Flow Measurement and Instrumentation 79 (2021) 101960

Contents lists available at ScienceDirect

Flow Measurement and Instrumentation

journal homepage: www.elsevier.com/locate/flowmeasinst

Identification of gas-liquid two-phase flow patterns in a horizontal pipe

based on ultrasonic echoes and RBF neural network
Fachun Liang a, b, *, Yue Hang a, Hao Yu a, Jifeng Gao c
a
College of Pipeline and Civil Engineering in China University of Petroleum, Qingdao, 266580, China
b
Key Laboratory of Marine Spill Oil Identification and Damage Assessment Technology, Qingdao, 266033, China
c
Sinopec Petroleum Engineering Zhongyuan Corporation, Puyang, 457001, China

A R T I C L E I N F O A B S T R A C T

Keywords: This paper proposes a novel flow pattern identification method using ultrasonic echo signals within the pipe wall.
Two-phase flow A two-dimensional acoustic pressure numerical model is established to investigate the ultrasonic pulse trans­
Flow pattern mission behavior between the wall-gas and wall-liquid interface. Experiments were also carried out at a hori­
Ultrasonic echo
zontal air-water two-phase flow loop to measure the ultrasonic echo pulse signals of stratified flow, slug flow,
and annular flow. It is interesting to find that the attenuation of the ultrasonic pulse at the wall-liquid interface is
faster than the attenuation at the wall-gas interface. An RBF neural network is constructed for online flow pattern
identification. The normalized envelop area and the area ratios of the echo spectrum are selected as the input
parameters. The results show that the stratified flow, slug flow, and annular flow can be identified with an
accuracy of 94.0%.

1. Introduction The most widely used electrical methods for flow pattern identifi­
cation are the conductance probe and the capacitance probe method.
Gas-liquid two-phase flow exists widely in nature and many indus­ This measurement method has a strong ability to detect local flow
trial applications [1]. The distribution of gas and liquid, i.e. the flow characteristics. However, the probe is directly in contact with the fluid,
pattern, greatly affects the two-phase flow pressure loss and the heat and it will cause a certain degree of interference of the fluid flow. Besides,
mass transfer behavior [2]. Flow pattern identification is essential for the probe is easy to be contaminated, which affects the accuracy of the
the development of predictive models and the efficient and safe opera­ measurement results [6,7].
tion of the gas-liquid two-phase flow system. Many methods have been The principle of the ray method is that the radiation will attenuate
developed to measure the flow pattern, which can be roughly classified when it passes through the medium, and different media have different
into the optical method, electrical method, ray method, and ultrasonic degrees of attenuation. This method is a non-contact measurement and
method [3]. does not affect the flow inside the pipe, but its application is restricted in
The visual observation method can be viewed as the simplest optical some cases due to safety and economic considerations [8].
method by observing the fluid flow from the transparent window. The tomography method can obtain the flow structure over the entire
However, it is only suitable for low-velocity conditions. Besides, the cross-section according to the measured gamma-ray, conductance, or
visual observation method is strongly subjective and lacks a measurable capacitance signal. Gamma-ray tomography, electrical resistance to­
quantity. The high-speed camera method is a very popular optical mography (ERT), and electrical capacitance tomography (ECT) are
method today, which can capture the image of gas and liquid distribu­ widely used in flow pattern identification in recent years [9–11]. Con­
tion with more than 1000 frames per second. The limitation is that the ventional tomography cannot be used for rapid transient multiphase
light will generate multiple reflection and refraction on the phase flow because of its low imaging speed [12]. Wire mesh sensor (WMS) is a
interface, which affects the imaging resolution. In addition, the amount novel multiphase flow measurement instrument developed by Prasser
of image information acquired by a high-speed camera is too large, and et al. [13]. The WMS consists of two parallel wire grids positioned
even difficult to analyze and process [4,5]. orthogonally across the pipe section. The gas-liquid distribution is

* Corresponding author. College of Pipeline and Civil Engineering in China University of Petroleum, Qingdao, 266580, China.
E-mail address: [email protected] (F. Liang).

https://doi.org/10.1016/j.flowmeasinst.2021.101960
Received 15 January 2020; Received in revised form 1 May 2021; Accepted 5 May 2021
Available online 11 May 2021
0955-5986/© 2021 Elsevier Ltd. All rights reserved.
F. Liang et al. Flow Measurement and Instrumentation 79 (2021) 101960

Table 1
Parameters of materials (Temperature:25 ◦ C, Pressure:0.1 MPa).
Material Density kg/ Sound velocity m/ Acoustic impedance Mkg/
m3 s m2⋅s

plexiglass 1180 2730 3.221

Water 998 1497 1.494
Air 1.293 346 0.00045

Section 2: An acoustic pressure model is developed to investigate the

ultrasonic pulse propagation process at the wall-liquid and wall-gas
interface.
Section 3: The principle of flow pattern identification method using a
single ultrasonic transducer at the top of the pipe wall is described
and the echo characteristics of stratified flow, annular and slug flow
are analyzed.
Section 4: Experiments set up and comparison between the ultrasonic
echo series measured and simulation results are described.
Section 5: Flow pattern feature parameters extraction method and
RBF neural network for flow pattern automatic identification are
described and discussed.

2. Ultrasonic echo numerical simulation

2.1. Model building

Fig. 1. Schematic diagram of acoustic field model. The finite element software ANSYS is used to simulate the acoustic
field. We focus on the ultrasonic pulse propagating between the gas-
determined when the local void fraction at the crossing point is solid interface and liquid-solid interface. In order to reduce the
measured [14]. Compared with other techniques, it has higher temporal computational cost, a 2D model is established instead of a 3D model. As
and spatial resolutions. However, as an intrusiveness instrument, it is shown in Fig. 1, the inner diameter of the pipe is 30 mm and the wall
difficult to be used in a harsh environment. thickness is 10 mm. The material of the pipe wall is plexiglass, which has
Recently, the ultrasonic method has also been proposed for multi­ a sound velocity of 2730 m/s. Water and air are used as working fluids
phase flow measurement. As a kind of mechanical wave, the ultrasonic respectively. Detailed information of the physical parameters of plex­
wave has the characteristics of high frequency, short wavelength, and iglass, air, and water can be found in Table 1.
good directivity. There are three ultrasonic modes for multiphase flow At the outer wall of the pipe almost completely reflection of acoustic
diagnostics: the transmission method, the Doppler shift method [15,16], pressure occurs, therefore sound hard boundary is specified on the outer
and the pulse-echo method [17]. The pulse-echo method is based on the wall of the pipe. The initial excitation signal is applied to the node P on
reflection of an ultrasonic echo pulse from the gas-liquid interface and the top of the outer wall. The excitation pulse can be described using the
received great attention recently. Wada et al. [18] has detected the echo following equation:
signals reflected from a boundary between water and air in a vertical ( (
2(t − T0 )
)2 )
pipe. The bubble flow, slug flow and annular flow were identified based 100.exp − .sin(2πfi t) (1)
T0
on the transmit time and relative energy intensity of the pulse echo.
Murakawa et al. [19] used a combination of 2.0 and 8.0 MHz ultrasonic where: t is time, s; fi is the pulse emission frequency and T0 is the period
echo signals to measure bubble flow and slug flow. The main drawback of ultrasonic pulse, which is equal to 1/fi.
of the conventional pulse-echo method is that it is difficult to detect the The point P also acted as a signal receiver to receive echoes reflected
echo reflected from the gas-liquid interface due to the complex interface from the wall-fluid interface. The maximum amplitude of the incident
shapes of gas-liquid two phase flow [20]. pulse is 0.1 m2/s, and the emission frequency is 2.0 MHz.
From the literature review, we can conclude that despite the The mesh quality is essential for simulation accuracy and conver­
tremendous progress in flow pattern recognition in recent decades, it is gence speed. In the current work, a triangular mesh is applied. Ultra fine
still a challenge. The ultrasonic pulse-echo method has the advantages of mesh is used in the present work to ensure that the maximum mesh cell
non-invasive, fast responding speed, low cost and it’s suitable for size is less than one-sixth of the ultrasonic wavelength in both solid
metallic pipes, which makes it a promising method for flow pattern domain and fluid domain. It is well known that the wavelength is the
identification. However, previous researchers focused on the echo from ratio of the speed of sound to the frequency, so the maximum mesh size
the gas-liquid interface, which usually leads to large errors due to the in present work is determined by the following equation:
fluctuation of gas-liquid two phase flow. In the present work, a novel
c
echo pulse flow pattern identification method is proposed. The ultra­ hmax = (2)
6⋅f0
sonic reverberations which exist within the solid wall are used for flow
pattern, though it is often ignored in the previous studies. When ultra­
where, hmax denotes the maximum mesh size; c represents the sound
sonic pulse transmits in the pipe wall from the outer wall, a part of
velocity; f0 is the frequency of the ultrasonic pulse.
energy will reflect at the wall-fluid interface. The echo reflection coef­
In the present work, the emission frequency is 2.0 MHz. As listed in
ficient is related to the acoustic impedance of the fluid. Therefore, the
Table 1, the sound speeds of plexiglass, water and air are 2730 m/s,
echo traces from the inner wall of the pipe contain rich information of
1497 m/s, and 346 m/s respectively. According to Eq. (2), the maximum
gas-liquid two phase flow, which can be used for flow pattern
element size in the pipe wall, water and air are 0.228 mm, 0.125 mm and
identification.
0.0288 mm. The minimum mesh size is set to be 0.015 mm.
The following sections of this paper are arranged as follows:

2
F. Liang et al. Flow Measurement and Instrumentation 79 (2021) 101960

Fig. 2. The ultrasonic pulse propagation process.

3
F. Liang et al. Flow Measurement and Instrumentation 79 (2021) 101960

Fig. 2. (continued).

2.2. Numerical simulation results of echo series In contrast, for the plexiglass-water system, a part of ultrasonic pulse
will keep traveling in the water domain. This phenomenon can be
2.2.1. Ultrasonic wave propagation process explained by the transmission and reflection coefficients of the inner
Fig. 2 demonstrates the transmission process of the ultrasonic pulse wall.
in the pipe section. Under the effect of the incident pulse, the ultrasonic The ultrasonic transmission coefficient of the wall-fluid interface can
wave is generated and transmits in the pipe wall. When the ultrasonic be calculated as follows [21]:
pulse reaches the pipe’s inner wall, reflection occurs due to the great
2Zf
acoustic impedance difference between plexiglass and fluid. Then the Tp = (3)
Zf + Zw
ultrasonic wave will be reflected and transmit back in the pipe wall. This
echo is called the first echo. The first echo keeps traveling in the solid
where Tp is the transmission coefficient; Zw is the acoustic impedance of
wall until it reaches the outer wall of the pipe where it will be received
solid wall (kg/m2⋅s) and Zf is the acoustic impedance of fluid (kg/m2⋅s).
by point P. Similarly, on the outer wall, part of the ultrasonic waves will
As listed in Table 1, the acoustic impedances of plexiglass and air are
be reflected back again and still travels in the pipe wall. The ultrasonic
3.221 Mkg/m2⋅s and 0.00045 Mkg/m2⋅s, respectively. According to Eq.
pulse will be reflected back and forth between the inner and outer wall
(3), the transmission coefficient is 0.0279%, which indicates almost all
until all energy is lost.
the incident waves are reflected at wall-gas interface because of the
It is interesting to observe that the acoustic field of plexiglass-water
great acoustic impedance difference between plexiglass and air. In the
system is different from that of plexiglass-air system. In the case of
case of wall-liquid interface, the corresponding transmission coefficient
plexiglass-air system, no ultrasonic waves are found in the fluid domain.

4
F. Liang et al. Flow Measurement and Instrumentation 79 (2021) 101960

Fig. 4. Flow pattern identification principle.

2⋅δ
Δt = (5)
Cw

where Cw is the ultrasonic wave speed in the pipe wall and δ is the
thickness of the pipe wall.
Fig. 3 shows that the amplitude of each echo reflected from wall-gas
interface is larger than that at solid-liquid interface. As the number of
reflections increases, the difference in echo amplitude between the two
cases increases. The third echo at the wall-gas interface is still very
obvious, while the echo at wall-liquid interface is difficult to distinguish
after the second one. According to the numerical simulation results,
obvious differences in the echo spectrum exist between wall-gas and
wall-liquid interface. The echo series can be used to identify which
phase is now contacting the inner wall of the pipe. This finding can be
used for flow pattern identification.

3. The principle of flow pattern identification

For horizontal gas-liquid two phase pipe flow, the gas and liquid
distribution information at the top of the wall can be used to identify the
flow pattern. As shown in Fig. 4, it is now possible to determine which
kind of fluid is contacting the top wall based on the trajectory of the
ultrasonic echo measured by the transducer mounted on the top of the
wall.
As shown in Fig. 5, different flow patterns have different gas-liquid
distribution schemes at the pipe cross-section. In the case of stratified
flow, the liquid tends to flow at the bottom of the pipe due to the effect of
gravity. The gas always occupies the upper part of the pipe section and
the ultrasonic transducer will detect a trail of echoes from wall-gas
Fig. 3. Comparison of echo series between wall-liquid and wall-gas interface.
interface.
For annular flow, a liquid film is formed at the pipe wall and the gas
is 63.37%. This means that more than half of the incident ultrasonic core with entrained droplets flow continuously in the center of the tube.
waves can penetrate the wall-water boundary and propagate in the The inner wall of the pipe is always covered by a layer of liquid film and
water. As a result, only a small part of the ultrasonic waves are reflected the ultrasonic transducers will receive the echo from wall-liquid inter­
and the ultrasonic transducer at point P will receive a much lower face. In the horizontal pipe, the liquid film distribution is not uniform
amplitude of reflected echo. due to the effect of gravity and the bottom liquid film tends to be thicker
than that of the top. But the measured echoes are reflected from the
2.2.2. Echo spectrum characteristics at wall-gas and wall-liquid interface pipe’s inner wall, which are independent of liquid film thickness.
Fig. 3 demonstrates the echoes reflected from plexiglass-water and Slug flow is characterized by alternating liquid slugs and gas bubbles
plexiglass-air interface respectively. The acoustic pressure of echo is flowing above the liquid films. So a slug unit can be divided into two
normalized by the following equation: zones: liquid film zone and slug zone. The liquid film zone is similar to
Pt that of stratified flow, where the liquid film flow at the bottom of the
P*t = (4) pipe and the upper part of the pipe is a gas pocket. The slug zone is
Pmax
mainly liquid containing small bubbles that occupy the whole pipe
where Pmax is the maximum acoustic pressure amplitude of the incident cross-section. When the Taylor bubble passes the test point, the received
pulse; Pt is acoustic pressure amplitude at time t; P*t is normalized echo series is identical to that of stratified flow. On the other hand, when
acoustic pressure and is within the range of 0–1. the slug is coming, the whole pipe cross-section is blocked by the liquid
The first wave is the incident pulse, and then followed by the first slug, and the echo characteristic is similar to that of the wall-liquid
echo, the second echo and so on. The Time of Flight (TOF) (Δt) of the interface. So in the case of slug flow, the gas and liquid will contact
ultrasonic wave through the pipe wall can be determined by recording the top wall intermittently, and the slug flow pattern can be identified by
the time difference between the echo pulses reflected from the inner the change of the echo series over time.
wall.

5
F. Liang et al. Flow Measurement and Instrumentation 79 (2021) 101960

Fig. 5. Gas-liquid distribution characteristics of various flow patterns.

Fig. 6. Schematic of experimental setup.

4. Experimental verification separator is close to atmospheric pressure. Water from the pump and
compressed air from the compressor are mixed at the mixer before
4.1. Experimental setup entering the test pipe section. The distance between the test section and
the mixer is about 150 times of the diameter to ensure flow fully
The experiment was carried out on a gas-liquid two-phase flow loop. developed. The test section is made of transparent plexiglass for easy
Air and tap water are used as working fluids. Fig. 6 represents the observation of flow patterns. The length of the test section is 4.0 m, and
schematic of the experimental setup. The experimental system consists it has an inner diameter of 30 mm and pipe thickness of 10 mm. There
of three parts: gas-liquid circulation system, experimental test section was a cyclone gas-liquid separator located at the end of the loop, where
and data acquisition system. Experiments were performed under the all the gas and liquid were separated. The gas was directly discharged
conditions of room temperature and the outlet pressure of the cyclone into the atmosphere and the water flowed back into the water tank for

6
F. Liang et al. Flow Measurement and Instrumentation 79 (2021) 101960

Fig. 7. Ultrasonic echo measurement system.

type vortex flowmeter. M1L was measured by a Yokogawa AXF010G-

type electromagnetic flowmeter. In the present experiments, the su­
perficial gas velocity was ranged from 0.6 m/s to 30.0 m/s, and the
superficial liquid velocity varied from 0.02 m/s to 0.25 m/s. The flow
patterns are observed through the transparent test section. The flow
patterns that occurred in the experiments included wavy flow, slug flow,
and annular flow.
As shown in Fig. 7, an ultrasonic transducer is attached to the top
wall. The transducer can emit an ultrasonic pulse and receive the re­
flected echo from the inner wall simultaneously. The transducer used is a
contact longitudinal wave sensor with a center frequency of 2.0 MHz
and an outer diameter of 6 mm. The ultrasonic transmitter is the 5072
PR pulse generator of Olympus, and the oscilloscope is the DPO2024
oscilloscope of Tektronix.

4.2. Comparison of experiment and simulation results

Fig. 8 demonstrates the echo spectrum from experimental measure­

ment and numerical simulation. It should be pointed out that the pres­
sure pulse signal is used as the excitation signal in the numerical
simulation, while electrical signal pulse is emitted by the ultrasonic
transducer in the experiments. So the amplitude of these echoes is quite
different in magnitude. However, it is found that the echo intervals of
experimental and simulation results are consistent with each other. As
expected, there are three echoes reflected from wall-gas interface
measured, and in the case of wall-liquid interface, only two echoes are
observed. It is concluded that the difference of echo spectrum charac­
teristics can help to distinguish which kind of fluid (gas or water) is now
in touch with the inner wall and then identify the flow pattern according
to the principle described in Section 3.

5. Flow pattern identification using artificial neural network

5.1. Flow pattern feature parameters extraction

In order to reduce the noise of the raw signal, a filter algorithm is

adopted first [22]. Then Hilbert transform is conducted to eliminate the
negative value of the signal. The next step is to extract the feature pa­
rameters hidden in the ultrasonic echo spectrum.
In the case of slug flow, when the tailor bubble approaches the test
point, the gas phase is in contact with the upper wall of the pipeline. So
the relatively larger S is obtained due to strong reflection of echo at gas-
wall interface (see Fig. 9a). In contrast, when the liquid slug approaches
Fig. 8. Comparison of echo spectrum between experiment and simulation. the test point, the liquid will cover the top wall and a relatively smaller S
is obtained (see Fig. 9b). For stratified flow, the top wall always contacts
circulation. the gas and higher S is obtained (see Fig. 9c). Similarly, in the case of
The inlet gas mass flow rate (M1G) and liquid mass flow rate (M1L) annular flow, the top wall is covered by the liquid film and a smaller S is
were measured before mixing. M1G was metered by a Yokogawa DY025- achieved (see Fig. 9d).
Fig. 9 (a)–(d) indicate there exists an obvious difference in envelop

7
F. Liang et al. Flow Measurement and Instrumentation 79 (2021) 101960

Fig. 9. Raw pulse echo signal treatment.

8
F. Liang et al. Flow Measurement and Instrumentation 79 (2021) 101960

Table 2
Envelope area of echo signals under different flow regimes.
Stratified flow Annular flow Slug flow

Slug zone Liquid film zone

Sn 0.72–0.97 0.1–0.29 0.04–0.14 0.68–0.93

Fig. 11. Area ratio K under different flow patterns.

envelop area alone can’t distinguish slug flow from wavy flow and
annular flow. Here, area ratio (K) is proposed to help distinguish slug
flow. K is defined as:
( )
max sn,T , sn,T+ΔT , ..., sn,m.ΔT
K= ( ) (8)
min sn,T , sn,T+ΔT , ..., sn,m.ΔT
Fig. 10. Envelope area of ultrasonic echo signals under different flow regimes.
where Sn,T represents the normalized echo envelope area at time T;
Sn,m.ΔT represents the normalized echo envelope area at time T + m.ΔT;
area between wall-gas and wall-liquid interface. Therefore, the envelope
ΔT is the measurement time interval between two echo pulses; m is the
area of the processed echo signal spectrum can be used as a flow pattern
number of echo signal.
indicator. The envelope area is calculated as follows:
In order to capture the change of the ultrasonic echo under the slug
length(t)
∑ flow, care must be taken to ensure the ultrasonic echo acquisition time
S= Ai ⋅Δts (6) exceeds a whole slug period during the experiment. The period of slug
flow can be determined by slug frequency, which is the number of slugs
i=1

where S is the envelope area of the echo spectrum; Δts is the sampling passing a specific point along the pipeline over a certain period. Ac­
point time interval; Ai is the signal envelope amplitude; t is the signal cording to slug flow frequency prediction correlation proposed by Jep­
time series. son et al. [23], the slug frequency is calculated as:
Then the normalized envelope area can be obtained using the USL
following equation: fs = (0.00759Um + 0.01) (9)
D
S − Sl where fs is the slug frequency, D is the pipe diameter; USL is the super­
Sn = (7)
Sg − Sl ficial liquid velocity, and Um is the mixture velocity.
As mentioned above, the inner diameter of the test section is 30 mm
where Sn is the normalized area, Sl is the envelope area of wall-liquid
and the range of superficial gas and liquid velocities are 6.0 m/s - 30.0
interface in static calibration experiment, Sg is the envelope area of
m/s and 0.01 m/s - 0.25 m/s respectively. According to Eq. (9), the slug
wall-gas interface in the static calibration experiment.
frequency is within the range of 0.02 and 2.0. Therefore, the minimum
The normalized ultrasonic signal envelope area under various flow
measurement duration should be no less than 50s. In the present work,
patterns is obtained and the results are shown in Table .2.
120s of measurement duration is selected and the sampling interval is
Combining Table .2 with Fig. 10, it can be seen that the normalized
10s.
envelope area under each flow regime is very obvious. When the flow
As mentioned above, K is the maximum envelop area ratio within a
pattern in the pipe is stratified flow, the envelope area of its echo
measurement period. For slug flow, the top wall will contact liquid and
spectrum is generally large, which is due to the fact that the inner wall of
gas intermittently, so a larger is achieved. However, for annular flow
the pipe at the top is always in contact with air and higher echo
and stratified flow, the fluid in contact with the top wall remains un­
amplitude is obtained. Under the annular flow condition, the entire
changed. As a result, annular flow and stratified flow tend to have a
inner wall of the pipe is covered by a layer of liquid. The energy of the
smaller K. As shown in Fig. 11, the slug flow has a much larger K than
reflected echo at the wall-liquid interface is much smaller than that of
that of stratified flow and annular flow.
wall-gas interface. As a result, lower echo amplitude is detected and
leads to much smaller envelope areas in the case of annular flow.
For slug flow, when the liquid slug reaches the test point, the liquid 5.2. Flow regime identification using RBF neural network
will be in touch with the top wall and the envelope area of echo signal
measured by the probe is very similar to that of annular flow. When In order to identify the flow pattern automatically according to the
Taylor bubble reaches the test point, the gas liquid distribution is similar flow pattern indicators: normalized envelop area (Sn) and area ratio (K),
to stratified flow and the gas is now in touch with the top wall. So the an RBF (Radial basis function) neural network is developed. The RBF
echo received by the probe is similar to that of wavy flow. Therefore, network is proposed by Moody and Darken in 1989 based on the

9
F. Liang et al. Flow Measurement and Instrumentation 79 (2021) 101960

Fig. 12. Schematic diagram of the RBF neural network for flow pattern identification.

Table 3
Test samples of RBF neural network.
No. Usl Usg Sn K Flow Identification result
1 1
/m⋅s− /m⋅s− regime of RBF

1 0.0478 2.296 0.96 1.05 Stratified Stratified

2 0.0478 3.454 0.94 1.11 Stratified Stratified
3 0.0478 4.750 0.89 1.16 Stratified Stratified
4 0.0478 6.175 0.85 1.20 Stratified Stratified
5 0.0478 7.811 0.79 1.15 Stratified Stratified
6 0.0345 20.729 0.11 1.52 Annular Annular
7 0.0345 22.454 0.18 1.33 Annular Annular
8 0.0345 24.378 0.22 1.18 Annular Annular
9 0.0345 25.781 0.25 1.26 Annular Annular
10 0.0345 27.103 0.27 1.37 Annular Annular
11 0.117 3.075 0.05 18.12 Slug Slug
12 0.117 3.487 0.09 12.86 Slug Slug
13 0.117 4.316 0.89 9.21 Slug Slug
14 0.117 5.037 0.69 7.33 Slug Slug
15 0.117 5.518 0.12 20.12 Slug Slug
16 0.184 2.207 0.91 12.86 Slug Slug
17 0.184 3.114 0.79 9.88 Slug Slug
18 0.184 3.659 0.14 7.80 Slug Slug
19 0.184 4.188 0.92 23.13 Slug Slug
20 0.184 4.764 0.75 5.05 Slug Slug

localized and overlapping receptive domains in the human cortex [24].

Because of its simple structure, excellent function approximation ability, Fig. 13. Test matrix in Mandhane flow pattern map [25].
and fast learning speed, it has good performance in pattern recognition
and system identification.
The schematic diagram of the RBF neural network flow regime Table 4
identification system is shown in Fig. 12. It is composed of three layers: Identification results of RBF neural network.
input layer, hidden layer, and output layer. The input layer has two Stratified flow Annular flow Slug flow
nodes: normalized envelope area Sn and area ratio K. The output layer
Number of flow regime 25 25 50
has three nodes and the corresponding output vectors are stratified flow Number of right identification 24 23 47
(0 0 1), annular flow (0 1 0), and slug flow (1 0 0). Recognition accuracy/% 96 92 94
In essence, flow pattern identification based on a neural network is a
training-recognition process. The RBF neural network is trained by using
the feature vector defined by Sn and K as an input sample and the real The recognition results are shown in Table .4. It shows most flow
flow pattern is obtained through visual observation. Table .3 shows patterns are identified correctly and the overall recognition accuracy is
some samples used during network training. 94.0%.
From the total of 300 sets of data obtained from the experiments, 200 It should be pointed out that only stratified flow, annular flow and
sets of data were selected as training samples, and the RBF neural slug flow are tested in the present work. For bubble flow and plug flow,
network was trained by the cross-training method to obtain the flow due to the effect of buoyancy forces, the gas tends to accumulate in the
regime recognition feature database. The other 100 sets were used as upper part of the pipe. The top wall will touch gas and liquid intermit­
test samples to evaluate the network identification performance (see tently, and bubble flow and plug flow may be identified as slug flow.
Fig. 13). Each point represents an experimental condition. The position However, the fluctuation characteristics of bubble flow, plug flow and
of the test point is specified by the corresponding superficial gas and slug flow are different. The fluctuation frequency of attenuation spec­
liquid velocity. It is found that test points with incorrect flow pattern trum will be tried for flow pattern identification in the future study.
recognition are located near the flow pattern transition boundary.

10
F. Liang et al. Flow Measurement and Instrumentation 79 (2021) 101960

6. Conclusion [2] G.F. Hewitt, Measurement of two phase flow parameters, 1978. Nasa Sti/recon
Technical Report A 79.
[3] P.L. Spedding, D.R. Spence, Flow regimes in two-phase gas-liquid flow, Int. J.
(1) An acoustic pressure model is established to study the echo pulse Multiphas. Flow 19 (2) (1993) 245–280.
characteristics in a pipe cross-section. Numerical simulation re­ [4] L. Shi, J. Cai, Z. Zhou, Gas-liquid two-phase flow pattern identification based on
sults show that more reflection echo pulses are obtained at wall- image processing, J. Zhejiang Univ. Eng. Sci. 39 (8) (2005) 1128.
[5] M.J. Da Silva, S. Thiele, L. Abdulkareem, B.J. Azzopardi, U. Hampel, High-
gas interface than that of wall-liquid interface due to the fact that resolution gas–oil two-phase flow visualization with a capacitance wire-mesh
more energy is lost when ultrasonic pulse travels across the sensor, Flow Meas. Instrum. 21 (3) (2010) 191–197.
interface into the water. This difference in echo spectrum can be [6] G.P. Lucas, J.C. Cory, R.C. Waterfall, A six-electrode local probe for measuring
solids velocity and volume fraction profiles in solids-water flows, Meas. Sci.
used to determine the type of medium in contact with the inner Technol. 11 (10) (2000) 1498.
wall of the pipe. [7] M. Fossa, Design and performance of a conductance probe for measuring the liquid
(2) Experiments are carried out in an air-water two-phase flow loop, fraction in two-phase gas-liquid flows, Flow Meas. Instrum. 9 (2) (1998) 103–109.
[8] B.K. Arvoh, R. Hoffmann, M. Halstensen, Estimation of volume fractions and flow
stratified flow, annular flow and slug flow are tested. A probe regime identification in multiphase flow based on gamma measurements and
mounted on the top wall of the test section is used to measure the multivariate calibration, Flow Meas. Instrum. 23 (1) (2012) 56–65.
ultrasonic echo sequence. The measured flight time is in good [9] C. Sætre, G.A. Johansen, S.A. Tjugum, Tomographic multiphase flow measurement,
Appl. Radiat. Isot. 70 (7) (2012) 1080–1084.
agreement with the simulation. [10] F. Dong, Z.X. Jiang, X.T. Qiao, L. Xu, Application of electrical resistance
(3) A novel flow pattern identification method is proposed using the tomography to two-phase pipe flow parameters measurement, Flow Meas. Instrum.
RBF neural network. Normalized envelop area (Sn) and area ratio 14 (4–5) (2003) 183–192.
[11] R. Banasiak, R. Wajman, T. Jaworski, P. Fiderek, H. Fidos, J. Nowakowski,
(K) are used as two feature parameters to identify flow patterns.
D. Sankowski, Study on two-phase flow regime visualization and identification
Compared with the conventional methods, the ultrasonic echoes using 3D electrical capacitance tomography and fuzzy-logic classification, Int. J.
from wall-fluid interface instead of gas-liquid interface are used Multiphas. Flow 58 (2014) 1–14.
for flow pattern identification and the echo signals are not [12] M. Parsi, R.E. Vieira, C.F. Torres, N.R. Kesana, B.S. McLaury, S.A. Shirazi,
U. Hampel, Experimental investigation of interfacial structures within churn flow
affected by the fluctuation of the gas-liquid interface. The present using a dual wire-mesh sensor, Int. J. Multiphas. Flow 73 (2015) 155–170.
method has the advantages of low cost and no interference and [13] H. Prasser, A. Böttger, J. Zschau, A new electrode-mesh tomograph for gas/liquid
can be used in harsh environments. flows, Flow Meas. Instrum. 9 (1998) 111–119.
[14] R.E. Vieira, M. Parsi, C.F. Torres, B.S. McLaury, S.A. Shirazi, E. Schleicher,
U. Hampel, Experimental characterization of vertical gas–liquid pipe flow for
Author statement annular and liquid loading conditions using dual wire-mesh sensor, Exp. Therm.
Fluid Sci. 64 (2015) 81–93.
[15] B.M. Abbagoni, H. Yeung, Non-invasive classification of gas–liquid two-phase
Under supervision by Fachun Liang, Yue Hang constructed the horizontal flow regimes using an ultrasonic Doppler sensor and a neural network,
acoustic model, Hao Yu carried out experiments. Jifeng Gao developed Meas. Sci. Technol. 27 (2016), 084002.
the RBF neural network model for flow pattern identification. Fachun [16] S.G. Nnabuifea, K.E.S. Pilarioa, L. Lao, Y. Cao, Identification of gas-liquid flow
regimes using a non-intrusive Doppler ultrasonic sensor and virtual flow regime
Liang revised the manuscript according to comments of the reviewers. maps, Flow Meas. Instrum. 68 (2019) 101568.
All authors have made substantial contributed to the manuscript. [17] J.S. Chang, E.C. Morala, Determination of two-phase interfacial areas by an
ultrasonic technique, Nucl. Eng. Des. 122 (1–3) (1990) 143–156.
[18] S. Wada, H. Kikura, M. Aritomi, Pattern recognition and signal processing of
Declaration of competing interest
ultrasonic echo signal on two-phase flow, Flow Meas. Instrum. 17 (4) (2006)
207–224.
The authors declare that they have no known competing financial [19] H. Murakawa, H. Kikura, M. Aritomi, Application of ultrasonic multi-wave method
interests or personal relationships that could have appeared to influence for two-phase bubbly and slug flows, Flow Meas. Instrum. 19 (3–4) (2008)
205–213.
the work reported in this paper. [20] Y. Murai, Y. Tasaka, Y. Nambu, Y. Takeda, Ultrasonic detection of moving
interfaces in gas–liquid two-phase flow, Flow Meas. Instrum. 21 (3) (2010)
Acknowledgements 356–366.
[21] L.E. Kinsler, A.R. Frey, A.B. Coppens, J.V. Sanders, Fundamentals of Acoustics,
Wiley-VCH, 1999.
This work is supported by National Natural Science Foundation of [22] R. Demirli, J. Saniie, Model-based estimation of ultrasonic echoes. Part I: analysis
China (Grant No. 51574272) and National Key Research and Develop­ and algorithms, IEEE Trans. Ultrason. Ferroelectrics Freq. Contr. 48 (3) (2001)
787–802.
ment Program of China(Grant No. 2016YFC0802301),which are grate­ [23] W.P. Jepson, R.E. Taylor, Slug flow and its transitions in large-diameter horizontal
fully acknowledged. pipes, Int. J. Multiphas. Flow 19 (3) (1993) 411–420.
[24] J. Moody, C.J. Darken, Fast learning in networks of locally-tuned processing units,
Neural Comput. 1 (2) (1989) 281–294.
References [25] J.M. Manhdhane, G.A. Gregory, K. Aziz, A flow pattern map for gas-liquid flow in
horizontal pipes, Int. J. Multiphas. Flow 1 (1974) 537–553.
[1] E. Michaelides, C.T. Crowe, J.D. Schwarzkopf, Multiphase Flow Handbook, CRC
Press, 2016.

11