energies

Article
A Reliable Acoustic EMISSION Based Technique for
the Detection of a Small Leak in a Pipeline System
Thang Bui Quy , Sohaib Muhammad and Jong-Myon Kim *
School of Computer Engineering, University of Ulsan, Ulsan 44610, Korea; bqthangcndt@gmail.com (T.B.Q.);
md.sohaibdurrani@gmail.com (S.M.)
* Correspondence: jmkim07@ulsan.ac.kr; Tel.: +82-52-259-2217




Received: 10 March 2019; Accepted: 15 April 2019; Published: 18 April 2019


Abstract: This paper proposes a reliable leak detection method for water pipelines under different
operating conditions. This approach segments acoustic emission (AE) signals into short frames based
on the Hanning window, with an overlap of 50%. After segmentation from each frame, an intermediate
quantity, which contains the symptoms of a leak and keeps its characteristic adequately stable even
when the environmental conditions change, is calculated. Finally, a k-nearest neighbor (KNN) classifier
is trained using features extracted from the transformed signals to identify leaks in the pipeline.
Experiments are conducted under different conditions to confirm the effectiveness of the proposed
method. The results of the study indicate that this method offers better quality and more reliability
than using features extracted directly from the AE signals to train the KNN classifier. Moreover, the
proposed method requires less training data than existing techniques. The transformation method is
highly accurate and works well even when only a small amount of data is used to train the classifier,
whereas the direct AE-based method returns misclassifications in some cases. In addition, robustness
is also tested by adding Gaussian noise to the AE signals. The proposed method is more resistant to
noise than the direct AE-based method.

Keywords: acoustic emissions; k-nearest neighbor; leak detection; pipeline diagnostics; reliability

1. Introduction
Water pipeline systems are required to accomplish commercial and domestic activities. A leak
in such a system can cause financial loss, waste of resources, and even human deaths due to the
collapse of the system. Thus, numerous approaches have been introduced to diagnose pipeline
leaks. The approaches include passive and active systems [1], hardware-based and software-based
methods [2]. Previous works [3,4] also provided an overview of research on leak detection. Among
those approaches, acoustic emission (AE)-based techniques, which are passive and hardware-based,
are promising because AE sensors can detect small leaks quickly, offering high sensitivity regarding
fault growth in a pipeline system. Thus, many researchers have exploited AE-based techniques for
pipeline diagnostics [5–11].
The first stage of the diagnosis process is to spot an existing fault. Because of the complexity of
AE activity, it is difficult to establish a leak detection model directly from mathematical equations;
therefore, different methods have been proposed to train the model using recorded datasets [6,7].
Although these techniques offer high accuracy, they could be inefficient in diverse circumstances
because the classifiers are trained using features extracted directly from AE signals. Thus, it makes the
classifiers dependent on the absolute signal levels that are influenced by the flow rate and pressure
in a pipeline system [12,13]. Moreover, AE signals also vary with temperature, which is uncertain in
industries. As a result, AE measurements recorded under such uncertain working conditions have
different signal levels, and thus, a fault diagnosis model based on the absolute amplitude values of the

Energies 2019, 12, 1472; doi:10.3390/en12081472 www.mdpi.com/journal/energies

Energies 2019, 12, 1472 2 of 18

signals is unreliable. In [14], the authors presented a study that investigates the inherent properties of
the vibro-acoustic signals instead of relative amplitude values. It delivers useful insights and results for
water leak detection, but the study focused on pipelines buried under soil less influenced by external
factors while AE signals from pipelines above soil in factories are prone to noise. This issue can be
addressed by acquiring many training datasets under different working conditions to provide enough
information for the classifier. However, it is not an optimal solution because the design of experiments
can become immensely expensive or even prohibitively complicated. Therefore, this study extracts
features from transformed signals independent of an absolute level instead of directly from AE signals.
The obtained classifier can detect small leaks in a pipeline with high accuracy and reliability.
For the experiments, this study uses a water pipeline system deployed in a laboratory. Three AE
sensors are mounted on different parts of the system with the known distances among the sensors.
A pinhole is drilled through the pipe wall to simulate leakage of the pipeline and a valve is attached
to the wall to control the flow of water through the leak. When the valve is closed, and system is
working in a normal state (no leakage) then there is no variation in the levels of the recoded signals.
If a small leak appears between sensors 2 and 3, an imbalance is created between AE channels 1 and 2
by reporting dissimilar amounts of noise from the leak. Since sensor 2 is nearer the leak than sensor 1,
the leak noise on channel 2 is greater than that on channel 1, as explained by the theory of wave
attenuation [15–17]. Thus, this uneven behavior of the signals can be used to detect small leaks in a
pipeline system. The leak-signal-to-normal-noise ratio and the attenuation law of wave propagation
can be used to detect the leak; these characteristics are demonstrated in Section 3.
To demonstrate the advantages of the balance/imbalance-based approach, this paper uses the theory
of Kullback-Leibler distance [18] to measure the separability between the two classes (i.e., NORMAL
and ABNORMAL) in different experiments. Furthermore, the paper applies a k-nearest neighbor
(KNN) model to classify between normal and leak. The KNN classifier is trained by two approaches,
one using features extracted directly from the AE signals and the other using features extracted from
the transformed signals. The KNN classifier in both the approaches are trained on the data of one
working condition and tested using data of other conditions. In addition, Gaussian noise is added to
the AE measurements to evaluate the robustness of the classifiers trained by the two approaches.

2. Experimental Setup

2.1. Data Acquisition

Figures 1 and 2 show a setup of the AE signals acquisition from a water pipeline system. The testbed
installation and the sensors configuration parameters are listed in Table 1.

Table 1. Experimental parameters.

No Quantity Detail
1 Distance from Sensor 2 to Sensor 3 (D) 2000 (mm)
2 Distance from Sensor 2 to Sensor 1 (d) 500 (mm)
3 Distance from Sensor 2 to Leak (d2) 900 (mm)
4 Distance from Sensor 3 to Leak (d3) 1100 (mm)
5 Thickness of pipeline 6.02 (mm)
6 Outer diameter of pipeline 114.3 (mm)
7 Material of pipeline Stainless steel 304

The experiments are based on four leaks with different diameters, i.e., 2.0, 1.0, 0.5, and 0.3 mm
(mm), abbreviated as L1, L2, L3, and L4, respectively. The water flow is controlled using pressures of 7,
13, and 18 bar, given as P1, P2, and P3, respectively. The experiments are conducted under a stable
temperature of approximately 29 ◦ C.
Energies 12,x1472
2019,12,
Sensors 2019, FOR PEER REVIEW 33 of
of 18
18
Sensors 2019, 12, x FOR PEER REVIEW 3 of 18

D
Pressure D Leak d1
Pressure
meter Leak d1
meter 3 AE Sensor 2 1
3 AE Sensor d3
d3
Valve
Valve
d2 AE Sensor
AE Sensor
2 d
d
1AE Sensor
d2 AE Sensor
A pipe sample
A pipe sample
(water-induced flow)
(water-induced flow)

flow
flange flow flange
flange flange

Figure 1.
1. Data acquisition setup.
Figure 1. Data
Figure Data acquisition
acquisition setup.
setup.

AE AE AE
AE AE AE
SENSOR 3 LEAK SENSOR 2 SENSOR 1
SENSOR 3 LEAK SENSOR 2 SENSOR 1

LEAK
LEAK

AE SENSOR 3
AE SENSOR 3

Figure 2. The testbed for pipeline leak detection.

Figure 2.
Figure 2. The testbed for pipeline leak detection.
2.2. AE Sensors
2.2.
2.2. AE
AE Sensors
Sensors
To acquire AE signals, R15I-AST sensors are used to provide high sensitivity during the data
To
To acquire
acquire AE signals,
signals, R15I-AST
AErecorded R15I-AST sensors
sensors are used to provide
provide high sensitivity
sensitivity during the data
recording and the signals tend to beare used
free of to highcomponents.
low-frequency during
The the data
sensors’
recording
recording and
and the
the recorded
recorded signals
signals tend
tend to
to be
be free
free of
of low-frequency
low-frequency components.
components. The
The sensors’
sensors’
characteristics are summarized in Table 2.
characteristics
characteristics are
are summarized
summarized in in Table
Table 2.
2.
Table 2. Specifications of R15I-AST.
Table 2. Specifications of R15I-AST.
Table 2. Specifications of R15I-AST.
No Parameter Value
No
No Parameter
Parameter Value
Value
1 Peak sensitivity, ref (V/(m/s)) 109 (dB)
11 Peak
Peak sensitivity, ref (V/(m/s)) 109 (dB)
2 Peaksensitivity, ref(V/µbar)
sensitivity, ref (V/(m/s)) 109(dB)
−22 (dB)
232 Peak
Peak sensitivity,
Operating ref
ref (V/µbar)
frequency
sensitivity, range
(V/µbar) −22
−22 (dB)
50–400 (kHz)
(dB)
343 Operating
Resonant frequency
frequency,
Operating range
ref (V/(m/s))
frequency range 50–400
75 (kHz)
50–400 (kHz)
(kHz)
6 Resonant
44 Resonant frequency,
frequency, ref
ref (V/µbar)
(V/(m/s)) 150
75 (kHz)
(kHz)
Resonant Directionality
frequency, ref (V/(m/s)) +/−1.5 75 (kHz)
676 Resonant
Resonant frequency,
frequency, ref
ref (V/µbar)
(V/µbar) 150 (dB)
(kHz)
8 Temperature range −35 to (kHz)
150 75 (◦ C)
77 Directionality
Directionality +/−1.5
+/−1.5 (dB)
(dB)
88 Temperature
Temperature range range −35
−35 to
to 7575 (°C)
(°C)
2.3. Data Record
2.3.
2.3. Data
The Record
Data normal
Record and abnormal states of the system refer to the closed and open positions of the valve
installed
The on the leak, respectively. The data for everyrefer
pair of the
(P, L) combinations are recorded for 2valve
min
Thenormal
normaland andabnormal
abnormalstatesstatesof ofthe
thesystem
system referto to theclosed
closedand andopen
openpositions
positionsof ofthe
the valve
with a sampling
installed frequency of 1 megahertz (MHz)everyafter the water flowcombinations
are stable. Thus, there are a total2
installed on on the
the leak,
leak, respectively.
respectively. The The data
data for
for every pairpair ofof (P,
(P, L)L) combinations are are recorded
recorded for for 2
of
min 72 datasets for the three signal channels in the experimental conditions (3 channels × 3 pressures ×
min with
with aa sampling
sampling frequency
frequency of of11 megahertz
megahertz(MHz)(MHz)after after the
the water
water flow
floware arestable.
stable. Thus,
Thus, there
there are
are
4a leaks × 2 classes). Figure 3 presents mode setting of water flow. In the first stage, the pump is turned
a total
total of
of 72
72 datasets
datasets for for the
the three
three signal
signal channels
channels in in the
the experimental
experimental conditions
conditions (3 (3 channels
channels ×× 33
on, the
pressures leak is deactivated (closed) and the pressure is controlled around 7 Bar. Then, AE signals the
are
pressures ×× 44 leaks
leaks ×× 22 classes).
classes). Figure
Figure 33 presents
presents modemode setting
setting of of water
water flow.
flow. InIn the
the first
first stage,
stage, the
recorded
pump in 2 min. In the second stage, the leak is activated (open). The acquisition device waits for
pumpis isturned
turnedon, on,the
theleak
leakis isdeactivated
deactivated(closed)
(closed)andandthethepressure
pressureis iscontrolled
controlledaround
around77Bar. Bar.Then,
Then,
the
AE flow stabilization and records AE thesignals in 2stage,
min. The leak processactivated
continues, as shown in Figure 3.
AE signals
signals are
are recorded
recorded in in 22 min.
min. InIn the second
second stage, the the leak is is activated (open).
(open). The
The acquisition
acquisition
Figure
device 4 shows the AE signals from channel 2 for leak L3 under three different flow pressures.
device waits
waits for
for the
the flow
flow stabilization
stabilization andand records
records AE AE signals
signals in in 22 min.
min. TheThe process
process continues,
continues, as as
shown
shown in in Figure
Figure 3. 3. Figure
Figure 44 shows
shows the the AE
AE signals
signals from
from channel
channel 22 for for leak
leak L3 L3 under
under three
three different
different
flow
flow pressures.
pressures.
Sensors 2019, 12, x FOR PEER REVIEW 4 of 18
Energies 2019, 12, 1472 4 of 18
Sensors 2019, 12, x FOR PEER REVIEW 4 of 18
Deactivated leak,
Pressure = 7 Bar Pressure = 13 Bar
Deactivated leak,
Pressure = 7 Bar Pressure = 13 Bar

Activated leak
Flow rate

Activated leak
Activated leak
Deactivated leak,
Flow rate

Activated leak
Pressure = 18 Bar Pump OFF,
Deactivated leak,Activated leak deactivated leak
Pressure = 18 Bar Pump OFF,
Pump ON, deactivated leak Activated leak
deactivated leak
Pump ON, deactivated leak

Figure 3. Input flow rate experimented with leak L3.

Figure 3. Input flow rate experimented with leak L3.
Figure 3. Input flow rate experimented with leak L3.

Figure 4. Acoustic emission (AE) signals from channel 2 for leak L3.
Figure 4. Acoustic emission (AE) signals from channel 2 for leak L3.
3. Small Leak Detection Figure Methodology
4. Acoustic emission (AE) signals from channel 2 for leak L3.
3. Small Leak Detection Methodology
The mathematical modeling of the leaks and their symptoms is based on the testbed, as shown in
3. Small
Figure The Leak Detection
1. mathematical Methodology
modeling of the leaks and their symptoms is based on the testbed, as shown
in Figure 1.
The mathematical modeling of the leaks and their symptoms is based on the testbed, as shown
3.1. Symptoms of Leak Presence
in Figure 1.
3.1. Symptoms
There exists of Leak Presencein pipelines, even when they are operating in a healthy state. This activity
AE activity
3.1.There
could Symptoms
be due toofthe
exists Leak
AE Presencein sources,
mechanical
activity pipelines, sucheven as when
pumpsthey or particles
are operating in the in flow, hitting state.
a healthy a pipeline
This
wall, or hydraulic
activityThere
couldexists sources
be due AEtoactivity caused
the mechanical by pressure
in pipelines, sources, pulses
evensuch at
when vortexes
asthey
pumps in
are or the fluid inside
particlesininathe
operating the pipeline
flow, state.
healthy hitting[12].
Thisa
Let assume
pipeline wall,that
or n (i =
hydraulic
i 1, 2) are
sources AE signals
caused acquired
by pressure by sensors
pulses
activity could be due to the mechanical sources, such as pumps or particles in the flow, hitting a at1, 2 in
vortexesthe normal
in the state,
fluid and
inside their
the
mean
pipeline and[12].
pipeline variance are
Letorassume
wall, 0 and
hydraulic thatNsources
n1i ≈(i N = 21,2) Nare
≈caused > 0,AE respectively.
by signals
pressure Whenby
acquired
pulses a small
at sensors
vortexes leak 1,2in occurs
inthe in
thefluid theinside
normal testbed
state, the
pipeline,
andpipeline it
their meanmakes a small
andassume
[12]. Let variance disturbance
aren0i (iand
that in the
N1 ≈are
= 1,2) flow
N2AE around
≈ Nsignals the leak,
> 0, respectively.
acquired by which
When introduces
sensors a small 1,2 in a
leaknew
theoccursAE
normal source
in state,
the
into
andthe
testbed system.
their meanThe
pipeline, it
and previous
makes
variance source
a small ni can
are disturbance
0 and N1 be ≈ Ndeemed
in
2 ≈the 0,asrespectively.
N >flow background
around the When noise;which
leak, asuppose
smallintroduces
leakthatoccurs
thisanoise
new
in the
is
AE uncorrelated
sourcepipeline,
testbed with
into the it the
system.
makesThesource
a small of
previousthe signal
source nin
disturbance obtained
i can
thebe from
deemed
flow the
around leak.
as the Therefore,
background
leak, which the
noise; model for
supposeathat
introduces AE
new
measurements
this noise is from
uncorrelated the sensors
with the for this
source scenario
of the can
signal be given
obtained
AE source into the system. The previous source ni can be deemed as background noise; suppose that as:
from z i =
the s i +
leak.n i (i = 1,
Therefore, 2), where
the s
modeli is
the
for leak
AE AE signal
measurements received
from by
the sensor
sensors i. If
for the
this variance
scenario of z
can i and
be
this noise is uncorrelated with the source of the signal obtained from the leak. Therefore, the model s
giveni is Zas:
i and
z i = S
s ii ,
+ respectively,
n i (i = 1,2), then
where sthe
i is
uncorrelation
theforleak
AE AE between
signal received
measurements n and
ifromby s , Z
i sensor
the i =
sensors S + N
i i. Iffor can
thethis be explored
variance
scenario of can by
zi and setting g
si is Zias:
be given =
and Z /Z
z2i S and it
= i,s1irespectively, is transformed
+ ni (i = 1,2), where then the si is
as follows:
uncorrelation between n and s
the leak AE signal received by sensorZi. If the , Z = S + N can be explored
variance by setting g = Z /Z and it is transformed
S of zi +
and 1 si is Zi and Si, respectively, then the
i i i i 2 1
2 S2 + N /N
asuncorrelation
follows: between ni and si, Z gi =
= SZi + N = can be = 2 by setting
explored . g = Z2/Z1 and it is transformed (1)
1 S1 + N S1 /N + 1
as follows: Z 2 S2 + N S2 / N+ 1
If the background noise ni (i = 1, 2)g is = bandlimited
= = white noise, . then its variance is always N over (1)
Z1Z 2 S1S+2 N + N S1S/2N/ + N1+ 1
its entire frequency range. Next, consider g= =
the measurement = model. when the measurement consists (1) of
If the background
the background noise nnoise and n
thei (i leak
= 1,2) is
signal Zs1 (i S=1 +
bandlimited 1,N2)white
at + 1 then
S1a/frequency
Nnoise, ω. itsIn variance
fact, the is always
spectrum N ofover
the
i i
its entire frequency
If the background range. Next,
noise ni (iconsider
= 1,2) is the measurement
bandlimited whitemodelnoise, when then its the measurement
variance is always consists
N over
its entire frequency range. Next, consider the measurement model when the measurement consists
Energies 2019, 12, 1472 5 of 18

leak noise is a broadband range of frequencies; however, all the components demonstrate an identical
Sensors 2019, 12, x FOR PEER REVIEW 5 of 18
behavior toward the leak phenomenon.
The AE attenuation characteristic from a power law [17], S1 andω.
of the background noise ni and the leak signal si (i = 1,2) at a frequency
S2Inare related by:
fact, the spectrum of the
leak noise is a broadband range of frequencies; however, all the components demonstrate an identical
behavior toward the leak phenomenon. S1 = S2 e−αd . (2)
The AE attenuation characteristic from a power law [17], S1 and S2 are related by:
In (2), α = α(ω) is the attenuation coefficient− din wave propagation which is dependent on
S1 = S 2e . (2)
frequency, d is the distance between sensors 1 and 2, and Si (i = 1, 2) is the leak signal with frequency ω.
By substituting ( )into
In (2),  = (2) is (1),
the and abbreviating
attenuation r = Sin2 /N,
coefficient wavewhich is the signal
propagation to is
which noise ratio measured
dependent on
by sensor 2 at frequency ω, and symbolizing β = e−αd , (1) transforms as follow:
frequency, d is the distance between sensors 1 and 2, and Si (i = 1,2) is the leak signal with frequency
ω.
r + 1r = S / N , which is the signal to noise ratio
By substituting (2) into (1), and abbreviating
g= . 2 (3)
βr + 1
measured by sensor 2 at frequency ω, and symbolizing  = e− d , (1) transforms as follow:
r +to
Next, take the partial derivative of g according 1 r:
g= . (3)
r +1
∂g
Next, take the partial derivative of g according
1−β
to r:
= . (4)
∂rg (βr  1)
1 −+
2
= . (4)
r (  r + 1)
2

Since 0 < β < 1 for 0 < d < ∞, then ∂g/∂r > 0. As a result, g(r) is a monotonically increasing
function Since 0   to1 r.forIf 0r d, r , ,then
according g /),
then g(r r g(r
0 . ).
As Naturally,
a result, g(r)a is a monotonically
normal state has increasing
r = 0 at every
1 2 1 2
function according to r. If r1 ≠ r2, then g(r1) ≠ g(r2). Naturally, a normal state has r = 0 at every frequency
frequency ω, and an abnormal state always has r , 0; thus, the function g(r) is applicable for leak
ω, and an abnormal state always has r ≠ 0; thus, the function g(r) is applicable for leak detection.
detection. Figure 5 shows the dependence of g(r) on r with different β values at a particular frequency.
Figure 5 shows the dependence of g(r) on r with different β values at a particular frequency. It can be
It can be easily observed that all the curves g(r) increase from the normal state when r increases.
easily observed that all the curves g(r) increase from the normal state when r increases.

Figure 5. The dependence of g(r) on r with different β at frequency ω.

Figure 5. The dependence of g(r) on r with different β at frequency ω.
3.2. Robustness of g(r) in Leak Manifestation
3.2. Robustness of g(r) in Leak Manifestation
This section investigates the reliability of leak manifestation using the function g(r) when the
This section investigates the reliability of leak manifestation using the function g(r) when the
level of background noise increases. Gaussian noise is added to the signals to emulate the presence
level of background noise increases. Gaussian noise is added to the signals to emulate the presence
of noise. Suppose
of noise. Supposethat
thatanan amount
amount ofofnoisenoise Δn∆n with with the mean
the mean 0 and 0variance
and variance ∆N is
ΔN is added to added
the noiseto the
noisebackground
background while
while thethe leak
leak signal
signal remains
remains the the same.
same. At this
At this moment,
moment, the background
the background n is n is
noise noise
replaced by n0 = n + ∆n; its variance is N 0 = N + ∆N, and r is replaced by r0 = S /N 0 .
replaced by n′ = n + Δn; its variance is N′ = N + ΔN, and r is replaced by r′ = S2/N′. 2
Setting γ = ∆N/N,
Setting = N / Nproduces
, produces r0 = r /((11 ++ γ
r ' =r/ ) ,),which
whichis is a function
a function of two
of two variables
variables γ). Consider
(r,Consider
(r, γ).
the partial derivative of r’ with respect to
the partial derivative of r’ with respect to γ: γ:

∂r0r = − r r 2 .
'

(5)
= − (1 +  ) 2 . (5)
∂γ (1 + γ)
Energies 2019, 12, 1472 6 of 18

Now, the function g is replaced by:

Sensors 2019, 12, x FOR PEER REVIEW 6 of 18
r+1+γ
g0 = g ( r0 ) = (6)
Now, the function g is replaced by: βr + 1 + γ
r +1+ γ
= g ( r 'by:
and its partial derivative according to γ is ggiven
'
)= (6)
β r +1+ γ
and its partial derivative according to γ∂g (β − 1)r
is0given by:
= . (7)
∂γ∂g ' (βr( β+−11)+
r γ)2
= . (7)
∂γ (βr +1+ γ )
2

If 0 < β < 1 and ∂g0 /∂γ < 0, then g’ decreases when γ increases. In other words, the discrimination
If 0 < β < 1 and ∂g ' / ∂γ < 0 , then g’ decreases when γ increases. In other words, the
quality of g’ becomes poorer as the intensity of the background noise is higher. This characteristic is
discrimination quality 0 of g’ becomes poorer as the intensity of the background noise is higher. This
similar to r0 ; however,  g is more reliable than r0 because the decline in g0 is smaller than the decline in r0 .
characteristic is0 similar to r′;0  however, g′ is more reliable than r′ because the decline in g′ is smaller
Shortening  ∂r in=r′.∆1 ,  ∂g  = ∆2 , and taking the proportion ∆2 produces:
than the decline ∂γ ∂γ ∆1
∂r ' ∂g ' Δ2
Shortening = Δ1 , = Δ 2 , and taking the proportion produces:
∂γ ∂γ ∆2 ( 1 + γ ) 2 Δ1
= (1 − β) (8)
∆1 Δ (βr(1+
+ γ1) + γ)
2 2
2
= (1 − β ) (8)
(βr +1+ γ )
2
Δ1
∆
Obviously, if the parameter β in (10) is selected suitably, << 1 ∀r, γ. As a result, g(r’)
then ∆2
Δ1
varies more Obviously,
slowly than if γ is increasing.
if ther0parameter β in (10) is
It selected
turns outsuitably, 2
thennoise
that if the << 1 ∀r , γ . As increases
background a result, g(r’)
to some
Δ1
extent, the variable r 0 exceeds the limitation of leak discrimination, whereas the function g(r 0 ) still
varies more slowly than r′ if γ is increasing. It turns out that if the noise background increases to
provides
someenough
extent, thedifferentiation.
variable r′ exceeds the limitation of leak discrimination, whereas the function g(r′)
In (8), if β approaches 0, the ratio ∆2/∆1 converges 1 and the variation of the function g is similar
still provides enough differentiation.
to r if the Inbackground noise changes,
(8), if β approaches theΔ2/Δ1
0, the ratio characteristic
converges of g isthe
1 and novariation
longer robust. In contrast,
of the function (3) reveals
g is similar
that iftoβr approaches
if the background 1, g converges
noise changes,1 for
theany r. This means
characteristic of g isthat the function
no longer robust. Ing contrast,
does not(3)manifest
reveals any
that ifstate
abnormal β approaches 1, g converges
of the system. Hence, 1the any r. This βmeans
forparameter shouldthatbethe function
chosen g does not
optimally manifest
to trade off any
between
abnormal state of the system. Hence, the parameter β should
the two above cases so that both high sensitivity and reliability can be achieved.be chosen optimally to trade off between
the two above cases so that both high sensitivity and reliability can be achieved.
3.3. Detection Procedures
3.3. Detection Procedures
This paper proposes an algorithm to detect a small leak in a pipeline based on g(r) function
This paper proposes an algorithm to detect a small leak in a pipeline based on g(r) function
because it can indicate the presence of leak as described in the previous section. Figure 6 shows a
because it can indicate the presence of leak as described in the previous section. Figure 6 shows a
generic framework
generic for for
framework leak detection
leak detectionusing
usingtwo
two approaches: directAE-based
approaches: direct AE-basedandand g(r)-based.
g(r)-based.

Training set

NORMAL
AE signals Frame g(r)- Feature
Classification
division construction extraction
ABNORMAL
Direct AE-based Testing set

Figure6.6.Leak
Figure Leak detection
detection procedures.
procedures.

The The AE signals

AE signals from
from sensors1,1,2,2,and
sensors and33are
are the
the inputs
inputstotothe
theleak detection
leak framework.
detection In theIn the
framework.
direct AE-based method, there is no g(r)-construction block. After dividing the
direct AE-based method, there is no g(r)-construction block. After dividing the AE signals intoAE signals into
frames,
frames, they are provided to the feature extraction block. The feature extraction process is carried out
they are provided to the feature extraction block. The feature extraction process is carried out after the
after the completion of g(r)-construction process.
completion of g(r)-construction process.
3.3.1. Frame Division
3.3.1. Frame Division
The recorded AE signals are segmented into a series of frames by the frame division block. The
The recorded
AE waves AE signals
propagate aredifferent
through segmented into afrom
distances series ofleak
the frames bysensors.
to the the frame division
Thus, block. are
their arrivals The AE
waves propagate
lagged. through
It indicates thatdifferent distances
the frame indexes from the leak
associated withtodifferent
the sensors. Thus,
channels aretheir
not arrivals
exactly are
lagged. It indicates that the frame indexes associated with different channels are not exactly correlation.
Energies 2019, 12, 1472 7 of 18
Sensors 2019, 12, x FOR PEER REVIEW 7 of 18

Hence, at the detection

correlation. Hence, at stage, the position
the detection stage,ofthe theposition
leak is obscure,
of the leak and
is the time of
obscure, andarrival of the
the time of signals
arrival at
theofsensors
the signals at the sensors is unknown. Thus, one way to deal with this problem is to select size.
is unknown. Thus, one way to deal with this problem is to select a reasonable frame a
Figure 7 uses frame
reasonable an example of two
size. Figure signals
7 uses s1 and s2of
an example totwo
explain
signalsthes1method.
and s2 toIn this figure,
explain ∆tlag and
the method. t f rame
In this
the lagΔtime
arefigure, tlag and
and tframe size (in time) of the two signals, respectively. Due to the existence of the lag
frame are the lag time and frame size (in time) of the two signals, respectively. Due
time ∆t
to thelag , a lag part of signal
existence of the lag time s1 cannot be correlated
Δtlag , a lag part of signal with any part
s1 cannot beof signal s2 with
correlated in the same
any partframe index
of signal
i because it has already propagated in frame (i − 1) of s2 . Thus, a formula for the frame size can be
s2 in the same frame index i because it has already propagated in frame (i − 1) of s2. Thus, a formula
defined
for theas:
frame size can be defined as:
t f rame = ∆tlag + text (9)
t frame = Δtlag + text (9)
where text is an amount of time to extend the frame size. Obviously, the bigger text is, the smaller the
where text is an amount of time to extend the frame size. Obviously, the bigger text is, the smaller
lag compared with the remains, which reduces the effect of the lag on the correlation.
the lag compared with the remains, which reduces the effect of the lag on the correlation.

Figure 7. Frame division dependent on lag time.

Figure 7. Frame division dependent on lag time.
Next, the parameter ∆tlag is calculated. The location of the leak in the pipeline is unknown;
Next, the parameter Δ tlag is calculated. The location of the leak in the pipeline is unknown;
however, the leak lies somewhere within the tested pipeline. According to this condition, the following
however, the leak lies somewhere
equations can calculate the maximum within the tested
lag time and this pipeline. According
result is to this condition,
used to calculate the
the reasonable
following
frame size. equations can calculate the maximum lag time and this result is used to calculate the
reasonable frame size. (D + d)
∆tlag = (10)
Δtlag =
( D +Cd )
(10)
In (10), C is the wave speed. According to [19], C AE signals can be propagated in fluid in the
frequencies range
In (10), C isofthe
20wave
kHz to 80 kHz
speed. besides propagating
According to [19], AE signalsthrough
canthebepipe wall in high
propagated frequencies.
in fluid in the
Furthermore, AE signals
frequencies range of leaks
of 20 kHz to 80are
kHzfrom the propagating
besides flow turbulence andthe
through interaction
pipe wallof in particles at the leak
high frequencies.
Furthermore,
point. Thus, AEAE signalsmight
signals of leaks are from
contain thekinds
both flow turbulence and interaction
of propagation. of particles
In other words, the at the leak
wave speed
in point.
waterThus, AE signals
is smaller than in might
solidcontain both[20].
materials kinds of propagation.
Hence, the valueIn ofother words,
C should bethe wave speed
calculated withinthe
water is smaller
propagation thanThe
in water. in solid
wavematerials
speed can[20]. Hence, theas
be calculated value of C[21,22]:
follows should be calculated with the
propagation in water. The wave speed can be calculated as follows [21,22]:
v
K
u
t
CC==  K  (11)
 1+
ρ DpDKp K ψ
 (11)
ρ 1 + e ψ E
 e E 
h  ih    i
ψ =ψ 1/
= 11/ (+ e / Dpp )  11+
+ 22( ee/D
/ Dpp) ((11+ +
u ) (u1)+ 1e /+De/D
1 +e/D )p  p (12)
(12)

where
whereK and ρ are ρtheare
K and volumetric compressibility
the volumetric modulus
compressibility and the
modulus andliquid density
the liquid of the
density ofmedium inside
the medium
the pipeline, e and Dp are the thickness and inner diameter of the pipe, ψ is a factor related to the pipe
inside the pipeline, e and Dp are the thickness and inner diameter of the pipe, ψ is a factor
supporting condition, and u is Poisson’s ratio.
related to the pipe supporting condition, and u is Poisson’s ratio.
Thus, the frame size is counted as:
Thus, the frame size is counted as:
t frame== ((11++
t f rame ζ )ζΔ)t∆t
lag lag (13)(13)

t text
ζ = ζ ext
= . . (14) (14)
Δtlag ∆tlag

InIn (14),
(14), ζ must
ζ must keep
keep thethe lagged
lagged part
part ofof the
the signals,which
signals, whichisisnot
notvery
verylarge
largeasascompared
comparedtotothe
therest
rest of the
of the signal. signal.
Energies 2019, 12, 1472 8 of 18

Sensors 2019, 12, x FOR PEER REVIEW 8 of 18

3.3.2. g(r)-Construction
3.3.2. g(r)-Construction
The quantity g(r) in Section 3.1 is formulated by dividing the variance of one frequency for the
The quantity g(r) in Section 3.1 is formulated by dividing the variance of one frequency for the
leak signal.leak
Insignal.
this section, g(r) vector
In this section, is constituted
g(r) vector fromthe
is constituted from the signal
signal of sensors
of sensors 1 and 2.1The
andproportions
2. The proportions
of the amplitudes over all the frequencies are considered in the g(r) vector (Figure 8).
of the amplitudes over all the frequencies are considered in the g(r) vector (Figure 8).

AE sensor 2 FFT ABS

g(r)
DIVIDER LOG

AE sensor 1 FFT ABS

Figure 8. g(r)-construction.
Figure FFT,fast
8. g(r)-construction. FFT, fast Fourier
Fourier transform.
transform.

In FigureIn 8, time
Figure domain
8, time domain signals
signals areare converted
converted to the domain
to the frequency frequency
by fastdomain by fast Fourier
Fourier transform
(FFT), taking only their amplitudes as components of the divider for
transform (FFT), taking only their amplitudes as components of the divider for every every frequency. After the frequency.
transformation, we have a new signal in the form of g ( r ) containing information about leakage
After the transformation, we have a new signal in the form of g(r) containing information about
symptoms.
leakage symptoms.
3.3.3. Feature Extraction
3.3.3. Feature Extraction
In this study, the three features given in Table 3 are used to compare the performance of the
direct
In this AE-based
study, methodfeatures
the three with that given
of the g(r)-based
in Tableone. These
3 are features
used are selectedthe
to compare because of their
performance of the
effectiveness in both time and frequency domains.
direct AE-based method with that of the g(r)-based one. These features are selected because of their
effectiveness in both time and frequency domains. Table 3. Typical features.

No Feature Definition Abbreviation

Table 3. Typical features. N

1 Root mean square RMS = x /N 2

i RMS
i =1

No Feature Definition
STE =  x
Abbreviation
N

2 Short time energy STE

2

s i =1

N
N
μ =  x / NP
3 mean
Average amplitude AVA
1 Root square x2i /N RMS
i

RMS = i =1

i=1
3.3.4. Classification N
x2i
P
2 Short time energy STE = STE
KNN is a popular kernel function used to identify instancesi=1 belonging to different classes during
N
the diagnosis.
3 The theory of KNN
Average amplitudeis presented clearly
µ=
Pin |x[18,23]. This paper uses
AVAthe KNN-based
i |/N
classifier to solve a binary classification problem, i.e., iwhether
=1 a sample belongs to the normal or
abnormal conditions of the pipeline.
3.3.4. Classification
4. Experimental Results
KNN is aFor popular kernel
a comparison function
between usedAE-based
the direct to identify instancesapproaches,
and g(r)-based belongingthetonecessary
different classes during
dataset
parameters
the diagnosis. The are listed of
theory in Table
KNN 4, where P is one ofclearly
is presented the pressures {P1, P2, This
in [18,23]. P3} and L is one
paper of the
uses leaks
the KNN-based
{L1, L2, L3, L4} defined in Section 2.1.
classifier to solve a binary classification problem, i.e., whether a sample belongs to the normal or
abnormal conditions of the pipeline. Table 4. Dataset parameters.

Parameters Training Testing

4. Experimental Results
Sample frequency (MHz) 1 1
For a comparison between the Window type
direct AE-based andHanning Hanning
g(r)-based approaches, the necessary dataset
Window size (points) 3000 3000
parameters are listed in Table 4, where P is one of the pressures {P1, P2, P3} and L is one of the leaks
Overlapping of window 50% 50%
{L1, L2, L3, L4} defined in Section
{P, L,2.1.
‘NORMAL’} (frame) 6665 73,331
{P, L, ‘ABNORMAL’} (frame) 6665 73,331
Table 4. Dataset parameters.

Parameters Training Testing

Sample frequency (MHz) 1 1
Window type Hanning Hanning
Window size (points) 3000 3000
Overlapping of window 50% 50%
{P, L, ‘NORMAL’} (frame) 6665 73,331
{P, L, ‘ABNORMAL’} (frame) 6665 73,331
Sensors 2019, 12, x FOR PEER REVIEW 9 of 18
Energies 2019, 12, 1472 9 of 18

After the AE signals are divided into frames, they are transferred directly to the feature
extraction
After theblockAEin the direct
signals AE-based
are divided intomethod, and are
frames, they to the g(r)-construction
transferred directly toblock in the extraction
the feature proposed
method (Figure 6). Figure 9 illustrates the AE signals of the pair (P1,
block in the direct AE-based method, and to the g(r)-construction block in the proposed method L1), and Figure 10 presents the
signals 6).
(Figure in aFigure
frame9index.
illustrates the AE signals of the pair (P1, L1), and Figure 10 presents the signals in a
The
frame index. R15I-AST AE sensors used in the experiment are manufactured by MISTRAS in accordance
withThe the industrial
R15I-AST AE standards,
sensorsand usedthey
in thecanexperiment
capture signals with low levelbyasMISTRAS
are manufactured shown in in Figures 9 and
accordance
10. These signals demonstrate that noise levels of all the channels are nearly
with the industrial standards, and they can capture signals with low level as shown in Figures 9 and 10. same in the normal
condition,
These signalsbutdemonstrate
differ from those of thelevels
that noise abnormalof all condition.
the channels Channel 1 is same
are nearly mounted at normal
in the the furthest place
condition,
from
but the leak,
differ from and
thosetheof signal level atcondition.
the abnormal this point Channel
is the lowest, which resembles
1 is mounted the noise,
at the furthest placewhile the
from the
two others get higher amplitudes for being closer to the leak. Moreover, such
leak, and the signal level at this point is the lowest, which resembles the noise, while the two others get a low level of signals is
prone to noise in industry. Therefore, a classification algorithm relying on absolute
higher amplitudes for being closer to the leak. Moreover, such a low level of signals is prone to noise in levels is unreliable
if it is only
industry. trained by
Therefore, a limited number
a classification of datasets.
algorithm relying In onaabsolute
real application, the leak position
levels is unreliable is obscure;
if it is only trained
AE signals can vary between low level like noise (when AE sensors are
by a limited number of datasets. In a real application, the leak position is obscure; AE signals canfar from the leak) and higher
vary
levels (when they are closer to the leak). To address the issue, a classifier
between low level like noise (when AE sensors are far from the leak) and higher levels (when they are is trained by the quantity
g(r) formed
closer to the by AETo
leak). signals
addressfromthecritical
issue, aconditions;
classifier isfor example,
trained by the when sensors
quantity g(r) are mounted
formed by AEfar from
signals
the leak, and pressure is low. The proposed model can detect a leak correctly
from critical conditions; for example, when sensors are mounted far from the leak, and pressure is low. even if the environment
fluctuates
The proposed widely.
model can detect a leak correctly even if the environment fluctuates widely.

Figure 9. Datasets of the pair (P1, L1).

Figure 9. Datasets of the pair (P1, L1).
The g(r)-construction block in Figure 8 converts the signals of channels 1 and 2 from the time
Thetog(r)-construction
domain block inand
the frequency domain Figure
then8 transforms
converts the signals
into of channels
the quantity g(r).1 Figure
and 2 from the time
11 presents a
domain
frame of to
g(r)the frequency
in which domaing(r)
the normal andfluctuates
then transforms
around 0,into
butthe
the quantity
abnormalg(r).
g(r) Figure 11 presents
has a different a
trend.
frame of g(r) in which the normal g(r) fluctuates around 0,
This behavior is independent of distance and resistant to noise. but the abnormal g(r) has a different trend.
This behavior is independent of distance and resistant to noise.
Energies 2019, 12, 1472 10 of 18
Sensors2019,
Sensors 2019,12,
12,xxFOR
FORPEER
PEERREVIEW
REVIEW 10ofof18
10 18

NORMAL ABNORMAL
CHANNEL 1
CHANNEL 1
CHANNEL 2
CHANNEL 2
CHANNEL 3 NORMAL ABNORMAL
CHANNEL 3

Figure10.
Figure 10.AAframe
frameofofdatasets
datasetsofofthe
thepair
pair(P1,
(P1,L1).
L1).

NORMAL ABNORMAL
NORMAL ABNORMAL

Figure
Figure11. AAframe
11.A of
frameof g(r)
ofg(r) for
g(r)for datasets
fordatasets ofofthe
datasetsof pair
thepair (P1,
(P1,L1).
pair(P1, L1).
Figure 11. frame the L1).
4.1. Effectiveness of the g(r)-Based Approach Compared with the Direct AE-Based Method
4.1.Effectiveness
4.1. Effectivenessofofthe
theg(r)-Based
g(r)-BasedApproach
ApproachCompared
Comparedwith
withthe
theDirect
DirectAE-Based
AE-BasedMethod
Method
The effectiveness of the g(r)-based approach compared with the direct AE-based method is
The effectiveness
The effectiveness ofof the
the g(r)-based
g(r)-based approach compared
compared with with the
the direct AE-based
AE-based method
method isis
illustrated by state scattering in 3D space, approach
as shown in Figures 12–15. Thedirect
plot uses the features and a
illustrated
illustrated by state scattering in 3D space, as shown in Figures 12–15. The plot uses the features and
separabilitybycomparison
state scattering
that in 3D on
relies space, as shown in Figures
Kullback-Leibler 12–15. The
(KL) distances [18]plot uses the
between thefeatures and
two classes
aseparability
separabilitycomparison
aunder comparisonthat thatrelies
relieson
onKullback-Leibler
Kullback-Leibler(KL)(KL)distances
distances[18]
[18]between
betweenthe
thetwo
twoclasses
classes
various experimental conditions. The KL distance is given as follows:
under various experimental conditions. The KL distance is given
under various experimental conditions. The KL distance is given as follows: as follows:
dkldd=kl==
DDD ++DD2121
1212+ (15)
(15)
(15)
kl 12 21

( x(x|w
w ) )
) lnlnpp( xxww1 )1 1
pp
x( xww11)1 )
pp(p(x|w
X
D12 ==
D12 =D (16)
(16)
(16)
12 ln
( (wx|w
p ( xp 2) 2 )
2)

pp( x(x|w
w )2 )
lnp ( x w2 2)
X
 ( x( xww222) ln
D21== ppp(x|w
D21 D= )) ln (17)
(17)
(17)
pp x( xww1 )1
( )
21 p
( x|w
1 )
whereww1w,1,w
, 2wware x = [xx1=, x[2x, .,.x. , ,...,
xn ]TxisT]Ttheis
where
where 1
features for 2
n data
2
the the
are
are
twotwo
the
classes
two
(i.e.,(i.e.,
classes
classes
NORMAL
(i.e., NORMAL
NORMAL
andand
ABNORMAL)
and ABNORMAL)
ABNORMAL)
frames, and p is the probability density function.
andand
and x = [ x 1 ,
1 x 2 ,...,
2 xn n] is
thefeatures
the featuresforfornndata
dataframes,
frames,and
andppisisthetheprobability
probabilitydensity
densityfunction.
function.
Figures 12–15 show state scattering based on the features extracted from direct AE signals and g(r).
Figures 12–15 show state scattering based on the features
Figures 12–15 show state scattering based on the features extracted extractedfrom
fromdirect
directAE AEsignalssignalsand and
g(r).
g(r).
Sensors 2019, 12, x FOR PEER REVIEW 11 of 18
Energies 2019, 12, 1472 11 of 18
Sensors 2019, 12, x FOR PEER REVIEW 11 of 18

Figure 12. State scattering based on features extracted from the AE signal on channel 1 (CH1).
Figure 12. State scattering based on features extracted from the AE signal on channel 1 (CH1).
Figure 12. State scattering based on features extracted from the AE signal on channel 1 (CH1).

Figure 13. State scattering based on features extracted from the AE signal of channel 2 (CH2).
Figure 13. State scattering based on features extracted from the AE signal of channel 2 (CH2).
Figure 13. State scattering based on features extracted from the AE signal of channel 2 (CH2).
Table 5 shows the KL distances for each pair of pressure, leak (P, L) values. Since the KL distances
depend on features and their values can be relatively large, this study uses a logarithm calculation
to convert the distances into dB for demonstration. Obviously, the KL distances based on features
extracted directly from the AE signals vary in a wide range. In contrast, the KL distances calculated
Energies 2019, 12, 1472 12 of 18

using features extracted from g(r) remain around a certain value for each type of feature under diverse
conditions. Moreover, the average values also demonstrate that the class separability of the g(r)-based
approach is greater
Sensors 2019, 12, x FORthan
PEERthat of the direct AE-based one. From the state scattering and the KL distances,
REVIEW 12 of 18
it isSensors
clearly observed
2019, the g(r)-based approach is more effective than the direct AE-based12
that REVIEW
12, x FOR PEER one.
of 18

Figure 14. State scattering based on features extracted from the AE signal of channel 3 (CH3).
Figure 14. State scattering based on features extracted from the AE signal of channel 3 (CH3).
Figure 14. State scattering based on features extracted from the AE signal of channel 3 (CH3).

Figure 15.State
Figure15. Statescattering
scattering based on features
based on featuresextracted
extractedfrom g(r).
fromg(r).
Figure 15. State scattering based on features extracted from g(r).
Table 5 shows the KL distances for each pair of pressure, leak (P, L) values. Since the KL distances
Table
depend on 5features
shows theandKL distances
their valuesfor
caneach pair of pressure,
be relatively leak
large, this (P, L)uses
study values. Since thecalculation
a logarithm KL distances
to
depend on features and their values can be relatively large, this study uses a logarithm
convert the distances into dB for demonstration. Obviously, the KL distances based on features calculation to
convert the distances into dB for demonstration. Obviously, the KL distances based on
extracted directly from the AE signals vary in a wide range. In contrast, the KL distances calculated features
extracted
using directly
features from the
extracted AE signals
from vary around
g(r) remain in a widea range.
certainIn contrast,
value the KL
for each distances
type calculated
of feature under
using features
diverse extracted
conditions. from the
Moreover, g(r)average
remainvalues
aroundalso
a certain value that
demonstrate for each typeseparability
the class of feature under
of the
diverse conditions. Moreover, the average values also demonstrate that the class separability of the
Energies 2019, 12, 1472 13 of 18

Table 5. Kullback-Leibler distances in decibels (dB) (dkl (dB) = 10*log10 (dkl )).

CH1 CH2 CH3 g(r)

P L
RMS STE AVA RMS STE AVA RMS STE AVA RMS STE AVA
P1 L1 11 1 16 1 22 −10 8 25 2 59 112 60
P1 L2 9 −1 14 11 42 5 16 44 11 63 118 63
P1 L3 −6 13 2 3 34 −6 8 35 4 70 109 70
P1 L4 −2 6 5 11 44 6 17 46 12 63 110 63
P2 L1 0 24 −11 12 42 8 16 42 12 70 102 71
P2 L2 2 36 −7 15 58 11 21 61 17 73 104 73
P2 L3 3 29 −4 24 65 20 29 67 25 60 116 60
P2 L4 4 33 −2 13 49 8 18 51 14 68 110 68
P3 L1 14 48 10 23 64 19 27 65 23 72 101 73
P3 L2 10 48 5 17 62 12 22 65 18 76 99 77
P3 L3 7 −3 11 5 36 0 9 35 4 73 107 74
P3 L4 6 37 1 12 48 8 19 52 15 69 108 70
MEAN 5 23 3 12 47 7 18 49 13 68 108 69

4.2. Classifier Training and Testing

4.2.1. Training
This paper uses datasets of pressure and leak {P, L} pairs to train the KNN classifier used in the
two approaches. Table 6 presents the performance of the two approaches. The P4 column in the
table represents the combination of datasets with all the pressures and leaks. Although Table 6 shows
high accuracies with both methods, their performances become different when applying them under
different conditions, as shown in the next section.

Table 6. Trained classifiers and their accuracy (%).

Direct AE-Based g(r)-Based

P1 P2 P3 P4 P1 P2 P3 P4
L1 97.50 100.0 100.0 98.80 100.0 100.0
L2 100.0 100.0 100.0 100.0 100.0 100.0
99.00 100.0
L3 100.0 100.0 100.0 100.0 100.0 100.0
L4 100.0 100.0 100.0 100.0 100.0 100.0

4.2.2. Cross Testing

The previous section trained 12 classifiers using datasets for conditions corresponding to each
pair {Pi , Lj } for each method: direct AE-based and g(r)-based. This section uses datasets in conditions
{Pm , Ln } to test them, where i, m ∈ {1, 2, 3} and j, n ∈ {1, 2, 3, 4}. The purpose of the test is to verify the
reliability of each approach if the training is carried out using only some of the datasets. Table 7 shows
the results of the experiment carried under this configuration. In each table, the trained classifiers are
symbolized in the columns, and the test datasets are listed in the rows.
It is obvious from Tables 7–10 that when the testing datasets use the same conditions as the training
datasets, the accuracies reach 100% or nearly 100%. However, if the testing conditions differ from the
training conditions, the accuracies are not promising, and misclassification can occur when using the
direct AE-based method. In Table 8, an accuracy of 00.00% indicates misclassification. In the same
situation; however, the g(r)-based approach identifies leaks without misclassification. Furthermore,
the average accuracy of the g(r)-based approach is higher than that of the direct AE-based method in
every test.
Energies 2019, 12, 1472 14 of 18

Table 7. Accuracy (%) of classifiers trained with leak L1.

Direct AE-Based g(r)-Based

P1 P2 P3 P1 P2 P3
P1 98.55 69.64 62.66 99.78 99.82 99.82
L1 P2 99.54 99.98 65.85 99.82 99.88 99.88
P3 96.71 99.94 100.0 99.88 99.82 99.80
P1 99.88 100.0 64.45 99.98 99.76 99.76
L2 P2 98.23 99.86 100.0 100.0 99.88 99.86
P3 99.76 100.0 100.0 99.98 99.96 99.96
P1 100.0 100.0 62.91 99.96 99.90 99.88
L3 P2 99.42 99.96 67.25 100.0 99.82 99.82
P3 99.32 99.98 82.05 99.76 99.82 99.84
P1 99.96 64.27 62.94 99.98 99.90 99.88
L4 P2 99.80 99.94 63.18 99.94 99.96 99.94
P3 99.70 99.94 63.83 99.19 99.84 99.84
99.24 94.46 74.59 99.86 99.86 99.86
MEAN
89.43 99.86

Table 8. Accuracy (%) of classifiers trained with leak L2.

Direct AE-based g(r)-Based

P1 P2 P3 P1 P2 P3
P1 65.77 62.54 00.00 86.92 99.54 99.90
LS1 P2 100.0 63.25 62.69 62.91 99.52 99.92
P3 100.0 77.53 64.53 62.65 99.58 99.92
P1 100.0 62.65 62.51 100.0 99.98 99.84
LS2 P2 99.94 100.0 99.36 70.40 100.0 99.92
P3 100.0 100.0 100.0 63.00 99.96 99.98
P1 100.0 62.58 62.53 98.33 99.88 99.92
LS3 P2 99.96 64.15 62.91 98.06 100.0 99.92
P3 100.0 62.99 62.74 98.06 99.64 99.74
P1 63.97 62.58 00.00 64.73 99.94 99.98
LS4 P2 99.98 62.74 62.58 98.56 99.84 99.96
P3 99.96 63.04 62.68 95.72 98.95 99.52
94.13 70.34 58.54 83.28 99.74 99.88
MEAN
74.34 94.30

Table 9. Accuracy (%) of classifiers trained with leak L3.

Direct AE-Based g(r)-Based

P1 P2 P3 P1 P2 P3
P1 70.51 73.06 63.19 97.96 95.10 97.83
LS1 P2 99.98 99.96 76.28 81.45 66.99 79.70
P3 99.92 99.88 100.0 76.79 64.07 74.84
P1 100.0 100.0 99.82 99.98 100.0 99.98
LS2 P2 99.86 99.80 100.0 99.90 99.64 99.88
P3 100.0 100.0 100.0 98.95 78.53 98.64
P1 100.0 100.0 63.87 99.32 99.07 99.30
LS3 P2 99.96 99.96 72.89 99.98 99.70 99.96
P3 99.98 99.96 100.0 99.21 98.85 99.13
P1 64.34 64.60 63.52 98.50 92.05 98.42
LS4 P2 99.94 99.92 100.0 99.52 99.27 99.50
P3 99.94 99.94 80.45 97.83 97.35 97.79
94.54 94.76 85.00 95.78 90.89 95.41
MEAN
91.43 94.03
Energies 2019, 12, 1472 15 of 18

Table 10. Accuracy (%) of classifiers trained with leak L4.

Direct AE-Based g(r)-Based

P1 P2 P3 P1 P2 P3
P1 97.00 63.79 67.29 99.78 98.08 98.33
LS1 P2 99.72 99.52 100.0 99.88 82.62 86.62
P3 97.60 100.0 99.98 99.96 77.92 82.90
P1 99.94 100.0 100.0 99.98 99.98 99.98
LS2 P2 98.64 100.0 99.94 100.0 99.90 99.94
P3 99.86 100.0 100.0 99.98 99.03 99.36
P1 100.0 98.62 100.0 99.98 99.32 99.42
LS3 P2 99.62 88.80 99.98 100.0 99.98 100.0
P3 99.56 100.0 100.0 99.78 99.21 99.23
P1 99.96 63.72 64.06 100.0 98.52 98.93
LS4 P2 99.80 100.0 99.94 99.96 99.54 99.60
P3 99.80 99.44 99.94 99.30 97.87 97.96
99.29 92.82 94.26 99.88 96.00 96.86
MEAN
95.46 97.58

4.2.3. Testing Classifiers Trained by Combined Datasets

To detect a fault, a classifier that depends on data must be trained using many datasets representing
various experimental conditions to provide high reliability. Table 11 shows accuracies of the two
approaches when using the dataset of each pair {P, L}.

Table 11. Accuracy (%) of classifiers of two approaches by combined datasets.

Direct AE-Based g(r)-Based

P1 P2 P3 P1 P2 P3
L1 96.49 99.72 97.75 99.94 100.0 100.0
L2 99.96 98.76 99.88 99.90 100.0 100.0
L3 100.0 99.66 99.64 100.0 99.92 99.96
L4 99.96 99.80 99.80 100.0 100.0 99.90
99.10 99.48 99.27 99.96 99.98 99.97
MEAN
99.28 99.97

The average accuracies are 99.28 and 99.97 for the direct AE-based and g(r)-based approaches,
respectively, as shown in Table 11. Thus, a model trained by many datasets can work in varied
conditions. However, the g(r)-based method is more efficient than the direct AE-based because it still
achieves the expected accuracy with only a small number of training datasets.

4.2.4. Evaluating Two Approaches Using Combined Datasets with Added Noise
To evaluate robustness of the proposed method, we add Gaussian noise to the acquired signals to
increase the intensity of background noise. The parameter γ is counted in (dB) for each channel of the
three AE sensors. Figure 16 illustrates the signals from channel 2 before and after added noise with
γ = 0 (dB). This evaluation is conducted using the KNN classifier trained by the combination of all
datasets. The testing accuracies are shown in Table 12.
Energies 12,x1472
2019,12,
Sensors 2019, FOR PEER REVIEW 16 of
16 of 18
18

Before adding noise

NORMAL ABNORMAL
mV

sample sample
After adding noise with γ = 0 dB
NORMAL ABNORMAL
mV

sample sample

Figure 16. AE on channel 2 before/after adding noise.

Figure 16. AE on channel 2 before/after adding noise.
Table 12. Accuracy with added noise.
Table 12. Accuracy with added noise.
Model
γ (dB) Model
γ (dB) Direct AE-Based g(r)-Based
Direct AE-Based g(r)-Based
−30 99.28 99.97
−30
−20 99.28
99.28 99.97
99.97
−20
−10 99.28
99.28 99.97
99.97
−10
−5 99.28
99.25 99.97
99.96
−50 99.08
99.25 99.83
99.96
5 00.00 98.46
0 99.08 99.83
10 00.00 94.38
5
15 00.00
00.00 98.46
88.83
10
20 00.00
00.00 94.38
76.90
15
25 00.00
00.00 88.83
70.74
30
20 00.00
00.00 66.96
76.90
35 00.00 00.00
25 00.00 70.74
30 00.00 66.96
Table 12 shows that if γ is small, 35 the accuracies
00.00 of the two00.00 methods are approximately the same
as in Table 11. When γ increases, the accuracy of both approaches decreases, and at a certain value,
they Table
cannot12guarantee classification.
shows that if γ is small,However, the direct
the accuracies AE-based
of the method
two methods arefails in classification
approximately the when
same
thein
as added
Tablenoise reaches
11. When 5 (dB), whereas,
γ increases, the accuracy of bothg(r)-based
the proposed approaches method worksand
decreases, untilatγareaches
certain35 (dB).
value,
Thus, the g(r)-based
they cannot approach
guarantee producesHowever,
classification. a classifierthethat can detect
direct a small
AE-based leak more
method failsrobustly than the
in classification
direct the
when AE-based
added classifier.
noise reaches 5 (dB), whereas, the proposed g(r)-based method works until γ reaches
35 (dB). Thus, the g(r)-based approach produces a classifier that can detect a small leak more robustly
5. Conclusions
than the direct AE-based classifier.
AE-based techniques have the advantage of detecting the physical changes within a structure
5. Conclusions
such as a water pipeline with high accuracy. Due to the complication of the AE phenomenon,
model-based
AE-based fault diagnosis
techniques of the
have thewater pipeline
advantage of isdetecting
difficult,the
whereas a data-driven
physical fault diagnosis
changes within a structureis
relatively
such easy to
as a water implement.
pipeline However,
with high if the
accuracy. training
Due to therelies on extracting
complication of the features directly from
AE phenomenon, AE
model-
signals, the resulting classifier is unreliable because such data cannot reflect all
based fault diagnosis of the water pipeline is difficult, whereas a data-driven fault diagnosis is possible information
under divers
relatively easyworking conditions.
to implement. However, This if
paper introduced
the training reliesanonintermediate step in directly
extracting features a water from
pipeline
AE
fault diagnosis framework in which the signals are preprocessed before the extraction
signals, the resulting classifier is unreliable because such data cannot reflect all possible information of features.
under divers working conditions. This paper introduced an intermediate step in a water pipeline
Energies 2019, 12, 1472 17 of 18

The preprocessing step calculates an intermediate quantity g(r) by applying an attenuation equation
to the acquired AE signals. The quantity g(r) is more stable than the original AE signals in leakage
identification. The results showed that the g(r)-based approach achieves higher accuracy and higher
reliability than the direct AE-based method. This approach can be applied in a variety of systems
because the g(r) function always possesses the same behavior reflecting leaks in a system. This paper
also proposed a way to select a reasonable window size when dividing signal frames. The proposed
technique uses the speed of acoustic emission signal to calculate a possible maximum lag time between
signals on the sensor channels. From that computation, the frame size can be adjusted appropriately.

Author Contributions: All authors contributed equally to the conception of the idea, as well as implementing
and analyzing the experimental results, and writing the manuscript.
Funding: This work was supported by the Korea Institute of Energy Technology Evaluation and Planning
(KETEP) and the Ministry of Trade, Industry & Energy (MOTIE) of the Republic of Korea (No. 20172510102130).
This research was also supported by Basic Science Research Program through the National Research Foundation
of Korea (NRF) funded by the Ministry of Education(2016R1D1A3B03931927).
Conflicts of Interest: The authors declare no conflict of interest.

References
1. Chan, T.K.; Chin, C.S.; Zhong, X. Review of Current Technologies and Proposed Intelligent Methodologies
for Water Distributed Network Leakage Detection. IEEE Access 2018, 6, 78846–78867. [CrossRef]
2. Datta, S.; Sarkar, S. A review on different pipeline fault detection methods. J. Loss Prev. Process Ind. 2016,
41, 97–106. [CrossRef]
3. Puust, R.; Kapelan, Z.; Savic, D.A.; Koppel, T. A review of methods for leakage management in pipe networks.
Urban Water J. 2010, 7, 25–45. [CrossRef]
4. Murvay, P.-S.; Silea, I. A survey on gas leak detection and localization techniques. J. Loss Prev. Process Ind.
2012, 25, 966–973. [CrossRef]
5. Martini, A.; Troncossi, M.; Rivola, A. Vibroacoustic Measurements for Detecting Water Leaks in Buried
Small-Diameter Plastic Pipes. J. Pipeline Syst. Eng. Pract. 2017, 8, 04017022. [CrossRef]
6. Li, S.; Song, Y.; Zhou, G. Leak detection of water distribution pipeline subject to failure of socket joint based
on acoustic emission and pattern recognition. Measurement 2018, 115, 39–44. [CrossRef]
7. Mandal, S.K.; Chan, F.T.S.; Tiwari, M.K. Leak detection of pipeline: An integrated approach of rough set
theory and artificial bee colony trained SVM. Expert Syst. Appl. 2012, 39, 3071–3080. [CrossRef]
8. Xiao, Q.; Li, J.; Bai, Z.; Sun, J.; Zhou, N.; Zeng, Z. A Small Leak Detection Method Based on VMD Adaptive
De-Noising and Ambiguity Correlation Classification Intended for Natural Gas Pipelines. Sensors 2016,
16, 2116. [CrossRef] [PubMed]
9. Nicola, M.; Nicola, C.I.; Vintila, A.; Hurezeanu, I.; Duta, M. Cross-Correlation, Acoustic emission, Leakage
detection, LabVIEW environment. J. Mech. Eng. Autom. 2018, 8, 59–67.
10. Sun, J.; Wen, J. Target location method for pipeline pre-warning system based on HHT and time difference of
arrival. Measurement 2013, 46, 2716–2725. [CrossRef]
11. Martini, A.; Troncossi, M.; Rivola, A. Leak Detection in Water-Filled Small-Diameter Polyethylene Pipes by
Means of Acoustic Emission Measurements. Appl. Sci. 2016, 7, 2. [CrossRef]
12. Antaki, G. Piping and Pipeline Engineering: Design, Construction, Maintenance, Integrity, and Repair; Dekker
Mechanical Engineering; CRC Press: Boca Raton, FL, USA, 2003; Volume 159, ISBN 978-0-8247-0964-8.
13. Rienstra, S.W.; Hirschberg, A. An Introduction to Acoustics. 298. Available online: https://www.win.tue.nl/
~{}sjoerdr/papers/boek.pdf (accessed on 2 December 2018).
14. Martini, A.; Rivola, A.; Troncossi, M. Autocorrelation Analysis of Vibro-Acoustic Signals Measured in a Test
Field for Water Leak Detection. Appl. Sci. 2018, 8, 2450. [CrossRef]
15. He, P. Simulation of ultrasound pulse propagation in lossy media obeying a frequency power law. IEEE Trans.
Ultrason. Ferroelectr. Freq. Control 1998, 45, 114–125. [PubMed]
16. Szabo, T.L.; Wu, J. A model for longitudinal and shear wave propagation in viscoelastic media. J. Acoust.
Soc. Am. 2000, 107, 2437–2446. [CrossRef] [PubMed]
Energies 2019, 12, 1472 18 of 18

17. Chen, W.; Holm, S. Modified Szabo’s wave equation models for lossy media obeying frequency power law.
J. Acoust. Soc. Am. 2003, 114, 2570. [CrossRef] [PubMed]
18. Koutroumbas, K.; Theodoridis, S. Pattern Recognition, 4th ed.; Academic Press: New York, NY, USA, 2008.
19. BS EN 15856 Non-Destructive Testing—Acoustic Emission—General Principles of AE Testing for the
416 Detection of Corrosion with Metallic Surrounding Filled with Fluid 2010. Available online: https:
//shop.bsigroup.com/ProductDetail/?pid=000000000030165681 (accessed on 2 December 2018).
20. Measurement & Control. Sound Speed & Pipe Size Data 2014. Available online: https://www.rshydro.co.uk/
files/SNSP-and-Pipe-Data.pdf (accessed on 2 December 2018).
21. Twyman, J. Wave speed calculation for water hammer analysis. Obras y Proyectos 2016, 20, 86–92. [CrossRef]
22. Watters, G.Z. Analysis and Control of Unsteady Flow in Pipelines; Butterworths: Boston, MA, USA, 1984;
ISBN 0-250-40492-3.
23. Devroye, L.; Györfi, L.; Lugosi, G. A Probabilistic Theory of Pattern Recognition; Springer: New York, NY, USA,
1996; ISBN 978-0-387-94618-4.

© 2019 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access
article distributed under the terms and conditions of the Creative Commons Attribution
(CC BY) license (http://creativecommons.org/licenses/by/4.0/).