÞ ¼ ð13Þ depth to be predicted and if the soil and pipe characteristics given
t  tsd
in Table 1 are treated as independent variables, then the pitting
This means that the intensity of the Markov process k(t) is in-
parameters in Eq. (9) can be estimated as follows [14]:
versely proportional to the exposure time, just like the determinis-
tic intensity of the damage process in Eq. (10). j ¼ j0 þ jph ph þ jcc cc þ jrp rp þ jre re þ jbc bc þ jsc sc ð17Þ
Now, it is relatively easy to show that the probability parameter m ¼ m0 þ mpp pp þ mwc wc þ mbd bd þ mct ct ð18Þ
ps ¼ efqðtÞqðt0 Þg in Eq. (3) can be expressed as:
 m The reader is referred to the original paper [14] for details of the
t 0  tsd
ps ¼ ; t P t0 P t sd ð14Þ derivation of Eqs. (17) and (18). There, an in-depth discussion on
t  t sd the relation between the model parameters and the environmental
variables can be found. It is worth noting that, in Eq. (18), the var-
iable ct is treated as a discreet ordinal variable, whose value is as-
1
signed according to a scoring model described in detail in Ref. [14]
A sufficient condition for the equality of the stochastic and deterministic means is
and summarized in a footnote to Table 1.
that for any positive integer q, the structure of the process starting from qm
individuals (states) is identical to that of the sum of q separated systems each starting Once the soil class has been identified and the values of the
from m (see page 159 in [3]). predictor variables have been measured, Eq. (18) can be used to
Author's personal copy

2200 F. Caleyo et al. / Corrosion Science 51 (2009) 2197–2207

Table 1
Variables and coefficients of the predictive model for pitting corrosion [14].

Variable or parameter Symbol (units) Coefficient name Coefficient or parameter value by soil class
Clay Clay loam Sandy clay loam All
Maximum pit depth dm(mm) – – – – –
Pipeline age t(years) – – – – –
Redox potentiala rp(mV) jrp 9.0  105 1.1  104 1.8  104 1.8  104
pH ph jph 5.9  102 1.2  101 6.4  102 6.5  102
Pipe/soil potentialb pp(V) mpp 4.9  101 4.6  101 5.1  101 5.2  101
Resistivity re(X–m) jre 2.2  104 3.0  104 2.1  104 2.6  104
Water content wc(%) mwc 3.7  103 1.7  102 4.5  104 4.6  104
Bulk density bd(g/ml) mbd 1.0  101 9.9  102 1.6  101 9.9  102
Chloride content cc(ppm) jcc 8.4  104 1.8  103 8.6  104 8.7  104
Bicarbonate content bc(ppm) jbc 1.3  103 4.9  104 6.8  104 6.4  104
Sulphate content sc(ppm) jsc 5.3  105 2.1  104 1.1  104 1.2  104
Coating typec ct mct 4. 7  101 57  101 4.3  101 4.3  101
Pitting starting time tsd(years) – 3.0 3.1 2.6 2.9
Constant prop. term j0(mm/yrm_0) – 5.5  101 9.8  101 6.0  101 6.1  101
Constant exponent term m0 – 8.8  101 2.8  101 9.6  101 8.9  101
Typical prop. factor jt(mm/yrm_t) – 0.178 0.163 0.144 0.164
Typical exponent factor mt – 0.829 0.793 0.734 0.780
a
Relative to the standard hydrogen electrode.
b
Relative to a Cu/CuSO4(sat.) reference electrode.
c
The pipeline coating conditions described by the ct variable and the assigned scores (in parenthesis) are: non-coated (1.0), asphalt-enamel-coated (0.9), wrap-tape-coated
(0.8), coal-tar-coated (0.7) and fusion-bonded-epoxy-coated (0.3) [14].

determine m and thereby identify q(t). The identification of q(t) al- the simulated maximum pit depth distributions was fitted to Eq.
lows the transition probability function pm,n(t0, t) to be determined (9) using the corresponding value of tsd given in Table 1. The model
for the soil and pipe characteristics in case. parameters obtained by this method are shown in the last two
It will be shown later that the proposed model can be used to rows of Table 1. They are assumed to be unbiased estimates of j
predict the progression of other pitting processes. For this to be and m for typical conditions for each soil category [14,16]. The mod-
done, the values of the pitting exponent and starting time must el output for the typical values of the predictor variables in a given
be known for the process. These can be obtained, for example, from soil category can be interpreted as the typical or average response
the analysis of repeated in-line inspections of the pipeline, from of the model in that soil.
the study of corrosion coupons or from the analysis of laboratory Fig. 2 shows the time evolution of q(t) (Fig. 2a) and ps (Fig. 2b)
tests. predicted from the exponent associated with the pitting damage
process for each soil category (mt in Table 1). Given that q(t) is com-
3. Results and discussion pletely determined by the extent of the damage (Eqs. (9) and (12)),
its value is unique for each soil category, being larger for higher soil
3.1. Modelling pitting corrosion in soils corrosivities. In addition, the probability ps is unique for each soil
type. Its value at a given moment in time t P t0 increases with
In a previous work [16], the authors used a Monte-Carlo simu- increasing t0 and decreases when the soil corrosivity and the
lation framework based on Eqs. (9), (17) and (18) to predict the length of the interval (t0, t) increases.
time evolution of the probability distribution of maximum pit The results shown in Fig. 2 determine the stochastic behaviour
depth caused by external corrosion in underground pipelines. of the pitting corrosion rate. Based on the properties of the Neg-
The observed distributions of the independent variables consid- Bin(m, ps) distribution, it follows that a reduction in the value of
ered in Table 1 were used as inputs in Monte-Carlo simulations ps with soil corrosivity entails larger mean and variance values of
of the pitting process for each soil category. The reader is directed the pitting rate in highly corrosive soils. The increase in ps with
to reference [16] for details of the used distributions and the t0 indicates that the mean and variance of the pitting rate decrease
Monte-Carlo simulation framework. Random values drawn from as the lifetime of the pitting damage increases. Finally, the form of
these distributions were used to evaluate Eqs. (9), (17) and (18) q(t) in Fig. 2a suggests that, for pits with equal lifetimes, the deeper
5000 times for different exposure times. The regression coefficients the pit, the smaller the value of ps and, therefore, the larger the
used in the simulations are those given in Table 1. mean and variance of the pitting rate.
For each soil category and exposure time considered, the simu- The properties of the stochastically predicted pitting rate are
lated pit depth distribution was fitted to the generalized extreme illustrated in Fig. 3 for the case in which the pitting damage occurs
value (GEV) distribution [17]: in clay soils. To produce this figure, it was assumed that pm(t0) = 1
for every m considered. Eqs. (3), (5) and (14) were used to compute
 h z  li1=f  f(t; m, t0, t) using the values of mt and tsd given in Table 1 for clay
GðzÞ ¼ exp  1 þ f ð19Þ
r soils. The pipe wall thickness was divided in 0.1-mm-thick states
so that the pitting damage is represented through Markov chains
where f, r and l are the shape, scale and location parameters, with states ranging from m = 1 to m = N = 100. Unless otherwise
respectively. specified, this scheme of discretisation of the pipe wall thickness
Fig. 1 shows the time evolution of the shape (Fig. 1a), scale is used hereafter to represent the pitting damage penetration.
(Fig. 1b) and location (Fig. 1c) parameters of the fitted GEV func- The results shown in Fig. 3 fully agree with the behaviour pre-
tions. In order to estimate the pitting proportionality and exponent dicted in the foregoing paragraphs for the damage rate as well as
factors associated with the typical (average) values of the predictor with the general body of knowledge about pitting corrosion rate.
variables in each soil category, the time evolution of the mean of The mean and variance of the pitting rate distribution decrease
Author's personal copy

F. Caleyo et al. / Corrosion Science 51 (2009) 2197–2207 2201

Fig. 2. Time evolution of (a) q(t) and (b) ps predicted from the exponent mt (Table 1)
associated with the pitting damage process for each soil category.

Fig. 1. Time evolution of (a) the shape, (b) the scale and (c) the location parameters
of the GEV distribution fitted to the Monte-Carlo-simulated pit depth distribution
for each soil category.

Fig. 3. Dependence of the pitting rate distribution f ðm; m; t0 ; t) on pit depth and
lifetime.
with increased pit lifetime. In Fig. 3, the distributions for the 15–20
year period (first two distributions from the left) have lower mean
and variance values than those predicted for the 5–10 year period the pitting rate mean and variance. Fig. 3 also reveals one of the
(the two distributions on the right of Fig. 3). For pits with similar advantages of the proposed Markov chain approach, that is, that
lifetimes, the pitting rate distribution mean and variance increase a unique probability distribution of pitting rate can be correlated
with increasing pit depth; the deeper the pit (larger m) the larger to the damage according to its depth and lifetime.
Author's personal copy

2202 F. Caleyo et al. / Corrosion Science 51 (2009) 2197–2207

The 10-year pit depth distribution predicted for each soil cate- At the same time, Fig. 4 reveals that the starting distributions are
gory using the Monte-Carlo framework was used as the initial unique for each soil class. The variation in the shape of the
damage distribution pm(t0 = 10) from which the damage distribu- Monte-Carlo-simulated pit depth distributions is adequately
tion pn(t) was predicted for different times using the proposed reproduced by the proposed Markov chain model for all soil types,
Markov chain model. Fig. 4 shows the GEV distribution fitted to from the beginning of the damage process and over the entire
pm(10) for each soil category. modelling time span considered, as can be seen from Figs. 5 and 6.
Fig. 5 shows the results of the Markov chain modelling of pitting Fig. 7 shows the 30-year pitting rate distribution for the All soil
corrosion in the least corrosive (sandy clay loam, Fig. 5b) and most category estimated using Eqs. (5) and (6) with t0 = 30 and t = 50
corrosive (clay, Fig. 5a) soils. The pitting exponent and starting years. It is worth noting that the application of Eq. (5) involves a
time used to carry out the predictions for each soil category were careful selection of the estimation interval Dt. On one hand, a
those given in Table 1. The Monte-Carlo-simulated pit depth distri- too-short (<5 years) estimation interval might result in a large
butions are represented by grey shaded histograms, while the con- uncertainty in the estimation of the pitting rate distribution. On
tinuous lines represent the Markov chain-predicted distributions the other hand, too-long estimation intervals would lead to exces-
pn(t). sively averaged estimates of the pitting rate distribution over the
The goodness-of-fit of the results shown in Fig. 5, and of those selected time span. A 20-year estimation interval was selected in
not shown here for the other soil classes, was assessed using the order to predict the long-term distribution of the pitting rate in
Kolmogorov–Smirnov test. For the sake of simplicity, the one-sam- this category and to compare it with the long-term rate recently re-
ple test was conducted based on the null hypothesis that each ported by the authors in [16]. The adequacy of the Markov chain
Monte-Carlo-estimated distribution (empirical sample) comes approach for estimating the pitting corrosion rate by means of
from the corresponding Markov chain-modelled distribution Eqs. (5) and (6) is supported by the good agreement between the
(underlying reference distribution). In all the analyzed cases (a to- two predictions shown in Fig. 7.
tal of 20, five exposure times for each one of the four soil classes), it
was not possible to reject the proposed null hypothesis at 5% sig- 3.2. Modelling pitting corrosion from repeated in-line inspections
nificance level. This could be interpreted as evidence that there is
no reason to reject the assertion that the Monte-Carlo-estimated Another application example of the proposed Markov chain
and Markov chain-modelled distributions are close enough to con- model is shown in Fig. 8. An 82-km-long operating pipeline, used
sider the proposed Markov model accurate. to transport sweet gas since its commissioning in 1981, was in-
The proposed Markov chain model is capable of reproducing the spected in 2002 and 2007 using magnetic flux leakage in-line
time evolution of the pit depth distribution accurately. The quality inspection (ILI). The pipeline in case is coal-tar coated with a wall
of the fitting in the clay loam and in all soil categories was very thickness of 9.52 mm (0.37500 ). The distributions shown in Fig. 8a
similar to that observed in Fig. 5. This can be corroborated by the were obtained by calibrating the ILI tools using a methodology de-
results shown in Fig. 6, which shows the time evolution of the scribed elsewhere [18]. The use of calibrated pit depth distribu-
mean (Fig. 6a) and standard deviation (Fig. 6b) of the Monte-Car- tions allows minimizing the negative impact on the results of the
lo-simulated and Markov-modelled pit depth distributions for analysis of the measurement errors associated with the inspection
the soils of interest. tools. The cross-hatched histogram in Fig. 8a shows the depth dis-
In addition to the qualitative (Fig. 5) and quantitative (Fig. 6) tribution of NT02 = 3577 pits2 located and measured by ILI at the
accuracy of the Markov chain predictions, the most striking feature beginning of 2002. The grey shaded histogram shows the depth dis-
of these results is that the model is able to reproduce the changes tribution of NT07 = 3851 pits located and measured by ILI in mid-
observed in the shape of the pitting depth distribution with 2007.
increasing exposure time for all soil categories. Fig. 1a shows that, In order to apply the proposed model, the empirical depth dis-
depending on the exposure time, any of the three maximal ex- tribution observed in 2002 was used as the initial distribution so
treme value distributions, i.e., Weibull, Fréchet or Gumbel, arises that t0 = 21 years, Dt = 5.5 years and pm(t0 = 21) = Nm/NT02, with
as the best fit to the pitting depth data in the investigated soils. Nm being the number of pits with measured depth in the m-th
state. The soil characteristics along the pipeline were assumed to
be those of the All category. Therefore, the values of tsd and m were
taken as 2.9 years and 0.780, respectively (Table 1). The application
of Eqs. (3), (14) and (16) produced the Markov chain-predicted pit-
ting depth distribution in Fig. 8a, shown with the thick line.
The good agreement between the empirical pit depth distribu-
tion observed in 2007 and the Markov chain-modelled distribution
also points to the accuracy of the proposed model. The validation of
its capability to correctly predict the evolution of the pit depth dis-
tribution in this pipeline points to the fact that the proposed model
can be used to predict its reliability over a given time span. For
example, the corrosion rate distribution f(t;t0, t) associated with
the entire population of pit defects in the 2002–2012 period can
be predicted using Eqs. (5) and (6), and then, it can be used to esti-
mate the evolution of the pitting damage starting from any year in
this period. Fig. 8b shows the pitting rate distribution predicted for
the entire population of pits in the pipeline using this approach
with: t0 = 21 years, t = 31 years, m = 0.780 and tsd = 2.9 years. The
obtained pitting rate distribution is very close to a GEV distribution

Fig. 4. GEV distribution fitted to the 10-year Monte-Carlo-simulated pit depth


2
distributions, used as the initial pit depth distribution for the Markov chain In both ILI reports, pits were identified as corrosion-caused metal losses with a
modelling for each soil category. The Markov chain states are 0.1-mm-thick. diameter equal to or less than two times the pipe wall thickness.
Author's personal copy

F. Caleyo et al. / Corrosion Science 51 (2009) 2197–2207 2203

Fig. 5. Results of the Markov chain modelling of pitting corrosion compared to the Monte-Carlo-simulated distribution in (a) clay and (b) sandy clay loam soils. The Markov
chain states are 0.1-mm-thick.

(Eq. (19)) with the shape parameter f being negative, although it is ilies of the GEV distribution seem to be appropriate for describing
very close to zero. Therefore, the Weibull and the Gumbel subfam- the pitting rate in the pipeline over the selected estimation period.
Author's personal copy

2204 F. Caleyo et al. / Corrosion Science 51 (2009) 2197–2207

Fig. 8. Results of the Markov chain modelling of pitting damage evolution from
Fig. 6. Comparison of the time evolutions of (a) the mean and (b) standard repeated in-line inspections. (a) Measured and modelled pit depth distributions. (b)
deviation of the Markov chain-modelled and Monte-Carlo-simulated pit depth Predicted pitting corrosion rate for the 2002–2012 period. The Markov chain states
distributions for all soils. The Markov chain states are 0.1-mm-thick. are 0.1-mm-thick.

3.3. Modelling pitting corrosion in immersion test coupons

With the purpose of illustrating the versatility of the proposed


model, the final application example involves the modelling of
laboratory-induced pitting corrosion in coupons of 1  1 cm2
from API-5L X52 pipeline steel immersed for different periods
of time in a solution with characteristics close to those found in
soils. The experimental details and the results of these pitting
experiments were reported recently by Rivas et al. [19]. The grey
shaded histograms shown in Fig. 9a represent the pit depth distri-
butions of the stable pits induced in the investigated coupons for
the selected immersion times (1, 3, 7, 15, 21 and 30 days). The
observed time evolution of the mean pit depth is shown in
Fig. 9b. It was fitted to a power law function with a pitting expo-
nent m = 0.4175, while the pitting starting time was assumed to
be negligible (tsd = 0) with respect to the duration of the immer-
sion tests.
The coupon thickness was divided in non-overlapping 1-lm-
thick discrete states to represent the pitting damage through Mar-
kov chains with states ranging from m = 1 to m = N = 150. The
empirical pit depth distribution for a 1-day exposure was used as
Fig. 7. Comparison of the Monte-Carlo-simulated [16] and Markov chain-predicted
pm(t0). The total number of pits in the 1-day coupons was
long-term pitting corrosion distributions in the All soil category. The Markov chain NT1 = 1019, and the non-parametric mean and standard deviation
states are 0.1-mm-thick. of this distribution were found to be 17.6 and 5.5 lm, respectively
Author's personal copy

F. Caleyo et al. / Corrosion Science 51 (2009) 2197–2207 2205

Fig. 9. Empirical and Markov chain-modelled distributions for immersion pitting corrosion tests in pipeline steel. (a) Pit depth distributions. (b) Time evolution of the means.
(c) Time evolution of the standard deviations. The Markov chain states are 1-lm-thick.

(see Table 3 in [19]). From pm(t0 = 1) = Nm/NT1, where Nm is the The relatively good quality of the Markov chain estimations is
number of pits with depth in the m-th state, the damage distribu- further corroborated in Fig. 9b and c, where the time evolutions
tion pn(t) was predicted for the immersion times at which the of the mean and standard deviation of the Markov chain-predicted
empirical depth distribution had been obtained. pit depth distributions are compared with the empirical values. It
The pit depth probability distributions predicted for the immer- can be seen that, for all immersion times, the predicted mean
sion times greater than 1 day using the proposed Markov chain and standard deviation are within the 95% prediction bands of
model are shown in continuous thick lines in Fig. 9a. Qualitatively, the power law model fitted to the experimental data. It is impor-
the agreement between the observed and Markov chain-predicted tant to underline that the largest disagreement is observed
pit depth distributions is reasonably good, even when some of the between the experimental and modelled mean values for the
experimental depth distributions (t P 15 days) are not strictly uni- 15- and 21-day pit depth distributions (Fig. 9b). This disagreement
modal. The bimodal character of these distributions was explained is most likely related to the fact that the empirical distributions for
by Rivas et al. [19] on the basis of the physicochemical dependence these exposure times are associated with a noticeably bimodal
among the growing pits in the test coupons. However, this fact character due to the pit interdependence [19]. Nonetheless, the re-
does not seem to impair the suitability of the proposed Markov sults shown in Fig. 9 are interpreted by the authors as a confirma-
chain framework for modelling the results of these immersion tion of the accuracy and adaptability of the proposed Markov chain
tests. modelling framework.
Author's personal copy

2206 F. Caleyo et al. / Corrosion Science 51 (2009) 2197–2207

4. Concluding remarks pretation, when Markov chains are used for modelling pitting cor-
rosion, is the aim of ongoing research by the authors.
A new Markov chain model for pitting corrosion has been devel-
oped and validated using both synthetic and experimental pitting Acknowledgement
corrosion data. The use of a continuous-time, non-homogenous lin-
ear growth (pure birth) Markov process for modelling pitting cor- This work was performed during a stay of A. Valor at the Na-
rosion is particularly attractive due to the existence of a closed tional Polytechnic Institute (ESIQIE-IPN), Mexico, under research
form solution (Eq. (3)) to the system of Kolmogorov’s forward project No. 428817805. The comments provided by the reviewers
equations that describe this type of Markov process. The use of this are greatly appreciated.
solution avoids a reduction of the number of states for the sake of
mathematical simplicity or the requirement of assumptions about Appendix A
the staying time of the pitting damage in a given state.
Critical to the development of the proposed model is the The stochastic expression for the instantaneous corrosion rate
assumption that the Markov chain-derived stochastic mean of can be obtained based on the properties of the negative binomial
the pitting damage equals the deterministic, empirical mean of distribution [20] and the expression for the probability ps.
the process. Such an assumption allows the transition probability Let us assume that the pit depth is in the m-th state at t0. In this
function to be identified only from the pitting starting time and case, the average value of the damage rate over an interval Dt,
the pitting exponent (Eq. (14)). Furthermore, at least for the situa- starting at t0, can be approximated by:
tions in which the power law model accurately describes the pit-
E½n  m
ting process, this assumption requires that the functional form of tðm; t0 ; DtÞ ¼ ðA-1Þ
the stochastic and deterministic instantaneous pitting rates also
Dt
agree. This supports the idea that the intensities of the transitions According to Eq. (3), pm,n(t0, t0 + Dt) has a negative binomial dis-
in the Markov process are closely related to the pitting damage tribution NegBin(r, p) with parameters r = m and p = ps. Therefore,
rate. its expected value over Dt is given by [20]:
The fact that the transition probability function depends only E½n  m ¼ mð1  ps Þ=ps ðA-2Þ
on the pitting starting time tsd and the pitting exponent m can be
explained by considering this function to be directly related to For the conditions considered in this deduction, the expression
the corrosion rate (see Eqs. (5), (14) and (15)). The pitting propor- for ps is:
tionality coefficient j in the power law model for pit depth appears
ps ¼ efqðt0 þDtÞqðt0 Þg ðA-3Þ
implicitly in the model when the initial distribution pm(t0) is mea-
sured or assumed. One of the advantages of the Markov chain ap- If expressions (A-2) and (A-3) are substituted into (A-1), the
proach over deterministic [14] and other stochastic models [12] for average damage rate over Dt becomes:
pitting corrosion is that the Markov chain model is able to capture
ðefqðt0 þDtÞqðt0 Þg  1Þ
the dependence of the pitting rate on the pit depth and lifetime. tðm; t0 ; DtÞ ¼ m ðA-4Þ
Dt
Therefore, it allows for an estimation of not only the probability
distribution of the pitting rate associated with the entire pit popu- Based on the expression for q(t) (Eq. (12)) and using the fact
lation but also such a distribution for a subpopulation within spe- that, for sufficiently small values of x, the exponential function ex
cific lifetime and depth ranges. can be approximated by 1 + x, it can be shown that:
A knowledge of the dependence of the pitting starting time and  
mm Dt
exponent on the environmental characteristics determining the tðm; t0 ; DtÞ ! ln 1 þ ðA-5Þ
Dt!0 Dt t0
pitting damage allows for a generalisation of the model for a range
of pitting environments. This was illustrated in the present study This expression can be further simplified by taking into account
by modelling external pitting corrosion in underground pipelines that, for sufficiently small values of x, the function ln(1 + x) can be
in a range of soils. Importantly, this approach could be extended approximated by x. Therefore:
to investigate pitting corrosion in laboratory conditions. For exam- m
ple, the effects of chloride concentration and pH on pitting corro- tðm; t0 Þ ! m ðA-6Þ
Dt!0 t0
sion evolution in low carbon steels and aluminium are currently
being investigated by the authors using the proposed Markov chain This expression represents the stochastically-derived instanta-
model. neous corrosion rate for a pitting damage, whose depth is in the
Two questions that were not answered in this study are of par- m-th state at the moment in time t0.
amount significance for future work: How can the validity of the
assumption about the equality of the stochastic and deterministic References
pit depth means be assessed? What kind of pit populations can be
[1] T. Shibata, Statistical and stochastic approaches to localized corrosion,
described by means of the proposed Markov chain model? Corrosion 52 (1996) 813–830.
In the theory of stochastic processes, the answer to the first of [2] P.M. Aziz, Application of the statistical theory of extreme values to the analysis
these questions is known for situations in which counting pro- of maximum pit depth data for aluminium, Corrosion 13 (1956) 495–506.
[3] D.R. Cox, H.D. Miller, The Theory of Stochastic Processes, first ed., Chapman &
cesses are considered (see page 189 in [3]). In the case of pitting Hall, CRC, Boca Raton Florida, 1965.
corrosion, the prediction and interpretation of the equality of the [4] J.W. Provan, E.S. Rodrı́guez III, Part I: Development of Markov description of
stochastic and deterministic pit depth means are not straightfor- pitting corrosion, Corrosion 45 (1989) 178–192.
[5] A. Valor, F. Caleyo, L. Alfonso, D. Rivas, J.M. Hallen, Stochastic modelling of
ward. Still, the equality of the means seems to hold true for all pitting corrosion: a new model for initiation and growth of multiple corrosion
the cases studied in this work. The second question seems to be re- pits, Corros. Sci. 49 (2007) 559–579.
lated to the existence and type of the limiting distribution of the [6] T.B. Morrison, R.G. Worthingham, Reliability of high pressure line pipe under
external corrosion, in: Proc. 11th ASME Int. Conf. Offshore Mechanics and Artic
state occupation probabilities after a sufficiently long period of Eng., vol. V, Part B, Book No. H0746B, Calgary, Canada, June 7–12, 1992, pp.
time [3]. Both aspects are of a fundamental nature, and their inter- 401–408.
Author's personal copy

F. Caleyo et al. / Corrosion Science 51 (2009) 2197–2207 2207

[7] H.P. Hong, Inspection and maintenance planning of pipeline under external [13] Y. Katano, K. Miyata, H. Shimizu, T. Isogai, Predictive model for pit growth on
corrosion considering generation of new defects, Structural Safety 21 (1999) underground pipes, Corrosion 59 (2003) 155–161.
203–222. [14] J.C. Velazquez, F. Caleyo, A. Valor, J.M. Hallen, Predictive model for pitting
[8] F. Bolzoni, P. Fassina, G. Fumagalli, L. Lazzari, E. Mazzola, Application of corrosion in buried il and gas pipelines, Corrosion 65 (2009) 332–342.
probabilistic models to localized corrosion study, Metallurgia Italiana 98 [15] E. Parzen, Stochastic Processes, Classics in Applied Mathematics, Society for
(2006) 9–15. Industrial and Applied Mathematics (SIAM), PA, 1999.
[9] T. Zhang, L. Xiaolan, S. Yawei, M. Guozhe, W. Fuhui, Electrochemical noise [16] F. Caleyo, J.C. Velazquez, A. Valor, J.M. Hallen, Probability distribution of pitting
analysis on the pit corrosion susceptibility of Mg–10Gd–2Y–0.5Zr, AZ91D alloy corrosion depth and rate in underground pipelines: a Monte Carlo study,
and pure magnesium using stochastic model, Corros. Sci. 50 (2008) 3500– Corros. Sci. (2009), doi:10.1016/j.corsci.2009.05.019.
3507. [17] S. Coles, An Introduction to Statistical Modelling of Extreme Values, Springer
[10] S.A. Timashev, M.G. Malyukova, L.V. Poluian, A.V, Bushiskaya, Markov Series in Statistics, Springer-Verlag, London, 2001.
description of corrosion defects growth and its application to reliability [18] F. Caleyo, L. Alfonso, J.H. Espina, J.M. Hallen, Criteria for performance
based inspection and maintenance of pipelines, in: Proc. 7th ASME Int. Pipeline assessment and calibration of in-line inspections of oil and gas pipelines,
Conf. IPC2008, Calgary, Canada, Paper IPC2008-64546, September 26–29, Meas. Sci. Technol. 18 (2007) 1787–1799.
2008. [19] D. Rivas, F. Caleyo, A. Valor, J.M. Hallen, Extreme value analysis applied to
[11] T.B. Morrison, G. Desjardin, Determination of corrosion rates from single in- pitting corrosion experiments in low carbon steel: comparison of block
line inspection of a pipeline, NACE One day seminar, Houston, TX, December 1, maxima and peak over threshold approaches, Corros. Sci. 50 (2008) 3193–
1998. 3204.
[12] J.L. Alamilla, E. Sosa, Stochastic modelling of corrosion damage propagation in [20] E. Parzen, Modern Probability Theory and its Applications, John Wiley and
active sites from field inspection data, Corros. Sci. 50 (2008) 1811–1819. Sons, Inc., Wiley Classics Library Edition, New York, 1992.

View publication stats

You might also like