A Study on the Detection of Internal Leaks through Valves in Gas Pipeline Based on Support Vector Machine

chardfriendlyΤεχνίτη Νοημοσύνη και Ρομποτική

16 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

105 εμφανίσεις

A Study on the Detection of Internal Leaks through Valves in Gas Pipeline
Based on Support Vector Machine



Guoliang

C
a
o
1
,

Changhang Xu
2
, Guoming Chen
3

,
Huandi Shi
4

and Bixia Pan
5


1

Master;
College of Electromechanical Engineering, China Universi
ty of Petroleum
(East China); No.66, West Changjiang Road, Qingdao 266580; Email:

xiaoc2008@126.com;

18954220069

2

Associate
Professor
,

College of Electromechanical Engineering, China University
of Petroleum (East China); No.66, West Changjiang Road, Qingd
ao 266580; Tel

1
5866820823
; Email:
changhangxu
@126.co
m
,

Corresponding author

3 Professor
,

College of Electromechanical Engineering, China University of
Petroleum (East China); No.66, West Changjiang Road, Qingdao 266580; Tel

13954671082; Email:offshore
@126.
com

4 Master; College of Electromechanical Engineering, China University of Petroleum
(East China); No.66, West Changjiang Road, Qingdao 266580; Email:

hualishitou
@126.com;

18954237930

5

Master; College of Electromechanical Engineering, China University of

Petroleum
(East China); No.66, West Changjiang Road, Qingdao 266580; Email:

pbx922@hotmail.com
;

18958999210


ABSTRACT


As the traditional fluid control equipment, valves have wide applications in many
aspects, including the long distance pipeline, urban u
nderground pipeline and
chemical pipeline. However, internal leak
s

through valve is a common phenomenon
in engineering leading to medium Leaking, medium pollution, the failure of
emergency
shutdown

and other problems, which threats the safety of pipeline
s
eriously. On
-
line detection of internal leaks through valves, not only effectively
prevents accidents, but also decreases the cost of disassembling the valve and
pneumatic test blindly for conventional measurement methods. In this paper, several
common fai
lures of valves in gas pipeline to cause internal leaks through valves,
including wear, damage of seal, and their combinations, are simulated and acoustic
emission(AE) data are collected during the experiments. Then, acoustic emission
signals are analyzed
in frequency domain and time domain, and are decomposed into
a series of frequency bands based on wavelet packet. At last, a method, which
combines the feature of wavelet packet
-
energy entropy with support vector machine
(SVM), is put forward to identify t
he leak
s

failure modes. The results show that this
method can be effectively applied to the on
-
line detection of internal leaks through
valves in gas pipeline.


KEYWORDS


Internal leaks through valves; Gas Pipeline; Wavelet packet; Support vector machine


INTRODUCTION


Valves are indispensable fluid control equipment in many aspects, including the long
distance pipeline, urban underground pipeline and chemical pipeline. They play an
important role in ensuring the safety of workers, equipment and process. B
eing
affected by valve quality, operation and maintenance method,
and
rugged
environments, many problems arise
during

valve service time, such as seal failure,
erosion of the ball, foreign objective, etc. These problems directly lead to internal
leak
s

thro
ugh valve. Internal leak through valve is a common phenomenon in
engineering and leak
s

rate has become one of the most key indicators of valve quality.
Leak of a key valve affect
s

normal production operations and reduces the production
efficiency. What is
worse, it will cause fire and explosion accident which directly
threatens the security of stuff and equipment. Statistics show that there is serious
internal leak
s

problem in about 20% valves, and the direct economic losses caused
by
the

problem account fo
r about 5% of the total.

A convenient approach to on
-
line detecting of valve remains a longstanding
challenge in practical engineering. The existing detection methods primarily rely on
one's experience but less effective. On
-
line detection of internal leak
s through valves,
not only can effectively prevent accidents, but also decreases the cost of
disassembling the valve and pneumatic test blindly because of conventional
measurement methods.

As a dynamic nondestructive test method, acoustic emission technolo
gy is suitable
for online monitoring of the valve. Much research has been completed to investigate
and estimate leak
s

rate through acoustic emission. There have been many attempts at
predicting the noise emitted from control valves including Kaewwaewnoi mo
del and
E.Meland model. However, it is proved that modeling internal leak
s

in a valve with
unknown leak geometry is inherently difficult (Kaewwaewnoi et al., 2010 and
Meland et al., 2012). What’s more, people research on the problem of internal leak
s

throu
gh valve in viewpoint of the features of AE signal full waveform. The minimum
detectable leak
s

rate through a globe valve and the frequency spectra obtained are
also analyzed with different AE sensors. Broadband sensor is more suitable because
it reduced t
he need for noise reduction (Mohamed et al., 1998) .The AE signals
generated by leaks of a valve with artificial damage are studied and the spectrum of
frequency is analyzed. The results show that AE method is effective for the detection
of leaks (Meland e
t al., 2011). However, too little attention has been devoted to the
method of identifying failure mode in a valve. These previous results could be
preferable accurate and better convinced if they identify the failure mode in the first
phase. Further studie
s are still necessary to identify the leak
s

failure modes. An
identification method of the leak
s

failure mode, which combines the feature of
wavelet packet
-
energy entropy with support vector machine (SVM), is performed in
this study.


THEORIES


Wavelet pac
ket transform.
Wavelet packet transform (WPT) is a multi
-
resolution
analysis technique used to decompose the signals into different frequency segments.
It is a further decomposition based on wavelet transform. Application of the
transform to both the detai
l and the approximation coefficients results in an
expansion of the structure of the wavelet transform tree algorithm to the full binary
tree (Walczak et al., 1997). Fig. 1 shows the relation between 3rd level wavelet
transform and wavelet packet transform

of a time
-
domain signal.



Figure 1. 3rd level wavelet packet transform of a time
-
domain signal


A wavelet packet function can be expressed as
, where

indicates the
modulat
ion parameter, j indicates the scale parameter, and k indicates the translation
parameter (Sun et al., 2002).


The wavelets

meet the following double scale equations:



Where
and
are quadrature mirror filters associated with the predefined
scaling function and the mother wavelet function. A time
-
domain signal

, whose
maxim
um frequency is
, is decomposed into

resolution levels. The whole
frequency band is evenly divided into

subspaces on the
th level, including the
fre
quency range of

,

,

,

.

Wavelet packet node energy defined by Yen and Lin, can provide a more robust
signal feature for classification than using th
e wavelet packet coefficients directly

(Yen et al., 2000)
. The signal

is decomposed into several signals of different
frequency, marked
, the signal energy

can be defined as:


A feature vector can be constructed as
with the signal energy.


Support vector machine.

Support vector machine (SVM), which can solve practical
problems that relate to small sample, nonlinear, high dim
ension, local minimum
point, etc. has become the focus of international research in machine learning field.

Support vector machine (SVM) transforms the input space into high
-
dimensional
feature space by nonlinear transformation, and uses linear function

for
data fitting in the high dimensional feature space. According to the structural risk
minimization principle, the parameters in the linear function can be obtained by
minimizing the following objective function (Chen et al., 20
06 and Lu et al., 2012).


Where

indicates generalized constant. Cost function


indicates insensitive
loss function. The optimization problem can be converted into the followi
ng form of
constrained minimization, by means of introducing two set of non
-
negative slack
variable
,


This equation is converted into a dual form. The following Lagrange funct
ion can be
constructed according to objective function and constraint conditions.


The optimization objective function towards the dual problem is achieved by
calculating derivatives of

,

,

,

.


Maximizing the above equation to get:


According to the Mercer theorem, inner product kernel is designed as:


Support vector machine (SVM) regression model can be expressed as:


There are various kernel functions. Commonly used linear kernel functions include
polynomial kernel function and radial basis function (RBF), etc.


VALV
E LEAKS EXPERIMENT


Experimental rig
.
The experimental rig is composed of two parts. The first part of
the rig is the fluid circulation system controlled by the compressor, gasholder, pipes,
electromagnetic valve

and so on,
and it
can provide different lev
els of pressure (Fig.
2). Another part used specially to detect internal leaks through valves is a branch line
from the fluid circulation system. It contains the valve used in the experiments, pipes,
metal hose for connecting, control valve, piezometers, m
etering device, etc. (Fig. 3).
As the
shut
-
down valves
, ball valve is commonly used in the oil and gas industry.
Therefore, a 25mm ball valve is selected as research object.





Figure 2. The fluid circulation system


Fig
ure 3. Valve leak
s

test system


Valve failure mode
s
.
There are various failure modes leading to internal leaks
through valves. Common failure models include two aspects: Ball can't be adjusted
to full closed position and seal is damaged. Besides, failure c
aused by combination of
these two
factor
s is also very
probable
. In order to induce these failures,
valve is
adjusted to small opening (Fig. 4)
and
some damage is done to the valve seal (Fig. 5)
as common failure models.





Figure 4.
V
alve is adjusted to small opening

F
i
gure 5.
D
amage
to

valve s
eal


AE measurement system.

Two PCI
-

2 AE detection systems from PAC are used to
acquire data. The system includes AE acquisition software (AE
-
win), acquisition
card, preamplifier w
hich has a fixed gain o
f 40dB, and a wide
-
band AE sensor (WSα)
covered the range
of
0.1
-
1MHz. The sampling rate is 1MHz and AE sensor is
mounted on the surface of the valve body.


Experimental contents.

Three aspects of the experimental contents should be
addressed. Firstly, 27
groups of AE data are acquired when valve is adjusted to
various openings (15°, 20° and 25°) and inlet pressure varies from 0.1Mpa to 1Mpa.
Secondly, 11 groups of AE data are acquired when
the damaged conditon

to the seal
is changed and inlet pressure vari
es from 0.1Mpa to 1Mpa. Thirdly, 10 groups of AE
data are acquired when damage to the valve seal and valve adjusted to small opening
occur at the same time.


RESULT AND DISCUSSION


AE wave and frequency characteristic.



(a) Valve is adjusted to small op
ening


(b) Some damage is done to the valve seal


(c) Combination of two failure modes

Figure 6. AE wave and power spectrum of three different types of failure modes


FFT is used to convert the AE wave into the frequency domain. Fig. 6 shows AE
wave an
d power spectrum obtained from three different types of failure modes with
the pressure of 0.3Mpa. As shown in Figure, AE signal
of

valve leak
s

is continuous
and it contains a large amount of information. The spectra of three different types of
failure mod
es are similar in the frequency range of 60
-
180 kHz, whereas the peak
frequencies of all failure modes are about 85 and 160 kHz. While there is some
difference in signal amplitude and details of frequency range, i
t is difficult
to
distinguish the certain f
ailure mode both in time domain and frequency domain.
Therefore, in order to obtain the characteristic parameter of different types of failure
modes, it is necessary to analy
ze

AE signal based on advanced methods.


Feature of wavelet packet
-
energy
.

In the
present study, the AE signals are
decomposed into five resolution levels with ‘db8’ wavelets, producing a total of 32
subspaces which correspond to a frequency range of 0
-
7.8125, 7.8125
-
15.625,
15.625
-
23.4375, 23.4375
-
31.25, ... 484.375
-
492.1875, 492.1875
-
500 kHz.
Meanwhile, the total signal energy can be decomposed into a summation of wavelet
packet component energies that correspond to different frequency bands. Fig. 7
shows component signals of the 5th level WPT and wavelet packet component
energies for
a group of AE data. 46 groups of other AE data are processed by the
mentioned methods. A data matrix with wavelet packet component energies can be
constructed as:


Where n is the number of trials (n=47), and j is resolution level

(j=5). Each row
vector includes wavelet packet component energies for a group of AE data.



(
a
)
Component signals of the 5rd level WPT


(
b
)
Wavelet packet component energies for a group of AE data

Figure 7. Component signals of the 5rd level WPT and wavel
et packet
component energies for a group of AE data

Identification of the leak
s

failure modes
.

A Class labels vector is constructed to
illustrate three failure modes, where 1, 2, 3 represent that valve is adjusted to small
opening, some damage is done to t
he valve seal, combination of two failure modes,
respectively. Firstly, the data are divided into two subsets: A training set, used to
select genes and adjust the weights of the classifiers, and an independent test set used
to estimate the performance of t
he system obtained. According to the requirement of
the SVM, data set is classified as shown in Table 1.


Table 1. Classified Data Set for SVM

Failure Mode

(code)

Total Number of
Samples

Number of
Training Samples

Number of
Testing Samples

valve is adjus
ted to
small opening

1


27

22

5

some damage is done
to the valve seal

2


11

6

5

combination of two
failure modes

3


10

5

5


Secondly, data in matrix are normalized to a range of 0
-
1.0 and principal component
analysis (PCA) is used to reduce the dimensio
n of the classification data. The result
of PCA is shown in Fig. 8, where the black bar stands for variance explained of each
principal component and the blue line presents accumulated variance explained.



Figure 8. The result of PCA


Thirdly, RBF which

can be adjusted efficiently with parameter c and kernel
parameter g, is selected as kernel function. In order to find out the optimal parameter
values of RBF for training data, a grid search technique with 5
-
fold cross
-
validation
is used. Finally, SVM mod
e is established based on the optimal parameter values
(c=5.6569, g=0.8839, accuracy=100%), which are calculated according to the curve
in Fig. 9.


Figure 9. Optimal parameter values of RBF for training data


As shown in Fig. 10, two
-
class classification
of SVM is applied to the identification
of failure mode (1), (2) and the prediction accuracy is 90% (there are 10 test samples,
and only one of them is wrongly identified ). According to methods mentioned above,
three
-
class classification is applied to all

three failure modes. Results show that
accuracies of training and testing model are 95 % and 80% (there are 15 test samples,
and three of them are wrongly identified) respectively in Fig. 11.


Figure 10. Identification of failure mode (1), (2)



Figure
11. Identification of all three failure modes

CONCLUSION


In this paper, the identification of various leak
s

failure modes is investigated based
on acoustic emission technology. The time
-
domain characteristic and frequency
range of three different failure
modes are similar, and the peak frequencies of these
failure modes are about 85 and 160 kHz. It is difficult to intuitively distinguish the
certain failure mode. The method we proposed, which combines the feature of
wavelet packet
-
energy entropy with suppo
rt vector machine (SVM), is authentically
effective. It performs better towards identification of two failure modes that ball can't
be adjusted to full closed position and seal is damaged. However, further studies are
needed to improve identification accur
acy of three or more failure modes.


ACKNOWLEDGEMENTS


This work was financially supported by “the Fundamental Research Funds for the
Cen
tral Universities” (14CX06056A),

National science and technology support
project
(2011BAK03B08)
,

Shandong Province Natu
ral Science Foundation

(ZR2011EL048)

and Graduate Innovation Fund of China University of Petroleum
(East China) (JD12
-
04).


REFERENCES



Chen Q.S., Zhao J.W., Zhang H.D.,
et al.

(2006). Indentification of Authenticity of
Tea with Near Infrared Spectroscopy

Based on Support Voctor Machine.
J. Acta
Optica Sinica
, 933
-
937.

Kaewwaewnoi W, Prateepasen A and Kaewtrakulpong P (2010). Investigation of the
relationship between internal fluid leakage through a valve and the acoustic
emission generated from the leakag
e.
J. MEASUREMENT
, 274
-
282.

Lu J.Y., Wang B.F. and Liang D.K. (2012). Identification of impact location by using
FBG based on wavelet packet feature extraction and SVR.
J. Optics and
Precision Engineering
, 712
-
718.

Meland E, Henriksen V, Hennie E, et al.

(2011). Spectral analysis of internally
leaking shut
-
down valves.
J .MEASUREMENT
, 1059
-
1072.

Meland E, Thornhill N F, Lunde E, et al.(2012). Quantification of valve leakage rates.
J. AIChE Journal
, 1181
-
1193.

Sharif M A and Grosvenor R I (1998). Internal
valve leakage detection using an
acoustic emission measurement system.
J. Transactions of the Institute of
Measurement and Control
, 233
-
242.

Sun, Z. and C. Chang (2002). Structural Damage Assessment Based on Wavelet
Packet Transform.
J. Journal of Structur
al Engineerin
, 1354
-
1361.

Walczak, B. and D. L. Massart (1997). Noise suppression and signal compression
using the wavelet packet transform.
J.
Chemometrics and Intelligent Laboratory
Systems
,

81
-
94.

Yen, G. G. and K.
-
C. Lin. (2000). Wavelet packet feature

extraction for vibration
monitoring.

J. Industrial Electronics
, 650
-
667.