Making Use of External Corrosion Defect Assessment (ECDA) Data To Predict DCVG %IR Drop and Coating Defect Area
Making Use of External Corrosion Defect Assessment (ECDA) Data To Predict DCVG %IR Drop and Coating Defect Area
Making Use of External Corrosion Defect Assessment (ECDA) Data To Predict DCVG %IR Drop and Coating Defect Area
DOI: 10.1002/maco.201810085
| Accepted: 14 March 2018
ARTICLE
Nik N. Bin Muhd Noor1 | Keming Yu1 | Ujjwal Bharadwaj2 | Tat-Hean Gan2
KEYWORDS
Bayesian quantile regression, buried pipelines, coating defect, direct current voltage gradient, external
corrosion direct assessment
the coating (defect) which leaves the metal substrate exposed symmetric and have some degree of skewness. Distributions
to the environment. In the event of coating discontinuity, such as the ones above, are asymmetrical and hence need
corrosion is likely to occur which could undermine the whole more complex solution in describing the entirety of the
structural integrity of the pipeline. response variable's distribution.[15]
Buried pipelines are normally protected with a cathodic The Bayesian approach toward quantile regression was
protection (CP) system. This system acts as backup to the elaborated by Ref. [16]. Bayesian inferences is more
coating system and comes into play when defects are present on advantageous than the classical approach in mainly two
the pipeline's coating.[8] The monitoring of the CP system, the instances: 1) Bayesian statistics does not rely on asymptotic
coating and the overall integrity of the pipeline is normally variances of the estimators and 2) the estimated parameter
addressed by conducting an external corrosion direct assessment includes the parameter uncertainty in the form of a posterior
(ECDA).[9] As part of the ECDA process, an indirect assessment distribution. Since the mechanism of cathodic protection are
which is commonly used is the direct current voltage gradient complex, uncertainty of parameter values becomes an inherent
(DCVG). This technique is used to identify the location of trait. The Bayesian approach helps us to quantify this
coating defects and to classify its severity. Based on a defect uncertainty. The findings from this paper can then be used
severity, a decision can be made on whether to proceed with as prior information for the next iteration of the ECDA process.
further direct assessment which requires excavation of the defect This paper is divided into several sections. Section 2
location. The DCVG technique is considerably accurate in describes the data that was obtained from a recent ECDA
locating a defect location but lacks the accuracy in predicting its project conducted by TWI Ltd. Section 3 outlines the
size (area).[10] The prediction of coating defect area has not yet methodology used followed by Section 4 which presents the
been a popular research theme within the academic sphere and results from the models. Section 5 is discussion and Section 6
the pipeline industry. The authors found only a handful of is the conclusion and future work.
literature relating to this topic. The most noticeable of which was
done by Ref. [11]. In this paper, a quantile regression was used to
2 | MEOC DATA
model the relationship between the coating defect area and its
possible contributors. The paper also sheds light on the
The Middle Eastern Oil Company (MEOC) has appointed TWI
challenges faced by pipeline operators when interpreting
Ltd. to conduct an ECDA on its network of pipelines. There are
DCVG indications. McKinney,[12] has produced a model which
a total of nine (9) pipelines, all of which are non-piggable. The
estimates the coating defect area based on simulated data. The
ECDA conducted by TWI Ltd. complied with the ANSI/
approach taken is deterministic where a finite element method
NACE SP0502-2010: Standard Practice Pipeline External
(FEA) was used. Moghissi et al.[13] have identified that there is
Corrosion Direct Assessment Methodology. The ECDA is
no simple solution toward prioritizing coating defects for further
divided into four parts. Data from these parts were gathered and
assessment. Data were collected from the closed interval
annotated by the authors to be used in the modeling process.
pipeline survey, DCVG and current attenuation assessments and
were used to derive basic formulations to model the relationship
between coating defect area and its possible contributing factors. 2.1 | Pre-assessment
The approach taken in Ref. [13] uses similar methods as those
The data in this section includes the design data of the pipe which
found in the work by Ref. [12].
included its design philosophy, material selection, and the pipe
The motivation for the work reported in this paper is to
supplement the body of knowledge highlighted above.
Statistical models are proposed to better explain the inner
workings of a DCVG indication for the prioritization of
coating defects for subsequent direct examination of the
affected pipeline.
Quantile regression is used to fully characterize the
dependent variable without relying on assumptions of the
response distribution e.g., normally distributed. As compared
to the mean regression, quantile regression is much more
robust to outliers since it employs absolute values of the error
terms.[14] Judging by the distribution of the response variable
from the MEOC data (which will be described in detail in the
following section) which is represented by Figures 1 and 2,
the total coating defect area (TCDA) and the %IR (IR drop) FIGURE 1 Probability density plot for TCDA. Reproduced with
variable demonstrate a distribution which is not normal nor permission from TWI Ltd.
BIN MUHD NOOR ET AL.
| 3
TABLE 1 (Continued)
Symbol Variables considered Type of variable/summary statistics
1st quantile 36 inches
Median 36 inches
Mean 35.3 inches
3rd quantile 36 inches
Max. value 42 inches
TCDA Total coating defect area Quantitative
Min. value 0 cm2
1st quantile 1200 cm2
Median 9985 cm2
Mean 44 893 cm2
3rd quantile 77 865 cm2
Max. value 269 894 cm2
Backfill type
θ Rock Qualitative
κ Sand + clay Qualitative
λ Stones + clay Qualitative
Coating type
μ Coal tar Qualitative
ξ Polyethylene Qualitative
CW Cold wrap Qualitative
Backfill geometry
ρ Angular Qualitative
σ Round + angular Qualitative
R Rounded Qualitative
pH of water in soil
φ Acidic Qualitative
χ Alkaline Qualitative
ψ Neutral Qualitative
pH of water underneath coating
ω Acidic Qualitative
ϋ Alkaline Qualitative
ï Neutral Qualitative
and is associated to the aforementioned quantile regression maximizing the likelihood function which is formed by
minimization problem (minimization β) joining independently distributed asymmetric Laplace
densities (ALD).
min∑ρp yt x0t β ; ð1Þ The probability density function of the ALD is given as
t follows[18]
Variables Posterior mean 0.025 0.975 Posterior mean 0.025 0.975 Posterior mean 0.025 0.975
(Intercept) 14.2 −2.95 38.4 86.1 77 95.5 23.6 −28 73.5
IR drop (%IR) 0.0000022 −0.000012 0.0000384 0.0000687 0.0000517 0.0000837 0.0000532 0.00000513 0.0000788
Soil resistivity (SR) −0.0000235 −0.00045 0.000325 −0.000567 −0.000881 −0.000293 −0.000346 −0.00064 0.0000823
Percentage of pit depth to wall thickness (POPD) 0.00611 −0.102 0.118 0.0439 −0.0321 0.12 0.108 −0.0198 0.264
Deposits under coatings (DUC) 0.0079 −0.0358 0.0511 −0.0372 −0.0764 0.00186 −0.0704 −0.139 −0.00845
Depth of cover (DOC) 0.0549 0.0111 0.108 0.0933 0.0675 0.122 0.0364 −0.0251 0.109
Time in service (TIS) −0.336 −0.779 0.207 −0.374 −0.561 −0.189 1.19 0.137 2.65
Pipe size (PS) −0.0818 −0.329 0.104 −1.31 −1.53 −1.11 0.285 −0.159 0.739
Backfill type (Rock) 5.2 −9.36 47.8 50.8 42.7 57.6 10.7 −2.52 47.9
Backfill type (sand + clay) −1.03 −12.7 3.38 16.3 5.88 30.6 −11.6 −27.8 1.94
Backfill type (stones + clay) 1.72 −1.43 6.36 0.562 −1.46 3.69 0.411 −3.95 6.33
Coating type (coal tar) −3.26 −8.38 8.85 −0.215 −2.8 1.82 11 −5.26 35
Coating type (polyethylene) −6.28 −20.7 5.22 0.368 −2.97 4.66 2.44 −25.9 21.6
Backfill geometry (angular) 0.754 −3.01 5.65 −19.9 −23.2 −16.4 −8.69 −36.8 2.98
Backfill geometry (round + angular) −2.64 −7.12 0.369 −0.835 −4.32 1.22 −0.446 −6.5 5.37
pH of water in soil (acidic) 1.17 −10.3 16.5 −8.1 −14.8 0.286 −4.8 −44 9.5
pH of water in soil (alkaline) 8.41 −0.222 15.3 0.753 −1.25 4.34 −0.804 −7.01 4.32
pH of water in soil (neutral) 7.24 −0.67 15.8 7.03 1.03 11.1 −14 −20.3 −0.277
pH of water underneath coating (acidic) −0.943 −10.3 3.26 −3.24 −20.6 1.32 −11 −42.4 4.92
pH of water underneath coating (alkaline) −2 −5.33 0.492 −7.78 −10.1 −5.28 0.991 −1.36 5.77
pH of water underneath coating (neutral) 2.56 −4.74 13 −0.125 −3.71 3.08 −1.99 −11.9 4.81
FIGURE 3 Example of a trace plot and posterior histogram of the 0.5 quantile for the estimated coefficient, TCDA for Model 1. Reproduced
with permission from TWI Ltd.
4.2 | Refined %IR (Model 1a) The trend of the estimated coefficients for the variable soil
resistivity is also similar to Model 1. From 0.25 quantile
The results of the estimated coefficients by BQR for Model 1a
upwards, the trend is negative with its most negative at the 0.5
is presented in Table 4. In achieving convergence for all the
quantile. The reason for this can be considered consistent with
variables, iterations of up to 300 000 were determined with
the assessment for Model 1 when one looks at the rock variable
the initial 5000 steps regarded as burn-in. For the variable of
with most of the estimates showing high positive values. The
interest, the TCDA, the maximum estimated value occurs at
peak is also found at the 0.5 quantile suggesting that the effect
the 0.5 quantile. This prediction is similar to the one predicted
of having coarse grained soil affects %IR values at its median
by Model 1. As for the overall estimated trend, it follows the
quantile. There is also the factor of heterogeneity of the soil
same pattern as Model 1 with Model 1a being more
itself which also contributed to the non-linearity effect toward
pronounced. The value of the coefficient at the maximum
certain quantiles of the %IR distribution. Equations of various
is 0.0000828. This means that a 1 cm2 of TCDA will have an
quantiles are presented in equations below.
effect on the %IR by 0.0000828%. At the 0.05 quantile, the
coefficient value is at its lowest with a value of −0.0000353.
%IR0:05 ¼ 4:74 0:0000353TCDA þ 0:000000565β
The negative value signifies that with a 1 cm2 increase in þ0:0508ε 0:158η þ 5:23θ 0:939κ
TCDA will yield a 0.0000353% decrease in %IR. þ1:56λ þ 2μ 0:113ξ þ 0:434ρ
3:65σ þ 1:2φ þ 8:11χ þ 5:06ψ ð9Þ
0.0001
%IR0:95 ¼ 64:9 þ 0:000073TCDA 0:000296β
0.00008 þ0:0228ε þ 0:432η þ 6:45θ 8:21κ
1:46λ 6:78μ 15:3ξ 6:73ρ
0.00006 1:24σ 4:67φ 0:575χ 12:6ψ ð11Þ
0.00004
%IR
0.00002
0
0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
4.3 | Contributing factors to %IR (Model 2)
-0.00002
-0.00004
The estimated coefficients for Model 2 are given in Table 5.
QUANTILE
However, the reference variable is substituted to be backfill
FIGURE 4 Example of a quantile plot of the TCDA variable for type – rock, coating type – polyethylene and backfill
Model 1. Reproduced with permission from TWI Ltd. [Color figure geometry – angular. This is due to investigate on the factors
can be viewed at wileyonlinelibrary.com] regarding soft soils which included clay with rounded grain
BIN MUHD NOOR
TABLE 4 Bayesian quantile regression (BQR) estimates with 95% credible intervals for quantiles 0.05, 0.5, and 0.95 for Model 1a
ET AL.
Quantiles
Variables Posterior mean 0.025 0.975 Posterior mean 0.025 0.975 Posterior mean 0.025 0.975
(Intercept) 4.74 −1.53 14.4 87.5 78.2 97 64.9 56 84.8
Total coating defect area (TCDA) −0.0000353 −0.00004 0.0000375 0.0000828 0.0000661 0.0000989 0.000073 0.0000008 0.0000833
Soil resistivity (SR) 0.000000565 −0.000344 0.000364 −0.000668 −0.000863 −0.000416 −0.000296 −0.000641 0.0000428
Depth of cover (DOC) 0.0508 0.0148 0.0888 0.0722 0.0447 0.099 0.0228 −0.0243 0.0562
Pipe size (PS) −0.158 −0.388 0.0759 −1.77 −2.01 −1.53 0.432 −0.199 0.64
Backfill type (rock) 5.23 −9.26 47.9 53.4 45.1 60.9 6.45 −4.09 36.7
Backfill type (sand + clay) −0.939 −7.93 3.08 25.4 16.1 36.8 −8.21 −22.3 2.47
Backfill type (stones + clay) 1.56 −0.754 5.66 0.619 −0.927 3.38 −1.46 −8.41 2.96
Coating type (coal tar) 2 −0.739 6.36 5.54 3 8.17 −6.78 −11.9 0.445
Coating type (polyethylene) −0.113 −8.54 6.18 6.77 −0.593 14.7 −15.3 −34.4 2.78
Backfill geometry (angular) 0.434 −3.6 4.73 −18.2 −22.7 −14.4 −6.73 −30.5 2.56
Backfill geometry (round + angular) −3.65 −7.75 −0.335 0.251 −1.89 2.89 −1.24 −10.2 5.22
pH of water in soil (acidic) 1.2 −10.2 16.7 −6.07 −12.7 0.598 −4.67 −42.2 9.47
pH of water in soil (alkaline) 8.11 −0.303 13.7 1.76 −1.08 5.82 −0.575 −6.11 3.91
pH of water in soil (neutral) 5.06 −1.05 12.4 1.14 −1.34 4.89 −12.6 −19.3 0.294
structure. A total of 400 000 iterations were made to get to iterations were made to achieve convergence with the initial
the point of convergence with the initial 5000 readings as 5000 recordings regarded as burn-ins. At the 0.05 quantile,
burn-in. Table 5 shows the TCDA variable coefficients has the predicted TCDA coefficient showed similar results to the
an upward trend with a slight dip at the 0.25 quantile. The one obtained for Model 2. The coefficient value drops at the
highest value is reached at the 0.95 quantile with a value of 0.25 quantile and rising steadily after this all the way up to the
0.000229%. For a 1 cm2 increase in the size of coating 0.95 quantile where it reaches its maximum. Maximum
defect area, a 0.000229% increase in percentage IR is predicted value stands at 0.000221 which means a 1 cm2
expected. This is higher than the maximum obtained by increase in TCDA will give an increase of 0.000221% in %IR.
Model 1. Additionally, this happens at the 0.95 quantile Previously for Model 2, similar characteristics were observed
which goes well with established understanding of the with only slight differences in the predicted values.
technique as compared to Model 1 where the maximum Soil resistivity plays a role in Model 2a where an
occurred at the 0.5 quantile. This is mainly due to the increasing trend is observed starting from the 0.25 quantile all
contribution of the careful judgement of the authors which the way up to the 0.95 quantile. The highest predicted value is
obliterated four points from the original set. at the 0.95 quantile with a Bayes estimate of 0.000482. At the
Estimated coefficients for the soil resistivity variable 0.95 quantile, a 1 unit increase in the value of soil resistivity
showed increasing trends starting from the 0.25 quantile up to will mean a 0.000482% increase in %IR.
the maximum which is at the 0.95 quantile. The maximum The presence of clay as the backfill material will affect the
Bayes estimate is 0.000373. Therefore a 1 unit increase in soil %IR differently across the quantile of the %IR distribution
resistivity, an increase of 0.000373% of %IR is expected. when compared to the soil resistivity variable. Clay affects the
Moreover, large uncertainties were observed in the upper and 0.75 quantile the most with the 0.05 the least affected. The
lower ends of the quantiles as compared to the median region. value of the maximum estimate coefficient is 57. This is not
The variable clay showed increasing trends across the % far off than the estimated value at the same quantile for Model
IR distribution with a dip at the 0.95 quantile. The maximum 2. The upward trend up to the 0.75 quantile reflects that clay
estimated coefficient was noticed to be at the 0.75 quantile has a positive effect in the contribution of the %IR reading.
with a value of 60.8. This can be translated as the effect of the Models of various quantiles are presented in the following
presence of clay to the %IR will be the most at the 0.75 equations.
quantile of the %IR distribution.
The following are selected models (Model 2) for the %IR0:05 ¼ 30:3 þ 0:0000956TCDA 0:000132β
contribution of %IR based on various quantiles. þ0:0561ε 0:301η 18:9C
19:1κ 6:25λ 6:48CW
þ8:51ξ þ 1:22R 0:0602σ
%IR0:05 ¼ 14:7 þ 0:0000741TCDA þ 0:0000293β þ4:13φ 1:26χ 11:3ψ ð15Þ
þ0:0334γ 0:0209δ þ 0:0668ε 0:116ζ
0:126η 2:26C 1:69κ þ 11:4λ
9:26CW þ 7:57ξ þ 1:73R 0:246σ %IR0:5 ¼ 86:2 þ 0:0000768TCDA 0:000178β
þ3:98φ 1:84χ 11:4ψ þ 2:8ω 0:0665ε 0:452η þ 0:785C
1:62ϋ 4:96ı̈ ð12Þ 20:3κ 3:21λ 32:1CW
0:72ξ 0:279R 0:243σ
þ6:63φ 18:9χ 11:6ψ ð16Þ
%IR0:05 ¼ 79:4 þ 0:0000618TCDA þ 0:000206β
þ0:161γ 0:0373δ þ 0:00696ε 0:234ζ
0:3η 3:35C 16:7κ þ 3:29λ %IR0:95 ¼ 31:3 þ 0:000221TCDA þ 0:000482β
31:4CW þ 1:45ξ þ 1:02R þ0:0639ε 0:00579η þ 18:6C þ 10:9κ
0:156σ þ 7:4φ 21:2χ þ25:9λ 8:59CW þ 29:7ξ
11:2ψ þ 1:02ω 8:36ϋ 6:67ı̈ ð13Þ 0:829R þ 1:22σ þ 11:9φ þ 4χ 16:8ψ
ð17Þ
%IR0:95 ¼ 22 þ 0:000229TCDA þ 0:000373β
þ0:0558γ 0:05δ þ 0:0982ε þ 0:179ζ
þ0:186η þ 16:3C þ 4:9κ þ 25:4λ
5:69CW þ 29:7ξ 1:58R þ 1:58σ 4.5 | Total coating defect area (TCDA) models
þ9:2φ 2:62χ 16:8ψ 4:98ω
3:43ϋ 13:1ı̈ ð14Þ With the establishment of the %IR models utilizing both
the Oriset and the Filtset data (Models 1, 1a, 2, and 2a), the
construction of the TCDA model will further increase the
capability of operators and decision makers in prioritizing
4.4 | Refined %IR (Model 2a)
coating defects based on their severity. To add to this
Table 6 shows the estimated coefficients predicted by the enhancement, we propose TCDA models (Models 3 and 4)
BQR method with the Filtset data for Model 2a. 400 000 which predict the coating defect area based on variables from
TABLE 5 Bayesian quantile regression (BQR) estimates with 95% credible intervals for quantiles 0.05, 0.5 and 0.95 for Model 2
Quantiles
BIN MUHD NOOR
Variables Posterior mean 0.025 0.975 Posterior mean 0.025 0.975 Posterior mean 0.025 0.975
(Intercept) 14.7 −1.76 40.681215 79.4 43.7 92.1 22 −415 69.1
Total coating defect area (TCDA) 0.0000741 0.0000676 0.000128 0.0000618 0.0000493 0.0000769 0.000229 0.000168 0.000232
Soil resistivity (SR) 0.0000293 −0.000316 0.000383 0.000206 0.0000175 0.000329 0.000373 −0.0000612 0.000848
Percentage of pit depth to wall thickness (POPD) 0.0334 −0.085 0.144788 0.161 0.088 0.232 0.0558 −0.0308 0.238
Deposits under coatings (DUC) −0.0209 −0.0789 0.029226 −0.0373 −0.0655 −0.0111 −0.05 −0.115 0.00565
Depth of cover (DOC) 0.0668 0.0269 0.108152 −0.00696 −0.0324 0.0161 0.0682 0.012 0.119
Time in service (TIS) −0.116 −0.328 0.085368 −0.234 −0.338 −0.128 0.179 −0.22 0.788
Pipe size (PS) −0.126 −0.31 0.079494 −0.3 −0.472 −0.153 0.186 −0.575 1.93
Backfill type (clay) −2.23 −13.6 3.720633 3.35 −2.41 23.9 16.3 −7.2 260
Backfill type (sand + clay) −1.63 −12.8 5.18233 −16.7 −23.9 3.6 4.9 −26.6 246
Backfill type (stones + clay) 11.4 −0.0934 18.838816 −3.29 −9.69 18.8 25.4 0.026 262
Coating type (PVC cold wrap) −9.26 −29.1 2.434434 −31.4 −38.2 −22 −5.69 −17.9 0.765
Coating type (Coal tar) 7.57 −11 20.359542 −1.45 −8.49 7.7 29.7 16.4 38.7
Backfill geometry (round) 1.73 −0.761 6.591509 1.02 −0.55 3.55 −1.58 −7.62 16.5
Backfill geometry (round + angular) −0.246 −4.13 3.671415 −0.156 −2.4 2.01 1.58 −2.14 10.4
pH of water in soil (acidic) 3.98 −10.1 38.354927 7.4 −0.427 14.5 9.2 −0.768 30.3
pH of water in soil (alkaline) −1.84 −7.75 0.907865 −21.2 −24.3 −17.7 2.62 −1.7 12.2
pH of water in soil (neutral) −11.4 −19.6 -0.207547 −11.2 −14.5 −8.07 −16.8 −26.6 −0.241
pH of water underneath coating (acidic) 2.8 −6.1 18.678275 1.02 −2.44 9.92 −4.98 −27.4 5.78
pH of water underneath coating (alkaline) −1.62 −4.69 0.344062 −8.36 −10.1 −6.7 −3.43 −7.89 0.356
pH of water underneath coating (neutral) −4.96 −16.5 0.619875 −6.67 −14.2 0.543 −13.1 −33.1 2.77
TABLE 6 Bayesian quantile regression (BQR) estimates with 95% credible intervals for quantiles 0.05, 0.5, and 0.95 for Model 2a
Quantiles
Variables Posterior mean 0.025 0.975 Posterior mean 0.025 0.975 Posterior mean 0.025 0.975
(Intercept) 30.3 −2.29 501 86.2 77.7 94.1 31.3 −308 79.5
Total coating defect area (TCDA) 0.0000956 0.0000936 0.000138 0.0000768 0.0000547 0.0000989 0.000221 0.000186 0.000222
Soil resistivity (SR) −0.000132 −0.000408 0.000202 −0.000178 −0.000395 0.0000902 0.000482 0.0000361 0.00108
Depth of cover (DOC) 0.0561 0.0195 0.0981 −0.0665 −0.0907 −0.0337 0.0639 0.0178 0.0949
Pipe size (PS) −0.301 −0.333 0.00686 −0.452 −0.622 −0.309 −0.00579 −0.571 0.214
Backfill type (clay) −18.9 −393 4 0.785 −2.71 5.56 18.6 −6.5 323
Backfill type (sand + clay) −19.1 −392 3.51 −20.3 −26.2 −13.6 10.9 −24.2 300
Backfill type (stones + clay) −6.25 −381 17.7 −3.21 −7.47 0.767 25.9 −0.151 324
Coating type (PVC cold wrap) −6.48 −51 2.72 −32.1 −36.8 −28.5 −8.59 −29.3 0.1
Coating type (Coal tar) 8.51 −37.8 18.5 −0.72 −5.65 2.67 29.7 9.06 38.9
Backfill geometry (round) 1.22 −5.87 5.43 −0.279 −2.77 1.51 −0.829 −6.18 2.5
Backfill geometry (round + angular) −0.0602 −5.43 3.04 −0.243 −3.27 2.1 1.22 −1.56 7.78
pH of water in soil (acidic) 4.13 −9.69 39.6 6.63 −0.543 14 11.9 0.0936 31.7
pH of water in soil (alkaline) −1.26 −6.84 1.42 −18.9 −21.5 −15.8 4 −1.39 12.3
pH of water in soil (neutral) −11.3 −18.2 −0.912 −11.6 −15.9 −6.64 −16.8 −24.1 −0.0591
207 000
−7690
−70.9
limited to only quantitative values due to the large amounts of
0.975
0.482
−212
2790
7410
28.6
subjective interpretations on the qualitative variables.
Another reason for this is to avoid higher computational
−0.0154
128 000
cost as Bayesian inference with the Metropolis-Hastings
−8200
−162
−387
−122
2640
5670
0.25
algorithm (MCMC – MH) is known to take large amounts of
computational memory when dealing with large quantities of
Credible intervals
Posterior mean
−8030
−93.1
0.151
−257
−111
2740
6040
interpreting what the signal mean, one can make sound
0.95
−2319.907
−53.681
4.6 | TCDA Model 3
100.219
264.055
728.063
−0.476
20.721
0.975
−2362.104
228.826
−0.529
82.292
10.358
702.75
−2351.485
707.098
−0.524
19.543
−0.821
0.0559
0.0392
6.79
5.16
11.3
−0.9835
−0.0523
5.4327
4.6761
9.8316
models for the 0.05, 0.5, and 0.95 quantile for Model 3.
Credible intervals
Posterior mean
−0.00321
−0.901
−465
0.05
6.04
4.92
10.4
TCDA cm2 Þ0:5 ¼ 78687:177 þ 84:428α 0:524β
þ232:204γ þ 19:543δ 69:776ε
Percentage of pit depth to wall thickness (POPD)
−3763.1
78394.4
−155.8
1828.3
4720.8
−91.7
are narrow as compared to previous predicted models.
0.975
455.7
12.3
4.7 | TCDA Model 4
−1.72E + 02
−2.24E + 02
−1.14E + 02
−5.17E + 03
6.65E + 04
1.46E + 03
1.17E + 01
3.54E + 03
As was previously mentioned, the data considered for this
0.25
assessment is the Filtset. Results of the analyses is highlighted
in Table 8. As was expected, the %IR variable showed a
Credible intervals
Posterior mean
−2.19E + 02
−9.78E + 01
−5.07E + 03
Starting at the 0.05 and 0.25 quantile, the increase of the
7.77E + 04
1.48E + 03
1.22E + 01
4.06E + 02
4.61E + 03
estimated coefficients is subtle but for the 0.5 quantile the
0.95
changes are much more abrupt with the values tapering back at
the 0.75 and 0.95 quantile. The maximum value occurs at the
−1684.384 0.95 quantile with an estimated coefficient of 1481.9. In other
29374.475
−19.534
476.451
219.652
961.977
−0.452
34.787
−2.07E + 01
−1.69E + 03
148 190 cm2 in TCDA which is the maximum size the model
TABLE 8 Bayesian quantile regression (BQR) estimates with 95% credible intervals for quantiles 0.05, 0.5, and 0.95 for Model 4
2.91E + 04
4.72E + 02
2.19E + 02
3.40E + 01
9.60E + 02
size is 2.21 cm2. The maximum predicted defect sizes for all
the quantiles are shown in Figure 5.
Credible intervals
Posterior mean
−2.00E + 01
−1.68E + 03
2.19E + 02
3.42E + 01
9.63E + 02
−2.98E − 03
−8.31E − 01
1.07E − 01
6.79E + 00
2.84E − 02
5.19E + 00
1.12E + 01
0.95 quantile.
0.975
−9.99E − 01
−6.00E − 02
5.46E + 00
4.71E + 00
9.81E + 00
Posterior mean
−4.64E + 02
−3.36E − 03
−9.16E − 01
−8.71E − 03
2.21E − 02
6.06E + 00
4.95E + 00
1.03E + 01
Quantiles
160000
140000
Percentage of pit depth to wall thickness (POPD)
120000
100000
Reproduced with permission from TWI Ltd.
80000
TCDA
60000
Deposits under coatings (DUC)
40000
20000
Depth of cover (DOC)
0
Time in service (TIS)
-20000
QUANTILE
Pipe size (PS)
TCDA cm2 Þ0:5 ¼ 28978:772 þ 475:876α 0:453β models. These variables take on 10.5%, 35.2%, 109.5 cm,
þ219:149γ þ 34:219δ 20:033ε 32.5 years, 35.3 inch, and 2722.1 Ω-cm, respectively which
1681:945ζ þ 963:069η ð22Þ
represents the mean value of each variable.
Figure 6 shows the predictions made by Model 1 of the
TCDA cm2 Þ0:95 ¼ 77655:6 þ 1481:9α þ 12:2β %IR with increasing TCDA. Generally, the models highlight
þ406γ 218:8δ 97:8ε
5066:9ζ þ 4610:6η ð23Þ an upward trend which is in parallel with the current
understanding of the system. However, the slope of the
models indicates a small effect of TCDA toward %IR. This
can clearly be seen at the lower quantiles (0.05 and 0.25)
where the line is almost flat. Also, the median quantile has the
5 | DISCUSSION highest prediction value and the steepest slope which
corresponds to the estimated coefficient values in Figure 4.
5.1 | Contributing factors to %IR – (Models 1, A refined version of Model 1 is given by Model 1a which is
1a, 2, and 2a) presented in Figure 7. Similarly, the models take on the mean
values of each contributing variable.
5.1.1 | TCDA variable The prediction of the resulting %IR in Figure 7 shows an
The low coefficient values estimated for the TCDA variable improvement in terms of the effect of TCDA on %IR with
(Models 1, 1a, 2, and 2a) was unexpected since the concept of steeper slopes being observed. Similar to Model 1, the median
a DCVG technique relied primarily upon coating defects to of the %IR received the largest effect from the TCDA. The
generate voltage drops. The results show coating defects in estimated %IR values based on the median is also higher with
general have a mild effect on the %IR reading. Other known Model 1a as compared to Model 1. The removal of certain
and unknown factors might also be a contributor toward %IR. variables which do not contribute to the %IR has improved the
One of these factors could be SR and the nature of the backfill %IR estimation for the top three quantiles. For the 0.25
geometry. Other factors could include the presence of quantile, small effects of the TCDA toward %IR are seen
interference in the form of stray currents especially if the which is similar to the previous Model 1. However, the
pipeline is situated adjacent to other pipelines or is located estimated values here are higher. The 0.05 quantile show
near overhanging power cables. Although an interruption decreasing trend where increasing TCDA relates to a
technique was used to eliminate foreign currents contributing decreasing of %IR.
to %IR indication, large structures such as buried pipelines The inconsistency (higher TCDA does not reflect higher
need longer periods for it to depolarize and considered IR %IR values) for Models 1 and 1a with respect to the 0.05
free.[19] To picture this more clearly, the following figures quantile could possibly be attributed to the outliers present at
show the relationship between TCDA and %IR while keeping higher and lower quantiles of the TCDA distribution – large
other variables constant. As was previously mentioned, other defect areas are paired to low reading of the %IR and vice
factors which gave rise to the %IR readings such as the POPD, versa. Additionally, credible intervals at higher and lower
DUC, DOC, TIS, PS, and SR were used to generate the quantiles for Models 1 and 1a are much wider indicating
90.000 100.000
90.000
80.000
80.000
70.000
70.000
60.000 60.000
%IR
50.000 50.000
%IR
40.000 40.000
30.000 30.000
20.000
20.000
10.000
10.000
0.000
0.000 0 50000 100000 150000 200000 250000 300000
-10.000
0 50000 100000 150000 200000 250000 300000
-20.000
TCDA (CM^2) TCDA (CM^2)
FIGURE 6 TCDA versus %IR for Model 1. Each color represent FIGURE 7 TCDA versus %IR for Model 1a. Each color
a different quantile. Reproduced with permission from TWI Ltd. represent a different quantile. Reproduced with permission from TWI
[Color figure can be viewed at wileyonlinelibrary.com] Ltd. [Color figure can be viewed at wileyonlinelibrary.com]
16
| BIN MUHD NOOR ET AL.
higher uncertainty as compared to the median quantile where examination process. These defects were not present
the maximum estimated values have occurred. Inconsistent during the indirect assessment (DCVG measurements).
results can also be summarized in the following bullet points. Deposits of scales due to the cathodic protection current on
the metal substrate will mask the true size of a coating
Interference in the form of stray or telluric currents which defect. Measurements are perceived to be small based on
will interfere with the %IR signal. Currents from adjacent the %IR reading. This is an erroneous representation of the
ICCP system, electrified railway tracks (DC traction true size of the defect.
system), overhead power cables, etc. have the potential
of compromising the %IR signal. This can be seen at certain The assessment on Models 2 and 2a utilizes the Filtset
locations of the pipeline. data. The estimated %IR readings based on Model 2 (Figure 8)
Adjacent transmitting power cables could compromise the and 2a (Figure 9) are given as follows. Similar to the previous
DCVG signal in the form of AC currents. AC currents can Models 1 and 1a assessments, the mean of POPD, DUC,
also lead to accelerated corrosion to the pipelines running DOC, TIS, PS, and SR was used to generate these models.
below.[20] In the case of the MEOC pipelines, power cables Immediately, it can be seen that in Figure 8 the estimated
can be seen running closely along and perpendicular to the values of the %IR are much improved than Models 1 and 1a.
direction of the buried lines. The effect of TCDA on %IR is also greater which reflects the
The heterogeneous nature of soils compromises or alters underlying intention of a DCVG assessment. The highest
the measured voltage signal. The calculation of the %IR predicted value of the %IR is at the 0.95 quantile which
value requires input in the form of the pipeline-to- indicates TCDA has the highest effect on higher readings of
electrolyte interface resistance. The resistant value is the %IR. Additionally, narrower credible intervals were
related to the SR value which is measured at test posts. obtained highlighting in lesser uncertainty of the estimated
However, DCVG readings are conducted away from test coefficients. Therefore, the removal of four excavation points
posts where the magnitude of SR changes. The changes improves the overall estimation of the role of TCDA on %IR.
will contribute to the inconsistencies of the %IR measure- However, looking at quantiles 0.05 and 0.25 shows an
ments where the heterogeneity of soil is not considered in apparent effect of TCDA toward %IR. However, these
the %IR formula. Although SR measurements were taken estimates are below the zero line. For the 0.25 quantile, all the
for every excavated area, this was not included into the %IR predicted readings of %IR are negative and it sits lower than
calculation. the 0.05 quantile. Although the apparent outliers were
Defects occurring at the 6 o’clock position will tend to removed for this assessment, there are other factors that
attenuate the voltage signal and will not correspond to the might give an overall effect on the %IR predictions. Model 2a
true size of a defect.[11] tries to find this answer by further refining the model through
Based on the report provided by TWI Ltd., there is a the omission of variables which in theory does not contribute
possibility that some of the coating defects were caused by to the generation of %IR. Model 2a's prediction of %IR is
the excavator during excavation of bell holes for the direct given as follows in Figure 9.
100 100
90 90
80 80
70 70
60
60
50
50
40
%IR
%IR
40
30
30
20
20
10
0 10
FIGURE 8 TCDA versus %IR for Model 2. Each color represent FIGURE 9 TCDA versus %IR for Model 2a. Each color
a different quantile. Reproduced with permission from TWI Ltd. represent a different quantile. Reproduced with permission from TWI
[Color figure can be viewed at wileyonlinelibrary.com] Ltd. [Color figure can be viewed at wileyonlinelibrary.com]
BIN MUHD NOOR ET AL.
| 17
From Figure 9 the omission of certain variables has 1a, the two variables seem to complement each other and is
improved the overall prediction of the %IR based on TCDA. only understood when both of them are looked at together.
Significant effects of the TCDA toward the %IR is seen across The decrease in the estimated value at the 0.95 quantile for the
all the quantiles. The effect of higher TCDA on higher clay variable and the increase of the predicted value at the
readings of %IR is seen with the highest predicted values of % 0.95 quantile for the backfill geometry – round variable is the
IR occurring at the 0.95 quantile. This can also be said with cause of a possible mixture of fine to coarse grain soils in the
other quantiles where lower values of TCDA effects the backfill. Moreover, there are also the possibility of foreign
lowest part of the %IR readings. However, for the predicted currents interfering with the measured signal as was
values with regard to the 0.05 quantile, shows prediction mentioned above. Coupled this with the heterogeneous
values of %IR of less than zero. This small inconvenience can nature of soils, unexpected outcomes like this are not unusual
be stipulated as the cause of the linear approach taken by the to find.
authors when modeling the relationship. Overall, Model 2a is
an acceptable model in the prediction of %IR (based on
5.2 | TCDA model – (Models 3 and 4)
established literature on the DCVG technique) with the added
bonus of simplicity and brevity due to its utilization of fewer
5.2.1 | %IR variable
variables.
The estimated coefficients for Model 3 has shown that the
trend does not sit well with current industrial understanding
5.1.2 | SR and backfill type variable
on DCVG. A better way of visualizing this is by plotting the
The SR estimated coefficients for Models 1 and 1a show a predicted TCDA based on increasing %IR using model. Other
decreasing trend with its lowest value occurring at the 0.5 variables in the model were kept constant where the mean of
quantile region. However, the estimated effect of the rock the POPD, DUC, DOC, TIS, PS, and SR similar to previous
variable on the contribution of %IR indicated an inverse trend assessments in this paper were used as the contributing
with the maximum estimated coefficients occurring also factors.
within the region of 0.25–0.5 quantile. Since these two Figure 10 shows the linear effect of %IR toward the
variables are somewhat related, the opposite predictions seem resulting TCDA estimation. At the lowest quantiles (0.05 and
to complement each other and highlights the heterogenous 0.25) the effect is almost zero which is represented by the flat
nature of soil. Highly resistive electrolyte which contain line. The trend in Figure 10 is not surprising if one is to look at
materials such as rocks will produce large amounts of voltage the Oriset data where small indications of %IR has been
drop as current passes through it. These voltage drops are paired to very large coating defects and vice versa. The same
likely to be detected by the DCVG instrument which indicates scenario is encountered during the construction of Models 1
a defect more severe than it actually is. This is confirmed by and 1a. The irregularities we see here can also be explained by
the works of Mckinney[12] in his thesis which states that the bullet points given in Section 5.1.1.
prioritization of DCVG indication will be more accurate if SR The most probable cause for this trend is due to the
is taken into account. The higher quantiles highlight a disturbance coming from stray and telluric currents. Most of
relatively weak effect of the rock variable to %IR. However,
this can be understood by also observing the value estimated
for the general SR variable which highlights a stronger effect.
With respect to Models 2 and 2a, the reference variable for %IR VS TCDA (BQR MODEL 3)
the models were changed and the variable backfill type – clay, 0.05 0.25 0.5 0.75 0.95
300000
shows increasing trend until it reduces at the 0.95 quantile. 280000
Clay is considered to have high degree of compactness thus 260000
240000
possessing low resistance toward current flow. The low 220000
200000
TCDA (CM^2)
resistance would not produce large voltage drops and hence 180000
160000
one would not expect the raising trend of the estimated 140000
120000
coefficients. However, if we were to look at the backfill 100000
80000
geometry – round variable, the estimates are much more 60000
40000
streamlined with common understanding. The presence of 20000
0
rounded soil grains creates an environment which is less 0 10 20 30 40 50 60 70 80 90 100
%IR
resistant to electrical currents (similar to clay). Across the
quantile, the estimated coefficient values show a downward FIGURE 10 %IR versus TCDA for Model 3. Each color
trend with a slight increase at the highest quantile. This is the represent a different quantile. Reproduced with permission from TWI
inverse of the clay variable's trend. Similar to Models 1 and Ltd. [Color figure can be viewed at wileyonlinelibrary.com]
18
| BIN MUHD NOOR ET AL.
the pipes under assessment were situated within a network of coefficients. As such, the models are more general and are
pipelines which run in parallel and perpendicular with the one sufficient for the case of subsequent inspection of the MEOC
that is under investigation. Currents from adjacent CP pipelines.
systems which are protecting other pipelines have the
potential of leaving its intended path and are being picked
5.2.2 | POPD variable
up by the DCVG instrument. Kutz[19] has explained this
problem in greater detail. Findings from Model 3 indicated that at large coating defect
Another interesting finding was that the pipes were area the possibility of finding deeper corrosion pits are more
originally protected by a sacrificial anode system. The anodes likely. With larger TCDA, the amount of current provided by
were attached to the pipe via cad welds. Based on the pre- the cathodic protection system also should be large. When the
assessments photographs, cad welds were still visible and not level of protection current is inadequate or obstruction of the
insulated. Since these cad welds and its connecting rod are current's path in the form of a shielding electrolyte is present,
exposed to the environment, they provide an exit point for the one is to expect corrosion activity to be highly likely.
currents to leave the surface of the pipeline. The exiting However, a dip at quantile 0.75 tells us that at pipelines with
currents can also meddle with the voltage gradient generated by medium to large TCDA corresponds to corrosion pits with
the coating defects which in turn produces misleading shallower depths which goes against the normal assumption
information toward the interpretation of %IR. Apart from that a pit's depth is directly proportional to the size of TCDA.
disturbing the potential gradient signal, the exposed cad welds At first glance, Model 4 does not exhibit such issues. At the
and its associating rod could also lead to accelerated corrosion. same quantile, the coefficient predicted shows a smooth
However, corrosion was not observed at these points. increase from the median quantile to the largest quantile.
The relationship of %IR and TCDA based on Model 4 is Moreover, for Model 4, a consistent upward motion can be
illustrated in Figure 11. The trend in Figure 11 illustrates the seen across the TCDA quantiles. Between the 0.25 and the
general industrial understanding of the relationship between 0.75 quantile, shows a plateau of estimates suggesting that for
%IR and TCDA. As the quantiles increase, so does the effect these defect sizes, the effect of an increasing POPD is
of %IR on TCDA which leads to the conclusion of higher %IR minimal. The increase in values from the 0.05 quantile to the
affecting larger coating defect areas in a positive way. It can 0.25 quantile can be judged as an initial step toward the
also be said that the sensitivity of the DCVG technique relies corrosion process. At this stage, corrosion is initiated and
on the size of the coating defect. Medium to large defects give coating defects grow in tandem. The plateau is an indication
a reasonable approximation of the defect size. However, the that the pit growth rate is faster than the growth of TCDA.
interpretation based on the %IR on smaller defects should be This will produce deeper pits at smaller TCDA which
treated with caution due to large amounts of zero readings solidifies the notion that pit depth is not proportional to the
present at lower quantiles. As was mentioned earlier, outliers size of coating defect – at least not linearly. This finding was
were omitted based on careful judgment. Due to this, Model 4 also observed in Ref. [11]. Deeper pits at smaller coating
does not suffer from the problems faced by Model 3, Model 1, defect should be treated with caution as defects of such
and Model 1a where outliers play a role in the estimation of characteristics will normally go unnoticed with the conse-
quence of failure being very severe. The effect of direct
proportionality between pit depths and coating defect size can
be seen between the 0.75 and 0.95 quantile. However,
%IR VS TCDA (BQR MODEL 4) between these quantiles the credible interval increases in
0.05 0.25 0.5 0.75 0.95 wideness indicating a less certain prediction. The OLS
300000
280000
prediction is also located in the negative region which means
260000
240000
that all the above observation would be missed with the
220000
200000
average approach.
TCDA (CM^2)
180000
160000
140000
120000
100000
5.2.3 | SR variable
80000
60000 Model 4's predicted coefficient quantile trend can be
40000
20000 interpreted as highly resistive soil having a large effect on
0
0 10 20 30 40 50 60 70 80 90 100 the size of coating defects. Coarse grained soil is known to be
%IR
highly resistant to electrical current flow hence soils such as
FIGURE 11 %IR versus TCDA for Model 4. Each color sand, silt or even rocks poses high units of SR. These types of
represent a different quantile. Reproduced with permission from TWI soil with its angular particle geometry have the possibility of
Ltd. [Color figure can be viewed at wileyonlinelibrary.com] damaging the pipe coatings through the process of abrasion.
BIN MUHD NOOR ET AL.
| 19
Pipe or soil movement have the possibility of creating no known methods to check the convergence of MCMC at
abrasion between the coating interface and the electrolyte. this moment.[24] The authors had to rely on the graphical
Another factor to consider is the stresses created by the self- representation of the trace plots which lacks mathematical
weight of the backfill.[21,22] The backfill weight applies justification.
stresses on to the pipe's coating creating a wrinkling affect
normally found at 8 and 4 o’clock position of the pipe. The
wrinkling of the coating combined with the abrasion effects of 6 | CONCLUSION AND FUTURE
the angular particle size (high SR) will sometime result in the WORK
coating tearing apart.
This paper has showed that Bayesian techniques on quantile
regression is an essential tool for engineers in assessing
5.3 | Why use Bayesian quantile regression?
uncertain data. ECDA pipeline data particularly for the
The coefficient estimates illustrated by both the Bayesian DCVG technique incorporates large amounts of uncertainty
and classical method in this paper are somewhat similar. due to the unknown factors such as the factors highlighted
Both approaches consider parameter uncertainty with the in Section 5.1.1, the heterogeneity of soils, the levels of CP
Bayesian approach being more reliable as it does not rely on current, and human factors. As was mentioned earlier,
asymptotic approximation of the variances. Classical Bayesian techniques allow an assessor to quantify the full
approach such as bootstrapping in the construction of spectrum of uncertainty in the prediction of parameters.
confidence intervals uses estimation of the asymptotic In certain countries, the law dictates that an ECDA should
variances and depend on the model error density which is be performed on a periodic schedule to ensure the safe
difficult to reliably estimate. Hence, the coverage probabil- continual operation of the pipeline.[25] The NACE SP0502-
ities of the true parameter of these methods is sufficient at 2010[9] highlights the importance of periodic assessments
best but not necessarily 100% reliable. This is supported by where “through successive applications of the ECDA method,
a paper from Ref. [23] which shows the classical approach an operator will be able to identify and address locations of
estimated a lower probability of containing the parameter corrosion activity which has occurred, is occurring and at
value from the confidence interval as compared to the locations where there is a potential to occur.” This makes the
Bayesian approach. This seems to suggest that a Bayesian ECDA a continuous updating process. The Bayesian principle
method is better in terms of coverage and thus includes all fits this philosophy nicely since updating the findings from
parameter uncertainty. Other advantages of the Bayesian this paper is made possible with future ECDA. It is expected
method are that it provides a simple explanation based on that future findings will produce better estimates with every
the credible interval. For this paper, the credible intervals iteration of the ECDA process.
are set to be 95% and thus the true value of the coefficients The MEOC data was divided into two for the purpose of
can be explained as “having a probability of 0.95 of falling investigating the influence of outliers occurring at the upper
within the credible intervals.” For the classical method, the quantiles of the TCDA distribution. These outliers are
interpretation is not as direct. thought to be produced from one of the factors highlighted
Additionally, the BQR method uses the ALD as the in Section 5.1.1. One of the dataset had a total of four points
likelihood function. Since the likelihood function (ALD) removed based and the results of the removal can be seen in
disregard the original distribution of the data, specifying a three of the six models produced namely Models 2, 2a, and
specific distribution is not needed. The paper[16] goes on to 4. Although it is widely known that median regression
say that the use of the ALD is a “very natural and effective are robust to outliers, this was not the case for the other
way for modelling Bayesian quantile regression.” After the three models (Model 1s, 1a, and 3). A dip in the largest
Bayesian process, the resulting posterior statistics such as quantile for Model 3 with regard to the TCDA variable
the mean estimates of the quantiles and the calculated credible suggest that it was influenced by outliers. This was not seen
intervals can be used as new information for future ECDA. in Model 4 (after removal of outliers). Clearly, the quantile
This process is often referred to as Bayesian updating. regression applied here does not eliminate the problem of
In the process of conducting this research, the authors outliers entirely. An alternative way to solve this is to
found some drawbacks in employing the Bayesian method. construct the model with a non-linear approach. However,
One of them being the problem of convergence. As was seen with the already established Bayesian approach here, this is
in the results of Model 4, up to 11 million iterations were not necessary. All that is required is new data and the
needed to achieve convergence. This is due to the nature of Bayesian method will update the findings here.
the sampling algorithm (Metropolis-Hastings) which uses the For the estimation of the effect of soil resistivity (SR) on
accept and reject approach in the goal of achieving %IR, it is concluded that one must look at both the soil
convergence at the stationary distribution. Also, there are resistivity measurements and the effect of soil grain geometry
20
| BIN MUHD NOOR ET AL.
together. The two variables seem to have some relation and Equipment and Reliability (ICPER 2010), Kuala Lumpur, Malaysia,
provide a much more holistic picture on the effect its having 2010.
on the contribution toward %IR. Tests such as the variance [5] M. Romano, M. Dabiri, A. Kehr, JPCL 2005, 22, 7.
[6] K. R. Larsen, MP 2016, 55, 32.
inflation factors[26] for multicollinearity effects could be used
[7] S. Papavinasam, M. Attard, R. W. Revie, MP 2006, 45, 28.
for future work to see whether the variables are statistically
[8] J. M. Esteban, K. J. Kennelley, M. E. Orazem, R. M. Degerstedt,
correlated. Corrosion 1997, 53, 264.
As for the case of pit depths (POPD), the rate of growth [9] NACE, SP0502-2010: Pipeline External Corrosion Direct
between the depth of pits and the size of coating defects is not Assessment Methodology, NACE International, Houston 2010.
proportional. At some point in time the rate of corrosion is [10] F. Varela, M. Yongjun Tan, M. Forsyth, CSET 2015, 50, 226.
faster which resulted in very deep pits occurring in smaller [11] F. Anes-Arteche, K. Yu, U. Bharadwaj, C. Lee, B. Wang, Mater.
coating defect area. This is illustrated in Model 4. In this Corros. 2017, 68, 329.
[12] J. P. McKinney, Master Thesis, University of Florida, USA, 2006.
situation, the chances of locating small coating defects is low
[13] O. C. Moghissi, J. P. McKinney, M. E. Orazem, D. D’Zurko,
and hence elevating the risk of failure of the pipeline. It can be presented at Corrosion 2009, 22–26 March, Atlanta, Georgia,
said that small coating defects should not be taken lightly 2009.
especially if the environment for corrosion is highly likely. [14] R. Koenker, K. Hallock, JEP 2001, 15, 43.
Overall, each model represents a unique trait which [15] L. Hao, D. Q. Naiman, Quantile Regression. Sage Publication,
or which does not agree with established theories. The Thousand Oaks, CA, USA 2007.
differences are largely due to the influence of external factors [16] K. Yu, R. A. Moyeed, Statist. Probab. Lett. 2001, 54, 437.
which disrupts the obtained DCVG indications and thus [17] ASTM G187-12a Standard Test Method for Measurement of Soil
Resistivity Using Two Electrode Soil Box Method, American
influences the outcome of the analyses. Fortunately, these
Society for Testing Materials, West Conshohocken 2013.
uncertainties were considered and by continually updating the [18] K. Yu, J. Zhang, Commun. Stat. Theory Methods 2005, 34,
results through successive iterations of the ECDA, one can 1867.
only improve the understanding of the state of the pipeline [19] M. Kutz, Handbook of Environmental Degradation of Materials.
translating into the reduction of operating costs, enhancement William Andrew, Elsevier, Amsterdam 2005.
of safety and keeping failure risks at bay. [20] M. Shwehdi, U. Johar, International conference (ICNIR2003)
EMFI, 2003, pp. 20–22.
[21] R. Norsworthy, Corrosion 2009, March 22nd, Atlanta, Georgia,
ACKNOWLEDGMENT 2009.
[22] B. C. Yen, G. D. Tofani, OGJ 1985, 83, 34.
The authors would like to thank TWI Ltd. for providing the [23] J. L. Devore, Probability and Statistics for Engineering and the
necessary data for the work presented in this paper. Sciences. Cengage Learning, Boston 2011.
[24] K. Yu, P. Van Kerm, J. Zhang, Sankhya Ser. A 2005, 67, 359.
[25] U. S. DOT 49 CFR 192.925 − What are the requirements for using
ORCID External Corrosion Direct Assessment (ECDA)? U. S. Department
of Transportation, New Jersey 2011.
Nik N. Bin Muhd Noor http://orcid.org/0000-0001-9527-
[26] J. Neter, M. H. Kutner, C. J. Nachtsheim, W. Wasserman, Applied
2874 Linear Statistical Models. McGraw-Hill/Irwin, Chicago, USA
2004.
REFERENCES
How to cite this article: Bin Muhd Noor NN, Yu
[1] D. Furchtgott-Roth, Pipelines Are Safest For Transportation of Oil K, Bharadwaj U, Gan T-H. Making use of external
and Gas. Manhattan Institute for Policy Research, New York 2013.
corrosion defect assessment (ECDA) data to predict
[2] C. Lam, Master Thesis, UWO Ontario, Canada, 2015.
[3] M. Nasir, K. wee Chong, S. Osman, W. S. Khur, Mater. Sci. Eng.
DCVG %IR drop and coating defect area. Materials
2015, 78, 012026. and Corrosion. 2018;1–20.
[4] N. Noor, S. R. Othman, L. K. Sing, N. Yahya, M. N. M. A. Napiah, https://doi.org/10.1002/maco.201810085
Z. Abdullah, presented at 2nd International Conference on Plant