Quantitive Assessment of Corro

TECHNICAL ASPECTS OF THE RISK MANAGEMENT OF CORRODIBLE SYSTEMS
Quantitive Assessment of Corrosion Probability—

A Bayesian Network Approach
F. Ayello,‡,* S. Jain,* N. Sridhar,* and G.H. Koch*
ABSTRACT 16th and 17th centuries.1 One of the earliest recorded

definitions of risk, and the explicit reference to prob-
Corrosion processes mainly affect the probability of failure, ability, was made by the 17th century Port Royal
which then leads to consequences, such as, fire, explosion, monastery in France2—“Fear of harm ought to be
or environmental damage. This paper focuses on the use of proportional not merely to the gravity of the harm,
Bayesian network models for assessing the probability of cor-
but also to the probability of the event.” The specific
rosion. The Bayesian network approach incorporates cause-
definition of risk embedded in the Port Royal Monas-
effect relationships of complex systems in the form of condi-
tional probabilities. This method considers both knowledge
tery statement has carried through the subsequent
uncertainties (i.e., modeling uncertainties) and data uncertain- generations into our standard definition of risk as
ties to make more informed decisions. The Bayes theorem al- probability multiplied by consequence. However, such
lows the model to predict the probability of events from their a definition does not enable one to adequately com-
causes, and, if a particular event is known to have occurred, pare a low-probability/high-consequence scenario to a
predict probable causes of that event. Two case studies, the high-probability/low-consequence scenario. Two such
first one involving internal corrosion and the second involving scenarios may result in the same “risk number,” but
external corrosion of oil and gas pipeline, are presented, along may mean very different things to different stakehold-
with validation using field measurements. The extension of ers. However, in a broader sense, risk is defined in
the approach to predicting stress corrosion cracking of pipe-
terms of a triplet of questions:
lines is discussed.
—What can go wrong?
KEY WORDS: corrosion rate, external corrosion, internal corro- —How likely is it?
sion, oil and gas, pipelines, stress corrosion cracking —How does it affect us?
given by the mathematical expression:3
INTRODUCTION
Risk = {Si, pi, Ci} (1)
Risk assessment is used to calculate the future prob-
ability of a loss or damage and seems to have been where Si is a set of scenarios or threats (what can
practiced as far back as 3,200 B.C. by people living go wrong?), pi is their probability of occurrence (how
in the Tigris-Euphrates valley. However, the quanti- likely is it?), and Ci is their consequence (how does
tative underpinning of risk assessment arose in the it affect us?). In this broader definition of risk, for
Submitted for publication: January 2, 2014. Revised and ac-
a given system or a design option, there may be a
cepted: April 30, 2014. Preprint available online: May 22, 2014, set of scenarios, each with its own pair of probabil-
doi: http://dx.doi.org/10.5006/1226. ity (frequency) and consequence, which then can be
‡ Corresponding author. E-mail: [email protected].
* Det Norske Veritas (USA), Inc., 5777 Frantz Road, Dublin, OH portrayed as a curve of probability vs. consequence.3
43017. Comparing two design or maintenance options re-
ISSN 0010-9312 (print), 1938-159X (online)
1128 14/0000183/$5.00+$0.50/0 © 2014, NACE International CORROSION—NOVEMBER 2014
quires a full evaluation of a set of such curves. The denominator in Equation (3) is called the probability
reader is referred to a clear exposition of these issues of observation and is the sum of all the conditional
in Kaplan and Garrick.3 probabilities of B given events, Aj multiplied by the
The estimation of probability has engendered the probabilities of Aj. The prior and posterior probabili-
greatest amount of controversy in the literature. The ties can also be considered as “cause” and “conse-
“frequentists” or statistical view of probability is that quence” of a process. The term “consequence” in this
probability is an objective number that can be ap- sense should not be confused with consequence of a
proached if a sufficiently large number of controlled risk model. It is strictly a relation between two events
observations are made. The relationship between fre- in a process, one leading to the other. An example
quency and probability is defined by the well-known would be the presence of water in a pipeline leading to
Bernoulli’s limit theorem:4 corrosion.
Bayesian networks use Bayesian inference (Equa-
p(|f(x) – p(x)| > ε) → 0 as N → ∞ (2) tion [6]) on a larger scale. Bayesian networks allow
calculating the probability of not one but numerous
where the frequency, f(x), of a population of data interconnected parameters. Bayesian networks are
approaches the probability, p(x), of that same data often represented graphically where many random
as the number of trials approaches infinity. Stated variables are connected by cause-consequence depen-
another way, frequency is a measurable quantity dencies. This approach is particularly useful to per-
based on repeated observations, whereas probability form risk assessment of corrodible systems, because
represents the degree of belief or confidence in the of the ability to consider a great number of events
measured frequency, also referred to as probability of that can lead to failure. Since, the Bayes theorem
frequency.3 The “subjectivists” refer to probability as shows what is and is not evidence, and describes the
simply a degree of belief in an event. This view stems strength of the evidence, a succession of Bayesian
from the idea that not all phenomena can be repeated inferences will provide the most probable scenarios
in a controlled manner to derive statistical distribu- leading to failure, quantify the certainty of the sce-
tions. This is especially true of complex systems. narios, and provide a mathematical framework to
Therefore, probability can be assigned to the strength reduce uncertainties through observation (i.e., data
of an expert’s belief about an event and then can later gathering). The main limitation of Bayesian networks
be corrected using repeated observations. This is at is that to solve the succession of Bayesian infer-
the heart of the Bayes theorem and it is often referred ences numerically only a directed acyclic graph can
to as belief network. In this paper, we combine these be used. In other words, the strength of the cause-
two perspectives—where possible statistical distribu- consequence relationships must be preserved and no
tions are derived through the use of mechanistic mod- feed-back loops are allowed between causes and con-
els that are, in turn, based on experimental data with sequences to alter the strength of these relationships.
their associated uncertainties—but we also include di- However, knowledge of a consequence can be used to
rect probability distributions representing the degree update the probability of a cause without altering the
of belief of an expert in a given observation. These two strength of their relationship. The heart of the Bayes-
streams of probabilities are linked in a Bayesian net- ian network model is the derivation of the conditional
work that can be updated through laboratory or field probability tables or P(B|Aj), from fundamental mod-
observations. els and subject matter experts.
The Bayes theorem states that the posterior prob- In this paper, we describe the application of
ability of an event (i.e., probability of the event after Bayesian network modeling to two case studies in-
an observation is made) is related to the prior proba- volving internal and external corrosion of oil and gas
bility of the event (i.e., before the observation is made) pipelines. A description of how the Bayesian network
through the probability of observing the event and the model is set up is given first. The computed probabili-
conditional probability of observation given the event ties are compared to measured inspection results to
occurred, as given by Equation (3): validate the approach. The Bayesian network model is
extended for the more complex stress corrosion crack-
ing (SCC) phenomena of pipelines and illustrated for
P(B
B A i ) × P( A i )
P( A i B
B) = n a hypothetical pipeline. However, at present there is
∑ P(BB A ) ⋅ P(
j P( A ) j a lack of specific pipeline data to validate the SCC
j=1 (3) model.
where P(Ai|B) is the posterior probability of event Ai BAYESIAN NETWORK APPROACH

given the observation B, P(Ai) is the prior probability
of the event Ai before the observation B, and P(B|Ai) Principle
is called the likelihood function and is the probability Pipelines are exposed to a number of risk fac-
of the observation B given that event Ai occurred. The tors or threats, such as intrusion by third parties,
CORROSION—Vol. 70, No. 11 1129

TABLE 1
Beginning (columns 1 through 5) and End (columns 3067 through 3072) of the Conditional Probability Tables
for Internal Uninhibited Uniform Corrosion Rate Probability
Temperature Celcius 20-40 20-40 20-40 20-40 20-40 … 80-100 80-100 80-100 80-100 80-100
CO2 Bar 0-0.1 0-0.1 0-0.1 0-0.1 0-0.1 … 10-100 10-100 10-100 10-100 10-100
Fe2+ ppm 0-10 0-10 0-10 0-10 0-10 … 50-100 50-100 50-100 50-100 50-100
pH — 4-5 4-5 4-5 4-5 4-5 … 7-8 7-8 7-8 7-8 7-8
O2 ppb 0-10 0-10 0-10 0-10 10-100 … 100-1,000 1,000- 1,000- 1,000- 1,000-
10,000 10,000 10,000 10,000
H2S ppm 0-10 10-100 100-1,000 1,000- 0-10 … 1,000- 0-10 10-100 100-1,000 1,000-
10,000 10,000 10,000
Corrosion 0-0.01 0 0 0 0 0 … 0 0.65 0 0 0
rate probability 0.01-0.1 0.065 0.525 0.42 0 0.08 … 0 0.35 0 0 0
in mm/year 0.1-1 0.865 0.475 0.58 1 0.86 … 0 0 0 0 0
1-2 0.07 0 0 0 0.06 … 0.16 0 0.8 0.57 0.185
2-5 0 0 0 0 0 … 0.84 0 0.2 0.43 0.815
5-10 0 0 0 0 0 … 0 0 0 0 0
10-20 0 0 0 0 0 … 0 0 0 0 0
fabrication defects, cyclic loading, and external and on a deep understanding of the underlying phenom-
internal corrosion. The intrusion and fabrication- ena; a statistical analysis based on field data would
related factors may be considered to be essentially not be enough to create these relationships since cor-
time-invariant factors, although some parameters that relations are not enough evidence of causations.
affect intrusion probability (e.g., population density)
change over time. The probabilities of these failure Step 2: Quantify the Cause-Consequence
modes may be estimated from the probabilities (i.e., Relationships Identified in Step 1
frequencies) of related factors, such as surrounding Next, causal dependencies are quantified in joint
population centers, vehicle traffic, type of welding probability tables involved with the Bayes theorem
used to manufacture the pipeline, etc. The probabili- (Equation [3]).8 This is the most time-consuming step,
ties of time-dependent factors, corrosion, hydrogen because it requires quantifying all cause-consequence
embrittlement, and SCC have to be assessed using a relationships. This quantification uses three types of
number of models. The next section shows two exam- information:
ples: internal corrosion and external corrosion. These —tested and recognized physics-based models
models were created using common understanding of (most preferable)
corrosion processes. The Bayesian network construc- —expert’s knowledge (less preferable)
tion is done in two steps.5-6 —field data (even less preferable)
It must be emphasized here that, while field data is
Step 1: Create a Graphical Representation the least preferable from the point of view of calculat-
of the Chain of Events Leading to Corrosion ing joint probability distributions, they are essential
and Ultimately to Failure in conducting probabilistic analyses and validation.
First, a graphical representation of the Bayesian Of course, no Bayesian network model is created us-
network is developed, where all events/parameters ing only one type of information, and usually all three
that can have a direct or indirect effect on corrosion are required to complete a Bayesian network model.
and failure are linked using arrows, each arrow rep- Quantification of causal dependencies is a crucial
resenting the relationship of a cause pointing toward step in making a successful Bayesian network model.
a consequence.7 It is very important that arrows rep- The information collected during step 1 is put in con-
resent only cause-consequence relationships and not ditional probability tables under the form of a succes-
correlations, since they will be quantified during the sion of probabilities (Ai in Equation [3]). A conditional
next step using Equation (3). A simple anecdote dem- probability table is created for every event of the
onstrates that correlation is not causation. There is Bayesian network. All conditional probability tables
a direct correlation between children’s reading skills are made of columns representing a set of input com-
and their shoes sizes, mainly because younger chil- binations (a node has multiple connections, therefore
dren with smaller feet sizes may not be as advanced multiple B in Equation [3]) and must have as many
in reading skills. Therefore, a Bayesian network esti- columns as there are possible sets of inputs (making
mating children learning skills should not link shoe the tables too large to publish). An example of a con-
size with children reading skills directly, even if shoe ditional probability table is shown in Table 1.
size is a relatively good marker of children’s reading Physics-Based Models — One of the most reli-
abilities. Consequently, the graphical representation able ways to derive conditional probability tables is to
of the Bayesian network risk model has to be based use fully tested and recognized physics-based models
1130 CORROSION—NOVEMBER 2014

because these represent our understanding of the model, probabilistic analysis can be conducted by
underlying phenomena, and recognized models can repeatedly running such an integrated model using a
be assumed to have been peer-tested. Physics-based probabilistic driver, such as the Monte Carlo method,
models (such as multiphase flow or corrosion rate each run being called a realization. In such a case,
models in the following examples) are run multiple the Bayesian method provides no advantage and
times over all sets of possible inputs in a Monte Carlo could indeed be less rigorous. However, in complex
fashion. It is recognized that a specific phenomenon, systems where different phenomena are connected
such as carbon dioxide (CO2) corrosion, may be rep- in diverse ways and cannot be described by an inte-
resented by a number of models that may produce grated model, Bayesian networks provide a rigorous
different calculated results for the same input param- mathematical method to combine different types of
eters. In the Bayesian network construct, multiple probabilistic knowledge to make informed decisions.
models can be run multiple times and results can be The other major advantage of Bayesian models is the
combined in the conditional probability tables, using reversibility of the Bayesian inference as shown in
weighting functions for different models. This is very Equation (3). Unlike other models where there are
useful because trust is increased in the areas where inputs and outputs, in Bayesian models there are
all models provide the same values and justified doubt only unknown and known probabilities.3 If two events
emerges in the areas where models diverge. If the ve- are linked, then knowing the probability of one event
racity of different models for the same phenomenon improves the knowledge of the probability of the
is not known a priori, equal weighting functions can other event. Finally, Bayesian networks are graphical
be applied that can then be corrected later through models, making the visualization of complex chains
observations. of events easy to understand, unveiling the probable
Expert’s Knowledge — Quantification of the mechanisms of failure. A Bayesian network model
causal relationships is possible using an expert’s developed for corrosion helps the user understand
knowledge. This is necessary since there are many corrosion phenomena and implicitly suggests ways to
mechanisms of pipeline failure with no reliable mech- control or combat corrosion. The next section devel-
anistic model. SCC and microbiologically influenced ops in detail two Bayesian network models developed
corrosion are examples of complex phenomena where for pipelines.
many overlapping mechanisms may operate and a
detailed model of the complete phenomenon is diffi- BAYESIAN NETWORK FOR PIPELINES—
cult to achieve. Yet, some experts understand parts of CASE STUDIES
these problems quite well and this knowledge should
not be discarded because it is difficult to quantify. Internal Corrosion Threat
This knowledge is added scenario by scenario in usu- Model — A cause-consequence relationship model
ally smaller conditional probability tables. Conditional was created to assess the internal corrosion in crude
probability tables derived from expert’s knowledge oil transportation pipelines. The Bayesian network
have larger uncertainties, and this uncertainty is car- model (schematic shown in Figure 1 and full model
ried by the Bayesian network model all the way to the shown in Figure 2) is divided into seven parts. The
final results. first four parts take into account threats to the pipe-
Field Data — Field data can be used to populate line: uniform corrosion, localized corrosion, erosion,
the conditional probability tables. Every instance for and microbiologically influenced corrosion. The next
each input set is counted and used to generate col- two parts of the model evaluate flaws (depth and
umns of the conditional probability tables. Using field length) in the pipeline. Finally, the last part deter-
data has many drawbacks. First, field data usually do mines the remaining strength by comparing pipeline
not cover all possible sets of input parameters, and operating pressure with burst pressure to evaluate
when it does, some combination of inputs have many the probability of pipeline failure at any given location
measurements while others have few. Second, field and time.
data vary from field to field and estimation of the un- To illustrate how the internal corrosion model
certainties associated to the data is difficult. And most works, a small portion of the model has been high-
importantly, while field experts are easily challenged, lighted in Figure 2 and shown in detail in Figure 3.
field data are rarely put into context, providing a false This figure shows the part of the corrosion model that
sense of security. calculates the uninhibited corrosion rate. (Corrosion
inhibition, flow, and steel wettability are not shown
Advantages here for clarity reasons.) First, the figure shows that
If an analytical or a numerical model is available each event is not represented by a number but a
to describe a phenomenon, then probabilistic analysis probability density function that describes the relative
of this phenomenon can be performed using distrib- likelihood for each variable to take on any given value,
uted inputs into this model. If several such phenom- and no state, no matter how unlikely, is overlooked.
ena exist, and they can be integrated into a numerical A red bar shows a node that has a precisely known

FIGURE 1. Layout of the internal corrosion model.
FIGURE 2. Bayesian network created for pipelines’ internal corrosion damage assessment. Every event is linked to the
next by cause-consequence relationships.
state (e.g., in Figure 3, hydrogen sulfide [H2S] concen- hence, they are represented by uniform (or flat) dis-
tration is known to be null), the blue bars show nodes tributions. Consequently, because of the uncertainty
with uncertain (or incomplete) information (e.g., the in the data, the uniform uninhibited corrosion rate is
temperature is known to be below 60°C, but the exact known with uncertainty. Calculations are performed
value is unknown), and green bars show unknown using Equation (3), and the conditional probability
variables (Fe2+ and O2 concentrations are unknown); tables obtained by running multiple corrosion models

FIGURE 3. Zoom of the rectangle in Figure 2; parameters according to Table 6 (temperature in °C, CO2 partial pressure
in bar, Fe2+ concentration in ppm, O2 in ppb, H2S in ppm, and corrosion rate in mm/y).
multiple times (1,000) for all possible combinations partial pressure, and pH), some data is uncer-
of the inputs parameters. The conditional probability tain (O2 concentration is lower than 100 ppb,
table has 3,072 rows and is shown in Table 1. H2S concentration is lower than 100 ppm, but
Generating this conditional probability table the exact values for O2 and H2S are unknown),
required more than 3 million corrosion simulations and some data is completely unknown (Fe2+
for each corrosion model used. Once the conditional concentration). Consequently, the uniform un-
probability tables are generated, it is not necessary inhibited corrosion rate is known with uncer-
to run the models again because the knowledge con- tainty and is most probably between 0.1 mm/y
tained in the models is transferred into the probability and 1 mm/y, but could be as high as 5 mm/y
tables. to 10 mm/y with 7% probability. Depending on
Some examples of the Bayesian calculations the rest of the model (i.e., corrosion inhibition,
demonstrated in Figure 3 are shown in Tables 2 flow, wettability of the steel), this high uninhib-
through 6. In these tables, the green bars represent ited corrosion rate value might be acceptable or
unknown variables and the red bars represent known not. If the high uniform uninhibited corrosion
variables. rate is unacceptable, for example, because of
1. In Table 2, the probability distributions of all the lack of corrosion inhibition or the presence
the variables are unknown; therefore, we as- of high water-cut, then a sensitivity analysis on
sign uniform (equal) distribution within broad the model can help prioritize what data should
bounds. As expected, the calculated uniform be gathered to reduce the uncertainty on the
uninhibited corrosion rate distribution is also uniform uninhibited corrosion rate. It should
broad with a finite probability as low as 0 and be noted that such sensitivity analysis is not
as high as 10 mm/y and higher. However, it generic, because the sensitivity analysis will
should be noted that despite assigning uniform take into account what is known of the system;
distribution to input variables, the corrosion when a parameter changes, so do the results of
rate is not distributed uniformly. This example the sensitivity analysis.
also demonstrates that the Bayesian model can 4. Table 5 shows a specific scenario where un-
still run even without information related to known data do not create uncertainty. In this
input data distributions. scenario four input parameters are known
2. In Table 3, all input parameters are known; and two are unknown. This is a special case,
hence, the uniform uninhibited corrosion rate because the H2S concentration is high, which
is also known with a high degree of certainty. implies the formation of an iron sulfide (FeS)
However, the certainty is not 100%, because protective film (H2S concentration has other
the factors are bracketed within narrow ranges effects, including undesirable effects on local-
of values. Therefore, a quantitative corrosion ized corrosion). In this scenario, not knowing
model will yield a different corrosion rate if the temperature and Fe2+ concentration has no
value of CO2 is 1 bar vs. 10 bars, whereas the determining effect on the final corrosion rate.
probability for this range is 100%. Uncertainty in two of the input parameters
3. Table 4 shows more realistic conditions where does not generate uncertainty in the output
some data is known (i.e., temperature, CO2 parameter. This scenario demonstrates that

TABLE 2
Calculation with No Known Data (Ref. Figure 3)
T CO2 Fe2+ O2 H2S pH Corr. Rate
C % Bar % ppm % ppb % ppm % — % mm/year %
20-40 25 0-0.1 25 0-10 33 0-10 25 0-10 25 4-5 25 0-0.01 3
40-60 25 0.1-1 25 10-50 33 10-100 25 10-100 25 5-6 25 0.01-0.1 15
60-80 25 1-10 25 50-100 33 100-1,000 25 100-1,000 25 6-7 25 0.1-1 43
80-100 25 10-100 25 1,000-10,000 25 1,000-10,000 25 7-8 25 1-2 18
2-5 8
5-10 4
>10 9
TABLE 3
Calculation with Only Known Data (Ref. Figure 3)
20-40 100 0-0.1 0 0-10 100 0-10 0 0-10 0 4-5 0 0-0.01 0
40-60 0 0.1-1 0 10-50 0 10-100 100 10-100 0 5-6 100 0.01-0.1 0
60-80 0 1-10 100 50-100 0 100-1,000 0 100-1,000 0 6-7 0 0.1-1 99.5
80-100 0 10-100 0 1,000-10,000 0 1,000-10,000 100 7-8 0 1-2 0.5
2-5 0
5-10 0
>10 0
TABLE 4
Combination of Known (red), Uncertain (blue), and Unknown (green) Data (Ref. Figure 3)
20-40 100 0-0.1 0 0-10 33 0-10 50 0-10 50 4-5 0 0-0.01 0
40-60 0 0.1-1 0 10-50 33 10-100 50 10-100 50 5-6 100 0.01-0.1 0
60-80 0 1-10 100 50-100 33 100-1,000 0 100-1,000 0 6-7 0 0.1-1 45
80-100 0 10-100 0 1,000-10,000 0 1,000-10,000 0 7-8 0 1-2 21
2-5 27
5-10 7
>10 0
TABLE 5
Special Scenario, Fe2S Film Dominates Uniform Uninhibited Corrosion Rate. In this Scenario Unknown Data Do Not
Generate Uncertainty (Ref. Figure 3)
20-40 25 0-0.1 0 0-10 33 0-10 0 0-10 0 4-5 0 0-0.01 0
40-60 25 0.1-1 100 10-50 33 10-100 100 10-100 0 5-6 100 0.01-0.1 0
60-80 25 1-10 0 50-100 33 100-1,000 0 100-1,000 0 6-7 0 0.1-1 100
80-100 25 10-100 0 1,000-10,000 0 1,000-10,000 100 7-8 0 1-2 0.6
2-5 0
5-10 0
>10 0
TABLE 6
Bayesian Networks have No Inputs No Outputs, only Known and Unknown Data. Unknown Data (green) is Inferred Using
Model (Figure 3) and Known Data (red). In this Example, Unknown Data is Inferred with High Certainty.
20-40 0 0-0.1 0 0-10 0 0-10 0 0-10 0 4-5 100 0-0.01 0
40-60 0 0.1-1 0 10-50 0 10-100 50 10-100 0 5-6 0 0.01-0.1 0
60-80 100 1-10 100 50-100 100 100-1,000 50 100-1,000 0 6-7 0 0.1-1 0
80-100 0 10-100 0 1,000-10,000 0 1,000-10,000 100 7-8 0 1-2 0
2-5 100
5-10 0
>10 0

not all data are of equal importance or even

necessary at all times. This also demonstrates
how Bayesian networks can help pipeline op-
erators to prioritize what data to gather. Of
course under different conditions, temperature
and Fe2+ concentration could become important
parameters.
5. Table 6 shows that even what is usually con-
sidered as an “output” (corrosion rate) can be
used as an “input” to the model. In this sce-
nario, two input parameters are known and the
corrosion rate is also known (through labora-
tory measurements or corrosion coupons and
FIGURE 4. Comparison between 2010 ILI data and model’s flaw
probes installed in the pipeline). This scenario depth prediction.
shows the effectiveness of Bayesian networks,
where any parameter can be used to improve
the probability estimates of consequences or carbon steel was installed in 2004. The first in-line
causes. inspection (ILI) was performed in 2010 and revealed
When the CO2 partial pressure, the H2S concentra- extensive localized corrosion damage (the line has
tion, and the corrosion rate are known to the values been repaired and no loss of containment has oc-
fixed in Table 6, then the probability distributions of curred). The line data cannot be shared because of
temperature, Fe2+ concentration, O2 concentration, proprietary reasons, and results are shown to demon-
and pH can be inferred. It may be counterintuitive to strate how model validation is performed.
see parameters that are typical inputs and outputs Flow characteristics, fluids composition (oil and
combined to calculate missing inputs. This is very water), and corrosive species concentrations were pro-
useful for risk assessments when often much of the vided to the model. To test the methodology, it was de-
data is missing, and deduction can help to achieve cided to use only the data that was available in 2004
a reasonable conclusion. For example, in Table 6, and try to predict the condition of the line in 2010.
although the temperature was not measured, it is It should be noted that 2010 ILI was provided to the
known to be between 60°C to 80°C, which might be authors only after completion of the project. The re-
good or bad depending on the type of microbial activ- sults of the risk assessment were then compared to
ity. Also, even though pH was not measured, it is also the results of the 2010 ILI. Comparison between flaw
inferred to be between 4 and 5 and may affect other depth prediction and actual ILI observation (Figure
aspects of the overall corrosion model presented in 4) indicates that the median flaw depth prediction for
Figure 2. 2010 was 25% of wall loss, which is close to the me-
Of course, these calculations depend on the dian value of flaw depth (24% wall loss that was mea-
physical model used to relate these factors.9 It is well- sured with ILI). There is an overall good agreement
known that different corrosion models may predict between the Bayesian network results and the ILI
different corrosion rates even for the same inputs. The observations, both on the extent of corrosion damage
Bayesian network model can include such model un- and on the mechanism of corrosion (widespread local-
certainties in addition to parameter uncertainties. The ized attack from sulfate-reducing bacteria [SRB]). The
examples above show only the principle of the internal Bayesian network distribution prediction is somewhat
corrosion threat assessment model. The overall model “flatter” than the actual observation, which is the
tracks more than 100 events that could have direct or result of the uncertainty in the 2004 data. If, within
indirect effects on internal corrosion. The fundamen- 2004’s data uncertainty, all parameters combined to
tals of the models are shown in Appendix A. Moreover, form the least corrosive conditions, then negligible
as new data, models, and knowledge become avail- corrosion would have occurred (with a probability of
able, it is easy to update both the Bayesian network 15%). On the other hand, if all the worst parameters
and the conditional probability tables to make better combined to form the most corrosive conditions, then
predictions and the structure of the Bayesian model the largest flaws would have been 60% to 70% of
will not change. pipe-wall thickness (with a probability of 2%). How-
Field Validation — Kuwait Oil Company (KOC, ever, none of these extreme scenarios happened and
Ahmadi, Kuwait) operates more than 300 crude oil flaw depth distribution is similar to the predicted flaw
and gas pipelines, which range in age from 30+ years depth.
to recently constructed pipelines. A pipeline operated The result of the model matches ILI data in part
by this company was of particular interest, because because the model does not try to predict a single
of an unexpected high wall loss in only 6 years. The value, but all possible outcomes. This makes the pre-
25 km, 36-in outside diameter (OD) line made of diction more useful to engineers because failure is of-

FIGURE 5. Layout of the external corrosion model.
FIGURE 6. Bayesian network model for external corrosion assessment. Every event is linked to the next by cause-
consequence relationships.
ten the outcome of some unlikely scenarios as a result and the CP should provide protection to the pipeline.
of unlikely parameters combinations. These unlikely However, the coatings can disbond or get damaged
scenarios are found in the tails of the distributions. because of the environmental variables such as soil
Also, there was a good agreement between flaw length stress as a result of depth of cover, drainage, soil
predictions and actual ILI observations. Finally, the expansion, or mechanical damage. If the CP system
flaw size was related to rupture probabilities using the does not work perfectly, the pipeline is exposed to the
Monte Carlo simulation of ASME B31G equations10 for corroding environment. Figure 5 shows the schematic
determining the remaining strength of corroded pipe- layout of the model and Figure 6 shows the Bayesian
lines. No pipeline failures were predicted by the model network model to predict the probability of failure
despite the extent of corrosion damage; this is also in caused by external corrosion. The probability of fail-
agreement with field observations. ure depends on the corrosion flaw depth and length,
which depend on the corrosion rate and the time
External Corrosion Threat when the corrosion initiated. The corrosion initiates
Model — Most of the corrosion failures of hazard- only after the coating is damaged. First, the Bayesian
ous liquid and natural gas pipelines occur because of network model calculates the probability of coating
external corrosion. It has been estimated that between damage. The probability of coating damage depends
1994 and 1999, more than half of the corrosion fail- on several factors ranging from soil stress, applied CP
ures in the U.S. onshore pipelines were from external (cathodic effect), and construction practices to the op-
corrosion.11 In general, these pipelines are under- erating temperature. Second, the model evaluates the
ground and protected by a coating and a cathodic pro- environment that forms at the damaged coating site.
tection (CP) system. The combination of the coating The environment depends on the CP potential, the soil

TABLE 7
Most Important Parameters Used in the External Corrosion
Simulations, Blue Uncertain Data, Red Certain Data (Ref.
Figure 6 and Appendix B)
Simulation Certainty
Parameters Value (%)
CP on potential –500 to –650 VCSE 0
–650 to –700 VCSE 0
–700 to –750 VCSE 0
–750 to –800 VCSE 0
–800 to –850 VCSE 0
–850 to –950 VCSE 0
–950 to –1,200 VCSE 50
–1,200 to –1,500 VCSE 50
Soil type Sand (10%) 10
Clay 85
Loam 5
FIGURE 7. Probability of coating damage predicted by the
Topography Depressed 10
external corrosion model.
Level 90
Dent Yes 95
No 5
Area affected Affected 30 strength, and flaw dimensions. Finally, the Bayesian
by weld Not affected 70
network model uses the critical pressure to estimate
Surface Blasting 0
preparation Brushing 0 the probability of failure for each pipe section at one-
No preparation 100 year intervals. The conditional probability tables for
Sulfates 0 ppm to 50 ppm 20 the Bayesian network model were developed based on
50 ppm to 100 ppm 20 subject matter expert beliefs, mechanistic models, or
100 ppm to 150 ppm 20 data from the field or experiments. More information
150 ppm to 200 ppm 20
200 ppm to 1,000 ppm 20 about the exact sources is provided in the Apendix B.
Chlorides 0 ppm to 150 ppm 33 Field Validation — The model shown in Figure 6
150 ppm to 1,500 ppm 33 was used to calculate the corrosion flaw characteris-
1,500 ppm to 10,000 ppm 33 tics (depth and length) for a pipeline operated by the
MIC Yes 50 China National Petroleum Corporation (CNPC, Beijing,
No 50
Month of 0-3 0
China). The pipeline suffered extensive damage and
soil wetness 3-6 0 was compared to the data obtained using a magnetic
6-9 100 flux leakage ILI tool.(1) The pipeline was installed in
9-12 0 1973. The inputs for the model were provided by
Coating type Asphalt 100 CNPC and are summarized in Table 7.
Coal tar 0
FBE 0
The external corrosion model was used to predict
PVC 0 the probability of coating damage as well as corrosion
Tar glass 0 flaw depth and length. Figure 7 shows the probability
Temperature 80°C to 70°C 0 of the coating damage as a function of the year of the
70°C to 60°C 25 pipeline. The double inflection in probability increase
60°C to 50°C 25
50°C to 40°C 25
with time is due to the mix of soil types that affects
40°C to 30°C 25 coating damage, with clay-type soil causing a faster
30°C to 20°C 0 increase in damage than a sandy soil. However, coat-
20°C to 10°C 0 ing damage is not measured. The predicted and the
measured distributions of corrosion flaws depth and
length are shown in Figure 8. The model is in good
properties such as total dissolved solids, oxygen avail- agreement with the ILI data at multiple locations and
ability, and CO2 pressure. Third, the model assesses time periods. The pipeline was severely corroded be-
corrosion rates. Corrosion rates depend on the chem- cause of the original construction quality that likely
istry and determine the extent of corrosion damage. led to poor coating adhesion; the pipeline has a long
Fourth, the model calculates the pressure for failure operating history of nearly 40 years (the asphalt
of the pipeline; it depends on corrosion flaw depth and coating tends to become porous or deteriorates with
length and can be estimated from a numerical frac- time); the soil moisture fluctuated greatly leading to
ture mechanics model outlined by Jaske, et al.12 This periods of time when the CP was not working, and
model was used with Monte Carlo simulation to derive
(1) To maintain the privacy of the company and follow the legal re-
the conditional probability tables of critical pressure quirements, the geographic details about the pipeline and the
for an exhaustive set of diameter, thickness, yield tools used are not provided.

—the ability to deal with the complete lack of

data by initially representing the data through
a uniform probability distribution within broad
bounding values of the parameters
—the ability to update the calculations in a
straightforward manner from observations
The major weaknesses of the Bayesian method are
that it considers the event nodes to be discrete, and it
does not allow feedback between cause-consequence
relationships. When feedback loops are necessary
(e.g., the presence of a flaw can accelerate local corro-
sion or cracking), it is necessary to run the Bayesian
(a) model multiple times to obtain convergence.
Case Studies: Internal and External Corrosion

This paper illustrates two failure processes en-
demic to on-shore oil and gas pipelines and presents
validation of the Bayesian technique through two
field cases. The validation of risk assessment models
using field data is fraught with many difficulties,
both technical (uncertainties in input data and mod-
els) and non-technical (sensitivities to communication
of results). Although the non-technical factors ought
not to be the subject of a research article, they can
nevertheless form a significant barrier in improving
a Bayesian model that especially relies on improving
(b) predictions through observations. In conducting the
FIGURE 8. External corrosion model forward prediction compared validation studies, care was taken not to include the
to the wall loss (WT) and flaws length obtained from ILI in 2011. field failure data in the model so that independent
validation is assured. In both cases presented, the
field inspection data were not revealed to the model-
possibly the pipeline experienced some bacterial activ- ing team until after the probabilities of failure were
ity because it passes through a populated area close presented.
to sewage dumps, garbage landfills, and other places In the case of internal corrosion, the Bayesian
that could have had high moisture retention harbor- network model represents the effect of well-known
ing bacterial activity. phenomena (e.g., flow regime or corrosion) and less
understood phenomena (e.g. corrosion inhibition or
DISCUSSION microbiologically influenced corrosion). Final results
will depend on conditional probability matrixes. For
Bayesian Network Methodology Advantages abiotic corrosion without inhibition, detailed physics-
There are many different approaches to risk based and accurate models are available. Therefore,
assessment, each with its own advantages and dis- the conditional probability tables are derived from
advantages.13-15 The Bayesian network represents a these models using Monte Carlo simulations. How-
highly flexible approach that can include physics- ever, for less understood phenomena, conditional
based models as well as expert opinions and data in probability tables must be derived from subject matter
one framework. As with any quantitative risk assess- experts, increasing the uncertainties. A validation of
ment model of a complex system, the creation of a the internal corrosion assessment was conducted us-
Bayesian network model requires deep understanding ing a pipeline of KOC that was known to suffer from
of the phenomena modeled and significant effort in internal corrosion. The subject matter expertise was
constructing conditional probability matrixes. Consid- particularly useful in the case of KOC’s pipeline. Al-
erable time is spent in eliciting expert knowledge as though the complete Bayesian network model shown
well as exercising various detailed models to extract in Figure 2 was used to determine failure probability
conditional probability data. The major strengths of for the KOC pipeline; the resulting probabilities were
the Bayesian approach lie in the following: mainly determined by the widespread localized attack
—the transparency with which the cause-conse- by SRB. All threat models predicted negligible corro-
quence connections are represented, so that sion rates except the microbiologically influenced cor-
continuous improvement can be made to the rosion part of the internal corrosion model; this is in
overall system representation by the users agreement with actual observations.

FIGURE 9. Layout of the model for SCC in pipelines.
In the case of external corrosion, predictions of eters and, therefore, help the users prioritize data col-
pipeline failure rates depend on a complex interac- lection. The use of importance analysis is illustrated
tion between different phenomena. Most of these phe- for the assessment of SCC in the next section.
nomena are considered intractable (such as coating
degradation, CP effectiveness in different soil types, Extension of the Approach—Stress Corrosion
crack interactions, electrolyte chemistry development Cracking
under coating, etc.) because of the uncertainty in the Similar to the case of internal corrosion and ex-
variables and complex mechanisms. However, Bayes- ternal corrosion, the concept of Bayesian network
ian network methodology allows dividing the problem models can be applied to the assessment of threat
into smaller problems and connecting them according as a result of SCC on pipelines. Initiation of SCC de-
to the overall degradation mechanism. When condi- pends on numerous interconnected factors such as
tional probability tables could not be derived from soil conditions, wetting and drying cycles, presence of
physics-based models, they were generated using field organic decay products, coating types, disbondment
and laboratory data. The external corrosion assess- geometries, pipe surface condition, stresses, tempera-
ment was compared to field observations for a pipeline ture, cathodic protection, chemistry under-coating,
known to suffer from extensive coating damage. In etc. Most of these parameters are imperfectly known
conducting this assessment, the external corrosion and may change during the long operation times of
rates were calculated by combining a set of field data16 the pipeline. Consequently, the probability of SCC
available for carbon steel corrosion in different soils failure is a complex function of variables that interact
with the mechanism of corrosion growth. The set of on different scales.17 Figure 9 shows the layout of the
external corrosion data from buried steel coupons had Bayesian network model that can be used to assess
missing information on a number of input conditions, the SCC threat. The probability of failure due to SCC
which results in great model uncertainties. Conse- depends on the crack sizes and distribution. Crack
quently, the uncertainty in the prediction of external growth rates are influenced by the environment, the
corrosion flaws depths and lengths was expected to be effective potential, and the stresses. Whether SCC
high. However, when the model was put to the test, a will occur also depends on whether the coating has
reasonable match between the prediction and the ILI disbonded. Therefore, all of these processes should be
observations were seen. More testing of both internal analyzed.
and external corrosion models is undergoing to im- Two forms of SCC have been observed under dis-
prove the confidence in the models. bonded coating of underground pipelines: high-pH
An important barrier for conducting risk assess- SCC and near-neutral pH SCC. High-pH SCC (also
ment of pipelines is the sheer lack of significant data known as classical SCC) is intergranular and is usu-
for modeling endemic to such geographically extended ally accompanied by moderately concentrated solu-
systems as pipelines. Many parameters, both for ex- tions (0.1 M to 1 M) in the pH range from 8 to 10.
ternal and internal corrosion, are either not measured Near-neutral pH SCC is transgranular and is observed
or measured too infrequently in space and time. In in relatively dilute solutions with pH in the range
some cases, pipeline location and depth of burial are from 6 to 8.18 The high-pH solution is the result of
poorly known. There are uncertainties in aligning ILI reduction of oxygen or water due to CP. Therefore, the
flaw indications from one period to the next, result- coating should allow for the CP and oxygen to pass
ing in uncertainties in estimated flaw growth rates. through. Shielding coatings such as polyethylene tape
However, Bayesian network modeling is beneficial for and PVC do not allow CP to get through, and, there-
these cases because one can assess the importance fore, are more resistant to the formation of high-pH
and sensitivity of model calculations to input param- solutions.19 The soil resistivity and the seasonal varia-

of high temperature. Figure 10 shows a typical profile

of temperature and pressure near a compressor, with
the corresponding crack growth rates and the risk of
failure calculated using the preliminary SCC Bayes-
ian network model. The predictions are in reasonable
agreement with published failure frequencies. Exten-
sive pipeline data is not available at present to vali-
date the model. However, limited validation has been
performed using openly available data. However, the
value of the Bayesian network model is in prioritiz-
ing data collection by performing evidence sensitivity
analysis with respect to the knowledge of different
FIGURE 10. Preliminary results: temperature, pressure, crack variables. The evidence sensitivity is defined in terms
growth rates, and probability of failure as a result of high-pH SCC of the calculated difference in probabilities between
(compressor station: mile 0). two outcomes, in which for one of the outcomes, the
evidence is removed. A cost of omission of information
about a variable Y, when prediction X, is defined as:
P( X, Y)
Y ) = ∑ P( Y))∑ P( X Y) log
I( X Y) (4)
Y X P( X)P( Y)
where I(X|Y) is the sensitivity factor (also referred to

as the mutual entropy), P(Y) and P(X) are the probabil-
ities for Y and X, respectively, and P(X,Y) is the joint
probability for X and Y.
An evidence (or a factor) is considered to be un-
important, if the calculated probability of the outcome
is no different if this evidence is removed from con-
FIGURE 11. Sensitivity factor (cross entropy) of crack growth rates
sideration. Conversely, if the calculated probability is
to different variables.
affected significantly by the absence of that evidence,
it is considered to be important. For example, in the
tions (that affect the groundwater movement) also case of high-pH SCC, a sensitivity analysis with re-
play an important role in the effectiveness of CP. For spect to crack growth rates is shown in Figure 11.
example, in conductive groundwater, CP may pene- The model shows that the operating temperature and
trate to greater distances within shielded coatings and pressure are the most important variables that affect
cause an increase in pH. Since, a high pH is associ- the crack growth rate; therefore, if cracks have al-
ated with CP and high-pH SCC occurs at less negative ready initiated, temperature and pressure should be a
potentials, it has been suggested that alternate wet priority to collect.
and dry conditions caused by seasonal water table
fluctuations may contribute to the coexistence of high CONCLUSIONS
pH and somewhat anodic potentials.19 The carbonate
solutions will form only if there is sufficient carbon ❖ This paper is focused on Bayesian network model-
species and soluble cations (Na+/K+) in the ground- ing of oil and gas pipelines. Two Bayesian network
water/soil.20 A comprehensive model for predicting models were created to assess the probability of pipe-
the chemistry at the damaged coating considering all line failure due to internal and external corrosion.
these factors have proven to be difficult and therefore Both models were validated against field data for two
the probabilities of these factors are linked through a different pipelines. The predicted locations of high-
Bayesian network based on expert judgment. The use- probability failure were in reasonable agreement with
fulness of such a model then depends on the updating field observations. The uncertainties attendant upon
of the model through failure data. Ideally, if multi- such predictions were identified. The Bayesian net-
year failure data is available, then the model could be work model is extended for the prediction of SCC. Al-
updated using some failure data and then validated though a field case for the validation of the SCC model
using the remaining data. However, at the time of has not yet been identified, the Bayesian network
writing this paper, such an extensive database was model can be used to identify factors that are impor-
not available to the research team. Therefore, a more tant to calculate risk and prioritize data collection.
limited validation using literature data was attempted. REFERENCES
It is well known that most of the failures caused
by high-pH SCC occur near the compressor because 1. V.T. Covello, J. Mumpower, Risk Anal. 5, 2 (1985): p. 103-120.

2. P.L. Bernstein, Against The Gods – The Remarkable Story of Risk 16. R.E. Ricker, “Analysis of Pipeline Steel Corrosion Data From NBS
(Hoboken, NJ: John Wiley & Sons, 1998). (NIST) Studies Conducted Between 1922-1940 and Relevance to
3. S. Kaplan, B.J. Garrick, Risk Anal. 1, 1 (1981): p. 11-27. Pipeline Management,” NISTIR 7516 (Gaithersburg, MD: National
4. T. Bedford, R. Cooke, Probabilistic Risk Analysis: Foundations and Institute of Standards and Technology, 2007).
Methods (Cambridge, U.K.: Cambridge University Press, 2009). 17. NACE Technical Committee Report 35103, “External Stress Cor-
5. P. Naïm, P.-H. Wuillemin, P. Leray, O. Pourret, A. Becker, Ré- rosion Cracking of Underground Pipelines” (Houston, TX: NACE,
seaux bayésiens, 3rd ed. (Paris, France: Libraire Eyrolles, 2007), 2003).
ISBN 978-2-212-11972-5. 18. J.A. Beavers, B.A. Harle, “Mechanisms of High-pH and Near-
6. P. Friis Hansen, “Structuring of Complex Systems Using Bayes- Neutral pH SCC of Underground Pipelines,” International Pipeline
ian Network,” Proc. JCCS. Workshop (Copenhagen, Denmark: Conf. (New York, NY: ASME, 1996), p. 555-564.
Danish Technical University, 2004). 19. M. Wilmott, B. Erno, T. Jack, “The Role of Coatings in the De
7. T. Bedford, R. Cooke, Probabilistic Risk Analysis Foundations and velopment of Corrosion and SCC on Gas Transmission Pipe-
Methods (Cambridge, U.K.: Cambridge University Press, 2001), p. lines,” Int. Pipeline Conf. (New York, NY: ASME, 1998), p. 399-
286-298. 408.
8. D. Koller, N. Friedman, Probabilistic Graphical Models: Principles 20. J.A. Beavers, R.G. Worthington, “The Influence of Soil Chemistry
and Techniques (Cambridge, MA: MIT Press, 2009), p. 697-776. on SCC of Underground Pipelines,” Int. Pipeline Conf. (New York,
9. FreeCorp, “Corrosion Protection Software,” Institute for Corrosion NY: ASME, 2002), p. 1-8.
and Multiphase Technology, Ohio University, http://www. 21. L. Banach, Pipeline Industry 4 (1988): p. 37.
corrosioncenter.ohiou.edu/software/freecorp/. 22. S. Papavinasam, R.W. Revie, “Protective Pipeline Coating Evalua-
10. ASME B31G-2009, “Manual for Determining the Remaining tion,” CORROSION/2006, paper no. 06047 (Houston, TX: NACE
Strength of Corroded Pipelines” (New York, NY: ASME, 2009). 2006), p. 1.
11. J.A. Beavers, N.G. Thompson, “External Corrosion of Oil and 23. J.A. Beavers, N.G. Thompson, “Effects of Coatings on SCC of
Natural Gas Pipelines,” ASM Handbook, vol. 13C, “Corrosion: Pipelines: New Developments,” Proc. 16th Int. Conf. on Offshore
Environments and Industries” (Materials Park, OH: ASM Mechanics and Arctic Engineering: Pipeline Technology, vol. 5
International, 2006), p. 05145. (New York, NY: ASME, 1995), p. 249-263.
12. C.E. Jaske, P.H. Vieth, J.A. Beavers, “Assessment of Crack-Like 24. J.A. Beavers, N.G. Thompson, “Corrosion Beneath Disbonded
Flaws in Pipelines,” CORROSION/2002, paper no. 02089 (Hous- Coatings: A Review,” CORROSION/96, paper no. 96208 (Houston,
ton, TX: NACE 2002), p. 1. TX: NACE 1996), p. 1.
13. API Recommended Practice 580, “Risk Based Inspection” (Wash- 25. J.A. Beavers, N.G. Thompson, K.E.W. Coulson, “Effects of Sur-
ington, DC: American Petroleum Institute [API], 2009). face Preparation and Coatings on SCC Susceptibility of Line Pipe
14. ISO 31000-2009(E), “Risk Management – Principles and Guide- Phase 2—Field Studies,” Proc. 16th Int. Conf. on Offshore Me-
lines” (Geneva, Switzerland: International Organization for Stan- chanics and Arctic Engineering: Pipeline Technology, vol. 5 (New
dardization [ISO], 2009). York, NY: ASME, 1993), p. 225-239.
15. ISO 31010-2009, “Risk management – Risk Assessment Tech- 26. “Report of Public Inquiry Concerning Stress Corrosion Cracking
niques” (Geneva, Switzerland:International Organization for Stan- on Canadian Oil and Gas Pipelines,” National Energy Board
dardization [ISO], 2009). (Calgary, Alberta, Canada, 1996).
APPENDIX A
Internal Corrosion Model
Node States Source Causes Consequences
Temperature (Probability of 20 to 40°C Pipeline operator or database — Uninhibited
internal temperature) 40 to 60°C corrosion rate
60 to 80°C Favorable
80 to 100°C environment
CO2 (Probability of CO2 partial 0 to 0.1 bar Pipeline operator or database — Uninhibited
pressure) 0.1 to 1 bar corrosion rate
1 to 10 bar
10 to 100 bar
Fe2+ (Probability of Fe2+ 0 to 10 ppm Pipeline operator or database — Uninhibited
concentration) 10 to 50 ppm corrosion rate
50 to 100 ppm
O2 (Probability of oxygen 0 to 10 ppb Pipeline operator or database — Uninhibited
concentration) 10 to 100 ppb corrosion rate
100 to 1,000 ppb GAB
1,000 to 10,000 ppb GAnB
H2S (Probability of H2S 0 to 10 ppm Pipeline operator or database — Uninhibited
concentration) 10 to 100 ppm corrosion rate
100 to to 1,000 ppm Anodic film
1,000 to 10,000 ppm
pH (Probability of pH) 4 to 5 Pipeline operator or database — Uninhibited
5 to 6 corrosion rate
6 to 7 Passive film
7 to 8
Uninhibited corrosion rate 0 to 0.01 mm/year Conditional probability table Temperature CO2 Inhibited
(Probability of uniform 0.01 to 0.1 mm/year built using Monte Carlo Fe2+ corrosion rate
uninhibited corrosion rate) 0.1 to 1 mm/year simulation on FREECORP O2
1 to 2 mm/year corrosion model H2S
2 to 5 mm/year pH
5 to 10 mm/year
10 to inf mm/year
Corrosion inhibition 0 to 10% efficiency Pipeline operator or database Future improvement: Inhibited
(Probability of corrosion 10 to 20% efficiency crude oil chemistry, corrosion rate
inhibition due to corrosion … corrosion inhibition
inhibitors or naturally occurring 80 to 90% efficiency efficiency
surface active compounds 90 to 100% efficiency
present in crude oil)
Inhibited corrosion rate 0 to 0.01 mm/year Conditional probability table Uninhibited corrosion rate Uniform
(Probability of uniform 0.01 to 0.1 mm/year built using Monte Carlo Corrosion inhibition corrosion
inhibited corrosion rate) 0.1 to 1 mm/year simulation based on the
1 to 2 mm/year equation:
2 to 5 mm/year Corrosion inhibition (%) =
5 to 10 mm/year 100·(CR uninhibited–
10 to inf mm/year CR inhibited)/CR uninhibited (continued on next page)

APPENDIX A (continued)
Uniform corrosion (Probability 0 to 0.01 mm/year Conditional probability table Inhibited corrosion rate Wall loss in%
of wall loss due to uniform 0.01 to 0.1 mm/year built using Monte Carlo Water wet
generalized corrosion in one 0.1 to 1 mm/year simulation based on the
time step) 1 to 2 mm/year equation:
2 to 5 mm/year Uniform corrosion = Inhibited
5 to 10 mm/year corrosion rate* water
10 to inf mm/year probability
ID (Probability of pipe section 4 inch Pipeline operator or database — Flow type
internal diameter) 6 inch
Pipe inclination (Probability –45 to –10 degrees Pipeline operator or GIS — Flow type
of pipe section inclination) –10 to –5 degrees database
–5 to –1 degrees
–1 to 0 degrees
0 to 1 degrees
1 to 5 degrees
5– to degrees
10 to 45 degrees
Oil density (Probability of 700 to 800 kg/m3 — Flow type
crude oil density) 800 to 900 kg/m3
900 to 1,000 kg/m3
Oil viscosity (Probability of 0.001 to 0.1 Pa.s Pipeline operator or database — Flow type
crude oil viscosity) 0.1 to 1 Pa.s
1 to 10 Pa.s
Gas density (Probability of 10 to 50 kg/m3 Pipeline operator or database — Flow type
gas density) 50 to 100 kg/m3
100 to 1,500 kg/m3
Liquid velocity (Probability of 0 to 0.1 m/s Pipeline operator or database — Flow type
liquid velocity) 0.1 to 0.7 m/s Sand deposits
0.7 to 1.5 m/s Water layer
Gas velocity (Probability of 0 to 0.1 m/s Pipeline operator or database — Flow type
gas velocity) 0.1 to 1 m/s
1 to 6 m/s
Flow type (Probability of Stratified smooth Conditional probability table ID Water layer
flow type) Stratified wavy built using Monte Carlo Pipe inclination
Annular simulation on Society of Oil density
Dispersed bubble Petroleum Engineers/Ovadia Oil viscosity
Slug flow Shoham multiphase flow Liquid velocity
Single phase model Gas density
Gas velocity
Water cut (Probability of 0% Pipeline operator or database — Water layer
amount of water in the flow) 0 to 1%
1 to 5%
5 to 10%
10 to 20%
20 to 40%
40 to 100%
Water layer (Probability of Water wetting Conditional probability table Water cut
water layer formed on the Oil wetting built using experts knowledge: Flow type
bottom of the pipe) water wetting probability Liquid velocity
increase with water cut
concentration and is most
probable is stratified flows
Steel hydrophilic (Probability Yes Pipeline operator or database — Water wet
that steel is hydrophilic) No
Water wet (Probability of Yes Conditional probability table Water layer Steel hydrophilic
water layer formed and wet No built using experts knowledge: Water layer
the bottom of the pipe) the steel surface will be water
wet if steel is hydrophilic or
water layer is formed on the
bottom of the pipe and pipe
was not cleaned recently by a
cleaning pig
Cleaning pig (Probability of 0 Discrete number of cleaning pig — Sand deposits
cleaning pigs run to remove 1 to 2 per year from pipeline operator Water layer
sand, deposits, and water 3 to 4
from the pipe) 5 to 6
7 to 8
9 to 10
11 to 12
Sand (Probability of sand in Sand Pipeline operator or database — Sand deposits
the flow) Trace Erosion
No sand
Sand deposits (Probability of High Conditional probability table Cleaning pig Localized
sand deposit formation) Medium built using experts knowledge Sand inhibitor
Low based on liquid velocity and Liquid velocity Protection form
cleaning pig history flow
Passive film (Probability of Yes Conditional probability table pH Galvanic cell
passive film formation) No built using experts knowledge
based on steel type and pH
Anodic film (Probability of Yes Conditional probability table H2S Galvanic cell
anodic film formation) No built using experts knowledge
based on steel type and H2S
concentration
Localized inhibitors Yes Conditional probability table Corrosion inhibition Galvanic cell
(Probability of localized No built using experts knowledge Sand deposits
inhibitor concentrations) based on corrosion inhibition
and sand deposits
(continued on next page)

Chlorides (Probability of Yes Pipeline operator or database — Galvanic cell
chlorides present) No
Galvanic cell (Probability of Yes Conditional probability table Localized inhibitors Localized
galvanic cells) No built using experts knowledge. Passive film corrosion
Galvanic cell probablity Anodic film
increase with the presence of Chlorides
localized inhibition, passive
films, anodic films, and chlorides
Wax deposits (Probability of Yes Pipeline operator or database Future improvement: Corrosion
wax deposits) No crude oil chemistry, under deposits
temperature, flow type
Asphaltenes deposits Yes Pipeline operator or database Future improvement: Corrosion
(Probability of asphaltenes No crude oil chemistry, under deposits
deposits) temperature, flow type
Corrosion under deposits Yes Conditional probability table Wax deposits Localized
(Probability of corrosion under No built using experts knowledge. Asphaltenes deposits corrosion
deposits) Corrosion under deposits Sand deposits
probability increase with the Water layer
presence of water and deposits
(wax, asphaltenes and sand)
Localized corrosion 0 to 0.01 mm/year Conditional probability table Corrosion under deposits Wall loss in %
(Probability of severity of 0.01 to 0.1 mm/year built using experts knowledge, Galvanic cell Flaw length
localized corrosion) 0.1 to 1 mm/year the severity of localized
1 to 5 mm/year corrosion increase with the
presence of corrosion under
deposits and galvanic cells
Valve (Probability of valve in Yes Pipeline operator or database — Erosion
pipe section) No
Bend (Probability of bend in Yes Pipeline operator or database — Erosion
pipe section) No
T-Piece (Probability of T-piece Yes Pipeline operator or database — Erosion
in pipe section) No
Diameter change (Probability Yes Pipeline operator or database — Erosion
of change of diameter in pipe No
section)
Erosion (Probability of erosion) 0 to 0.01 mm/year Conditional probability table Valve Wall loss in %
0.01 to 0.1 mm/year built using experts knowledge, Bend Flaw length
0.1 to 1 mm/year the severity of erosion T-piece
1 to 5 mm/year increase with the presence Diameter change
sand, geometry changes, and Sand
fluid velocity Liquid velocity
GAB (Probability that General High Pipeline operator or database O2 Protection from O2
Aerobic Bacteria [GAB] Medium
are present) Low
GAnb (Probability that General High Pipeline operator or database O2 SRB presence
Anaerobic Bacteria [GAnB] Medium
are present in pipe section) Low
Protection from flow High Conditional probability table Sand deposits Shelter
(Probability that SRBs are Medium built using experts knowledge, T-piece
protected from flow) Low a geometry change can protect Erosion
bacteria from flow, erosion does
the opposite
Protection from O2 (Probability High Conditional probability table GAB Shelter
that SRBs are protected from Medium built using experts knowledge,
oxygen) Low GAB can form a protective layer
Shelter (Probability of a SRB High Conditional probability table Protection from flow SRB conditions
shelter from flow) Medium built using average of flow and Protection from O2
Low oxygen protection
Sulfate (Probability of enough High Pipeline operator or database — Source of energy
sulfate to sustain SRB growth) Medium
Low
Hydrocarbons (Probability of High Pipeline operator or database — Source of energy
enough hydrocarbon to keep Medium
SRB alive) Low
Source of energy (Food to High Conditional probability table Sulfate SRB conditions
sustain SRB activities) Medium built using average of sulfides Hydrocarbons
Low and hydrocarbons
Favorable environment Good for SRB Conditional probability table Water layer SRB conditions
(Probability of favorable Bad for SRB built using experts knowledge. Temperature
environment for SRB in the Favorable environment is
pipe section) defined as moderate tempera-
tures and is increased by sand
deposits
SRB conditions (Probability of Good for SRB Conditional probability table MIC Shelter
favorable SRB conditions) Bad for SRB built using experts knowledge. Source of energy
SRBs need a favorable Favorable
environment, a source of environment
energy, and shelter
SRB presence (Probability that Yes Based on GAnB measurements GAnB MIC
SRB present in pipe section) No
MIC (Probability of 0 to 0.01 mm/year Conditional probability table SRB conditions Wall loss in %
microbiologically influenced 0.01 to 0.1 mm/year built using experts knowledge, SRB presence Flaw length
corrosion due) 0.1 to 1 mm/year the severity of erosion increase Future improvement:
1 to 5 mm/year with the presence, and good add effect of iron
conditions for bacteria to stay reducing bacteria and
alive acid producing bacteria (continued on next page)

Pipe wall thickness 5 to 7 mm Pipeline operator or database — Wall loss in %
(Uncorroded measured pipe 7 to 9 mm Bursting pressure
wall thickness) 9 to 11 mm
11 to 13 mm
Wall loss in% (Probability of 0 to 1 Conditional probability table Uniform corrosion Flaw depth
wall loss in one year) 1 to 2 calculated using pipe thickness Localized corrosion
… and the sum of all threats Erosion
24 to 25 uniform corrosion, localized MIC
>25 (corrosion, erosion, and MIC)
Flaw depth (Probability of 0 to 2 Conditional probability table Wall loss in % Bursting pressure
depth of corroded region) 2 to 4 calculated using the sum of Last flaw depth
… wall loss during current time
96 to 98 step and flaw depth of the
98 to 100 previous time step
Last flaw depth (Probability 0 to 2 Value from the previous time Flaw depth of the Flaw depth
of depth of corroded region 2 to 4 step previous time step
from the previous time step) …
96 to 98
98 to 100
Flaw length (Probability of 0 to 10 mm Conditional probability table Localized corrosion Bursting pressure
longitudinal length of the 10 to 100 mm built using experts knowledge, Erosion
corroded region) 100 to 1,000 mm the length of the law depends MIC
on the past length and mecha- Last flaw length
nism of corrosion (localized
corrosion, mic or erosion)
Last flaw length (Probability 0 to 10 mm Value from the previous time Flaw length of the Flaw length
of longitudinal length of the 10 to 100 mm step previous time step
corroded region during the 100 to 1,000 mm
previous time step)
OD (Probability of nominal 4 inch Pipeline operator or database — Bursting pressure
outside diameter) 6 inch
UTS (Ultimate tensile 250 to 350 MPa Pipeline operator or database — Bursting pressure
strength) 350 to 450 MPa
450 to 550 MPa
550 to 650 MPa
Bursting pressure 0 to 5 Conditional probability table OD Failure
(Probability of bursting 5 to 10 calculated from ASME B31G. UTS
pressure) … Pipe wall thickness Flaw depth
180 to 190 Flaw length
190 to 200
>200
Operating pressure 0 to 5
(Probability of operating 5 to 10 Pipeline operator or database — Failure
pressure) …
180 to 190
190 to 200
>200
Failure (Probability of pipe Yes Conditional probability table Bursting pressure —
section failure) No calculated from comparison Operating pressure
between operating pressure
and bursting pressure

APPENDIX B
External Corrosion Model
Coating type (Probability of Asphalt Pipeline operator or database — Shielding effect
finding a particular coating) Tape Cathodic effect level
FBE External force level
Coal tar Coating damage
Composite coatings
Depth of cover (Probability 5 m to 3 m Pipeline operator or database — Soil stress level
of pipe being located in a depth 3 m to 1m
range) 1 m to 0 m
0m
Weld position (Probability that At weld
the location has a weld) Away from the weld Pipeline operator or database — External forces level
Dents (Probability that the At dent
location has a dent) Away from the dent Pipeline operator or database — External forces level
Surface preparation Water blasting Pipeline operator or database — External forces level
(Probability that a certain Abrasive blasting with
surface preparation was walnut sheels
used during construction) Abrasive blasting with
silica sheels
Power wire brush
No prep
Annual rainfall (Probability of 1,000 mm/y to 100 mm/y Pipeline operator or GIS database — Drainage
rainfall being in a certain range) 100 mm/y to 10 mm/y
10 mm/y to 0 mm/y
Topography (Probability of Undulating Pipeline operator or database — Drainage
being in a certain topography) Ridged
Inclined
Depressed
Level
Side Slope
Temperature (Probability of the 80°C to 70°C Pipeline operator or weather database — Coating damage
operating temperature) 70°C to 60°C
60°C to 50°C
50°C to 40°C
40°C to 30°C
30°C to 20°C
20°C to 10°C
Coating age (Probability of age 0 years, 1 year, etc. Pipeline operator or database — Coating damage
of the coating)
Soil porosity 60% to 45% Pipeline operator or database — O2 availability
(Probability of soil porosity being 45% to 20%
in a certain range) 20% to 10%
10% to 0%
Soil type Sand Pipeline operator or GIS database — Drainage
(Probability of soil type) Clay Soil stress
Loam
Mixed soils
Soil resistivity (Probability of >10,000 Ω·cm Pipeline operator or GIS database — Effective potential
soil resistivity being in a certain 2,000 Ω·cm to 10,000 Ω·cm
range) 1,000 Ω·cm to 2,000 Ω·cm
500 Ω·cm to 1,000 Ω·cm
0 Ω·cm to 500 Ω·cm
Millscale (Probability that there Yes Pipeline operator or database — Effective potential
is mill scale on the pipeline) No
AC line (Probability that Yes Pipeline operator or database Presence of an AC line Effective potential
there is an AC line nearby) No
Effective potential (Probability –500 to –650 VCSE CIS or CPIS survey AC line
of effective potential being in –650 to –700 VCSE Or Mill scale
a certain range) –700 to –750 VCSE CPT table can be developed Shielding coating
–750 to –800 VCSE based on the Field data Concentration cells
–800 to –850 VCSE Applied CP
–850 to –950 VCSE CP effectiveness
–950 to –1,200 VCSE
–1,200 to –1,500 VCSE
Wet/Dry cycle length 9 to 12 month wet GIS database — CP effectiveness
(Probability of wetness of 6 to 9 month wet
the soil throughout the year) 3 to 6 month wet
0 to 3 month wet
O2 availability (Probability of High Subject matter expert Soil porosity pH under coating
O2 availability) Medium Drainage
Low
Total dissolved solids 0 to 4 me/L Government database or — pH under coating
(Probability of total dissolved 4 to 50 me/L Pipeline operator
solids being in a certain range) 50 to 110 me/L
110 to 800 me/L
CO2 pressure (Probability of 0.09 to 0.06 kPa Subject matter expert — HCO3 concentration
CO2 partial pressure being in 0.06 to 0.045 kPa
a certain range) 0.045 to 0.039 kPa
0.039 to 0.03
0.03 to 0.01 kPa
HCO3 concentration 50 mM to 5 mM Mechanistic models from pH under coating Max corrosion rate
(Probability of HCO3 5 mM to 0.5 mM literature FREECORP CO2 partial pressure Uniform corrosion rate
concentration being in a 0.5 mM to 0.05 mM Initiation rate
certain range at the pipe) 0.05 mM to 0.005 mM
surface 0.005 mM to 0.0005 mM
Chloride concentration 0 ppm to 150 ppm Pipeline operator or — Max corrosion rate
(Probability of chloride 150 ppm to 1,500 ppm Government database Initiation rate
concentration being in a certain 1,500 ppm to 10,000 ppm
range at the pipe surface) 10,000 ppm to 80,000 ppm (continued on next page)

APPENDIX B (continued)
Sulfates concentration 0 ppm to 50 ppm Pipeline operator or — Max corrosion rate
(Probability of sulfates 50 ppm to 100 ppm Government database Uniform corrosion rate
concentration being in a certain 100 ppm to 150 ppm Initiation rate
range at the pipe surface) 150 ppm to 200 ppm
200 ppm to 1,000 ppm
MIC Yes ECDA- field data — Max corrosion rate
(Probability of MIC occurring) No Uniform corrosion rate
Initiation rate
External forces (Probability High Subject matter expert Coating type Composite effect
that the external forces are Medium combined with literature Weld position
high, medium, or low) Low information [21-25] Dents
Surface preparation
Soil stress
Composite effect (Probability Very high Model: Pool of cathodic effect External forces Coating damage
that composite forces are high, High and external forces Cathodic effect
medium, or low) Medium
Low
Cathodic effect (Probability that High Subject matter expert and Coating type Composite effect
effective potential on the pipe Medium combination with literature Effective potential
surface has a high, medium, or Low information [20-21]
low impact on disbondment)
Coating damage (Probability Yes Field data: Statistical model Coating age Years since initiation
that coating damage exists) No fits to excavation data Temperature
Composite effect
Drainage Well drained Field data: statistical model Annual rainfall O2 availability
(Probability that drainage Imperfectly drained fits to excavation data Topography Soil stress
falls in a certain class) Poorly drained Soil type
Very poorly drained
Very poorly-very poorly
Drained
Soil stress High Subject matter expert Depth of cover External forces
(Probability that the soil stress Medium Soil type
for coating damage is high, Low Drainage
medium, or low)
Shielded coatings (Probability Shielding Literature [17,19,22-23,25-26] Coating type Effective potential
that a coating will fall in shielding No shielding or pipeline operators
or no shielding category)
Concentration cells (Probability Yes Subject matter expert [27] Soil type Effective potential
of formation of galvanic cells) No on the surface
pH under coating (Probability of pH = 6 to 8 Field data: Statistical fit to field data Total dissolved dolids HCO3 concentration
finding a pH in a certain range) pH = 8 to 9 O2 availability Max corrosion rate
pH = 9 to 11 Effective potential Uniform corrosion rate
Initiation rate
CP effectiveness (Probability of 9 to 12 month Same as wet dry cycle length: Wet/dry cycle length Max corrosion rate
number of months the CP was 6 to 9 month Taken from GIS weather Uniform corrosion rate
working for) 3 to 6 month database Initiation rate
0 to 3 month
Years since corrosion initiation 1 year, 2 years ….. Iteration in the model Time since initiation Corrosion length
(Probability of years since the Corrosion depth
corrosion initiated) Loss of thickness
Faults per mile
Year of previous initiation 1 year, 2 years Iteration in the model — Years since
(Probability of years since the corrosion initiation
corrosion initiated in the last
year)
Corrosion depth (Probability 8 mm to 7.6 mm Model: Time since initiation Failure pressure
of corrosion depth being in a 7.6 mm to 7.2 mm Corrosion rate x time0.5 Max corrosion rate
certain range) 7.2 mm to 6.8 mm
6.8 mm to 6.4 mm
6.4 mm to 6 mm
6 mm to 5.6 mm
5.6 mm to 5.2 mm
5.2 mm to 4.8 mm
4.8 mm to 4.4 mm
4.4 mm to 3.6 mm
3.6 mm to 3.2 mm
3.2 mm to 2.8 mm
2.8 mm to 2.4 m
2.4 mm to 2 mm
2 mm to 1.6 mm
1.6 mm to 1.2 mm
1.2 mm to 0.8 mm
0.8 mm to 0.4 mm
0.4 mm to 0 mm
Corrosion length (Probability of 1,000 mm to 190 mm Model: Time since initiation Failure pressure
corrosion length being in a 190 mm to 180 mm Max corrosion rate
certain range)

180 mm to 170 mm
∑
Corrosion rate x time
170 mm to 160 mm x P{pits coalescence to n pits}
160 mm to 150 mm n
150 mm to 140 mm
140 mm to 130 mm
130 mm to 120 mm
120 mm to 110 mm
110 mm to 100 mm
100 mm to 90 mm
90 mm to 80 mm
80 mm to 70 mm
70 mm to 60 mm
60 mm to 50 mm
50 mm to 40 mm
40 mm to 30 mm
30 mm to 20 mm
20 mm to 10 mm (continued on next page)
10 mm to 0 mm

APPENDIX B (continued)
Max corrosion rate (Probability 1 mm/y to 0.95 mm/y Field data [16] pH under coating Corrosion length
of localized corrosion rate 0.95 mm/y to 0.90 mm/y HCO3 concentration Corrosion depth
being in a certain range) 0.90 mm/y to 0.85 mm/y Chlorides/salinity
0.85 mm/y to 0.80 mm/y Sulfates
0.80 mm/y to 0.75 mm/y MIC
0.75 mm/y to 0.70 mm/y Effective potential
0.70 mm/y to 0.65 mm/y CP effectiveness
0.65 mm/y to 0.60 mm/y
0.60 mm/y to 0.55 mm/y
0.55 mm/y to 0.50 mm/y
0.50 mm/y to 0.45 mm/y
0.45 mm/y to 0.40 mm/y
0.40 mm/y to 0.35 mm/y
0.35 mm/y to 0.30 mm/y
0.30 mm/y to 0.25 mm/y
0.25 mm/y to 0.20 mm/y
0.20 mm/y to 0.15 mm/y
0.15 mm/y to 0.10 mm/y
0.10 mm/y to 0.05 mm/y
0.05 mm/y to 0.0 mm/y
Uniform corrosion rate 1 mm/y to 0.95 mm/y Field data [16] pH under coating Loss of the thickness
(Probability of uniform corrosion 0.95 mm/y to 0.90 mm/y HCO3 concentration
rate being in a certain range) 0.90 mm/y to 0.85 mm/y Chlorides/salinity
0.85 mm/y to 0.80 mm/y Sulfates
0.80 mm/y to 0.75 mm/y MIC
0.75 mm/y to 0.70 mm/y Effective potential
0.70 mm/y to 0.65 mm/y CP effectiveness
0.65 mm/y to 0.60 mm/y
0.60 mm/y to 0.55 mm/y
0.55 mm/y to 0.50 mm/y
0.50 mm/y to 0.45 mm/y
0.45 mm/y to 0.40 mm/y
0.40 mm/y to 0.35 mm/y
0.35 mm/y to 0.30 mm/y
0.30 mm/y to 0.25 mm/y
0.25 mm/y to 0.20 mm/y
0.20 mm/y to 0.15 mm/y
0.15 mm/y to 0.10 mm/y
0.10 mm/y to 0.05 mm/y
0.05 mm/y to 0.0 mm/y
Loss of thickness of the pipe 8 mm to 7.6 mm Model: Uniform corrosion rate x time Uniform corrosion rate Failure pressure
(Probability of loss of general 7.6 mm to 7.2 mm Time since corrosion initiation
thickness of the pipe being in 7.2 mm to 6.8 mm
a certain range) 6.8 mm to 6.4 mm
6.4 mm to 6 mm
6 mm to 5.6 mm
5.6 mm to 5.2 mm
5.2 mm to 4.8 mm
4.8 mm to 4.4 mm
4.4 mm to 3.6 mm
3.6 mm to 3.2 mm
3.2 mm to 2.8 mm
2.8 mm to 2.4 mm
2.4 mm to 2 mm
2 mm to 1.6 mm
1.6 mm to 1.2 mm
1.2 mm to 0.8 mm
0.8 mm to 0.4 mm
0.4 mm to 0 mm
Failure pressure 60 bar to 56 bar Model [15] Loss of thickness P > Pc
(Probability of failure pressure 56 bar to 52 bar Corrosion length
being in a certain range) 52 bar to 48 bar Corrosion depth
48 bar to 44 bar (Also material properties)
44 bar to 40 bar
40 bar to 36 bar
36 bar to 32 bar
32 bar to 28 bar
28 bar to 24 bar
24 bar to 20 bar
20 bar to 16 bar
16 bar to 12 bar
12 bar to 8 bar
8 bar to 4 bar
4 bar to 0 bar
Operating pressure (Probability 60 bar to 56 bar Pipeline operator or database — P>Pc
of operating pressure being in 56 bar to 52 bar
a certain range) 52 bar to 48 bar
48 bar to 44 bar
44 bar to 40 bar
40 bar to 36 bar
36 bar to 32 bar
32 bar to 28 bar
28 bar to 24 bar
24 bar to 20 bar
20 bar to 16 bar
16 bar to 12 bar
12 bar to 8 bar
8 bar to 4 bar
4 bar to 0 bar
P > Pc (Probability that the Yes Model — Failure pressure
operating pressure exceed No Operating pressure
the critical pressure)

Initiation rate (Probability of the 10,000/mile/y to 1,000/mile/y Field data: Visual survey of —
localized corrosion initiation 1,000/mile/y to 100/mile/y pipeline images
being in a certain range) 100/mile/y to 10/mile/y
10/mile/y to 0/mile/y
0/mile/y
Faults per mile (Probability of 10,000/mile to 1,000/mile Initiation rate x time pH under coating Faults per mile
faults per mile being in a certain 10,000/mile to 1,000/mile HCO3 concentration
range) 1,000/mile to 100/mile Chlorides/salinity
100/mile to 10/mile Sulfates
10/mile to 0/mile MIC
0/mile Effective potential
CP effectiveness
Probability of Failure Yes Mode1 P>Pc —
(Probability that a pipeline No if faults per mile x Prob(P>Pc)
will fail) >1 : Yes, Else, No

Reproduced with permission of copyright owner. Further reproduction
prohibited without permission.

Quantitive Assessment of Corro

Uploaded by

Copyright:

Available Formats

Quantitive Assessment of Corro

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Quantitive Assessment of Corro

Uploaded by

Copyright:

Available Formats

TECHNICAL ASPECTS OF THE RISK MANAGEMENT OF CORRODIBLE SYSTEMS

Quantitive Assessment of Corrosion Probability—

F. Ayello,‡,* S. Jain,* N. Sridhar,* and G.H. Koch*

ABSTRACT 16th and 17th centuries.1 One of the earliest recorded

where P(Ai|B) is the posterior probability of event Ai BAYESIAN NETWORK APPROACH

CORROSION—Vol. 70, No. 11 1129

1130 CORROSION—NOVEMBER 2014

CORROSION—Vol. 70, No. 11 1131

FIGURE 1. Layout of the internal corrosion model.

1132 CORROSION—NOVEMBER 2014

CORROSION—Vol. 70, No. 11 1133

1134 CORROSION—NOVEMBER 2014

not all data are of equal importance or even

CORROSION—Vol. 70, No. 11 1135

FIGURE 5. Layout of the external corrosion model.

1136 CORROSION—NOVEMBER 2014

CORROSION—Vol. 70, No. 11 1137

—the ability to deal with the complete lack of

Case Studies: Internal and External Corrosion

1138 CORROSION—NOVEMBER 2014

FIGURE 9. Layout of the model for SCC in pipelines.

CORROSION—Vol. 70, No. 11 1139

of high temperature. Figure 10 shows a typical profile

where I(X|Y) is the sensitivity factor (also referred to

1140 CORROSION—NOVEMBER 2014

CORROSION—Vol. 70, No. 11 1141

1142 CORROSION—NOVEMBER 2014

CORROSION—Vol. 70, No. 11 1143

1144 CORROSION—NOVEMBER 2014

CORROSION—Vol. 70, No. 11 1145

1146 CORROSION—NOVEMBER 2014

CORROSION—Vol. 70, No. 11 1147

You might also like