Experimental Thermal and Fluid Science 70 (2016) 29–50

Contents lists available at ScienceDirect

Experimental Thermal and Fluid Science

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

The use of an ultrasonic technique and neural networks for identification

of the flow pattern and measurement of the gas volume fraction
in multiphase flows
M.M.F. Figueiredo a, J.L. Goncalves a, A.M.V. Nakashima b, A.M.F. Fileti b, R.D.M. Carvalho a,b,⇑
a
Instituto de Engenharia Mecânica (IEM-UNIFEI), Itajubá, MG, Brazil
b
Faculdade de Engenharia Química (FEQ-UNICAMP), Campinas, SP, Brazil

a r t i c l e i n f o a b s t r a c t

Article history: In the oil industry, the well stream often consists of a full range of hydrocarbons and a variety of
Received 28 October 2014 non-wanted components such as water, carbon dioxide, salts, sulfur, and sand. The need for multiphase
Received in revised form 11 August 2015 flow metering (MFM) arises when it is necessary or desirable to meter the flow upstream of the
Accepted 14 August 2015
separators. The ultrasonic technique fulfils many of the requirements for MFM in the oil industry (mainly,
Available online 21 August 2015
non-invasive, non-radiative, robust, and relatively non-expensive) and has the capability to provide the
information required. The drawback of current ultrasonic techniques, as is the case with other MFM
Keywords:
methods, is the need for prior signal calibration. A broader solution to this issue could be the use of
Ultrasonic technique
Neural networks
artificial neural networks (ANNs). ANNs provide a non-linear mapping between input and output
Multiphase flow variables and the cross-correlation among these variables and could be an alternative tool for automatic
Flow pattern identification of flow patterns. In this context, the objectives of the current investigation are two-fold:
Gas volume fraction (i) to present and analyze acoustic attenuation data for vertical, upward oil-continuous multiphase flows
in 1-in. and 2-in. acrylic pipes and flow patterns ranging from bubbly flows to annular flows; (ii) to
develop neural networks for flow pattern recognition and gas volume fraction (GVF) measurement using
the ultrasonic attenuation data as input. The results shown testify to the ability of the neural networks
and the ultrasonic technique to perform these tasks.
Ó 2015 Elsevier Inc. All rights reserved.

1. Introduction A detailed discussion of different instruments and techniques

A simplified view of an oil and gas production plant is shown in
Fig. 1. The wellheads on the left feed into production and test man-
ifolds; the remainder of the figure is the actual gas and oil separa-
tion plant (GOSP). Often the well stream will consist of a full range
of hydrocarbons and a variety of non-wanted components such as
water, carbon dioxide, salts, sulfur, and sand [3]. Separation of
these components is accomplished mainly by gravity production
separators. The need for multiphase flow metering (MFM) arises
when it is necessary or desirable to meter the flow upstream of
the separators. MFM enables measurement of unprocessed multi-
phase streams very close to the well, thereby providing continuous
monitoring of well performance and better reservoir exploitation
and drainage.
⇑ Corresponding author at: UNIFEI, Av. BPS, 1303, Bairro Pinheirinho, Itajubá,
MG 37.500-903, Brazil. Tel.: +55 35 8867 4154.
E-mail address: ridimarcar@gmail.com (R.D.M. Carvalho).
that have been considered for MFM applications can be found in
[3–5]. Table 1 below groups these instruments and techniques
according to the flow parameter involved in their principle of oper-
ation. The Remarks column summarizes information in the litera-
ture [3,6] involving the application of these techniques to MFM.
It can be seen that most techniques either call for additional
research and development before they can be reliably applied to
multiphase flows or still need further development despite already
being used in MFM applications. Gamma densitometry, despite
requiring carefully implemented procedures to handle radioactive
sources, is used in many commercial multiphase flow meters
(MPFMs) [3]. However, it is subject to errors due to phase distribu-
tion in time and space and thus needs to be used in combination
with other devices that properly account for the flow upstream
conditions. Scheers [6] points out an almost 20% difference
between actual and measured gas volume fraction (GVF) in the
annular flow pattern using gamma ray densitometry.
Many other techniques in Table 1 can be seen to use radioactive
sensors. In this connection, Arora [8] discusses the advantages of

http://dx.doi.org/10.1016/j.expthermflusci.2015.08.010
0894-1777/Ó 2015 Elsevier Inc. All rights reserved.
30 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50

Nomenclature

Latin symbols S total extinction area (m2)

a bubble or particle radius (m) T normalized amplitude of transmitted signal (–)
A projected cross sectional area (m2) TI intensity transmission coefficient (–)
A parameter defined in Atkinson and Kytömaa [1] T temperature (K)
(Pa s/m2) U amplitude of particle velocity (m/s)
A parameter defined by Isakovich [2] V voltage (V)
B parameter defined in Atkinson and Kytömaa [1] (kg/m3)
c speed of sound (m/s) Greek symbols
cp specific heat at constant pressure (J/kg K) a spatial absorption coefficient (Np/m or dB/m)
E acoustic pulse energy (V2 s / J) ac classical absorption coefficient (Np/m or dB/m)
I acoustic intensity (W/m2) ak thermal absorption coefficient (Np/m or dB/m)
jl spherical Bessel function of the first kind as viscous (Stokes) absorption coefficient (Np/m or dB/m)
k thermal conductivity (W/m K) b thermal expansion coefficient (K1)
k wave number (m1) c ratio of specific heats (–)
K wave vector (m1p ) ffiffiffiffiffiffiffi C interfacial area per unit volume (m2/m3)
j imaginary unity, 1 d boundary layer thickness (m)
n thermal wave vector (m1) g coefficient of shear viscosity (Pa s)
nl spherical Bessel function of the second kind j suspension effective bulk modulus of elasticity (Pa]
N border between transducer near and far fields (m) m particles fraction (–)
P pressure or pressure amplitude (Pa) q density (kg/m3)
r specific acoustic resistance (rayl) q suspension density (kg/m3)
RI intensity reflection coefficient (–) q parameter defined in Atkinson and Kytömaa [1] (kg/m3)
Re Reynolds number (–) x angular frequency (rad/s)
x spatial coordinate (m)

non-radioactive MPFMs over MFM techniques involving radioac-
tive sensors. Non-radioactive MPFMs are much cheaper in terms
of construction because they do not use radioactive elements. In
addition, the operational expenditure is much less than that of
radioactive instruments because of the enormous costs associated
with the various health, safety, and environmental requirements
for the latter. Furthermore, the costs related to export/import of
non-radioactive equipment are much lower because they do not
require specialized packing and clearance procedures. Finally,
mobile non-radioactive MPFMs can be used with ease for well tests

Fig. 1. Oil and gas production overview – adapted from [7].

 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50 31

Table 1
Classification of MFM instruments and techniques (adapted from [4]).

Measured flow Measurement principle Remarks

parameter
Mean fluid Vibrating tube Techniques not quite established yet; difficulties regarding their application to two-phase flows
density Weighing of tube
Phase fractions Electrical impedance Rapid response; however, if phases configuration is not known a priori, phase fraction cannot be
determined unambiguously; problems with phase inversion in oil–water flows
Acoustic attenuation Technique still under development; the relative importance of acoustic phenomena in the multiphase
flow depends on the flow pattern, thus requiring different data reduction methods
Single-beam gamma densitometry* Their use is restricted to truly homogeneous flows
Broad beam gamma densitometry* Advantages over single-beam offset by need for careful determination of collimator shape
Multi-beam gamma or X-rays Well established for void fraction and flow pattern measurements; errors due to phase distribution
densitometry*
Microwave resonance Resonant frequency measurements restricted to oil–water mixtures in sample lines; not effective in
water –continuous fluids; in the presence of gas technique is unlikely to work predictably
Gamma-ray scattering* Many measurements necessary to obtain void fraction; unlikely to find much use outside research lab
Neutron absorption* Neither technique offers sufficient advantage over gamma-ray, particularly in subsea applications
Neutron scattering*
Flow velocity Turbine flow meter Requires independent measurements of void fraction and quality or slip ratio; careful calibration required
unless flow is truly homogeneous
Vortex shedding meter Unlikely to be part of multiphase monitoring packages, except maybe for single-phase flow
measurements associated with component separation
Acoustic (pulse and return) meter Its application to multiphase flows still needs to be established by careful experiments
Acoustic cross correlation In-situ calibration is essential requirement for measuring multiphase flow velocity
Pulsed-neutron activation (PNA)* Phase fractions must be known beforehand; tagging of oxygen in water or impurities in oil and gas could
then yield flow velocity; much work needed for application in flows without homogenization
Electromagnetic flow meter Advantages when applied to homogenized gas–air–water mixture; viability and application range in
oil–water–gas flows still have to be established
Gamma-ray*/Neutron*, All of them use the cross correlation process, thus requiring extensive in-situ calibration
Capacitance/conductivity cross
correlation
Volume/mass EMFM/TMFM/PNA*, vibrating tube (EMFM – Electromagnetic flow meter; TMFM – True mass flow meter) TMFM is promising for allocation
flow rate (custody transfer) and possibly fiscal metering; problems relating to phase fraction measurement
Momentum Venturi meter Its application to heterogeneous flows requires calibration for particular conditions at hand
flux Pressure fluctuation signals For a given flow pattern, statistical characteristics of pressure fluctuations may be used to determine
phases momentum fluxes and component velocities; practical applications require extensive in-situ
calibration
*
Requires radioactive sensor.

during the fluid drilling phase for early production checks without
the need for any of the special safety requirements associated with
radioactive MPFMs.
The use of impedance devices to characterize multiphase flows
dates back to the 1960s [3]; it has also been used in commercial
MPFMs [4,6]. However, impedance methods can be highly sensitive
to the flow pattern within the channel [3] and are subject to the
phase inversion problem, i.e. when the flow changes from oil con-
tinuous to water continuous detection has to change from capaci-
tance to conductance. In addition, the electrodes have to be in
direct contact with the flow (the technique is invasive), which
might lead to difficulties in building the measuring instrument in
some cases. Tan et al. [9] presented a measurement system that
combines conductance-ring sensors (CR) for phase fraction mea-
surement and a differential pressure cone meter for overall flow
rate measurement. The system is intended for application to
water–oil (liquid–liquid) flows, but the authors claim the proposed
methodology can be extended to gas–liquid two-phase flows.
However, no experimental data are provided for these flows.
It is noteworthy the wire-mesh impedance technique [10,11],
which can be used to determine the instantaneous void fraction
distribution over the pipe cross-section. Even though intrusive,
invasive, and still restricted to the laboratory environment, this
technique exhibits high temporal resolution, low cost, and simplic-
ity when compared with other imaging systems. In connection
with traditional electrical capacitance tomography (ECT), it is
cheaper, faster, and safer than radiation based tomographic tech-
niques and can be readily automated and digitized; nonetheless,
it is subject to the phase inversion problem and its spatial resolu-
tion is inferior to that of nuclear based tomography [3].
The ultrasonic techniques represent a promising alternative to
the costly, complex, and hazardous techniques in Table 1. In this
regard, Poette and Reynier [12] presented a study comparing
non-intrusive techniques for three-dimensional mapping of multi-
phase flows occurring during tank sloshing. This is a key issue for
assuring propellant feeding during flight and manoeuvers of
launchers and satellites. Particular attention was paid to electrical
and ultrasonic tomography since they are both non-intrusive, non-
invasive, low cost, fast and simple to operate, and suitable for real
time measurements. The ultrasonic technique was deemed more
promising for the intended application as it is less affected by tem-
perature variations and also represents an advantage in terms of
safety. An extensive review on the use of the ultrasonic technique
in the context of MFM can be found in [13]. Ultrasonic signals can
fulfill the requirements of remote sensing, long distance signal
transmission, and harsh operating conditions normally encoun-
tered in the oil industry. In multiphase flows, as the concentration
and size distribution of the dispersed media (solid particles and/or
gas bubbles) change the signal amplitude and transit time vary due
to the combined effect of acoustic attenuation and transmission
phenomena; these issues are further discussed in Section 3.2. For
detailed information regarding acoustic phenomena, including
sound absorption in heterogeneous flows, the reader is referred
to specific textbooks on the subject, e.g. [14–16].
Many investigations in the literature deal with the ultrasonic
technique for measurement of velocity of the dispersed phases
[17–24]. Of more immediate interest to the present investigation,
acoustic attenuation was used to measure the void fraction of bub-
bly water–air flows [25]; a semi-empirical model of the interaction
of the sound wave with the two-phase flow was developed. Chen
32 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50
et al. [26] theoretically and experimentally investigated the feasi-
bility of using acoustic resonant spectroscopy (ARS) to measure
the gas volume fraction of gas–liquid flows. The experiment con-
sisted of a cylindrical cavity with open ends immersed in trans-
former oil through which air was bubbled; their effect on the
resonant frequency (RF) and resonant amplitude (RA) of the assem-
bly was captured by a hydrophone. Air bubbles of volume smaller
than 0.1% of the cylindrical cavity size were capable of causing a
prominent RF shift; this shift was maximum when the bubbles
were located in the middle of the cavity. However, both the RF
and RA were observed to decrease dramatically with increasing
gas flow rate, thus decreasing the sensitivity of the technique. Xing
et al. [27] presented an investigation on the potential use of a
single-path gas ultrasonic flow meter for measuring gas–liquid
two-phase flows of low liquid loading, e.g. flows of wet natural
gas. A correction method based on the use of traditional void frac-
tion models was proposed whereby metering of individual flow
rates of stratified and annular flows could be achieved with less
than 5% rms error.
Murai et al. [28] presented an ultrasonic method for measuring
the spatial distribution of the void fraction aimed at solving a prob-
lem that occurs in the ultrasonic sensing of bubbles in liquids, i.e.
the difficulty of the ultrasound to provide information beyond
the bubble closest to the transducer. Two signal processing
schemes were used, the echo intensity and the Doppler methods;
the former technique was suited for small spherical bubbles while
the latter one was better suited for large bubbles. Data were pre-
sented for water–air bubbly and slug flows. Later on, Murai et al.
[29] experimentally compared three different ultrasonic tech-
niques for detection of moving gas–liquid interfaces, namely, echo
intensity, local Doppler, and velocity variance. Data were presented
for free traveling waves, bubbly channel flow, rotating free surface,
free-rising bubbles, and slug flow in pipes. It was concluded that
the use of the three different techniques can extend the applicabil-
ity of ultrasound for detection of gas–liquid interfaces that move
passively and actively in two-phase flows.
The ultrasonic technique has also been used to measure the dis-
persed phases concentrations in three-phase and two-phase flows
(air–oil; glass beads – oil; and air–glass beads–oil) [30,31]. Investi-
gators in [30] observed the signal attenuation to increase exponen-
tially with increasing interfacial area of both the solid and gas
phases. The air bubbles were said to have caused an increase in
the acoustic wave transit time. Besides confirming the observa-
tions in [30], investigators in [31] also verified that the statistical
scatter in the transit time and in the signal amplitude increased
as the gas holdup increased. More recently, Carvalho et al. [32] pre-
sented a detailed discussion of the ultrasonic signal waveforms and
how they relate to the computation of the acoustic parameters; the
authors then correlated the attenuation of the acoustic signals with
void fraction for vertical, upward water–air flows in 2-in. ID acrylic
pipes. Tanahashi et al. [33] extended on this work and demon-
strated that the ultrasonic technique has the potential to provide
real-time information about the structure of intermittent horizon-
tal slug flows as well as estimates of the individual gas and liquid
flow rates for this particular flow pattern. Gonçalves [34] obtained
ultrasonic data for vertical, upward, liquid–gas–solid flows in 1-in.
and 2-in. ID acrylic pipes. Mineral oil was used as the continuous
phase and the flow patterns evolved from bubbly to annular flow.
The acoustic data reduction method initially presented in [32] was
improved for application to these multiphase flows more represen-
tative of the oil industry.
Zhai et al. [35] used an ultrasonic technique to measure the oil
fraction in water-continuous water–oil flows in a vertical 20-mm
pipe; the oil fraction was as high as 65%. The authors used two
ultrasonic sensors at 0° and 180° in transmission mode only and
observed the ultrasonic field to be affected by the size of oil dro-
plets. In addition, by using symbolic dynamic filtering (SDF) based
on analytic signal-space partitioning (ASSP), the relationship
between the average amplitudes of the ultrasonic signals with
the oil fraction could be made linear for oil fractions up to approx-
imately 35%. Meribout et al. [36] used ultrasonic sensors in combi-
nation with impedance sensors to determine the oil and water
fractions of oil–water–gas flows. The gas fraction was obtained
from a previously calibrated Venturi meter and pressure probes;
the total mass flow rates were divided in ranges for further data
reduction and determination of the oil and water fractions. No
information was provided on the oil used. The ultrasonic instru-
mentation consisted of low (4 MHz) and high frequency
(23 MHz) sensors to be used with high and low gas fractions,
respectively. From preliminary results, the authors verified that
more accurate oil and water cut measurements could be obtained
by using the ultrasonic technique in the 33–46% and 82–100%
water cut ranges; in the remaining ranges, the capacitance/conduc-
tance probes provided more accurate results. Data reduction to
obtain individual flow rates was accomplished by means of artifi-
cial neural networks (ANNs), one for each total mass flow rate
range. It was then possible, for the operating conditions used and
a 500 Hz sampling rate, to track real-time variations in the water
and gas fraction as well as in the total mass flow rate.
The drawback of current ultrasonic techniques, as is the case
with most MFM methods, is the need for prior signal calibration.
A lengthy discussion on calibration of multiphase flow meters
(MPFMs) can be found in [37]. As discussed by this author, the pri-
mary measurement elements that make up a MPFM can usually be
calibrated according to standard procedures, similar to those used
for single-phase flow measurements. However, contrary to single-
phase meters, the output of the primary measurements of a MPFM
is used as the input to advanced signal-processing algorithms for
calculation of individual phase flow rates. Calibration is then lim-
ited by the fact that each oil field is different and no flow loop fluid
used for calibration will be representative of an application unless
those particular field fluids are brought into the loop and operated
at field pressure and temperature. Hence, a specific product used as
calibration fluid for a particular application may not be representa-
tive of any other product or well stream [37].
In view of these calibration issues, a broader solution could be
the use of artificial neural networks (ANNs). ANNs are considered
an alternative tool for the automatic identification of flow patterns
[3] and since the early 80s they have been used extensively for
applications such as adaptive control, model based control, process
monitoring, fault detection, dynamic modeling, and parameter
estimation [38]. ANNs provide a non-linear mapping between
input and output variables and the cross-correlation among these
variables. A detailed description of a combination of ANNs and
MFM, using a venture tube and a density meter, can be found in
[3]; the ANNs provided higher accuracy for flow rates predictions
than fluid mechanics models. Shaikh and Al-Dahhan [39] devel-
oped an ANN to predict the overall gas holdup in bubble column
reactors using dimensionless fluid mechanics groups as input. Rosa
et al. [40] performed a direct comparison of three different types of
ANNs (multilayer perceptron, MLP; radial basis function, RBF; and
probabilistic neural network, PNN) and expert systems to predict
the flow patterns of vertical upward water–air flows based on
measurements from a wire resistivity probe. The ANNs exhibited
similar performances with percentages of correct predictions in
the 96–100% range. The expert systems exhibited performances
similar to the ANNs, but the rule creation and the encoding process
can be cumbersome with this methodology. Filletti and Seleghim
[41] numerically simulated the propagation of an acoustic pulse
through a water continuous water–oil mixture by discretization
 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50
of the wave equation in a fluid; the numerical results were then
used to train a neural network to provide the dispersed phase frac-
tion and interfacial area. The effect of the geometrical distribution
of the phases was taken into account by the local acoustic wave
velocity in the discretization scheme. Inoue et al. [42] developed
an ANN for identification of two-phase flow patterns (discrete bub-
bles, stratified flow, slug flow, and annular flow) using as input five
dimensionless groups that account for the competing forces occur-
ring in a multiphase flow. Care was taken to include in these
groups only variables that are routinely available in the oil indus-
try. An initial database was created for training and validating the
ANN; this database includes data from the literature for several
fluid and geometry combinations and the laboratory data of
Carvalho et al. [32] and Gonçalves [34].
In this context, the general objective of the present work is to
develop ANNs for flow pattern and GVF predictions based on
acoustic measurements using a relatively simple ultrasonic appa-
ratus to be discussed below. This apparatus should simultaneously
capture different aspects of the acoustic attenuation and transmis-
sion phenomena (pulse-echo, direct transmission, and scattering);
in addition, the apparatus should have the potential to provide real
time measurements of the flow pattern and the GVF in practical
applications of the oil industry. The specific objectives are two-
fold: (i) to present and analyze a comprehensive set of acoustic
attenuation data for vertical, upward, oil-continuous multiphase
flows in 1-in. and 2-in. acrylic pipes for flow patterns ranging from
bubbly flows to annular flows; (ii) to develop neural networks for
flow pattern recognition and GVF determination using the acoustic
attenuation data as input.
Fig. 2. Schematic view of the multiphase flow test rig.
33
2. Experimental apparatus and procedure
2.1. Multiphase flow test rig
The multiphase flow test rig is depicted in Fig. 2. Experiments
can be run in either a 1-in. or a 2-in. ID vertical acrylic pipe; the
ultrasonic station is located 6 m from the bottom end of the pipe
and a fast closing valves device is used for measurement of the
GVF. This system closes up approximately 1-meter section of the
pipe around the ultrasonic test section. A high-speed filming sta-
tion is located just downstream of the ultrasonic test section. The
continuous phase can be either water or oil; a separate tank with
a mixer is used for suspensions. The flow rates can be adjusted
by frequency inverters that control the pumps RPM. At the bottom
of each acrylic pipe there is a porous medium through which com-
pressed air is injected into the flow.
The liquid flow is measured by a Rheonik RHM12 model,
Coriolis-type flow meter (from 0.75 up to 100 kg/min ± 0.2%),
which includes a ±1 °C accurate thermometer. The air flow rate is
measured by a 3–30 ± 0.1 LPM Cole-Parmer, variable area, acrylic
flow meter. Pressure inside the pipes was measured by 0–4
psi ± 0.2 psi Bourdon pressure gauges; atmospheric pressure mea-
surements came from a barometer. The air at the injection point
was assumed to be at ambient temperature as measured by a
±0.5 °C accurate bulb thermometer; this assumption underesti-
mates the compressed air temperature but the error in the calcula-
tions of the air flow rate at the ultrasonic station is not significant.
The high-speed camera is a 512  512 resolution, 5000 fps RED-
LAKE Motion Pro X4 model.
34 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50

2.2. Ultrasonic instrumentation and stored in Microsoft Excel spreadsheets for plotting and
analysis.
The ultrasonic instrumentation was kept as simple as possible
(Fig. 3) as it is intended for use in real-time applications of the
oil and gas industry. The use of only four transducers allows for 2.3. Experimental procedure
less CPU usage during testing and smaller data files. Further details
about the selection of the number of transducers and their location Table 2 shows the flow types tested in the multiphase flow test
are given in Section 3.2.6. rig. Data were obtained for the 1-in. and 2-in. pipes for various
A schematic view of the ultrasonic data acquisition system is combinations of phases. The use of polyamide was intended for
shown in Fig. 4. Panametrics Videoscan 2.25 MHz, 13 mm diameter velocity measurements using the Doppler technique; these results
ultrasonic transducers were mounted on acrylic adapters as shown will be discussed in a separate paper. Table 3 shows the ranges of
in Fig. 3. The 0° sensor was a dual element, model DHC706-RM operating conditions tested. The GVF was made to vary between 0%
pulse-echo transducer and the 45°, 135°, and 180° sensors were and 85% for the 1-in. pipe while for the 2-in. pipe it varied from 0%
single element, model V106-RM receivers. The pulser used was to 35% due to limitations in the test rig. In addition, due to the dif-
Olympus 5077PR, which allows for adjustments of the pulse width ficulties in working with crude oils, experiments were run with
and thus a more efficient excitation of V106-RM transducer crystal. mineral oil as the continuous phase (q = 814.5 kg/m3 and
This increases the sound intensity of the emitted signals, which are l = 41.36 mPa s at 20 °C). Glass beads, 80–600 lm in diameter,
sinusoidal and last approximately 1.5 ls. However, how long the were used to simulate the sand contents. The values for the liquid
V1060-RM pulse-echo transducer will oscillate in a given situation
depends on factors such as damping material, test material proper- Table 2
ties, and voltage. A detailed discussion of the signals obtained in Types of flow tested in the multiphase flow test rig.
the present investigation is given in Section 3.3 (Analysis of Ultra-
Pipe ID Flow type Component
sonic Signals). Ultrasonic data were acquired by means of a National
(in.) identification
Instruments PXIe-1062Q acquisition board using the resident Lab- Mineral Air Polyamide Sand Water
Oil (0.1% w/w) (1% w/w) (10% v/v)
VIEWÒ software. Connections #1 through #4 in Fig. 4 send the
received acoustic signals to the PXI, which in turn transfers the 1 O–A X X
O–A–P X X X
data to the computer for further manipulation. Connection #5
O–A–P–S X X X X
ensures synchronization between the pulser and the PXIe-1062Q. O–A–P–S–W X X X X X
Hence, every time an ultrasonic pulse is sent out, the pulser signals
2 O–A X X
it to the PXI so that proper time windows for data acquisition from O–A–W X X X
each sensor are then set up as discussed in Section 3.3. Connection O–A–W–S X X X X
#8 ensures synchronization between acquisition of the high-speed
O–A: Oil–Air, O–A–P: Oil–Air–Polyamide, O–A–P–S: Oil–Air–Polyamide–Sand,
filming and the ultrasonic data. The data was reduced in MatLabÒ O–A–P–S–W: Oil–Air–Polyamide–Sand–Water, O–A–W: Oil–Air–Water, O–A–W–S:
Oil–Air–Water–Sand.

Table 3
Operating conditions for the experiments in the multiphase flow test rig.

Pipe ID Flow type Parameter

(in.)
GVF jl jg Toil ptestsection
(%) (m/s) (m/s) (°C) (kgf/cm2)

1 O–A 0–83 0.10–0.30 0–0.17 24 0.5

O–A–P 0–85 0.10–0.26 0–0.37 22 0.4
O–A–P–S 0–84 0.10–0.30 0–0.10 21 0.4
O–A–P–S–W 0–82 0.10–0.28 0–0.16 24 0.7
2 O–A 1–33 0.05–0.17 0–0.02 30 0.4
O–A–W 0–36 0.04–0.17 0–0.04 31 0.5
O–A–W–S 0–36 0.04–0.17 0–0.02 27 0.5
Fig. 3. Schematic view of the ultrasonic instrumentation.

Fig. 4. Schematic view of the ultrasonic data acquisition system.

 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50
superficial velocities, jl, are only approximate as the solid contents
would flow along with the liquid through the coriolis meter.
The ultrasonic data acquisition rate was set to 20 MHz in order
to avoid any aliasing effects. The sampling interval and the pulse
generation rate were 7.5 s and 2 kHz, respectively, which makes
for 15,000 pulses in each acoustic sample. The 7.5-s sampling
interval allows for representative samples of the flows topology
to be obtained as discussed above. The 2 kHz pulse generation rate
by the 5077PR pulser was determined by considering the flow dis-
placement between pulses for the maximum superficial velocities
in Table 3 and transducers overheating limitations.
An attempt was made to vary the GFV in approximately 2.5%
increments and keep a fixed number of five acoustic samples for
each condition. (Two consecutive samples were separated by
approximately two minutes, which was the time necessary to
transfer the previously acquired data to the computer hard disk.)
However, because the test rig was not a closed loop system, deple-
tion of fluids in the tanks forced in-time adjustments to the testing
procedure. In addition, the liquid and gas flow rates for the desired
GVF were previously calculated based on the homogeneous model,
which did not correspond to the actual GVF measured by the fast
closing valves device. Therefore, the resulting GVF increments
and the number of acoustic samples for each condition varied as
the flow patterns evolved. Temperatures measurements were
made at the beginning and end of each test series in order to obtain
the fluid properties.
3. Results and discussion
3.1. Discussion of visual data
Knowledge of the flow patterns greatly simplifies the analysis
and understanding of events showing up in the ultrasonic signals.
(a) (b) (c)
Fig. 5. Oil–air flow patterns in the 100 pipe: (a) bubble flow; (b) cap bubbles; (c) stable slug flow; (d) unstable slug flow; (e) churn flow; (f) annular flow.
(a) (b)
Fig. 6. Oil–air flow patterns in the 200 pipe: (a) small cap bubbles; (b) large cap bubbles; (c) stable slug flow; (d) unstable slug flow.
35
Fig. 5 shows the observed oil–air flow patterns in the 100 pipe. For
GVF values from 0% to approximately 4%, bubbly flow was the
observed pattern (Fig. 5a); bubbles were approximately 2 mm in
size. For GVF around 4%, coalescence started taking place and cap
bubbles were already present. Dispersed bubbles and cap bubbles
coexisted in the pipe up to about 14% GVF (Fig. 5b). The cap bub-
bles were approximately 20 mm in diameter and just as long while
the dispersed bubbles were in the 0.5–2.5 mm range. For GVF val-
ues in the 14–30% range, slug flow was observed with smooth
stable interfaces (Fig. 5c). The Taylor bubbles were slightly over
20 mm in diameter and about 200 mm long. In the liquid slugs,
the dispersed bubbles were the same size as before, i.e. in the
0.5–2.5 mm range. From 30% to 45% GVF, the Taylor bubbles inter-
face became irregular and started breaking up into small bubbles,
approximately 2 mm in size (Fig. 5d). For purposes of analyzing
the ultrasonic signals in the sections to follow, the flow patterns
in the 14–30% range will be called simply slug flow while the pat-
terns in the 30–45% range will be called unstable slug flow. For GVF
in the 45–75% range, churn flow existed in the pipe (Fig. 5e). Above
75% annular flow was the observed pattern (Fig. 5f); the liquid film
was approximately 2.5 mm thick and the dispersed bubbles were
very small, from 0.2 to 0.8 mm. In the experiments with three-
phase oil–air–solids flows and four-phase oil–air–water–solids
flows, the observed patterns were essentially the same as those
for oil–air flows.
In the 200 pipe, the flow patterns were less clear due to water
emulsification in the oil. However, due to the increased turbulence
bubbles could be observed to coalesce more easily than in the 100
pipe to form small cap bubbles (Fig. 6a). These bubbles increased
in size for GVF up to 12% and were surrounded by smaller bubbles
about 1–10 mm in size (Fig. 6b). Smooth interface Taylor bubbles
first appeared at 14% GVF (Fig. 6c) and increased in size up to about
35% GVF (Fig. 6d), also surrounded by 1–10 mm bubbles.
(d) (e) (f)
(c) (d)
36 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50
For each operating condition, the flow pattern was identified by
careful analysis of the visual data. Flows for which there were only
small discrete bubbles in a narrow size range were classified as
bubble flows; the same classification was applied even if small
cap bubbles were present. When these cap bubbles would grow
into much bigger cap bubbles, taking up a substantial portion of
the pipe cross-section, or incipient Taylor bubbles, the flows were
considered having entered the intermittent regime. This was also
the classification of flows with developed Taylor bubbles at higher
GVF values. Once the Taylor bubbles became unstable and started
breaking into small bubbles and the flow as a whole exhibited a
pulsating behavior, churn flow was the classification adopted.
Finally, when the gas phase was segregated into a continuous core
in the pipe inner region surrounded by a liquid film on the pipe
wall, annular flow was the obvious classification. These four flow
patterns were the ones used for identification by the neural net-
works discussed in Section 4. It was preliminarily verified that
the use of churn flow as a separate category allowed for more accu-
rate results by the pattern recognition algorithm than when these
flows were included in the intermittent flow pattern category.
3.2. Acoustic dissipation mechanisms and definition of acoustic
parameters
The derivation for the acoustic field generated by a baffled plane
circular piston can be found in [14], which is a geometry that
approximates the one in the present investigation. Closed-form
solutions of this field exist for the acoustic axis (a line perpendicular
to the face of the piston and passing through its center) and at suf-
ficiently large distances from the piston, in the far field (Fig. 7a). In
the near field, strong interference effects are present and the pres-
sure amplitude fluctuates between zero and 2q0 cU 0 . In the far field,
the acoustic beam diverges and this behavior is quantified by a half
angle spread between 6 dB points, a/2 (Fig. 7b) The border
between the near and far fields is denoted as N. Therefore, the
acoustic field generated by circular pistons results in a wave that
is definitely non-planar, especially in the near field.
For the Olympus-Panametrics Videoscan, ½00 diameter, 2.25 MHz
transducers in mineral oil, calculations yielded N = 63 mm and
a/2 = 1.5°. These calculations also showed that for both the 1-in.
and 2-in. pipes the multiphase flows is entirely located in the near
field. In view of the practical difficulties to construct an experimen-
tal ultrasonic apparatus to place the multiphase flow in the far field,
preliminary experiments were carried out to verify this effect on the
measured attenuation. The experimental data showed no measur-
able differences; hence, for ease of construction the ultrasonic
assembly was constructed as shown in Fig. 3.
On the other hand, analysis of the acoustic phenomena becomes
more difficult as expressions for acoustic parameters are given for
the far field. The discussion to follow is based on plane or near
Fig. 7. Acoustic field generated by a plane circular cylinder: (a) axial response; (b) half angle spread [43].
plane waves (far field) and it is meant only as a guide for reasoning
on the relative importance of the various acoustic phenomena in
the present apparatus.
3.2.1. Acoustic dissipation in the continuous phase
In multiphase flows, acoustic dissipation results from the com-
bined effects of dissipation in the continuous phase as the sound
wave propagates through the continuous medium and dissipation
at the interfaces present in the flow. In the continuous phase, the
acoustic intensity decays according to the following equation [14]:
2
IðxÞ ¼ ðP0 eax Þ =2q0 c ¼ Ið0Þe2ax ð1Þ
where P0 and Ið0Þ are, respectively, the acoustic pressure amplitude
and intensity at x = 0 and a is the absorption coefficient in nepers
per meter (Np/m).
Absorption in the continuous phase occurs due to many differ-
ent mechanisms depending on the specific fluid at hand, e.g. vis-
cous and thermal absorption associated with the compressions
and rarefactions of the fluid by the acoustic wave, molecular ther-
mal relaxation, structural relaxation (related directly to volume
changes – and not to temperature changes – in polar liquids such
as water), chemical relaxation, and so on. The classical acoustic
theory considers the viscous and thermal absorption due to the
compressions and rarefactions of the fluid to be the only mecha-
nisms present so that the absorption coefficient is given by:
 
x2 4 ðc  1Þk
ac ¼ as þ ak ¼ gþ ð2Þ
2q0 c3 3 cp
The classical absorption theory has been verified to be in good
agreement with results for viscous liquids; therefore, it is expected
to hold for oils. For the mineral oil used in these experiments,
assumed at 20 °C, and a 2.25 MHz wave, the result obtained for
the absorption coefficient is a = 20.6 dB/m. Calculations were
repeated for a 1 MHz wave and the result a = 4.06 dB/m is consis-
tent with experimental value for light oil samples presented in
[44], a = 4.33 dB/m.
3.2.2. Acoustic dissipation by gas structures
As a beam of sound enters a liquid containing gas bubbles, it can
be attenuated by interfacial phenomena such as reflection, refrac-
tion, absorption, and scattering. When subjected to a sound wave, a
bubble can be driven into radial oscillations, which, depending on
their amplitude, can reradiate a significant amount of energy out of
the sound beam [14]. This energy is radiated uniformly in all direc-
tions. In addition, temperature gradients near the interface can
cause significant energy losses (absorption). Modeling of these pro-
cesses assume the bubble radius to be much smaller than the
acoustical wavelength, i.e. the bubbles are assumed to be lumped
acoustic elements. The bubbles oscillations are most intense at res-
onance, at which point their frequency is given by [14]:
 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50
sffiffiffiffiffiffiffiffiffiffiffi
1 3cP b
x0 ¼
a q0
For air bubbles in mineral oil at 20 °C and atmospheric pressure,
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
this leads to k0 a ¼ ð1=cÞ 3cPb =q0 ¼ 0:0153, which shows the
radius of resonant bubbles under these conditions to be much
smaller than the acoustical wavelength. For the 2.25 MHz wave
used in the present investigation, the bubble radius for resonance
is 1.6 lm. From the discussion of the visual data above, there
was no indication of the presence of microbubbles. Discrete bub-
bles were in the millimeter range and the assumption of lumped
acoustic elements no longer holds.
When the scattering object is large compared with the wave-
length of the scattered sound, the sound is said to be reflected
and diffracted rather than scattered. Behind the object, there is a
shadow where the pressure amplitude is vanishingly small; in
front or to the sides, in the ‘‘illuminated region”, there is a combi-
nation of the incident wave and the wave reflected from the sur-
face of the scattering object [16]. There are substantially less
studies to determine the interaction of bubbles with acoustic
waves having wavelengths less than or of the same order of mag-
nitude as the bubble diameter. Nishi [45] obtained an analytical
solution, valid for all wavelengths, for a spherical bubble in a vis-
cous liquid subjected to a plane acoustic wave. From this work, it
is possible to obtain the total scattering area (or total extinction
area) of a bubble, which is defined as the total energy removed
from a wave by radiation and absorption divided by the intensity
of the incident wave. The resultant expression is,
4p X
1 2
jl ðkaÞ
SðkaÞ ¼ 2
ð2l þ 1Þ 2
k l¼0 jl ðkaÞ þ n2l ðkaÞ
where k is the wave number of the ultrasound surrounding the
bubble.
When the frequency of the sound wave is somewhat higher
than the resonance value, the extinction cross section approaches
4pa2 (the surface area of the sphere). For frequencies much higher
than the resonant frequency, the radiation is no longer spherically
symmetric and the total extinction cross section and the bubble
geometric cross section become nearly equal [14]. Hence, even in
the case of high frequency ultrasound, the total scattering area var-
ies significantly with bubble size and the intensity scattered by a
given bubble population will be very dependent on the bubble size
distribution. Jones et al. [25] presented a model for the interaction
of an ultrasonic wave with a bubbly mixture. The resultant equa-
tion is,
 
CS
T ¼ exp  x ¼ exp½2ax
8A
C S
a¼
16 A
where T is the amplitude of the transmitted signal normalized by
the amplitude of the transmitted signal in the absence of bubbles
(single-phase flow) and A is the geometrically projected cross sec-
tional area of the bubble.
Several assumptions restrict the applicability of this model,
namely, (i) the bubbles are isolated scatterers, i.e. there is no acous-
tical interaction between them, (ii) the bubbles are spherical, and
(iii) forward scattering (contribution of the scattered energy to
the total transmitted signal) is negligible. The authors estimate
the applicability of their model to be restricted to bubbly flows with
void fractions less than ten percent. Varadan et al. [46] presented a
model for the acoustic dissipation in a bubble swarm that accounts
for multiple interactions between the scatterers; comparison with
experimental data for bubbles in water showed excellent agree-
ment. However, the model does not lend itself to ready
37
calculations of the absorption coefficient in the liquid-gas medium;
ð3Þ calculations were then made based on the model by Jones et al. [25]
and the result is intended only as a rough guide for reasoning. For
GVF = 2.5% (Fig. 5a) the estimated bubble equivalent diameter is
d = 1.3 mm; the estimated value for the spatial absorption coeffi-
cient for the bubble swarm in mineral oil at 20 °C is a = 20.3 dB/m.
For larger gas structures, such as Taylor bubbles and the gas
core of annular flows, reflections of the sound beam at the oil–air
interfaces are expected to be the main acoustic dissipation mech-
anism. Even though the measurements in the present investigation
did not involve plane waves and the interfaces are not flat, equa-
tions for normal incidence of harmonic plane waves on flat inter-
faces provide clues to understanding the behavior exhibited by
the ultrasonic data. Under these simplifying assumptions, the
intensity reflection and transmission coefficients are given, respec-
tively, by:
 2  2
1  r1 =r 2 r2 =r 1  1
RI ¼ ¼ ð6Þ
1 þ r1 =r 2 r2 =r 1 þ 1
r 1 =r2 r 2 =r 1
TI ¼ 4 ¼4 ð7Þ
ð1 þ r 1 =r2 Þ2 ðr 2 =r1 þ 1Þ2
In these equations, the subscripts 1 and 2 refer to the media
making up the interface. Table 4 shows values of the acoustic
reflection and transmission coefficients at the interfaces of interest
in this study. In view of the fairly high value of the transmission
coefficient at oil–acrylic interfaces (T = 0.777), a significant portion
of the acoustic energy of the pulses emitted by the 0° sensor is
expected to enter the multiphase flow. Oil–air interfaces are seen
ð4Þ
to have a reflection coefficient very close to unity and are thus cap-
able of very strong reflections of the acoustic beam.
In view of the dissipation mechanisms discussed above, Fig. 8
depicts four distinct situations of great importance in the present
investigation, namely, (a) when the cross-section is sparsely popu-
lated with small bubbles; (b) when the cross-section is densely
populated with discrete bubbles; (c) when a large gas structure
is present such as cap or incipient Taylor bubbles; and (d) when
the gas core of annular flows takes most of the pipe cross section
leaving only a thin liquid film around it. From the discussion above,
it is suggested that when the cross-section is populated with only
dispersed bubbles (Fig. 8a) there is very little acoustic energy scat-
tered to the sides. The energy is not uniformly scattered as even
small bubbles are much larger than the size required for reso-
nance; there would thus be a tendency for a more intense scattered
acoustic field toward the emitter transducer despite the random-
ness of the bubbles motion. As the acoustic field is ‘‘pierced” by
ð5aÞ only very few small bubbles, a substantial amount of the acoustic
energy is transmitted through the continuous phase to the oppo-
ð5bÞ site side of pipe wall. A reflection then takes place at the acrylic–
oil interface and the reflected wave travels back to the emitter.
When the cross-section gets more densely populated with rela-
tively larger bubbles as in bubbly flow (Fig. 8b), the energy scat-
tered to the sides increases and the scattered acoustic field as a
whole is less affected by the randomness of the bubbles motion.
Hence, its non-uniformity could be more easily measured around
Table 4
Values of the reflection and transmission coefficients for plane waves impinging on
flat interfaces of interest in this study.
Interface RI TI
Oil–Acrylic 0.223 0.777
Oil–Air 0.999 0.001
Oil–Glass 0.645 0.355
Oil–Water 0.014 0.986
38 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50
Fig. 8. Schematic representations of the sound field: (a) dispersed bubbles; (b) bubbly flow; (c) cap bubbles and incipient Taylor bubbles; and (d) annular flow.
the pipe circumference. Blockage of the acoustic beam increases
and the amount of energy transmitted through the continuous
phase decreases accordingly; and so does the amount of energy
carried by the wave reflected off the opposite oil–acrylic interface.
A distinct picture would represent the acoustic dissipation
mechanisms when there are significantly larger coalesced gas
structures in the pipe cross-section (Fig. 8c) such as cap bubbles
or incipient Taylor bubbles. The acoustic field is now made mostly
by reflections of the ultrasonic beam that bounces back and forth
between the oil–air and the oil–acrylic interface along the pipe cir-
cumference. Upon each reflection at the latter interface, a substan-
tial amount of energy is lost to the pipe wall. Again, the acoustic
intensity is not uniform along the pipe circumference, being higher
in the regions closer to the emitter transducer. The portion of the
acoustic energy that is directly transmitted through the continuous
phase decreases as the coalesced gas structures get bigger. Addi-
tionally, dispersed small bubbles also contribute to scatter this
energy. The portion to the acoustic beam reflected straight back
off the oil–air interface of the larger gas structure increases, thus
conveying very signification information about the development
of the flow patterns.
Finally, the dissipation mechanisms for the annular flow pattern
could be understood as a further development of the mechanisms
pictured above. As shown in Fig. 8d, the liquid film on the wall is
much thinner than in the previous case and the ultrasonic beam
now has to bounce back and forth many more times as it pro-
gresses forward. Consequently, it dies off much faster. The thinner
the liquid film, the less energy gets to the rear portions of the pipe
wall. The portion to the acoustic beam reflected straight back off
the oil–air interface of the gas core would be even larger than in
the previous case.
These schematic views of the acoustic dissipation mechanisms
are consistent with the trends in the ultrasonic data for the various
sensors discussed in Sections 3.4 and 3.5.
3.2.3. Acoustic dissipation by solid particles
As discussed in [47], the most well-known acoustic theory for
heterogeneous systems was developed by Epstein and Carhart
[48] and Allegra and Hawley [49], and is termed the ‘‘ECAH the-
ory”. This theory takes into account the four most important mech-
anisms (viscous, thermal, scattering, and intrinsic) for acoustic
dissipation in a monodisperse system of spherical particles and is
valid only for dilute systems. The term ‘‘dilute” means there is no
consideration of particle–particle acoustic interaction. Particles
oscillating under the sound wave give rise to viscous and thermal
boundary layers, which damp in the particle vicinity; there is no
particle interaction when the corresponding boundary layers are
sufficiently separated.
Kytömaa [50] further developed this theory for application
specifically to suspensions of solids in liquids. This author identi-
fies three different regimes of sound propagation in suspensions
depending on the dominant length scale involved in the process.
The multiple-scattering regime (short wavelength regime) occurs
for ka >> 1 (k is the wave number and a is the particle radius),
which means that the particles are large compared to the wave-
length and sound scatters uniformly in all directions. On the other
hand, when the particles are much smaller than the wavelength
(ka << 1), the particle moves under the action of the acoustic wave
and the length scale is the thickness of the boundary layer sur-
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
rounding the particle, d ¼ 2g=ql x.
This boundary layer thickness, often called the Stokes layer
thickness, in turn leads to a Reynolds number defined as
Re ¼ a=d. When Re << 1, the boundary layer is thick compared to
the particle radius and the viscous relaxation time is shorter than
the period of excitation. This particular situation is referred to as
the viscous regime in which the drag is well represented by the
Stokes law. When Re >> 1, the particle radius is much larger than
the boundary layer thickness and the fluid outside it is governed
by inviscid inertial effects (the inertial regime). The dominant loss
term is the in-phase component of the Basset force. Additional
effects such as unsteady thermal conduction through a thermal
boundary layer surrounding the particle and resonant behavior
are small in liquid–solid systems and can be neglected.
In the long wavelength limit (ka << 1 and multiple scattering
effects not present), traveling wave solutions were sought in the
form q ¼ q0 exp½jðxt þ kxÞ. In this equation, k ¼ x=c þ ja is the
 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50 39

complex wave number and a is the spatial attenuation coefficient. Table 6

A rather lengthy procedure to obtain k and a was then obtained. Contributions of the individual absorption mechanisms to the overall dissipation of a
2.25 MHz acoustic wave in the multiphase flows at 20 °C.
Atkinson and Kytömaa [1] further developed this model for the
particular case when viscous interaction between the solid parti- Loss type as (dB/m) ascattering (dB/m) %
cles are absent, i.e. when the wave frequency is high enough to Mineral oil intrinsic dissipation [14] 20.6 – 42.1
make the boundary layer so thin that the inertial regime prevails. Air bubbles in mineral oil [25] 20.3 41.5
This condition occurs for: Glass beads in mineral oil [1] 6.64 13.6
Polyamide particles in mineral oil [1] 0.14 0.3
2 3
2
Water droplets in mineral oil [2] 1.23 2.5
4g 6 ðm=mmax Þ 7
3
x 4  5 ð8Þ Total 100.0
a2 ql 1 2
1  ðm=mmax Þ3

The maximum particle packing fraction is mmax  0:635  0:005 The acoustic dissipation by the polyamide particles was also
for a random close-packed structure of monodisperse spheres and calculated using the model by Atkinson and Kytömaaa [1]. The
mmax  0:555 for random loose packing [51] For the conditions average particle size adopted was 200 lm based on information
associated with large x, the spatial absorption coefficient, a, can provided by the manufacturer. Results for the absorption coeffi-
be approximated by: cients of both the glass beads and the polyamide particles can be
2
seen in Table 6.
F ðk Þ
aðxÞ ¼ F ðkÞ  qffiffiffiffiffiffiffiffiffiffiffiffiffi ð9Þ
2
2 Rðk Þ 3.2.4. Acoustic dissipation by water droplets
Even though the dilute system assumption in the ECAH theory
restricts its general applicability to volume fractions of a few per-
2 x2 A2 q þ x2 ½B þ ð1  mÞq ½Bq þ qs ql ð1  mÞ
Rðk Þ ¼ ð10Þ cent, there is some evidence that, in very specific situations, it does
j A2 þ x2 ½B þ ð1  mÞq 2 provide a correct interpretation of experimental data for volume
fractions as large as 30% [47]. The validity of the ECAH theory for
2 x2 Axð1  mÞðqs ql  q q Þ moderately concentrated system has only been demonstrated in
F ðk Þ ¼ ð11Þ
j A2 þ x2 ½B þ ð1  mÞq 2 systems where the thermal losses were dominant such as emul-
sions. Herrmann et al. [52] presented an experimental study of
In the equations above, R and F represent the real and imagi- sound attenuation in quasi-monodisperse emulsions of two differ-
nary parts, respectively, of a complex number. For the expressions ent oils in water. The oils used were n-hexadecane and polysilox-
of parameters A and B as well as the suspension properties j , q
, ane and their concentrations were up to approximately 60% by
and q , the reader is referred to Atkinson and Kytömaa [1]. Theo- volume; many emulsion samples were prepared with particle size
retical results were compared to experimental data for a wide covering a wide range from 20 to 900 nm. The transducers used
range of ka. As ka increases from small values to unity, geometric spanned the 0.5–10 MHz frequency range. The authors verified
and multiple scattering effects, which were unaccounted for in that existing theories for sound attenuation in heterogeneous
the model, caused the attenuation to be up to one order of magni- media exhibited excellent agreement with the experimental data
tude greater than the theoretical predictions. for concentrations up to 15%.
Table 5 below shows the particle size distribution in the present One of the theories Herrmann et al. [52] used to compare their
investigation and the corresponding acoustic and hydrodynamic experimental data with is the model due to Isakovich [2]. This
parameters; in all cases d = 2.7 lm. Approximately 92.5% of the author proposed a simple formulation of the thermal-elastic effect
particles are in the 212–590 lm range with 1.1 < ka < 2.9, d a, in the cases where other losses can be neglected. In the case of
and 39 < Re < 110. These results for ka do not correspond to the spherical particles, Isakovich [2] obtained the following expres-
multiple scattering regime for which ka  1 while the results for sions for the acoustic wave vector:
   
2 1 2
Re and the ratio x= a42lq ðm=mmax Þ3 1  ðm=mmax Þ3  1 show "  2 #1=2
l
3mq0 c2
x b1 b2
the acoustic dissipation to be in the inertial regime. A weighted KðxÞ ¼ 1þj 2 T0  A ð12Þ
c a x q1 cp1 q2 cp2
average spatial attenuation coefficient of approximately 4 dB/m
was obtained for this suspension according to the model by Atkin-
son and Kytömaa [1]. Even though values of ka are greater than k1 k2 ½ð1  jÞn2 a þ 1fð1  jÞn1 a  tanh½ð1  jÞn1 ag
A¼
unity in the present case, the solids concentration is very small k1 fð1  jÞn1 a  tanh½ð1  jÞn1 ag þ k2 tanh½ð1  jÞn1 a½ð1  jÞn2 a þ 1
so that multiple scattering effects should not be very intense. ð13Þ
Therefore, for purposes of general reasoning the model can be
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi
expected to give a satisfactory estimate of the attenuation of the
ni ¼ xqi cpi =ki ð14Þ
sound wave by the glass beads.
In these equations, T 0 and q0 refer to the equilibrium values of
Table 5
the temperature and density, respectively, m is particles fraction by
Particle size distribution and corresponding acoustic and hydrodynamic parameters volume, and the indices 1 and 2 pertain, respectively, to the parti-
for 1% w/w suspension of glass beads in mineral oil. cle and the fluid matrix. The model by Isakovich [2] emphasizes the
Size range % wt. ka Re " x # ( 104) aðxÞ
role played by the quantity ni a, which is the product of the thermal
(lm) 4g
2
ðm=mmax Þ3 (dB/m) wave vector in medium i by the particle radius. Physically, this pro-
a2 ql 1 2
ð1ðm=mmax Þ3 Þ duct expresses the ratio of the particle radius to the extent attained
590–595 0.6 2.9–3.0
110 4.8–4.9
2.40 by the thermal wave. The propagation of the acoustic wave is con-
425–590 54.3 2.1–2.9 79–110 2.5–4.8 3.33–2.40 trolled by an attenuation coefficient, a, given by corresponding to
212–425 38.2 1.1–2.1 39–79 0.62–2.5 6.64–3.33 the imaginary part of the wave vector K. Herrmann et al. [52]
150–212 6.0 0.7–2.1 28–39 0.31–0.62 9.34–6.64
observed deviations between the experimental data for 10% v/v
<150 0.9 <0.7 <28 <0.31 >9.34
emulsions of n-hexadecane in water at high frequencies probably
40 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50

due to small viscous–inertial effects. These effects, not accounted
for in the model by Isakovich [2], arise from the non-negligible dif-
ference in the densities of water and n-hexadecane (997.0 and
773.0 kg/m3, respectively).
The result for the absorption coefficient from the Isakovich
model [2] for 10% v/v water in mineral oil and 100 lm particle
diameter is shown in Table 6.
3.2.5. Comparison of the theoretical predictions for the acoustic
dissipation mechanisms in oil-continuous multiphase flows
Table 6 shows a breakdown of the overall dissipation mecha-
nisms for a 2.25 MHz acoustic wave in a bubbly multiphase flow
similar to that in Fig. 5a. Also shown in the table are the references
from which the models were taken for calculations of the acoustic
absorption coefficients. First of all, the intrinsic dissipation in the
mineral oil as a continuous phase and scattering by the air bubbles
are seen to be most important contributing factors; the latter is
fairly high even though the GVF is only 2.5%. The contribution of
the glass beads is also important, its absorption coefficient being
over half of the previous ones. Acoustic dissipation by water dro-
plets is fairly small as the water concentration is still in the dilute
range. Even if small viscous–inertial effects arising from the differ-
ence in densities between the mineral oil and water (810.0 and
998.2 kg/m3, respectively) were present, the absorption coefficient
would still be very small. Finally, dissipation by the polyamide is
negligible as only trace amounts of these particles were added to
the flow.
Assuming that for the small concentrations involved the dissi-
pations mechanisms are additive, it can be concluded that the vari-
ations in the GVF will have a very significant effect on the overall
multiphase flow acoustic dissipation. Moreover, as the GVF
increases, the contribution of the gas phase will become even lar-
ger while the intrinsic dissipation in the mineral oil and by the
remaining dispersed phases will remain about the same. Changes
in these values will now be due mostly to changes in the acoustic
path as the gas phase occupies increasing fractions of the pipe
cross section (Fig. 8).
From these results, it is expected that acoustic attenuation data
for the multiphase flows in Table 2 can provide accurate measure-
ments of the GVF and reliable indications of the flow pattern. For
small GVF values (bubbly flows and incipient intermittent flows),
differences in the overall acoustic attenuation for the different flow
types in Table 2 are expected, mainly due to the presence or
absence of the glass beads. The artificial neural networks to be
developed for flow pattern recognition and GVF determination in
multiphase flows should be able to handle these differences. In
the remainder of this paper, the work developed to accomplish this
goal is described in detail.
3.2.6. Selection of the number and position of ultrasonic transducers
From the results in Table 6 and in Fig. 8, the acoustic dissipation
in the multiphase flows can be expected to be dominated by the
gas phase and the intensity of the acoustic field along the pipe cir-
cumference to be non-uniform. In view of these results, it was rea-
soned that four ultrasonic transducers were enough for properly
probing the resultant acoustic field in the multiphase flows. In
addition, the use of four transducers is compatible with the original
goal of using the ultrasonic technique for real-time measurements
of multiphase flows in the oil industry.
Only one emitter was deemed necessary and located at 0° as
shown in Fig. 3. One opposite receiver at 180° was intended to
measure mainly the attenuation of the acoustic field that traverse
the multiphase flow. Forward scattering, if significant at all, should
also be picked up by this transducer. A dual element (emitter–
receiver) transducer was selected for the 0° location to measure
the attenuation of the reflected wave at the opposite oil–acrylic
interface as it propagates backwards in the multiphase flow. Addi-
tionally, this transducer should capture the backward reflections
off large gas structures. A minimum of two receivers was judged
necessary to capture the non-uniformity of the scattered (diffuse)
field; they were positioned at the center points of the first and
second quadrants at 45° and 135°, respectively.
Finally, the flow was assumed to be radially symmetric from a
statistical point of view so that only half of the pipe circumference
was instrumented. Preliminary experiments were run to ensure
this set-up would allow for representative samples of the flow to
be taken. This was checked by increasing the sampling time and
observing the statistical scatter in the acoustic attenuation data
as a function of GVF for a fixed flow condition. The minimum sam-
pling time corresponded to the condition when no significant
decrease in the scatter was observed for further increases in the
sampling time. As related previously, in the present investigation
a 7.5-s sampling time was used, which allowed for representative
ultrasonic samples of the various multiphase flows to be taken.
3.3. Analysis of the ultrasonic signals
The acoustic trace signals captured by the ultrasonic apparatus
are made up of longitudinal and shear waves and base noise; how-
ever, not all waves carry meaningful information about the multi-
phase flow. Table 7 shows simplified calculations of the transit
times for the corresponding expected acoustic paths in the 2-in.

Table 7
Acoustic paths and transit times for the waves in the 2-in. pipe and mineral oil as the continuous phase.

Acustic Acoustic Length of acoustic Transit time Total Transit Ultrasonic Remarks
Wave path path (mm) (ls) Time (ls) Transducer
Acrylic Oil Acrylic Oil
LW SW LW
LW1 A-B-H-B-A 56.0 0.0 108.0 20.5 75.9 96.5 0° Longitudinal wave generated in A that goes
through the two-phase flow
A-B-?-B-A 56.0 0.0
20.5
20.5 < t < 96.5 0°
A-B-?-D-C 50.0 1.0
19.0
20.5 < t < 96.5 45°
A-B-?-F-E 50.0 0.0
18.3
20.5 < t < 96.5 135°
A-B-H-G 50.0 0.0 54.0 18.3 38.0 56.3 180°
LW2 A-B-A 56.0 0.0 0.0 20.5 0.0 20.5 0° Longitudinal wave generated in A that travels
through the acrylic only
SW1 A-B-H-B-A 28.0 28.0 108.0 29.8 75.9 105.8 0° Shear wave generated in B that travels through
the acrylic only
SW2 A-B-H-G 28.0 22.0 54.0 25.6 38.0 63.6 180° Shear wave generated in H from LW1; it travels
through the acrylic only
 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50
pipe; points A through H were shown in Fig. 3. The longitudinal
and shear wave velocities in the acrylic used in these calculations
were 2730 and 1430 m/s, respectively [43]. The velocity of the lon-
gitudinal wave in the mineral oil was 1422 m/s as measured by
Gonçalves [34].
Acoustic path A-B-H-G is associated with the longitudinal wave
LW1 that is transmitted from the emitter transducer at 0° to the
receiver directly opposite at 180° (Fig. 9). It can be observed that
the 56.5-ls calculated transit time agrees closely with the instant
this wave starts arriving at this sensor. It has been previously ver-
ified [32] that the pulse duration (the time interval between the
moment the wave first hits the receiver and the moment it dies
out) remains essentially constant despite increases in the GVF,
which helps establish clear bounds for the reception of the differ-
ent waves.
Upon reaching point H, LW1 gives rise to shear wave SW2 that
hits the 180° sensor shortly after LW1 itself. Therefore, both LW1
and SW2 are affected by and carry information about the multi-
phase flow; the red rectangle in Fig. 9 indicates the corresponding
time window for reduction of data acquired by this sensor. Outside
this window there is only base noise and the corresponding data
was discarded.
Acoustic path A-B-H-B-A is taken by the portion of the longitu-
dinal wave LW1 that is reflected at the oil–acrylic interface at H
and transmitted back to the emitter at 0° (Fig. 10). Once again,
the 96.5-ls transit time calculated in Table 7 agrees closely with
the instant this wave starts arriving at this sensor. Upon getting
back to point B, reflected LW1 gives rise to shear wave SW1 that
hits the 0° sensor shortly after LW1. Therefore, both LW1 and
SW1 carry meaningful information about the multiphase flow
and should be included in the time window for data reduction.
2 LW1
Signal Amplitude (v)
1
0
-1
-2
-3
0 10 20
Fig. 9. Time window for reduction of signals acquired by the 180° sensor in the 2-in. pipe (oil–air flow at 2% GVF).
3
LW2
2 LW1
Signal Amplitude (v)
-1
-2 Markup
-3
0 10 20
Fig. 10. Time window for reduction of signals acquired by the 0° sensor in the 2-in. pipe (oil–air flow at 2% GVF).
41
However, for the 0° sensor the random reflections at interfaces
in the multiphase flow also have to be taken into account. Even
though it is not possible to set up any precise time for the reflec-
tions to happen, they can only happen when the wave is traversing
the flow. This occurs during the wave ‘‘round trip” from the first
acrylic–oil interface (point B in Fig. 3) to the opposite acrylic–oil
interface (point H) and back to point B; this corresponds to the
20.5–96.5 ls time interval associated with general path A-B-?-B-
A in Table 7. Accordingly, the time window in Fig. 10 was extended
to encompass these possible random reflections as depicted by the
red rectangle. Ahead of this rectangle, one can see an electronic
markup of the emitted pulse (t = 0 ls) and LW2, the portion of
the longitudinal wave generated in A that is reflected back to A
upon hitting point B. As LW2 never traverses the flow, it was
excluded from the time window for data reduction.
Acoustic paths A-B-?-D-C and A-B-?-F-E generally represent
paths for the portions of the longitudinal wave LW1 that were scat-
tered to the side transducers by inhomogeneities in the flow.
Again, no specific paths or transit times can be associated with
these events as the inhomogeneities are randomly distributed over
the pipe cross section. The corresponding time windows for the 45°
and 135° sensors thus are the one shown in Fig. 10 for the random
reflections.
The appropriateness of these time window was verified by the
experimental observation of which regions of the entire acquisition
window were sensitive to GVF variations. Only minor adjustments
were necessary in the case of the side sensors as shown in Fig. 11.
Multiple reflections in the pipe wall, even if they occur during the
time windows discussed above, are attenuated very quickly and
are not expected to significantly affect the results and trends dis-
cussed in Sections 3.4 and 3.5.
SW2
30 40 50 60 70 80 90 100 110
Time (µs)
Random reflections
SW1
30 40 50 60 70 80 90 100 110
Time ( µ s)
42
1.0
Signal Amplitude (10 v)
-2
0.5
0.0
-0.5
-1.0
0 10
1.0
Signal Amplitude (10 v)
-2
0.5
0.0
-0.5
-1.0
0 10
Fig. 11. Response signals and time windows for the 2-in. pipe for oil–air flows at different GVF values: (a) 45° sensor; (b) 135° sensor.
A similar procedure was followed to establish the time win-
dows for the 1-in. pipe. It is noteworthy that the approximate
100-ls duration of the events in Figs. 9–11 is much shorter than
the 500-ls time period for pulse generation. Therefore, one can
be sure that the effects of one pulse completely die out before
another pulse is emitted. The time windows so established allowed
for a reduction in the sizes of raw data files as well as in the com-
puting time to reduce the data.
The energy carried by each acoustic pulse, as received by each
transducer, was calculated from the waveforms by integrating
the squared signal voltage over the signal duration. The energy
ratio for any transducer was defined as the energy carried by an
individual acoustic pulse divided by the energy carried by a refer-
ence signal, i.e. the maximum energy likely to be received by that
same transducer. For the 0° and 180° transducers, the maximum
energy was received when the pipe was filled only with the contin-
uous phase (single-phase mineral oil); the corresponding energy
ratios are then given by the following equation,
R t2
En; t
ðV multiphase Þ2 dt
¼ R t21
E1;
t1
ðV single-phase Þ2 dt
By normalizing the energy ratios with the single-phase signal,
the effect of intrinsic dissipation of the transmitted wave is taken
out of the data trends. This will help make the procedure devel-
oped herein easily applicable to other fluids as the continuous
phase; in addition, it will make it easier to train and validate the
neural networks to be discussed in Section 4 with data for different
oils.
For the 45° and 135° sensors, there is no energy reflected to the
sides when the pipe is filled with the continuous phase only. In
these cases, the denominator in the equation above was replaced
with the condition for maximum side reflection in two-phase
oil–air flows, which was experimentally verified to occur at
approximately 28% GVF. In this way, the energy ratios for the 45°
M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50
5% Random reflections
13%
33%
(a)
20 30 40 50 60 70 80 90 100 110
Time (µs)
5% Random reflections
13%
33%
(b)
20 30 40 50 60 70 80 90 100 110
Time ( µ s)
and 135° transducers also span the 0–1 range. The energy ratios
were calculated for all 15,000 individual pulses in each acoustic
sample and the average then taken for the sample. Furthermore,
as various samples were taken for each operating condition, it
was also possible to calculate the mean average energy ratio for
a given condition. This is the acoustic parameter of interest in ana-
lyzing the data in the 1-in. and 2-in. pipes, discussed next. For a
further discussion on the acoustic signals and calculation of the
energy ratios, the reader is referred to [32].
3.4. Results from acoustic data for the 1-in. pipe
Figs. 12–15 show plots of the mean average energy ratios, or
simply energy ratios, as a function of GVF for the various transduc-
ers for all types of multiphase flows in the 1-in. pipe. The horizon-
tal error bars represent the standard deviations of the average GVF
measurements using the quick-closing valve device. In this sense,
they represent the variations in the flow samples trapped in the
device, one flow sample for each acoustic sample. The vertical error
ð15Þ bars are the corresponding standard deviations of the mean aver-
age energy ratios and thus represent variations in the acoustic phe-
nomena. The flow patterns indicated in the plots correspond to
those discussed in Section 3.1. It can be observed that the scatter
in the GVF is very small for bubbly flow, increases slightly with
the presence of cap bubbles and Taylor bubbles (slug flow), and
can be very large as the flow progresses into the unstable slug flow
and churn flow; the scatter then decreases in the annular flow pat-
tern. This behavior reflects the repeatability in the flow samples in
the quick-closing valve device for each flow pattern.
In the 180° sensor (Fig. 12), the GVF is the main parameter con-
trolling the acoustic attenuation as the data points for all flow
types remain close to each other. It is thought that the near zero
transmission value for the oil–air interfaces as opposed to 99%
and 36% for the oil–water and oil–glass interfaces, respectively,
could account for this behavior. Reflections and scattering of the
 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50 43

1.1
bubbly
O-A
1.0 flow O-A-P
dispersed bubbles O-A-P-S
0.9
O-A-P-S-W
+
0.8
cap bubbles
0.7
slug flow unstable slug churn flow annular flow
Energy Ratio
0.6 flow

0.5

0.4

0.3

0.2

0.1

0.0
0 10 20 30 40 50 60 70 80 90 100
GVF (%)

Fig. 12. Energy ratios at the 180° sensor for multiphase flows in the 100 ID pipe.

1.1
O-A
1.0 O-A-P
O-A-P-S
0.9
O-A-P-S-W
0.8

0.7
Energy Ratio

0.6

0.5

0.4
dispersed bubbles
0.3 +

cap bubbles
0.2

0.1 bubbly slug flow unstable slug churn flow annular flow
flow flow
0.0
0 10 20 30 40 50 60 70 80 90 100
GVF (%)

Fig. 13. Energy ratios at the 0° sensor for multiphase flows in the 100 ID pipe.

ultrasonic beam of the ultrasonic beam by the air bubbles would
then be the dominant acoustic dissipation mechanisms. The
greater dispersion in the data for GVF < 15% would be due to the
smaller bubble population, which lessens the importance of reflec-
tions at the oil–air interfaces and thus allows the water and solid
dispersed phases to play a more prominent role in the acoustic dis-
sipation process.
The energy ratios at the 0°, 45°, 135° sensors (Figs. 13–15,
respectively) increased significantly over the GVF range tested. At
the 0° sensor this increases is mainly due to acoustic backward
reflections off the bubbles interface whereas for the 45° and 135°
sensors scattering of the acoustic beam accounts for the increase
in the energy ratios. Therefore, increasing the gas phase concentra-
tion increases the energy ratio in all these sensors. Conversely, for
the 180° transducer the energy ratios decreased exponentially with
GVF as scattering and backward reflections of the acoustic beam
continually decrease the amount of energy that hits this sensor.
It is noteworthy the data stratification at the 0° sensor – i.e.
systematically different values of the energy ratios for different
flow types for the same GVF –, especially for GVF < 50%. This indi-
cates an effect of the suspended particles and water droplets on
signal attenuation. For this sensor, the acoustic path can be twice
as long as that for the 180°, thus enhancing the action of these dis-
persed phases. In very general terms, it can also be stated that the
smaller the GVF the larger the data stratification according to the
flow type (number of phases). For bubbly flow up to the slug flow
regime, this stratification is more noticeable and decreases for the
remaining flow patterns.
In connection with the energy ratios error bars, it was previ-
ously verified that the errors introduced by the ultrasonic sensors
and associated electronics – measuring system errors [53] – were
not significant at all. In addition, as the ultrasound is non-
intrusive and the pulses are low energy, system-sensor interaction
errors and system disturbance errors are also negligible. The vari-
ations in the energy ratios are believed to be caused mostly by
stochastic variations in the multiphase flows themselves. They
44 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50

1.1

1.0

0.9

0.8

Energy Ratio 0.7

0.6

0.5

0.4
bubbly slug flow unstable slug churn flow annular flow
0.3 flow flow

0.2 O-A
dispersed bubbles
O-A-P
0.1 + O-A-P-S
cap bubbles O-A-P-S-W
0.0
0 10 20 30 40 50 60 70 80 90 100
GVF (%)

Fig. 14. Energy ratios at the 45° sensor for multiphase flows in the 100 ID pipe.

1.2
O-A
1.1
O-A-P
1.0 O-A-P-S
O-A-P-S-W
0.9

0.8
Energy Ratio

0.7

0.6

0.5

0.4
slug flow unstable slug churn annular
0.3 bubbly
flow flow flow
flow
0.2 dispersed bubbles

0.1 +

cap bubbles
0.0
0 10 20 30 40 50 60 70 80 90 100
GVF (%)

Fig. 15. Energy ratios at the 135° sensor for multiphase flows in the 100 ID pipe.

are generally small for the bubbly flow and annular flow. These
flow structures, because they are fairly stable and repetitive, do
not lead to large variations in the intensity of the acoustic reflec-
tion and scattering. On the other hand, intermittent flow struc-
tures, like cap bubbles and Taylor bubbles, do cause these
variations, which in turn can lead to very large standard deviations.
The present authors believe these data helped establish the
potential of the ultrasonic technique to measure the GVF for all
flow regimes, including those associated with high GVF. It is note-
worthy that the annular flow pattern, which presents difficulties
for measurements with gamma ray techniques as mentioned in
the introduction, lends itself to fairly easy measurements with
the ultrasound. As shown schematically in Fig. 8d, this flow pattern
has a very repeatable structure and the gas core leads to more
directional (less random) reflections of the ultrasonic beam. In
addition, the liquid film around the gas core also constitutes a
more homogeneous medium for the acoustic processes to take
place, which are thus more repeatable.
3.5. Results from acoustic data for the 2-in. pipe
Figs. 16–19 show the results for the 2-in. pipe. The scatter in the
GVF are small when there are only dispersed bubbles and cap bub-
bles present and increases with the appearance and growth of Tay-
lor bubbles. Once again, this behavior reflect the repeatability of
the flow samples trapped in the quick-closing-valve device; when
larger gas structures are present, repeatability decreases.
The plots for all types of multiphase flows at the 180° sensor
(Fig. 16) show that data stratification occurred only for oil–air
flows at GVF < 13%. For this sensor, acoustic energy is received
mostly by direct transmission through the liquid phase; hence,
for low GVF values there will be little obstruction of the cross-
 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50 45

1.1
O-A
1.0 O-A-W
O-A-W-S
0.9

0.8

0.7
Energy Ratio
0.6 slug flow

0.5

0.4

0.3 dispersed bubbles

0.2
+
0.1
cap bubbles
0.0
0 10 20 30 40
GVF (%)

Fig. 16. Energy ratios at the 180° sensor for multiphase flows in the 200 ID pipe.

1.1

1.0

0.9

0.8

0.7
Energy Ratio

0.6

0.5

0.4
dispersed bubbles
0.3
slug flow
0.2 +
O-A
0.1 cap bubbles O-A-W
O-A-W-S
0.0
0 10 20 30 40
GVF (%)

Fig. 17. Energy ratios at the 0° sensor for multiphase flows in the 200 ID pipe.

section and substantial amounts of acoustic energy will hit the
180° sensor. If the continuous phase contains suspended particles
and water droplets, their effect on acoustic dissipation will be
more significant. As the cross-section becomes increasingly clut-
tered with the gas phase and less energy hits the 180°, the changes
in signal amplitude will be less perceptible.
For the 0° sensor (Fig. 17), data stratification for the oil–air
flows occurred throughout the entire GVF range tested. Beam
reflection at the liquid–gas interfaces are just as important in this
case as the acoustic energy that propagates through the liquid
phase all the way to the opposite pipe wall and back to this sensor.
For low GVF values, direct transmission prevails and the effect of
solid particles and water droplets is similar to that observed for
the 180° sensor. The increased stratification for the 0° sensor could
be due to the longer acoustic path (twice as long), which enhances
the attenuation by the dispersed phases.
For increasing GVF values, reflection at the liquid–gas interfaces
play an increasingly more important role and the resultant acous-
tic paths get shorter. Therefore, the effect of the dispersed phases
in the oil would be less significant. However, for the slug flow this
is an intermittent phenomenon as it depends on the passage of
Taylor bubbles. During the passage of liquid slugs, direct transmis-
sion through the liquid phase is once again an important mecha-
nism for energy reception and so is effect of the dispersed
phases. One could then expect that data stratification for the slug
flow would remain at an approximately constant level. Had the
annular flow pattern been attained for the 2-in. pipe, data stratifi-
cation would probably have decreased; in fact, this is the behavior
exhibited in Fig. 13 for the 100 pipe.
For the 45° and 135° sensors (Figs. 18 and 19), energy reception
is totally dependent on scattering of the acoustic beam by the bub-
bles. The acoustic paths in the liquid phase are shorter and the
46 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50

1.1

1.0

0.9

0.8

Energy Ratio 0.7

0.6

0.5

0.4
dispersed bubbles
0.3
slug flow
0.2 +
O-A
0.1 O-A-W
cap bubbles
O-A-W-S
0.0
0 10 20 30 40
GVF (%)

Fig. 18. Energy ratios at the 45° sensor for multiphase flows in the 200 ID pipe.

1.1

1.0

0.9

0.8

0.7
Energy Ratio

0.6

0.5

0.4

0.3 dispersed bubbles

0.2
slug flow
+
O-A
0.1 O-A-W
cap bubbles
O-A-W-S
0.0
0 10 20 30 40
GVF (%)

Fig. 19. Energy ratios at the 135° sensor for multiphase flows in the 200 ID pipe.

amount of energy received by these sensors is less than that
received by the other sensors, as experimentally verified. There-
fore, one would not expect significant data stratification for these
sensors.
Finally, the scatter in the energy ratios is generally small, except
for the oil–air flows in the presence of cap bubbles. This is in agree-
ment with the fact cap bubbles appeared more randomly and can
significant alter the relative importance of the acoustic reflection
and scattering. In the next section, a check was made on the ability
of neural networks to take into account the different aspects in the
behavior of the multiphase flows discussed above and successfully
predict the flow pattern and the GVF.
4. Development and results from the artificial neural networks
In developing the artificial neural network (ANN) for
application to multiphase flows, one has to keep in mind that the
GVF continuously spans the 0–100% range while the flow patterns
are discrete concepts or schematic visualizations of the multiphase
flow topology. Hence, the GVF and the flow patterns constitute
very different types of outputs that are not easily handled by one
single ANN. Accordingly, one ANN was developed for flow pattern
identification and another one for GVF prediction. Both have the
acoustic energy ratios as input and use the same basic architecture
(Fig. 20). The ANNs developed were of the MLP type as these are
frequently used for pattern recognition tasks. In addition, as cited
in the introduction, in the comparison by Rosa et al. [40] the RBF
and PNN types did not exhibit significantly superior performances.
In selecting the flow patterns for recognition by the ANNs, a
careful study was conducted on the flow patterns proposed by dif-
ferent authors [42]. These flow patterns were then grouped under a
broader classification according to the patterns normally used in
daily practice in the oil industry. For vertical upward flows, the
three main categories were dispersed bubbles, intermittent, and
annular flows; this division also provided satisfactory results in
 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50
Fig. 20. The ANN architecture for both flow pattern detection and GVF determination.
flow pattern prediction having the dimensionless groups as input
[42]. In the present investigation, it was preliminarily verified that
better results were achieved when the following categories were
used: dispersed bubbles, intermittent flow, churn flow, and annu-
lar flow. The first category regroups dispersed bubbles and cap
bubbles as discussed in Section 3.1; the second category corre-
sponds to slug flow and unstable slug flow; the third and fourth
categories are simply churn flow and annular flow as generally
described in the literature.
The four energy ratios comprise the four ANN inputs, one for
each sensor (0°, 45°, 135°, and 180°); the input layer thus has four
neurons. Each entry corresponds to the average energy ratio over
the 7.5-s sampling interval for a given condition. First, the com-
plete data set discussed in Sections 3.4 and 3.5 was divided into
two subsets, 75% of the data points for training the ANN and 25%
for validation, corresponding to 347 and 116 experimental data
points, respectively. Care was taken to make the 25% validation
set representative of all the operating conditions, diameters, and
flow patterns tested in this investigation. As the multiphase flow
acoustic database does not constitute a continuous smooth func-
tion but rather is made of discrete points, the number of hidden
layers should be greater than unity [54]. The number of hidden lay-
ers was limited to two to keep the ANN feed-forward architecture
as simple as possible; the number of neurons was empirically
determined by testing several combinations of one to six neurons
in the first hidden layer with one to three neurons in the second
hidden layer. From the results for the summation of squared errors
(SSE), the best combination was verified to be five and two neurons
in the first and second hidden layers, respectively. This result was
valid for both ANNs, the one to determine the GVF and the one to
determine the flow pattern. The output layer has just one neuron,
as the output variable is either the GVF or the flow pattern. The
logarithmic sigmoid and the purely linear functions were used
for the hidden and output layers, respectively.
The ANNs were trained using the Bayesian regularization
Levenberg–Marquardt algorithm; this was considered one the
most reliable method among those available in MatLabÒ Neural
Network Toolbox because it avoids data over-fitting. Training
was accomplished in relatively short times and, due to the Baye-
sian term in the objective function, changes in weights and biases
were more conservative, thus avoiding training data over-fitting.
In connection with the SSE goal attainment criteria, considera-
tion was made of the typical uncertainty in mass flow rate mea-
surements using MFM. Falcone et al. [3] summarize uncertainty
ranges indicated by several authors for different scenarios; the
uncertainty in liquid mass flow rate using MFM is in the 10–20%
47
range. How these uncertainty values translate into GVF uncertainty
depends on the uncertainty in other parameters and instruments
that make up the complete multiphase metering procedure. In
the present investigation, the first guideline was that the SSE in
the ANN results should not exceed, in terms of GVF, the standard
deviations in the GFV measurements. Moreover, the uncertainty
in the ANN results should be compatible with flow pattern assess-
ments. For bubbly flow, which represents the most difficult case, it
was reasoned that 1% maximum variation in the GFV was still
within acceptable limits for distinguishing nuances in this flow
pattern. Accordingly, the SSE limit was set to 102 for GVF predic-
tion and to 101 for flow pattern identification for the entire 75%
training set. The training procedure would then stop when the
SSE results were below the limiting values or the number of epochs
reached the maximum, whichever came first. The maximum num-
ber of epochs was set to 103.
The flow patterns were given numbers for identification by the
ANN as shown in Table 8. As discussed by Inoue et al. [42], the
accuracy of the flow pattern detection was observed to vary
depending on the classification order of the flow patterns. The min-
imum error was empirically determined to occur when the several
flow patterns were ordered as they should occur for a fixed liquid
flow rate and increasing gas flow as shown on the flow pattern
map for vertical flows in [37]. A ±0.1 interval was adopted for each
flow pattern index so that, for instance, the vertical dispersed bub-
bles regime is centered on 1.1 and spans the [1.0,1.2] interval. (The
0–1.0 interval was reserved for horizontal flows.) By not fixing
exact discrete values for each flow pattern, but rather allowing
them to vary over prescribed intervals, the accuracy of the ANN
predictions increased.
Next, to determine the weights and bias, 100 training attempts
were made using a routine developed in MatLab. The weights and
bias were initialized with random numbers generated by the init
(net) command and, for each attempt, the SSE was calculated and
compared with the previous attempt. In case it was smaller, the
corresponding weights and bias were saved. The weights and bias
Table 8
Flow patterns indexes and corresponding ranges.
Flow pattern Flow pattern index Flow pattern range
Dispersed bubbles 1.1 1.0–1.2
Intermittent flow 1.3 >1.2–1.4
Churn flow 1.5 >1.4–1.6
Annular flow 1.7 >1.6–1.8
48 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50

values finally saved were those that yielded the smallest SSE value Table 9
among all 100 attempts using the validation data set. The ANN out- Evaluation of ANN results using unseen data points as input.

puts were compared with the GVF values measured by the quick- Flow pattern ANN flow pattern predictions ANN GVF
valves device and with the flow patterns from visual observations predictions
and high-speed filming. % Correct SSE ±Index SSE ±GVF
Figs. 21 and 22 show the ANN results for flow pattern identifi- variation variation
cation and GVF prediction, respectively, using unseen data points Dispersed bubbles 100.0 0.013 0.02 179.4 2.0
as input; also shown are the approximate flow pattern bounds Intermittent flow 97.6 0.068 0.04 496.8 3.6
based on the visual observations. Table 9 summarizes the results Churn flow 100.0 0.029 0.03 944.2 5.7
Annular flow 83.0 0.015 0.05 377.7 7.9
obtained. As can be seen, incorrect flow pattern predictions
occurred for the intermittent flow and annular flow regimes. One Overall result 98.3 0.125 0.03 1998.0 4.2

intermittent flow data point was incorrectly predicted as in the

dispersed bubbles flow pattern. This could be due to difficulties
in visually assessing the flow pattern as cap bubbles grow into Tay-
lor bubbles, i.e. setting practical limits for the bubbly flow/inter-
mittent flow transition. In addition, the standard deviations in
the energy ratios for these two regimes were large in some cases,
which could also have contributed to the incorrect prediction.
Regarding the incorrect predictions of annular flow data points
as churn flow, one should recall that the uncertainty in the GVF
measurements in these cases was rather large as shown in the
plots in Sections 3.4 and 3.5. The SSE results for the individual flow
patterns are smaller than the 101 goal and are not likely to have
contributed to the incorrect predictions. Anyway, the 98.3% overall
1,8
Annular flow
1,6
ANN Prediction
result for the correct flow pattern predictions is evidence of the
excellent ANN performance.
Regarding the GVF prediction, Fig. 22 shows that the predicted
values agree satisfactorily with the targets. The SSE results and the
corresponding GVF variations steadily increased as the flow pat-
tern progressed from dispersed bubbles to annular flow, which is
in the reverse order with the number of data points available to
training and validating the ANN. This could have influenced the
results. The 102 goal for the SSE and corresponding 1% maximum
GVF variation were not attained, but all the values in Table 9 are
still compatible with the 10–20% uncertainty range in liquid mass
flow rate measurements using MFM metering [3].
These results show that the ANNs could successfully handle the
effect of diameter on flow turbulence and bubbles coalescence,
which affected the flow patterns development (Section 3.1), and
the effect of flow composition (Sections 3.4 and 3.5), which caused
some stratification in the energy ratios. In other words, the ANNs
successfully predicted the flow pattern and the GVF even though
the only inputs were the acoustic energy ratios. However, due to

Churn flow
lack of data for GVF higher than 35% in the 2-in. pipe, the ability
of the ANNs to handle the effect of diameter could be checked only
1,4
up to slug flow. In a new large-scale test rig, ultrasonic data will be
obtained for all flow patterns, including annular flow, in a 2-in.
Intermittent flow
pipe; these new data will be presented in a future paper. In addi-
1,2 tion, the present data are restricted to small concentrations of
solids and water. In the oil industry, solids (sand) concentrations
Dispersed bubbles
are normally very small, but the water cut can increase to very high
values. Therefore, data should be taken in the future to check
1,0
whether the ANNs could still provide reliable flow pattern and
1,0 1,2 1,4 1,6 1,8
GVF results without making the water concentration one specific
Observed Flow Pattern input. Anyhow, the feasibility of using a combination of the ultra-
sonic technique and ANNs for flow pattern recognition and GVF
Fig. 21. ANN results for flow pattern identification in the 1-in. and 2-in. pipes.
measurements in multiphase flows has been confirmed by the pre-
sent results.
100 In order to bring this new MFM procedure closer to the field,
Dispersed acoustic data for real fluids in large-scale facilities in the oil indus-
bubbles try will be acquired in the near future. In addition, a combination of
80 the ANNs based on the ultrasonic data with that by Inoue et al. [42]
Intermittent flow
based on the dimensionless groups is currently under way. An
ANN Prediction

extension of the initial database by Inoue et al. [42] is also being

60 worked on so that the ANNs based on the dimensionless groups
can be thoroughly trained and validated.

40
5. Conclusions
Churn flow Annular flow
20
In this paper, acoustic attenuation data for vertical, upward oil-
continuous multiphase flows in various flows patterns have been
0 presented and analyzed. The flows were two-phase (oil–air) to
0 20 40 60 80 100 four-phase (oil–water–air–sand) and the patterns ranged from
Measured GVF bubbly to annular flows in the 1-in. pipe and from bubbly to slug
flows in the 2-in. pipe. The ultrasonic instrumentation consisted
Fig. 22. ANN results for GVF prediction in the 1-in. and 2-in. pipes. of four transducers in a simple arrangement intended to be used
 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50
in real-time in actual applications of the oil industry. For the 1-in.
pipe, plots of the energy ratio (normalized attenuation parameter)
as a function of GVF exhibited some degree of stratification accord-
ing to the number of phases present; this effect was most notice-
able for the 0° sensor. For the 2-in. pipe, data stratification was
significant only for the 0° sensor.
The acoustic attenuation measured by the four transducers
were used as input to artificial neural networks developed in this
investigation to identify the flow pattern and determine the GVF
of multiphase flows. Despite the stratification in the data for the
various flow types and two different pipe diameters, the ANNs
could successfully identify the flow patterns and determine the
GVF for all types of multiphase flows tested. Ultrasonic data for
the 2-in. pipe was limited to bubbly and slug flow; however, the
ability of the ANNs to handle the effect of diameter in annular
flows will be tested in a new facility and discussed as part of a
future paper.
The extension of the present work to real applications in the oil
industry is currently under way. In broad terms, it involves obtain-
ing ultrasonic data for real fluids and operating conditions of the oil
industry, including higher water fractions, and further training and
validating the ANNs with these new data. The combination of the
ANNs based on the ultrasonic data with that by Inoue et al. [42]
based on dimensionless groups is also planned on.
Acknowledgements
The authors gratefully acknowledge PETROBRAS for the finan-
cial support through Grant FUNCAMP 46088, and PRH/ANP,
FAPESP, and CNPq for the scholarships. The authors also kindly
acknowledge Dr. Valdir Estevam, from PETROBRAS, for his encour-
agement and support in the development of an ultrasonic multi-
phase flow meter.
References
[1] C.M. Atkinson, H.K. Kytömaa, Acoustic wave speed and attenuation in
suspensions, Int. J. Multiphase Flow 18 (1992) 577–592.
[2] M.A. Isakovich, On propagation of sound through emulsions (in Russian), J. Exp.
Theor. Phys. (in Russian, Pis’ma v Zh. Eksp. Teor. Fiz.) 18 (1948).
[3] G. Falcone, G.F. Hewitt, C. Alimonti, Multiphase Flow Metering, Principles and
Applications, First ed., vol. 54, Elsevier, 2010.
[4] J.L. Goncalves, T.A. Paiva, J.R. Abud Jr., R.D.M. Carvalho, O.J. Venturini,
Development of a Multiphase Flow Metering Procedure Based on the
Ultrasonic Technique, in: 15th International Conference on Multiphase
Production Technology (MPT 2011), Cannes, France, 2011.
[5] Y.A. Wahaba, R.A. Rahimb, M.H.F. Rahimanc, S.R. Awd, F.R.M. Yunusb, C.L.
Gohb, H.A. Rahimb, L.P. Lingb, Non-invasive process tomography in chemical
mixtures – a review, Sens. Actuat. B: Chem. 210 (2015) 602–617.
[6] L. Scheers, Multi-Phase Flow Metering: On its way from nursing to mature
technology, Hydrocarbon Production Accounting Workshop, 2008, Moscow,
Russia.
[7] H. Devold, Oil and Gas Production Handbook, An Introduction to Oil and Gas
Production, 1.3 ed, 2006, ABB ATPA Oil and Gas.
[8] R. Arora, Production well testing using non-radioactive multiphase flow
meters, Explor. Prod. 7 (2009) 86–89.
[9] C. Tan, H. Wu, F. Dong, Horizontal oil–water two-phase flow measurement
with information fusion of conductance ring sensor and cone meter, Flow
Meas. Instr. 34 (2013) 83–90.
[10] H.-M. Prasser, A. Böttger, J. Zschau, A new electrode-mesh tomograph for gas–
liquid flows, Flow Meas. Instr. 9 (1998) 111–119.
[11] M.J. Da Silva, Impedance Sensors for Fast Multiphase Flow Measurement and
Imaging, 2008, Electrical and Computing Engineering Department,
Technischen Universität Dresden, Dresden, Germany.
[12] C. Poette, P. Reynier, Evaluation of 3D mapping experimental non-intrusive
methods for multiphase flows, Int. J. Multiphys. 8 (2014) 69–90.
[13] R.D.M. Carvalho, A.C. Bannwart, The Use of the Ultrasonic Technique for the
Study and Online Monitoring of Multiphase Flows, in: L. Cheng, D. Mewes
(Eds.), Advances in Multiphase Flow and Heat Transfer, Bentham Science
Publishers, 2010.
[14] L.E. Kinsler, A.R. Frey, A.B. Coppens, J.V. Sanders, Fundamentals of Acoustics,
Fourth ed., John Wiley & Sons, 2000.
[15] D.T. Blackstock, Fundamentals of Physical Acoustics, John Wiley & Sons Inc,
2000.
49
[16] P.M. Morse, Ingard K.U., Theoretical Acoustics, Princeton University Press,
1986.
[17] H. Murakawa, H. Kikura, M. Aritomi, Application of ultrasonic doppler method
for bubbly flow measurement using two ultrasonic frequencies, Exp. Therm.
Fluid Sci. 29 (2005) 843–850.
[18] H. Murakawa, H. Kikura, M. Aritomi, Application of ultrasonic multi-wave
method for two-phase bubbly and slug flows, Flow Meas. Instr. 19 (2008) 205–
213.
[19] S. Longo, The effects of air bubbles on ultrasound velocity measurements, Exp.
Fluids 41 (2006) 593–602.
[20] B.M. Wrobel, R.W. Time, Ultrasonic measurement and characterization of a
low concentration system of solid particles in liquid, in high shear flow, Appl.
Acoust. 73 (2012) 117–131.
[21] X. Dong, C. Tan, Y. Yuan, F. Dong, Oil–water two-phase flow velocity
measurement with continuous wave ultrasound Doppler, Chem. Eng. Sci.
(2015).
[22] J. Claesson, A. Rasmuson, J. Wiklund, T. Wikström, Measurement and analysis
of flow of concentrated fiber suspensions through a 2-D sudden expansion
using UVP, J. Am. Inst. Chem. Eng. 59 (2013) 1012–1021.
[23] D. Harbottlen, M. Fairweather, S. Biggs, The minimum transport velocity of
colloidal silica suspensions, Chem. Eng. Sci. 66 (2011) 2309–2316.
[24] H. Kawai, Development of a mixing device and photosynthetic bioreactor with
a Taylor Vortex Flow (velocity measurement in solid–liquid flows with the
ultrasound velocity profiler – in Japanese, Jpn. J. Multiphase Flow 24 (2010)
275–281.
[25] S.W. Jones, A. Amblard, C. Favreau, Interaction of an ultrasonic wave with a
bubbly mixture, Exp. Fluids 4 (1986) 341–349.
[26] D. Chen, X. Wang, C. Che, J. Cong, D. Xu, X. Wang, A transient method for
measuring the gas volume fraction in a mixed gas–liquid flow using acoustic
resonance spectroscopy, Sci. China: Phys., Mech. Astronomy 53 (2010) 1412–
1418.
[27] L. Xing, C. Hua, H. Zhu, W. Drahm, Flow Measurement Model of Ultrasonic
Flowmeter for Gas–Liquid Two-Phase Stratified and Annular Flows, Advances
in Mechanical Engineering, Hindawi Publishing Corporation, 2014. Volume
2014.
[28] Y. Murai, S. Ohta, A. Shigetomi, Y. Tasaka, Y. Takeda, Development of an
ultrasonic void fraction profiler, Meas. Sci. Technol. 20 (2009) 114003–114013.
[29] Y. Murai, Y. Tasaka, Y. Nambu, Y. Takeda, S.R. Gonzalez A, Ultrasonic detection
of moving interfaces in gas liquid two-phase flow, Flow Meas. Instr. 21 (2010)
356–366.
[30] Y. Zheng, Q. Zhang, Simultaneous measurement of gas and solid holdups in
multiphase systems using ultrasonic technique, Chem. Eng. Sci. 59 (2004)
3505–3514.
[31] M. Vatanakul, Y. Zheng, M. Couturier, Application of ultrasonic technique in
multiphase flow, Ind. Eng. Chem. Res. 43 (2004) 5681–5691.
[32] R.D.M. Carvalho, O.J. Venturini, E.I. Tanahashi, F. Neves Jr., F.A. Franca,
Application of the ultrasonic technique and high-speed filming for the study
of the structure of air–water bubbly flows, Exp. Therm. Fluid Sci. 33 (2009)
1065–1086.
[33] E.I. Tanahashi, T.A. Paiva, F.A. Grangeiro, R.D.M.d. Carvalho, O.J. Venturini, A.C.
Bannwart, V.C. Bizotto, Application of the Ultrasonic Technique for
Monitoring Intermittent Liquid–Gas Flows and Liquid–Solid Flows, in: 7th
North American Conference on Multiphase Technology (MPNA 2010), 2010,
Banff, Canada.
[34] J.L. Goncalves, Development of an Ultrasonic Technique for the Measurement
of the Concentration of the Dispersed Phases in Multiphase Flows in the Oil
Industry (in Portuguese), 2013, Ph.D. in mechanical engineering, Department
of Mechanical Engineering, UNIFEI, Itajubá, MG – Brazil.
[35] L.-S. Zhai, N.-D. Jinn, Z.-K. Gao, Z.-Y. Wang, D.-M. Li, The ultrasonic
measurement of high water volume fraction in dispersed oil-in-water flows,
Chem. Eng. Sci. 94 (2013) 271–283.
[36] M. Meribout, N.Z. Al-Rawahi, A.M. Al-Naamany, A. Al-Bimani, K. Al-Busaidi, A.
Meribout, A multisensor intelligent device for real-time multiphase flow
metering in oil fields, IEEE Trans. Instr. Meas. 59 (2010) 1507–1519.
[37] E. Dahl (Ed.), Handbook Multiphase Flow Metering, 2005, Norwegian Society
for Oil and Gas Measurement (NFOGM) and The Norwegian Society of
Chartered Technical and Scientific Professionals (Tekna).
[38] N. Bhat, T.J. McAvoy, Use of neural nets for dynamic modeling and control of
chemical process systems, Comput. Chem. Eng. 14 (1990) 573–583.
[39] A. Shaikh, M. Aldahhan, Development of an artificial neural network
correlation for prediction of overall gas holdup in bubble column reactors,
Chem. Eng. Process. 42 (2003) 599–610.
[40] E.S. Rosa, R.M. Salgado, T. Ohishi, N. Mastelari, Performance comparison of
artificial neural networks and expert systems applied to flow pattern
identification in vertical ascendant gas–liquid flows, Int. J. Multiphase Flow
(2010).
[41] E. Filletti, P. Seleghim Jr., Nonintrusive measurement of interfacial area and
volumetric fraction in dispersed two-phase flows using a neural network to
process acoustic signals—a numerical investigation, Int. J. Num. Meth. in
Biomed. Eng. 26 (2010) 234–251.
[42] E.H. Inoue, R.D.M. Carvalho, V. Estevam, A.C. Bannwart, A.M.F. Fileti,
Development of a Neural Network for the Identification of Multiphase Flow
Patterns, in: 14th International Conference on Intelligent Systems and Control
(ISC 2013), 2013, Marina del Rey, CA, USA.
[43] Olympus-Panametrics, Ultrasonic Transducers Technical Notes, Olympus
Corporation, 2011.
50 M.M.F. Figueiredo et al. / Experimental Thermal and Fluid Science 70 (2016) 29–50
[44] R.J. Kalivoda, Understanding the limits of ultrasonics for crude oil
measurement, Oil and Gas Engineer – Production/Processing, 2013.
[45] R. Nishi, The scattering and absorption of sound waves by a gas bubble in a
viscous liquid, Acustica 33 (1975) 65–74.
[46] V.K. Varadan, V.V. Varadon, Y. Ma, A propagator model for scattering of
acoustic waves by bubbles in water, J. Acoust. Soc. Am. 78 (1985).
[47] Acoustics and Electroacoustics for Emulsions, 2015, <http://www.
dispersion.com/acoustics-and-electroacoustics-for-emulsions>, (cited 2015).
[48] P.S. Epstein, R.R. Carhart, The absorption of sound in suspensions and
emulsions, J. Acoust. Soc. Am. 25 (1953) 553–565.
[49] J.R. Allegra, S.A. Hawley, Attenuation of sound in suspensions and emulsions:
theory and experiments, J. Acoust. Soc. Am. 51 (1972) 1545–1564.
[50] H.K. Kytömaa, Theory of sound propagation in suspensions: a guide to particle
size and concentration characterization, Powder Technol. 82 (1995) 115–121.
[51] G.Y. Onoda, E. Liniger, Random loose packings of uniform spheres and the
dilatancy onset, Phys. Rev. Lett. 64 (1990) 2727–2730.
[52] N. Herrmann, P. Boltenhagen, P. Lemarechal, Experimental study of sound
attenuation in quasi-monodisperse emulsions, J. de Physique II (6) (1996)
1389–1403.
[53] R.J. Moffat, Describing the uncertainties in experimental results, Exp. Therm.
Fluid Sci. 1 (1988) 3–17.
[54] S. Haykin, Neural Networks: A Comprehensive Foundation, second ed.,
Prentice Hall, 1998.