sensors

Review
A Survey on Distributed Fibre Optic Sensor Data Modelling
Techniques and Machine Learning Algorithms for Multiphase
Fluid Flow Estimation
Hasan Asy’ari Arief 1, * , Tomasz Wiktorski 2 and Peter James Thomas 1

1 NORCE Norwegian Research Centre AS, 5008 Bergen, Norway; [email protected]

2 Department of Electrical Engineering and Computer Science, University of Stavanger,
4036 Stavanger, Norway; [email protected]
* Correspondence: [email protected]

Abstract: Real-time monitoring of multiphase fluid flows with distributed fibre optic sensing has
the potential to play a major role in industrial flow measurement applications. One such application
is the optimization of hydrocarbon production to maximize short-term income, and prolong the
operational lifetime of production wells and the reservoir. While the measurement technology itself
is well understood and developed, a key remaining challenge is the establishment of robust data
analysis tools that are capable of providing real-time conversion of enormous data quantities into
actionable process indicators. This paper provides a comprehensive technical review of the data



analysis techniques for distributed fibre optic technologies, with a particular focus on characterizing


fluid flow in pipes. The review encompasses classical methods, such as the speed of sound estimation
Citation: Arief, H.A.; Wiktorski, T.; and Joule-Thomson coefficient, as well as their data-driven machine learning counterparts, such as
Thomas, P.J. A Survey on Distributed
Convolutional Neural Network (CNN), Support Vector Machine (SVM), and Ensemble Kalman Filter
Fibre Optic Sensor Data Modelling
(EnKF) algorithms. The study aims to help end-users establish reliable, robust, and accurate solutions
Techniques and Machine Learning
that can be deployed in a timely and effective way, and pave the wave for future developments in
Algorithms for Multiphase Fluid
the field.
Flow Estimation. Sensors 2021, 21,
2801. https://doi.org/10.3390/
s21082801
Keywords: multiphase fluid flow; machine learning; speed of sound; distributed acoustic sensor;
distributed temperature sensor
Academic Editors: Nikolai Ushakov,
Leonid B. Liokumovich and Arthur H.
Hartog
1. Introduction
Received: 17 March 2021 Increasing field recovery is essential in the oil and gas industry. Equinor, a Norwegian
Accepted: 12 April 2021
state-owned energy company, estimates an untapped potential of around four billion
Published: 15 April 2021
barrels of oil from a 10% increase of oil recovery on the Norwegian Continental Shelf (NCS)
alone [1]. Assuming the average oil price in the first quarter of 2021 around 61 USD per
Publisher’s Note: MDPI stays neutral
barrel [2], this equates to an economical value of 24.4 billion USD per percent of increased
with regard to jurisdictional claims in
recovery; an enormous revenue increase from one region alone. In addition, there are
published maps and institutional affil-
also environmental benefits for such recovery enhancements, including: reducing carbon
iations.
footprints due to production, and reducing the need for new oil field developments [3].
Production optimization is defined as the maximization of short and long-term pro-
duction of oil, while at the same time minimizing production costs [4]. A typical starting
point for production optimization is through continuous monitoring of the downhole pro-
Copyright: © 2021 by the authors. duction well variables (e.g., Water in Liquid Ratio (WLR), Gas Volume Fraction (GVF), fluid
Licensee MDPI, Basel, Switzerland.
flow rate, water or gas breakthrough, and sand production). Typically, these measurements
This article is an open access article
are then combined with simulations in order to optimize production control parameters
distributed under the terms and
(e.g., Inflow Control Valve (ICV) and/or Inflow Control Device (ICD) parameters, pressure
conditions of the Creative Commons
setting, and controlling water/gas injection) for stimulating production [5]. It is a rigorous
Attribution (CC BY) license (https://
process with a continuous loop of monitoring, simulating, and optimizing the production
creativecommons.org/licenses/by/
of oil, while preventing and addressing anomalies and production disturbances. Therefore,
4.0/).

Sensors 2021, 21, 2801. https://doi.org/10.3390/s21082801 https://www.mdpi.com/journal/sensors

Sensors 2021, 21, 2801 2 of 29

it is crucial to have robust, reliable, and accurate monitoring capabilities to achieve the
most optimized oil production system.
Some of the most powerful approaches for optimizing oil production employ mul-
tiphase fluid flow monitoring [6–9]. Such techniques are used to estimate and monitor
the phase-fraction of oil, water, and gas inside the production wells. Metering devices,
such as optical flow meter [10], Electrical Impedance Tomography (EIT) [11], differential
pressure meters [12], and gamma densitometer [13] are used to estimate the multiphase
fluid flow. It should be noted that some of the flow meter devices [14,15] use radioactive
sources, thus, they require careful supervision during installation, operation, and disposal,
not to mention a permit and experts for handling the radioactive materials [7]. Other
non-radioactive devices such as the optical flow meters, however, are expensive and only
provide a point-based measurement. For long-range pipeline and well applications, this
significantly limits the monitoring capabilities of the overall system.
An inexpensive, non-invasive meter solution, called the Virtual Flow Meter (VFM),
can also be used for multiphase flow estimation. VFMs combine pressure, temperature, and
other process control data from an existing sensors with multiphase fluid models, in order
to estimate the fluid flow rates. However, the VFMs require extensive parameter tuning
and active sensor calibration [16–18]. Parameter tuning in VFMs becomes complicated due
to the complexity and ever-changing variables in the downhole environment. As a result,
VFMs have not experienced widespread adoption as a multiphase flow meter solution [7].
Distributed fibre optic sensors are receiving an increasing amount of attention. The
sensors consist of a passive optical fibre and an interrogator unit located at one end of
the fibre, depicted in Figure 1. Such systems are able to measure parameters at every
position where the fibre optic cable is installed. They have been developed to measure
temperature [19], strain [20], acoustic [21], and even chemical [22] parameters, such as
humidity [23] and the presence of oil [24] in a distributed manner. The sensors operate
by using the interrogator to send pulses of laser light along with the fibre. As the light
pulses propagate, they interact with the glass fibre in several different ways, leading to the
generation of three distinct types of scattered light (propagating back to the interrogator),
known as Rayleigh, Raman and Brillouin scatter. Each backscatter type can be distinguished
by its frequency content and is influenced by different physical parameters, forming
the basis of different Distributed Sensors (DSs), such as Distributed Temperature Sensor
(DTS) [19] and Distributed Acoustic Sensor (DAS) [21] .

Figure 1. The Distributed Acoustic Sensor carries light pulse travelling inside the fibre cable that are
backscattered to the Interrogation Unit that recovers the acoustic signal profile along the cable.

In 1996, DTS technology began to find its use in real-world applications [25]. Subse-
quently in 2009, a DAS system was used to successfully measure an increase in oil produc-
tion following a fracking intervention [26]. Since then, the DSs have formed the basis of a
great deal of research, both in the development and understanding of the technologies [27],
as well as their practical application. Due to their long-range and distributed monitoring
capabilities, the distributed fibre sensors have been deployed for Vertical Seismic Profiling
(VSP) [28], hydraulic fracture detection [29], early-warning systems for earthquake and
Sensors 2021, 21, 2801 3 of 29

seismic activity monitoring [30], traffic pattern analysis and monitoring [31], and in-well
flow profiling that not only used in the laboratory or field trials but also in real-time
operations [6,32–34].
In general, applied DS systems require advanced data processing algorithms to es-
timate physical quantities of certain parameters within their surroundings. For the mul-
tiphase fluid flow measurement, these algorithms are often needed to be carried out in
real time. This challenge is particularly great for the DAS systems where the raw data
quantities can often exceed 1 GB/s. Speed of sound estimation from acoustic data [9] and
Joule-Thomson coefficient from temperature data [35] are the state-of-the-art techniques
that are used to estimate the multiphase fluid flow from DS data. In recent years, the ability
to perform similar tasks using data-driven machine learning algorithms have been explored.
For example, Convolutional Neural Network (CNN) and Artificial Neural Network (ANN)
have been used to perform flow regime classification and multiphase estimation [36,37].
Subsequently, the ANN with Long Short-Term Memory (LSTM) algorithm has also been
used to perform fluid flow rate estimation [38–41].
In this paper, we provide a structured and comprehensive review of the recent ad-
vances in the multiphase fluid flow estimation based on the distributed fibre optic sen-
sor technologies. In Section 2 we begin by reviewing the challenges and importance of
multiphase fluid flow and discuss existing technologies and devices that perform such
measurement. In Section 3, we review the distributed sensor technologies and their work-
ing mechanism. In Section 4 we present the state-of-the-art in regards to data modelling
algorithms that characterize multiphase fluid flows by combining DS measurements with
physical flow modelling. Section 5 presents a similar state-of-the-art overview for machine
learning-based algorithm for characterizing multiphase fluid flows. Finally, Section 6
discusses the performance and limitations of the state-of-the-art, in addition to potential
future research directions.
To the best of our knowledge, this is the first paper that presents and analyzes the
state-of-the-art in multiphase fluid flow characterisation with an emphasis on distributed
fibre optic sensor and machine learning algorithms. There are several excellent reviews
for multiphase flow estimation using other technologies, such as VFMs [17,18,42] and
hardware-based flow meters [7,43].

2. Multiphase Fluid Flow

Multiphase fluid flows are defined as the simultaneous flow of more than one fluid
phase through channels, such as porous media or pipelines [44]. Multiphase fluid flows
in the context of the oil and gas industry, are the combinations of water, oil, and gas.
Knowledge of the state of the multiphase flow can be used to optimize the production wells
by, for example, changing the ICV or ICD settings in a way that maintains oil uplift and
avoids problems such as water or gas coning, sand production, sharp pressure, or flow rate
drops [9]. Moreover, with real-time knowledge over extended lengths of the flow column,
anomalies can be detected as soon as they arise irrespective of their location, allowing
for a targeted solution that can be implemented instantly without shutting down the
whole production system. In this way, real time information about the state of multiphase
flows provides a transformative capability for the reservoir and production engineers for
reducing operational disturbances and maximizing oil production [9,10].
However, providing an accurate, reliable, and cost-effective measurement of multi-
phase fluid flow is a huge challenge in the oil and gas industry. In addition to the flow
variables (such as flow pattern, phase density, fraction, and velocity), other factors are influ-
encing the multiphase measurements, for example, (1) operating pressure and temperature,
(2) pipe properties such as diameter, shape, inclination, surface roughness, and (3) the
presence of other components such as valves, T-junctions, and bends [7,9,10]. The challenge
becomes greater for the downhole in-well measurement, due to the High Pressure and High
Temperature (HPHT) environments. Other complicating variables include the formation
Sensors 2021, 21, 2801 4 of 29

of emulsion and hydrate phases, as well as the presence of sand that flow carries from
the formation.
The multiphase flow in a pipeline is characterised by the flow regime and the cor-
responding flow rate. The multiphase flow can be represented as the fraction of fluid
phases flowing simultaneously at a given time and location. The flow rate, on the other
hand, represents the volume of fluid flowing per unit time at a given location. They can be
estimated using the speed of sound propagating in the fluid, flow velocity, pressure and
temperature in a given flow regime, and fluid density.
Several different approaches can be used to measure the multiphase flow, including
but not limited to, conventional dedicated hardware-based flow meters [7,10,14,15,43],
virtual flow meters [17,18,39,42,45–47], and distributed sensor flow estimations [8,37,40,48].
This paper focuses on providing a comprehensive review of the last approach, using
distributed sensors with physical flow modelling and machine learning algorithms for
multiphase flow estimation. A brief summary of the other two approaches is provided
in the following subsections. For completeness, it must be stated that phase separation
(followed by single-phase fluid flow measurement, depicted in Figure 2) can also be
employed for characterising the multiphase flow [49]. While this type of multiphase fluid
characterisation is accurate, it is extremely resource and time demanding. Therefore there
has been a lot of R&D effort towards developing multiphase flow meters.

Figure 2. Schematic of a typical test separator within an oil production system. The multiphase flow
is denoted as M while the single phase flow is denoted as S. Reprinted from ref. [7].

2.1. Hardware-Based Flow Meter

Conventional hardware-based multiphase flow meters (MPFMs), can provide accurate
measurements of density, velocity, and phase fraction of fluid flow in the pipes. Electrical
capacitance tomography [50], electrical impedance tomography [51], and optical tomogra-
phy [52], are commonly encountered technologies that are used to build MPFM sensors.
These technologies acquire a series of simultaneous images from a cross-section of the
pipes, and cross-correlate these to determine the values of flow velocity [7].
Sensors 2021, 21, 2801 5 of 29

Similarly, gamma densitometry is another technological advancement that can be used

to build tomographic MPFM sensors. Such devices rely on the different ray attenuation
properties of the different phases for performing the multiphase measurement [7]. While
the gamma-ray instruments can provide very accurate measurements [53], their reliance on
radioactive sources presents additional challenges with respect to installation, operation,
and disposal.
Optical flow meters represent another family of MPFM sensors. They use the flow
velocity and speed of sound of the flowing fluid to estimate the phase-fraction of the
fluids. The flow velocity is obtained by tracking the vertical structures in the flow, and
the speed of sound is obtained from the acoustic waves of the turbulent flow [10]. The
optical flow meters are commercially mature devices and can provide reliable point-based
measurements [9].

2.2. Virtual Flow Meter (VFM)

VFMs work by combining numerical models with available field data (e.g., pressure
and temperature from both the bottomhole and wellhead of well chokes, as mentioned
in [42]) to estimate the multiphase flow. Based on its modelling paradigms, the VFMs
can be divided into first principles VFM and data-driven VFM. Flow modelling in near-
well region, pipelines, and production chokes, together with pressure and temperature
measurements are utilized by the first principles VFMs to estimate fluid flow rates [54].
The steady-state optimization algorithms, such as choke model, can be used to provide a
point-based fluid flow estimation [42]. On the other hand, the dynamic state optimization
algorithms, such as Kalman Filter, can be used to provide dynamic modelling of the first
principle VFMs [55].
Data-driven VFMs, however, work by using machine learning algorithms to estimate
the multiphase flow. The general framework is to fit the field data and mathematical models
(lead by machine learning algorithms),without basing them on exact understanding of the
physical parameters and information of the production wells and the reservoir. Several
machine learning algorithms, including Support Vector Machine (SVM), LSTM, ANN, and
gradient boosting algorithms [39,56–59], can be used to perform the multiphase estimation.
Combining both first principles and machine learning algorithms can help to improve
accuracy as well as the transparency of the VFM-based approaches, providing insight
into the physical origins of the results [60]. For an in-depth review of the first principles
and data-driven VFMs, we direct the reader to an excellent review by Bikmukhametov
et al. [42].

3. Distributed Sensor Technologies

Physically, a distributed fibre optic sensor is a passive cable that’s capable of spatially
resolved measurements, similar to that achievable with a densely spaced array of point
sensors [61]. As a surveillance and monitoring infrastructure, it provides a lot of benefits,
for example, (1) it is less difficult to install compared to installing a large number of sensors,
(2) it works in passive mode reducing the power source utilization and is easy to maintain
since most of the system’s complex elements can be made easily accessible, and (3) it can
work in harsh HPHT environment with sufficient coating materials.

3.1. Distributed Sensor Working Mechanism

The main component of Distributed Sensors (DSs), in addition to the fibre cable, is the
Interrogation Unit (IU) or Interrogator. It sends pulses of laser light through a fibre optic
cable and then analyses the properties of the elastic and inelastic backscattered light [62] as
depicted in Figure 1. These measurements are characterised by the frequency, phase, and
amplitude properties of the backscattered light [6], that are in turn related to temperature,
pressure, vibration, and strain changes within the cable and therefore it’s surroundings. The
distributed sensors work based on three distinct scattering processes that take place as laser
Sensors 2021, 21, 2801 6 of 29

pulses interact with the glass fibre: Rayleigh backscattering [63], Raman backscattering [64],
and Brillouin backscattering [65], depicted in Figure 3.

Figure 3. The backscattering phenomenons used for distributed fibre optic sensing. Reprinted from
ref. [22].

Environmental conditions, such as temperature and strain changes, directly affect the
backscattered signals. The DTS exploits these phenomenons by measuring the intensity
of anti-Stokes Raman scattering signal, as a function of the local temperature changes
within the fibre cable. The DAS, however, operates differently. The Rayleigh backscat-
tering used by DAS is modulated in intensity and phase as a function of acoustic/strain
perturbations [66]. Since the fibre strain is also influenced by the temperature, Rayleigh
backscattering can also be used as a second form of DTS. Unlike Raman-based DTS, the
Rayleigh DTS uses the same fibre type as that for DAS, meaning that a single fibre can
be used to perform DTS and DAS together [67]. For a more detailed explanation of these
backscattering phenomenons and fibre-based distributed sensors, we suggest excellent
distributed sensor reviews by Gohari et al. [6], Lu et al. [22], and Schenato [66].
It is important to highlight that for measurements at any given location, the DAS data
have a low Signal to Noise Ratio (SNR) relative to their array-point sensor counterparts.
One of the strengths of DAS, over the point-based sensors, is their ability to provide data
continuously along the cable. This feature, together with the ability to make measurements
with high temporal frequency, indicates large data rates. This data rate is ultimately only
limited by the time taken for the backscatter from each pulse to leave the fibre, before
the next pulse can be introduced. This in turn highlights the importance of powerful and
efficient data processing algorithms. One common method of reducing the influence of
noise is through an algorithm known as F-K filtering.
The design of the DAS cable itself can have a significant effect on the system per-
formance (as can the cable installation itself) that should be done to optimise the signal
coupling. Helically wound fibre cable is an example of how cable design can influence the
sensing properties [68]. Firstly, the helical trajectory through the cable ensures the cable has
“broadside” sensitivity to disturbances perpendicular to the cable. It is more sensitive com-
pared to the straight cables which mostly being influenced by axial disturbances. Secondly,
the fibres are wound about a compliant material that serves as a sensitivity amplifier [68].
In addition, engineered-fibres are becoming the subject of much research in recent years,
where the motivation is to create fibres with enhanced Rayleigh scattering, therefore in-
creasing the signal [69]. Such engineered-fibres are particularly useful for increasing the
range of the DAS systems, that are typically limited to a few 10 km by attenuation of the
light by the glass. It is worth noting that the DSs are not meant to be the replacement of
conventional point flow meter devices, but they work as complimentary equipment to
provide distributed measurements in a cost-effective way.

3.2. Applications for Distributed Sensors

The early applications for DSs technologies were in DTS, which were used to measure
temperature and pressure changes in a field trial in West Coalinga Field, CA, USA, in
1996 [25]. Since then, significant improvements have been made both in technologies and
Sensors 2021, 21, 2801 7 of 29

hardware systems. The DTSs have been used for transformer monitoring [70], wildfire
behavior characterization [71], leakage detection [72], structure monitoring [73], fire de-
tection [74], and cooling effect and temperature log in oil and gas industry [75] as was
mentioned in [76].
DAS systems, began to emerge in the late 2000s, have been used for early-warning
system for earthquake and seismic activity monitoring [30], hydraulic fracture detec-
tion [29], traffic pattern analysis and monitoring [31], gas leak detection [77], pipeline
surveillance [78], Vertical Seismic Profiling (VSP) [28], Steam Assisted Gravity Drainage
(SADG) monitoring [79], and in-well flow profiling that not only used in the laboratory
or field trials but also in the real time operations [6,8,32,33]. Combining DTS and DAS
data has also been explored, for example, to address the three-phase flow estimation of
oil, water, and gas for the downhole well simulations [8], which was less accurate and
seems unsolvable when only using DAS or DTS alone. Figure 4 shows a sketch of smart
wells with several ICVs and fibre-based distributed sensors as a straight-line and helically
wound cable around the pipe.

Figure 4. A sketch of a smart well with ICVs and fibre cables mounted around the flow pipe.

A single DAS IU can generate 20 Terabyte (TB) of data per day [80]. In an experiment
by Ajo-Franklin et al. [80], 128 TB data was generated by sampling 12,000 channels at a
frequency of 500 Hz over a period of 3 months. It means that as a permanent monitoring
device for an in-well downhole operation, a single fibre-based distributed sensor can
generate more than several Petabytes (PB) of data during its lifetime. Therefore, advanced
data management strategies, data compression algorithms, feature extraction techniques,
including comprehensive signal processing algorithms are required to process the DS data
to provide a realtime monitoring solution over a long period of time.

4. Physical Flow Modelling

Physical flow modelling can be defined as using physical phenomena to extract
valuable information from a given data. In the case of multiphase fluid flow estimation,
the physical flow modelling is used to approximate the changes of physical phenomena to
the value and phase-fraction of the multiphase fluids. Temperature changes, flow velocity
changes, and speed of sound changes, are some of the physical phenomena that fluctuate
when multiphase (or even single-phase) fluids flow. Distributed fibre optics can be be used
to measure these values, with consideration of the complexity of data processing required
being dependent on the DS type and the parameter of interest. For example the processing
workflow for measuring flow temperature with DTS [75] is relatively simple compared to
measuring speed of sound and flow velocity using DAS data [8,9].
Sensors 2021, 21, 2801 8 of 29

The common framework of using physical flow modelling for estimating the multi-
phase flow from distributed sensor is as follows:
1. gather a block of data as measured intervals, corresponding to a specific range of time
and location, see Figure 5,
2. extract the physical parameter values, including speed of sound and flow velocity
(more can be found in Table 1), then
3. estimate the multiphase values using extracted physical flow features and data from
fluid mixture databases. Examples of publicly available database can be found
in [81–84].

Table 1. Comparison of physical flow modelling techniques used with distributed fibre optic sensors
for flow rate estimation and multiphase classification

Modelling Technique Year Data Sources Note

Doppler effect [85] 2012 Field surveys Early work on DAS for well and reservoir
with DAS monitoring.
Doppler effect and Root 2014 Field trial with Early implementation of DAS on real oil field.
Mean Square (RMS) of DAS
acoustic energy [48]
SoS and eddy velocity 2015 Flow-loop Ability to visualize the generation and convection of
estimation [86] experiment with eddies using waterfall plot of distance versus time.
DAS
SoS and J-T coefficient 2016 Production Integral image algorithm for estimating SoS of
value matching [8,34] oilfield with DAS multiphase fluids and Ability to accurately measure
and synthetic DTS two-phase flows.
Forward model [87] 2017 Simulated DAS Simulating DAS data taking into account formation
wellbore properties, flow characteristic, noise
processes and optical fibre parameters.
Thermal-and-hydraullic 2018 DAS and DTS Thorough analysis on combining DAS and DTS data
modelling [88] for identifying gas flow.
Multiphysics analysis 2019 Flow-loop Applied on steam flow profiling experiment with
and clustering experiment high resolution DTS and DAS data.
optimization [89]
Statistical analysis and 2019 Flow-loop Designing and commissioning an advanced
SAGD modelling [90] experiment and multi-phase flow injection experiment.
simulation model
SoS analysis [91] 2019 DAS and DTS Applicable for HPHT horizontal gas producer.

4.1. Data Acquisition

The DS data must be processed partially due to the size. As was mentioned in [34],
a one-minute of DAS data often could not be loaded to regular desktop PC hardware.
Therefore, the data should be divided into series of blocks for processing; the size of the
blocks will represent the spatial and temporal resolution of the flow profile. It should
be pointed out that spatial overlapping of consecutive blocks is often employed during
processing in order to increase the spatial resolution and enhance the repeatability [48].
After dividing the data into several blocks, it is often required to transform each blocks
of DAS data from time-space domain, corresponding to the time of data being recorded
and the spatial locations within the fibre cable, to the frequency and wavenumber domain,
also called F-K domain. The two dimensional Fast Fourier Transforms (FFT) algorithm [40]
can be used to perform such transformation (F ( f , k )) as depicted in Figure 6. It is defined
in Equation (1) where t, x denote the time and location of the input data, while f , k denote
the frequency and wavenumber, respectively. Please note that the complex physical values,
such as speed of sound and flow velocity can be extracted from the F-K domain.
Z Z
F( f , k) = f (t, x )e−i(kx−2π f t) dx dt. (1)
Sensors 2021, 21, 2801 9 of 29

Figure 5. Raw acoustic data collected from a multilateral well. Reprinted from ref. [34].

Figure 6. F-K plot generated using 2D-FFT with V-shape line fitting. Reprinted from ref [34].

4.2. Physical Flow Data Extraction

4.2.1. Speed of Sound
Sounds travel at different speed depending on density, pressure, temperature, and
molecular structures within the travelling medium. For multiphase fluids, sound travels
faster within the water phases than in the oil and gas phase components [34]. The Speed
of Sound (SoS) measures how fast sounds travel within a medium. In fact, the SoS has
been used as a strong feature to estimate the type of medium where the sound travels. The
method is called phase estimation [9,48]. Figure 7 shows a possible range of mixture fluid
percentages given the SoS within those fluids.
Sensors 2021, 21, 2801 10 of 29

Figure 7. The relation between speed of sound and fraction of fluid mixture. Adapted from ref. [5].

The SoS can be measured by applying a line fitting algorithm [40,92] from an F-K
plot of the DAS data, depicted in Figure 8. It is based on the slopes of the lines in the F-K
domain, where the frequency and the high Fourier coefficients form the speed of sound [34].
It is formulated in Equation (2) where cm denotes the multiphase SoS, while λ, f , and k
denote wavelength, frequency, and wavenumber respectively. It will be recalled that cm
from the slope lines are consisted of the upgoing (cu ) and downgoing (cd ) SoS from the
propagating acoustic wave. The positive slope of the line represents the cu , while the
negative one represents the cd .
2π f
cm = λ f = . (2)
k

Figure 8. Comparison of the speed of sound between a clear line image (left) and not a clear one
(right). Reprinted from ref. [34].
Sensors 2021, 21, 2801 11 of 29

4.2.2. Flow Velocity

The flow velocity is defined as the speed of travelling fluids inside the pipe; the
flow rate is calculated from the flow velocity by including the flow pipe diameter in the
calculation. The flow velocity is derived from the Doppler Effect (or Doppler Shift) principle
of cu and cd [5]. It is important to highlight that when a flow approaches a sensor (at a given
location), the sound waves that reach the sensor have a shorter wavelength and a higher
frequency. However, when the flow moves away from the sensor, the sound waves that
reach the sensor have a longer wavelength and lower frequency. This phenomenon is called
the Doppler Effect. It measures the changes in an apparent frequency of a wave when the
flow (of the acoustic source) moves relative to a stationary sensor location. The flow velocity
can be calculated using this phenomenon. It is defined in Equation (3) where v is the flow
velocity and cd is assumed to be negative, see [5] for a more comprehensive derivation.

cu + cd
v= . (3)
2

4.2.3. Joule-Thomson Effect

The Joule-Thomson effect is characterized by temperature changes that happens when
fluids flow through a valve or porous plug with no heat exchange in the environment [34].
Wang’s work [93] has been focused on analyzing Joule-Thomson Coefficient (C JT ) using
DTS data. The C JT in the fluid mixtures is defined as a function of well temperature and
pressure, compressibility factor, and fluid mass-weighted phase fractions, more detail can
be found in [34]. Figure 9 shows the C JT as a function of phase fraction and can be used to
estimate accurate two-phase flow from water, oil, and gas.

Figure 9. The relation between Joule-Thomson coefficient and fraction of fluid mixture. Adapted
from ref. [34].

4.3. Multiphase Estimation

Determination of the SoS, flow velocity, and Joule-Thomson coefficient represent
the state-of-the-art with regards to fluid flow characterisation using DAS and DTS data.
These data, are then calibrated using information from fluid mixture databases to calculate
multiphase flow parameters such as WLR and GVF. The National Institute of Standards and
Sensors 2021, 21, 2801 12 of 29

Technology (NIST) in the US has provided a comprehensive database of thermophysical

and geophysical properties of hydrocarbon mixtures [82,84].
The physical values, when accurately estimated, can provide high accuracy multi-
phase flow information as depicted in Figure 10. However, prediction uncertainties are
expected due to factors such as volatilities of the surrounding physical environment, sensor
noise, systematic errors in the measurement method, corruptions within the data, and
other problems during value extraction process that might arise [94]. Thus, including
error estimations and uncertainty values, when providing the multiphase information for
realtime monitoring solution is often a requirement.

Figure 10. Example of two-phase flow characterisation using DAS data that has been used to calculate
the oil fraction. The figure also shows the location of the ICVs.

A summary of physical flow modelling techniques that have been used with dis-
tributed fibre optic data for flow rate and multiphase estimation is provided in Table 1.
Johannessen et al. [85] provided an early work on using DAS data to extract qualitative
information on the flow regime, speed of sound and an estimate for flow velocity in some
part of the wells. Even though quantitative analysis in [85] was limited, the work showed
an interesting qualitative analysis that tied together well acoustic signatures and well
behaviour. Xiao et al. [48], also presented DAS data obtained using cables installed within
producing wells. The work presented several data analysis technique for modelling and
enhancing DAS performance, including RMS of acoustic energy, amplitude estimation,
FFT transformation, and SoS with flow analysis. As the number of the spatial channel of
the recorded DAS data increased, the SoS calculation can provide higher accuracy and
precision, influencing the overall accuracy of estimating fluid flow rates. Finfer et al. [86]
provided experimental results from single- and multiphase tests for assessing DAS suitabil-
ity for monitoring fluid velocity and flow composition. In [86], F-K transformations were
employed for measuring flow velocity from DAS data, The paper also proposed the use of
multiphase multipoint flow sensing and provided practical guidance on how to set up a
DAS system for real-time multiphase measurement. Fidaner et al. [87], on the other hand,
developed a forward model to connect between two-phase flow in the wellbore and DAS
data using a set of analytical expressions, such as physical fluid mechanism, propagation
of the acoustic signal, and phase changes in optical signals due to pressure change. The
Sensors 2021, 21, 2801 13 of 29

wavelet analysis method was used to capture the most relevant components of DAS data
for multiphase flow rate estimation. These components were then trained using ANN to
obtain a more realistic flow rate estimation model.
The work of Abukhamsin et al. [8,34] covers many measurement aspects discussed
in this review. In particular Abukhamsin et al discussed the use of DAS, DTS, and the
combination of the two on addressing the challenges of characterizing three-phase flows.
Even though, the DTS was derived from commercial thermal simulator [34], the work
shows a promising result on combining SoS from DAS and C JT from simulated DTS to
provide an accurate multiphase estimation. Hemink et al. [88], on the other hand, showed
that C JT from actual DTS does not always provide reliable results as a straightforward
thermal model to identify gas-injection. Instead, the work proposed an improvement by
considering the temperature response measured by DTS where the fibre is clamped and
bending away from the tubing. Using the DTS trace, the identification of the annular-fluid
interfaces (brine/gas, gas/flowing gas) was possible, as well as the depths of active lifting
points. Shirdel et al. [89] employed several signal processing algorithms, including DAS
spectrogram, DTS and DAS waterfall analysis, and steady-state injection, to interpret
the DTS and DAS data to provide a quantitative step-injection-flow profiling. It shows
that those algorithms are tied together with an independent physical principle related to
multiphase flow, acoustic effects, data array, and others. The work can be used to analyze
complex flow regimes and heat transfer of wet-steam flow in horizontal wells providing a
good basis for benchmarking multiphase estimation algorithms.
Another flow-loop experiment and simulation model was presented by
Soroush et al. [90]. The work focused on analysing the potential of fibre optics tech-
nology to perform inferential multi-phase flow measurement. The results showed that the
flow regime and existence of gas-phases could be determined by DAS-based on the signal
frequency content. These results are crucial for SAGD wellbore monitoring on detecting
steam breakthrough. Another work is from Cerrahoglu et al. [91] on identifying cluster flow
from DAS and DTS on horizontal dry gas-producing wells from the HPHT environments.
They showed that using SoS analysis from the cable bottom section, nearly 50% of the total
gas rate comes from below the cable. A result that might be skewed when performing a
full spatial channel analysis based on SoS calculation.

5. Machine Learning
Advances in the machine learning field in the past few years have generated a lot of
interest for potential applications within the oil and gas industry, especially in the realm of
production monitoring and automatic surveillance. For example, several machine learning-
based techniques have been used for multiphase flow and flow rate estimation, including
feed-forward Neural Network (NN) [58], Recurrent Neural Network (RNN) [39,41], Sup-
port Vector Machine (SVM) [57], gradient boosting algorithm with regression trees [95],
and Kalman Filter (KF) [41]. An example schematic of modelling the DAS data using
CNN algorithm can be seen in Figure 11. Most of those works, however, used the machine
learning algorithms on the VFM domain, while only a handful of research (including the
work of Jalilian et al. [96], Silkina [36], and Vahabi et al. [37,40]) have been focused on using
machine learning on the DS data for flow rate and phase estimation.
Sensors 2021, 21, 2801 14 of 29

Figure 11. An example schematic for modelling DAS data using a preprocessed spectrogram plot on
CNN-based models. Reprinted from ref. [97].

Even though the actual implementations may vary, the common pipeline that is used
for multiphase estimation based on the machine learning techniques can be simplified
as has been depicted in Figure 12. It starts with the data acquisition and preprocess-
ing, and concludes with an inference process which can include prediction, smoothing,
and extrapolation.
Sensors 2021, 21, 2801 15 of 29

Figure 12. The common workflow of machine learning-based techniques for multiphase fluid
flow characterisation.

5.1. Data Preprocessing

According to Forbes, more than 60% of a data scientist’s time is utilized on the data
understanding and preprocessing step [98]. It is the main backbone within the overall
machine learning workflow. The data preprocessing includes gathering, cleansing, slicing,
and transforming the input data to be forwarded and processed into the next step [37,99].
With the large size of the DS data, offline preprocessing sometimes is required to simplify
the learning process and speed up the overall implementation, similar to that demonstrated
in Vahabi et al.’s work [37].
Several techniques can be used to speed up the data preprocessing step. For example,
GPU-based implementation can be used for preprocessing the DAS data [100]. Generative
Adversarial Network (GAN) can be used to simplify the data generation process which
can help reduce the total processing time [101]. Moreover, reducing the data transforma-
tion procedures and only use simple bandpass filtering can also help speed up the data
preprocessing step, similar to Shi et al.’s work [99].

5.2. Feature Engineering

In machine learning, features are defined as the measurement values that can be
obtained from the object of interest. Acoustic amplitude/gain, temperature, time, and
measurement locations, are some of the features that can directly be obtained from the DSs
data. Depending on the objective of the machine learning task, the features can be used
to classify an object or to predict the next possible changes within the particular object
of interest.
Feature engineering, on the other hand, refers to the techniques used to transform the
existing features to a new domain where the new features are generated. These new features
can be used to enrich the ability of a machine learning model to achieve a better outcome.
Based on the way new features are generated, they can be divided into handcrafted and
non-handcrafted feature engineering techniques [102].
The handcrafted feature engineering techniques derive new properties using various
algorithms or physical formulations based on the understanding of the physical phe-
nomenons captured in the input data. The FFT transformation [37], bandpass filtering [99],
F-K filtering [103], and physical flow feature transformations in Section 4.2 are considered
handcrafted feature engineering techniques. On the other hand, the non-handcrafted
techniques generate new properties without understanding the physical phenomenons
surrounding the input data. These techniques perform cross-correlation, feature combina-
tion and multiplication, and high-dimensional transformation of the input data. The aim
is to provide new representations of the input data that are useful for achieving accurate
Sensors 2021, 21, 2801 16 of 29

prediction. Multi Layer Perceptron (MLP) [104], Binary Descriptor [105], and multi-stage
CNNs [37] are considered as the non-handcrafted engineering techniques that can be used
for multiphase estimation. Dimensional reduction algorithm, such as Principle Compo-
nent Analysis (PCA), can be used to reduce the number of features while selecting highly
relevant features for multiphase estimation objective [57,106].
The wavelet components from acoustic data, temperature changes and thermal lo-
cation from DTS data, low-frequency acoustic signal, spectrogram plot, F-K plot, as well
as the mean and variance from a time window of DAS and DTS data, have been used as
the main features for modelling distributed fibre optic data with machine learning. The
spectrogram plot from acoustic data [97], for example, can provide rich interpretations of
different classification schema, depicted in Figure 13. F-K plots, on the other hand, are the
representation of the SoS values, which have a strong correlation with the phase-fraction
information of multiphase fluid. Providing a clear V-shape sign from F-K plots, however,
is not a trivial process. It requires a longer range of spatial channels, as well as sufficient
acoustic fidelity within each channel [48].

Figure 13. The behaviours of time-dependant frequencies captured in spectrogram representation

from different classification schema. Reprinted from ref. [97].

5.3. Learning Algorithms

Machine learning algorithms are used to recognize pattern of an object given the input
features (handcrafted or otherwise). Support Vector Machine [57,59], Kalman Filter families
(including the Extended Kalman Filter (EKF) and Ensemble Kalman Filter (EnKF) [41,107]),
and Neural Network families [57,106,108], are some of the machine learning algorithms
that can be used for multiphase flow characterisation.
The SVMs use kernel functions to transform the input features to the higher dimension,
therefore, the data can be linearly separated in the new dimension. The SVM kernel
Sensors 2021, 21, 2801 17 of 29

functions work as the non-linear feature transformation to allow the SVM algorithm to
handle non-linear systems that are often the case in the petroleum industry [42]. The EKF
and EnKF, on the other hand, work by tracking the dynamic model of the data using
the state-space estimation based on variance-covariance matrices in time. The original
Kalman Filter was developed for the linear system, while these extension algorithms avoid
the linearization by estimating the covariance matrix instead of using the true matrix,
called ensembles. Both EKF and EnKF are used due to their robustness to noise and data
corruption, fast implementation, and their dynamic non-linear estimation can provide
accurate results. The study by Loh et al. [41] shows that the EnKF updated model can
provide a more accurate prediction compared to the ones without EnKF, allowing possible
application for a realtime monitoring solution.
Lastly, the NN families use a stack of weighted linear structures (called layers) with
intermediate non-linear functions to perform automatic classification/prediction. There
are several operations that can be used in the NN-models, including but not limited to
(a) pooling, (b) unpooling, (c) convolution, and (d) transposed convolution operation,
depicted in Figure 14. The NN structures consist of thousands of parameters that are
optimized by using the gradient descent algorithm and backpropagation parameter update
operation. ANN, CNN, and RNN are some types of NN algorithms that can be used
for multiphase estimation. For example, Vahabi et al. [37] used CNN to perform phase
classification using the F-K plot from the DAS data as input, providing a high accuracy
classification (99.3% accuracy on test data). It should be noted that the CNN can model
different type of data representation, e.g., F-K and spectrogram plots, depicted in Figures 6
and 11, respectively.

Figure 14. Illustration of (a) pooling, (b) unpooling, (c) convolution, and (d) transposed convolution
operations in NN models. Reprinted from ref. [109].

It is worth mentioning that the NN families and SVM algorithms are considered
black-box approaches. It means that a model generated by these algorithms is hard to
interpret, and often the results give limited insight into the underlying physical processes.
Therefore, several works [42,60] have tried to combine the physical features and first-
principle methods with the black-box algorithms to ensure the reliability of the predictions,
and facilitate the building of trust among stakeholders within the industry.
A summary of machine learning algorithms that have been used with distributed fibre
optic data for flow rate and multiphase estimation is provided in Table 2. Silkina [36] used
ANN to correctly identify the flow conditions of multiphase fluids, providing almost 100%
accuracy. A simple two-layer MLP was used to classify 11 different classes combining water
and pine oil with different air-flow rates. However, the random split between training
and test data on a sequentially generated dataset indicates a potential information leakage,
undermining the overall accuracy performances. Park et al. [110] considered the total
Sensors 2021, 21, 2801 18 of 29

spectral power of the signal within a bounded range from DAS as a regression model along
with the measured flow rates. The model was trained using a robust regression algorithm
to reduce the effect of corrupted data and outliers. Even though the presented results are
limited in term of accuracies and performances, they addressed important issues within
modelling corrupted data and addressing outliers within the acoustic data.

Table 2. Comparison of machine learning algorithms on using distributed fibre optic sensors for flow
rate estimation and multiphase classification.

Modelling Technique Year Data Sources Note

ANN [36] 2014 Flow loop Early report and experiment for using DAS data and
experiment ANN for flow regime classification and flow rate
with DAS estimation.
ANN [87] 2017 Simulated DAS The wavelet coefficients are the input and flow
pattern are the output.
Robust regression and 2018 DAS Frequency Band Extracted (FBE) bands analysis is
band switching used to improve the prediction accuracy.
algorithm [110]
MLP [104] 2018 Gas producing Mainly focus on using DTS for forecasting gas
well with DAS production while DAS data was only recorded during
and DTS hydraulic fracturing of the well.
Decision Tree, Adaptive 2019 Real field DAS Training was conducted under limited amount of
Boosting, and Random data.
Forest (RF) [111]
ANN, SVM, and RF [112] 2019 Gas production A well defined data-driven machine learning
well with DAS experiment, including the use of sensitivity analysis
and DTS for analyzing feature importance.
ANN [96] 2019 DAS Autoencoder ANN is used for modelling acoustic and
flow rate data.
CNN, ANN [37] 2019 Real well Resulting on high accuracy flow regime classification
underwater from F-K images of DAS data.
DAS
Cross-correlation, 2020 Real well Providing fast flow velocity estimation from a large
K-means, and Radial underwater volume of DAS data.
integration [40] DAS

Ghahfarokhi et al. [104] used an averaged daily data from 1320 DTS measurements
along the lateral of the gas-producing well in the Marcellus Shale, in Northern West Virginia
to forecast daily gas production. An MLP model was trained and deployed, and Sensitivity
Analysis (SA) was conducted to analyse weight behaviour. Similar to [104], Bhattacharya
et al. [112] used DAS and DTS among other datasets to predict daily gas production using
ANN, SVM, and RF. A high accuracy (96%) was achieved by employing 18 features to RF
model for prediction. However, the utilization of DAS and DTS in the project was limited,
since they were presented as spatially averaged point measurements before being fed into
to the main classifier. Therefore, the results did not fully benefit from the the distributed
and real-time nature of the sensors.
Another interesting work was presented by Vahabi et al. [37,40]. The DAS data, col-
lected from real oil, water and gas well pipes under the sea, was used to identify fluid
types [37] and to estimate fluid flow velocities [40] using machine learning algorithms.
The F-K transformation technique was employed to provide input data for the CNN and
ANN models to classify the type of fluid in pipes. The highest accuracy of 99.3% can be
achieved by CNN, which indicates a potential for further classifying multiphase fluids
using DAS data under a real production environment. Other machine learning algorithms,
such as Cross-correlation, K-Means, and Radial Integration, were also employed to deter-
mine fluid flow velocity in pipes. Flow velocity from the wellhead was used as the true
label, and with some physical assumptions, the machine learning algorithms performed
Sensors 2021, 21, 2801 19 of 29

quite well on estimating the flow velocity from input data derived according to the F-K
transformation [40].

5.4. Inference and Uncertainty Estimation

The inference process generates predictions as the final outcome and (often) is coupled
with prediction uncertainty estimation. For a realtime solution, inferring results from a
machine learning model not only requires the model to generate accurate prediction, but
also to deliver the results in timely manner. The CPU implementation of ANN model
can perform automatic flow regime classifications within 0.02 s, while the CNN model
can provide them with 99.3% accuracy within 0.01 s [37]. It must be mentioned that for
a realtime monitoring solution, the time for data acquisition and preprocessing must be
included in the overall prediction times. And those, however, are still a bottleneck for the
CNN implementation, as mentioned in [37].
In addition to the inference time, the uncertainty estimation is also an important factor
for delivering monitoring solutions based on the machine learning algorithm. This is
partly due to the stochastic nature of machine models when making the prediction and the
black-box property of some machine learning algorithms. Therefore, having the prediction
coupled with uncertainty estimation can increase the confidence and reliability of the
prediction from a machine learning model.
Several techniques can be used to measure the prediction uncertainties. These tech-
niques include Bootstrap algorithm, Bayesian statistic, and Dropout technique for NN-
based models [113]. These techniques use input data, posterior information, parameter
values, and the like, as the control variables to measure the changes in the final predictions.
The more varied the predictions given the changes in the control variables, the larger its
prediction uncertainty. On the other hand, the more uniform the predictions given those
changes, the smaller the prediction uncertainty of the aforementioned model.

6. Discussion and Comparison

Table 1 compares several techniques that have been used for modelling distributed
fibre optic data for production monitoring in the petroleum industry. It is important to
highlight that most of these methods were tested on confidential datasets, from real produc-
tion fields and flow-loop experiments. Some of them can provide high accuracy predictions
on their predefined objectives, either for flow rate estimation or single/multiphase classi-
fication. It should be noted that the Doppler Effect technique has been used extensively
as the state-of-the-art method for flow rate estimation, while the ANN-based machine
learning algorithms have been used for flow regime and multiphase classification.
In this section, we will discuss how some of these methods perform for realtime
monitoring in the real well environments, as well as providing future research directions for
multiphase flow estimation. We will first summarize the main differences between physical
flow modelling and machine learning algorithms. We then discuss the performance and
limitations of these methods, and finally lay out potential future research directions for
data-driven machine learning algorithms on distributed fibre optic sensors.

6.1. Physical Flow Modelling and Machine Learning Algorithms

One of the main differences between physical flow models and machine learning
approaches is within the learning process. The physical models use the hydrocarbon
mixture database to estimate the multiphase flow, while the machine learning techniques
use the so-called learning algorithm to extract patterns from the data. It is important to note
that the hydrocarbon mixture database is a robust, accurate, and well-proven reference
data that can provide high accuracy multiphase information. In the real field operations,
especially in the HPHT environments, relying solely on this database limits the ability of the
physical flow modelling algorithm to provide reliable distributed monitoring solution. This
is due to the limitations in the modelling process itself, that must often make assumptions
concerning system complexity such ignoring the influence of irregular surfaces inside the
Sensors 2021, 21, 2801 20 of 29

pipes, the nature of surrounding HPHT environments, and complex fluid behaviours such
as the formation of hydrate and emulsion phases.
Data-driven machine learning algorithms, on the other hand, work by analyzing
patterns in the data and can model (theoretically) any complex system accurately, even
with limited to none information of the underlying system (see universal function ap-
proximations [114]). Therefore, those techniques are capable to approximate and model
any well-understood phenomenons, as well as address the unknowns. This capability is
advantageous for fluid flow estimation, especially in the field of distributed fibre optic
sensors, where the data and their references/labels are abundant. For example, with an
adequate number of layers and enough data points, the NN-based algorithms can provide
a very accurate prediction with high certainty. In some cases [115], they can perform better
then human predictions. It is known that training a large structure of NN model with
large volume of data can be extremely time consuming and resource intensive [116], not to
mention the limitation of those methods to provide clear explanation of their results.
Due to their black-box nature, the NN-algorithms mentioned in Table 2 mostly ignore
the temporal correlations among data points. It has been known that temporal correlation
is an important characteristic for modelling sequential and time-series data. Several state-
of-the-art algorithms for sequence modelling, including Convolutional LSTM [117] and
Attention model [118], have provided significant improvement in terms of accuracy, by
considering the temporal dependencies within the input data. In addition, a graphical
model, such as Conditional Random Field (CRF) principle [119], can also be used to
cross-correlate the spatiotemporal structure of the distributed sensors, tying together the
spatiotemporal relationship among each spatial channel and their neighbouring channels
in the spatial domain as well as in the time domain.
Another contrast between physical flow models and machine learning algorithms is
the way features are generated. The physical models use physical formulations and first
principle methods to generate representative features, while the majority of the machine
learning algorithms (in Table 2) use NN-based feature generators. Physical formulations are
fast and reliable methods for understanding and explaining the fluid dynamics, hydrocar-
bon mixture phenomenons, and thermophysical events. NN-based feature generators, on
the other hand, have the ability to learn from the data without being boxed by rigid-known
formulations. They can learn and address the unknowns and (eventually) provide robust
and accurate estimations [120].

6.2. Challenges
As can be seen in Table 2, there has been limited research focusing on estimating multi-
phase flow by combining distributed fibre optic sensor and machine learning technologies.
This is in part, due to the complexity of the physical system but perhaps more significantly,
a lack of access to relevant annotated distributed fibre optic datasets. The NN-based ma-
chine learning algorithms are data intensive techniques that work well when sufficient data
are available for training. Unfortunately, to the best of our knowledge, only limited amount
of available DAS or DTS datasets that can be used for developing multiphase fluid flow
characterisation techniques. Thus, collaborations with oil and gas companies as well as
research institutions are a necessity for advancing this field further. For many applications,
easy access to annotated datasets is a trend that is helping to accelerate machine learning
research, for example, ImageNet for image classification [121], Pascal VOC2012 for image
segmentation [122], ISPRS Vaihingen dataset for remote sensing [123], and KITTI dataset
for autonomous vehicle applications [124], have become the backbone of many advances
within those research areas.
Another challenge for multiphase flow characterisation using distributed fibre optic
sensors is providing fast and near real-time classification. This is mostly due to the large
volume of DAS/DTS data that are being generated and used for processing. As was
mentioned in [37], 40 TB of DAS data were generated during 24 h of measurements, this
is an equivalent to around 28 GB of data per minute that must be processed to provide a
Sensors 2021, 21, 2801 21 of 29

real-time monitoring capability. Resource intensive data processing algorithms and high
performance infrastructures are required in order to model such large data within the
machine learning environment.
In addition to the dataset access problem and real-time processing issue, the black-box
nature of the NN-based algorithms also limits the progress and adaptation of the machine
learning techniques for multiphase fluid flow characterisation. Reservoir engineers and
stakeholders within the field have some reservations with the use of black-box algorithms
for estimating the multiphase fluid flow compared to the first principle-based methods and
physical flow modelling techniques [42].

6.3. Relevant Work from Other Industries

Estimating the phase fraction and flow velocity using acoustic sensors, Doppler Effect,
and machine learning is not only applicable for multiphase flow estimation within the
oil and gas industry. The medical industry, for example, has used similar technology to
detect anomalies and blood flow rate inside the veins using techniques called Biomedical
Photoacoustic Imaging (BPI) [125] and Venous Doppler Ultrasound (VDU) [126]. The BPI,
for example, uses sound wave formation and optical absorption in biological tissues to
form a biomedical image modality that can be used to measure hemoglobin concentration
and oxygen saturation. Combining BPI and machine learning have also been explored, for
example, to remove the photoacoustic reflection artefacts [127], to measure the prediction
uncertainty [128], or to reconstruct the photoacoustic faster using ANN [129].
The transportation industry, on the other hand, has also been using DAS and machine
learning for analysing traffic flow and detecting objects within the flow [31,130]. The
phase-fluid components and fluid flow rates can be thought as analogues to traffic object
movements and traffic flow respectively. Blood property concentration and blood flow rate
are similar analogues from the medical industry. By mimicking the process and reconstruct-
ing the technologies and advances from other industries, we can further accelerate the
research for multiphase fluid flow estimation within the oil and gas and process industries.
Table 3 shows several machine learning algorithms that are used for modelling distributed
sensor data from different fields and industries.

6.4. Future Research Directions

Despite the existence of distributed temperature measurements and thermal models
for more than 20 years, as pointed out in [34], DTS is limited in terms of characterizing
three-phase fluids. DAS on the other hand, with the ability to measure rapidly varying
dynamic physical properties, offers a richer variety of possibilities. As a result, the current
state-of-the-art for distributed fibre optic multiphase estimators is becoming increasingly
weighted towards DAS, and this is trend is likely to continue into the future.
Tables 2 and 3 encompass the rapid development of modern NN-based algorithms on
modelling distributed fibre optic data. Getting access to data relevant for applying such
techniques to multiphase fluid flow characterisation is key to accelerating this field where
there is clear potential for significant advances. Such advances could be effectively done
through interdisciplinary collaborations between industry and academia, with natural
mechanisms for monitoring the progress and benchmarking the quality, accuracy, and
processing time of new data modelling techniques.
In order to provide high accuracy multiphase fluid flow characterisation, modelling
the spatiotemporal aspects of distributed systems should also be considered as an interest-
ing research direction. The fibre optic sensor as a monitoring system is a spatiotemporal
structure consisting of large volume of temporal data with dense spatial resolution. The
Attention Model with its positional encoder combined with the RNN architecture for
its temporal encoder could potentially help understand those spatiotemporal structure
resulting on a higher accuracy classifier.
Sensors 2021, 21, 2801 22 of 29

Table 3. Recent work concerning machine learning algorithms for modelling distributed fibre optic
sensor data for non-multiphase fluid flow objectives

Algorithms Year Objectives Note

Gaussian Mixture Model 2019 Pipeline in- The contextual information at the feature level
(GMM) and Hidden Markov tegrity threat was incorporated in a Gaussian Mixture Model
Model (HMM) [131] detection and Hidden Markov Model (GMM-HMM)-based
pattern classification system for acoustic trace de-
cision strategy.
k-Nearest Neighbor (kNN) 2019 Event identifica- The disturbance events, such as knocking, press-
and SVM [132] tion ing, watering, climbing, and false disturbance
event, are identified for 25.05 km long OTDR sys-
tem using combination of kNN and SVM.
HMM [133] 2019 Pipeline safety The HMMs were trained to identify sequential
monitoring state process of events and extract the temporal
information of the data, and provided an average
accuracy of 98.2%.
Dual Path Network [134] 2019 Railway safety The proposal provides proof-of-concept on us-
monitoring ing distributed sensor and machine learning al-
gorithm for actual railway safety monitoring. The
F1-scores for all classes reached up to 97% in the
test data.
CNN [135] 2019 Microseismic The synthetic microseismic events injected into
event detection recorded ambient noise and was trained using
CNN to detect seismic events in the test DAS data.
NN [29] 2019 Fracture-hit de- The NN was trained on Low-frequency dis-
tection tributed acoustic sensing (LFDAS) to detect frac-
ture hits to monitor wells during hydraulic frac-
turing operations.
DNN [136] 2019 Human move- The DAS signal was enhanced using ultrafast
ment identifica- laser; the data was trained using supervised and
tion unsupervised machine learning algorithms to de-
tect human movement and pipeline monitoring.
SVM [137] 2020 Train tracking The vibrations of moving objects are used to iden-
tify and track trains in real-time; the algorithm
runs on GPU to speed up the calculations.
CNN, LSTM, K-means [138] 2020 Human locomo- High spatial resolution and bandwidth data was
tion identifica- shown to be effective on increasing the machine
tion learning accuracy.
LSTM [139] 2020 Railway intru- A real field experiment with noise background
sion detection sound was conducted in this study, resulting on
shortening the average detection response time to
8.25 s.
Random Matrix Theory 2020 Event activity de- Events were detected along with their location
(RMT) [140] tection on the fibre, then they were extracted from the
random noise using Spiked RMT models.
CNN [141] 2020 Earthquake de- The CNN shows a promising results for providing
tection a reliable earthquake detection despite low signal-
to-noise ratio of the fibre telecom infrastructure.

The machine learning algorithms, especially the NN-based algorithms, have the abil-
ity to extract useful features from vast amounts of data while providing high accuracy
predictions, and thus can simplify the data modelling process. Despite their potential
effectiveness for certain applications, their black-box nature can present challenges with
regard to gaining stake-holder trust and confidence. This challenge is less for more conven-
tional approaches relying on physical flow models that are based on scientifically proven
phenomena through experiment. The combination between the two will greatly advance
this field on providing high accuracy prediction with explainable outcome, for example,
(1) using the physical flow components as input features for the machine learning algo-
rithm, (2) using the physical flow models for self-calibrating machine learning predictions,
or (3) using the machine learning algorithms to validate the first principle methods.
Sensors 2021, 21, 2801 23 of 29

7. Summary
Accurate and real-time multiphase fluid flow characterisation techniques employing
distributed measurement capabilities will provide a game-changing functionality for pro-
duction optimization in the oil and gas industry. The state-of-the-art in terms of physical
flow modelling techniques and machine learning algorithms has been presented and dis-
cussed in this paper. An extensive review and comparative summary of the structure of
the state-of-the-art has been provided. The characteristics, performance, and trade-offs
between different algorithms were discussed. A comprehensive analysis of the potential
of machine learning algorithms for modelling the fibre optic sensor data for multiphase
estimation has been included. Finally, potential future research directions for multiphase
fluid characterisation using distributed fibre optic sensors and machine learning algorithms
were discussed.

Author Contributions: Conceptualization, H.A.A., T.W. and P.J.T.; methodology, H.A.A., T.W. and
P.J.T.; investigation, H.A.A.; writing—original draft preparation, H.A.A., T.W. and P.J.T.; writing—
review and editing, H.A.A. and P.J.T.; supervision, T.W. and P.J.T. All authors have read and agreed
to the published version of the manuscript.
Funding: This work was funded by the Research Council of Norway’s (RCN) Petromaks2 programme
(Grant number 308840) and industry partners, namely Equinor and Lundin.
Institutional Review Board Statement: Not applicable.
Informed Consent Statement: Not applicable.
Data Availability Statement: Not applicable.
Acknowledgments: The authors would like to thank Bjørnar Ystad and Yngve Heggelund for valu-
able discussions and suggestions. We gratefully acknowledge all the anonymous reviewers and
their constructive notes and reviews, through which the manuscript was enriched and improved.
The computations and model reproductions were performed on resources provided by UNINETT
Sigma2—the National Infrastructure for High Performance Computing and Data Storage in Norway.
Conflicts of Interest: The authors declare no conflict of interest.

Abbreviations
The following abbreviations are used in this manuscript:

ANN Artificial Neural Network

BPI Biomedical Photoacoustic Imaging
CNN Convolutional Neural Network
CRF Conditional Random Field
DAS Distributed Acoustic Sensor
DS Distributed Sensor
DTS Distributed Temperature Sensor
EIT Electrical Impedance Tomography
EKF Extended Kalman Filter
EnKF Ensemble Kalman Filter
F-K Frequency and Wavenumber
FBE Frequency Band Extracted
FFT Fast Fourier Transform
GAN Generative Adversarial Network
GMM Gaussian Mixture Model
GPU Graphical Processing Unit
GVF Gas Volume Fraction
HMM Hidden Mixture Model
HPHT High Pressure High Temperature
ICD Inflow Control Device
ICV Inflow Control Valve
ISPRS International Society for Photogrammetry and Remote Sensing
Sensors 2021, 21, 2801 24 of 29

IU Interrogation Unit
J-T Joule Thomson
KF Kalman Filter
KITTI Karlsruhe Institute of Technology and Toyota Technological Institute
kNN k-Nearest Neighbor
LFDAS Low-Frequency Distributed Acoustic Sensor
LSTM Long Short-Term Memory
MLP Multi Layer Perceptron
MPFM Multiphase Flow Meter
MSEEL Marcellus Shale Energy and Environment Laboratory
NCS Norwegian Continental Shelf
NIST National Institute of Standards and Technology
NN Neural Network
OTDR Optical Time-Domain Reflectometer
PC Personal Computer
PCA Principle Component Analysis
RF Random Forest
RMT Random Matrix Theory
RNN Recurrent Neural Network
SA Sensitivity Analysis
SADG Steam Assisted Gravity Drainage
SoS Speed of Sound
SNR Signal to Noise Ratio
SVM Support Vector Machine
VDU Venous Doppler Ultrasound
VFM Virtual Flow Meter
VOC2012 Visual Object Classes Challenge 2012
VSP Vertical Seismic Profiling
WLR Water in Liquid Ratio

References
1. Equinor. Improving Recovery Rates (IOR). Equinor ASA. 2020. Available online: https://www.equinor.com/en/how-and-why/
increasing-value-creation.html (accessed on 9 November 2020).
2. Business Insider. Crude Oil Price Today. Bus. Insid. 2021. Available online: https://markets.businessinsider.com/commodities/
oil-price (accessed on 9 November 2020).
3. Lake, L.W.; Johns, R.; Rossen, W.R.; Pope, G.A. Fundamentals of Enhanced Oil Recovery; Society of Petroleum Engineers: Richardson,
TX, USA, 2014.
4. Izadmehr, M.; Daryasafar, A.; Bakhshi, P.; Tavakoli, R.; Ghayyem, M.A. Determining influence of different factors on production
optimization by developing production scenarios. J. Pet. Explor. Prod. Technol. 2018, 8, 505–520. [CrossRef]
5. Bukhamsin, A.; Horne, R.N. Using Distributed Acoustic Sensors to Optimize Production in Intelligent Wells. In Proceedings of
the SPE Annual Technical Conference and Exhibition, Amsterdam, The Netherlands, 27–29 October 2014.
6. Gohari, K.; Jutila, H.; Kshirsagar, A.; Chattopadhyay, A.; Mascagnini, C.; Gryaznov, A.; Kidd, P.; Zarei, F. DAS/DTS/DSS/DPS/
DxS-Do We Measure What Adds Value? In Proceedings of the SPE Europec Featured at 78th EAGE Conference and Exhibition,
Vienna, Austria, 30 May–2 June 2016.
7. Hansen, L.S.; Pedersen, S.; Durdevic, P. Multi-phase flow metering in offshore oil and gas transportation pipelines: Trends and
perspectives. Sensors 2019, 19, 2184. [CrossRef] [PubMed]
8. Bukhamsin, A.; Horne, R. Cointerpretation of Distributed Acoustic and Temperature Sensing for Improved Smart Well Inflow
Profiling. In Proceedings of the SPE Western Regional Meeting, Anchorage, AK, USA, 23–26 May 2016.
9. Ünalmis, Ö.H. Sound speed in downhole flow measurement. J. Acoust. Soc. Am. 2016, 140, 430–441. [CrossRef] [PubMed]
10. Unalmis, O.H.; Trehan, S. In-well, optical, strain-based flow measurement technology and its applications. In Proceedings of the
SPE Europec/EAGE Annual Conference, Copenhagen, Denmark, 4–7 June 2012.
11. Liu, D.; Khambampati, A.K.; Kim, S.; Kim, K.Y. Multi-phase flow monitoring with electrical impedance tomography using level
set based method. Nucl. Eng. Des. 2015, 289, 108–116. [CrossRef]
12. Matsui, G. Identification of flow regimes in vertical gas-liquid two-phase flow using differential pressure fluctuations. Int. J.
Multiph. Flow 1984, 10, 711–719. [CrossRef]
13. Babelli, I.M. Development of Multiphase Meter Using Gamma Densitometer Concept; King Abdulaziz City for Science and Technology:
Riyadh, Saudi Arabia, 1997.
14. Weatherford. VSRWet-Gas Flowmeter. Data Sheet. 2018. Available online: https://www.weatherford.com/en/documents/
brochure/products-and-services/production-optimization/vsr-wet-gas-flowmeter/ (accessed on 1 December 2020).
Sensors 2021, 21, 2801 25 of 29

15. Schlumberger. Vx Spectra. Data Sheet 17-TP-302930. 2017. Available online: https://www.slb.com/~/media/Files/testing/
brochures/multiphase/vx_spectra_surface_multiphase_flowmeter_br.pdf (accessed on 1 December 2020).
16. Paz, E.F.D.; Balino, J.L.; Slobodcicov, I. Virtual Metering System for Oil and Gas Field Monitoring Based on a Differential Pressure
Flowmeter. In Proceedings of the SPE Annual Technical Conference and Exhibition, Florence, Italy, 20–22 September 2010.
17. Mokhtari Jadid, K. Performance Evaluation of Virtual Flow Metering Models and Its Application to Metering Backup and Production
Allocation; WIT Transactions on Engineering Sciences: Wessex, UK, 2016; pp. 99–111.
18. Amin, A. Evaluation of commercially available virtual flow meters (VFMs). In Proceedings of the Offshore Technology Conference,
Houston, TX, USA, 4–7 May 2015.
19. Dakin, J.P.; Pratt, D.J.; Bibby, G.W.; Ross, J.N. Distributed optical fibre Raman temperature sensor using a semiconductor light
source and detector. Electron. Lett. 1985, 21, 569–570. [CrossRef]
20. Horiguchi, T.; Kurashima, T.; Tateda, M. Tensile strain dependence of Brillouin frequency shift in silica optical fibers. IEEE
Photonics Technol. Lett. 1989, 1, 107–108. [CrossRef]
21. Juarez, J.C.; Maier, E.W.; Choi, K.N.; Taylor, H.F. Distributed fiber-optic intrusion sensor system. J. Light. Technol. 2005, 23,
2081–2087. [CrossRef]
22. Lu, X.; Thomas, P.J.; Hellevang, J.O. A review of methods for fiber-optic distributed chemical sensing. Sensors 2019, 19, 2876.
[CrossRef]
23. Thomas, P.J.; Hellevang, J.O. A fully distributed fibre optic sensor for relative humidity measurements. Sens. Actuators B Chem.
2017, 247, 284–289. [CrossRef]
24. Totland, C.; Thomas, P.J.; Størdal, I.F.; Eek, E. A fully distributed fibre optic sensor for the detection of liquid hydrocarbons. IEEE
Sens. J. 2020, 21, 7631–7637. [CrossRef]
25. Karaman, O.S.; Kutlik, R.L.; Kluth, E.L. A field trial to test fiber optic sensors for downhole temperature and pressure mea-
surements, West Coalinga Field, California. In Proceedings of the SPE Western Regional Meeting, Anchorage, AK, USA, 22–24
May 1996.
26. Molenaar, M.M.; Hill, D.; Webster, P.; Fidan, E.; Birch, B. First downhole application of distributed acoustic sensing for hydraulic-
fracturing monitoring and diagnostics. SPE Drill. Complet. 2012, 27, 32–38. [CrossRef]
27. Lu, X.; Thomas, P.J. Numerical modeling of Fcy OTDR sensing using a refractive index perturbation approach. J. Light. Technol.
2019, 38, 974–980. [CrossRef]
28. Mateeva, A.; Mestayer, J.; Cox, B.; Kiyashchenko, D.; Wills, P.; Lopez, J.; Grandi, S.; Hornman, K.; Lumens, P.; Franzen, A.; et al.
Advances in distributed acoustic sensing (DAS) for VSP. In SEG Technical Program Expanded Abstracts 2012; Society of Exploration
Geophysicists: Las Vegas, NV, USA, 2012; pp. 1–5.
29. Jin, G.; Mendoza, K.; Roy, B.; Buswell, D.G. Machine learning-based fracture-hit detection algorithm using LFDAS signal. Lead.
Edge 2019, 38, 520–524. [CrossRef]
30. Daley, T.M.; Freifeld, B.M.; Ajo-Franklin, J.; Dou, S.; Pevzner, R.; Shulakova, V.; Kashikar, S.; Miller, D.E.; Goetz, J.; Henninges,
J.; et al. Field testing of fiber-optic distributed acoustic sensing (DAS) for subsurface seismic monitoring. Lead. Edge 2013, 32,
699–706. [CrossRef]
31. Liu, H.; Ma, J.; Xu, T.; Yan, W.; Ma, L.; Zhang, X. Vehicle Detection and Classification Using Distributed Fiber Optic Acoustic
Sensing. IEEE Trans. Veh. Technol. 2019, 69, 1363–1374. [CrossRef]
32. Paleja, R.; Mustafina, D.; Park, T.; Randell, D.; van der Horst, J.; Crickmore, R. Velocity tracking for flow monitoring and
production profiling using distributed acoustic sensing. In Proceedings of the SPE Annual Technical Conference and Exhibition,
Houston, TX, USA, 28–30 September 2015.
33. Cannon, R.T.; Aminzadeh, F. Distributed acoustic sensing: State of the art. In Proceedings of the SPE Digital Energy Conference,
The Woodlands, TX, USA, 5–7 March 2013.
34. Bukhamsin, A. Inflow Profiling and Production Optimization in Smart Wells Using Distributed Acoustic and Temperature
Measurements. Ph.D. Thesis, Stanford University, Stanford, CA, USA, 2016.
35. Wang, X.; Lee, J.; Thigpen, B.; Vachon, G.P.; Poland, S.H.; Norton, D. Modeling flow profile using distributed temperature sensor (DTS)
system. In Proceedings of the Intelligent Energy Conference and Exhibition, Amsterdam, The Netherlands, 25–27 February 2008.
36. Silkina, T. Application of Distributed Acoustic Sensing to Flow Regime Classification. Master’s Thesis, Institutt for Petroleumste-
knologi og Anvendt Geofysikk, Trondheim, Norway, 2014.
37. Vahabi, N.; Selviah, D.R. Convolutional Neural Networks to Classify Oil, Water and Gas Wells Fluid Using Acoustic Signals.
In Proceedings of the 2019 IEEE International Symposium on Signal Processing and Information Technology (ISSPIT), Ajman,
United Arab Emirates, 10–12 December 2019; pp. 1–6.
38. Al-Naser, M.; Elshafei, M.; Al-Sarkhi, A. Artificial neural network application for multiphase flow patterns detection: A new
approach. J. Pet. Sci. Eng. 2016, 145, 548–564. [CrossRef]
39. Andrianov, N. A machine learning approach for virtual flow metering and forecasting. IFAC-PapersOnLine 2018, 51, 191–196.
[CrossRef]
40. Vahabi, N.; Willman, E.; Baghsiahi, H.; Selviah, D.R. Fluid Flow Velocity Measurement in Active Wells Using Fiber Optic
Distributed Acoustic Sensors. IEEE Sens. J. 2020, 20, 11499–11507. [CrossRef]
41. Loh, K.; Omrani, P.S.; van der Linden, R. Deep learning and data assimilation for realtime production prediction in natural gas
wells. arXiv 2018, arXiv:1802.05141.
Sensors 2021, 21, 2801 26 of 29

42. Bikmukhametov, T.; Jäschke, J. First principles and machine learning Virtual Flow Metering: A literature review. J. Pet. Sci. Eng.
2020, 184, 106487. [CrossRef]
43. Yan, Y.; Wang, L.; Wang, T.; Wang, X.; Hu, Y.; Duan, Q. Application of soft computing techniques to multiphase flow measurement:
A review. Flow Meas. Instrum. 2018, 60, 30–43. [CrossRef]
44. Bai, Y.; Bai, Q. Subsea Engineering Handbook; Gulf Professional Publishing: Houston, TX, USA, 2018; pp. 455–487.
45. Camilleri, L.A.; Zhou, W. Obtaining Real-Time Flow Rate, Water Cut, and Reservoir Diagnostics from ESP Gauge Data. In Offshore
Europe; Society of Petroleum Engineers: Aberdeen, UK, 2011.
46. Cheng, B.; Li, Q.; Wang, J.; Wang, Q. Virtual Subsea Flow Metering Technology for Gas Condensate Fields and its Application in
Offshore China. In Proceedings of the International Conference on Offshore Mechanics and Arctic Engineering, Madrid, Spain,
17–22 June 2018; V008T11A030.
47. Ma, X.; Borden, Z.; Porto, P.; Burch, D.; Huang, N.; Benkendorfer, P.; Bouquet, L.; Xu, P.; Swanberg, C.; Hoefer, L.; et al. Real-time
production surveillance and optimization at a mature subsea asset. In Proceedings of the SPE Intelligent Energy International
Conference and Exhibition, Aberdeen, UK, 6–8 September 2016.
48. Xiao, J.J.; Farhadiroushan, M.; Clarke, A.; Abdalmohsen, R.A.; Alyan, E.; Parker, T.R.; Shawash, J.; Milne, H.C. Intelligent
distributed acoustic sensing for in-well monitoring. In Proceedings of the SPE Saudi Arabia Section Technical Symposium and
Exhibition, Al-Khobar, Saudi Arabia, 21–24 April 2014.
49. Corneliussen, S.; Couput, J.; Dahl, E.; Dykesteen, E.; Frøysa, K.; Malde, E.; Moestue, H.; Moksnes, P.; Scheers, L.; Tunheim,
H. Handbook of Multiphase Flow Metering. pp. 18–28. Available online: https://nfogm.no/wp-content/uploads/2014/02/
MPFM_Handbook_Revision2_2005_ISBN-82-91341-89-3.pdf (accessed on 14 April 2021).
50. Wang, F.; Marashdeh, Q.; Fan, L.S.; Warsito, W. Electrical capacitance volume tomography: Design and applications. Sensors 2010,
10, 1890–1917. [CrossRef] [PubMed]
51. Heikkinen, L.M.; Kourunen, J.; Savolainen, T.; Vauhkonen, P.J.; Kaipio, J.P.; Vauhkonen, M. Real time three-dimensional electrical
impedance tomography applied in multiphase flow imaging. Meas. Sci. Technol. 2006, 17, 2083. [CrossRef]
52. Arridge, S.R.; Schotland, J.C. Optical tomography: Forward and inverse problems. Inverse Probl. 2009, 25, 123010. [CrossRef]
53. Hampel, U.; Hoppe, D.; Diele, K.H.; Fietz, J.; Höller, H.; Kernchen, R.; Prasser, H.M.; Zippe, C. Application of gamma tomography
to the measurement of fluid distributions in a hydrodynamic coupling. Flow Meas. Instrum. 2005, 16, 85–90. [CrossRef]
54. Holmås, K.; Løvli, A. FlowmanagerDynamic: A Multiphase Flow Simulator for Online Surveillance, Optimization and Prediction
of Subsea Oil and Gas Production. In Proceedings of the 15th International Conference on Multiphase Production Technology,
Cannes, France, 15–17 June 2011.
55. De Kruif, B.; Leskens, M.; van der Linden, R.; Alberts, G. Soft-sensing for multilateral wells with down hole pressure and
temperature and surface flow measurements. In Proceedings of the Abu Dhabi International Petroleum Exhibition and Conference,
Abu Dhabi, United Arab Emirates, 3–6 November 2008.
56. Figueiredo, M.M.F.; Goncalves, J.L.; Nakashima, A.M.V.; Fileti, A.M.F.; Carvalho, R.D.M. 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. Exp.
Therm. Fluid Sci. 2016, 70, 29–50. [CrossRef]
57. Xu, L.; Zhou, W.; Li, X.; Tang, S. Wet gas metering using a revised Venturi meter and soft-computing approximation techniques.
IEEE Trans. Instrum. Meas. 2010, 60, 947–956. [CrossRef]
58. Meribout, M.; Al-Rawahi, N.; Al-Naamany, A.; Al-Bimani, A.; Al-Busaidi, K.; Meribout, A. Integration of impedance measure-
ments with acoustic measurements for accurate two phase flow metering in case of high water-cut. Flow Meas. Instrum. 2010, 21,
8–19. [CrossRef]
59. Kolla, S.S.; Xu, B.; Nadeem, A.; Luo, Q.; Shirazi, S.A.; Sen, S. Utilizing Artificial Intelligence for Determining Threshold Sand
Rates from Acoustic Monitors. In Proceedings of the SPE Annual Technical Conference and Exhibition, Virtual Conference, 5–7
October 2020.
60. Bikmukhametov, T.; Jäschke, J. Combining Machine Learning and Process Engineering Physics Towards Enhanced Accuracy and
Explainability of Data-Driven Models. Comput. Chem. Eng. 2020, 138, 106834. [CrossRef]
61. Glisic, B. Sensing solutions for assessing and monitoring pipeline systems. In Sensor Technologies for Civil Infrastructures; Woodhead
Publishing: Cambridge, UK, 2014; pp. 422–460.
62. Sakaguchi, S.; Todoroki, S.I.; Shibata, S. Rayleigh scattering in silica glasses. J. Am. Ceram. Soc. 1996, 79, 2821–2824. [CrossRef]
63. Posey, R.; Johnson, G.A.; Vohra, S.T. Strain sensing based on coherent Rayleigh scattering in an optical fiber. Electron. Lett. 2000,
36, 1688–1689. [CrossRef]
64. Kikuchi, K.; Naito, T.; Okoshi, T. Measurement of Raman scattering in single-mode optical fiber by optical time-domain
reflectometry. IEEE J. Quantum Electron. 1988, 24, 1973–1975. [CrossRef]
65. Tateda, M.; Horiguchi, T.; Kurashima, T.; Ishihara, K. First measurement of strain distribution along field-installed optical fibers
using Brillouin spectroscopy. J. Light. Technol. 1990, 8, 1269–1272. [CrossRef]
66. Schenato, L. A review of distributed fiber optic sensors for geo-hydrological applications. Appl. Sci. 2017, 7, 896. [CrossRef]
67. Nikitin, S.P.; Kuzmenkov, A.I.; Gorbulenko, V.V.; Nanii, O.E.; Treshchikov, V.N. Distributed temperature sensor based on a
phase-sensitive optical time-domain Rayleigh reflectometer. Laser Phys. 2018, 28, 085107. [CrossRef]
68. Kuvshinov, B.N. Interaction of helically wound fiber-optic cables with plane seismic waves. Geophys. Prospect. 2016, 64, 671–688.
[CrossRef]
Sensors 2021, 21, 2801 27 of 29

69. Silixa. Carina Sensing System, Breakthrough Performance Delivered by Constellation Fibres. 2018. Available online: https:
//silixa.com/products/carina-sensing-system-enabled-by-constellation-fibre/ (accessed on 1 December 2020).
70. Nokes, G. Optimising power transmission and distribution networks using optical fiber distributed temperature sensing systems.
Power Eng. J. 1999, 13, 291–296. [CrossRef]
71. Cram, D.; Hatch, C.E.; Tyler, S.; Ochoa, C. Use of distributed temperature sensing technology to characterize fire behavior. Sensors
2016, 16, 1712. [CrossRef]
72. Mishra, A.; Soni, A. Leakage detection using fiber optics distributed temperature sensing. In Proceedings of the Abu Dhabi
International Petroleum Exhibition and Conference, Abu Dhabi, United Arab Emirates, 13–16 November 2017.
73. Inaudi, D.; Glisic, B. Distributed fiber optic strain and temperature sensing for structural health monitoring. In Proceedings of the
3rd International Conference on Bridge Maintenance, Safety and Management, Porto, Portugal, 16–19 July 2006; pp. 16–19.
74. Downes, J.; Leung, H.Y. Distributed temperature sensing worldwide power circuit monitoring applications. In Proceedings of the
IEEE 2004 International Conference on Power System Technology (PowerCon 2004), Singapore, 21–24 November 2004; Volume 2,
pp. 1804–1809.
75. Smolen, J.J.; van der Spek, A. Distributed temperature sensing. In A Primer for Oil and Gas Production; Shell: Missouri City, TX,
USA, 2003.
76. Ukil, A.; Braendle, H.; Krippner, P. Distributed temperature sensing: review of technology and applications. IEEE Sens. J. 2011,
12, 885–892. [CrossRef]
77. Sharma, J.; Cuny, T.; Ogunsanwo, O.; Santos, O. Low-Frequency Distributed Acoustic Sensing for Early Gas Detection in a
Wellbore. IEEE Sens. J. 2020, 21, 6158–6169. [CrossRef]
78. Tejedor, J.; Macias-Guarasa, J.; Martins, H.F.; Pastor-Graells, J.; Corredera, P.; Martin-Lopez, S. Machine learning methods for
pipeline surveillance systems based on distributed acoustic sensing: A review. Appl. Sci. 2017, 7, 841. [CrossRef]
79. Soroush, M.; Mohammadtabar, M.; Roostaei, M.; Hosseini, S.A.; Fattahpour, V.; Mahmoudi, M.; Keough, D.; Tywoniuk, M.;
Cheng, L.; Moez, K. Fiber Optics Application for Downhole Monitoring and Wellbore Surveillance; SAGD Monitoring, Flow
Regime Determination and Flow Loop Design. In Proceedings of the SPE Canada Heavy Oil Conference, Virtual Conference, 28
September–2 October 2020.
80. Ajo-Franklin, J.B.; Dou, S.; Lindsey, N.J.; Monga, I.; Tracy, C.; Robertson, M.; Tribaldos, V.R.; Ulrich, C.; Freifeld, B.; Daley, T.; et al.
Distributed acoustic sensing using dark fiber for near-surface characterization and broadband seismic event detection. Sci. Rep.
2019, 9, 1–14.
81. Friend, D.G. Speed of sound as a thermodynamic property of fluids. In Experimental Methods in the Physical Sciences; Academic
Press: Cambridge, MA, USA, 2001; Volume 39, pp. 237–306.
82. Lemmon, E.W.; McLinden, M.O.; Friend, D.G. Thermophysical Properties of Fluid Systems. In NIST Chemistry WebBook, NIST
Standard Reference Database Number 69; Linstrom, P.J., Mallard, W.G., Eds.; National Institute of Standards and Technology:
Gaithersburg, MD, USA, 2020. [CrossRef]
83. Chaudhuri, A.; Osterhoudt, C.F.; Sinha, D.N. An algorithm for determining volume fractions in two-phase liquid flows by
measuring sound speed. J. Fluids Eng. 2012, 134, 101301. [CrossRef]
84. Huber, M.L. NIST Thermophysical Properties of Hydrocarbon Mixtures Database (SUPERTRAPP), Version 3.2; National Institute of
Standards and Technology: Gaithersburg, MD, USA, 2007.
85. Johannessen, K.; Drakeley, B.K.; Farhadiroushan, M. Distributed Acoustic Sensing—A new way of listening to your well/reservoir.
In Proceedings of the SPE Intelligent Energy International, Utrecht, The Netherlands, 27–29 March 2012.
86. Finfer, D.; Parker, T.R.; Mahue, V.; Amir, M.; Farhadiroushan, M.; Shatalin, S. Non-intrusive multiple zone distributed acoustic
sensor flow metering. In Proceedings of the SPE Annual Technical Conference and Exhibition, Houston, TX, USA, 28–30
September 2015.
87. Fidaner, O. Downhole Multiphase Flow Monitoring Using Fiber Optics. In Proceedings of the SPE Annual Technical Conference
and Exhibition, San Antonio, TX, USA, 9–11 October 2017.
88. Hemink, G.; van der Horst, J. On the Use of Distributed Temperature Sensing and Distributed Acoustic Sensing for the Application
of Gas Lift Surveillance. SPE Prod. Oper. 2018, 33, 896–912. [CrossRef]
89. Shirdel, M.; Buell, R.S.; Wells, M.J.; Muharam, C.; Sims, J.C. Horizontal-Steam-Injection-Flow Profiling Using Fiber Optics. SPE J.
2019, 24, 431–451. [CrossRef]
90. Soroush, M.; Roostaei, M.; Fattahpour, V.; Mahmoudi, M.; Keough, D.; Cheng, L.; Moez, K. Prognostics Thermal Well Management:
A Review on Wellbore Monitoring and the Application of Distributed Acoustic Sensing DAS for Steam Breakthrough Detection.
In Proceedings of the SPE Thermal Well Integrity and Design Symposium, Banff, AB, Canada, 19–21 November 2019.
91. Cerrahoglu, C.; Naldrett, G.; Vigrass, A.; Aghayev, R. Cluster Flow Identification During Multi-Rate Testing Using a Wireline
Tractor Conveyed Distributed Fiber Optic Sensing System With Engineered Fiber on a HPHT Horizontal Unconventional Gas
Producer in the Liard Basin. In Proceedings of the SPE Annual Technical Conference and Exhibition, Calgary, AB, Canada, 30
September–2 October 2019.
92. Viola, P.; Jones, M. Robust real-time object detection. Int. J. Comput. Vis. 2001, 4, 4.
93. Wang, Z. The Uses of Distributed Temperature Survey (DTS) Data. Ph.D. Dissertation, Stanford University: Stanford, CA, USA, 2012.
94. Willis, M.E.; Barfoot, D.; Ellmauthaler, A.; Wu, X.; Barrios, O.; Erdemir, C.; Shaw, S.; Quinn, D. Quantitative quality of distributed
acoustic sensing vertical seismic profile data. Lead. Edge 2016, 35, 605–609. [CrossRef]
Sensors 2021, 21, 2801 28 of 29

95. Bikmukhametov, T.; Jäschke, J. Oil production monitoring using gradient boosting machine learning algorithm. In Proceedings
of the 12th IFAC Symposium on Dynamics and Control of Process Systems, including Biosystems, Florianopolis, Brazil, 23–26
April 2019.
96. Jalilian, S.E.; Huang, D.; Leung, H.; Ma, K.F.; Hifi Engineering Inc. Method of Estimating Flowrate in a Pipeline. U.S. Patent
Application 16/310,375, 31 October 2019.
97. Vidana-Vila, E.; Navarro, J.; Borda-Fortuny, C.; Stowell, D.; Alsina-Pagès, R.M. Low-Cost Distributed Acoustic Sensor Network
for Real-Time Urban Sound Monitoring. Electronics 2020, 9, 2119. [CrossRef]
98. Press, G. Cleaning Big Data: Most Time-Consuming, Least Enjoyable Data Science Task, Survey Says. Forbes, 2016. Available
online: https://www.forbes.com/sites/gilpress/2016/03/23/data-preparation-most-time-consuming-least-enjoyable-data-
science-task-survey-says (accessed on 1 December 2020).
99. Shi, Y.; Wang, Y.; Zhao, L.; Fan, Z. An event recognition method for otdr sensing system based on deep learning. Sensors 2019,
19, 3421. [CrossRef]
100. Wang, S.; Jiang, J.; Wang, S.; Ma, Z.; Xu, T.; Ding, Z.; Lv, Z.; Liu, T. GPU-based fast processing for a distributed acoustic sensor
using an LFM pulse. Appl. Opt. 2020, 59, 11098–11103. [CrossRef]
101. Shiloh, L.; Eyal, A.; Giryes, R. Efficient Processing of Distributed Acoustic Sensing Data Using a Deep Learning Approach. J.
Light. Technol. 2019, 37, 4755–4762. [CrossRef]
102. Nanni, L.; Ghidoni, S.; Brahnam, S. Handcrafted vs. non-handcrafted features for computer vision classification. Pattern Recognit.
2017, 71, 158–172. [CrossRef]
103. Onajite, E. Seismic Data Analysis Techniques in Hydrocarbon Exploration; Elsevier: Amsterdam, The Netherlands, 2013.
104. Ghahfarokhi, P.K.; Carr, T.; Bhattacharya, S.; Elliott, J.; Shahkarami, A.; Martin, K. A fiber-optic assisted multilayer perceptron
reservoir production modeling: A machine learning approach in prediction of gas production from the marcellus shale. In
Proceedings of the Unconventional Resources Technology Conference, Houston, TX, USA, 23–25 July 2018; pp. 3291–3300.
105. Lu, J.; Liong, V.E.; Zhou, X.; Zhou, J. Learning compact binary face descriptor for face recognition. IEEE Trans. Pattern Anal. Mach.
Intell. 2015, 37, 2041–2056. [CrossRef] [PubMed]
106. Shaban, H.; Tavoularis, S. Measurement of gas and liquid flow rates in two-phase pipe flows by the application of machine
learning techniques to differential pressure signals. Int. J. Multiph. Flow 2014, 67, 106–117. [CrossRef]
107. Lorentzen, R.J.; Saevareid, O.; Naevdal, G. Soft Multiphase Flow Metering for Accurate Production Allocation (Russian). In
Proceedings of the SPE Russian Oil and Gas Conference and Exhibition, Moscow, Russia, 26–28 October 2010.
108. Babanezhad, M.; Nakhjiri, A.T.; Rezakazemi, M.; Marjani, A.; Shirazian, S. Functional input and membership characteristics in
the accuracy of machine learning approach for estimation of multiphase flow. Sci. Rep. 2020, 10, 1–15. [CrossRef] [PubMed]
109. Arief, H.A.; Strand, G.H.; Tveite, H.; Indahl, U.G. Land cover segmentation of airborne LiDAR data using stochastic atrous
network. Remote Sens. 2018, 10, 973. [CrossRef]
110. Park, T.; Paleja, R.; Wojtaszek, M. Robust Regression and Band Switching to Improve DAS Flow Estimates. In Proceedings of the
SPE Annual Technical Conference and Exhibition, Dallas, TX, USA, 24–26 September 2018.
111. Alkhalaf, M.; Hveding, F.; Arsalan, M. Machine Learning Approach to Classify Water Cut Measurements using DAS Fiber Optic
Data. In Proceedings of the Abu Dhabi International Petroleum Exhibition & Conference, Abu Dhabi, United Arab Emirates,
11–14 November 2019.
112. Bhattacharya, S.; Ghahfarokhi, P.K.; Carr, T.R.; Pantaleone, S. Application of predictive data analytics to model daily hydrocarbon
production using petrophysical, geomechanical, fiber-optic, completions, and surface data: A case study from the Marcellus
Shale, North America. J. Pet. Sci. Eng. 2019, 176, 702–715. [CrossRef]
113. Gal, Y.; Ghahramani, Z. Dropout as a bayesian approximation: Representing model uncertainty in deep learning. In Proceedings
of the International Conference on Machine Learning, New York, NY, USA, 19–24 June 2016; pp. 1050–1059.
114. Park, J.; Sandberg, I.W. Universal approximation using radial-basis-function networks. Neural Comput. 1991, 3, 246–257.
[CrossRef]
115. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on
Computer Vision and Pattern Recognition, Las Vegas, NV, USA, 27–30 June 2016; pp. 770–778.
116. Brown, T.B.; Mann, B.; Ryder, N.; Subbiah, M.; Kaplan, J.; Dhariwal, P.; Neelakantan, A.; Shyam, P.; Sastry, G.; Askell, A.; et al.
Language models are few-shot learners. arXiv 2020, arXiv:2005.14165.
117. Shi, X.; Chen, Z.; Wang, H.; Yeung, D.Y.; Wong, W.K.; Woo, W.C. Convolutional LSTM Network: a machine learning approach
for precipitation nowcasting. In Proceedings of the 28th International Conference on Neural Information Processing Systems,
Montreal, QC, Canada, 7–12 December 2015; Volume 1, pp. 802–810.
118. Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A.N.; Kaiser, Ł.; Polosukhin, I. Attention is all you need.
In Proceedings of the 31st International Conference on Neural Information Processing Systems, Long Beach, CA, USA, 4–9
December 2017; pp. 6000–6010.
119. Arief, H.A.; Indahl, U.G.; Strand, G.H.; Tveite, H. Addressing overfitting on point cloud classification using Atrous XCRF. ISPRS
J. Photogramm. Remote Sens. 2019, 155, 90–101. [CrossRef]
120. Kim, B.; Kim, H.; Kim, K.; Kim, S.; Kim, J. Learning not to learn: Training deep neural networks with biased data. In Proceedings
of the IEEE Conference on Computer Vision and Pattern Recognition, Long Beach, CA, USA, 16–20 June 2019; pp. 9012–9020.
Sensors 2021, 21, 2801 29 of 29

121. Deng, J.; Dong, W.; Socher, R.; Li, L.J.; Li, K.; Li, F. Imagenet: A large-scale hierarchical image database. In Proceedings of the 2009
IEEE Conference on Computer Vision and Pattern Recognition, Miami, FL, USA, 20–25 June 2009; pp. 248–255.
122. Everingham, M.; Eslami, S.A.; Van Gool, L.; Williams, C.K.; Winn, J.; Zisserman, A. The pascal visual object classes challenge: A
retrospective. Int. J. Comput. Vis. 2015, 111, 98–136. [CrossRef]
123. Niemeyer, J.; Rottensteiner, F.; Soergel, U. Contextual classification of lidar data and building object detection in urban areas.
ISPRS J. Photogramm. Remote Sens. 2014, 87, 152–165. [CrossRef]
124. Geiger, A.; Lenz, P.; Urtasun, R. Are we ready for autonomous driving? The kitti vision benchmark suite. In Proceedings of the
2012 IEEE Conference on Computer Vision and Pattern Recognition, Providence, RI, USA, 18–20 June 2012; pp. 3354–3361.
125. Beard, P. Biomedical photoacoustic imaging. Interface Focus 2011, 1, 602–631. [CrossRef] [PubMed]
126. Rossvoll, O.; Hatle, L.K. Pulmonary venous flow velocities recorded by transthoracic Doppler ultrasound: relation to left
ventricular diastolic pressures. J. Am. Coll. Cardiol. 1993, 21, 1687–1696. [CrossRef]
127. Allman, D.; Reiter, A.; Bell, M.A.L. A machine learning method to identify and remove reflection artifacts in photoacoustic
channel data. In Proceedings of the 2017 IEEE International Ultrasonics Symposium (IUS), Washington, DC, USA, 6–9 September
2017; pp. 1–4.
128. Godefroy, G.; Arnal, B.; Bossy, E. Compensating for visibility artefacts in photoacoustic imaging with a deep learning approach
providing prediction uncertainties. Photoacoustics 2020, 21, 100218. [CrossRef]
129. Hauptmann, A.; Cox, B.; Lucka, F.; Huynh, N.; Betcke, M.; Beard, P.; Arridge, S. Approximate k-space models and deep learning
for fast photoacoustic reconstruction. In Proceedings of the International Workshop on Machine Learning for Medical Image
Reconstruction, Granada, Spain, 16 September 2018; pp. 103–111.
130. Chambers, K. Using DAS to investigate traffic patterns at Brady Hot Springs, Nevada, USA. Lead. Edge 2020, 39, 819–827.
[CrossRef]
131. Tejedor, J.; Macias-Guarasa, J.; Martins, H.F.; Martin-Lopez, S.; Gonzalez-Herraez, M. A contextual GMM-HMM smart fiber optic
surveillance system for pipeline integrity threat detection. J. Light. Technol. 2019, 37, 4514–4522. [CrossRef]
132. Jia, H.; Liang, S.; Lou, S.; Sheng, X. A k-Nearest Neighbor Algorithm-Based Near Category Support Vector Machine Method for
Event Identification of ϕ-OTDR. IEEE Sens. J. 2019, 19, 3683–3689. [CrossRef]
133. Wu, H.; Liu, X.; Xiao, Y.; Rao, Y. A dynamic time sequence recognition and knowledge mining method based on the hidden
markov models (hmms) for pipeline safety monitoring with OTDR. J. Light. Technol. 2019, 37, 4991–5000. [CrossRef]
134. Wang, Z.; Zheng, H.; Li, L.; Liang, J.; Wang, X.; Lu, B.; Ye, Q.; Qu, R.; Cai, H. Practical multi-class event classification approach for
distributed vibration sensing using deep dual path network. Opt. Express 2019, 27, 23682–23692. [CrossRef] [PubMed]
135. Binder, G.; Chakraborty, D. Detecting microseismic events in downhole distributed acoustic sensing data using convolutional
neural networks. In SEG Technical Program Expanded Abstracts 2019; Society of Exploration Geophysicists: San Antonio, TX, USA,
2019; pp. 4864–4868.
136. Peng, Z.; Jian, J.; Wen, H.; Wang, M.; Liu, H.; Jiang, D.; Mao, Z.; Chen, K.P. Fiber-optical distributed acoustic sensing signal
enhancements using ultrafast laser and artificial intelligence for human movement detection and pipeline monitoring. In Optical
Data Science II; International Society for Optics and Photonics: San Francisco, CA, USA, 2–7 February 2019; Volume 10937, 109370J.
137. Wiesmeyr, C.; Litzenberger, M.; Waser, M.; Papp, A.; Garn, H.; Neunteufel, G.; Döller, H. Real-time train tracking from distributed
acoustic sensing data. Appl. Sci. 2020, 10, 448. [CrossRef]
138. Peng, Z.; Wen, H.; Jian, J.; Gribok, A.; Wang, M.; Huang, S.; Liu, H.; Mao, Z.H.; Chen, K.P. Identifications and classifications of
human locomotion using Rayleigh-enhanced distributed fiber acoustic sensors with deep neural networks. Sci. Rep. 2020, 10,
1–11. [CrossRef] [PubMed]
139. Li, Z.; Zhang, J.; Wang, M.; Zhong, Y.; Peng, F. Fiber distributed acoustic sensing using convolutional long short-term memory
network: A field test on high-speed railway intrusion detection. Opt. Express 2020, 28, 2925–2938. [CrossRef]
140. Bencharif, B.A.E.; Ölçer, I.; Özkan, E.; Cesur, B. Detection of Acoustic Signals from Distributed Acoustic Sensor Data with
Random Matrix Theory and Their Classification Using Machine Learning. Available online: https://www.spiedigitallibrary.org/
conference-proceedings-of-spie/11525/115251S/Detection-of-acoustic-signals-from-Distributed-Acoustic-Sensor-data-with/
10.1117/12.2581696.full?SSO=1 (accessed on 14 April 2021).
141. Huot, F.; Biondi, B. Detecting earthquakes through telecom fiber using a convolutional neural network. In Proceedings of the
SEG Annual Meeting and International Exhibition, Virtual Conference, 11–16 October 2020; pp. 3452–3456.