Professional Documents
Culture Documents
Stationarity Index Based Segmentation Approach in Leak Locating Systems
Stationarity Index Based Segmentation Approach in Leak Locating Systems
Abstract—This paper presents a practical criterion for velocity and the distance between the sensors, the leak can
extracting piecewise stationary segments from pipeline leak be located [4],[5],[6],[7].
signals, in order to better locate the leak on the pipe. The However, in practice, real leak data prove to perform a
segmentation technique, that enables detecting and avoiding number of random abrupt amplitude changes, causing
the non stationary abrupt data changes, is based on certain deviations from the original ideal assumptions and
computing the stationarity index (SI) function of the lower estimation results [12]. These burst interferences
acquired noisy leak signals. A comparative study, involving can be produced internally, from a sudden pressure and
different segmentation thresholds on real leak signals, flow velocity variation (turbulent flow), or externally by
captured in an experimental pipeline installation, indicate non stationary disturbing noises such as traffic, human
an optimum threshold. According to the proposed criterion, voice, etc., [8]. A useful idea in preprocessing real leak
the signals acquired with pipeline’s end obstructed,
signals is to detect and avoid these abrupt data changes by
practically needed no segmentation and gave the best
extracting piecewise stationary segments from the
estimates.
acquired signals for a better implementation of the
Index Terms—Time delay estimation, leak detection,
estimation algorithms.
stationarity index. The time-frequency representation (TFR) of the
recorded data is an important tool for characterizing non
stationary signals. Usually, the TFR [1] is a two-
I. INTRODUCTION dimensional plot in the time-frequency plane that displays
An important problem for pipeline transportation the estimated power spectral density of the signals passed
systems is to find leaks as soon as possible as they may through a moving window. The window’s length is
occur along the pipe. Leaks in pipelines carrying different comparable to the time over which the data can be
fluids can cause serious environment pollution or injures assumed to remain stationary.
of the pipe bedding, roads or nearby buildings. One of the An efficient detection method, introduced by Laurent
most known methods in the literature for leak locating is and Doncarli [9],[10], based on the time-frequency
based on the analysis of the acoustic noise generated by representation TFR (t,f), define the “stationarity index”,
the fluid passing through the leak. The acoustic leak signal (SI) function which characterizes the abrupt spectral
can be captured by non intrusive sensing devices placed signals’ changes. According to this definition, the
on the pipeline. The leak locating principle consists of stationarity index at each time instant t is computed as the
estimating the time delay at which the leak signal reaches distance measure of two TFR sub-images around that time
at two separate locations on the pipe [2],[3]. The current instant. Between different distance measures proposed in
mathematical model for this problem is described by a the literature [9], [10], [11], the Kolmogorov distance,
linear set of equations: which generally proved to give the best results define the
stationarity index as follows:
r1 (t ) = s (t ) + n1 (t ), (1)
p
r2 (t ) = s (t − D ) + n 2 (t ) SI (t ) = ∫ ∫ TFR(t − p + τ , f ) − TFR(t + τ , f ) dfdτ ,(2)
τ =0
where r1(t), r2(t) are the received signals, n1(t), n2(t) are
the disturbing additive noises at the sensor locations, s(t) where p is the sub-image’s width and τ ∈ [0, p ].
is the original leak noise and D is the time delay, desired
to be estimated. A typical non stationary leak data pair, its TFRs and the
derived stationary indices are presented in Fig.1. It can be
Most of the techniques from the literature assume that seen that the SI functions perform sharp peaks
the random received signals are described by ideal corresponding to signals’ burst portions and are nearly
features (stationary, white, Gaussian) [2],[3],[4],[5],[6], constant otherwise.
[7],[13]. In this case, a typical method for estimating the
time delay is to compute the cross-correlation function This paper investigates how this function can be applied
between the received signals. The argument at which the particularly to leak signals, as a preprocessing stage for
cross-correlation function’s maximum occurs is the further estimation algorithms implementations..
estimated time delay. Knowing the noise’s propagation
Authorized licensed use limited to: Jesus Esquivel. Downloaded on March 17,2010 at 10:03:55 EDT from IEEE Xplore. Restrictions apply.
quantities have to be compared, according to the proposed
criterion. The first one, (denoted with Q1) is the absolute
difference between the SI’s mean and trimmed mean,
while the second one, (denoted with Q2) is the peak-to-
peak SI’s amplitude, SIpp divided by a constant, k, as in the
following relationships:
Q1 = m0 − m ,
(4)
Q2 = SI pp / k
Authorized licensed use limited to: Jesus Esquivel. Downloaded on March 17,2010 at 10:03:55 EDT from IEEE Xplore. Restrictions apply.
As shown in Fig.2, two parameters must be specified x 10
4
Stationarity Index-Case Two
for computation, namely r and k. The first one represents 2
SI
1
Fig.3 and fig.4 present some typical algorithm 0.8
implementation situations. In the first case, Q1 is higher
than Q2, as shown in the upper part of fig.3. Therefore, the 0.6
6000
4000
0
0 2 4 6 8 10 12 III. THE EXPERIMENTAL MODEL
The leak signals used for this study have been captured
in an experimental bended water transportation pipeline
with configuration depicted in fig.5. The studied system,
having a total length of 12.82 meters was implemented
from metal pipes of 2.54 centimeters diameter each. The
leaks, placed on the second and third installation sides
were simulated by faucets for flow rate adjustments. The
experimental pipeline installation was also equipped with
a supplementary faucet placed at the pipe’s end creating
the alternative of obstructing the output flow.
The signal pairs were acquired at the same locations on
the pipe considering two working modalities: without and
with pipeline end obstruction faucet closed, respectively.
The measuring points were equally distributed at 0.3
meters intervals along the pipe.
The acquisition system was composed of a pair of non
intrusive vibration sensors KD Radebeul, two amplifiers
M60T with adjustable amplification between 40 and 60
dB, anti-aliasing low pass filters and a dSPACE DS1102
Fig.3 Typical situation when segmentation is needed board connected to a PC [17]. The sampling frequency
was set to 25 KHz. The processing algorithms were
implemented using the MATLAB® environment [16].
Authorized licensed use limited to: Jesus Esquivel. Downloaded on March 17,2010 at 10:03:55 EDT from IEEE Xplore. Restrictions apply.
Typical Cross-Correlation Function
0.4
Bended Water Transportation Pipe
0.3
0.2
0.1
Amplitude
0
-0.1
-0.2
End -0.3
Obstruction
-0.4
Faucet
-0.5
Simulated
Leaks -2.5 -2 -1.5 -1 -0.5 0 0.5 1 1.5 2 2.5
Time [ms]
Sink
C (f )
-0.2
1
W( f ) = ⋅ 12 ,
-0.4
0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 1.1 1.2 S 12 ( f ) 1 − C12 ( f ) (5)
S 12 ( f )
2
C12 ( f ) =
0.4
.
0.2 S 11 ( f ) ⋅ S 22 ( f )
Amplitude
0
In the above relationships, S12(f) is the estimated cross-
-0.2 spectrum between the received signals, r1 and r2, while
-0.4
S11(f) and S22(f) are auto-spectra of r1 and r2 , respectively.
0.3 0.4 0.5 0.6 0.7 0.8
time[s]
0.9 1 1.1 1.2 Also, C12(f) defines the acquired signals’ squared
coherence function. Finally, the ML windowed cross-
correlation function is computed as the inverse Fourier
Fig.6 Induced transient signals for propagation velocity estimation , transform of the product between the ML window and the
TOF method signals’ cross-spectrum, W(f) and S12(f). The time delay is
Authorized licensed use limited to: Jesus Esquivel. Downloaded on March 17,2010 at 10:03:55 EDT from IEEE Xplore. Restrictions apply.
estimated as the windowed cross-correlation’s peak, using Table II summarizes the slope values computed by the
a three point interpolation, according to: four techniques for seven threshold levels also (in cases
where segmentation was necessary).
Dˆ i = Dˆ −
1 ( ) ( )
R12 Dˆ + 1 − R12 Dˆ − 1 (6)
For each implemented algorithm, the estimated time
delay variation versus distance is presented in fig.9 to
( ) ( ) ( )
2 R12 Dˆ + 1 − 2 R12 Dˆ + R12 Dˆ − 1 fig.11. Also, the corresponding best straight line
approximations for each data set, along with the expected
In the second processing technique, the received signals delay line are shown superimposed. Fig. 12 presents a
are passed through whitening filters prior computing the synthetic results’ general view for a better algorithm
ML cross correlation function. Here, the received signals performance comparison.
are first modeled as AR random processes, then In conclusion, the results presented here bring some
transformed into signals of white noise type, also called useful insights regarding some algorithms used for
the innovations representations [1],[15]. In the following processing real leak signals and leak detection. The tables
study, this technique was denoted with WML. and fig.9 to fig.12 indicate that the best procedure for
The described ML and WML algorithms were also which both the proportionality factor is minimum and the
implemented for leak signals acquired with obstruction at approximation line’s slope reaches the expected slope the
the pipeline’s end. Here, the last two techniques were most, is OWML. This technique adds together both a
denoted with OML and OWML. “hardware” procedure by closing the pipeline’s end and an
optimum “software” algorithm combination.
B. Comparative Results Regarding the other case where the signals were
In order to compare the above described algorithms, the acquired without closing the pipeline’s end, the proposed
leak signals have been acquired by keeping one sensor segmentation criterion decided for most of the signals, that
fixed while changing the other’s sensor position gradually, the segmentation algorithm needed to be applied. In this
along the pipe at the equally spaced measuring points’ case, the results show the existence of an optimum
locations. Therefore, assuming a uniform propagation, the segmentation threshold (here at the second level) for
estimated time delays obtained from one point to another which, especially the WML algorithm, comes nearer to
need to have values proportionally distributed along the the desired values for proportionality and slope.
pipe. In this experiment, the leak signals have been
acquired from the straight sides of the pipeline installation V. CONCLUSIONS
(fig.5). A measure of the data “proportionality degree” Leak signals acquired in real pipeline systems are
can be obtained by subtracting the best approximation line affected by burst-type interferences that worsen the leak
from the estimated delays along the pipe and computing locating accuracy by inducing certain signal deviations
the error’s power. A minimum value is desired for this from the ideal characteristics. A way for obtaining better
factor. estimation results is to avoid the abrupt signal changes by
The segmentation algorithm, described in section II, has extracting piecewise stationary segments from the
been applied to the received data considering seven acquired data. This work proposes a practical
threshold levels, for signals that, according to the segmentation algorithm for leak signals, based on the
proposed criterion, demanded this procedure. An stationarity index computation. An associated criterion
interesting observation was that the leak signals acquired for deciding if the acquired signals need segmentation is
with the pipeline’s end obstructed, generally needed no also proposed. A comparative study was conducted in an
segmentation. Table I resumes the estimated delay’s experimental pipeline installation. The results indicate
proportionality degree factor for the four previously that the algorithm that gives the best estimates includes
described techniques. combination of “hardware” and “software” methods.
Also, another study was accomplished taking into Finding the signal’s innovations representation prior to
account the slopes of the estimated delays’ approximation proper applying the estimation algorithm proved to
lines, compared to the expected slope, computed from the improve the obtained results, as noticed in [12], also.
propagation velocity information, previously estimated. Future work will include studies on mean varying leak
Authorized licensed use limited to: Jesus Esquivel. Downloaded on March 17,2010 at 10:03:55 EDT from IEEE Xplore. Restrictions apply.
signals. Proportionality Error Power
8
ML
Estimted Delay vs. Distance -ML Algorithm WML
6
OML
Amplitude
400 thresh1 WOML
thresh2 4
thresh3
200 thresh4 2
thresh5
Estimated delay [microsec.]
0 thresh6 0
thresh7 1 2 3 4 5 6 7
expected line
-200 ML lines Aproximation Line Slope
6
-400
4
Amplitude
-600 ML
2
WML
-800 OML
0
WOML
Expected slope
-1000 -2
1 2 3 4 5 6 7
Threshold No.
-1200
-200 -150 -100 -50 0
Length [cm.]
Fig. 12 Algorithm comparison from proportionality degree and slope
vicinity points of view, respectively
Fig. 9 Experimental and expected delays vs. distance -ML Algorithm
thresh7
1000
expected line [3] IEEE Trans. On Acoustics, Speech and Signal Processing -
WML lines Special Issue on Time Delay Estimation, June 1981.
500 [4] K. Watanabe, H.Koyama, H. Tanoguchi and D.M. Himmelblau,
“Location of pinholes in a pipeline”, Computers & Chemical
0 Engineering, vol. 17, No. 1, pp. 61-70, January, 1993.
[5] O. Hunaidi,W.T.Chu, “Acoustical characteristics of leak signals in
-500 plastic water distribution pipes”, Applied Acoustics, Vol. 58
(1999),pp. 235-254.
-1000 [6] O. Hunaidi, A.Wang, “Leak finder- new pipeline leak detection
system”, 15th World Conference on Non-Destructive Testing,
-200 -150 -100 -50 0
Rome Italy, Oct. 2000,pp.1-6.
Length [cm.]
[7] O. Hunaidi,W.T.Chu, A.Wang, W. Guan, “Detecting leaks in
Plastic Pipes”, Journal AWWA, Vol 92, No.2, pp.82-94, February
Fig. 10 Experimental and expected delays vs.distance -WML 2000.
Algorithm [8] Y. Wen, P.Li, J. Yang, Z. Zhou, “Information processing in buried
pipeline leak detection system”, IEEE, Proc. of International
Conference on Information Acquisition, 2004.
Estimted Delay vs. Distance -ML Algorithm, End Obstructed Pipe
200 [9] H. Laurent, C. Doncarli, “Abrupt changes detection in the time-
OML frequency plane”, Proc. IEEE Int. Symp Time-Frequency and
WOML Time-Scale Analysis, 1996, pp.285-288.
0 expected line
OML line [10] H. Laurent, C. Doncarli, “Stationarity index for abrupt changes
WOML line detection in the time-frequency plane”, IEEE Signal Processing
Estimated delay [microsec.]
-200
Letters, vol. 5, No2, pp.43-45, February 1998.
[11] J.M. Smulko, “Abrupt changes detection of broad-band signals”
-400
IEEE Instrumentation and Measurement Technology Conference,
Budapest, Hungary, pp. 1139-1142 Mai 21-23, 2001.
-600
[12] M. Daneti, “Modeling burst interferences-a practical tool for
studying leak signals”, IEEE Proc. of The 2-nd International
-800 Design and Test Workshop pp.111-112, December 2007.
[13] M. Daneti, S. Ionel, “An overview of time delay estimation
-1000 algorithms”, Trans. on Electronics and Communications, Tom
45(59), Fasc.1, 2000.
-1200
-200 -150 -100 -50 0
[14] M. Daneti, S. Ionel, “A dSPACE implementation of a time
Length [cm.] estimation algorithm”, Trans. on Electronics and
Communications, Tom 45(59), Fasc.1-2, 2002.
[15] J. G. Proakis, C. M. Rader, F. Ling,C.Nikias, Advanced Digital
Fig.11 Experimental and expected delay data and approximation lines - Signal Processing, Maxwell Macmillan International Edidtions,
OML and WOML algorithms, end obstructed mode 1992.
[16] The Mathworks Inc.,” Matlab, Simulink”, 1999
[17] dSPACE, “DS1102 user’s guide”, 3-rd version 1999
Authorized licensed use limited to: Jesus Esquivel. Downloaded on March 17,2010 at 10:03:55 EDT from IEEE Xplore. Restrictions apply.