Recognition of Pipeline Safety Events Applied to Optical Fiber

Pre-warning System

Qian Sun, Hao Feng, Jian Li and Shijiu Jin

State Key Laboratory of Precision Measuring Technology and Instruments, Tianjin University, Tianjin, China

Keywords: Pipeline Safety, Recognition, Feature Extraction, Wavelet Energy Spectrum, Support Vector Machine.

Abstract: Recognition of pipeline safety events is a key problem in the research of the optical fiber pre-warning

system. In this paper a feature extraction method combined with wavelet energy spectrum (WES) and

wavelet information entropy (WIE) is proposed. In order to avoid kernel function being dominated by trivial

relevant or irrelevant features, a support vector machine (SVM) approach is also put forward based on the

feature weighting, i.e. Feature Weighted SVM (FWSVM). The experiment shows that the method proposed

in this paper is effective for recognition of the pipeline safety events and can be applied in optical fiber pre-

warning system.

1 INTRODUCTION

By virtue of a great many advantages, pipelines have

become the principal means of oil and gas

transportation. However, pipeline leakage takes

place due to some natural or artificial damages.

Leakage accidents may cause loss of life and

properties along with environmental pollution (Yang

et al., 2004).

Pre-warning technique based on distributed

optical fiber for the long oil and gas pipeline can

give an alarm when the pipeline is threatened;

therefore it is an important means to reduce the

economic expense and to ensure people's security.

Optical fibers are used to compose distributed

optical fiber vibration signal sensor based on Mach-

Zehnder interferometer principle (Zhou et al., 2007).

How to accurately recognize the types of pipeline

safety events is a key problem in the research of the

optical fiber pre-warning system. In (Qu et al.,

2006), a distributed optical fiber alarming system for

the safety of oil and gas pipeline has been

developed. Unfortunately, the system, with low

intelligence, cannot tell which kind of activity

causes the leaking accident, and the man on duty can

not take corresponding action at once. In order to

overcome the defect a recognition method based on

feature weighted support vector machine is purposed

in this paper.

The remainder of this paper is organized as

follows. Considering that the non-stationary and

random characteristics of the pipeline safety

detection signals, a feature extraction method

combined with WES and WIE is proposed in section

2. In section 3 a support vector machine approach

based on the feature weighting, i.e. Feature

Weighted SVM (FWSVM) is put forward in order to

avoid kernel function being dominated easily by

trivial relevant or irrelevant features. The experiment

in section 4 analyzes the performance of the method

proposed in this paper. Finally the conclusions and

future research are given in Section 5.

2 FEATURE EXTRACTION

2.1 Wavelet Energy Spectrum

According with energy mode, the result of wavelet

decomposition is called wavelet energy spectrum

(Qu et al., 2008). The energy of the signal at

different scales can be arranged as a feature vector.

In other words the characteristic bands of the signal

are extracted and can be used for classifier. The

original sequence

()

can be expressed as

() () () ()

jJ j

nDnAnDn









(1)

Sun Q., Feng H., Li J. and Jin S..

Recognition of Pipeline Safety Events Applied to Optical Fiber Pre-warning System .

DOI: 10.5220/0004277300730077

In Proceedings of the International Conference on Photonics, Optics and Laser Technology (PHOTOPTICS-2013), pages 73-77

ISBN: 978-989-8565-44-0

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

()

is the component of the signal

()

different scales.

The signal energy in a certain scale is the sum of

squares of wavelet coefficients in the same scale.

Definition is shown as

() , 1,2, ,

EDkj J









(2)

In the formula (2),

is the length of the sample

points,

(), 1,2, ,

Dkk N 

is the reconstructed

wavelet coefficient at scale

[, , , ]

EE E 

is the WES of signal

()

scales.

2.2 Wavelet Information Entropy

In the information theory, entropy is used to

represent the average information of the source

output. It can provide signal useful information of

potential and dynamic process. Its value is the

measure of signal average uncertainly and

complexity (EI-Zonkoly et al., 2011). The traditional

entropy can express the uncertainty of a signal in the

whole period, but cannot analyze the local nonlinear

characteristic of the non-stationary signal. Therefore

an time window is defined to calculate the WIE

value, and to observe the change of the WIE

following the sliding window. In this paper signal is

decomposed to

scales, and the discrete wavelet

coefficient at scale

()

. The length of the

short-time window is

, the sliding step is



, and

the calculation of the signal energy within a certain

time window at each scale is

()





(3)

The total energy within the time window is the sum

of the energy component at each scale.

tot j







(4)

The signal relative energy of each scale within the

time window is

tot



(5)

is the energy distribution of different scale. The

WIE of the signal within time window is defined as

log ( )

WT j j

Spp









(6)

Therefore, the change law of WIE along with

time window sliding can be obtained.

3 FEATURE WEIGHTED

SUPPORT VECTOR MACHINE

Support vector machine has excellent learning,

classification and generalization abilities which use

structural risk minimization instead of empirical risk

minimization. The basic idea of SVM is to transform

the input space into a high-dimensional feature space

through non-liner transformation, and optimal

separating hyperplane can be obtained in this feature

space (Kurek et al., 2010). In a training sample set

with classification mark

(, ), ,

ii i

yx R

{1, 1},





1, ,il





. The optimal separating

hyperplane can be created by the following

optimization problem,

1,1

max ( , )

.. 0,

0,1,,

iijijij

iij

aaayykxx

st a y

aCi l

















 











(7)

is the Lagrange multiplier of

. The

corresponding decision function is shown as

() ( ,)

ii i

xsi

kx x b















(8)

(, ) () ( )

ij i j

kx x x x







is called kernel function

which should be selected as integral operator of

feature space.

SVM makes the nonlinear separable problem

become linearly separable. According to the

functional theory, as long as a kernel function

satisfies mercer condition, it corresponds to a dot

product in the transform space. Different algorithm

formed by different kernel function. Common kernel

functions are shown as follows

PHOTOPTICS2013-InternationalConferenceonPhotonics,OpticsandLaserTechnology

(1) Polynomial kernel function

(, ) ( 1), 1,2,

ij i j

kx x x x d  

(9)

(2) Gauss radial basis kernel function

(, ) exp( )

ij i j

kx x x x





(10)

(3) Sigmoid kernel function

(, ) tanh(( ) ), 0, 0

ij i j

kx x bx x c b c

(11)

According to some criteria, the features of the data

set are given certain weights, this is called feature

weighting. The key issue of feature weighting is to

obtain the weight vector



. Calculation of the

weight vector is important for analyzing feature

correlation. The basic idea of feature correlation

analysis in classification learning process is to

calculate a certain metric in order to quantify the

correlation of feature and a given category. In this

paper class separability criterion is used to compute

weight vector. Calculating the divergence of the

one-dimensional

of each pair of classes, the

criterion value of each feature is

() min

ck d

Characteristic which has bigger criterion value is

well differentiable, in other words , that is a greater

contribution to the classification. Assuming that

each sample of data set is described by n features,

the vector

((1),(2), , ())Ccc cn 

describes the

weight of each feature. The weight vector



constructed by vector

The support vector machine based on feature

weighted kernel function is called feature weighted

support vector machine. The calculation of the

weighted kernel function can avoid kernel function

being dominated easily by trivial relevant or

irrelevant feature; therefore the better classification

results can be obtained (Zhang et al., 2009). The

weighted kernel function can be computed as

(, ) ( , )

Pi j i j

kxx kxPxP

(12)

k is defined as the kernel function of X×X, X

 R

P is the transformation matrix,

()Pdiag







the weight vector. Common weighted kernel

functions are shown as follows

(1) Feature weighted polynomial kernel function

(, ) ( 1)

TT d

pi j i j

kxx xPxP

(1)1,2,

TT d

xPPx d

(13)

(2) Feature weighted gauss radial basis kernel

function

( , ) exp( )

pi j i j

kxx xPxP



 

exp( (( ) ( )))

ij ij

xxPPxx



  

(14)

(3) Feature weighted sigmoid kernel function

( , ) tanh( ( ) )

pi j i j

kxx bxPxP c





tanh( ( ) ), 0, 0

bx PPx c b c





(15)

The construction steps of feature weighted

support vector machine shown as follows

Step1. Extract signal characteristics by WES and

WIE described above. The normalized eigenvector is

[, , , ]

TTT T





(16)

Step2. Calculate criteria value

()ck

of each

characteristic. The weight vector



and linear

transformation matrix

are constructed as

((1),,()),Cc cn



 

(17)

()p diag





(18)

Step3. Constitute the weighted kernel function

with matrix p by formula (13) to (15) to replace

standard SVM kernel function. In this paper

FWSVM is constructed by the sigmoid weighted

kernel function.

Step4. Recognize the pipeline safety events by

FWSVM, and evaluate the performance of the

classifier.

4 EXPERIMENTS

Three pipeline safety events including truck passing,

excavator and digging were experimented based on

optical fiber pre-warning system.

Optical fiber sensor was buried above oil

products pipeline. The distance between pipeline and

optical fiber is 500mm. The single mode optical

fiber and semiconductor laser source were used in

this system. The wavelength of laser source is

1550nm, and the power is 1mw. The optical

interference signal is converted into electric signal

by photodetector. Datas of these three events were

collected through data acquisition module. Figure 2

shows the normalized voltage valve of three cases.

RecognitionofPipelineSafetyEventsAppliedtoOpticalFiberPre-warningSystem

Figure 1: Optical fiber pre-warning system.

Figure 2: Signals of three cases.

Figure (a) is the signal generated by truck passing at

the surface of the soil within 50cm of the pipeline on

both sides. Figure (b) and figure (c) are the signal

generated respectively by excavator and digging at

the surface of soil just above the pipeline.

4.1 Signal Feature Extraction

In this paper, db6 wavelet function is used to

decompose signals for 7 layers, and the wavelet

energy spectrum of vibration signals caused by the

three safety events is obtained by formula (2). The

normalized energy of 8 frequency bands is shown in

figure 3. Different energy on different frequency

band provides a basis for the identification of the

signal.

Figure 3: WES of three cases.

Pipeline safety events first make incursions into

the soil near pipeline, and cause the vibration of soil.

The wavelet information entropy increases when the

optical fiber feels the vibration, and if the vibration

is more intense the entropy will be greater. In this

paper the length of time window

100L 

, sliding step





. The WIE is obtained by formula (6). Through

experimental analysis the maximum WIE of the

three pipeline safety events are significantly

different, thus the maximum WIE can be also used

as pattern recognition feature. The maximum WIE

of three conditions are shown as table 1.

Table 1: Maximum WIE of pipeline safety events.

Pipeline events Maximum WIE

Truck passing 0.1969

Digging 0.3112

Excavator 0.2230

4.2 Recognition of Pipeline Safety

Events by FWSVM

WES is used as characteristic of each safety event.

The energy of 8 frequency bands is marked as

feature 1 to 8. The maximum of WIE is marked as

feature 9. The three pipeline safety events truck

passing, digging and excavator are labeled

respectively as class 1, class 2 and class 3. FWSVM

based on weighted sigmoid kernel function is used

for recognizing the three events. 30 sets training

sample of each event are obtained from field

experiments for FWSVM classifier; besides,

randomly select 20 sets sample of each event to test

the classifier. Recognition results are shown in the

following table 2 and table 3. In table 2 the accuracy

is the percentage of the recognition correct number

of each event. In table 3 the accuracy is the

percentage of the total recognition correct number of

three safety events.

0 0.2 0.4 0.6 0.8 1 1. 2 1.4 1.6 1.8 2

x 10

-1

-0.5

0.5

Sampling Number

(a)

Signal of Truck Passing/(V)

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

x 10

-1

-0.5

0.5

Sampling Number

(b)

Signal of Excavator/(V)

0 0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2

x 10

-1

-0.5

0.5

Sampling Number

(c)

Signal of Digging/(V)

1 2 3 4 5 6 7 8

0.1

0.2

0.3

0.4

0.5

0.6

0.7

Frequency Band

Wavelet Energy Spectrum

Truck passing

Digging

Excavator

PHOTOPTICS2013-InternationalConferenceonPhotonics,OpticsandLaserTechnology

Table 2: Recognition accuracy of different pipeline safety

events.

Type of samples

Truck

passing

Digging Excavator

Number of

samples

20 20 20

Truck passing 20 0 0

Digging 0 18 1

Excavator 0 2 19

Accuracy 100% 90% 95%

Table 3: Recognition accuracy of different method.

Pattern recognition method Accuracy

WES-SVM 81.7%

WES-WIE-SVM 85%

WES-WIE-FWSVM 95%

As can be seen from table 2 and table 3, the

FWSVM proposed in this paper is very effective for

recognizing pipeline safety events. The recognition

accuracy can reach 95%. Traditional SVM cannot

satisfy the accuracy requirement because of its bad

stability. Therefore FWSVM can be applied to the

pattern recognition module of optical fiber pre-

warning system.

5 CONCLUSIONS

This paper studies the recognition of pipeline safety

events applied to optical fiber pre-warning system.

In order to solve the typical recognition problem of

pipeline safety events, the FWSVM is used in this

paper. Firstly, wavelet energy spectrum and wavelet

information entropy are used to extract features of

signals, then the FWSVM is used for recognizing the

three typical safety events. Through field

experiment, the results show that FWSVM has high

identification accuracy. The accuracy 95% is much

higher than traditional SVM. The calculation of

feature weighted kernel function can avoid kernel

function being dominated by trivial relevant or

irrelevant features. Therefore this method can satisfy

the requirement of optical fiber pre-warning system.

In the future work, recognition of more types of

pipeline safety events still need further research and

fieldexperiment.

REFERENCES

Yang, J., Wang, G. Z., 2004. Leak detection and location

methods for gas transport pipeline, Instruments in

Chemical Industry, Vol.31, No.2, pp.1-3.

Zhou, Y., Jin, S. J., Feng, H., 2007. Study on oil and gas

pipeline leakage real-time inspection system based on

distributed optical fiber, Conference Committee of the

International Symposium On Measurement

Technology and Intelligent Instruments, pp.507-510.

Qu, Z. G., Jin, S. J., Zhou, Y., 2006. Study on the

Distributed Optical Fiber Pre-warning System for the

Safety of Oil and Gas Pipeline, Piezoelectectrics

&Acoustooptics, Vol.28, No.6, pp.1-3.

Qu, W., Jia, X., Pei, S. B., Wu, J., 2008. Non-stationary

signal noise suppression based on wavelet analysis,

Congress on Image and Signal Processing, pp.303-

306.

EI-Zonkoly, A. M., Desouki, H., 2011. Wavelet entropy

based algorithm for fault detection and classification

in FACTS compensated transmission line,

International Journal of Electrical Power & Energy

Systems, Vol.33, No.8, pp.1368-1374.

Kurek, J., Osowski, S., 2010. Support vector machine for

fault diagnosis of the broken rotor bars of squirrel-

cage induction motor, Neral Computing &

Applications, Vol.19, No.4, pp.557-564.

Zhang, Y., Liu, X.D., Xie, F. D., Li, K. Q., 2009. Fault

classifier of rotating machinery based on weighted

support vector data description, Expert Systems with

Applications, Vol.36, No.4, pp.7928-7932.

RecognitionofPipelineSafetyEventsAppliedtoOpticalFiberPre-warningSystem