You are on page 1of 9

Determination of Corrosion Growth Rates using Pipeline Operators Forum

Defect Classifications
T.Fletcher, F Samie.; Wood Group Integrity Management
Abstract
Over the service lives of pipelines, operators may check for the presence of defects using intelligent pigs. Over
time, defects may corrode and grow in size until they threaten the integrity of the pipeline. Assessing how quickly
defects may corrode can be complex due to various factors including non-linearity in corrosion growth rates,
differing corrosion mechanisms for different kinds of defect and differences in inspection tool tolerances. To
simplify such assessments, a single corrosion growth rate may be applied to all defects in a pipeline; this
approach is frequently taken as the basis for determining a re-inspection date by intelligent pig. The Pipeline
Operators Forum provides guidance on classifying defects based on pipeline wall thickness, defect axial length
and defect circumferential width. Examples include pitting-type defects, general-corrosion-type defects and
pinholes. This paper examines the use of these classifications to estimate corrosion growth rates by fitting the
dimensions of defects to different statistical distributions. The use of these methods helps to refine estimates of
corrosion growth rates and refine corrosion management strategies including more accurate intelligent pig reinspection intervals.
Introduction
Oil and gas pipelines rarely fail, however failures
can and do occur. Statistics compiled by
[1]
CONCAWE
indicate that mechanical failure,
corrosion and third party activity were the main
causes of oil pipeline failures in Europe since 1971.
Of these, 131 (27% of the total) were related to
corrosion. An analysis of United Kingdom onshore
[2]
pipeline failures by UKOPA
indicate that
corrosion accounted for 39 (21% of the total) loss of
containment incidents between 2006 and 2010.
Assessing the degree of corrosion is needed to
determine the condition of a pipeline. The
dimensions of any metal loss corrosion anomalies
present in a pipeline may be measured using
inspection technologies such as ultrasonic (UT) or
magnetic flux leakage (MFL) tools mounted on pigs.
The acceptability of a metal loss corrosion anomaly
may be assessed by estimating the stress in its
remaining ligament of pipe wall at given boundary
conditions; such as the pipelines maximum
allowable operating pressure (MAOP). If the
estimated stress is within specified criteria, then the
metal loss anomaly may be considered to be
acceptable; otherwise the anomaly is considered to
have failed and require remediation in the form of
further assessment, repair or replacement.
Examples of such assessment methodologies
include the DNV-RP-F101 Part B Allowable Stress
[3]
[4].
Design and the Detailed RSTRENG criteria
Estimating a corrosion growth rate for a pipeline
enables such assessment methodologies to be
used to make a prediction of the time taken for a
metal loss corrosion anomaly to grow in size to
reach an acceptable limit. These methods
commonly take either the reported length or width
of an anomaly as constant values and determine
the greatest acceptable depth. The difference
between the greatest acceptable depth and
reported depth is divided by a corrosion growth rate

Page 1 of 9

enabling an estimate of a failure date to be made.


This approach is illustrated in Figure 1.
Predicted failure dates for metal loss corrosion
anomalies may be used to prioritise corrosion
management measures such as re-inspection
intervals by intelligent pig, physical inspections of
external anomalies, use of chemical inhibition and
cleaning. Refining such management measures
can result in cost savings to a pipeline operator. For
this reason, estimating accurate corrosion growth
rates makes the prediction of failure dates more
precise, this in turn helps to lower corrosion
management costs.
Corrosion Growth Rate Estimation: Current
Methods
Corrosion rates can be estimated by assuming
linear growth between two given dates. Where
intelligent pig inspection data exists, the reported
depth of a metal loss anomaly may be used to
determine a corrosion growth rate. A basic
approach to calculate such a figure assumes that
the anomalys depth was 0% of wall thickness at
the time of the pipeline installation and grows in a
linear manner until the inspection date. The
anomalys depth is divided by the number of years
of service and a corrosion growth rate figure
determined. Taking the deepest reported anomaly
as the basis of determining a single corrosion
growth rate to apply to the whole pipeline therefore
provides a conservative value on which to base
failure date predictions.
If a pipeline has been inspected more than once by
intelligent pig then the difference in depths for an
anomaly at two or more dates may be determined.
This difference may then be used to calculate a
linear corrosion growth rate. Complexities arise in
using this approach due to accurately matching
anomaly locations between inspections and
accounting for tool types and tolerances.

Where a known corrosion mechanism exists or is


assumed, then a corrosion growth rate may be
estimated by applying a corrosion model. Many
models exist and are generally used in the form of
inputting data describing operating parameters and
fluid composition into a software package.
Examples include Shells Hydrocorr, BPs
Cassandra and NORSOKs M-506 corrosion model.
A number of models are in the public domain and
may be freely downloaded from the internet; many
[5,
of these are based on the work of deWaard et al
6, 7]
.
Large differences in estimated corrosion growth
rates may be obtained when comparing different
models; particularly at high operating temperatures
where protective film formation may occur on the
inner pipe wall.
Anomaly Classification
The Pipeline Operators Forum provides guidance
on the classification of metal loss anomalies based
on pipeline wall thickness, anomaly width and
anomaly length. An illustration of the algorithm used
[8]
to classify anomalies is shown in Figure 2 . The
following seven anomaly dimension classes are
specified:

Circumferential Slotting
Circumferential Grooving
General
Pitting
Axial Grooving
Axial Slotting
Pinhole

Each of these encompasses a range of different


shapes. Although the classification of anomalies
using this algorithm may be considered arbitrary,
anomalies classified as the same type may be
assumed more likely to share similar corrosion
mechanisms than anomalies of differing types.
Anomaly Distributions
It can be useful to inspect the depths of each
anomaly type by plotting their reported depths using
a histogram. Two examples of these are shown in
Figure 3; one for general type anomalies and one
for pitting type anomalies. Note that Figure 3 plots
frequency density as opposed to frequency on the
y-axis (frequency density is defined as frequency
divided by class width). A subjective inspection of
such histograms enables inferences to be made
regarding the likelihood of an anomaly having a
depth greater than a given value, and by extension,
a corresponding corrosion growth rate.
Figure 3 shows that the distribution of depths for
different anomaly classes may be dissimilar: in this
example, pitting type anomalies have a distribution
that is skewed towards shallower depths, whereas
general type anomalies have an approximately
symmetrical shaped distribution. In this case, after
inspecting the reported depths on the x-axes of
Page 2 of 9

both plots, we could infer that general type


anomalies have a higher mean corrosion growth
rate compared to pitting type anomalies.
The method presented in this paper describes a
quantitative method for estimating corrosion growth
rates by analysing the shape of these anomaly
depth distributions.
Probability Distribution Fitting
In statistics, a probability distribution describes the
probability of each outcome in a sample; for
example, the probability of scoring 5 heads when
tossing a coin 10 times. The outcomes of randomly
tossing a coin and recording the number of heads
and tails obtained can be modelled by the widely
recognised normal distribution with its characteristic
bell shaped curve. There are many other types of
probability distributions that are used to model
physical phenomena; examples include the Weibull
distribution,
commonly
used
for
reliability
engineering applications; or the gamma distribution,
commonly used in weather forecasting.
Distribution fitting is defined as the process of fitting
a probability distribution to variable measurement
data. The purpose of modelling a data distribution is
to allow the estimation of the likelihood of
occurrence at a given interval. For example,
modelling anomaly depth distributions from an inline inspection enables the probability of an
anomaly type with a given percentage depth to be
determined.
Two main techniques exist for modelling data
distributions. The first uses methods that evaluate
the parameters that describe a distributions shape;
for example, the mean and standard deviation of a
normal distribution. The second uses regression
methods; these apply mathematical transforms to
linearise the cumulative probability of a data set.
Many open source and commercial software
packages exist that enable users to easily apply
these kinds of methods. Two examples of modelling
using parameter evaluation are illustrated in Figure
4. For these figures, a random sample of 100
normally distributed data points are plotted as two
density histograms and a normal distribution model
is fitted to each plot. In this case, the parameters
describing the distribution of the data were known,
having a mean of 5 and a standard deviation of 2.
These values correspond to the estimated mean of
5.2 and standard deviation of 1.77 for the red
Normal distribution curve approximation.
Goodness of Fit
In statistical modelling, goodness of fit measures
how well a model describes a set of observed data.
Determining how accurately distributions model
data sets provides a measure of reliability.
Goodness of Fit by Eye
The two examples in Figure 4 show histograms of
identical normally distributed data plotted using

different frequency intervals (bins). In each case,


inspecting the normal distribution goodness of fit to
each data set by eye appears to give a good match.
However, the goodness of fit appears to be different
for each plot. This is because, in each case, a
different number of bins are used to plot the
histogram; even though the data sets are identical.
Changing the number of bins of a density histogram
may change the apparent shape of the distribution.
For this reason subjectively fitting a probability
distribution to a histogram by eye can be
misleading and lead to inaccurate probability
estimates. To help improve the accuracy of curve
fitting, probability distributions may be fitted to
density plots rather than histograms as their shapes
are constant (two examples of density plots are
shown in Figure 5 along with probability distribution
fits).

3. Compare the test statistic with a critical value


(tables of critical values are available for many
distributions
such
as
normal,
Weibull,
exponential and gamma). The critical value used
is normally quoted at a given significance level
(). Commonly used significance levels are 0.05
(representing 5%) and 0.01 (representing 1%).
4. If the test statistic is smaller than the critical
value the null hypothesis is accepted, otherwise
it is rejected.
Determination of Depth Probability
In general terms, for a continuous probability
density function,
) the probability that lies
between two values,
and
is given by the
following relationship:
Equation 3

Goodness of Fit by Testing


Whilst assessing the goodness of fit of a probability
distribution by eye provides a useful indication of
fitting accuracy, it does not provide any quantitative
measure. Statistical hypothesis testing enables the
accuracy of a goodness of fit to be determined.
Generally, hypothesis testing for modelling works
by comparing the differences between the models
expected values and the data distributions values.
These differences are represented as a single
value known as a test statistic. If the probability of
obtaining this value (commonly called the p-value)
is greater than the test statistic, then the hypothesis
that the model describes the data distribution may
be accepted. Otherwise it may be rejected.
There are a number of tests that may be used to
assess how well a model fits a data distribution.
The main hypothesis testing method presented in
[9]
this paper is known as the Anderson-Darling test .
To apply this test the following procedure is used:
1. State a null hypothesis, H0 that the data follow a
stated distribution.
2

2. Calculate the test statistic, A using:


Equation 1

where:
=

number of data points

and:
Equation 2
[

cumulative distribution function of the


stated distribution

Data point (from ordered data)

Page 3 of 9

The integral of the function is equal to one, i.e:


Equation 4

Applying these relationships to the anomaly


classifications discussed above means that the
probability of an anomaly having a depth of less
than a given value can be determined with a certain
confidence level.
Confidence Intervals
A confidence interval provides a range between
which a population parameter is likely to fall. The
limits of a confidence interval show the upper and
lower boundaries that define the range. For
example, if a corrosion growth rate for internal
general type anomalies was calculated as 0.25
mm/yr at a 95% confidence level; then the actual
rate may be greater or less than this value. A
confidence interval provides a way of showing how
likely a corrosion growth rate is to lie between two
given values.
Confidence levels are commonly calculated so that
the likelihood of the population parameter falling
within the given range is 95% (however, they can
be calculated at any given percentage). Using this
figure for the example above and assuming a lower
limit of 0.24 mm/yr and an upper limit of 0.26 mm/yr
would result in a corrosion growth rate of 0.25
mm/yr 0.01 at a 95% confidence level.
Calculating confidence intervals is straightforward
for distributions that are approximately normally
distributed and is commonly an in-built function for
many statistical software packages. Generally, low
sample sizes result in large confidence intervals

and large sample sizes result in small confidence


intervals. This means that calculating a corrosion
growth rate for a small population of anomalies
results in a larger range of possible values
compared to a large population of anomalies.
Calculating a confidence interval for some
asymmetrical probability distributions such as the
Weibull distribution may become complex.
However, approximations may be used based on
[10, 11]
the normal distribution or numerical methods
.
Discussion
The depth distributions of anomaly classes may be
modelled using probability distributions. Doing so
enables the probability of an anomaly having a
depth greater or less than a given value to be
determined for a given confidence level.
Assuming a linear corrosion growth between two
dates (for example, the installation and inspection
dates) enables depth probabilities to be used to
estimate corresponding corrosion growth rates.
In general, lower test statistic values are calculated
for anomaly classes with large populations. This
means that it may be impractical to successfully
model an anomaly class with a large population
using a probability distribution. In practice,
sectioning a pipeline by distance or clock position
may be used to subdivide anomaly classes into
smaller populations. Doing so enables successful
modelling to be carried out using an array of
probability distributions; however, this technique
can result in large numbers of corrosion growth
rates being estimated for a single pipeline. Whilst
applying these to remnant life calculations may
become complex, doing so results in more refined
estimates being made.
Modelling anomaly class populations is based on
the results of in-line inspections by intelligent pig.
The reported anomaly dimensions from these
inspections are themselves subject to tool
tolerances and have a degree of error. In the case
of magnetic flux leakage inspections, anomaly
dimensions may be subject to a degree of
interpretation by trained personnel. The method
presented in this paper makes no allowance for
these issues, for this reason, any tolerances quoted
by an inspection tool vendor need to be included in
the classification of anomaly types.
In addition to the inspection tool tolerances
discussed above, the dimensions of metal loss
anomalies reported by in-line inspections are
normally provided at a given confidence level. This
means that the length, width and depth values of an
anomaly may be greater or less than the reported
figures. The axial and circumferential positions of
anomalies reported by intelligent pig are also
subject to error due to pig slippage. These points
illustrate that there are uncertainties in measuring
any factor describing an anomaly characteristic. By
extension, these uncertainties also apply to the
Page 4 of 9

calculation of a corrosion growth rate. The


examples provided in this paper are calculated
using a 95% confidence level.
Estimating corrosion growth rates using a
probability
distribution
fitting
technique
is
necessarily based on historical inspection data. Any
predicted corrosion growth rates therefore make the
assumption that the past operation of the pipeline is
applicable to the future. Whilst this may in many
cases be a safe assumption to make, it may not be
in all cases. The technique makes no allowances
for any differences in future operation: for example
changes in production fluid composition, chemical
inhibition or cleaning pigging frequency. However,
given this limitation, the technique does allow an
alternative and more refined method of estimating
corrosion growth rates and measuring the
effectiveness of corrosion management systems.
Conclusions
A new method of estimating corrosion growth rates
for metal loss anomalies in pipelines has been
developed based on POF classifications and
probability distribution fitting.
The method can be used to determine a corrosion
growth rate for different types of metal loss anomaly
in different sections of a pipeline at a given
confidence level and confidence interval.
A detailed assessment of corrosion growth is made
possible by estimating a range of corrosion growth
rates based on the results of in-line inspection data.
A comparison of these corrosion growth rates after
any additional in-line inspections enables the
effectiveness of any corrosion management
strategy to be evaluated in more detail compared to
using a single corrosion growth rate figure.
Example
The following example illustrates the application of
probability distribution fitting. The data used in this
example is taken from a hypothetical subsea crude
oil pipeline with internal corrosion. The corrosion
was considered to be caused by water drop out due
to an inefficient separator. Although the following
data is hypothetical, this example is based on a
number of real life examples.
In-line inspection data from the first 500 m section
of heavily corroded pipeline was analysed and
classified using the POF algorithm illustrated in
Figure 2. For illustration, only general and pitting
type anomalies are analysed in this example and
no subdivision by clock position or spool is
considered. 155 general type and 57 pitting type
anomalies were classified.
The population of 155 general type anomalies were
found to have a good fit by eye to a Weibull
distribution with a shape and scale parameter of
2.65 and 5.36 respectively. Similarly, the population
of 57 pitting type anomalies was found to have a
good fit by eye to a gamma distribution with a

shape and rate parameter of 4.00 and 1.27


respectively. These results are illustrated in Figure
5.
The goodness of fit of each model was tested
against each data distribution. The results of these
tests are shown in Table 1.
Table 1
Anomaly
Class

Null
Hypothesis

Test
Statistic

p-value

General

Weibull

0.34

0.90

Pitting

Gamma

0.19

0.99

As the test statistics were lower than the p-values in


each case, the null hypotheses that the general
type anomalies were drawn from a Weibull
distribution and the pitting type anomalies were
drawn from a gamma distribution were both
accepted.
The Weibull and gamma models were used to
estimate a depth for general and pitting type
anomalies at a 95% confidence level. This process
is illustrated graphically in Figure 6. This resulted in
a depth of 8.1 mm 5.4% for general type
anomalies and 6.1 mm 11.0% for pitting type
anomalies. The larger confidence interval for pitting
type anomalies reflects a lower sample size (n =
57) compared to general type anomalies (n = 155).
Assuming that these anomalies had zero depth at
installation then a linear corrosion growth rate
corresponding to the above depths could be
estimated based on the in-line inspection date and
used to predict the failure dates for these
anomalies.
In practice, the example described above would
include an analysis of all anomaly types and include
an investigation into the effect of sectioning the
pipeline by distance and clock position.

Page 5 of 9

References
1. P.M. Davis, J Dubois et al. Performance of European cross-country oil pipelines. Statistical
summary of reported spillages in 2010 and since 1971. CONCAWE, Brussels, Belgium,
December 2011;
2. R.A McConnell, Dr J. V Haswell. UKOPA Pipeline Product Loss Incidents (1962-2010).
UKOPA/11/0076, Ambergate, UK, November 2011;
3. DNV Recommended Practice, DNV-RP-F101, Corroded Pipelines, October 2006;
4. PRCI Report PR 3-805, A Modified Criterion for Evaluating the Remaining Strength of
Corroded Pipe, 22 December 1989;
5. C. deWaard, U. Lotz, D.E. Milliams, Paper 577, CORROSION/91, 1991. Predictive Model
for CO2 Corrosion in Wet Natural Gas Pipelines;
6. C. deWaard et al., Paper 69, CORROSION/93, 1993, Prediction of CO2 Corrosion of
Carbon Steel;
7. C. deWaard et al., Paper 128, CORROSION/95, 1995, Influence of Liquid flow Velocity on
CO2 Corrosion: a Semi-empirical Model;
8. Anon, Specifications and requirements for intelligent pig inspection of pipelines. Version
2009, Pipeline Operators Forum, 2009;
9. Stephens, M.A. EDF Statistics for Goodness of Fit and Some Comparisons, Journal of the
American Statistical Association, 69, pp.730-737, 1974;
10. Hong,Y, Meeker, W, Escobar, L. Normal Approximations for Computing Confidence
Intervals for Log-Location-Scale Distribution Probabilities, Iowa State University,
Louisiana State University, June 2006;
11. Symynck, J, De Bal, F. Monte Carlo Pivotal Confidence Bounds for Weibull Analysis with
th
Implementations in R, The XVI International Scientific Conference, Tehnomus, Romania,
May 2011.

Page 6 of 9

Figures
Figure 1
Corrosion Growth Required to Cause Failure

Feature Depth

Increase in Depth Required to Cause Failure

Figure 2

Circumferential Slotting

8
7
6
5

W/A

Circumferential Grooving

Illustration of Pipeline Operators Forum Metal Loss Anomaly Classification Algorithm

General

4
3
2

Pitting

Axial Grooving

1
Axial Slotting
0
Pinhole

L/A

If Wall Thickness < 10 mm, A = 10 mm


If Wall Thickness 10 mm, A = Wall Thickness
W/A =Anomaly Width / A
L/A = Anomaly Length / A
Page 7 of 9

Figure 3
Density Histograms of General Type and Pitting Type Anomaly Depths

Figure 4
Distribution Fitting

Page 8 of 9

Figure 5
Density Plots of General Type and Pitting Type Anomalies with Probability Distribution Fits

Figure 6
Depth Comparison of General and Pitting Type Anomalies at a 95% Confidence Level

Page 9 of 9

You might also like