prediction of pccp failure based on hydrophne detectings2is.org/issues/v6/n4/papers/paper15.pdf ·...

17
INTERNATIONAL JOURNAL ON SMART SENSING AND INTELLIGENT SYSTEMS VOL. 6, NO. 4, SEPTEMBER 2013 1582 PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING Yuan Zhang, Yibo Li College of Precision Instruments and Optoelectronic Engineering Tianjin University , Weijin Road 72# Tianjin, 300072, China Emails:[email protected] Submitted: May 12,2013 Accepted: Aug.03,2013 Published:Sep.05,2013 Abstract: Prestressed Concrete Cylinder Pipe (PCCP) is a widely used water pipe all over the world. A major cause of PCCP failure is the internal wire break, which will emit acoustic signal. In this paper, a hydrophone-based PCCP real-time monitoring and failure-prediction system was proposed. By applying wavelet energy normalization analysis to signal feature extraction and Support Vector Machine (SVM) to signal recognition, a high prediction accuracy of 98.33% was achieved. The result showed that the hydrophone-based PCCP failure prediction system is much more effective and economic in real application compared with electromagnetic method and acoustic fiber optical. Index terms: Wire break signal, acoustic, PCCP, hydrophone, wavelet analysis, SVM

Upload: buibao

Post on 20-Mar-2018

227 views

Category:

Documents


6 download

TRANSCRIPT

Page 1: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

INTERNATIONAL JOURNAL ON SMART SENSING AND INTELLIGENT SYSTEMS VOL. 6, NO. 4, SEPTEMBER 2013

1582

PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE

DETECTING

Yuan Zhang, Yibo Li

College of Precision Instruments and Optoelectronic Engineering

Tianjin University , Weijin Road 72#

Tianjin, 300072, China

Emails:[email protected]

Submitted: May 12,2013 Accepted: Aug.03,2013 Published:Sep.05,2013

Abstract: Prestressed Concrete Cylinder Pipe (PCCP) is a widely used water pipe all over the world.

A major cause of PCCP failure is the internal wire break, which will emit acoustic signal. In this

paper, a hydrophone-based PCCP real-time monitoring and failure-prediction system was proposed.

By applying wavelet energy normalization analysis to signal feature extraction and Support Vector

Machine (SVM) to signal recognition, a high prediction accuracy of 98.33% was achieved. The

result showed that the hydrophone-based PCCP failure prediction system is much more effective

and economic in real application compared with electromagnetic method and acoustic fiber optical.

Index terms: Wire break signal, acoustic, PCCP, hydrophone, wavelet analysis, SVM

Page 2: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

Yuan Zhan and Yibo Li, PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING

1583

I INTRODUCTION

Prestressed Concrete Cylinder Pipe (PCCP) is a large-diameter water pipe widely used in the

world. Figure 1 illustrates the structure of PCCP, which primarily consists of five parts: the

internal concrete core, a thin steel cylinder, external concrete, highly prestressed wires and

external mortar layer [1][2]. Highly prestressed wire is winded on the circular surface of the

concrete core by certain tensile stress. When the size of inner concrete core is fixed, the PCCP

can bear different inner pressure and external loads by adjusting the prestressed wire diameter

and screw pitch. Therefore, the PCCP’s bearing capacity mainly depends on the wire winded

in it. Once the wire breaks due to artificial destruction or natural corrosion, the PCCP will

face threat of rupture[3], which brings about not only great economic loss but also potential

casualties [4][5][6]. Therefore the prediction of wire break in the PCCP is of great

significance.

Figure 1   Internal structure of the PCCP

Conventionally PCCP is inspected manually via visual observation or sounding [7]. Recently

new methods including Remote Field Eddy Current/Transformer Coupling (RFEC/TC) [8][9],

Acoustic Fiber Optical (AFO), and hydrophone-based acoustic inspection are implemented in

Page 3: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

INTERNATIONAL JOURNAL ON SMART SENSING AND INTELLIGENT SYSTEMS VOL. 6, NO. 4, SEPTEMBER 2013

1584

the PCCP failure detection[10][11][12]. RFEC/TC is an off-line inspection method with rather

low efficiency. In addition, the water in the PCCP has to be evacuated before the RFEC/TC

inspection, which will consume large human and material resources. AFO is a real-time

inspection method with quite high accuracy. However, AFO is only applicable to newly built

pipelines since the interior surface of old PCCP is unavailable for AFO installation, and it is

also expensive to lay fibers for AFO. Hydrophone-based inspection, an economic method

compared with AFO, is not only accurate due to the high sensitivity of the hydrophone to

acoustic signal but also convenient to install the hydrophone into the wells or holes that are

reserved and present on the outside surface of the PCCP when constructed. Due to its distinct

advantages, the hydrophone-based inspection has become a promising PCCP failure

prediction method with great economic potential.

In this paper, a hydrophone-based PCCP failure prediction method is proposed and fully

developed. Acoustic signal is sampled using a data acquisition device produced by the

National Instruments (NI) Corporation [13]. Signal features are extracted via wavelet

decomposition method. SVM is used to differentiate real wire break signal from other

interfering signals. The accuracy is up to 98.33%.

II ACOUSTICAL SIGNAL PROPAGATION AND IDENTIFICATION

a. Acoustic signal propagation in PCCP

PCCP is a typical cylinder waveguide, in which three families of guided waves: longitudinal,

torsional and flexural modes can propagate [14][15]. Longitudinal mode is axial-symmetric,

and studies have shown that L(0,1) mode and 1α mode exist in water-filler pipes surrounded

by any medium, whereas occurrence of 2α mode and 3α mode depends mainly on the

surrounded medium of the pipe [16][17]. L(0,1) mode wave has been demonstrated to suffer

strong attenuation due to leakage and scattering when encountering pipe joints and fittings

[14-16]. The attenuation of flexural mode F(1,1) wave is also large if the mode phase velocity

is greater than the longitudinal bulk velocities in the soil. Compared to L(0,1) and F(1,1)mode

Page 4: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

Yuan Zhan and Yibo Li, PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING

1585

wave,α mode wave experiences less attenuation due to its predominantly axial water-borne

displacements. It is predicted that α mode has the most chance to be the dominant mode in

the signal received over long propagation distance since α mode signal can propagate a long

time without much energy loss into the soil or pipe wall [16]. Based on this prediction, the

maximum spacing distance between sensors can be 260m within which acoustic signal still

propagate[18].

b. Signal feature extraction

Wavelet analysis is a time-frequency analysis method developed by Morlet in the 1980th and

it’s especially suitable for instable signals[19][20]. Wavelet decomposition is widely used in

signal feature extraction[21][22][23].The results of wavelet decomposition are coefficients

that include several details representing the high-frequency information and one

approximation representing the low-frequency information[24]. The wavelet decomposition

can be used for feature extraction by introducing the energy-mode concept. Suppose the

sampling rate of signal is fs, if a j layer wavelet is used to decompose the signal, the signal

can be decompose into j+i unequal frequency bands.

The jth layer wavelet coefficients can be expressed in cdk (k=1,2,3...j), which represent the

high frequency bands information, and caj which represent the low frequency energy

information. The time domain energy of signal x(t) can be reached by

dttxtx ∫+∞

∞−= 22 )()( (1)

According to the Parseval energy integration theory, x(t) in equation (1) can be connected

with the wavelet transform coefficients cdk and caj,, then we can get the following equation

∑=

=

+=jk

kjk cacdE

1

22 (2)

According to the equation (2), the wavelet transform coefficients cdk and caj have energy

calculation function.

The feature vector extraction based on wavelet decomposition method can be accomplished

according to the following steps:

First, decompose the signal by using wavelet analysis

Page 5: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

INTERNATIONAL JOURNAL ON SMART SENSING AND INTELLIGENT SYSTEMS VOL. 6, NO. 4, SEPTEMBER 2013

1586

Second, choose n frequency bands which are sensitive to energy and calculate each band’s

energy and normalize each band. Suppose E1,E2,E3....Ej are the energy in accordance with cdk

(k=1,2,3...j), and Ej+1is the energy in accordance with caj. Then we can get the following

equations

2

∑=m

mkk cdE ,(k=1,2,3,...j) (3)

m is the vector element number of cdk

2

1 ∑=+m

jj caE (4)

∑+=

=

= 1

1

'jk

kk

kk

E

EE (5)

Third, choose the above normalized energy as the feature vector of the signal, that is

],...,,[ '1

'2

'1 += jEEEE (6)

c SVM method

SVM (Support Vector Machine) is a classification and regression statistical theory proposed

by V.Vapnik in the AT&T Bell lab [25][26]. It’s widely used in signal recognition and

classification [27][28][29][30].The main concept of SVM lies in two notions: first, SVM is

used under the condition of liner separable cases. For the linear inseparable cases, non-linear

mapping algorithm is applied to transform the low-dimension input linear inseparable samples

into high-dimension feature space, making it possible to use liner algorithm to analyze the

non-linear characteristics of samples; second, when structural risk minimization theory is

applied in constructing the optimal split hyperplane among feature space, the studying

machine can get global optimization.

Given an input set liRx ni ,...2,1},{ =∈ that compose of two types of modes. If xi belongs to

the first type, then yi is 1, otherwise yi is -1. Here liy }1,1{ −⊂ is the SVM. Then the training

set can be expressed by liyx ii ,...,3,2,1},,{ = . The goal of SVM is to construct a target

function which can separate the two modes to extremity based on the risk structure

Page 6: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

Yuan Zhan and Yibo Li, PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING

1587

minimization theory. Under the linear separable condition, a hyperplane that is able to classify

the samples can be expressed as below:

0=+⋅ bxω (7)

Here “ ⋅ ” is dot product, ω is a normal vector of the hyperplane, b is the offset.

The hyperplane is reached by the following second-optimization

2

21)(min ωωφ = (8)

which meets the constraint condition

.,...,3,2,1,1)( nibxy ii =≥+⋅ω (9)

Under the condition of large feature number, the second-optimization problem can be changed

into dual problem.

)(21)(max

1,1jiji

n

jiji

n

ii xxyyW ⋅−= ∑∑

==

αααα (10)

ii

n

ii xy∑

=

=1

* αω (11)

ii xwyb ⋅−=* (12)

Which meets the following constraint condition

niy ii

n

ii ,...3,2,1,0,0

1=≥=∑

=

αα (13)

Here the ),...,,( 21 nαααα = is the Lagrange multiplier, *ω is the normal vector of the

hyperplane, *b is the offset of the hyperplane. KKT condition plays an important role in

such problem solution and analysis, which is shown in the equation (14)

nibxyii ,...,3,2,1,0}1)({ ==−+⋅ωα (14)

According to equation (14), the samples with 0=iα have no effect to the classification,

while samples with 0>iα are available in the classification process. The final classification

function is:

Page 7: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

INTERNATIONAL JOURNAL ON SMART SENSING AND INTELLIGENT SYSTEMS VOL. 6, NO. 4, SEPTEMBER 2013

1588

*

1

)()( bxxyxf ii

n

ii +⋅= ∑

=

α (15)

III METHODOLOGIES

Since wire break is one of the primary causes of PCCP failure, precise prediction of PCCP

failure depends on correct identification of PCCP wire break signal. In our approach,

hydrophones were placed along the PCCP pipeline used to detect the wire break signal.

However, in real monitoring system, interferential signals, such as the walking noise of

human being along the pipe, external interference noise of tapping due to construction, repair,

and the internal surging noise in the pipeline caused by air or unstable pressure inside the pipe,

can easily disturb or mix with wire break signal since the SNR of the interfering signals itself

is undesirable. Figure 2 shows the flow chart of the working process.

Abrupt acoustical signal received by

hydrophone

Data pre‐procesing

Wavelet feature extraction

SVM parameter selection

SVM classificationWire break

Warning signal

Continue regular work

No

Figure 2 Flow chart of the system working process

For the purpose of distinguishing the wire break signal from other interferential signals,

frequency-based feature extraction is an option since different types of acoustical signals were

dominated in different frequencies. Our research adopted wavelet analysis because wavelet

Page 8: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

Yuan Zhan and Yibo Li, PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING

1589

transform is good at scoping every detailed part of the original signal in both time-domain and

frequency-domain. By decomposing a signal into several sub- frequency bands, the energy

distribution in different bands can be utilized as eigenvectors of PCCP failure signal in feature

extraction.

After feature extraction, SVM was applied to the linear inseparable interfering noise and wire

break signal for classification. Since SVM is a structural-based risk minimization

classification tool, cross-validation was used for parameters optimization, which will enhance

the prediction accuracy tremendously. First, the original sample was divided into K average

groups, with one group as testing set, while the other K-1 groups as training set each time for

K times. The classification rates of the K models acquired are indicators of the property of the

model.

IV EXPERIMENT

a. System introduction

Figure 3 Experiment system display

Figure 3 is a sketch of the experiment system. The diameter of PCCP in this experiment is

1.4m. Hydrophone is inserted into PCCP through the joint well in the pipelines with sealing

Page 9: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

INTERNATIONAL JOURNAL ON SMART SENSING AND INTELLIGENT SYSTEMS VOL. 6, NO. 4, SEPTEMBER 2013

1590

measures. In real PCCP, the joint wells along the pipeline can be used for hydrophone

installation. The inner structure of hydrophone facilitates its high sensitivity to acoustic signal

in water. In the experiment, sensitivity of the hydrophone to broadband (1HZ-100KHZ)

acoustic signal is -170dB. Since the aimed acoustic signal in this experiment are all within the

human’s ear limitation (20HZ~20KHZ), hydrophone is capable of receiving the desired signal.

The signal acquisition device is USB 4431 developed by National Instrument Corporation,

with a 24 bits resolution , maximum 102.4ksps simultaneous sampling rate and ± 10V input

range. The user interface of data acquisition is programmed using LabView. The sampling

rate in this experiment is 44 kbps.

Besides the surging signal results from the instable pressure in the pipeline, and the wire

break signal results from PCCP inner failure, we artificially added two interfering signals:

human walking noise and hammer tapping on the pipe surface to mimic real PCCP

environments.

b. Data processing

As mentioned in 1.2, α mode is the dominant mode in water-filled pipeline with little

energy loss into soil or scattering at joint or fitting. Therefore, acoustic signal received by

hydrophone is mainly α mode acoustic signal. Figure 4 shows the time-domain wave plot

of the four signals, all of which can be considered as sudden event for the PCCP. The length

of every signal is 6000 points.

0 1000 2000 3000 4000 5000 6000-6

-4

-2

0

2

4

6

8Wire Break Signal

Samples

Ampl

itute

/V

0 1000 2000 3000 4000 5000 6000-5

0

5Internal surging interference signal

Samples

Ampl

itute

/V

0 1000 2000 3000 4000 5000 6000-15

-10

-5

0

5

10

15External tapping interference signal

Samples

Am

plitu

te/V

0 1000 2000 3000 4000 5000 6000-1.5

-1

-0.5

0

0.5

1

1.5Walking noise signal

Samples

Am

plitu

te/V

Figure 4 Time domain wave of four signals

Page 10: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

Yuan Zhan and Yibo Li, PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING

1591

c. Feature extraction

Wavelet energy normalization method is applied to extract the feature from the four different

signals. In the wavelet analysis process, the Daubechies6 wavelet is used as the wavelet basis.

The signals are decomposed into 8 layers. Since the sampling rate of data collected is 44kbps,

after the wavelet decomposition process, the signal is divided into 9 bands with each band

characterized by a normalized energy. Figure 5 displays the four signals’ energy distributions

in different bands.

12

34

56

78

9

12

34

0

0.5

1

Layer

Energy distribution of four sinals in 9 frequency bands

Signal type

Nor

mal

ized

ene

rgy

Signal type 1: wire break , 2: internal surging signal, 3: external tapping signal, 4: walking

signal. Layer 1(cd1): 11000HZ~22000Hz, 2(cd2):5500~11000HZ, 3(cd3): 2750~5500HZ,

4(cd4):1375~2750Hz, 5(cd5):688~1375HZ, 6(cd6): 344~688HZ, 7(cd7): 172~344HZ, 8(cd8):

86~172HZ, 9(ca8): 0~86HZ.

Figure5 Energy distribution of four signal base on wavelet decomposition

Figure 5 is the normalized energy distribution of the four different signals. The wire break

signal has rather average distribution in each band. While the internal surging signal

dominates in the low frequency band ca8(0~86HZ). The external tapping signal has

concentrated energy in cd5(688~1375HZ) and cd6(344~688HZ). The walking signal

dominates in the cd8(86~172HZ) and ca8(0~86HZ).From the energy distribution histogram,

obvious and clear difference of the four signals can be quantified by a series feature vector

which consists of 9 dimensions of elements. Since frequency is a fundamental characteristic

Page 11: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

INTERNATIONAL JOURNAL ON SMART SENSING AND INTELLIGENT SYSTEMS VOL. 6, NO. 4, SEPTEMBER 2013

1592

of acoustic signal, different types of signals may dominate in different frequencies, the

frequency-based feature extraction method is appropriate for the four acoustic signals.

d. SVM identification

0 30 60 90 1200

2

4

6

8 x 10-3 db1

samples

norm

aliz

ed e

nerg

y

0 30 60 90 1200

0.02

0.04

0.06db2

samplesno

rmal

ized

ene

rgy

0 30 60 90 1200

0.1

0.2

0.3

0.4db3

samples

norm

aliz

ed e

nerg

y

0 30 60 90 1200

0.2

0.4

0.6

0.8db4

samples

norm

aliz

ed e

nerg

y

0 30 60 90 1200

0.2

0.4

0.6

0.8db5

samples

norm

aliz

ed e

nerg

y

0 30 60 90 1200

0.2

0.4

0.6

0.8db6

samples

norm

aliz

ed e

nerg

y

0 30 60 90 1200

0.1

0.2

0.3

0.4db7

samples

norm

aliz

ed e

nerg

y

0 30 60 90 1200

0.2

0.4

0.6

0.8db8

samples

norm

aliz

ed e

nerg

y

0 30 60 90 1200

0.5

1ca8

samples

norm

aliz

ed e

nerg

y

Figure 6 Scatter plots of 120 samples including four different signals

In the classification process, a total number of 120 sets of data are analyzed. Set 1 to 30 are

wire break signals, sample 31 to 60 are internal surging signals and sample 61 to 90 are

external tapping signals and sample 91 to 120 are walking signals. Before the SVM

classification process, feature extraction of 120 samples is conducted based on wavelet energy

normalization method. After the feature extraction, each sample has 9 element feature vector

representing energy intensity in 9 different frequency bands. Figure 6 displays the energy

distribution of 120 samples in different frequency bands. From the figure, we can see

db3,db5,db6 and ca8 have quite good classification.

Page 12: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

Yuan Zhan and Yibo Li, PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING

1593

-10-8

-6-4

-20

24

68

10

-10-8

-6-4

-20

24

68

1030

40

50

60

70

80

90

100

log2c

SVM Parameter selection result(3D View) Best c=32 g=32 CVAccuracy=96.25%

log2g

Acc

urac

y(%

)

Figure 7 Large step SVM parameter selection result

-10-8

-6-4

-20

24

68

10

-10-8

-6-4

-20

24

68

1030

40

50

60

70

80

90

100

log2c

SVM Parameter selection result(3D View) Best c=24.2515 g=73.5167 CVAccuracy=97.5%

log2g

Acc

urac

y(%

)

Figure 8 Small step SVM parameter selection result (3D View)

In the application of SVM, cross validation is conducted before the final classification. At first,

a large step SVM parameter selection is conducted. The parameter c and g change

Page 13: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

INTERNATIONAL JOURNAL ON SMART SENSING AND INTELLIGENT SYSTEMS VOL. 6, NO. 4, SEPTEMBER 2013

1594

from2^( -10) to 2^(10) with a step of 0.5 .The predicted accuracy is shown in Figure 7. The

approximate range of c and g is achieved, where c is 32 and g is 32. The accuracy is 96.25%.

Then a small step parameter selection is carried out. The parameter c and g change

from2^( -10) to 2^(10) with a step of 0.1 The result is shown in Figure 8. The best c is

24.2515and best g is 73.5167. The accuracy is 97.5%.

By comparison of the results shown in Figure 7 and Figure 8, the accuracy of small step is

higher than that of large step. Therefore a final optimal parameter selection is accomplished

by small step. The best c is 24.2515 and the best g is 73.5167.

After the determination of best c and g, the next step is to apply the best c and g in the SVM

model. Of the totally sampled 120 sets of data, each ten samples are chosen from the four

different types as training samples. Then a classification model is built based on the training

samples. A total of 120 samples are utilized to test the accuracy of the model built based on

SVM method. A final accuracy of 98.33%(118/120) is shown in Figure 9. We can see that

there are one internal surging signal and one walking noise signal predicted wrong. All the

wire break samples are predicted right.

0 20 40 60 80 100 1201

1.5

2

2.5

3

3.5

4

Samples

Sam

ple

labe

ls

Test labelPredict label

Figure9 Prediction result based on SVM

Page 14: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

Yuan Zhan and Yibo Li, PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING

1595

V CONCLUSIONS

(1)An efficient and economic PCCP failure prediction system is proposed. Hydrophone is

used as the sensor due to its low cost, easy installation, and high sensitivity to acoustic signal

in the PCCP. The result shows that the hydrophone is proper for PCCP failure prediction.

(2)The wavelet method is proved to be appropriate for the feature extraction of the wire break

signal due to its fundamental frequency characteristics. The features of wire break signal and

other three signals including human walking noise, internal surging noise and external tapping

noise, show very distinct differences.

(3)SVM is proven to be an efficient and accurate method for type classification. By selecting

appropriate parameters for the SVM method, high classification precision can be achieved.

Wire break signal can be easily identified with reduced false prediction.

ACKNOWLEDGEMENTS

This work was supported by the National Natural Science Foundation of China (Grant

NO.60974110) and Tianjin tap water pipeline company.

REFERENCES

[1]S. Ge and S. K. Sinha, "Effect of Wire Breaks on Prestressed Concrete Cylinder Pipe

(PCCP) Reinforced with Steel Liners—A Case Study", in Pipelines 2012@ sInnovations in

Design, Construction, Operations, and Maintenance, Doing More with Less, pp. 1297-1306,

2012.

[2] S.-p. Sun and M. Wang ",Characteristic of prestressed concrete cylinder pipe," Municipal

Engineering Technology, 2006.

[3] J. J. Galleher and A. E. Romer, "Who Says You Need Multiple Wire Breaks for a PCCP

Pipe to Fail?", in Pipelines 2012@ sInnovations in Design, Construction, Operations, and

Maintenance, Doing More with Less, pp. 278-287, 2012.

Page 15: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

INTERNATIONAL JOURNAL ON SMART SENSING AND INTELLIGENT SYSTEMS VOL. 6, NO. 4, SEPTEMBER 2013

1596

[4] M. Higgins and P. Paulson, "Fiber optic sensors for acoustic monitoring of PCCP", in

Proceedings of the 2006 Pipeline Division Specialty Conference-Pipelines 2006: Service to

the Owner, 2006.

[5] N. D. Faber, M. R. Coghill, and J. J. Galleher, "Beyond the Wires: A Sustainable Approach

to Prestressed Concrete Cylinder Pipe Management", in Pipelines 2012@ sInnovations in

Design, Construction, Operations, and Maintenance, Doing More with Less, pp. 1214-1222,

2012.

[6] M. S. Zarghamee, D. W. Eggers, R. Ojdrovic, and B. Rose, "Risk analysis of Prestressed

concrete cylinder pipe with broken wires", Pipeline Engineering and Construction

International Conference 2003, pp. 599-609,2004.

[7] R. You and H. B. Gong, "Failure Analysis of PCCP with Broken Wires", Applied

Mechanics and Materials, vol. 193, pp. 855-858, 2012.

[8] A. Biggar, "Detecting Wire Breaks from the Outside of PCCP", in Proceedings of the

ASCE International Pipelines Conference 2010: Climbing New Peaks to Infrastructure

Reliability—Renew, Rehab, and Reinvest 2010.

[9] M. S. Zarghamee and R. P. Ojdrovic, "Risk Assessment and Repair Priority of PCCP with

Broken Wires", 2004.

[10] G. E. Bell and P. Paulson, "Measurement and Analysis of PCCP Wire Breaks, Slips and

Delaminations", Pipelines 2010: Climbing New Peaks to Infrastructure Reliability - Renew,

Rehab, and Reinvest - Proc. of the Pipelines 2010 Conference, vol. 386, pp. 1016-1024, 2010.

[11] P. Cliff Moore, "Use of Acoustic Monitoring Data for PCCP Condition Assessment",

Pipelines 2009: Infrastructure's Hidden Assets - Proceedings of the Pipelines 2009 Conference,

vol. 360, pp. 45-54, 2009.

[12] M. Higgins and P. Paulson, "Fiber optic sensors for acoustic monitoring of PCCP", in

Proceedings of the 2006 Pipeline Division Specialty Conference-Pipelines 2006: Service to

the Owner, 2006.

[13]N.Afsarimanesh & P. Z. Ahmed, “LabVIEW Based Characterization and Optimization of

Thermal Sensors”, International Journal On Smart Sensing and Intelligent Systems, vol. 4, No.

Page 16: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

Yuan Zhan and Yibo Li, PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING

1597

4, pp. 726-739, 2011.

[14]R. Long, M. Lowe, and P. Cawley, "Attenuation characteristics of the fundamental modes

that propagate in buried iron water pipes", Ultrasonics, vol. 41, No. 7, pp. 509-519, 2003.

[15] R. Long, K. Vine, M. Lowe, and P. Cawley, "Monitoring acoustic wave propagation in

buried cast iron water pipes", in AIP Conference Proceedings, p. 1202, 2001.

[16] R. Long, P. Cawley, and M. Lowe, "Acoustic wave propagation in buried iron water

pipes", Proceedings of the Royal Society of London. Series A: Mathematical, Physical and

Engineering Sciences, vol. 459, No. 2039, pp. 2749-2770, 2003.

[17] B. K. Sinha, T. J. Plona, S. Kostek, and S. K. Chang, "Axisymmetric wave propagation in

fluid‐loaded cylindrical shells. I: Theory", The Journal of the Acoustical Society of America,

vol. 92, p. 1132, 1992.

[18] F. Stulen and J. Kiefner, "Evaluation of acoustic emission monitoring of buried pipelines",

in 1982 Ultrasonics Symposium, pp. 898-903, 1982.

[19] S. G. Mallat, "A theory for multiresolution signal decomposition: the wavelet

representation", Pattern Analysis and Machine Intelligence, IEEE Transactions on, vol. 11, No.

7, pp. 674-693, 1989.

[20] Y. Meyer and D. H. Salinger, “Wavelets and operators” vol. 1: Cambridge university

press, 1995.

[21] L. I. Kuncheva and J. J. Rodríguez, "Interval feature extraction for classification of

event-related potentials (ERP) in EEG data analysis", Progress in Artificial Intelligence, pp.

1-8, 2013.

[22] H.Hashim, S.Ramli, N.Wahid, & M.S. Sulainman , “Recognition of psoriasis features via

Daubechies D8 wavelet technique”, International Journal On Smart Sensing and Intelligent

Systems, vol.6, No.2, pp.711-732,2013.

[23] B. Lechner, M. Lieschnegg, O. Mariani, M. Pircher, &A. Fuchs,”A wavelet-based bridge

weigh-in-motion system”, International Journal On Smart Sensing and Intelligent Systems

vol.3, No.4, pp. 573-591, 2010.

[24] I. Daubechies, "Ten lectures on wavelets vol. 61: SIAM, 1992.

Page 17: PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTINGs2is.org/Issues/v6/n4/papers/paper15.pdf · PREDICTION OF PCCP FAILURE BASED ON HYDROPHNE DETECTING ... PREDICTION OF PCCP FAILURE

INTERNATIONAL JOURNAL ON SMART SENSING AND INTELLIGENT SYSTEMS VOL. 6, NO. 4, SEPTEMBER 2013

1598

[25] V. Vapnik, "The nature of statistical learning theory: springer, 2000.

[26] B. E. Boser, I. M. Guyon, and V. N. Vapnik, "A training algorithm for optimal margin

classifiers", in Proceedings of the fifth annual workshop on Computational learning theory, pp.

144-152, 1992.

[27] W. Li, P. Fu, and W. Cao, “Study on feature selection and identification method of tool

wear states based on SVM” , International Journal On Smart Sensing and Intelligent Systems,

vol.6, No.2, pp.448-465.2013.

[28] A.-M.Cretu & P.Payeur, “Biologically-inspired visual attention features for a vehicle

classification task”, International Journal On Smart Sensing and Intelligent Systems, vol.4,

No.3, pp.402-423,2011.

[29] W. Zu, J. Yuan, and W. Zhang, "Determining method for reliability distribution function

of transformer fault diagnosis based on SVM", Heilongjiang Dianli Jishu(Heilongjiang

Electric Power), vol. 35, No. 2, 2013.

[30] L. Jian, K. Weikang, S. Jiangbo, W. Ke, W. Weikui, Z. Weipu, et al., "Determination of

Corrosion Types from Electrochemical Noise by Artificial Neural Networks", Int. J.

Electrochem. Sci, vol. 8, No. pp. 2365-2377, 2013.