Vous êtes sur la page 1sur 6

1 Copyright 2012 by ASME

MODEL ERROR ASSESSMENT OF BURST CAPACITY MODELS FOR CORRODED


PIPES


W. Zhou
Western University
London, Ontario, Canada
G. (Terry) Huang
TransCanada Pipelines
Calgary, Alberta, Canada
S. Zhang
Western University
London, Ontario, Canada


ABSTRACT
The model errors associated with five representative burst
pressure prediction models, namely B31G, B31G Modified,
DNV, PCORRC and RSTRENG, for corroded pipelines are
evaluated based on a relatively large number full-scale burst
tests on corroded pipes reported in the literature. All the test
specimens in the database contain single isolated real corrosion
defects. The means, coefficients of variation (COVs) and
probability distribution of the model errors for the considered
burst capacity models are derived based on the test-to-predicted
burst pressure ratios for the collected test data. A numerical
example is used to illustrate the impact of the model error on
the probability of burst of the corroding pipeline.
INTRODUCTION
In the reliability-based design and assessment of oil and
gas pipelines, it is essential to accurately evaluate the model
errors associated with the deterministic pipe capacity models
corresponding to various limit states and incorporate these
model errors in the reliability analysis. Metal-loss corrosion is
one of the most common threats to the structural integrity of oil
and gas pipelines. The prediction of the burst pressures of
corroded pipelines is of significant relevance to the pipeline
industry. Many burst pressure prediction models have been
developed in the past, e.g. the well-known ASME B31G,
RSTRENG and PCORRC (see, e.g. [1]). It has been reported
that the calculated burst probability of a corroded pipeline is
highly sensitive to the model error associated with the burst
pressure prediction model [2]. This highlights the significance
of evaluating the model errors for the various burst pressure
prediction models for corroded pipelines.
To accurately evaluate the model error for a given burst
pressure prediction model, a systematic assessment that is
based on a relatively large number of full-scale burst test data is
needed. Chauhan et al. [1] recently carried out a study to
review the prediction accuracies of six commonly used burst
pressure prediction models for corroded pipelines. They
collected a total of 313 data points of full-scale burst tests of
corroded pipe specimens and used these data to evaluate the
model errors associated with the six burst pressure prediction
models. The burst test database established in their study
included 180 data points involving specimens containing
machined (i.e. artificial) defects and 133 data points involving
specimens containing real (i.e. natural) corrosion defects.
Note that there is a significant difference between these two
groups of data in that the defect depth profile (i.e. the profile
projected on the vertical plane in the longitudinal direction of
the pipeline) for a machined defect is typically rectangular (i.e.
uniform defect depth) whereas the defect depth profile for a
natural corrosion defect is irregular and cannot be characterized
by a simple geometry such as a rectangle or parabola. Based
on the test data reported by Chauhan et al. [1], we found that
the mean and COV of the test-to-predicted ratios for the B31G
Modified criterion calculated using 118 burst test data
involving real defects are equal to 1.342 and 25.5%
respectively, whereas the mean and COV of the test-to-
predicted ratios for the same model calculated using 167 burst
test data involving machined defects are equal to 1.071 and
14.9%. The large difference between the two sets of statistics
suggests that machines defects should be separated from natural
corrosion defects in the model error assessment; however, no
differentiation between the two groups of data was considered
in Chauhan et al.s study [1]. Therefore, the model errors
reported in their study may not be appropriate for practical
applications, that is, the reliability evaluation of pipelines
containing real corrosion defects.
The objective of the study described in this paper was to
evaluate the model errors associated with several representative
burst pressure prediction models for corroded pipelines using a
relatively large number of full-scale burst test data collected
from the literature. Only test specimens containing single
isolated real corrosion defects were considered in the model
error assessment. The model errors were characterized based
on the test-to-predicted ratios corresponding to the data points
Proceedings of the 2012 9th International Pipeline Conference
IPC2012
September 24-28, 2012, Calgary, Alberta, Canada
IPC2012-90335
2 Copyright 2012 by ASME
included in the database. The means, standard deviations and
coefficients of variation (COV) of the test-to-predicted ratios
for the considered burst capacity models were calculated. The
probability distributions for the model errors were
recommended based on the distribution fitting techniques.
Furthermore, the impact of the model error on the reliability
analysis of corroded pipelines is investigated through a
numerical example.
MODEL ERROR ASSESSMENT OF BURST CAPACITY
MODELS FOR CORRODED PIPES
Burst Capacity Models
We evaluated the model errors for five representative burst
models, namely ASME B31G [3], B31G Modified, RSTRENG
[4], the DNV model [5] and PCORRC [6]. These models are
briefly summarized in the following. For the B31G, B31G
Modified and DNV models, the burst pressure of a pipe at a
given corrosion defect, P
b
, is calculated as follows:

|
|
|
|
.
|

\
|

=
o
o
flow b
MA
A
A
A
D
t
P
1
1
2

(1a)

=
DNV
Modified B31G & 31G B
t D
D
D
(1b)

+ =
DNV
Modified B31G 95 . 68
B31G 1 . 1
u
y
y
flow
MPa

(1c)

>

=
20
20
3
2
2
max
2
max
Dt
L
Ld
Dt
L
Ld
A
B31G (1d)

=
DNV
Modified B31G 85 . 0
max
max
Ld
Ld
A
(1e)

>
+
=
20
20 8 . 0 1
2
2 2
Dt
L
Dt
L
Dt
L
M
B31G (1f)

> +
+
=
50 0.032 3.3
50 0.003375 - 6275 . 0 1
2 2
2
2 2
4 2
Dt
L
Dt
L
Dt
L
t D
L
Dt
L
M

B31G Modified (1g)

31 . 0 1
2
Dt
L
M + =
DNV (1h)

where t and D denote pipe wall thickness and outside diameter,
respectively;
y
and
u
denote the pipe yield strength and
tensile strength, respectively; L and d
max
denote the length (i.e.
in the pipe longitudinal direction) and maximum depth (i.e. in
the through pipe wall thickness direction) of the corrosion
defect, respectively, and A
0
= Lt is a reference area.
The prediction equation for RSTRENG is the same as that
for B31G Modified except that the parameter A is determined
using the effective area method based on the river-bottom
profile of the corrosion defect. B31G, B31G Modified and
RSTRENG are limited to defects with d
max
/t 0.8, whereas the
DNV model is limited to defects with d
max
/t 0.85. The burst
pressure equation corresponding to PCORRC is given by

(
(

|
|
.
|

\
|
|
|
.
|

\
|

|
.
|

\
|
=
2 / ) (
157 . 0
exp 1 1
2
max
max
u b
d t D
L
t
d
D
t
P
(2)

PCORRC is limited to defects with d
max
/t 0.85 and L 2D
[1].
Burst Test Database of Corroded Pipes
To evaluate the model errors associated with the
aforementioned burst capacity models, a database of full-scale
burst test data for pipes containing single isolated real corrosion
defects was established based on the burst test data published in
the literature. The data selection criteria are as follows:
The specimens are thin-walled, i.e. D/t 20, and subjected
to internal pressure only.
The specimens contain single isolated real corrosion
defects only.
The length of the specimen is greater than twice the outside
diameter of the specimen. This ensures that the boundary
condition of the test specimen is similar to that of the real
pipelines, which are typically long and have negligible
axial strains under internal pressure [7].
The test data should meet the applicability limit (e.g. the
d
max
/t ratio not exceeding a certain value) associated with at
least one of the burst capacity models considered.
Sufficient information regarding the actual properties (as
opposed to the nominal properties) such as the pipe
geometry and strength of the test specimen associated with
a given data point is reported such that it is applicable to at
least one of the burst capacity models considered.
3 Copyright 2012 by ASME
The literature search resulted in a total of 150 data points,
detailed information of which is given in Huang [8]. The
main sources of the test data include the two reports published
by PRCI [9-10], the burst tests conducted by various
researchers at University of Waterloo [11-12] and the
investigation reported by Souza et al. [13]. The ranges of the
general parameters characterizing the pipe geometry and
strength as well as the defect geometry for the collected test
data are summarized in Table 1. Note that the defect geometry
is characterized by the normalized defect depth (d
max
/t) and
length (L/(Dt)
0.5
) in Table 1, as commonly used in the burst
capacity models for corroded pipelines. Only a subset of the
collected test data was used to evaluate the model error for each
of the five burst capacity models described above because of
the limited applicability associated with both the models (e.g.
the d
max
/t limit) and test data (e.g. data without defect profiles
not applicable to RSTRENG).

Table 1 Ranges of Parameters of Burst Test Database


Basic Statistics of the Model Errors
The model error associated with a given burst pressure
prediction model was evaluated based on the test-to-predicted
ratios for the set of test data applicable to the model. The
uncertainties involved in the tests such as the measurement
error also contribute to the variability of the test-to-predicted
ratio in addition to the model error. However, such
uncertainties are difficult to quantify due to a lack of
information and therefore are ignored in the model error
evaluation.
The mean values and COVs of the model errors were
calculated for the five burst models considered and are
summarized in Table 2. The results in the table suggest that
B31G is highly conservative and associated with a poor
prediction accuracy (the mean and COV of the model error
equal to 1.473 and 41.2% respectively). B31G Modified is
markedly more accurate than the B31G model in that the model
error for the B31G Modified criterion has a less conservative
mean value (1.297) and a much smaller COV (25.8%). The
model error for the DNV model has a relatively high mean
value (1.233) and large COV (37.4%). PCORRC appears to
be on average the least conservative among the four burst
models in Table 2 that do not require a detailed defect profile.
Finally, it is safe to suggest that RSTRENG is the most accurate
of the five models considered because the model error for
RSTRENG was evaluated using a relatively large number (98)
of test data, and has a mean value (1.067) closest to unity and
the lowest COV value (16.5%).

Table 2 Statistics Parameters for the Model Errors



The test-to-predicted ratios were plotted on various
probability papers and the best-fit distributions were selected
by observations. The computer program, C-FIT

, provided by
C-FER Technologies was employed to carry out the probability
fitting analysis. As summarized in Table 2, the normal,
lognormal, Gumbel and Frechet distributions [14] were found
to fit the data well. Figure 1 shows the test-to-predicted ratios
plotted on the probability papers corresponding to the best-fit
distributions. Given that the model error contributes to the
overall uncertainty in the calculated burst pressure (i.e.
resistance), emphasis was placed on ensuring that the selected
distribution fits well the lower tail of the test-to-predicted
ratios. For example, the Frechet distribution was found to fit
the test-to-predicted ratios for the DNV model quite well (see
Fig. 1(c)). However, the lognormal distribution was selected
as the best-fit distribution for the DNV model because the
distribution fits the lower tail of the test-to-predicted ratios
better than the Frechet distribution.

(a) B31G

Parameter Lower bound Upper bound
D (mm) 273.1 914.4
t (mm) 4.57 11.28
SMYS (MPa) 172 448

y
(MPa) 196 515

u
(MPa) 277 658
d
max
/t 23.7% 83.8%
L/(Dt)
0. 5
0.372 22.833

Model
Applicable
Data Points
Mean COV
Fitted
Distribution
B31G 149 1.473 41.2% Frechet
B31G Modified 149 1.297 25.8% Gumbel
DNV 106 1.233 37.4% Lognormal
PCORRC 68 1.079 26.4% Gumbel
RSTRENG 98 1.067 16.5% Normal

4 Copyright 2012 by ASME

(b) B31G Modified


(c) DNV


(d) PCORRC


(e) RSTRENG

Figure 1 Test-to-predicted ratios plotted on the probability
papers

IMPACT OF MODEL ERROR ON RELIABILITY
ANALYSIS
We used a numerical example to investigate the impact of
the model error on the reliability analysis of corroded pipelines.
The example pipeline has an outside diameter of 610 mm
(NPS24), a wall thickness of 7.93 mm and an operating
pressure of 6.72 MPa, and is made from API 5L Grade X52
steel with a specified minimum yield strength (SMYS) of 359
MPa. Consider that one metal-loss corrosion defect has been
detected on the pipeline with a maximum depth of 1.19 mm
(i.e. d
max
/t = 0.15) and a length of 100 mm (i.e. L
n
= L/(Dt)
0.5
=
1.44). For simplicity, we assumed that the defect maximum
depth follows a linear growth path with a constant but uncertain
growth rate and that the defect length does not grow over time.
The first-order reliability method (FORM) [14-15] was
employed to evaluate the failure probability (i.e. probability of
burst), P
f
, of the pipeline at the corrosion defect as a function of
time elapsed since the defect detection. B31G Modified was
selected as the burst pressure prediction model in the FORM
analysis. The corresponding limit state function, g, is given
by

o b
P P g =
(4)

where P
b
is the burst pressure at the corrosion defect evaluated
using B31G Modified; is the model error, and P
o
is the
operating pressure.
For simplicity and emphasis on the impact of the model
error, the diameter, wall thickness, operating pressure,
maximum depth and length of the defect at the time of
detection were assumed to be deterministic quantities. The
pipe yield strength was assumed to follow a normal distribution
with a mean equal to 1.1SMYS and a COV of 3.5% [16]. The
growth rate of the maximum defect depth was assumed to be a
gamma distributed random variable with a mean of 0.2
5 Copyright 2012 by ASME
mm/year and a COV of 50%. The analysis was carried out for
three scenarios in terms of the probabilistic characteristics of
the model error:
ignoring the model error, i.e. = 1 (Scenario 1);
assuming the model error to follow a Gumbel distribution
with a mean of 1.297 and a COV of 25.8% based on the
results obtained in the present study as summarized in
Table 2 (Scenario 2), and
assuming the model error to follow a Gumbel distribution
with a mean of 1.184 and a COV of 24.1% based on the
results obtained in Chauhan et al.s study [1], in which the
real and machined defects were combined together
(Scenario 3).

Figure 2 Failure probability vs. time elapsed since the
defect detection

Figure 2 depicts the cumulative probabilities of burst
corresponding to three scenarios for the model error over a ten-
year forecasting period. Note that this figure only shows the
failure probabilities that are above 10
-8
. The results indicate
that the model error for B31G Modified has a substantial
impact on the failure probability. For example, up to Year 4,
the failure probabilities corresponding to Scenarios 2 and 3 are
over four orders of magnitude higher than those corresponding
to Scenario 1. The failure probabilities corresponding to
Scenarios 1 approach to and eventually exceed those
corresponding to Scenarios 2 and 3 as time increases. The
above observations highlight the importance of accounting for
the model error when evaluating the burst failure probability of
corroding pipelines. It is noted that at a given time there is
slight difference between the failure probabilities corresponding
to Scenario 2 (i.e. model error estimated using the real defects
only) and Scenario 3 (i.e. model error estimated using the real
and machined defects). For example, the probabilities of the
latter at a given time are in general one and half times as high
as those corresponding to the former.
For this particular example, Fig. 2 further suggests that the
growth rate of the defect depth has a great impact on the failure
probabilities if the model error is ignored (i.e. Scenario 1),
because the corresponding failure probability increases very
rapidly over time; on the other hand, the depth growth has a
relatively small impact on the failure probability if the model
error is accounted for in the reliability analysis. Finally, it is
worth pointing out that the impact of the model error on the
calculated failure probability will depend on the burst capacity
model employed in the reliability analysis, given that the
variability of the model errors for different burst capacity
models varies significantly (see Table 2).
CONCLUSION
This paper presents the evaluation of the model errors
associated with five representative burst pressure prediction
models, i.e., B31G, B31G Modified, DNV, PCORRC and
RSTRENG, for corroded pipelines. A relatively large number
of full-scale burst tests for corroded pipes was collected from
the literature and used to establish a database for the model
error evaluation. All the test specimens in the database
contain single isolated real corrosion defects. The
probabilistic characteristics (i.e. mean, COV and probability
distribution) of the model errors were derived based on the
ratios of the test and predicted burst pressures corresponding to
the data points included in the database.
RSTRENG is considered the best among the five models
considered because the corresponding model error has a small
bias (mean value equal to 1.067) and the lowest COV value
(16.5%). B31G is highly conservative whereas B31G
Modified is in general markedly more accurate than the B31G
model. PCORRC is the least conservative among the four
burst models that do not require a detailed defect profile.
A numerical example was used to illustrate the impact of
the model error on the probability of burst of the corroding
pipeline. B31G Modified was selected as the burst pressure
prediction model. The analysis results indicate that the model
error has a substantial impact on the calculated failure
probability; for example, the failure probability corresponding
to the case where the model error is accounted for can be over
four orders of magnitude higher than that obtained from the
case where the model error is ignored.
ACKNOWLEDGMENTS
Financial support provided by the Natural Sciences and
Engineering Research Council and Canada (NSERC) through
the Discovery Grant program and by the Faculty of Engineering
at Western University is gratefully acknowledged. The
constructive comments provided by the two anonymous
reviewers are greatly appreciated.

1.0E-08
1.0E-07
1.0E-06
1.0E-05
1.0E-04
1.0E-03
1.0E-02
1.0E-01
1 2 3 4 5 6 7 8 9 10
F
a
i
l
u
r
e

p
r
o
b
a
b
i
l
i
t
y

(

P
f
)
Time elapsed since the defect detection (year)
Scenario 1
Scenario 2
Scenario 3
6 Copyright 2012 by ASME
REFERENCES
[1] Chauhan, V., Brister, J. and Dafea, M., 2008, A Review of
Methods for Assessing the Remaining Strength of
Corroded Pipelines, Report Prepared for US Department
of Transportation, Pipeline and Hazardous Materials Safety
Administration, Advantica, Report No. 6781.
[2] Kariyawasam, S. and Peterson, W., 2008, Revised
Corrosion Management with Reliability Based Excavation
Criteria, Proceedings of the International Pipeline
Conference IPC 2008, IPC2008-64536, ASME, Calgary.
[3] ASME., 1991, Manual for Determining the Remaining
Strength of Corroded Pipelines, ASME B31G-1991, New
York.
[4] Kiefner, J. F. and Vieth, P. H., 1989, A Modified Criterion
for Evaluating the Remaining Strength of Corroded Pipe,
Report Prepared for American Gas Association, PR 3-805.
[5] DNV., 1999, Corroded Pipelines, DNV Recommended
Practice, DNV RP-F101, Det Norske Veritas, Hovik,
Norway.
[6] Leis, B. N. and Stephens, D. R., 1997, An Alternative
Approach to Assess the Integrity of Corroded Line Pipe -
Part I: Current Status; Part II: Alternative Criterion,
Proceedings of the 7
th
International Offshore and Polar
Engineering Conference, ISOPE, 4, pp. 624-641.
[7] Zhu, X. K. and Leis, B. N., 2006, Average Shear Stress
Yield Criterion and Its Application to Plastic Collapse
Analysis of Pipelines, International Journal of Pressure
Vessels and Piping, 83(9), pp. 663-671.
[8] Huang, G. X., 2011, Model Error Assessments of Burst
Capacity Models for Energy Pipelines, MESc. Thesis,
Department of Civil and Environmental Engineering, The
University of Western Ontario, London, Canada.
[9] Vieth, P. H. and Kiefner, J. F., 1993, Database of Corroded
Pipe Tests, Report Prepared for the Pipeline Corrosion
Supervisory Committee, Pipeline Research Committee of
Pipeline Research Council International, Inc., PRCI
Catalog No. L51689.
[10] Kiefner, J. F., Vieth, P. H. and Roytman, I., 1996,
Continued Validation of RSTRENG, Report Prepared for
the Line Pipe Research Supervisory Committee, Pipeline
Research Committee of Pipeline Research Council
International, Inc., PRCI Catalog No. L51749e.
[11] Chouchaoui, B. A., 1993, Evaluating the Remaining
Strength of Corroded Pipelines, PhD Thesis, Department
of Mechanical Engineering, University of Waterloo,
Waterloo, Canada.
[12] Cronin, D. S., 2000, Assessment of Corrosion Damage in
Pipelines, PhD Thesis, Department of Mechanical
Engineering, University of Waterloo, Waterloo, Canada.
[13] Souza, R. D., Benjamin, A. C., Freire, J. L. F., Vieira, R. D.
and Diniz, J. L. C., 2004, Burst Tests on Pipeline
Containing Long Real Corrosion Defects, Proc. of the Int.
Pipeline Conference, IPC2004, ASME, Calgary, 2, pp.
1159-1167.
[14] Melchers, R. E., 1999, Structural Reliability Analysis and
Prediction (2
nd
Ed), John Wiley & Sons, Chichester.
[15] Low, B. K. and Tang, W. H., 2007, Efficient Spreadsheet
Algorithm for First-order Reliability Method, Journal of
Engineering Mechanics, ASCE, 133(12), pp. 1378-1387.
[17] CSA., 2007, Oil and Gas Pipeline Systems, CSA Standard
Z662-07, Canadian Standard Association, Mississauga,
Ontario, Canada.

Vous aimerez peut-être aussi