1-s2.0-S0926580514001149-main

7/24/2019 1-s2.0-S0926580514001149-main

1/16

Articial neural network models for predicting condition of offshore oil

and gas pipelines

Mohammed S. El-Abbasy a,, Ahmed Senouci b, Tarek Zayed a, Farid Mirahadi a, Laya Parvizsedghy a

a Dept. of Building, Civil and Environmental Engineering, Concordia University, Montreal, Quebec, Canadab Dept. of Civil and Environmental Engineering, Qatar University, Doha, Qatar

a b s t r a c ta r t i c l e i n f o

Article history:

Received 10 August 2013Revised 17 March 2014

Accepted 3 May 2014

Available online 22 May 2014

Keywords:

Offshore oil and gas pipelines

Condition prediction

Articial neural network

Pipelines daily transport and distribute huge amounts of oil and gas across the world. They are considered the

safest method of transporting oil and gas because of their limited number of failures. However, pipelines aresubject to deterioration and degradation. It is therefore important that pipelines be effectively monitored to

optimize their operation and to reduce their failures to an acceptable safety limit. Numerous models have been

developed recently to predict pipeline conditions. Nevertheless, most of these models have used corrosion

features alone to assess the condition of pipelines. Hence, this paper presents the development of models that

evaluate and predict the condition of offshore oil and gas pipelines based on several factors besides corrosion.

The models were developed using articial neural network (ANN) technique based on historical inspection

data collected from three existing offshore oil and gas pipelines in Qatar. The models were able to successfully

predict pipeline conditions with an average percent validity above 97% when applied to the validation data set.

The models are expected to help pipeline operators to assess and predict the condition of existing oil and gas

pipelines and hence prioritize the planning of their inspection and rehabilitation.

2014 Elsevier B.V. All rights reserved.

1. Introduction

Pipelines remain a reliablemeans of transportingproductsvital to the

sustenance of national economies in theworldsuch as water,oil, and gas

[11]. The rst oilpipeline, which was built in 1879 in Pennsylvania, was

109 miles long and 6 in. in diameter [21]. Nowadays more than 60

countries have pipeline networks exceeding 2000 km in length. The

United States has the longest pipeline network followed by Russia[17].

Gas and oil pipelines are subject to deterioration and degradation.

Pipeline accidents can cause catastrophic environmental damage due

to oil spillage as well as economic losses due to production interruption

[12]. Several types of accidents of oil and gas pipelines have been

recorded in the CONCAWE (CONservation of Clean Air and Water in

Europe) report [9]. They are most frequently classied inve cause cat-

egories: 1) third party, which represents a damage causedby operations

carried out by others in the pipeline vicinity and not related to its man-

agement; 2) corrosion, which consists of an inside corrosion related to

the product being transported and an outside corrosion related to the

pipeline coating and cathodic protection; 3) mechanical, which consists

of fracturesor cracksoccurring when efforts go beyond the efforts of the

system permits; 4) operational error, which is caused by excessive

pressure or system malfunction; and 5) natural events such as land-slides, oods, erosion in general, subsidence, earthquakes, frost or light-

ning. It is therefore important that the condition of pipelines is

effectively monitored to optimize their operation and reduce their fail-

ures to an acceptable safety limit.

Most of the developed condition assessment models are either sub-

jective (i.e., depending only on expert opinion considering no historical

data), or not comprehensive (i.e., dealing with only one failure cause).

Therefore, the objective of this research is to develop a more compre-

hensive condition assessment model that allows pipeline operators to

take the necessary actions to prevent future catastrophic failures.

2. Research objectives

Themain objectives of thepresent study areto designa conditionas-

sessment and prediction models and develop expected deterioration

curves for offshore oil and gas pipelines.

3. Background

Variousattemptshave been carried outin the last two decades foroil

and gas pipelines' leakage, defect type, corrosion rate, failure type, and

failure probability predictions. Ren et al. [30]applied back propagation

neural network to predict the corrosion rate of natural gas pipelines.

Particle swarm optimization (PSO) technique was employed by Liao

et al.[23]to develop a model that predicts internal corrosion rate for

Automation in Construction 45 (2014) 5065

Corresponding author.

E-mail addresses:[email protected](M.S. El-Abbasy),[email protected]

(A. Senouci), [email protected] (T. Zayed),[email protected] (F. Mirahadi),

[email protected](L. Parvizsedghy).

http://dx.doi.org/10.1016/j.autcon.2014.05.003

0926-5805/ 2014 Elsevier B.V. All rights reserved.

Contents lists available at ScienceDirect

Automation in Construction

j o u r n a l h o m e p a g e : w w w . e l s e v i e r . c o m / l o c a t e / a u t c o n
http://dx.doi.org/10.1016/j.autcon.2014.05.003http://dx.doi.org/10.1016/j.autcon.2014.05.003http://dx.doi.org/10.1016/j.autcon.2014.05.003mailto:[email protected]:[email protected]:[email protected]:[email protected]:[email protected]://dx.doi.org/10.1016/j.autcon.2014.05.003http://www.sciencedirect.com/science/journal/09265805http://www.sciencedirect.com/science/journal/09265805http://dx.doi.org/10.1016/j.autcon.2014.05.003mailto:[email protected]:[email protected]:[email protected]:[email protected]:[email protected]://dx.doi.org/10.1016/j.autcon.2014.05.003http://crossmark.crossref.org/dialog/?doi=10.1016/j.autcon.2014.05.003&domain=pdf

7/24/2019 1-s2.0-S0926580514001149-main

2/16

wetgas gatheringpipelines. Themodel wasalso developedusing genet-

ic algorithm (GA) and ANN techniques which outperformed the PSO.

Dawotola et al.[10]developed a rupture risk management model for

crude oil pipelines using a methodology that incorporates structured

expert judgment and analytic hierarchy process (AHP). Noor et al. [27]

used semi-probabilistic and deterministic methodologies to predict

the remaining strength of submarine pipelines subjected to internal

corrosion. Bersani et al.[4]developed a risk assessment model using

historical data from the United States Department of Transportation(DOT) to predict the failure causedby third party activity. Historical fail-

ure data wasalso used to develop a tool to predict the class of eachspill-

age in oil pipelines using statistical analysis classication and regression

tree [5].Lietal. [22] presenteda methodto predictthe corrosionand the

remaining life of underground pipelines using a mechanically-based

probabilistic model that considers the effect of randomness in pipeline

corrosion using Monte Carlo simulation technique. Singh and Markeset

[33]presented a proposed methodology, based on fuzzy logic frame-

work, for theestablishment of a risk-based inspection program for pipe-

lines according to the estimation of its corrosion rate. Peng et al. [28]

developed a fuzzy neural network model, which is based on failure

tree and fuzzy computing, to predict the rate of failure for oil and gas

pipelines. Dawotola (2009)[41]proposed a combined AHP and Fault

Tree Analysis to support the design, theconstruction, and theinspection

and maintenance of oil and gas pipelinesby proposing an optimal selec-

tion strategy based on the probability and consequence of failure. Jinhai

et al.[19]proposed a leak fault-detection method based on the combi-

nation of Rough Set (RS) and ANN, calledhybrid fault-detection method

based on RS and ANN (HFDMRSNN). Carvalho et al. [7]used the ANN

technique for pattern recognition of magnetic ux leakage (MFL) sig-

nals in weld joints of pipelines obtained by intelligent pig to distinguish

the presence of defects and their type. AHP was used by Dey [12] to de-

velop a model to help decision makers select a suitable type of inspec-

tion or monitoring technique for pipelines. Hallen et al.[18]presented

a probabilistic analysis framework to evaluate the condition of a corrod-

ing pipeline and the evolution of its probability of failure with time.

Sinha and Pandey [34] developed a simulation-based probabilistic

fuzzy neural network model to estimate the failure probability of

aging oil and gas pipelines due to corrosion. Ahammed [1]presented amethodology to assess the remaining service life of a pressurized pipe-

line containing active corrosion defects. Belsito et al. [3]developed a

leakage detection system for liqueed gas pipelines using ANN for

leak sizing and location.

Most of the previously-mentioned models were either subjective

[10,12]or did not cover all the failure causes of oil and gas pipelines

[1,4,18,22,23,26,27,30,33,34] . In other words, they lack the objectivity

in predicting the different failure types of pipelines. As a result, Senouci

et al.[31]developed a regression and articial neural network (ANN)

models to predict possible failure types for oil and gas pipelines. The

model took into consideration the prediction of failure types beside cor-

rosion, such as mechanical, third party, natural hazard, and operational

failures. The model was built based on a historical data collected from a

report that was prepared by CONCAWE [9]. Later, [32] developedanother model for the same purpose using fuzzy logic technique and

compared the results with those obtained using the regression and

ANN models developed by Senouci et al. [31]. The results comparison

showed that the developed fuzzy-based model outperformed the

regression and ANN models with respect to model validity.

Despite the attempts made to predict the failure type of oil pipelines

considering causes other than corrosion, still none of the previously-

mentioned models can be used to assess the condition of pipelines.

Actually, such models were intended to either predict the corrosion

rate, pipe leakage, or failure/defect type focusing mostly on corrosion

related factors. In addition to that, the important issues ofinterdepen-

dencybetween different factors' relations anduncertaintyof factors'

severity weights were not addressed simultaneously. Furthermore,

none of the previous studies developed models that can forecast the

deterioration rate of oil and gas pipelines and hence can build up their

respective deterioration curve. Consequently, El-Abbasy et al.[14]de-

veloped a model that assesses the condition of oil and gas pipelines

based on several factors including corrosion using both Analytic Net-

work Process (ANP) and Monte-Carlosimulation.The model considered

factors' interdependency (using ANP), made decisions under uncertain-

ty (using simulation), and handled decisions involving large number of

variables (using integrated simulation/ANP). It was successfully tested

on an existing offshore gas pipeline in Qatar by comparing the resultsobtained using the model with the actual pipeline condition.

Thesimulation model built by El-Abbasy et al. [14] is considered as a

rst phase to evaluate or assess the condition of offshore oil and gas

pipelines. Two major limitations were found in the study conducted

by El-Abbasy et al.[14]. First, the factors that were used to predict the

pipeline condition were not sufcient due to the lack of collected data.

The factors related to the pipeline structural condition were also not

considered. Second, the model was developed and tested using inspec-

tion data for a single 12-inch gas pipeline. The limited historical inspec-

tion data did not allow the examination of the effect of changing

the pipe diameter or the type of the transported product on the devel-

oped model. Moreover, the limited inspection data did not allow the

development of sound deterioration curves for oil and gas pipelines.

To overcome the above limitations, an extensive data collection was

performed in a study conducted by El-Abbasy et al. [15]where seven

historical inspection data seta were collected for three different

pipelines with several sizes (i.e. diameters), materials, and types of the

carried product. Four factors were considered in addition to those

presented in the previous study[14].The additional factors included

the anode wastageand three factors related to the pipeline structural

condition namely, support condition, joint condition, and free

spans. The regression analysis technique was used to correlate be-

tween all these factors using the gathered data. The developed models

were validated yielding an average validity percentage above 96%. In

addition, the study proposed a standard condition assessment scale

or rating system for oil and gas pipelines. This rating system can be

used as a guideline to decide and plan the maintenance of pipelines

(i.e., lining, cathodic protection, replacement, etc.) and to prioritize

rehabilitation/upgrading projects within the approved budget.Although the study conducted by El-Abbasy et al. [15]provided

sound results, the use of the regression analysis technique as a machine

learning algorithm has still some limitations. Several other machine

learning algorithms can be used to predict pipeline condition including

ANN, Support Vector Machine (SVM), Nave Bayes (NB), Decision Trees

(DT), Random Forest (RF), and K-Nearest Neighbor (K-NN). Although

ANN is considered one of the oldest methods, it is still found to be a

competitive algorithm among the new ones. For instance, Tahyudin

et al. [36] compared theperformance of SVM, DT, ANN, NB, and Logistic

Regression (LR) to predict the graduation students on time. The results

showed that ANN and SVM were the best predictors with an accuracy

rate almost100%.Caruana and Niculescu-Mizil[6] carried out a compar-

ison between ten supervised learning algorithms being applied on

eleven binary classication problems using eight performance metrics.It was found that the ANN was among the top ve best algorithms.

Other studies showed the outperformance of the ANN technique.

Mohana and Thangaraj[24]showed that ANN is better than SVM for

modelling resource state prediction. Prabhakar[29]also showed the

better performance of ANN over SVM in predicting software effort.

Mollazade et al. [25] compared four different learning algorithms,

namely, ANN, SVM, DT, and Bayesian Network (BN) for grading raisins

based on visual features. Results of validation stage showed ANN had

the highest classication accuracy, 96.33%. After ANN, SVM with poly-

nomial kernel function (95.67%), DT with J48 algorithm (94.67%) and

BN with simulated annealing learning (94.33%) had higher accuracy,

respectively. On the contrary, Folorunsho[16]used medical dataset to

predict the diabetes probability of any patient using ANN and DT. It

was found that DT outperformed ANN with a lower error metrics and

51M.S. El-Abbasy et al. / Automation in Construction 45 (2014) 5065

7/24/2019 1-s2.0-S0926580514001149-main

3/16

higher correlation coefcient. Such learning algorithms were also ap-

plied in elds that are relevant to our study, i.e. pipe condition monitor-

ing. Kalanaki and Soltani [20] developed a break rate prediction models

for water pipes using SVM. Performance results were compared with

continuous genetic algorithm-based support vector regression (SVR-

GA), continuous ant colony algorithm-based SVR (SVR-ACO), PSO-

based SVR (SVR-PSO), ANN, and adaptive neuro-fuzzy inference sys-

tems (ANFIS). The results revealed the outperformance of using SVM

as a machine learning algorithm. On the other hand, Tabesh et al. [35]assessed pipe failure rate and mechanical reliability of water distribu-

tion networks using ANN, ANFIS, and Nonlinear Regression (NLR)

where the results indicated the robustness of using ANN over the

other methods.

Based on the above, it is quite clear that there is no a specic ma-

chine learning algorithm have an absolute outperformance over the

other. In other words, some algorithms may perform clearly better or

worse than the others depending on the problem nature as well as the

data format. While a given machine learning approach may be easier

to implement for a given problem, or more suited to a particular data

format, to tackledifcult problems what matters in theend is theexper-

tise a scientist has in a particular machine learning technology. What

can be obtained with a general-purpose machine learning method can

be achieved using another general-purpose machine learning method,

provided the learning architecture and algorithms are properly crafted

[8]. As a result, it was decided in this study to use ANN to develop con-

dition prediction models for oil and gas pipelines. The main advantage

of ANN is related to the capability for learning from specic predened

patterns. The learning capacity may include classication, prediction,

and controlling of any specic task. Besides that, ANN provides two

major advantages over the regression analysis technique which have

been used by El-Abbasy et al. [15]to predict oil and gas pipeline condi-

tion. First, ANN has the ability to detect implicitly any complex nonlin-

ear relationships between independent and dependent variables[38].

If a signicant amount of nonlinearity between the predictor variables

and the corresponding outcomes is present in a training data set, the

network will adjust the connection weights to reect these nonlinear-

ities. Thepredictor variables in a neural network usually undergo a non-

lineartransformationat each hiddennode andoutput node. Therefore, aneural network can potentially model much more complex nonlinear

relationships than a logistic regression model can. Empirical observa-

tions suggest that when complex nonlinear relationships are present

in data sets, neural network models may provide a tighter model t

than conventional regression techniques [37]. Second, the hidden

layer of a neural network gives it the power to detect interactions or

interrelationships between all of the input variables. When important

interactions exist and are not modeled explicitly in a logistic regression

model, then a neural network model may be expected to predict a

particular outcome better than a logistic regression model.

Hence, in an attempt to enhance pipeline condition prediction pro-

cess, this paper presents new condition prediction models using the

ANN technique. The performance of the models are then compared

with those developed using the regression analysis technique [15].

4. Research methodology

The developed methodology, which is shown in Fig. 1, started by

performing a brief literature review to search for the different tech-

niques and studies used by researchers for the condition assessment

of oil and gas pipelines. After that, a comprehensive data collection

wasperformed to in two different stages. The rst stage was concerned

with the identication of the factors needed to increase the efciency of

the condition prediction process. The second stage included gathering

historical data for different pipeline sizes and types in Qatar based on

previous inspections' records. The provided inspection dataset did not

include all the factors identied in the rst stage of the data collection

process. Furthermore, few parameters or factors in the inspections'

dataset were not available in the factors identied previously. As a re-

sult, the compatible factors as well as the additional factors in the in-

spections dataset were selected to develop the condition prediction

models.

Five condition prediction models were developed, namely, three

models with respect to the pipeline size (i.e. diameter) and the

transported product type (i.e., oil or gas) and two models with respect

to the type of the transported product only. The details and framework

of the model development using the ANN technique will be explained

later. Themodels were then tested, validated,and compared with previ-

ously developed models. Finally, a sensitivity analysis was carried out to

develop different deterioration curves by examining the degree of the

individual impact of each factor on the pipeline condition.

5. Data collection

5.1. Factors affecting oil and gas pipeline condition assessment

The factors related to corrosion or third party features are insuf-

cient to build an accurate condition assessment model. Therefore, it

is essential to identify other factors affecting pipeline condition. The

identication process was carried out in three main steps: (1) experts

prepared the most important factors that need to be considered in a

pipeline condition assessment, (2) another list was subsequently pre-

pared from the literature, and (3) a comparison was made between

the two lists to determine the overlapping criteria. The most important

START

LITERATURE REVIEW

Identify Factors

Affecting Oil and Gas

Pipeline Condition

Select

Compatible

Factors

Collect Historical

Inspection Data for

Oil/Gas Pipelines

Three models with respect to

diameter & product type

Models

Successful?

ModifyM

odels

DETERIORATION CURVES

END

No

Yes

DATA COLLECTION

MODEL DEVELOPMENT

MODEL TESTING & VALIDATION

Two models with respect to

product type only

Fig. 1.Research methodology.

52 M.S. El-Abbasy et al. / Automation in Construction 45 (2014) 5065

7/24/2019 1-s2.0-S0926580514001149-main

4/16

factors affecting pipeline condition were then identied. These factors

were divided into three main groups, namely, physical, external, and

operational. The physical factors comprise general pipeline characteris-

tics such as age, wall thickness, diameter, and applied coating condition.

The external factors deal with the surroundingenvironmental condition

of the pipeline while the operational factors deal with the adapted

operational strategies for the pipelines. The questionnaire that wasdeveloped for the factors' identication was sent to 55 experts in oil

and gas pipelines integrity management. Out of the 55 questionnaires

sent, 28 completed questionnaires were received out of which only 25

were taken into consideration from the targeted sample, which repre-

sents 45.5% of the total sample. The respondents were asset, inspection,

and operation managers as well as onshore/offshore inspection engi-

neers with a technical experience of 6 to more than 20 years. It should

be noted that 22 of the respondents were inspection/operation man-

agers with a technical experience of more than 20 years. It should also

be noted that the surveys were mainly collected from the Middle-East

region, mostly from Qatar and Saudi Arabia.

5.2. Historical inspection data

Inspections data for three pipelines were collected from a large

state-owned oil and gas industry corporation in Qatar. The three pipe-

lines were located offshore and were of different sizes and different

steel grades. The main characteristics of the three pipelines are shown

in Table 1. Different inspection types were received whether they

were internal or external. The internal inspections included reports

and excel sheets obtained from the Magnetic Flux Leakage (MFL)

inspection method. On the other hand, the external inspections includ-

ed reports generated from Remotely Operated Vehicle (ROV) inspec-

tions and cathodic protection monitoring system. The inspection

reports were provided for two to three different inspection time inter-

vals covering the full lengthof the offshore pipelines. As mentionedear-

lier, the provided inspection dataset did not include all the factors

identied in the rst stage of the data collection process. Moreover,few factors in the inspections' dataset were not available in the factors

identied previously. As a result, the compatible factors as well as the

additional factors in the inspections dataset were selected to develop

the condition prediction models. The factors selected to develop the

condition prediction models are shown inFig. 2.

6. Model development

The model building procedure is shown inFig. 3. The historical in-

spection datasets obtained from three pipelines in Qatar were used to

represent the factors affecting the pipeline condition. The 11 factors

shown inFig. 2were used to build the models. As mentioned before,

three models were developed based on the pipeline size (i.e. diameter)

and the transported product type (i.e., oil or gas). Moreover, two addi-tional models were developed based only on the transported product

type making a total ofve models as shown inFig. 3.

After studying the inspection data of the three pipelines, it was

found that some factors were not constant across the whole pipeline

length. For example, the metal loss depth or the cathodic protection

can be high in a certain sector of the pipeline and low in another. As a

result, the pipelines were divided into several 100-meter-long

Table 1

Collected pipelines' main characteristics.

Characteristics Pipeline (1) Pipeline (2) Pipeline (3)

Total length (km) 211 45 121

Inspected length (km) 77 45 85

Overall location 80 km (Offshore) + 131 km (Onshore) 45 km (Offshore) 89 km (Offshore) + 32 km (Onshore)

Inspected location Offshore Offshore Offshore

Diameter (inches) 12 20 24

Wall thickness (mm) 9.5 12.7 12.7

Steel pipe grade X65 B X52SMYS (Mpa) 448 241 358

MAOP (bars) 125 48 50

Design pressure (bars) 139 50 107

Product type Gas Oil Gas

Construction year 1990 1972 1979

Inspection years 1996 and 2004 2001 and 2009 1996, 2001, and 2008

FACTORS AFFECTING

PIPELINE CONDITION

Diameter

Age

Metal Loss

Coating

Condition

Operating

Pressure

Anode

WastageSupport

Condition

Free Spans

Joint

Condition

Crossings

Cathodic

Protection

Fig. 2.Factors included in model development.


7/24/2019 1-s2.0-S0926580514001149-main

5/16

segments. For instance, the 24-inch gas pipeline with an inspected

length of 85 km was divided into 850 100-meter-long segments. As

shown inTable 1, inspections were carried for the 24-inch gas pipeline

in 1996, 2001, and 2008. Therefore, the total number of the data points

for developing the 24-inch gas pipeline model will be 850 3 = 2550

data points. Using the same concept; the number of data points used

for developing the 12-inch gas, 20-inch oil, gas, and oil models will be

1540, 900, 4990, and 4990, respectively.

The data for the

gas pipeline

model was generated by combiningthe data of the 12-inch and 24-inch gas pipelines with the 20-inch oil

pipeline. The 20-inch oilpipeline wastreated as a gaspipeline by calcu-

latingits actual condition using thegas model developed by El-Abbasy

et al. [14]. This will allow the prediction models to accommodate a

wider range of different pipelines' diameter and age. The same proce-

dure was followed again to obtain the data for the oil pipeline

model. In other words, the 20-inch oil pipeline data was combined

together with the 12-inch and 24-inch gas pipeline data into one set.

However, this time, the 12-inch and 24-inch gas pipelines were treated

as oil pipelines by calculating their actual condition using the oil

modeldeveloped by El-Abbasy et al.[14].

6.1. ANN model building

The ANN model is developed and analyzed using the Neuroshell

2V4.0 package [37], which is a commercially available neural network

analysis and modeling software. The ANN application framework of thecondition prediction problem is shown inFig. 3. The inspection data for

selected factors are used to train theANN in order to obtain ANN-based

condition prediction models. The inspection data points were divided

randomly into three sets: (1) 60% for training; (2) 20% for testing; and

(3) 20% for validation. The training set is used to train the network

whereas the testing set is used to test the network during the develop-

ment/training and also to continuously correct it by adjusting the

weights of network links.The validation data set, which is not presented

Define Models to be Developed

(1) 12-Inch Gas Pipes (2) 20-Inch Oil Pipes (3) 24-Inch Gas Pipes (4) GasPipes (5) Oil Pipes

Investigated variables

Condition Prediction Model Using

ANN Technique

Define Training, Testing, & Validation Datasets

Select Learning Algorithm

Model Validation

Actual vs. Predicted

Outputs

Mathematical

Validation

Train Model

Define Design Parameters Define Training Criteria

Recall Model for Testing

Test Model

Testing Successful?

YES

NO

NO

YES

END

START

Fig. 3.Model development framework.


7/24/2019 1-s2.0-S0926580514001149-main

6/16

to the network during training, is used to validate the model. To classify

data into the abovementioned data sets, data patterns are randomly

selected from the entire data space, according to the set percentages.

It is worth noting that the factors' real numeric values were normalized

according to the normalization method shown inTable 2.

6.1.1. Network architecture

After dividing the datasets; the inputs and outputs were identied

and selected. The inputs or predictors represent the factors affecting

the pipeline condition (Fig. 2) while the output represents the pipeline

condition. It should be noted that the pipeline condition is represented

by a numeric scale value from 0to 10, where 0indicates that the

pipeline is at its worst condition and 10at its best condition. Such

scale was adapted from the study conducted by El-Abbasy et al. [15]

which mainly depended on experts' opinions. Each pipeline's condition

numeric range and its corresponding linguistic description, most possi-

ble associated features, and suggested required action is shown in

Table 3[15]. A supervised ANN, using the back propagation algorithm,

is used to develop the condition prediction models. The network archi-

tecture has a total ofve layersof neurons with oneinput, three hidden,

and one output layers as shown inFig. 4.

6.1.2. Design parameters and training criteria

The design parameters and training criteria dened for the ANN

models are shown inTable 4. The selection of the input and output

variables greatly affects the ANN architecture and depends upon the

nature of the problem. The input variables for the gasand oilpipe-

line models include the11 factorsshown in Fig. 2. However, the diam-

eter factor is excluded from thelist for the 12-inch gas, 20-inch oil,

and 24-inch gaspipelines models since the diameter is constant for

those models. Each variable is represented by one arti

cial neuron inthe network's input and output layers. The input layer consists of 10

neurons that represent the factors affecting the pipeline condition

while the output layer has only one neuron that represents the pipeline

condition. The number of neurons in the hidden layers determines how

well a problem can be learned. If too many are used, the network will

tend to memorize the problem, and thus not generalize well later. On

the other hand, if too few are used, the network will generalize well

but may not have enough powerto learn the patterns well [37]. In

other words, the hidden layers rely on the available model building

dataset and the nature of outputs. Therefore, getting the right number

of hidden neurons is a matter of trial and error. Several iterations are

made to generate the optimum number of neurons in the hidden layer

as shown inTable 4.

Different activation functions were applied to a hidden layer to

detect different features in a pattern, which is processed through a net-

work. For example, a network design may use a Gaussian function on

one hidden layer to detect features in the mid-range of the data and

use a Gaussian complement in another hidden layer to detect features

from the upper and lower extremes of the data. Thus, the output layer

will get different views of the data. Combining the two feature sets in

the output layer may lead to a better prediction.

The interaction among the network's layers is governed by the links

between them. Each link in the network has a learning rate, momen-

tum, and initial weights. Each time a pattern is presented to the

network, the weights leading to an output node are modied slightly

during learning to produce a smaller error the next time the same pat-

tern is presented. The amount of weight modication is the learning

Table 2

Factors' normalization method.

Factor Unit Normalization method

Age (AG) Years AG/70

Diameter (DI) Inches (DI 2)/(52 2)

Metal loss (ML) % ML 1

Coating condition (CC) % CC 1

Crossings (CR) Number CR/20

Cathodic protecion (CP) mV (CP 500)/(1300 500)

Operating pressure (OP) % of Design Pressure OP 1Anode wastage (AW) % AW 1

Support condition (SC) % SC 1

Free spans (FS) meters FS/70

Joint condition (JC) % JC 1

Table 3

Condition assessment scale[15].

Overall pipeline

condition

Linguistic

scale

Description Action(s) required

Fit for purpose

910 Excellent Internal & external corrosion:almost no signs

C orrosion rate: highlybelow expected average rate (0.08 mm/year).

Cathodic protection:very good and within acceptable limits (N950 mV).

Coating:very well maintained (b15% coat loss).

Inhibitor's efciency: N95%

Reassess in 10 years

Regular annual maintenance.

79 Very good Internal & external corrosion:few signs

Corrosion rate:below expected average rate (b0.08 mm/year).

Cathodic protection:good and within acceptable limits (850950 mV).

Coating:well maintained (1530% coat loss).

Inhibitor's efciency:90

95%.

Reassess in 8 years


57 Good Internal & external corrosion:average signs

Corrosion rate:nearly equal to the expected average rate (0.08 mm/year).

Cathodic protection:adequate (750850 mV).

Coating:still intact (3040% coat loss)

Inhibitor's efciency:8090%.

Reassess in 5 years


Schedule for CP & recoating within the next 2 years.

35 Moderate Internal & external corrosion:signicant signs

Corrosion rate:above expected average rate (N0.08 mm/year).

Cathodic protection:inadequate (675750 mV).

Coating:partially damaged (4060% coat loss)

Inhibitor's efciency:7080%.

Reassess in 2 years


S chedule for minor rehabilitation within the next 1 year.

S chedule for CP & recoating within the next 1 year.

Close to failure

03 Critical Internal & external corrosion:severe signs

Corrosion rate:highly above expected average rate (0.08 mm/year).

Cathodic protection:poor (b675 mV).

Coating:almost damaged (N60% coat loss)

Inhibitor's efciency: b70%.

S chedule for major rehabilitation &/or replacement immediately.


7/24/2019 1-s2.0-S0926580514001149-main

7/16

rate times the error. For example, if the learning rate is 0.5, the weight

changeis one half the error. Large learning rates often lead to oscillation

of weight changes and learning never completes, or to a convergence to

a non-optimum solution. One way to allow faster learning without os-

cillation is to make the weight change as a function of the previous

weight change to provide a smoothing effect. The momentum factor

determines the proportion of the last weight change that is added into

the new weight change. As neurons pass values from one layer of the

network to the next layer in back propagation networks, the values

are modied by a weight value in the link that represents connection

strengths between the neurons.

Finally, the training criteria include xing the maximum and mini-

mumabsoluteerrorand thenumber of training cycles without improve-

ments. When the given training criteria are met, the model training is

stopped. If the training criteria are not met, it is suggested to either

change the activation functions or retrain using different architecture.

SLAB 1 SLAB 4

SLAB 3

SLAB 2

SLAB 5

Link 1

Link 3

Link 6

Link 4

Link 2

Link 5

Input Layer Hidden Layers Output Layer

Fig. 4.ANN models' architecture.

Table 4

Models' design parameters and training criteria.

Slab no. No. of neurons Activation function Link

Link no. Learning rate Momentum Initial weight

Design parameters 1 10a/11b Linear (0,1) 1 0.05 0.1 0.3

2 12 Gaussian 2 0.05 0.1 0.3

3 12 Tanh 3 0.05 0.1 0.3

4 12 Gaussian Complement 4 0.05 0.1 0.3

5 1 Logistics 5 0.05 0.1 0.3

6 0.05 0.1 0.3

Training criteria Stop training when one of these is true about the training set Average Error b 0.0002

Epochs since minimum average error N 1000

Stop training when one of these is true about the test set Calibration Interval (events) 200Events since minimum average error N 20,000

a For: 12-inch gas, 20-inch oil, and 24-inch gas pipes models.b For: gas and oil pipes models.

Table 5

Summary of the models' results.

Model R 2 (%) r2 (%) Mean Squared Error

(MSE)

Mean Absolute Error

(MAE)

Min. Absolute Error Max. Absolute Error Correlation

Coeff., r (%)

12-Inch gas pipes 99.04 99.04 0.011 0.083 0.000 0.502 99.52

20-Inch oil pipes 99.59 99.59 0.007 0.060 0.000 0.565 99.80

24-Inch gas pipes 99.29 99.31 0.009 0.070 0.000 0.507 99.65

Gas pipes 99.23 99.25 0.011 0.079 0.000 0.548 99.63

Oil pipes 99.28 99.29 0.011 0.081 0.000 0.566 99.64


7/24/2019 1-s2.0-S0926580514001149-main

8/16

6.1.3. Model training and testing

Consequently, themodel is trained until the training criteria are met.

The training process uses supervised learningwhere the inputs and

outputs are known within the context of the problem. After the ANN

is trained, it can be recalled to predict the pipeline condition for any

given testing input values. The testing input data sets are introduced

to the trained model in order to generate the predicted output, which

is then compared to the actual one. If they are close, the model testing

is successful andvice versa. Thetrainingand testing processeswere per-formed successfully for all the models with reasonable results as shown

in Table 5. The ANN model values of R2 were close to 1.0 while the MSE

and MAE were close to 0.0. The results conrm the robustness of the

developed models.

6.1.4. Signicance ranking of variable

Quantitative estimates of the relative contribution factor for

the input variables affecting the pipeline condition are made. The con-

tribution factor for a certain variable is considered as a rough measure

of the importance of that variable in predicting the network's output.

Therefore, the contribution factor, which is derived from the contribu-

tion factor module of the referred software, determines the variable sig-

nicance ranking. These contribution factors are developed through an

analysis of the link weights of the trained neural network. The higher

the number, the more that variable contributes toward prediction or

classication of the output. Table 6shows the factors' ranking and

their corresponding contribution factor.

The relative signicance ranking of the considered parameters,

obtained from the above analysis, can be generalized. As shown in

Table 6, it can be observed that the cathodic protectionfactor can be

considered as the most important variable affecting the pipe condition.

Themetal lossand coating conditionfactors are also relatively im-

portant in determining the pipeline condition. This shows thatminimiz-

ing the corrosion rate by applying adequate coating and cathodic

protection for offshore pipelines is essential to improve their conditions.

It can also be observed that the operating pressurefactor has signi-

cantly more impact in gas than in oil pipelines. On the other hand,

crossingsand diameterare considered as the least important factor

affecting the pipe condition. The rest of the factors, namely, age,anode wastage, support condition, joint condition, and free

spans, contributes moderately to the pipeline prediction process.

It should be noted that ranking the variables throughout the contri-

bution factors is useful in rening the variables by excluding the least

important ones. However, in the current study, no variables were ex-

cluded since the difference between each variable's contribution factor

is not signicant.

7. Model validation

The performances of the developed ve models are validated using

the 20% validation dataset. The performance of each model is validated

with respect to certain mathematical validation diagnostics as recom-

mended in literature[2,40]. Eqs.(1) and (2)show the average validi-ty/invalidity percentages (i.e., AVP and AIP) in order to predict the

error. The model is sound for an AIP value closer to 0.0 and not robust

for an AIP value closer to 100. Similarly, the Root Mean Square Error

(RMSE) is estimated using Eq.(3). If the value of the RMSE is close to

0, the model is sound and vice versa. The Mean Absolute Error (MAE)

is also dened in Eq.(4). The MAE value varies from 0 to innity and

it should be close to zero for sound results[13]. Finally, the MAE value

is also used to dene the tness functionfiof a model[13]as shown

in Eq.(5). The equation for the tness function indicates that a model

is valid when itsfivalue is close to 1000 and invalid when its fivalue

is close to 0.

AIP Xn

i11

Ei

Ci

100

n

1

T

able

6

R

elativevariables'contributionfactorsandranking.

Rank

Model

12-Inchgaspipes

20-Inchoilpipes

2

4-Inchgaspipes

Gaspipes

Oilpipes

Variable

Contributionfactor

Variable

Contributionfactor

V

ariable

Contributionfactor

Variab

le

Contributionfactor

Variable

Contributionfactor

1

Cathodicprotection

0.1

6607

Cathodicprotection

0.1

6253

C

athodicprotection

0.1

4446

Cathod

icprotection

0.1

5307

Cathodicprotection

0.1

6674

2

Metalloss

0.1

3936

Metalloss

0.1

1382

M

etalloss

0.1

2901

Metalloss

0.1

2572

Metalloss

0.1

2364

3

Coatingcondition

0.1

2299

Coatingcondition

0.1

1304

C

oatingcondition

0.1

1631

Coatin

gcondition

0.1

1175

Coatingcondition

0.1

1634

4

Operatingpressure

0.1

113

Freespans

0.0

9705

O

peratingpressure

0.0

9748

Operatingpressure

0.1

0525

Anodewastage

0.0

8406

5

Freespans

0.1

071

Supportcondition

0.0

9625

F

reespans

0.0

9451

Freesp

ans

0.0

8967

Age

0.0

8241

6

Anodewastage

0.0

9615

Anodewastage

0.0

9619

A

ge

0.0

9407

Jointcondition

0.0

8723

Jointcondition

0.0

8173

7

Age

0.0

6716

Age

0.0

9446

Jointcondition

0.0

9094

Age

0.0

7222

Supportcon

dition

0.0

7867

8

Supportcondition

0.0

6697

Jointcondition

0.0

7997

A

nodewastage

0.0

8498

Anode

wastage

0.0

7003

Freespans

0.0

7575

9

Jointcondition

0.0

6396

Crossings

0.0

7437

S

upportcondition

0.0

7593

Supportcondition

0.0

6507

Diameter

0.0

7492

10

Crossings

0.0

5893

Operatingpressure

0.0

7232

C

rossings

0.0

7232

Crossings

0.0

6006

Operatingpressure

0.0

637

11

Diame

ter

0.0

5992

Crossings

0.0

5204


7/24/2019 1-s2.0-S0926580514001149-main

9/16

Table 7

Regression and ANN models' validation comparison.

Model Regression technique ANN technique

R2 (%) AVP (%) AIP (%) RMSE MAE fi R2 (%) AVP (%) AIP (%) RMSE MAE fi

12-Inch gas pipes 98.80 97.9 2.1 0.008 0.098 911 99.04 98.1 1.9 0.007 0.093 915

20-Inch oil pipes 99.40 96.2 3.8 0.015 0.152 868 99.59 97.4 2.6 0.012 0.099 910

24-Inch gas pipes 99.20 98.3 1.7 0.005 0.079 927 99.29 98.4 1.6 0.005 0.078 928

Gas pipes 99.10 97.8 2.2 0.005 0.099 910 99.23 97.8 2.2 0.004 0.093 915

Oil pipes 99.00 97.9 2.1 0.004 0.094 914 99.28 98.0 2.0 0.004 0.089 918

0123456789

PipeCondition

Event Number

0123456789

200150100500

PipeCond

ition

Event Number

(b) 20-INCH OIL PIPES

(c) 24-INCH GAS PIPES

(d) GAS PIPES

(e) OIL PIPES

(a) 12-INCH GAS PIPES

0123456789

PipeCondition

Event Number

0123456789

PipeCondition

Event Number

012345678910

0 50 100 150 200 250 300 350

0 50 100 150 200 250 300 350 400 450 500 550

0 100 200 300 400 500 600 700 800 900 1000

0 100 200 300 400 500 600 700 800 900 1000

PipeCondition

Event Number

ACTUAL CONDITION PREDICTED CONDITION

Fig. 5.Models' validation plots.


7/24/2019 1-s2.0-S0926580514001149-main

10/16

AVP 100AIP 2

RMSEffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiX

n

i1 CiEi

2=n

q 3

MAE

Xn

i1CiEij j

n 4

fi 10001 MAE 5

Where: AIP = Average Invalidity Percent; AVP = Average Validity

Percent; RMSE = Root Mean Squared Error; MAE = Mean Absolute

Error;fi= tness function;Ei= estimated value; Ci= actual value;

andn= number of events.

Table 7 summarizes the results of the developed ANN model valida-

tion parameters. These results were compared with those obtained

using the regression models of El-Abbasy et al. [15]. As shown in the

table, both techniques provided ve corresponding models with an

AVP close to 1, AIP, RMSE, and MAE close to 0 as well as fi close to

1000. These results can be considered as good indicators to the robust

performance of the models in predicting accurately the pipeline condi-tion. Both techniques are almost similar to each other in terms of R2,

AVP, AIP, RMSE, MAE, and fi. However, the results inTable 7show that

the ANN technique provides slightly better results than the regression

technique with respect to all parameters. This is due to the fact that

the ANN technique considers the nonlinear relation of the dependent

CATHODIC PROTECTION (mv)

COATING CONDITION

CATHODIC PROTECTION

SUPPORT CONDITION

JOINT CONDITION

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

500 580 660 740 820 900 980 1060 1140 1220 1300

5.79 6.09 6.40 6.71 6.99 7.24 7.45 7.62 7.7 4 7.81 7.83

5.88 6.23 6.55 6.86 7.14 7.40 7.63 7.85 8.0 3 8.19 8.33

7.38 7.44 7.50 7.56 7.61 7.65 7.69 7.72 7.7 4 7.76 7.78

7.53 7.56 7.58 7.61 7.64 7.66 7.69 7.71 7.7 4 7.76 7.78

5.5

6.5

7.5

8.5

PIPE

CO

NDITION

NORMALIZED VALUE

12-INCH GAS PIPES DETERIORATION CURVES - DIRECTLY PROPORTIONAL FACTORS

AGE (Years)

CROSSING (#)

FREE SPANS (m)

AGE

METAL LOSS

CROSSINGS

OPERATING PRESSURE

ANODE WASTAGE

FREE SPANS

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

0 7 14 21 28 35 42 49 56 63 70

0 2 4 6 8 10 12 14 16 18 20

0 7 14 21 28 35 42 49 56 63 70

7.86 7.79 7.73 7.67 7.61 7.57 7.52 7.49 7.46 7.43 7.41

8.32 8.04 7.79 7.55 7.34 7.14 6.95 6.77 6.58 6.39 6.18

7.77 7.76 7.74 7.73 7.72 7.71 7.70 7.69 7.69 7.68 7.68

9.18 9.03 8.87 8.71 8.55 8.40 8.24 8.08 7.92 7.76 7.61

7.85 7.80 7.75 7.71 7.66 7.61 7.56 7.52 7.47 7.42 7.38

7.85 7.71 7.58 7.47 7.36 7.27 7.18 7.11 7.05 7.01 6.97

6

7

8

9

PIPE

CONDITION

NORMALIZED VALUE

12-INCH GAS PIPES DETERIORATION CURVES - INVERSELY PROPORTIONAL FACTORS

FACTORS' REAL VALUES

CORRESPONDING PIPE CONDITION



Fig. 6.12-Inch gas pipes deterioration curves.


7/24/2019 1-s2.0-S0926580514001149-main

11/16

and independent variables as well as the correlation between the

factors affecting the pipeline condition.

Fig. 5shows the actual versus predicted output plot results for the

ve developed models. It is observed that the predicted values by the

developedmodelsare withinthe acceptable limitsand scatteredaround

the actual values of response variable. In other words, the majority of

the results obtained are matching with few instances of disagreement.

Therefore, the validation test results are satisfactory.

8. Deterioration curve building

A sensitivity analysis is executed in order to test the developed

models' sensitivity to changes in their inputs. For each model, the

input variables were varied one at a time to determine their respective

impact on the pipeline condition. For this purpose, the values of the

studied factor were varied whereas the values of the other factors

were kept constant at their average values. The data was presented to

the developed models to predict new outputs. The pipeline conditions

were obtained using the prediction models. Since there were no varia-

tions in any factor other than the one studied, the variations observed

in the predicted pipeline conditions were under the exclusive inuence

of thestudied factor. In other words, the effects of all other factors were

discounted and the impacts of the studied factor upon the pipeline con-dition were determined.

It is also important to express graphically the impact of each fac-

tor on the pipeline condition. As a result, deterioration curves were

developed for each model as shown in Figs. 6 to 10. These deteriora-

tion curves give a clearer understanding of the interrelationships of


COATING CONDITION

CATHODIC PROTECTION

SUPPORT CONDITION

JOINT CONDITION

6

7

8

9

PIPE

CONDITION

NORMALIZED VALUE

20-INCH OIL PIPES DETERIORATION CURVES - DIRECTLY PROPORTIONAL FACTORS

AGE (Years)

CROSSING (#)

FREE SPANS (m)

AGE

METAL LOSS

CROSSINGS

OPERATING PRESSURE

ANODE WASTAGE

FREE SPANS

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

500 580 660 740 820 900 980 1060 1140 1220 1300

6.52 6.72 6.93 7.13 7.33 7.52 7.72 7.91 8.10 8.28 8.47

6.42 6.82 7.22 7.60 7.96 8.29 8.59 8.83 9.03 9.19 9.29

7.66 7.77 7.87 7.96 8.04 8.11 8.18 8.24 8.29 8.34 8.39

8.15 8.17 8.18 8.20 8.22 8.24 8.27 8.30 8.33 8.36 8.39

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

0 7 14 21 28 35 42 49 56 63 70

0 2 4 6 8 10 12 14 16 18 20

0 7 14 21 28 35 42 49 56 63 70

8.52 8.49 8.45 8.42 8.38 8.35 8.31 8.27 8.24 8.20 8.17

8.41 8.37 8.26 8.08 7.84 7.56 7.25 6.95 6.67 6.45 6.32

8.36 8.35 8.33 8.32 8.30 8.29 8.27 8.26 8.24 8.23 8.21

8.45 8.45 8.45 8.44 8.44 8.43 8.42 8.40 8.39 8.37 8.35

8.42 8.37 8.31 8.26 8.20 8.15 8.09 8.04 7.98 7.92 7.87

8.46 8.21 8.00 7.83 7.70 7.59 7.50 7.43 7.38 7.35 7.32

6

7

8

9

PIPE

CONDITION

NORMALIZED VALUE

20-INCH OIL PIPES DETERIORATION CURVES - INVERSELY PROPORTIONAL FACTORS





Fig. 7.20-Inch oil pipes deterioration curves.


7/24/2019 1-s2.0-S0926580514001149-main

12/16

pipeline future condition and the studied factors. As shown in the

mentionedgures, each deterioration curve set was separated into

two groups of factors, namely, directly proportional and inversely

proportional.

Therst groupshows only thefactors that have a positive impact on

the pipeline condition. The factors in this group are the diameter, the

coating condition, the cathodic protection, the support condition, and

the joint condition. Smaller pipeline diameters have usually higher

probability of failure than those with larger ones[39].This is possibly

because smaller standard dimension ratios (SDR) affect the structural

performance of pipelines and make them more vulnerable to external

impact and third party damage. In addition, smaller diameter pipelines,

which have thinner wall thicknesses, allow faster corrosion rate. Good

and well maintained coating and cathodic protection enhances the

pipeline condition by protecting it against corrosion. Finally, well main-

tained supports and joints condition achieves a better structuralperfor-

mance of the pipeline.

On the other hand, the second group shows only the factors that

have a negative impact on the pipeline condition. The factors in this

group are the age, the metal loss, the anode wastage, the operating

pressure, thecrossings, and thefree spans. Theimpact of ageon pipeline

condition is negative in nature. An increase of the metal loss depth as a

percentage of its pipeline wall thickness due to corrosion or mechanical

damage degrades pipeline condition. As galvanic anodes waste, their

sizes decrease; resulting in a decrease in their well-known sacricial

action to protect the pipeline from corroding. Maximum allowable

operating pressure (MAOP) decreases the pipeline condition when it

gets close to the design pressure as it can induces more stresses on the


COATING CONDITION

CATHODIC PROTECTION

SUPPORT CONDITION

JOINT CONDITION

6

7

8

9

PIPE

CONDITION

NORMALIZED VALUE

24-INCH GAS PIPES DETERIORATION CURVES - DIRECTLY PROPORTIONAL FACTORS

AGE (Years)

CROSSING (#)

FREE SPANS (m)

AGE

METAL LOSS

CROSSINGS

OPERATING PRESSURE

ANODE WASTAGE

FREE SPANS

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

500 580 660 740 820 900 980 1060 1140 1220 1300

6.46 6.66 6.85 7.04 7.22 7.39 7.56 7.73 7.88 8.04 8.18

6.29 6.54 6.84 7.16 7.48 7.80 8.08 8.33 8.53 8.66 8.73

7.82 7.87 7.91 7.95 7.98 8.01 8.04 8.07 8.09 8.11 8.13

7.63 7.67 7.72 7.77 7.82 7.88 7.94 8.00 8.06 8.12 8.19

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

0 7 14 21 28 35 42 49 56 63 70

0 2 4 6 8 10 12 14 16 18 20

0 7 14 21 28 35 42 49 56 63 70

8.31 8.25 8.19 8.13 8.07 8.01 7.95 7.89 7.83 7.77 7.71

8.91 8.66 8.41 8.16 7.91 7.66 7.42 7.17 6.92 6.67 6.42

8.13 8.10 8.07 8.04 8.02 7.99 7.96 7.94 7.91 7.88 7.86

8.86 8.69 8.53 8.37 8.22 8.07 7.92 7.77 7.62 7.46 7.30

8.21 8.14 8.07 8.02 7.96 7.91 7.87 7.83 7.79 7.76 7.74

8.21 8.10 7.99 7.88 7.77 7.66 7.56 7.45 7.34 7.23 7.12

6

7

8

9

PIPE

CONDITION

NORMALIZED VALUE

24-INCH GAS PIPES DETERIORATION CURVES - INVERSELY PROPORTIONAL FACTORS





Fig. 8.24-inch gas pipes deterioration curves.


7/24/2019 1-s2.0-S0926580514001149-main

13/16

pipeline. Finally, as the number of other crossing and free spans

increase, the pipelines become less stable.

As shown inFigs. 6 to 10, the vertical axis represents the predicted

pipeline conditions while the horizontal axis the impact factor. Since all

the factors do not have the same units, the horizontal axis of the

deterioration curves are plotted using a normalized scale from 0 to 1 as

explained inTable 2.However, for better visualization to the reader,

the real values of the factors (age, diameter, crossings, cathodic protec-

tion, and free spans) are listed below their corresponding normalized

values as shown inFigs. 6 to 10.The real values of the remaining factors

(metal loss, coating condition, operating pressure, anode wastage, sup-

port condition, and joint condition) are not listed since they are the

same as the normalized values. In addition, the corresponding predicted

pipelinecondition values to eachfactor's different values are tabulatedas

shown in each deterioration curve. For example, let us assume that the

effect of the coating condition factor on the pipeline condition is

studied for the 20-Inch Oil pipelines. Let us also assume that the target

coating conditionpercentages are 20 and 60%.Therefore, simply by using

Fig. 7, the user will select the 0.2 and 0.6 normalized values on the hori-

zontal axis and read the corresponding predicted pipeline condition for

each case using the underneath table for the coating condition factor.

Alternatively, the user can just hit the coating conditioncurve for each

case and accordingly read the predicted pipeline condition from the

vertical axis. Based on that, the predicted pipeline conditions for coating

condition of 20 and 60% will be 6.93 and 7.72, respectively. As a reminder,

these results are based on the average values of factors other than the

coating condition. The same procedure can be followed to check or

predict the individual impact of other factors on the pipeline condition.

DIAMETER (Inches)


DIAMETER

COATING CONDITION

CATHODIC PROTECTION

SUPPORT CONDITION

JOINT CONDITION

5.5

6.5

7.5

8.5

PIPE

CONDITION

NORMALIZED VALUE

GAS PIPES DETERIORATION CURVES - DIRECTLY PROPORTIONAL FACTORS

AGE (Years)

CROSSING (#)

FREE SPANS (m)

AGE

METAL LOSS

CROSSINGS

OPERATING PRESSURE

ANODE WASTAGE

FREE SPANS

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

2 7 12 17 22 27 32 37 42 47 52

500 580 660 740 820 900 980 1060 1140 1220 1300

7.84 7.88 7.92 7.96 8.00 8.05 8.09 8.13 8.17 8.21 8.26

6.32 6.56 6.78 6.99 7.18 7.36 7.53 7.68 7.81 7.93 8.04

5.95 6.25 6.61 6.98 7.35 7.68 7.96 8.18 8.31 8.37 8.40

7.56 7.59 7.62 7.66 7.70 7.75 7.80 7.85 7.92 7.98 8.05

7.36 7.46 7.54 7.62 7.70 7.77 7.83 7.89 7.95 7.99 8.03

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

0 7 14 21 28 35 42 49 56 63 70

0 2 4 6 8 10 12 14 16 18 200 7 14 21 28 35 42 49 56 63 70

8.18 8.12 8.05 7.99 7.92 7.86 7.79 7.73 7.67 7.60 7.54

8.36 8.25 8.09 7.87 7.61 7.31 6.99 6.68 6.39 6.14 5.98

8.01 7.97 7.92 7.88 7.84 7.80 7.75 7.71 7.67 7.63 7.58

9.31 9.10 8.90 8.70 8.51 8.33 8.15 7.97 7.80 7.63 7.47

8.07 8.02 7.96 7.90 7.85 7.79 7.73 7.68 7.62 7.57 7.51

8.58 8.51 8.43 8.34 8.25 8.15 8.05 7.93 7.82 7.69 7.56

5.5

6.5

7.5

8.5

9.5

PIPE

CONDITION

NORMALIZED VALUE

GAS PIPES DETERIORATION CURVES - INVERSELY PROPORTIONAL FACTORS





Fig. 9.Gas pipes deterioration curves.


7/24/2019 1-s2.0-S0926580514001149-main

14/16

Besides examining the impact of changing each factor on the pipe-line condition, additional deterioration curves were developedfor oil

and gas pipelinesto show the effect of changing all the factors togeth-

er. These curves were developed by changing simultaneously all the

factors normalized values starting from their best possible to their

worst effects on the pipeline condition as shown in Fig. 11. For more

clarication, the curves were built by rst increasing the age factor by

oneyear starting from zero. Simultaneously, themetal loss, coating con-

dition, cathodic protection, crossings, anode wastage, free spans, joint

and support condition factors were changed proportionally according

to their rate of increase or decrease based on the received historical

inspection data. However, both the factors of diameter and operating

pressure were kept constant. The effect of such periodical change in

the factors' values was determined by calculating the pipeline condition

each year using the developed model. It should be noted that the

deterioration proles shown inFig. 11does not take into account anykind of rehabilitation action that may take place during the pipeline op-

eration. Also, due to the lack of inspections data received, the accuracy

of the deterioration proles can be enhanced throughout receiving

more additional inspections data.

According to the developed deterioration curves, it can be observed

that the cathodic protectionand coating conditionfactors have the

highest positive impacts on the pipeline condition while the diameter

factorhas thelowest positive impact. On theotherhand,the metal loss

and crossings factors have the highest and lowest negative impacts on

the pipeline condition, respectively. This indicates that the corrosion re-

lated factors (cathodic protection, metal loss, and coating condition) are

the most important contributors to the pipeline condition prediction.

However, other factors including age, support condition, joint condition,

anode wastage, and free spans, are still essential for pipeline condition

DIAMETER (Inches)


DIAMETER

COATING CONDITION

CATHODIC PROTECTION

SUPPORT CONDITION

JOINT CONDITION

6

7

8

9

PIPE

CONDITION

NORMALIZED VALUE

OIL PIPES DETERIORATION CURVES - DIRECTLY PROPORTIONAL FACTORS

AGE (Years)

CROSSING (#)

FREE SPANS (m)

AGE

METAL LOSS

CROSSINGS

OPERATING PRESSURE

ANODE WASTAGE

FREE SPANS

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

2 7 12 17 22 27 32 37 42 47 52

500 580 660 740 820 900 980 1060 1140 1220 1300

8.50 8.54 8.58 8.62 8.66 8.70 8.74 8.78 8.82 8.86 8.90

6.66 6.90 7.14 7.36 7.58 7.79 7.99 8.19 8.37 8.56 8.73

6.21 6.31 6.65 7.13 7.66 8.17 8.61 8.92 9.08 9.16 9.18

7.94 8.02 8.09 8.16 8.23 8.29 8.36 8.42 8.47 8.53 8.58

7.92 8.01 8.10 8.18 8.26 8.33 8.39 8.44 8.49 8.54 8.57

0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0

0 7 14 21 28 35 42 49 56 63 70

0 2 4 6 8 10 12 14 16 18 20

0 7 14 21 28 35 42 49 56 63 70

8.86 8.79 8.72 8.64 8.57 8.50 8.42 8.35 8.27 8.20 8.12

9.21 9.02 8.78 8.50 8.19 7.86 7.51 7.15 6.82 6.51 6.25

8.66 8.63 8.59 8.56 8.54 8.51 8.49 8.47 8.45 8.44 8.43

8.83 8.80 8.77 8.75 8.72 8.69 8.67 8.64 8.61 8.59 8.56

8.59 8.55 8.50 8.44 8.35 8.25 8.13 7.99 7.84 7.67 7.48

8.02 7.97 7.93 7.89 7.85 7.80 7.76 7.72 7.67 7.63 7.59

6

7

8

9

PIPE

CONDITION

NORMALIZED VALUE

OIL PIPES DETERIORATION CURVES - INVERSELY PROPORTIONAL FACTORS





Fig. 10.Oil pipes deterioration curves.


7/24/2019 1-s2.0-S0926580514001149-main

15/16

prediction as they together represent around 40% of their contribution

strength. Also, operating pressure has a signicantly higher effect on

gaspipeline condition than on oilpipelines.

9. Conclusions

A studyon the condition prediction of oil and gaspipelines was con-

ducted. Twenty quantitative and qualitative factors having a major im-

pact on thecondition of oil and gas pipeline were identiedof which 11

were used in this study. The ANN technique was used to develop ve

condition prediction models with respect to pipeline size and type of

transported product. The developed models are based on historical

inspection data collected from three different offshore pipelines in

Qatar. The collected data were prepared, organized, and subsequently

used to train, test, andvalidate themodels. Thecoefcientof determina-

tion (R2) results for the developed models were between 99.04 and

99.59%. In addition, all other necessary statistical diagnosis have been

checked showing sound results for the developed models. The models

have been validated and the results showed their robustness with anaverage validity percentage between 97.40 and 98.40%. The developed

ANN models were compared with previously developed regression

models that serve the same purpose of this study. It was found that the

performance of both techniques were close to each other, however, the

ANN technique provided better results. Finally, a sensitivity analysis

was carried out in order to build deterioration curves for the models de-

veloped. The deterioration curves showthe degree of the individual effect

of each factor on the pipeline condition. Cathodic protection and coating

condition were found to have the highest positive effect on the pipeline

condition while metal loss have the highest negative effect. On the

other hand, diameter and crossings were found to have the lowest posi-

tive and negative effects on the pipeline condition, respectively. The

models are expected to help pipeline operators to assess and predict

the condition of existing oil and gas pipelines and hence prioritize theirinspections and rehabilitation planning.

Acknowledgments

The authors gratefully acknowledge the support provided by

Qatar National Research Fund (QNRF) for this research project under

award No. QNRF-NPRP 09-901-2-343. The authors would also like to

acknowledge the operational pipeline engineers of Qatar Petroleum

for their invaluable suggestions and recommendations.

References

[1] M. Ahammed, Probabilistic estimation of remaining life of a pipeline in thepresenceof active corrosion defects, Int. J. Press. Vessel. Pip. 75 (4) (1998) 321329.

[2] H. Al-Barqawi, T. Zayed, Condition rating model for underground infrastructuresustainable water mains, ASCE J. Perform. Constr. Facil. 20 (2) (2006) 126135.

[3] S. Belsi to, P. Lombardi, P. Andreussi, S. Banerjee, Leak detection in liqueed gaspipelines by articial neural networks, Am. Inst. Chem. Eng. J. 44 (12) (1998)26752688.

[4] C. Bersani, L. Citro, R.V. Gagliardi, R. Sacile, A.M. Tomasoni, Accident occurenceevaluation in the pipeline transport dangerous goods, Chem. Eng. Trans. 19(2010) 249254.

[5] M. Bertolini, N. Bevilacqua, Methodology and theory oil pipeline cause analysis aclassication tree approach, J. Qual. Maint. Eng. 12 (2) (2006) 186198.

[6] R. Caruana, A. Niculescu-Mizil, An empirical comparison of supervised learning

algorithms, Proceedings of the 23rd international conference on Machine Learning,ACM, Pittsburgh, PA, USA, 2006.[7] A.A. Carvalho, J.M.A. Rebello, L.V.S. Sagrilo, C.S. Camerini, I.V.J. Miranda, MFL signals

and articial neural networks applied to detection and classication of pipe welddefects, NDT E Int. 39 (8) (2006) 661667.

[8] J. Cheng, A.N. Tegge, P. Baldi, Machine learning methods for protein structureprediction, IEEE Rev. Biomed. Eng. 1 (2008) 4149.

[9] P.M. Davis, J. Dubois, F. Gambardella, F. Uhlig, Performance of European Cross-country Oil Pipelines: Statistical Summary of Reported Spillages in 2008 and Since1971, CONCAWE, Brussels, June 2010.

[10] A.W. Dawotola, P.H.A.J.M. VanGelder, J.K. Vrijling, Decision analysis framework forrisk management of crude oil pipeline system, Adv. Decis. Sci. 2011 (2011).

[11] A.W. Dawotola, P.H.A.J.M. VanGelder, J.K. Vrijling, Risk Assessment of PetroleumPipelines using a combined Analytical Hierarchy ProcessFault Tree Analysis(AHP-FTA), Proceedings of the 7th International Probabilistic Workshop, Delft,Netherlands, 2009.

[12] P.K. Dey, A risk based model for inspection and maintenance of cross countrypetroleum pipelines, J. Qual. Maint. Eng. 7 (1) (2001) 2543.

[13] I. Dikmen, M. Birgonul, S. Kiziltas, Prediction of organizational effectiveness in

construction companies, ASCE J. Constr. Eng. Manag. 131 (2) (2005) 252261.[14] M.S. El-Abbasy, A. Senouci, T. Zayed, F. Mosleh, A condition assessment model for oil

and gas pipelines using integrated simulation and analytic network process, Journalof Structure andInfrastructureEngineeringTaylor andFrancis Group, 2014, (in-press).

[15] M.S. El-Abbasy, A. Senouci, T. Zayed, F. Mirahadi, L. Parvizedghy, Condition predic-tion models for oil and gas pipelines using regression analysis, ASCE J. Constr. Eng.Manag. 140 (6) (2014).

[16] O. Folorunsho, Comparative study of different data mining techniques performancein knowledge discovery from medical database, Int. J. Adv. Res. Comput. Sci. Softw.Eng. 3 (3) (2013) 1115.

[17] R. Goodland, Oil and Gas Pipelines Social and Environmental Impact Assessment:State of the Art, IAIA 2005 Conference, Fargo, ND, USA, 2005.

[18] J.M. Hallen, F. Caleyo, J.L. Gonzalez, Probabilistic Condition Assessment of CorrodingPipelines in Mexico, Proceedings of the 3rd Pan American Conference for Nonde-structive Testing (PANNDT), Rio de Janeiro, Brazil, 2003.

[19] L. Jinhai, Z. Huaguang, F. Jian, Y. Heng, A new fault detection and diagnosis methodfor oil pipeline based on rough set and neural network, Advances in NeuralNetworks, Lecture Notes in Computer Science (LNCS), vol. 4493, 2007, pp. 561569.

[20] M. Kalanaki, J. Soltani, Performance assessment among hybrid algorithms in tuningSVR parameters to predict pipefailure rates, Adv.Comput. Int. J. 2 (5)(2013)4046.

[21] J.L. Kennedy, Oil and Gas Pipeline Fundamentals, 2nd edition PennWell PublishingCompany, Oklahoma, 1993.

[22] S.X. Li, S.R. Yu, H.L. Zeng, J.H. Li, R. Liang, Predicting corrosion remaining life ofunderground pipelines with a mechanically-based probabilistic model, J. Pet. Sci.Eng. 65 (34) (2009) 162166.

[23] K. Liao, Q. Yao, X. Wu,W. Jia,A numerical corrosionrate predictionmethod fordirectassessmentof wetgas gathering pipelines internal corrosion, Energies 5 (10) (2012)38923907.

[24] R.S. Mohana, P. Thangaraj, Machine learning approaches in improving service levelagreement-based admission control for a software-as-a-service provider in cloud,

J. Comput. Sci. 9 (10) (2013) 12831294.[25] K. Mollazade, M. Omid, A. Are, Comparing data mining classiers for grading

raisins based on visual features, Comput. Electron. Agric. 84 (2012) (2012)124131.

[26] N.M. Noor, N.A.N. Ozman, N. Yahaya, Deterministic prediction of corroding pipelineremaining strength in marine environment using DNV RPF101 (Part A), J. Sustain.Sci. Manag. 6 (1) (2011) 6978.

[27] N.M. Noor, N. Yahaya, N.A.N. Ozman, S.R. Othman, The forecasting residual life ofcorroding pipeline based on semi-probabilistic method, UNIMAS E J. Civ. Eng. 1(2010) 246263.

[28] X.Y. Peng, P. Zhang,L.Q. Chen, Long-DistanceOil/Gas Pipeline FailureRate PredictionBased On Fuzzy Neural Network Model, CSIE 2009 Proceedings, World Congress onComputer Science and Information Engineering, vol. 5, Wilshire Grand, Los Angeles,LA, USA, 2009.

[29] M .D. Prabhakar, Prediction of software effort using articial neural network andsupport vector machine, Int. J. Adv. Res. Comput. Sci. Softw. Eng. 3 (3) (2013)4046.

[30] C.Y. Ren, W. Qiao, X. Tian, Natural gas pipeline corrosion rate prediction modelbased on bp neural network, Adv. Intell. Soft Comput. 147 (2012) 449455.

[31] A. Senouci, M.S. El-Abbasy, E. Elwakil, B. Abdrabou, T. Zayed, A model for predictingfailure of oil pipelines, Journal of Structure and Infrastructure Engineering, vol. 10,No. 3, Taylor and Francis Group, 2014, pp. 375387.

[32] A.Senouci, M.S.El-Abbasy,T. Zayed, A fuzzy-based model forpredicting failure of oilpipelines, ASCE J. Infrastruct. Syst. (2014) (in-press).

[33] M. Singh, T. Markeset, A methodology for risk-based inspection planning of oil andgas pipes based on fuzzy logic framework, Eng. Fail. Anal. 16 (7) (2009) 20982113.

0

1

23

4

5

6

7

8

9

10

0 10 20 30 40 50 60 70

PIE

CONDITION

AGE (Years)

GAS & OIL PIPELINES DETERIORATION PROFILES

GAS PIPES OIL PIPES

Fig. 11.Deterioration proles.

http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0005http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0005http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0005http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0005http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0010http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0010http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0010http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0010http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0015http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0015http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0015http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0015http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0015http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0015http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0015http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0015http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0015http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0020http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0020http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0020http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0020http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0020http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0025http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0025http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0025http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0025http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0025http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0025http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0200http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0200http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0200http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0035http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0035http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0035http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0035http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0035http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0035http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0035http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0035http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0035http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0040http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0040http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0040http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0040http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0045http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0045http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0045http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0050http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0050http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0205http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0205http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0205http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0205http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0205http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0205http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0060http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0060http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0060http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0060http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0065http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0065http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0065http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0065http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0210http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0210http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0210http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0215http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0215http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0215http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0080http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0080http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0080http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0080http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0080http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0220http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0220http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0225http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0225http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0225http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0230http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0230http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0230http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0230http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0230http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0100http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0100http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0100http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0100http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0105http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0105http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0110http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0110http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0110http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0110http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0110http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0110http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0110http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0115http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0115http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0115http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0115http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0115http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0120http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0120http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0120http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0120http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0120http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0130http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0130http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0130http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0130http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0130http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0130http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0130http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0135http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0135http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0135http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0135http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0135http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0235http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0235http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0235http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0235http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0145http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0145http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0145http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0145http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0145http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0145http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0145http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0150http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0150http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0150http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0150http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0240http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0240http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0240http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0240http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0240http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0245http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0245http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0165http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0165http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0165http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0165http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0165http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0165http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0245http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0245http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0240http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0240http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0240http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0150http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0150http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0145http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0145http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0145http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0235http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0235http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0235http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0235http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0135http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0135http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0135http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0130http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0130http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0130http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0125http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0120http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0120http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0120http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0115http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0115http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0115http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0110http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0110http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0110http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0105http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0105http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0100http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0100http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0230http://refhub.elsevier.com/S0926-5805(14)00114-9/rf0230http://refhub.elsevier.com/S0926-5805(14)0

Documents

1-s2.0-S0926580514001149-main