You are on page 1of 6

1 Copyright 2012 by ASME

WHY RELIABILITY-BASED APPROACHES NEED MORE REALISTIC CORROSION


GROWTH MODELING



Luc Huyse
Chevron ETC
Houston, TX, USA
Katherine A. Brown
Chevron ETC
Houston, TX, USA





ABSTRACT
Deterministic design and assessment methods are by
definition conservative. Although no claim is made regarding
the actual reliability level that is achieved using deterministic,
i.e. safety-factor based approaches, the safety factors have been
selected such that generally sufficient conservatism is
maintained. Reliability-based methods aim to explicitly
quantify the aggregated conservatism in terms of failure
probabilities or risk. Accurate reliability estimates are not
possible without accurate computational prediction models for
the limit states and adequate quantification of the uncertainties
in both the inputs and model assumptions. Although this
statement may seem self-evident, it should not be made light-
heartedly. In fact, just about every analysis step in the pipeline
integrity assessment procedures contains an inherent, yet
unquantified, level of conservatism. One such example is the
application of a maximum corrosion growth rate that is
constant in time.
A reliability-based framework holds the promise of a more
consistent and explicitly quantified safety level. This ultimately
leads to higher safety efficiency for an entire pipeline system
than under safety factor based approaches. An accurate
prediction of the true likelihood of an adverse event is
impossible without significant research into determining and
understanding the, usually conservative, bias in the engineering
models that are currently employed in the pipeline integrity
state-of-the-practice. This paper highlights some of the
challenges that are associated when porting the maximum
corrosion rate approach used in a deterministic approach to a
reliability-based paradigm. Issues associated with both defect
and segment matching approaches will be highlighted and a
better corrosion growth model form will be proposed.
INTRODUCTION
In order to assess the risk of pipeline failure due to a leak
or burst, information about the current state of the pipeline must
be combined with a corrosion rate that models how quickly the
anomalies grow. Information about the current state of the
pipeline can be inferred from inspections and is a critical
ingredient in the integrity management decision making
process. Inline inspection results are subject to various sources
of uncertainty.
The corrosion growth is a critical input to pipeline integrity
management. Corrosion rates can be estimated from standards
and industry guidelines, correlation and regression models
applied to similar service conditions, coupon losses, indirect
measurements, or a comparison of the results of multiple inline
inspections. This paper discusses some aspects related to the
accurate estimation of corrosion rates from the results of
multiple inline inspections.
In a deterministic approach one often assumes or postulates
a maximum corrosion rate. For reliability-based approaches,
two methods have been commonly proposed: a segment-based
and defect-based corrosion rate. These approaches make some
fundamentally different assumptions about the uncertainty
model of the corrosion growth and these differences are often
not explicitly acknowledged or perhaps mis-understood. Each
of these approaches is discussed in a little more detail in the
next sections. Shortcomings and unintended consequences of
each of these models when applied in a reliability-based
framework will be discussed.
Proceedings of the 2012 9th International Pipeline Conference
IPC2012
September 24-28, 2012, Calgary, Alberta, Canada
IPC2012-90319
2 Copyright 2012 by ASME

NOMENCLATURE

Table 1: list of symbols and their definitions
Symbol Definition
d true feature size (depth, length or width)
e sizing uncertainty (error)
i.i.d. independent and identically distributed
m measured feature size (reported by ILI tool)
pdf probability density function
Covar(.) covariance of two variables
E(.) expected value of a random variable
ILI inline inspection
Var(.) variance of a random variable
< . > Mean or expected value of a random variable

DETERMINISTIC MAXIMUM CORROSION RATE
NACE proposed a maximum corrosion rate for External
Corrosion Direct Assessment assessments of 0.4mm/yr
(16mpy), which corresponds to about 80
th
percentile of the
corrosion growth rate distribution. In absence of other data, this
rate has been used as a default corrosion rate in ECDA
applications.


Figure 1: Illustration of the near-exponential nature of the maximum
pit depth distribution in an ILI run.
Feature depth distributions from in-line inspection data
often follow a near-exponential distribution. The plot in Figure
1 illustrates this point. Although the 80
th
percentile of the depth
distribution is near 20%WT, the 99
th
percentile of the max wall
loss is twice as deep and the maximum observed wall loss can
readily be a factor 3 higher, depending on the number of
anomalies.
A deterministic approach is quite straightforward in
application: once the maximum rate has been selected or
calculated, all features are assumed to grow at this rate. In
practice, this means that the relative ranking of features remains
preserved throughout the entire remaining life of the pipeline:
what was the most severe anomaly at a certain point in time,
will remain so (until the anomaly is repaired or the pipe
segment is replaced). Empirical evidence from in-line
inspection records suggests that this is not realistic (see Dawson
and Kariyawasam, 2009). The main objective of the
deterministic assessment methods is to provide a minimum
level of safety; accuracy of the prediction is really a secondary
(although obviously desirable) concern.
PURPOSE OF CORROSION GROWTH ASSESSMENT
A variety of methods have been proposed for corrosion
growth rate analysis. A corrosion growth assessment may be
performed for various reasons and each subsequent level
requires an additional level of fidelity (Figure 2). Several
researchers have looked at using statistical approaches to
determine if active growth is occurring (Rust and Johnson,
2001, Worthingham et al.. 2002, Dawson et al. 2009). In this
case the primary objective of the analysis is to determine
whether corrosion is active; in other assessments the objective
may be to set re-inspection intervals or compute the remaining
life of the pipeline.

Figure 2: Increasing levels of fidelity in corrosion growth rate
assessment
Several factors contribute to the uncertainty in the
corrosion growth rate assessment: natural variability of the
corrosion growth rate process, misspecification of the corrosion
growth rate model, inaccuracy in the inspection data. It is often
difficult, if not impossible, to separate each of these effects
when performing an analysis.
Iscorrosionactive?
Estimategrowthrates
Setreinspectionintervals
Accuratelycomputeremaininglife
3 Copyright 2012 by ASME
SEGMENT MATCHING CORROSION RATE MODELING
The segment matching method (see Dawson and Wharf,
2008) estimates the corrosion growth statistics over an entire
segment. Its application is most appropriate when the number
of features remains fairly constant and no localization of the
corrosion growth is expected. In this case one determines the
mean growth rate by dividing the difference between the
averages of the maximum pit depths during each inspection
under consideration by the time elapsed between inspections.
A time-averaged corrosion rate is given by the equation:

< CR >=
< J(t
2
) > -< J(t
1
) >
t
2
-t
1

where <CR> stands for average corrosion rate, <d> is the
average, exact anomaly depth without sizing uncertainty, t
1

and t
2
are the times of the first and second ILI, respectively.
Note that this average growth rate does not provide any
information about the growth rate distribution for individual
defects. Instead, the standard deviation associated with the
segment matching growth rates represents the uncertainty
associated with the average growth rate and is really more a
reflection of the intrinsic tool accuracy and the number of
anomalies in each segment than a representation of the actual
uncertainty on the growth rate estimate. The standard deviation
represents the effect of the ILI sizing error on the estimate of
the average corrosion growth only and does not represent the
inherent variability in the corrosion growth process. It is
therefore not appropriate to use this standard deviation to
estimate the variance of the corrosion rates.
Another problem is that the corrosion may be inactive in
many locations, either because the CP is providing adequate
protection (for external corrosion) or because the build-up of
corrosion product passivates further corrosion (internal
corrosion). Therefore, many features do not grow, whereas
some features grow quite rapidly. The lumping of both groups
into a single, uni-modal distribution necessarily underestimates
the maximum growth rate.
ANOMALY MATCHING CORROSION RATE MODELING
In the anomaly matching method, individual features are
first matched and corrosion growth rates are calculated for all
individually matched pairs. The corrosion growth rate
distribution is defined by the distribution of the growth rates
among these matched pairs. Subsequent reliability analysis is
often performed by assuming that future growth is modeled by
this distribution. Although intuitively appealing, significant
problems arise when applying this approach in practice.
As a first problem: the random sizing error has a large
impact on the computed distribution and overestimates the
corrosion growth rate (see Huyse and van Roodselaar, 2010).
An ILI inspection does not result in the exact anomaly size
but in a measured value which is an approximate indication
thereof: m = d + e. The time-averaged corrosion rate, computed
directly from the matched ILI data pairs is:

CR
uw
=
m(t
2
) -m(t
1
)
t
2
-t
1
=
J(t
2
) +c(t
2
) -J(t
1
) - c(t
1
)
t
2
-t
1
= CR
tuc
+
c(t
2
) -c(t
1
)
t
2
-t
1


Not all features grow at the same rate and it therefore
makes sense to compute the statistics of the observed corrosion
rates in the matched pairs. The expected (or average) value of
the growth rate is:

E|CR
uw
] = E _
m(t
2
) - m(t
1
)
t
2
- t
1
_
= E|CR
tuc
] +
E|c(t
2
)] - E|c(t
1
)]
t
2
- t
1

which indicates that the average corrosion rate obtained by the
ILI measurements is unbiased (i.e. equal to the true average
corrosion rate) as long as the sizing errors have the same
average value during both inspections
1
.
The variance of the corrosion rate distribution is:

Ior|CR
uw
] = Ior _
m(t
2
) - m(t
1
)
t
2
- t
1
_
= Ior|CR
tuc
] + Ior _
c(t
1
)
t
2
- t
1
_
+ Ior _
c(t
2
)
t
2
- t
1
_ +2Co:or _CR
tuc
,
c(t
2
)
t
2
- t
1
_
- 2Co:or _CR
tuc
,
c(t
1
)
t
2
- t
1
_
- 2Co:or _
c(t
2
)
t
2
- t
1
,
c(t
1
)
t
2
- t
1
_

Since the covariance of the true corrosion rate and the
random component of the measurement error is zero and if
the same tool and sizing model is used the variance of the
first and second inspections are identical with independent
random error components, this expression can in many practical
cases be simplified to:
Ior|CR
uw
] = Ior|CR
tuc
] +2Ior _
c(t
1
)
t
2
-t
1
_

The variance equation indicates that the sizing uncertainty


increases the apparent variability of the corrosion growth rate:
Ior|CR
uw
] Ior|CR
tuc
]. This is readily understood: the
measurement uncertainty adds noise to the actual corrosion
growth which gets translated as additional scatter into the
corrosion rate statistics. Unless there is considerable feature
growth between subsequent inspections, the noise terms in
Ior|CR
uw
] will often exceed the true variance.

1
Note that this expression applies to the complete feature distribution; in
many cases only features that exceed a specified threshold are reported and this
reporting threshold will introduce some bias in the average raw corrosion rate
which must be corrected for.
4 Copyright 2012 by ASME
A second problem is that the computed growth in many
pairs is negative, either due to mismatching of pairs or due to
sizing uncertainty. Some authors have proposed to account for
these measurement uncertainties by adding more conservatism
to the growth rate distribution computed on the basis of the
differences between the raw ILI data [Nessim et al., 2008]. It
has been shown that if the sizing error is assumed independent
of the exact anomaly depth, then it is in fact, not independent of
the reported feature size but that for a generally unbiased tool
the largest ILI calls tend to be overestimates more often than
not (see Huyse and van Roodselaar, 2010 for a theoretical
development and Huyse et al, 2011 and Haines et al, 2012 for
validation based on ILI data). Consequently, the computed
corrosion growth rate distribution needs to account for these
effects. Unfortunately, the corrosion rate distributions created
following the approach in Nessim et al. (2008) are
mathematically inconsistent with the results of the ILI runs:
growing anomalies observed during the first ILI run using the
proposed distribution does not result in the distribution
observed during the second ILI run.
EFFECT OF DISTRIBUTION TYPE
As indicated before the distribution that models the
individual pit depths is often well described by a censored
exponential or Generalized Pareto distribution (GPD) although
the apparent shape parameter may change over time.
In practical applications of probabilistic corrosion growth
rate assessments, a normal or truncated normal distribution has
sometimes been used to describe the corrosion growth rate
whereby the mean and standard deviation are computed directly
from the matched defect pairs.
Repeated addition of a Normally distributed growth, which
is independent of the current anomaly size, to the anomaly
depth that is recorded at the time of inspection, will ultimately
result in a Normally distributed anomaly depth curve. If this
growth model were correct, the depth distribution will
necessarily become closer aligned with a normal distribution
over time. This is not supported by overwhelming evidence
from the field. On the contrary, repeated ILI inspections often
support the existence of a single, stable distribution for the
maximum pit depths over time.
Because of this discrepancy, predictions that rely on the
use of a normally distributed growth increment must therefore
be limited to very small depth increases to remain credible.
In addition, care should be taken to properly account for
the effects of the ILI reporting threshold when estimating
corrosion growth rates.
MODELING OF CORROSION GROWTH OVER TIME
For many of the probabilistic models currently used,
reliability predictions are often made assuming that
independent growth increments occur from year to year. This
assumption does, however, lead to results that are untenable
from a practical point of view.
It is well understood that the pipeline integrity at time of
inspection is generally governed by either the deepest pit (for a
leak limit state) or the anomaly with lowest calculated burst
pressure (which is itself governed by a function of depth and
length). If one assumes a constant corrosion growth rate, the
integrity far into the future, however, is typically governed by
the fastest growing pit. Accurately predicting the corrosion
growth rate is therefore a necessary condition for good integrity
management.
It was previously noted that many pipelines have pit depths
that are well described by an exponential-like distribution type.
Consider Figure 1: whereas only a single feature may be near
60%WT, many thousands of features may be around 10%
depth. If all growth increments are independent and identically
distributed (i.i.d.), the distribution of the maximum growth rate
for all these features of a given size d
i
is:

F
max
(x | d = d
i
) = [F
X
(x | d = d
i
)]
N
= [F
X
(x)]
N


where x is the corrosion growth rate, N is the number of
features with depth equal to d
i
; and F is the distribution function
for the growth rate of a feature. By definition F does not depend
on the actual depth d
i
(since the growth is assumed to be
independent of the current depth).
If an exponential distribution is assumed, increasingly
more anomalies will be found at smaller depths. E.g. there may
only be a single anomaly at 25%WT, whereas there could by
many anomalies at, say, 15%WT. Each of these anomalies is
expected to grow with a corrosion rate x that is modeled by the
random variable distribution F
X
(x). Each pit has an equal
chance to grow at a particular rate x
i
. However, since there are
many more pits at 15%WT than at 25%, one of the 15%WT
pits will likely grow at a much faster rate than the 25%WT pit
and, consequently, become the dominant pit over time.
By means of example, consider a pipeline with exponential
anomaly depth distribution (mean value = 3%WT) and assume
that the maximum depth recorded in the line is 25%WT. Based
on the previous discussion, we already know that the
distribution type cannot be normal but should be more closely
aligned with an exponential type distribution. Also assume that
the line is 5 years old (i.e. max time-averaged corrosion rate is
5%WT, but the average time-averaged corrosion rate is
0.6%WT/yr). The median predicted value for the fastest
growing pit that was 5, 10, , 25%WT at time of inspection is
shown in Figure 3 as function of time. The median maximum
corrosion rate at any given depth is computed by setting F
max

equal to 0.5 and solving for the corrosion rate. It can be seen
that, with this model, the deeper pits statistically speaking
grow at lower rates than the faster-growing pits at lower depths.

5 Copyright 2012 by ASME

Figure 3: Median predicted value as function over time. The corrosion
growth rate is the maximum rate obtained for all anomalies at 5, 10
25%WT depth
In a probabilistic sense, this would suggest that over time
the risk contribution of the shallower anomalies increases,
which is clearly not the case and is contrary to the experience in
the field. Indeed, the plot in Figure 3 would suggest that some
of the shallower defects should be repaired first because they
might grow at a much faster rate. This is completely opposite to
successful practice. It must therefore be concluded that the
assumption of an independent, but time-invariant growth rate is
not tenable.
EFFECT OF TIME BETWEEN INSPECTIONS
Many of the probabilistic corrosion growth rate models
currently in use consider the depth sizing error as the sole
source of uncertainty and are in fact deterministic as a function
of time. If a measurement at a single point in time were made
with sufficiently high precision, no variability would remain in
the model. However, reliability predictions are often made
assuming that independent growth increments occur from year
to year. This is obviously inconsistent: during the parameter
estimation phase it is assumed that the correlation over time is
perfect, yet during the model prediction phase it is assumed that
the growth increments are independent between successive
years.
If the growth increments really were independent from the
current defect sizes, the variance of the growth rate would be
reduced if more time elapsed between multiple ILIs. This effect
should then be taken into consideration when computing
corrosion growth rates from successive ILI runs. Because the
computed corrosion rate really is a time-averaged value, its
variance decreases with increasing length of time between the
two inspections. The exact expression of the variance reduction
function depends on the auto-correlation of the random process
that describes the instantaneous corrosion rate (see Vanmarcke,
1983).
Corrosion is really a random process, where periods of
inactivity (no growth) alternate with periods of increasing wall
loss (active corrosion) as shown in Figure 4. In a deterministic
approach, a single time-independent corrosion growth rate is
used; this would be represented by a straight line in Figure 4
and amounts to perfect correlation over time.
In a random process model, no perfect linear relationships
exist. Corrosion may progress uniformly over time (Line 1 in
Figure 4), accelerate (Line 2) or passivate (Line 4) over time.
Corrosion may also initiate instantaneously (at first date of
service) or after an initiation time (Line 3).


Figure 4: Random process types describing corrosion growth
PROPERTIES OF GOOD GROWTH RATE MODELS
Based on the previous discussion, a couple important
mathematical conditions can be formulated in order to increase
the realism of corrosion growth modeling:
- Corrosion growth rates should not be directly computed
from the raw ILI data; the effects of sizing error should be
properly accounted for.
- To avoid inconsistencies in the mathematical treatment and
unexpected results, the growth increments should be
formulated as a random degenerative process (or random
field when location specific models are built).
- Corrosion science should be the basis of the main
characteristics of the model formulation. Efficient model
parameter estimation relies on the use of hierarchical
models. The sensitivities to sizing error should be
investigated and accounted for.
Even when all these conditions are met, it remains an
impossible task to accurately predict the state of the pipeline far
into the future. The main objective of the studies is to obtain
information about the anticipated or likely states of the pipeline
over the next few years.
SUMMARY
This paper outlined some important differences between
deterministic and probabilistic corrosion growth assessments.
This paper highlights some of the challenges that are associated
when porting the maximum corrosion rate approach used in a
deterministic approach to a reliability-based paradigm.
Accurate prediction of the true likelihood of an adverse
event is impossible without significant research into
determining and understanding the, usually conservative, bias
0
10
20
30
40
50
60
70
80
90
0 5 10 15 20
M
e
d
i
a
n

p
r
e
d
i
c
t
e
d

d
e
p
t
h

[
%
W
T
]
Timesince last inspection [year]
Depth=25%WT
Depth=20%WT
Depth=15%WT
Depth=10%WT
Depth=5%WT
6 Copyright 2012 by ASME
in the engineering models that are currently employed in the
pipeline integrity state-of-the-practice. This is not a trivial task
and in principle requires a fundamental revisiting of every
assumption.
Fundamental modeling issues associated with both defect
and segment matching approaches were highlighted. The
common assumption of using either fully-correlated or fully
independent corrosion growth increments is discussed.
Mathematical requirements for a better corrosion growth model
form are formulated and it is proposed that a random process
approach be used to avoid inconsistencies in the modeling.
ACKNOWLEDGMENTS
This work was funded through the Pipeline Integrity
Technology Development Program at Chevron ETC. The
support from the Focus Area managers and the members of the
Technical Leadership Team is gratefully acknowledged.
REFERENCES
Achterbosch, G.G.J., L.A. Grzelak, 2006, Determination of
the Corrosion Rate of a MIC Influenced Pipeline Using 4
Consecutive Pig Runs, IPC 2006-10142, Proceedings of 6
th

International Pipeline Conference, Calgary, AB, Canada.
American Petroleum Institute, 2005, API 1163, In-line
Inspection Systems Qualification Standard, Washington, D.C.
ANSI/ASNT, 2005, In-line Inspection Personnel
Qualification and Certification, ANSI/ASNT ILI-PQ-2005,
Columbus, OH, USA.
Brown W.H., 1999, Using ILI Pigs to Establish Pipeline
Corrosion Rates Case Histories, Corrosion 1999 Paper
99523, NACE International, Houston, TX, USA.
Dawson, S.J., Kariyawasam, S.N., 2009, Understanding
and Accounting for Pipeline Corrosion Growth Rates, Joint
Technical Meeting, Paper No. 22, Milan, Italy May
Dawson, S.J., J. Wharf, M. Nessim, 2008, Development of
Detailed Procedures for Comparing Successive ILI Runs to
Establish Corrosion Growth Rates PRCI project EC 1-2.
Desjardins, G., 2001, Corrosion Rate and Severity Results
from In-Line Inspection Data, Corrosion 2001 Paper 01624,
NACE International, Houston, TX, USA.
Desjardins, G., 2010, Detection of Active Corrosion from
Repeated ILI Runs, Clarion Pipeline Pigging Conference,
Houston, TX, USA.
Desjardins, G., Sahney, R., Spencer, K., 2011, Draft Final
Reports for Phase II of Field Demonstration and Bench
Marking Work for Reliability Based Guidelines for Pipeline
Integrity, final reports on project EC1-6, Pipeline Research
Council International, Inc., January
Escoe, A.K., 2006, Piping and Pipelines Assessment
Guide, Gulf Professional Publishing, Elsevier, Oxford, UK.
Fenyvesi L., S. Dumalski, 2005, Determining Corrosion
Growth Accurately and Reliably, NACE Corrosion 2005 Paper
05154, NACE International, Houston, TX, USA.
Fontana, M.G. and N.D. Greene, 1986, Corrosion
Engineering, McGraw-Hill, 3
rd
Edition, New York, NY.
Gelman A. and J. Hill, 2007, Data Analysis: Using
Regression and Multilevel/Hierarchical Models, Cambridge
University Press, New York, NY, USA.
Haines H., 2012, Progress Report on Project EC-4-2, PRCI
Technical Exchange Meeting, Phoenix, AZ.
Huyse, L., and A. van Roodselaar, 2010, Effects of Inline
Inspection Sizing Uncertainties on the Accuracy of the Largest
Features and Corrosion Rate Statistics, 8
th
International
pipeline conference, IPC2010-31037, Calgary, AB, Canada.
Huyse, L., J. Monroe, A. van Roodselaar, 2011, Effects of
in-line inspection sizing uncertainties on in-the-ditch
validation: Review of results, IBP1185_11, Rio de Janeiro,
Brazil.
NACE, 2001, In-line Nondestructive Inspection of
Pipelines, TR 35100, NACE International, Houston, TX, USA.
NACE, 2002, Recommended Practice: In-line Inspection
of Pipelines, RP0102-2002, Houston, TX, USA.
NACE, Pipeline External Corrosion Direct Assessment
Methodology (ECDA), Houston, TX, USA.
Nessim, M., J. Dawson. R. Mora, S. Hassanein, 2008,
Obtaining Corrosion Growth Rates form Repeat In-Line-
Inspection Runs and Dealing with the Measurement
Uncertainties, IPC 2008-64378, Proceedings of 7
th

International Pipeline Conference, Calgary, AB, Canada
Pipeline Operator Forum, 2009, Specifications and
requirements for intelligent pig inspection of pipelines.
Rust, S.W. and E.R. Johnson, 2001, Statistical Method for
Identifying Active Corrosion Based on Data from Multiple In-
Lin Inspection Runs, Corrosion 2001 Paper 01622, NACE
International, Houston, TX, USA.
Vanmarcke, E., 1983, Random Fields: Analysis and
Synthesis, MIT Press, Cambridge MA, USA.
Worthingham, R.G., L.L. Fenyvesi, T.B. Morrison and G.J.
Desjardins, 2002, Analysis of Corrosion Rates on a Gas
Transmission Pipeline, corrosion 2002, Paper 02091.

You might also like