An Investigation into the Use of Artificial Neural Network to Monitor and Detect

Leakage Point(s) along a select Pipeline.

P.A. Adewuyi, B.Tech.1* and M.O. Okelola, M.Sc.2

1
Department of Mechatronics Engineering, Bells University of Technology, Ota, Nigeria.
2
Department of Electronic and Electrical Engineering, Ladoke Akintola University, Ogbomoso, Nigeria.

*E-mail: [email protected]

ABSTRACT INTRODUCTION

The problem of incessant leakages in oil pipeline The most popular way of transporting petroleum
from intentional or natural accident with its untold products is by the use of pipelines. The use of
results of pollution, loss of lives from fire pipelines eases the movement of petroleum
outbreak, and reduction in production output, products from one location to the other. The
spurred the burning desire to proffer a novel distances covered by these pipelines are in
solution with the use of artificial neural network as thousands of miles passing through cities and
a tool to detect oil leakage, pin-point the leakage villages. Some pipelines are laid overhead when
spot and provide useful communication link with passing through frost prone areas. There is no
the supervisory control and data acquisition denying the fact that pipelines have become
(SCADA). indispensable occupants of especially dedicated
land areas in Nigeria.
The data used for this work was obtained from
Atlas Cove, Lagos, Nigeria under the NNPC- However, as beneficial as petroleum products are
PPMC Mosinmi Area Depots. Pressure of oil to us as man, damages often occur to pipeline
flowing in pipelines between Atlas Cove Depot, networks as a result of factors ranging from lack
Lagos and Mosinmi Area Depot in Ogun state, of proper maintenance of the network, ageing of
Nigeria were used for the training of artificial the pipeline networks, pipeline vandalization, and
neural network. accidents. Once there is any damage to the
pipeline(s) by any form, leakages often occur
The data was randomly picked and divided into which usually cause harm to both humans,
training data (50 constants), validation data (50 animals, and the environment.
constants), and testing data (50 constants) as
contained in the neural network training algorithm. In October 17, 1998 – a pipeline explosion
The best network architecture that is, a four layer occurred at Jesse in the Niger Delta region of
neural network (20-20-15-10) obtained from trials Nigeria, killing about 1,200 villagers, some of
for this model was used. The use of graphical whom were scavenging gasoline, so far the worst
®
user interface, a tool in MATLAB , makes the in the history of pipeline vandalization in Nigeria
analysis easy and interactive. [1].

Artificial neural network (ANN) approach to In May 12, 2006 – an oil pipeline vandalization or
solving this identified oil leakage problem gives rupture occurred outside Lagos, Nigeria killing
satisfactory results as the error between the ANN around 200 people. Year 2000 is the worst hit
output and the target is very tolerable being where three different pipelines vandalization led
0.000566069 with a goal of 0.01. This could be to explosion that killed hundreds of people [7].
implemented physically as an artificial intelligence Ijegun pipeline explosion occurred in May 16,
unit in the monitoring of oil pipeline networks. 2008.

(Keywords: pipeline leaks, pipeline right of way, oil flow The loss of human lives, damages to the natural
pressure, artificial neural network, pipeline model) habitat of plants and animals due to pollution,
developing a cheap model for pipeline monitoring,
and inability of relevant authorities to respond fast

The Pacific Journal of Science and Technology –98–

http://www.akamaiuniversity.us/PJST.htm Volume 14. Number 1. May 2013 (Spring)
to cases of pipeline vandalization due to manual = upstream pressure
methods of monitoring pipelines leak points in = downstream elevation
Nigeria are the motivating factors responsible for = pipe segment length
this research work.

The rest of the paper contains sections such as; EXISTING TECHNOLOGIES TO PROTECT
description of the pipeline network, the existing PIPELINES
technologies to protect pipeline, artificial neural
network design, ANN layer explained, importance Some of the technologies involved in the
of Levenberg-Marquardt algorithm, results and protection of pipelines took the form of manual
discussion, conclusion, and recommendation. monitoring (that is walking the pipeline), satellite
monitoring, and the use of security cameras.
Modern technology that is commonly being used
DESCRIPTION OF THE PIPELINE NETWORK to protect pipelines is the one that uses
computers [8]. This computer arrangement
The pipeline network segment under accepts information from the field related to
consideration is that which links Atlas Cove depot pressures, flow rates, and temperatures to
in Lagos state, Nigeria to Mosinmi depot in Ogun estimate the hydraulic behavior of the product
state, Nigeria under the NNPC/PPMC Mosinmi being transported. Estimation is done, results
Area Depots. The pipeline covers a distance of compared to other field references to detect the
45.7km. The method presently being used in presence of anomaly or unexpected situation,
monitoring of this pipeline is manual as pipeline which may be related to a leak.
guards still patrol the network once there is a
pressure drop in the flow of oil along the pipeline.
The exact location of leak(s) cannot be ARTIFICIAL NEURAL NETWORK DESIGN
ascertained because there are no installations of
modern monitoring equipment along the pipeline. Artificial neural network is a network developed to
Security agents detailed to the monitoring of the emulate biological neural networks. ANN is
pipeline usually rely on information from the extremely simple abstractions of biological
surrounding communities as regards the state of neurons, realized as elements in a program or
the pipeline. perhaps as circuits made of silicon.

Millions of dollars’ worth of petroleum products Networks of these artificial neurons do not have a
are being lost on a daily basis resulting chiefly fraction of the power of human brain, but they can
from the activities of vandals who are so be trained to perform useful functions [8].
sophisticated in their operations in order to cover Artificial neural network is a brain mesh network.
their tracks.
Neurons are believed to be the basic units used
for computation in the brain. In neural networks,
FLOW EQUATION connections are important. Each neuron in the
brain has the order of hundreds or thousands of
The equation for a typical steady state isothermal connections to other neurons, making the total
14 15
liquid flow in a pipeline is [11]: number around 10 or 10 . This is fewer than the
number we would get if every neuron connected
is considered in relation to the size of the brain, a
(1) neuron connects directly to only a small fraction
of other neurons [2, 9].

where, = flow rate (turbulent region)

= friction factor ANN layer explained
= pipe internal diameter
= inlet or upstream pressure A four-layer neural network was used for this
= downstream or outlet pressure work. After the input, the hidden layer consists of
= Constant three layers. In the first layer, the transfer
= specific gravity of fluid function was ‘tansig’. The second layer transfer
function was also ‘tansig’. In the third layer

The Pacific Journal of Science and Technology –99–

http://www.akamaiuniversity.us/PJST.htm Volume 14. Number 1. May 2013 (Spring)
‘purelin’ function was used and in the fourth layer, BACKPROPAGATION ALGORITHM
‘purelin’ was also used as the transfer function
[5]. Start
dsSt
Neuron(s) Used Initialize Input
vvvParameters
The neurons in the architecture were generated
from the algorithm for the network. 20 neurons
Define and format network input and
were applied in the first hidden layer.
target data.
20 neurons were also used in the second hidden
layer. The third hidden layer consists of 15
neurons and the fourth layer has 10 neurons in its Divide data into 3 sets: Training data,
layer. Testing data and Validation data
The type of training used was supervised learning
and data fed is termed training set. The method
used for adjusting the weights for this problem is Create feed-forward back-propagation
“back-propagation” [3]. 4-layer network

Data used were divided into; training data, test

data, and validation data. During training a
relationship is established between the actual Display the network and generate the
past results which serve as the input data and the simulated network
corresponding pipeline desired values that
represent the targets data. For all these data,
pressure in bar is considered throughout.
Train the network
The input pressure data ranges from 40.8bar to
42.84bar while the target pressure ranges from
37.74bar to 40.8bar. These values were
normalized to -1 and 1 for better neural network
handling. Is the training
performance
The network output is converted back to the met?
normal input unit using a reverse function in the No
back-propagation algorithm.
Yes

Simulate network

End

Figure 1: Artificial Neural Network Flowchart.

The Pacific Journal of Science and Technology –100–

http://www.akamaiuniversity.us/PJST.htm Volume 14. Number 1. May 2013 (Spring)
IMPORTANCE OF LEVENBERG-MARQUARDT REGRESSION CURVE
ALGORITHM TO THIS WORK
Outputs vs. Targets, R=0.99969
Since the performance index of this work’s neural 40.5
network training is the mean squared error, Data Points
Best Linear Fit
Levenberg-Marquardt algorithm which is a A=T
variation of Newton’s method that was designed 40.45

Outputs A, Linear Fit: A=(0.95)T+(2)

for minimizing functions that are sums of squares
of other nonlinear functions was used.
40.4
This algorithm appears to be the fastest method
for training moderate-sized feed-forward neural
networks as we have here. It also has a very 40.35
efficient MATLAB implementation, because it has
a built-in function for solution of matrix equation
[3, 4]. 40.3

RESULTS AND DISCUSSION

40.25
40.25 40.3 40.35 40.4 40.45 40.5
Targets T
2
Performance of Suction and Discharge Training
10
Training Figure 3: Regression Graph.
Validation
Test
1
10
The regression curve above shows that the
neural network output developed responded very
0
well with the various output-target pairs (Table 1)
10 during training. With 0.99969 regression value, it
Square Error

shows that the degree of association

(relatedness) and corresponding changes of
-1
10 output-target pair is 99.9%, which is excellent.
The sample of the output-target pair is shown in
Table 1.
-2
10

Table 1: Output-Target Pair for Pressure Values

-3
10 between 40.0000bar and 40.1895bar.
0 1 2 3 4
Epoch
Output-Target Pair
Figure 2: Training, Validation, Test Curves. ANN OUTPUT TARGET
PRESSURE(bar) PRESSURE(bar)
40.0001 40.0000
The iteration converges at the 3rd epoch 40.0132 40.0134
indicating the uniqueness of the Levenberg- 40.0273 40.0268
Marquardt training algorithm in terms of fastness. 40.0410 40.0402
The performance goal was met at epoch 3 with a 40.0549 40.0536
40.0691 40.0670
mean square error of 0.000566069 for a set goal
® 40.0824 40.0804
of 0.01 as displayed in Matlab R2007a version’s 40.0951 40.0938
window below: 40.1097 40.1072
40.1260 40.1206
TRAINLM-calcjx, Epoch 0/300, MSE 45.835/0.01, 40.1372 40.1340
Gradient 85.4153/1e-010 40.1474 40.1474
TRAINLM-calcjx, Epoch 3/300, MSE 40.1626 40.1608
0.000566069/0.01, Gradient 0.220226/1e-010 40.1773 40.1742
TRAINLM, Performance goal met. 40.1894 40.1876

The Pacific Journal of Science and Technology –101–

http://www.akamaiuniversity.us/PJST.htm Volume 14. Number 1. May 2013 (Spring)
LEAK LOCATION DEMONSTRATION (PMS) is transported along the pipeline right of
way, 8% of 40bar is 37bar.
Table 2: Leak Demonstration Data
Therefore, a pressure of 37bar is the critical
LEAK DEMONSTRATION DATA pressure point used to determine leaks. That is,
INPUT ANN DISTANCE any pressure that is less or equal to 37bar
PRESSURE PRESSURE (km) indicates leakage condition along the pipeline
(bar) (bar) transporting PMS for the present pipeline being
40.0001 40.0001 1 considered.
40.0132 40.0132 2
40.0273 40.0273 3 The location of leak along the pipeline as well as
34.0410 40.0410 4 the magnitude of the leak is displayed in the
40.0549 40.0549 5
figure above. The artificial neural network output
40.0691 40.0691 6
40.0824 40.0824 7
is compared with the pipeline pressure state each
36.0951 40.0951 8 time oil flows in the pipeline.
40.1097 40.1097 9
40.1260 40.1260 10 For a PMS (Premium Motor Spirit) flow between
35.1372 40.1372 11 Atlas Cove depot, Lagos and Mosinmi Depot in
40.1474 40.1474 12 Ogun state, Nigeria, the pipeline length is 45.7km.
40.1626 40.1626 13 This length is scaled to ratio 1:3.1 for better
40.1773 40.1773 14 ®
MATLAB handling of the data. In other words,
40.1894 40.1894 15 1km on the graph translates to 3.1km on site.
Now, once there is any pressure difference
between the ANN output and the current flow
pressure, the magnitude of the pressure as well
Leak Point Demonstration as the exact point of pressure deviation is
41 displayed as shown in Figure 4.
Normal Flow
leakage The pressure deviation which translates to leak
40
points for the sample above occur at 4km
(12.4km on site) with a magnitude of 34.0410bar ,
39 8km(24.8km on site) with a magnitude of
36.0951bar and 11km(34.1km on site) with a
ANN Output(bar)

38 magnitude of 35.1372bar, respectively.

37 CONCLUSION

36 Artificial Neural Network could be used to monitor

pipeline, interpret the data and provide an
important leak information as well as leak(s) exact
35 location(s). ANN is fast with very small mean
square error.
34
0 5 10 15
Distance(km) RECOMMENDATIONS
Figure 4: Leak Point and Magnitude Curve.
Artificial Neural Network should be incorporated
as an integral part of pipeline monitoring system
LEAK LOCATION DEMONSTRATION with Supervisory Control and Data Acquisition
EXPLAINED (SCADA). Further research work should consider
pipeline that is subject to constantly changing
A good leak detection technology should be able temperature, multiple pipeline networks and effect
to detect leak(s) of about 3 gallons of maximum of soil texture on leak detection.
flow within 15 minutes or less. Since 40bar is the
maximum pressure at which premium motor spirit

The Pacific Journal of Science and Technology –102–

http://www.akamaiuniversity.us/PJST.htm Volume 14. Number 1. May 2013 (Spring)
REFERENCES
ABOUT THE AUTHORS
1. Amaechi, R. 2010. “FG Lost $20bn in Five Years
to Pipeline Vandalism”. Retrieved July 14, 2011. Philip A. Adewuyi, obtained a B.Tech. in
http://www.thisdaylive.com>HOME. Electronic and Electrical Engineering from Ladoke
Akintola University of Technology Ogbomoso,
2. Aroh, K.N., I.U. Ubong, C.L. Eze, I.M. Harry, J.C. Nigeria in 2008. He is presently studying for an
Umo-Otong, and A.E. Gobo. 2010. “Impact on
Public Health and Negation to Attainment of
M.Sc. degree in the department of Electrical and
Millennium Development Goal: The Ishiagu Electronics Engineering, University of Lagos,
Example”. Disaster Prevention and Management. Nigeria. His research interests are in power
19(1):70-87. system management, application of neural
network to industrial operations, fuzzy logic and
3. Cavuto, D.J. 1997. “An Exploration and renewable energy sources.
Development of Current Artificial Neural Network
Theory and Applications with Emphasis on Muniru O. Okelola, obtained a B.Tech. degree in
Artificial Life”. Unpublished Master of Engineering Electronic and Electrical Engineering from Ladoke
Thesis.The Cooper Union, Albert Nerken School of
Engineering, New York, NY.
Akintola University of Technology Ogbomoso,
Nigeria and an M.Sc. degree in
4. Ennett, C.M. and Frize, M. 1998. “An Investigation Electrical/Electronic Engineering from University
into the Strengths and Limitations of Artificial of Lagos, Lagos, Nigeria. He is presently a
Neural Networks: An Application to An Adult ICU Lecturer in the Department of Electronic and
Patient Database”. Proc AMIA Symp. Canada. Electrical Engineering at Ladoke Akintola
University of Technology Ogbomoso, Nigeria. His
5. Hagan, M.T., H.B. Demuth, and M. Beale. 1996. research interests are in power system quality
Neural Network Design. PWS Publishing and analysis, power electronics, and artificial
Company: Chicago, IL.
neural network.
6. Kristiansen, M. 2005. “Model Behaviour: The use
of Model Based Pipeline Leak Detection”. Energy
Solutions International World Pipelines.
SUGGESTED CITATION
7. Lee, K., D. Kwak, and H. Lee-Kwang. 1996.
“Fuzzy Inference Neural Network for Fuzzy Model Adewuyi, P.A. and M.O. Okelola. 2012. “An
Tuning”. IEEE Transaction on System, Man, and Investigation into the use of Artificial Neural
Cybernetics – Part B: Cybernetics. 26(4). Network to Monitor and Detect Leakage Point(s)
along a select Pipeline”. Pacific Journal of
8. Naghedolfeizi, M. and S. Arora. 2003. “Artificial Science and Technology. 14(1):98-103.
Neural Network Model for Predicting Degradation
Trends in System Components and Sensors”.
AUTOTESTCON 2003. IEEE Systems Readiness Pacific Journal of Science and Technology
Technology Conference Proceedings. 647-651.
Sept. 22-25.

9. Stafford, M. and N. Wiliams. 1996. “Pipeline Leak

Detection Study”. Health and Safety Executive.
OTH 94 431: London, UK.

10. The Mathworks, Inc. 2007. “Backpropagation

(Neural Network Toolbox)”. http://www.mathworks.com

11. Mohitpour, M., H. Golshan, and A. Murray. 2007.

Pipeline Design & Construction: A Practical
Approach. ASME Press: New York, NY.

The Pacific Journal of Science and Technology –103–

http://www.akamaiuniversity.us/PJST.htm Volume 14. Number 1. May 2013 (Spring)