Sunteți pe pagina 1din 17

Paper No.

CORROSION2004
04556
A STATISTICAL RISK M O D E L TO PREDICT THE O C C U R R E N C E OF SCC

O.O. Youzwishen and A. Van Aelst


CIMARRON Engineering Ltd.
300, 6025 - 1 lth Street S.E.
Calgary, Alberta, Canada T2H 2Z2

P.F. Ehlers and A. Nettel


StatCaR
Dept. of Mathematics and Statistics, University of Calgary
2500 University Drive NW
Calgary, Alberta, Canada T2N 1N4

ABSTRACT

Near-neutral stress corrosion cracking (SCC) is an operational integrity problem experienced by pipeline
transportation companies since the 1970' s. Pipeline operators have used a number of different methods to predict
and locate SCC. Current in-line inspection technology allows for the detection of SCC in pipelines using
ultrasonic measurement. However, these tools have size limitations (not available for small diameter pipelines)
and can only accurately detect cracks above a certain threshold dimension. To date, predictive models have
focused mainly on establishing quantitative relationships between environmental factors and SCC formation and
growth. In general, the models used to predict SCC growth have been more successful than the models used to
predict the location of SCC formation.

In contrast to previous models that attempted to determine direct relationships between environmental parameters
and SCC formation, a model has been developed by statistically analyzing data pertaining to locations along a
pipeline where SCC was and was not found during field investigations. The data was analyzed using statistical
regression techniques and a multi-variable logistic regression model was created. The model was then applied to
a pipeline and verification digs were conducted. The results of the verification digs indicate that the model is able
to accurately predict locations with SCC.

Keywords: stress corrosion cracking, predictive model, near-neutral, integrity, regression

Copyright
2004 by NACE International. Requests for permission to publish this manuscript in any form, in part or in whole must be in writing to NACE
International, Publications Division, 1440 South Creek Drive, Houston, Texas 77084-4906. The material presented and the views expressed in this paper are
solely those of the author(s) and not necessarily endorsed by the Association. Printed in U.S.A.

1
Publication Right
Government work published by NACE International with permission of the author(s). Requests for permission to publish this manuscript in any form, in part
or in whole must be made in writing to NACE International, Publications Division, 1440 South Creek Drive, Houston, Texas 77084-4906. The material
INTRODUCTION

Stress corrosion cracking (SCC) is an operational integrity problem experienced by North American pipeline
transportation companies since the 1970' s. SCC is classified according to the degree to which the environment in
contact with the pipe surface is acidic or alkaline. SCC occurs in alkaline environments is referred to as classical
or high pH SCC, while SCC that occurs in acidic or neutral environments is referred to as non-classical or near-
neutral pH SCC. The latter has been the focus of much of the SCC-related work conducted in Canada in the past
decade and the type of SCC that has been found on the subject in this study.

Pipeline operators have used a number of different methods to predict and locate SCC in recent years. Currently
available in-line inspection (ILl) technology allows for the detection of SCC in pipelines using ultrasonic
measurement. However, these tools have size limitations and have limited detection tolerances. An alternative to
direct inspection is predictive modeling. Predictive modeling can also be complimentary. To date, predictive
models have focused mainly on establishing quantitative relationships between environmental factors and either
SCC formation or growth. In general, the models used to predict SCC growth have been more successful than the
models used to predict where SCC will form along a pipeline. Along with a pipeline operator (CIMARRON
Engineering Ltd. Client) CIMARRON Engineering Ltd. and StatCaR (University of Calgary) have developed a
statistical model to predict the occurrence location along the pipeline of the SCC.

Phase I

The model was developed in a phased manner. In Phase I, a preliminary model was used to identify several
locations where external corrosion and SCC were likely. The basis of the Phase 1 model was observations of
SCC on the subject pipeline, taken during ILI correlation dig programs. A basic statistical analysis was conducted
on the observations in order to identify trends related to SCC formation. Some of the findings were:
• 85% of SCC found was less than 6% of the wall thickness in depth,
• 90% of SCC was found in corrosion and interlinked,
• Many SCC colonies were found close to girth welds, and
• SCC found at creek and road crossings and bends tended to be deeper (greater than 12%).

Using these observations and previously researched SCC mechanisms, a weighted parametric model was
developed. Certain weighting factors were applied to parameters such as external corrosion depth, soil resistivity,
crossings and topography. A total score was calculated for each particular site and dig locations that scored high
were selected for field examination. A total of eight sites were excavated to validate the Phase 1 model. Of the
eight sites that were excavated, only two were found to have SCC (a 25% success rate). In comparison, 93 sites
were excavated previously for an ILI correlation program on the pipeline. These sites were chosen in order to
verify the corrosion levels reported by the ILI tool, not for their likelihood to have SCC. Of the 93 sites, SCC was
found at 45 (an occurrence rate of 48%).

Results of Phase I '

Comparing the Phase 1 model results to the ILI correlation dig results highlights the inability of the weighted
parametric model to improve SCC predictive capability. One contributing factor to this was that the development
of the weighted model was based on a subjective assignment of weighting factors for certain variables (and their
values). There was a higher probability for error when the weighting factors were assigned based on a judgment,
rather than a thorough and unbiased analysis of the data.

P h a s e II M o d e l

Using the learnings of Phase I of the project, a program to create an unbiased SCC predictive model using
detailed statistical analysis was developed by CIMARRON and StatCaR (University of Calgary) in Phase 2. The
Phase 2 model was constructed by conducting a detailed examination of all available data pertaining to SCC on

2
the subject pipeline, followed by a rigorous statistical analysis. The development and application of the Phase 2
model is discussed in this paper.
MODEL DEVELOPMENT

Background SCC Research


To establish the current body of knowledge on SCC, a review of published SCC research was conducted. The
goal of the review was to determine what variables, if any, have shown strong correlation to the formation of SCC
and thus should be focused on in the model. To date, research on near-neutral pH SCC is limited. However, the
following observations have been made with respect to initiation of near-neutral pH SCC:
• It is often associated with pitting and general corrosion ~,
• Non-metallic inclusions can promote or initiate SCC ~2~,
• When pipe stresses are high, very little fluctuation of stress is required to induce SCC ~3~,
• Residual stresses combine with operational stresses to increase the overall stress state in the pipe and thus,
increase the probability of SCC formation ~3~,
• Problems with external coatings are more common near bends ~4~,
• There is no direct correlation between pipe temperature and near-neutral SCC formation (5~,
• Ground depressions (on level terrain) have been found to be an aggressive topography that leads to SCC
initiation (6) and
• Seasonal changes to moisture levels at pipe depth correlate stronger to SCC formation than instances
where the moisture level is static (either constantly wet or constantly dry) (7~.

Data Collection and Organization


Data from several sources was collected before the statistical analysis stage of the project was initiated. Data for
the pipeline of interest was acquired from several sources including:
• Site information collected from previous investigative digs (corrosion deposits, soil properties,
topography, drainage, etc.),
• System information from a GIS database (crossings, close interval cathodic protection (CP) and depth of
cover data, pipeline profile, GPS survey coordinates),
• A steady-state hydraulic analysis of the pipeline,
• Historical SCADA readings and
• Anomaly readings (dents, metal loss, laminations) from previous ILI's of the pipeline (ultrasonic (UT)
and magnetic flux leakage (MFL)).

The data collected for the pipeline was divided into two (2) categories, site-specific and non-site-specific. Data
placed into the site-specific category was that information which was specific to a location along the pipeline
where SCC was either found or not found (i.e. at the investigative dig sites). All data at a particular dig site was
labeled with the GPS survey chainage of the reference girth weld at the site. Non-site-specific data was data that
was distributed along the length of the pipeline and was not necessarily specific to a dig site. This data included
close interval CP data, ILI anomaly readings, crossings, etc. Each data point in the non-site-specific category was
spatially tied to the pipeline via a GPS survey chainage as well.

Site-specific data was also obtained for another pipeline. A significant number of investigative digs had been
conducted on this second pipeline and the information was used to determine if strong SCC predictors were
immediately evident.

Site-Specific Data. Major investigative dig programs were conducted on the subject pipeline in recent
years. The primary goal of these dig programs was to mitigate defects that were identified during ILI runs (dents,
laminations, metal loss). During these digs, SCC was discovered. Since SCC occurred at some sites and not at
others, data specific to the sites where SCC occurred could be compared to data specific to non-SCC sites, to
determine if there were factors that correlated to SCC initiation.

3
161 investigative digs were performed on the subject pipeline in the last seven (7) years. In order to properly use
the data in the statistical analysis, the type of data (i.e. variables measured) had to be consistent among the data
sets. Because of this, the resulting data set that was used for the statistical analysis consisted of 93 of the 161
sites. Variables such as drainage, topography, land usage and soil properties comprised the site-specific data set.

Non-site-specific Data. In addition to the site-specific measurements, other parameters were believed to
affect the formation of SCC. These parameters were typically related to physical features of the pipeline or other
non-environmental variables. They included features such as: CP level, proximity to bends, proximity to
crossings, pipe slope, etc. Since this data was available for the entire length of the pipeline and not collected at
particular dig sites, it was categorized as non-site-specific data.

Two types of non-site-specific data were collected. The first type was data that was continuous along the
pipeline. For example, variables such as CP level, cover depth, slope and internal pressure had values for the
entire length of the pipeline. To determine if these variables influenced the formation of SCC, their values at the
dig sites (i.e. where SCC was known to occur or not to occur) had to be determined. Because the variables were
continuous and measured at close intervals (less than 100 m), a linear interpolation was applied to determine their
values at the chainages of the dig sites.

The second type of non-site-specific data acquired was discrete data. This type of data only occurred at specific
points along the pipeline. Features such as crossings (water body, pipeline, road and railway) and bends only
occur at specific points and cannot be interpolated. Thus, the correlation between the proximity of these features
to SCC and the likelihood of SCC was determined in the statistical analysis.

The non-site-specific variables collected were as follows:


• CP On, CP Off, CP Shift,
• Pipe Slope,
• Chainage From U/S Pump Station, Max. Pressure, R-Ratio,
• Total Discharge Pressure Cycles per Year, Discharge Pressure Cycles per Year > 300 kPa, R-Ratio
Histograms,
• Crossings,
• Bends,
• Dents (UT), Metal Loss (UT), Metal Loss (MFL).

Site-Specific Data (Secondary Pipeline). In addition to site-specific dig information collected on the
subject pipeline, an extensive database from digs on a second pipeline with a history of SCC was obtained. Over
600 investigative digs were conducted on the second pipeline (with SCC being found at over 300 sites). Because
of some of the similarities between the two (2) pipeline systems (coating type, construction era, geographic
proximity, liquid service), the data from the secondary pipeline was analyzed to determine if any site-specific
variables correlated strongly with the initiation of SCC along the pipeline. Variables such as drainage,
topography, land usage and soil properties also comprised the site-specific data set for the secondary pipeline. It
should be noted that this data set was not integrated into the site-specific data set of the subject pipeline. A
separate, simple statistical analysis was performed on the secondary site-specific data set.

Statistical Analysis
The second stage of the project involved statistical analysis of the data collected in order to create a predictive
model for SCC on the pipeline of interest. The primary goals of this stage were: to determine what set of
parameters best predict the occurrence (or absence) of SCC, to generate a model that calculated the probability of
SCC at any location along the pipeline, and to test the model against known SCC locations in order to determine
accuracy.

Previously developed SCC predictive models have focused on the environment in the immediate area of the SCC
colony. As such, a large amount of data relating to soil conditions and composition and the environment in the

4
vicinity of the excavation site has been collected during investigative digs where SCC was found. To utilize this
data, the first portion of statistical analysis focused on the site-specific variables. Site-specific data sets from the
primary and secondary pipeline were analyzed for correlation to SCC.

Site-Specific Data Analysis (Secondary Pipeline). Since the site-specific database for the secondary
pipeline was much larger than that of the subject pipeline, it was analyzed first. The data was separated based on
whether or not SCC was found at the site. Comparative barplots for each variable were studied to determine if
any single variable had an influence on the occurrence of SCC. Figure 1 illustrates a typical barplot generated
from the site-specific data of the secondary pipeline. Figure 1 shows that for local topography, there is no
appreciable difference between the presence and absence of SCC. This figure is typical of the other variables
analyzed for the site-specific variables.

Boxplots were also generated for this data set. Boxplots are a convenient graphical device for comparing two (or
more) groups of data in terms of a numerical predictor variable. The boxes show the range of the middle 50% of
the data values. The central dot is at the median value of the numerical variable. The major portion of the data is
enclosed in the brackets. Points well outside the whiskers may be considered unusual or indications of a non-
symmetrical distribution. Figure 2 illustrates a boxplot for the depth of cover at the dig sites. This figure
indicates that depth, by itself, is a poor predictor of SCC occurrence since the median depth is about the same for
both conditions. Given the complexity of SCC, the lack of correlation to any one (1) variable is reasonable.

One (1) variable that did have strong correlation was the "occurrence of SCC with corrosion". Approximately
95% of the occurrences of SCC showed SCC in conjunction with corrosion. This indicates that corrosion (which
can be readily measured using ILI technology) may provide evidence of locations more susceptible to SCC.

Because little or no correlation between single environmental variables and the occurrence of SCC was observed,
two- and three-variable graphical analysis was performed. There was no significant correlation observed between
SCC and no SCC for two (2) and three (3) variable sets. Again, this is likely due to the complex nature of SCC.
Graphical analysis of the site data from the secondary pipeline did not provide additional insight into the
occurrence of SCC, above that which was obtained through the literature review. Therefore, the focus of the
statistical analysis was shifted to the subject pipeline.

Site-Specific Data Analysis (Subject Pipeline). Site-specific data for the subject pipeline was first
analyzed using one-, two- and three-variable graphical analysis in a similar manner as was done with the
secondary pipeline. Correlation of site-specific variables to SCC occurrence was also limited for the primary data
set. The strong correlation of SCC with corrosion observed on the secondary pipeline was also observed on the
subject pipeline. Figure 3 shows a typical barplot for the subject pipelines. A boxplot for the depth of cover
variable on the pipeline of interest is shown in Figure 4. Again, no correlation between depth of cover and SCC
formation is directly observed. Two- and three-variable graphical analysis was also performed. A typical
threevariable plot is shown in Figure 5. In this figure, depth of cover (in metres) and drainage (well, imperfect,
poor, very poor) form the base plane and the CP shift (negative of the on minus off potential in millivolts) forms
the vertical axis. For the various drainage categories, there is no clear distinction between the SCC sites and non-
SCC sites at differing levels of depth or CP shift. Figure 5 also shows that graphical analysis becomes
cumbersome when multiple variables are examined simultaneously.

Combined Data Analysis (Subject Pipeline). Because of limited correlation between site-specific
variables and SCC occurrence in the graphical analysis, a numerical multi-variable analysis was performed.
Several methods of multi-variable analysis were contemplated (principal component analysis, multiple
contingency tables, logistic regression, etc.). Logistic regression was chosen as the best method in order to
achieve the objectives of the model (ease of use, adaptability, etc.). Logistic regression fits a multi-variable linear
equation to the input data set. The result of the regression is an estimation of the probability of the occurrence of
SCC calculated as a function of a set of predictor variables.

Numerical multi-variable analysis allowed a larger number of variables to be examined simultaneously for their
correlation to SCC formation. Initially, the site-specific data set was augmented with the continuous variables

5
from the non-site-specific data set. The continuous variables were examined first because values could be easily
determined at the dig sites. Logistic regression of the site-specific data set resulted in the following predictors"
• Elevation,
• C P On,
• Maximum Pressure (primary pump stations operating only),
• Total Pressure Cycles per Year,
• Pressure Cycles Greater than 300 kPa per Year,
• Drainage,
• Presence of Cobbles,
• Morainal Soil Deposit,
• Site Location,
• Topography.

While these predictors proved to be the strongest amongst the other site-specific variables, they yielded high
levels of uncertainty and thus this model (designated as the Preliminary Model) was not considered acceptable for
accurately predicting SCC. The distribution of uncertainty for this model is shown on the right side of Figure 6.
Figure 6 illustrates the distribution of total error (plus and minus) for the predicted SCC probability from the
model. The model constructed with the site-specific and continuous non-site-specific data has an error
distribution that is skewed towards higher errors. Because of the level of inaccuracy of this model, discrete non-
site-specific variables were added to the data set and analyzed.

Because the discrete data was not related to specific dig sites, a new approach had to be created in order to link
this data to dig sites, where the status of SCC was known. A technique of looking upstream and downstream of a
dig site and determining if the variable in question was present and what its magnitude was (if present), was
established. The examination range was set at 20 metres upstream and downstream of the reference girth weld of
the dig site. The distance of 20 metres was deemed to be sufficiently close to the site to capture the effects of the
discrete variables on SCC at the site and yet large enough to ensure a reasonable chance of capturing available
discrete data.

The majority of the digs used in the site-specific database (58 out of 93) were situated within the first quarter of
the pipeline's length. In contrast, the non-site-specific data (bends, dents, crossings, metal loss) was evenly
distributed along the pipeline. To compensate for the skewness of the data, the models were generated in two
stages. In the first stage, the predictor selection stage, predictor variables were selected based only on the data
from the first 58 sites. The model coefficients were then generated using data from all of the 93 dig sites.

Several logistic regression models were created for the combined data set. The model with the strongest
predictors (designated at the Final Model) was composed of the following variables:
• CP on potential,
• CP shift (on minus instant off potential),
• Presence of a ground depression,
• Bend angle of the pipe,
• Direction of the bend (vertical or horizontal),
• Proximity to metal loss (and whether the metal loss was near a girth weld),
• Metal loss severity.

Using these variables, the probability of SCC occurrence at any location along the pipeline could be estimated.
The predicted SCC probability was calculated using the following equation:

Pr(SCC)- exp(g) , (1)


1+exp(g)

6
where exp() is the exponential function with natural base and g is the following linear function of the predictor
variables:

g = b o + blV 1 + bzV 2 + b3V3 + b4V4 + bsV5 + b6V6 + bTV7 + bsV8 + b9V9 • (2)

where bo, b~, ..., bN are the coefficients determined during the statistical analysis and V~, V2, ..., VN are the values
of the predictor variables. Equations 1 and 2 can be readily implemented into a spreadsheet program for
application to the subject pipeline.

Statistical Model Testing and Accuracy


The Final Model was examined in detail and tested to determine its accuracy. Because the model was multi-
variant in nature, it could not be examined graphically in its entirety. To examine the model graphically,
combinations of variables were fixed at certain values and influence of two (2) variables at a time on the
probability of SCC was analyzed. A typical two-variable plot is shown in Figure 7. In Figure 7, the effects of the
bend angle (for vertical bends) and metal loss severity (as determined by the Modified B31-G Estimated Repair
Factor (ERF)) on SCC probability are illustrated. To generate this plot, the pipeline was assumed to be in a
depression, the CP variables were fixed to their median values and metal loss was assumed to be adjacent to the
pipe girth weld. Figure 7 shows that as the severity of metal loss in the vicinity increases, the probability of SCC
decreases. This trend may be attributed to corrosion growth in excess of SCC growth (which would result in the
elimination of any SCC that may have developed at the location). While the occurrence of SCC on the pipeline
has been shown to be a factor of many interacting variables, the three-dimensional plots allow for some intuitive
validation of the model's predictions.

To determine the most accurate version of all the models generated, misclassification was used as a performance
diagnostic. The coefficients for each model under consideration were calculated, using data from the first 58 sites.
The resulting model was then used to calculate the probability of SCC at all 93 sites. Finally, the probability of
SCC was used to classify each site as either having SCC (if the probability was high) or not having SCC (if the
probability was low). A misclassification score was calculated based on the number of times each model was
correct in predicting whether or not SCC occurred at a known site. A model was rated more favorably if it
achieved a lower misclassification score.

It was determined to be unreasonable to classify a site with estimated probability of SCC equal to 0.51 as an SCC
site, while classifying a site with probability of SCC equal to 0.49 as not having SCC. Therefore it was decided to
leave sites "unclassified" if their estimated probability of SCC was not sufficiently high or low. The probability
levels at which the sites would be deemed "classified" (referred to as the cutoffs) were selected such that if the
probability of SCC exceeded the high cutoff, then the site was classified as having SCC. If the probability was
below the low cutoff, then the site was classified as not having SCC. Sites with probabilities between the cutoffs
were "unclassified". The two (2) cutoffs were chosen symmetrically above and below 0.5.

The logistic regression analysis led to four models as worthy of consideration (including the "Preliminary" and
"Final" Models presented in this paper). Figures 8 and 9 depict the misclassification results for the Preliminary
Model and Final Model, respectively. Each plot shows the misclassifications as functions of the cutoff setting.
The high/low cutoffs are taken to be symmetric and thus only the high cutoff is shown. For example, a cutoff of
0.6 means that the high cutoff is 0.6 and the low cutoff is 0.4. All sites are classified at a cutoff of 0.5 and no site
is classified at a cutoff of 1.0. Curves show the misclassification percentage for SCC sites, the misclassification
percentage for non-SCC sites and the overall misclassification percentage as well as the percentage of unclassified
sites. Clearly, all these percentages depend on the choice of cutoff value. Ideally, all curves would be low,
especially for cutoff values below 0.7 (the curves are of little value for cutoffs above 0.75, for example, since that
means not classifying any site with estimated probability of SCC between 0.25 and 0.75).

Figure 8 shows that the Preliminary Model had a reasonable non-classification curve, but that the
misclassification percentages were high. Therefore it was not the optimum model. Conversely, Figure 9 shows

7
that the Final Model is the preferred model because its non-classification rate was not excessive and the
misclassification rates were comparatively low.

The uncertainty in predicted values of SCC probability was a function of the magnitudes of the predictor
variables. In general, uncertainty was smallest where predictors had values near their average values and largest
where predictors were more variant. Figures 10 (Preliminary Model) and 11 (Final Model) illustrate the variance
of uncertainty versus the predicted SCC probability (dark markers on the plots). In general, the uncertainty
margins are larger for the Preliminary Model. This is reflected in Figure 6 as well. The left side of Figure 6
shows the distribution of uncertainty for the Final Model. In contrast to the uncertainty distribution for the
Preliminary Model (right side of Figure 6), the distribution for the Final Model is skewed towards the lower end
of the spectrum. The mean uncertainty for the Final Model was approximately +0.22 (based on a 90% confidence
measure).

Expanding the database of known SCC or non-SCC sites on the pipeline will reduce the uncertainty in the model.
For example, data from another 10, 25, 50 or 100 sites will yield reductions of approximately 5%, 10%, 20% and
30%. Moreover, the model coefficients may change sufficiently to improve prediction even further.
Improvements to the accuracy of the predictions can also be achieved by using data more evenly distributed along
the pipeline's length (rather than using only the first quarter of the pipeline).

RESULTS

The Final Model presented in this paper was applied to the subject pipeline in order to predict the most likely
locations of SCC formation. Input data was taken from the non-site-specific data sources cited previously and
consolidated into a single spreadsheet. Equations 1 and 2 were entered into the spreadsheet and SCC probabilities
were calculated. SCC probabilities were calculated at the mid-length of each pipe joint along the pipeline (at
approximate intervals of 20 metres). Since the uncertainty of the SCC probability was dependent on the input
variables, positive and negative error values were also calculated for each point using Equations 1 and 2 (with
modified coefficients in Equation 2). Table 1 summarizes typical calculation points and the resulting SCC
probabilities (and uncertainties).

A preliminary investigative dig program was conducted on the pipeline in order to verify the accuracy of the
Phase 2 model. Pipe joints with SCC probabilities greater than or equal to 0.9 were identified from the complete
version of Table 1 and examined for excavation feasibility. At total of ten sites with high SCC probabilities were
investigated. In addition one (1) site with a lower SCC probability (approximately 0.7) was excavated. Of the ten
sites with SCC probabilities in excess of 0.9, seven (7) were found to have SCC (a success rate of 70%). The sites
where SCC was found displayed typical coating disbondment patterns seen with near-neutral SCC ~8~(see Figure
12). Corrosion was generally light and would not normally be noted for investigation, based on conventional
metal loss ILI's (see Figure 13). Estimated SCC depths ranged from 5% to greater than 50% of the pipe wall (see
Figures 14 and 15 for examples). SCC was not discovered at the site with the lower SCC probability. Therefore,
the Phase 2 model correctly predicted the presence (or absence) of SCC eight (8) times out of 11 (for a success
rate of 73%). This is a significant improvement over predictive models previously applied to the pipeline and
additional investigative digs are being considered for the near future based on the model.

CONCLUSIONS

A simplified equation to determine the probability of SCC along a subject pipeline has been developed. Site-
specific data collected during investigative dig programs on the subject pipeline and a secondary pipeline was
vigorously analyzed using statistical methods. This data, on its own, generated very weak predictions of SCC.
When this data was combined with other non-site-specific variables and analyzed using logarithmic regression,
the ability to predict the occurrence of SCC was increased substantially. Misclassification analysis was conducted
to determine the optimized regression model. The predictors that indicated the greatest correlation to SCC were:
• CP on potential,
• CP shift (on minus instant off potential),

8
• Presence of a ground depression,
• Bend angle of the pipe,
• Direction of the bend (vertical or horizontal),
• Proximity to metal loss (and whether the metal loss was near a girth weld),
• Metal loss severity.

The model that has been created is easily applied to the subject pipeline in order to determine the probability of
SCC at a large number of locations. The model is also easily adaptable. With very little effort, the model can be
regenerated with new coefficients or even new predictor variables, when data from new excavation sites or new
in-line inspection data is obtained. Finally, the current model incorporates predictor variables that can be
ascertained or measured without excavation of the pipeline. This allows the model to be applied over the entire
length of the pipeline, without costly excavations.

Through a preliminary investigative dig program, the predictive model has been confirmed to be a valid method
for identifying SCC on the subject pipeline. Additional investigative digs are planned and the information
gathered will aid in further refining the current version of the model.

By examining the occurrence of SCC using statistical methods, the ability to predict the occurrence of SCC along
a subject pipeline has been achieved. By developing this model, the pipeline operator has greatly enhanced the
ability to assess the likelihood of SCC along its pipeline and has shown due diligence in mitigating the risks
associated with this pipeline integrity concern.

REFERENCES

° PII Technical Note, Stress Corrosion Cracking in Pipelines: Causes and Remedies, PII Group Ltd,
Houston, Texas, 2000.
. M. Elboujdaini, Y. Z. Wang, R. W. Revie, Initiation of Stress Corrosion Cracking on X-65 Linepipe
Steels in Near-Neutral pH Environment, 2000 International Pipeline Conference - Volume 2, ASME,
Calgary, Alberta, p. 967-978, 2000.
. W. Zheng, R. Sutherby, R. W. Revie, W. R. Tyson, G. Shen, Stress Corrosion Cracking of Linepipe
Steels in Near-Neutral pH Environment: Issues Related to the Effects of Stress, Environmentally Assisted
Cracking: Predictive Methods for Risk Assessment and Evaluation of Materials, Equipment, and
Structures, ASTM STP 1401, Kane, Ed., American Society for Testing and Materials, West
Conshohocken, PA, 2000.
. G. King, Physical Forces on Buried Pipeline Coatings, 2000 International Pipeline Conference- Volume
2, ASME, Calgary, Alberta, p. 703-708, 2000.
. J. A. Beavers, SCC Direct Assessment, NACE International Pipeline Integrity Management Seminar,
NACE International, Calgary, Alberta, p. 13, 2002.
. J. A. Beavers, R. G. Worthingham, The Influence of Soil Chemistry on SCC of Underground Pipelines,
2002 International Pipeline Conference- CD ROM, ASME, Calgary, Alberta, p. 1-8, 2002.
. J. K. Patel, A Brief Report on the Influencing Factors on SCC in Buried Pipelines, NIOBIUM Products
Company GnbH Publication, NIOBIUM Products Company GnbH, Germany, p. 109-129, 2001.
. National Energy Board, Public Inquiry Concerning Stres~ Corrosion Cracking on Canadian Oil and Gas
Pipelines, National Energy Board, p. 23-24, 1996.

9
TABLE 1
S A M P L E INPUT AND SCC P R O B A B I L I T Y C A L C U L A T I O N F R O M THE P H A S E 2 M O D E L

Metal
CP Bend Bend Metal Metal
Chainage CP On Bend Loss Error Error
• Angle L o s s + / - Loss P(SCC) (_)
(m) Shift Depr.9 +/- Direction o) Adj. to ERF (+)
(-mV) (-mV) 20m? ( 20m? GW?

4494.4 1379 188 no no n/a 0.0 yes yes 0.921 0.65 0.29 0.21
7251.3 1304 109 no yes over 9.3 yes no 0.944 0.87 0.27 0.10
7676.3 1328 121 yes yes sag 11.5 yes no 0.930 0.96 0.16 0.04
22848.1 1126 46 yes yes right 8.2 yes no 0.860 0.83 0.34 0.13
29064.0 1177 59 no yes over 7.3 no n/a 0.839 0.74 0.43 0.20
38441.1 1366 132 no yes sag 5.0 yes no 0.921 0.90 0.25 0.08
47360.6 1427 185 yes yes sag 6.5 yes no 0.855 0.97 0.13 0.02
63299.9 1353 64 no no n/a 0.0 yes yes 0.841 0.84 0.30 0.12
66187.6 1363 50 no no n/a 0.0 yes no 0.853 0.59 0.26 0.22
68502.0 1374 43 no yes over 8.4 yes no 0.842 0.96 0.20 0.04
68815.7 1345 72 no yes sag 13.6 no n/a 0.839 0.83 0.42 0.14
84218.3 1679 355 yes yes sag 7.5 yes no 0.846 0.97 0.14 0.02
100848.9 1610 255 no no n/a 0.0 yes no 0.842 0.56 0.26 0.23
107411.7 1653 363 yes yes over 5.4 yes no 0.841 0.97 0.14 0.02
107609.8 1624 397 no yes sag 12.6 no n/a 0.839 0.70 0.43 0.24
149240.8 1667 518 yes yes over 6.8 no n/a 0.839 0.82 0.42 0.15
149643.4 1683 534 no yes sag 16.2 no n/a 0.839 0.56 0.40 0.34

B SCC ii No SCC

n/a

Undulating
.c
g2.
x_
O'} Rolling
O
Q.
O
I-
l Level
o
0
_J
Inclined
/

Depressional
_! I

0 10 20 30 40 50 60
% of Sites

F I G U R E 1 - Distribution of SCC vs. No SCC for Local Topography (Secondary Pipeline)

10
I i I I

No SCC 0 ............... 0 0 0

0
SCC O0 (~ 0 0 0
.........

I I I I

1 2 3 4

Depth
F I G U R E 2 - Boxplot of S C C vs. No S C C for C o v e r D e p t h ( S e c o n d a r y Pipeline)

B SCC n No SCC

Well

c Poor

Imperfect

10 15 20 25 30 35 40 45
% of S i t e s

F I G U R E 3 - Distribution of S C C vs. No S C C for D r a i n a g e

11
I I I !

O
SCC O O

O
No SCC (30 O O O O

L l
I I I

1 2 3

Cover Thickness

F I G U R E 4 - Boxplot of S C C vs. No S C C for Cover D e p t h

Drainage vs Cover Thickness vs CP Shift


separated by SCC ...
..

• NO
...-
..-
..
.. :,~:i:: YES
. .- ' . .
_.-

. ''- .

".,.
0
O-

''....,.,.,.,.
0 .
.

°
t
• ...,.............,,.
..........
..:.~....................
-... .. . . . . . . . . . . . . . i i i~,
. .... ...... '~i!i!i!ii!i!ii~::' ,q
0

t.--

~ ° . . - ' '

%!!!

:::::::::::::::::::::::::::::::::::::::::::::::::::::::::
f

o0

F I G U R E 5 - Drainage vs. Cover Depth vs. CP Shift for S C C and No S C C

12
With mechanical data Without mechanical data

Min.:O.1001, 1st Qu.:0.3485, Median:O.4335, Min.:O.O01, 1st Qu.:0.3171, Median:O.5271 ,


Mean:O.4379, 3rd Qu.:0.5276, Max.:O.7253 Mean:O.4616, 3rd Qu. :0.6169, Max.:O.8689

I I I I I I I I I I I !

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0,2 0.4 0.6 0.8 1.0
Interval widths Interval widths

F I G U R E 6 - Uncertainty Distribution for 93 Predictions (Final and Preliminary Models)

~3
O o
O
¢J9
{D

~d
O

F I G U R E 7 - Probability of SCC vs. Bend Angle (Vertical) and Metal Loss (ERF)

13
...... SCC reading=NO SCC,Predicted SCC= SCC
....... SCC reading=SCC,Predicted SCC=NO SCC
1.0 --
Total misclassified
nonclassified /
/
./
....... \ "° . . . . . . . . f
0.8 -- /

0.6-
c'-
o
(1)
~,.\ -.
0.4 J
/ • "k\ "...

• ~.,~ "...

"\"'-....
/,... j'- --'J
7
0.2 --
r..-/
/
.,t
/--

O0

I I I I I I I

0.4 0.5 0.6 07 08 0.9 1.0

cutoff

F I G U R E 8 - M i s c l a s s i f i c a t i o n C u r v e s for P r e l i m i n a r y M o d e l

SCC reading=NO SCC,Predicted SCC= SCC


....... SCC reading=SCC,Predicted SCC=NO SCC
1.0 m
Total misclassified
nonclassified

/
0.8 ~
/--
/
/.-.
/-
T-"
/- /
0.6 -
t-- J
f
o
~ r
t---J
~- 0.4 J
/

\. ~ f
0.2 --

_/ ~ - - . . - . .x.....
_.J
0.0

I I I I I I I

0.4 0.5 0.6 0.7 0.8 0.9 1.0

cutoff

F I G U R E 9 - M i s c l a s s i f i c a t i o n C u r v e s for Final M o d e l

14
I I I I I I

0.0 0.2 0.4 0.6 0.8 1.0

Probability

FIGURE 1 0 - Confidence Intervals for Predicted Probabilities (Preliminary Model)

_ f

m m
i
~ . ..,.,.,,.,,,.,,.-

I I I I I I

0.0 0.2 0.4 0.6 0.8 1.0

Probability

FIGURE 11 - Confidence Intervals for Predicted Probabilities (Final Model)

15
F I G U R E 1 2 - Typical Coating Disbondment Pattern from Validation Digs

FIGURE 13 - Corrosion Deposits Under Coating Disbondment

16
FIGURE 1 4 - Light Metal Loss and SCC (After Blasting)

FIGURE 15 - Significant SCC Discovered (Greater Than 50% Deep)

17

S-ar putea să vă placă și