852 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 66, NO.

5, MAY 2017

Gas–Liquid Two-Phase Flow Measurement Using

Coriolis Flowmeters Incorporating Artificial
Neural Network, Support Vector Machine,
and Genetic Programming Algorithms
Lijuan Wang, Student Member, IEEE, Jinyu Liu, Yong Yan, Fellow, IEEE, Xue Wang, and Tao Wang

Abstract— Coriolis flowmeters are well established for the flowrate of a two-phase mixture is challenging in industry.
mass flow measurement of single-phase flow with high accuracy. Significant research based on traditional flowmeters for two-
In recent years, attempts have been made to apply Coriolis phase flow measurement has been conducted, such as Venturi,
flowmeters to measure two-phase flow. This paper presents data
driven models that are incorporated into Coriolis flowmeters V-cone, turbine, vortex, and slotted orifice meters [1]–[3].
to measure both the liquid mass flowrate and the gas volume The determination of gas volume fraction of two-phase flow
fraction of a two-phase flow mixture. Experimental work was is crucial for the optimization of some industrial processes.
conducted on a purpose-built two-phase flow test rig on both Resistive sensors, capacitive sensors, electrical capacitance
horizontal and vertical pipelines for a liquid mass flowrate tomography, electrical resistance tomography, and microwave
ranging from 700 to 14500 kg/h and a gas volume fraction
between 0% and 30%. Artificial neural network (ANN), support probes have been proposed for the phase fraction measurement
vector machine (SVM), and genetic programming (GP) models of two-phase flow [4]–[6]. These techniques are often referred
are established through training with the experimental data. The to as direct method, since the systems are designed to measure
performance of backpropagation-ANN (BP-ANN), radial basis the desired two-phase flow characteristics directly. Due to
function-ANN (RBF-ANN), SVM, and GP models is assessed and the difficult nature of two-phase flow and complexity of the
compared. Experimental results suggest that the SVM models
are superior to the BP-ANN, RBF-ANN, and GP models for sensing systems, the applications of such direct two-phase
two-phase flow measurement in terms of robustness and accuracy. flowmeters have achieved limited success in industry.
For liquid mass flowrate measurement with the SVM models, Indirect techniques based on traditional sensors incorpo-
93.49% of the experimental data yield a relative error less rating soft-computing algorithms, such as artificial neural
than ±1% on the horizontal pipeline, while 96.17% of the results network (ANN), support vector machine (SVM), least-squares
are within ±1% on the vertical installation. The SVM models
predict the gas volume fraction with a relative error less than SVM, and extreme learning machine together with genetic
±10% for 93.10% and 94.25% of the test conditions on the algorithms or particle swarm optimization, have also been
horizontal and vertical installations, respectively. applied to two-phase or multiphase flow measurement or
Index Terms— Artificial neural network (ANN), Coriolis mass flow regime identification [7]–[10]. Coriolis flowmeters, as
flowmeter, flow measurement, gas volume fraction, genetic pro- one of the most accurate single-phase mass flowmeters, have
gramming (GP), support vector machine (SVM), two-phase flow. been successfully applied to a range of industrial applica-
tions. In recent years, many researchers have attempted to
use Coriolis flowmeters for two-phase or multiphase flow
I. I NTRODUCTION measurement [11]. However, despite recent progress in sensor
and transmitter technologies, improving the accuracy for mass

G AS–LIQUID two-phase flow is widely seen in oil and

gas fields, chemical engineering, food processing, and
other industrial processes. The accurate measurement of the
flow metering of liquid with entrained gas still remains a
challenge. A bubble effect model was proposed to study gas–
liquid two-phase flow for Coriolis flowmeters [12], but it
Manuscript received July 3, 2016; revised November 3, 2016; accepted cannot deal with positive errors in the mass flow measurement.
November 7, 2016. Date of publication December 19, 2016; date of current Subsequently, Liu et al. [13] used a neural network to correct
version April 5, 2017. This work was supported by the UK CCS Research mass flow errors in a Coriolis mass flowmeter, which was
Centre which is a part of the RCUK Energy Programme founded by the
EPSRC. The Associate Editor coordinating the review process was Dr. based on a horizontal flow tube and the flow rate was limited to
Domenico Grimaldi. (Corresponding author: Yong Yan.) 1.5–3.6 kg/s. The multilayer perceptron and radial basis func-
L. Wang, J. Liu, and Y. Yan are with the School of Engineering and Digital tion (RBF) networks include four inputs, i.e., temperature,
Arts, University of Kent, Canterbury CT2 7NT, U.K.
X. Wang is with the School of Mathematics, Statistics and Actuarial Science, damping, density drop, and flowrate to estimate mass flow
University of Kent, Canterbury CT2 7NF, U.K. errors. Although most of the mass flow errors were reduced
T. Wang is with KROHNE Ltd., Wellingborough NN8 6AE, U.K. to within ±2%, the gas entrainment was not quantified and
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. different installation conditions were not considered. A method
Digital Object Identifier 10.1109/TIM.2016.2634630 based on fuzzy inference was proposed to correct the mass
This work is licensed under a Creative Commons Attribution 3.0 License. For more information, see http://creativecommons.org/licenses/by/3.0/
WANG et al.: GAS–LIQUID TWO-PHASE FLOW MEASUREMENT USING CORIOLIS FLOWMETERS 853

flow errors of a Coriolis flowmeter for the measurement of

two-phase flow [14]. The fuzzy system accepts damping, drop
in density, and apparent mass flowrate as inputs to generate
corrected mass flowrate. Lari and Shabaninia [15] applied a
neuro-fuzzy algorithm to the error correction of a Coriolis
mass flowmeter for air–water two-phase flow measurement.
However, the experimental data and the results were not Fig. 1. Principle and structure of the measurement system.
explained in detail in [14] and [15]. Hou et al. [16] developed
a digital Coriolis flow transmitter and tested a commercial
Coriolis flowmeter. The measurement errors achieved under
gas–liquid two-phase flow conditions were corrected using a
feedforward neural network with two inputs—apparent liquid
mass flowrate and apparent drop in density. Xing et al. [17]
applied a Coriolis flowmeter in combination with an ultrasonic
flowmeter to measure the individual mass flowrates of gas–
liquid two-phase flow under low liquid loading. The root-
mean-square errors of gas and liquid mass flowrates were
3.09% and 12.78%, respectively. Ma et al. [18] used a
25-mm bore Coriolis flowmeter together with SVM algorithms
to measure the overall mass flowrate of oil–water two-phase
flow and achieved relative errors within ±1%. The mass Fig. 2. Structure of a BP-ANN.
flowrate of individual phase was obtained with the maximum
error of ±8%. However, it is known that the gas entrained in
a liquid flow affects significantly the performance of Corio- examples by constructing an input–output mapping in order
lis flowmeters, especially under different flow regimes [11]. to perform estimations of desired outputs. Fig. 1 shows the
Moreover, very little research has been undertaken to date to principle and structure of the measurement system. The data
predict the gas volume fraction from the outputs of a Coriolis driven models accept variables from a Coriolis flowmeter and
flowmeter. a DP transducer, while the output gives the corrected mass
Owning to the good reproducibility of the measurement flowrate or predicted gas volume fraction. The analysis of
errors of Coriolis flowmeters under two-phase flow conditions, parametric dependence and input variable selection for the data
data driven models, such as ANN, SVM, and genetic pro- driven models based on the experimental data is presented
gramming (GP), have the potential to correct the liquid mass in Section III-C. Since the volume of data is often limited
flowrate and predict gas volume fraction. In this paper, experi- in practice, it is appropriate to design a separate model for
mental work was undertaken on a purpose-built 1-in (25 mm) each desired output. The structure of each data driven model
bore air–water two-phase flow test rig. Coriolis flowmeters based on ANN, SVM, and GP will be explained in detail in
(KROHNE OPTIMASS 6400 S25) in conjunction with DP Sections II.B–II.E.
transducers were applied to obtain liquid mass flowrate and
gas volume fraction on both the horizontal and vertical pipes. B. BP-ANN
Parametric dependence along with input variable selection for
BP-ANN is a multilayer feedforward neural network trained
the data driven models is investigated based on the partial
with a BP learning algorithm, which is one of the most com-
mutual information (PMI) algorithm [19], [20]. Four data
mon neural networks. A BP-ANN consists of an input layer,
driven models based on backpropagation-ANN (BP-ANN),
one or more hidden layers, and an output layer. The hidden
RBF-ANN, SVM, and GP, respectively, are established and
layer connects the input and output layers and represents their
validated through training and testing with the experimental
quantitative relationship. In general, a neural network with a
data. The performances of the four models are evaluated and
single-hidden layer of sufficient neurons is able to represent
compared in terms of robustness and accuracy. The basic
any nonlinear problem. In consideration of the simplicity
principle of BP-ANN modeling with some preliminary results
of the ANN structure, a single-hidden layer is chosen and
was reported at the 2016 IEEE International Instrumentation
investigated in this paper.
and Measurement Technology Conference [21]. This paper
As shown in Fig. 2, x = [x 1 , x 2 , . . . , x n ]T is an input sample
presents in detail the principles, structures, training, and per-
and y is the desired output. Assume y is the linear output of
formance comparisons of the BP-ANN, RBF-ANN, SVM, and
the hidden neurons and a transfer function f (x) is used on
GP models.
the neurons, the ANN is modeled as
II. M ETHODOLOGY  n 
L L 
A. Overall Measurement Strategy yBP = ωj Hj + b = ωj f ωi j x i + a j + b (1)
ANN, SVM, and GP are common data driven models for j =1 j =1 i=1

modeling a nonlinear system with multiple inputs and outputs where n and L are the numbers of input variables and hidden
[22]–[26]. These techniques learn from history data and give nodes. ω j is the weight connecting the j th hidden node and
854 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 66, NO. 5, MAY 2017

Fig. 4. Structure of an SVM.

Fig. 3. Structure of an RBF-ANN.

and structural risk minimization. Then, this method has been
extended to the domain of regression and prediction prob-
the output node, and ωi j is the weight connecting the i th lems [28]. As shown in Fig. 4, the input vector x is first
input node to the j th hidden node. a j and b are the biases mapped into an L-dimensional feature space using transfer
on the j th hidden node and the output node. In this paper, functions, and then, a linear model is constructed in this feature
the hyperbolic tangent sigmoid function is used as a transfer space.
function on hidden neurons and presented by The linear model in the feature space is given by
2 y = ωx + b (4)
f (x) = − 1. (2)
1 + e−2x
where ω = (ω1 , ω2 , . . . , ωL ) is the weight vector and b is the
The learning algorithm is described as a procedure that con-
bias term.
sists of adjusting the weights and biases of a network, to
Regression estimates can be obtained by minimizing the
minimize an error function between the network output and
empirical risk on the training data. SVM regression performs
desired output for a given set of inputs. The BP algorithm
a linear regression in the high-dimensional feature space using
has been widely applied to solve practical problems. However,
ε-insensitive loss and tends to reduce the model complexity
the BP algorithm has the disadvantage of slow convergence
by minimizing ω2 . This can be described by introducing
and long training time. In addition, the success of the BP
slack variables ξi and ξi (i = 1, 2, . . . , m) to measure the
algorithm depends on the user-dependent parameters, such as
deviation of training samples (X ∗ , D) outside ε-insensitive
initialization and structure of the ANN.
zone. X ∗ = (x 1 , x 2 , . . . , x m ) represents m input vectors of
training samples and D = (d1 , d2 , . . . , dm ) is the correspond-
C. RBF-ANN ing desired output. Thus, the optimization problem can be
RBF-ANN has a fixed three layer structure (Fig. 3) and uses formulated as
a type of RBF as an activation function to the hidden nodes. 1 m
 
The output of the network is a linear combination of RBFs  = min ω2 + C ξi + ξi (5)
2
of the inputs and neuron parameters. The RBF measures the i=1
distance between the input vectors and the weight vectors and where m is the number of training samples. C is a positive
is typically taken to be the Gaussian function. Thus, the output constant as a regularization parameter that allows tuning the
of the network is given by tradeoff between the flatness of the function and the tolerance
L L   of deviations larger than ε (a constant).
1
yRBF = ωj Hj = ω j exp − 2 x − C j 2 (3) Minimize the risk function of (5) subject to the following
2σ constraints:
j =1 j =1

where C j is the center vector for the j th hidden node and di − yi ≤ ε + ξi (6)
determined by the K-means clustering method. x − C j  is
yi − di ≤ ε + ξi (7)
the Euclidean norm and σ 2 is the variance of the Gaussian
function. ξi ≥ 0 (8)
An RBF network with enough hidden nodes can approxi-
ξi ≥ 0. (9)
mate any continuous function with arbitrary precision. More-
over, as a local approximation network, the RBF neural Equation (4) can be transformed into a dual problem and
network has the advantages of simple structure, less adjustive solved by Lagrange functional
parameters, and fast training.

L
 
y= αi − αi∗ · K (x, x i ) + b (10)
D. SVM i=1
SVM was developed by Cortes and Vapnik [27] to solve the where αi and αi∗ are Lagrange multipliers and K (x, x i ) is a
classification problem based on the statistic learning theory kernel function.
WANG et al.: GAS–LIQUID TWO-PHASE FLOW MEASUREMENT USING CORIOLIS FLOWMETERS 855

Fig. 5. Structure of a GP model.

There are some optional kernel functions for SVM, such

as linear, polynomial, RBF, and sigmoid function. One of
the most widely used kernel functions is the RBF. The final
product of a training process in the SVM method can be
presented by
L  
 ∗
 1
ySVM = αi − αi · exp − 2 x − x i  + b. (11)
2
2σ
i=1

E. GP
GP as an evolutionary computation technique is an exten- Fig. 6. Schematic of the two-phase flow test rig.
sion of genetic algorithms and is widely applied to symbolic
data mining (symbolic regression, classification, and opti-
mization) [29]–[31]. Unlike the traditional regression analysis,
GP-based symbolic regression automatically evolves both the
structure and the parameters of the mathematical model from
the available data. Meanwhile, it is superior to other machine
learning techniques due to the ability to generate an empir-
ical mathematical equation without assuming prior form of
the existing relationships. In this paper, multigene symbolic
regression is applied to establish a model for two-phase
flow measurement. The structure of a multigene symbolic
regression model is shown in Fig. 5.
The GP model can be regarded as a linear combination
of lower order nonlinear transformations of the input vari- Fig. 7. Photograph of the test Coriolis flowmeters on 1-inch pipelines.
ables. The output yGP is defined as a vector output of n
trees modified by the bias term b0 and scaling parameters
b1 , . . . , bn are expected to be transportable to other gas–liquid two-phase
flow conditions. The gas flow is set to enter to the liquid flow
yG P = b0 + b1 t1 + · · · + bn tn (12)
through a bypass on the pipe. The liquid mass flowrate is
where ti (i = 1, . . . , n) is the (m × 1) vector of outputs from controlled by adjusting the pump frequency from 15% to 80%.
the i th tree comprising a multigene individual. The gas flowrate is varied by adjusting the opening of the
The evolutionary process starts with initial population by valve in a gas flow controller. Two independent Coriolis
creating individuals containing GP trees with different genes flowmeters (KROHNE OPTIMASS 6400 S25 and Bronkhorst
generated randomly. The evolutionary process continues with mini CORI-FLOW M15) were installed before the mixer to
an evaluation of the fitness of the new population, two-point provide references for the individual mass flow rates of the
high-level crossover to acquire and delete genes, and low-level liquid and gas phases, respectively. Both reference meters’
crossover on subtrees. Then, the created trees replace the measurement uncertainties under single-phase conditions were
parent trees or the unaltered individual in the next generation verified according to the manufacturer’s technical specifica-
through mutation operators. The best program that appeared tion. In the downstream, two additional Coriolis flowmeters
in any generation, the best-so-far solution, defines the output (see Fig. 7) of the same type as the liquid reference meter
of the GP algorithm [30]. were installed in the vertical and horizontal test sections,
respectively. These are the meters under test to assess the
III. E XPERIMENTAL R ESULTS AND D ISCUSSION performance of ANN, SVM, and GP models under two-
phase flow conditions. In view of the effects of gravity and
A. Test Rig and Experimental Conditions buoyancy on two-phase fluid, both the horizontal and vertical
Fig. 6 shows the schematic of the two-phase flow test rig installations of the meters are considered. A DP transducer
that was used in this paper. The measurement data obtained was used to record the DP value across each flowmeter under
on this rig and subsequent conclusions drawn from the data test.
856 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 66, NO. 5, MAY 2017

Fig. 9. Original errors of the liquid mass flowrate from Test II. (a) Horizontal
pipeline. (b) Vertical pipeline.

Fig. 8. Original errors of the liquid mass flowrate from Test I. (a) Horizontal
pipeline. (b) Vertical pipeline. errors become positive and crossing the zero line and then
return to negative errors again along with increasing entrained
gas. This is believed to be due to the flow regime effects
The data logging frequencies, as set in the data loggers for on the fluid-tube coupling system at different flowrates. At a
the mass flowrate, density, damping, and DP, are 50, 10, 2, lower flowrate (<2000 kg/h), the flow was nearly slug flow
and 20 Hz, respectably. Each parameter was logged over a as observed during the test, while the flow regime became
period of 100 s with a time averaged value generated under gradually dispersed bubbly flow as the flowrate and entrained
each experimental condition. Gas volume fraction α is defined gas increase. For the Coriolis flowmeter on the horizontal
and calculated as follows: pipeline, the range of mass flow errors is different from
qv,g
α= × 100% (13) that on the vertical pipeline most likely due to the effects
qv,l + qv,g of gravity and buoyancy on the flow regime. Positive errors
where qv,g and qv,l are the calculated volume flowrates of occur at the mass flowrates of 700 and 1000 kg/h when the
gas and liquid phases from the reference flow meters and the gas volume fraction below 6%. By comparing the mass flow
temperature and pressure in the upstream of the horizontal errors at the same flowrate in Figs. 8 and 9, the errors are
test meter. generally reproducible for the same installation and thanks to
Density drop is determined from the density of the liquid the new-generation flow transmitter [32]. For the test data set,
flow (ρl ) and the apparent density (ρ) from the Coriolis Test II includes some experimental data that were collected at
flowmeter under test different flowrates from those in Test I. The new conditions as
ρl − ρ in Test II that were conducted on a different day and obtained
d= × 100%. (14) under different flowrate from Test I are useful to assess the
ρl
models’ generalization capability and reproducibility.
Two series of experimental tests, Tests I and II, were conducted
Fig. 10 shows the distribution of the relative errors of the
for the liquid mass flow rate ranging from 700 to 14 500 kg/h
measured liquid mass flowrate on both the horizontal and
and gas volume fraction from 0% to 30%. The fluid temper-
vertical pipelines. Each color (blue or green) in the figure
ature during the tests was around 20 °C. For the purpose of
represents training or test data sets, respectively. The Coriolis
ANN training, 237 data sets were collected from Tests I, while
flowmeter on the horizontal pipeline yields the liquid mass
24 data sets recorded from Tests II for testing the performance
flowrate with a relative error between −41% and 9%, while
of the data driven models.
the meter on the vertical pipeline gives an error from −25%
to 11%. The difference in errors between the vertical and
B. Analysis of Original Errors horizontal installations is due to the fact that the bubbles in
The typical original mass flow errors of the Coriolis flowme- a vertical flow are distributed evenly in the pipe cross section
ters in Test I are shown in Fig. 8. The Coriolis flowmeter on due to the effect of gravity, resulting in less interruption on
the vertical section gives negative errors at flowrates below the tube vibration inside the Coriolis flowmeter and hence
4000 kg/h. At a higher flowrate (>5500 kg/h), the mass flow different errors.
WANG et al.: GAS–LIQUID TWO-PHASE FLOW MEASUREMENT USING CORIOLIS FLOWMETERS 857

Fig. 10. Relative error histogram of the measured liquid mass flowrate. (a) Horizontal pipeline. (b) Vertical pipeline.

C. Analysis of Parametric Dependence TABLE I

VARIABLE S ELECTION P ROCEDURES FOR M ODELS H-L AND V-L
There are three important parameters from a Coriolis
flowmeter, including observed density drop, apparent mass
flowrate, and damping. The DP value from the DP transducer
is also included as a potential input variable in this paper. The
apparent mass flowrate from a Coriolis flowmeter and the DP
value across the meter correlate strongly with the liquid mass
flowrate under two-phase conditions. In addition, when gas
entrains in the liquid flow, a rapid rise in damping occurs for
the fluid-conveying tube and the mixture density also deviates
TABLE II
from the liquid density. This physical background for the fluid-
VARIABLE S ELECTION P ROCEDURES FOR M ODELS H-G AND V-G
tube coupling system determines that these four input variables
are more important than other variables. There exist strong
nonlinearities between the outputs of a Coriolis flowmeter and
the flowrate being measured under two-phase flow conditions,
as observed by other researchers [12], [13]. Such nonlinearities
are also shown in Fig. 8.
In order to investigate the parametric dependence of indi-
vidual input parameters and the combined effect of multiple
parameters on the output of a data model, PMI is utilized
The coefficient of determination, R 2 , indicates the goodness of
to measure the partial dependence between a potential input
fit. A combination of the four variables gives the highest R 2 ,
variable and the output, conditional on any inputs that have
which illustrates that the combined effect of the variables is
already been selected. The variable with the highest PMI score
more significant than that of an individual variable on the
is added to the input set, if the Akaike information crite-
output. For predicting the gas volume fraction, x 1 (observed
rion (AIC) value decreases as a result from the inclusion of this
density drop), plays a more important part than other variables.
variable. The detailed definitions of PMI and AIC are available
Variable x 3 (damping) is not used in models H-G and V-G,
in [19] and [20]. Suppose variables x 1 , x 2 , x 3 , and x 4 represent
since the AIC value becomes increasing and R 2 is reducing
observed density drop, apparent mass flowrate, damping, and
with the inclusion of x 3 . As a result of these variable selec-
DP, respectively, the variable selection procedures for the
tion procedures, the models for correcting the liquid mass
models for correcting the liquid mass flowrate and predicting
flowrate accept the four input variables (observed density drop,
the gas volume fraction are summarized in Tables I and II.
apparent mass flowrate, damping, and DP) and three variables
H-L and V-L represent the models established for the hori-
(observed density drop, apparent mass flowrate, and DP) are
zontal and vertical pipelines, respectively, to correct the liquid
taken as the inputs to the models for predicting the gas volume
mass flowrate, while H-G and V-G stand for the models for
fraction.
the horizontal and vertical pipelines to predict the gas volume
fraction, respectively. The selection sequence also represents
the sensitivity level of each variable to the desired output. D. Performance of the BP-ANN
For the liquid mass flowrate, x 2 (apparent mass flowrate) The BP-ANN model is established through training with
has more significant effect on the liquid mass flowrate. data set I and tested with data set II. For each installation
858 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 66, NO. 5, MAY 2017

Fig. 11. Performance of BP-ANNs with differenct numbers of neurons in the hidden layer. (a) BP-ANN: H-L. (b) BP-ANN: V-L. (c) BP-ANN: H-G.
(d) BP-ANN: V-G.

condition, a separate model is established for the correction paper, normalized root-mean-square error (NRMSE) is used
of the measured liquid mass flowrate and the prediction of to assess the performance of a data driven model, which is
gas volume fraction. The inputs of the BP-ANN for liquid defined as
mass flowrate correction include four variables, i.e., observed

1
1 
m
density drop, apparent mass flowrate, damping, and DP. The
N RM S E =  (yi − ŷi )2 (17)
inputs of the BP-ANN for gas volume fraction prediction ȳ m
i=1
include observed density drop, apparent mass flowrate, and DP.
The number of neurons (L) in the hidden layer is determined where yi is the reference mass flow rate of the liquid phase or
using (15) and (16), as proposed in [33] gas volume fraction, ȳ is the mean of yi , ŷi is the corrected
mass flow rate or predicted gas volume fraction from the data
L ≤ 2n + 1 (15) driven model accordingly, and m is the number of samples
m
L ≤ (16) used.
n+1 As the weights and biases between the neurons are ini-
where n and m are the numbers of input variables and training tialized randomly, a different BP-ANN is obtained for each
samples, respectively. However, (15) and (16) give only the training, resulting in different performances. A preliminary
range of L for BP-ANN models. The exact L for a model study of averaging NRMSE of more than 200 BP-ANNs did
can be selected by a trial-and-error method to compromise not show any noticeable difference. Therefore, in order to
between minimizing errors and achieving good generalization minimize the effect of random initialization of an ANN, the
capability. The output layer has one neuron for each model, average NRMSE of 200 BP-ANNs with the same structure is
since there is only one output variable. calculated to assess the effect of the hidden neurons on the
The BP-ANN transfer function between the input and performance of the ANN.
hidden layers is hyperbolic tangent sigmoid transfer function. For the models for liquid mass flowrate correction, the
The pure linear function is taken as the transfer function number of neurons in the hidden layer is set from 4 to 9 as
connecting the hidden layer to the output layer. The training per (15) and (16). The NRMSE values of the BP-ANNs are
function is Bayesian regularization, while the learning function summarized in Fig. 11. The error bars indicate the maximum
is gradient descent with momentum weight and bias learning and minimum errors of 200 BP-ANNs for the same structure.
function. Training stops when the maximum number of epochs In view of the errors on both training and test datasets, the
is reached or the performance is minimized to the goal. In this BP-ANN with seven neurons in the hidden layer performs
WANG et al.: GAS–LIQUID TWO-PHASE FLOW MEASUREMENT USING CORIOLIS FLOWMETERS 859

Fig. 12. Errors of the corrected liquid mass flowrate from the trained Fig. 13. Error of the predicted gas volume fraction from the trained
BP-ANNs. (a) Errors of the corrected mass flowrate on the horizontal pipeline BP-ANNs. (a) Errors of the predicted gas volume fraction on the horizontal
with training data set. (b) Errors of the corrected mass flowrate on the pipeline with training data set. (b) Errors of the predicted gas volume fraction
horizontal pipeline with test data set. (c) Errors of the corrected mass flowrate on the horizontal pipeline with test data set. (c) Errors of the predicted gas
on the vertical pipeline with training data set. (d) Errors of the corrected mass volume fraction on the vertical pipeline with training data set. (d) Errors of
flowrate on the vertical pipeline with test data set. the predicted gas volume fraction on the vertical pipeline with test data set.

better than other structures under both the horizontal and liquid mass flowrate from the BP-ANNs. For the horizontal
vertical conditions. The BP-ANN used for gas volume fraction and vertical pipelines, the relative errors are mostly less than
prediction has lower NRMSE when the number of the hidden ±2% (the red dashed lines in Fig. 12) with the training
neurons is 6. data set except some larger errors at the low flowrates of
Once the structure of a BP-ANN is determined, the trained 700 and 1000 kg/h. This is very likely due to larger bubbles
neural network that has the minimum error with the test data or slugs appearing in the flow tubes under low flowrate,
set is selected. Fig. 12 shows the errors of the corrected which affects the Coriolis flowmeter behaving differently from
860 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 66, NO. 5, MAY 2017

Fig. 15. Errors of the predicted gas volume fraction from the RBF-ANNs.
Fig. 14. Errors of the corrected liquid mass flowrate from the RBF-ANNs. (a) Errors of the predicted gas volume fraction on the horizontal pipeline
(a) Errors of the corrected mass flowrate on the horizontal pipeline with with training data set. (b) Errors of the predicted gas volume fraction on
training data set. (b) Errors of the corrected mass flowrate on the horizontal the horizontal pipeline with test data set. (c) Errors of the predicted gas
pipeline with test data set. (c) Errors of the corrected mass flowrate on volume fraction on the vertical pipeline with training data set. (d) Errors of
the vertical pipeline with training data set. (d) Errors of the corrected mass the predicted gas volume fraction on the vertical pipeline with test data set.
flowrate on the vertical pipeline with test data set.
volume fraction is below 5%. As the entrained gas increases,
smaller bubbles. The trained BP-ANN has relatively larger the errors from the training data set are mostly within ±10%
errors at low flowrates and hence results in unsatisfactory (the red dashed lines in Fig. 13). For the test data set, however,
performance with the test data set under the same experimental all the errors are less than ±10% on the vertical pipeline, even
conditions. under the low flowrate conditions.
Since the gas volume fraction under the experimental con-
ditions ranges from 0% to 30% and the intrinsic complexity of E. Performance of the RBF-ANN
two-phase flow, the relative errors of the predicted gas volume Fig. 14 shows the relative errors of the corrected liquid
fraction from the BP-ANNs are quite large when the gas mass flowrate from the RBF-ANNs. In order to achieve more
WANG et al.: GAS–LIQUID TWO-PHASE FLOW MEASUREMENT USING CORIOLIS FLOWMETERS 861

Fig. 17. Errors of the predicted gas volume fraction from the SVMs.
(a) Errors of the predicted gas volume fraction on the horizontal pipeline
with training data set. (b) Errors of the predicted gas volume fraction on
Fig. 16. Errors of the corrected liquid mass flowrate error from the SVMs. the horizontal pipeline with test data set. (c) Errors of the predicted gas
(a) Errors of the corrected mass flowrate on the horizontal pipeline with volume fraction on the vertical pipeline with training data set. (d) Errors of
training data set. (b) Errors of the corrected mass flowrate on the horizontal the predicted gas volume fraction on the vertical pipeline with test data set.
pipeline with test data set. (c) Errors of the corrected mass flowrate on
the vertical pipeline with training data set. (d) Errors of the corrected mass
flowrate on the vertical pipeline with test data set.

difference in the original errors between the lower and higher

flowrates on the vertical pipeline, the RBF-ANN yields errors
accurate results with the test data set, the RBF-ANN on the between ±2% with the training data set and ±1% with the
horizontal pipeline disregards the errors at lower flowrates test data set.
(<2000 kg/h) and the network is trained to well fit higher As shown in Fig. 15, the RBF-ANN for gas volume fraction
flowrates (>4000 kg/h). Consequently, the errors at higher prediction outperforms significantly the BP-ANN, particularly
flowrates with the training data set, and the errors with the under the low entrained gas. When the gas volume fraction is
test data set are reduced to ±1%. Due to the insignificant below 5%, the maximum relative errors from RBF-ANNs on
862 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 66, NO. 5, MAY 2017

Fig. 18. Errors of the corrected liquid mass flowrate error from the GPs.
(a) Errors of the corrected mass flowrate on the horizontal pipeline with Fig. 19. Errors of the predicted gas volume fraction from the GPs. (a) Errors
training data set. (b) Errors of the corrected mass flowrate on the horizontal of the predicted gas volume fraction on the horizontal pipeline with training
pipeline with test data set. (c) Errors of the corrected mass flowrate on data set. (b) Errors of the predicted gas volume fraction on the horizontal
the vertical pipeline with training data set. (d) Errors of the corrected mass pipeline with test data set. (c) Errors of the predicted gas volume fraction
flowrate on the vertical pipeline with test data set. on the vertical pipeline with training data set. (d) Errors of the predicted gas
volume fraction on the vertical pipeline with test data set.

both the horizontal and vertical pipelines are around ±30%.

The rest errors with the training data set are well within ±10%.
F. Performance of the SVM
The relative errors from the test data set are almost less
than ±10%, except at the flowrate of 1000 kg/h on the SVM models are also established for both installation
horizontal pipeline. This is probably due to the fact that the conditions. An important difference between the SVM and
samples at 1000-kg/h flow rate are far away from the center ANN models is that the SVM leads to a unique deterministic
vectors in the network. model for each data set, while ANNs depend on a random
WANG et al.: GAS–LIQUID TWO-PHASE FLOW MEASUREMENT USING CORIOLIS FLOWMETERS 863

TABLE III
NRMSE OF SVM W ITH D IFFERENT K ERNEL F UNCTIONS

Fig. 20. Performance comparison between ANNs, SVMs, and GPs. (a) ANNs, SVMs, and GPs with training data set. (b) ANNs, SVMs, and GPs with
test data set.

initial choice of synaptic weights and cannot produce the fixed and terminal sets {x 1 , x 2 , x 3 , x 4 } for models H-L and V-L
results. Through a direct comparison of the performances of and {x 1 , x 2 , x 4 } for models H-G and V-G.
SVM between the four kinds of kernel function (Table III), we The GP-based formulations for the four models are given
know that the SVM with RBF generates the smallest NRMSE in the following:
among the four models.
From Fig. 16(a) and (c), the SVM model performs well y H −L = 0.994x 2 − 2633x 1 + 4300 tanh(x 1 ) tanh(x 3 )
to fit with training data and limit the relative errors on the + 13.2x 1 x 4 + 0.00571x 2 x 3 − 0.0995x 2 x 3
horizontal and vertical pipelines to ±1% or less, except some × tanh(x 1 ) + 62.4 (18)
points at 700 and 1000 kg/h, which is a common problem for yV −L = x 2 + 57.6x 3 − 0.161x 4 + 29.8x 1 x 4
the ANN and SVM models. The generalization ability of the  
+ 871 tanh x 32 x 4 tanh(x 1 ) − 0.00913x 4(x 3 + x 1 x 4 )
SVM model is proven, as shown in Fig. 16(b) and (d). Most
errors from the SVM models with the test data are reduced − 0.122x 1 x 2 x 3 + 32.5 (19)
−6
to ±1%. y H −G = 0.783x 1 + 1.6e x 2 + 0.00278x 4 − 0.114x 1 x 4
Fig. 17 shows that for gas volume fraction prediction, a less + 0.159x 12 x 4 + 6.82e−5 x 43 − 0.0182 (20)
−7 −7
number of points from the SVM models have an error beyond yV −G = 1.01x 1 − 5.49e x 2 − 0.0217x 4 − 2.74e tanh(x 1 )
±10% with the training data set. Since the kernel function − 1.05e−4 x 1 x 2 + 2.74e−6 x 2 x 4 + 0.00253x 1 x 42
used in the SVM models is RBF, the performance of the SVM
− 1.05e−4 x 12 x 4 − 2.74e−7 x 1 x 2 x 4 + 0.00587. (21)
models has the common problem with the RBF-ANN. The
relative errors in the predicted gas volume fraction with the The errors of the corrected mass flowrate on the train-
test data set at the flowrate of 1000 kg/h are larger than other ing data set using GP are higher by −15% and 25%,
test data. respectively, under the horizontal and vertical installations
[Fig. 18(a) and (c)], which results in larger errors on the
G. Performance of the GP
test data set [Fig. 18(b) and (d)]. As can be seen that, larger
Four GP models are established in this paper for correcting errors normally occur at low flowrates, which indicate that the
the liquid mass flowrate and predicting the gas volume frac- GP models are unable to approximate all the data.
tion, respectively, for the horizontal and vertical installations As shown in Fig. 19, for the prediction of gas volume
of Coriolis flowmeters. The parameters that were set in the GP fraction, the outputs of GP models have large errors for low
algorithms include: a population size of 250, a tournament size gas entrainment and low flowrates. The relative errors with
of 25, an elitism of 0.7, maximum number of genes allowed test data reach 25% and −50% on the horizontal and vertical
in an individual 6, function set {×, −, +, tanh, mult3, add3}, pipes, respectively.
864 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 66, NO. 5, MAY 2017

Fig. 21. Relative error histogram of ANNs, SVMs, and GPs for corrected liquid mass flowrate. (a) BP-ANN: H-L. (b) BP-ANN: V-L. (c) RBF-ANN: H-L.
(d) RBF-ANN: V-L. (e) SVM: H-L. (f) SVM: V-L. (g) GP: H-L. (h) GP: V-L.

H. Perforamce Comparison Between BP-ANN, gas volume fraction prediction, GP produces larger errors than
RBF-ANN, SVM, and GP the other three techniques. Both BP-ANN and RBF-ANN have
1) Robustness: In order to assess the robustness of the four a similar mean NRMSE with the training data set, while the
kinds of models, the averaged NRMSE values are shown in SVM models yield less error. With the test data set, BP-ANN,
Fig. 20. The models for liquid mass flowrate correction and RBF-ANN, and SVM methods perform similarly on Models
WANG et al.: GAS–LIQUID TWO-PHASE FLOW MEASUREMENT USING CORIOLIS FLOWMETERS 865

Fig. 22. Relative error histogram of ANNs, SVMs, and GPs for gas volume fraction prediction. (a) BP-ANN: H-G. (b) BP-ANN: V-G. (c) RBF-ANN: H-G.
(d) RBF-ANN: V-G. (e) SVM: H-G. (f) SVM: V-G. (g) GP: H-G. (h) GP: V-G.

H-L and V-L. However, the SVM models are significantly all the time. This outcome suggests that the SVM models are
better than the BP-ANN, RBF-ANN, and GP models for the superior to both ANN and GP models in terms of robustness.
prediction of gas volume fraction. Moreover, BP-ANN and 2) Accuracy: Fig. 21 shows the relative error histograms of
RBF-ANN have uncertain parameters to optimize which could the ANNs, SVMs, and GPs for corrected liquid mass flowrate.
result in differences in performance. However, due to their It is clear that the error distributions of the GP and ANN
fixed structure, the SVM models produce repeatable results models are much wider and dispersive than the SVM models.
866 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 66, NO. 5, MAY 2017

TABLE IV
M EAN AND S TANDARD D EVIATION OF THE R ELATIVE E RROR D ISTRIBUTION FOR L IQUID M ASS F LOWRATE C ORRECTION

TABLE V
M EAN AND S TANDARD D EVIATION OF THE R ELATIVE E RROR D ISTRIBUTION FOR G AS V OLUME F RACTION P REDICTION

TABLE VI
A CCURACY C OMPARISONS OF ANN, SVM, AND GP M ODELS

Through comparing the mean value and standard deviation of error less than ±1% on the horizontal pipeline, while 96.17%
the errors between the eight error distributions (Table IV), we of the results are within ±1% on the vertical installation.
can see that the SVM models with the lowest mean value and The SVM models predict the gas volume fraction with a
standard deviation outperform the BP-ANN, RBF-ANN, and relative error less than 10% for 93.10% and 94.25% of the
GP models for liquid mass flowrate measurement on both the test conditions on the horizontal and vertical installations,
horizontal and vertical pipelines. Moreover, the data driven respectively. Therefore, the SVM models perform significantly
models (a mean value of 0.0008% and a standard deviation better than the BP-ANN, RBF-ANN, and GP models for two-
of 0.40%) on the vertical pipeline perform better than those phase flow measurement in terms of robustness and accuracy.
on the horizontal pipeline (a mean value of 0.0585% and a
standard deviation of 0.66%).
IV. C ONCLUSION
Fig. 22 shows the relative error histograms of the four
types of models for gas volume fraction prediction. GP models In this paper, experimental and analytical investigations
have a larger range of errors than all other models. The error have been carried out to assess the performance of BP-ANN,
distribution of the SVM model is much narrower than the ANN RBF-ANN, SVM, and GP for gas–liquid two-phase flow mea-
models for the measurement of gas volume fraction. It can be surement using Coriolis flowmeters. The results presented have
seen that most errors of the SVM models are concentrated suggested that the SVM models are superior to the two ANN
around zero line. Table V shows that the standard deviations models and the GP models for two-phase flow measurement in
of the SVM and RBF-ANN models are smaller than that of the terms of robustness and accuracy. The SVM models perform
BP-ANN and GP models on both the horizontal and vertical well consistently, while the performance of ANN and GP
pipelines. models depends on the user-defined parameters. For liquid
In order to assess the accuracy of the ANN, SVM, and GP mass flowrate measurement, the SVM models outperform
models, the percentage of experimental data for each model the BP-ANN, RBF-ANN, and GP on both the horizontal
that can achieve the accuracy of ±2% and ±1%, respectively, and vertical pipelines and the most corrected errors (>93%)
for liquid mass flowrate measurement and ±10% for gas are within ±1%. For the gas volume fraction prediction,
volume fraction prediction is calculated and summarized in the RBF-ANN and SVM models yield most relative errors
Table VI. For liquid mass flowrate measurement with the (>90%) less than ±10% and outperform the BP-ANN and
SVM models, 93.49% of the experimental data yield a relative GP. It must be stressed that the significantly reduced errors in
WANG et al.: GAS–LIQUID TWO-PHASE FLOW MEASUREMENT USING CORIOLIS FLOWMETERS 867

mass flowrate measurement from the Coriolis mass flowmeters [17] L. Xing et al., “A combination method for metering gas–liquid two-
and gas volume fraction prediction are achieved by using the phase flows of low liquid loading applying ultrasonic and coriolis
flowmeters,” Flow Meas. Instrum., vol. 37, pp. 135–143, Jan. 2014.
existing data from the Coriolis flowmeters and a simple DP [18] L. Ma, H. Zhang, H. Zhou, and Q. He, “Mass flow measurement of oil-
transducer without the use of any other devices. SVM has water two-phase flow based on coriolis flow meter and SVM,” J. Chem.
consistently outperformed ANN and GP in the correction of Eng. Chin. Univ., vol. 21, no. 2, pp. 201–205, Apr. 2007.
liquid mass flow errors and prediction of gas volume fraction. [19] R. May, H. Maier, G. C. Dandy, and T. Fernando, “Non-linear variable
selection for artificial neural networks using partial mutual information,”
This outcome has effectively extended the applicability of Environ. Model. Softw., vol. 23, nos. 10–11, pp. 1312–1326, 2008.
Coriolis mass flowmeters to liquid flow measurement with [20] S. Galelli, G. B. Humphrey, H. R. Maier, A. Castelletti, G. C. Dandy,
a significant volume of entrained gas. In the future work, and M. S. Gibbs, “An evaluation framework for input variable selection
algorithms for environmental data-driven models,” Environ. Model.
the data driven models will be extended for the measurement Softw., vol. 62, pp. 33–51, Dec. 2014.
of other liquids with different viscosities under two-phase or [21] L. Wang, J. Liu, Y. Yan, X. Wang, and T. Wang, “Gas-liquid two-
multiphase flow conditions. phase flow measurement using coriolis flowmeters incorporating neural
networks,” in Proc. IEEE Int. Instrum. Meas. Technol. Conf., Taipei,
Taiwan, May 2016, pp. 747–751.
R EFERENCES
[22] H. A. Farfani, F. Behnamfar, and A. Fathollahi, “Dynamic analysis
[1] Y. Li, J. Wang, and Y. Geng, “Study on wet gas online flow rate of soil-structure interaction using the neural networks and the support
measurement based on dual slotted orifice plate,” Flow Meas. Instrum., vector machines,” Expert Syst. Appl., vol. 42, no. 22, pp. 8971–8981,
vol. 20, nos. 4–5, pp. 168–173, Aug. 2009. Dec. 2015.
[2] J. L. G. Oliveira, J. C. Passos, R. Verschaeren, and C. van der Geld, [23] S. Belaid and A. Mellit, “Prediction of daily and mean monthly global
“Mass flow rate measurements in gas–liquid flows by means of a Venturi solar radiation using support vector machine in an arid climate,” Energ.
or orifice plate coupled to a void fraction sensor,” Experim. Thermal Convers. Manage., vol. 118, pp. 105–118, Jun. 2016.
Fluid Sci., vol. 33, no. 2, pp. 253–260, Jan. 2009. [24] M. A. M. Ramli, S. Twaha, and Y. A. Al-Turki, “Investigating the
[3] Z. Sun, “Mass flow measurement of gas–liquid bubble flow with the performance of support vector machine and artificial neural networks in
combined use of a Venturi tube and a vortex flowmeter,” Meas. Sci. predicting solar radiation on a tilted surface: Saudi Arabia case study,”
Technol., vol. 21, no. 5, pp. 1–7, Mar. 2010. Energ. Convers. Manage., vol. 105, pp. 442–452, Nov. 2015.
[4] Y. Xu, H. Wang, Z. Cui, F. Dong, and Y. Yan, “Separation of gas–liquid [25] F. Chen, H. Li, Z. Xu, S. Hou, and D. Yang, “User-friendly optimization
two-phase flow through independent component analysis,” IEEE Trans. approach of fed-batch fermentation conditions for the production of
Instrum. Meas., vol. 59, no. 5, pp. 1294–1302, May 2010. iturin a using artificial neural networks and support vector machine,”
[5] L. Wang, Z. Yao, B. Wang, H. Ji, and H. Li, “Flow pattern identification Electron. J. Biotechnol., vol. 18, no. 4, pp. 273–280, Jul. 2015.
of gas–liquid two-phase flow based on capacitively coupled contactless [26] A. H. Gandomi and A. H. Alavi, “A new multi-gene genetic program-
conductivity detection,” IEEE Trans. Instrum. Meas., vol. 61, no. 5, ming approach to nonlinear system modeling. Part I, materials and
pp. 1466–1475, May 2012. structural engineering problems,” Neural Comput. Appl., vol. 21, no. 1,
[6] F. Dong, Y. Xu, L. Hua, and H. Wang, “Two methods for measurement of pp. 171–187, Feb. 2012.
gas-liquid flows in vertical upward pipe using dual-plane ERT system,” [27] C. Cortes and V. Vapnik, “Support-vector networks,” Mach. Learn.,
IEEE Trans. Instrum. Meas., vol. 55, no. 5, pp. 1576–1586, Oct. 2006. vol. 20, no. 3, pp. 273–297, 1995.
[7] L. Xu, W. Zhou, X. Li, and S. Tang, “Wet gas metering using a [28] H. Drucker, C. Burges, L. Kaufman, A. Smola, and V. Vapnik, “Sup-
revised Venturi meter and soft-computing approximation techniques,” port vector regression machines,” Neural Inf. Process. Syst., vol. 9,
IEEE Trans. Instrum. Meas., vol. 60, no. 3, pp. 947–956, Mar. 2011. pp. 155–161, Apr. 1997.
[8] X. Li, X. Huang, B. Wang, and H. Li, “A new method for the online [29] J. Koza, Genetic Programming: On the Programming of Computers by
voidage measurement of the gas–oil two-phase flow,” IEEE Trans. Means of Natural Evolution. Cambridge, MA, USA: MIT Press, 1992.
Instrum. Meas., vol. 58, no. 5, pp. 1571–1577, May 2009. [30] J. Madár, J. Abonyi, and F. Szeifert, “Genetic programming for the
[9] H. Ji, J. Long, Y. Fu, Z. Huang, B. Wang, and H. Li, “Flow pattern identification of nonlinear input-output models,” Ind. Eng. Chem. Res.,
identification based on EMD and LS-SVM for gas–liquid two-phase vol. 44, no. 9, pp. 3178–3186, Mar. 2005.
flow in a minichannel,” IEEE Trans. Instrum. Meas., vol. 60, no. 5, [31] D. P. Searson, “GPTIPS 2: An open-source software platform for sym-
pp. 1917–1924, May 2011. bolic data mining,” in Handbook of Genetic Programming Applications.
[10] M. G. De Giorgi, A. Ficarella, and A. Lay-Ekuakille, “Cavitation New York, NY, USA: Springer, 2015, ch. 22.
regime detection by LS-SVM and ANN with wavelet decomposition [32] J. W. Kunze, R. Storm, and T. Wang, “Coriolis mass flow measurement
based on pressure sensor signals,” IEEE Sensors J., vol. 15, no. 10, with entrained gas,” in Proc. ITG/GMA Symp. Sensors Meas. Syst.,
pp. 5701–5708, Oct. 2015. Jun. 2014, pp. 1–6.
[11] T. Wang and R. Baker, “Coriolis flowmeters: A review of developments [33] G. J. Bowden, H. R. Maier, and G. C. Dandy, “Input determination
over the past 20 years, and an assessment of the state of the art and likely for neural network models in water resources applications. Part 2. Case
future directions,” Flow Meas. Instrum., vol. 40, pp. 99–123, Sep. 2014. study: Forecasting salinity in a river,” J. Hydrol., vol. 301, nos. 1–4,
[12] J. Hemp and G. Sultan, “On the theory and performance of pp. 93–107, Jan. 2005.
coriolis mass flowmeters,” in Proc. Int. Conf. Mass Flow Meas.,
1989, pp. 1–38.
[13] R. P. Liu, M. J. Fuent, M. P. Henry, and M. D. Duta, “A neural
network to correct mass flow errors caused by two-phase flow in a digital
coriolis mass flowmeter,” Flow Meas. Instrum., vol. 12, no. 1, pp. 53–63,
Mar. 2001.
[14] B. Safarinejadian, M. A. Tajeddini, and L. Mahmoodi, “A new fuzzy
based method for error correction of coriolis mass flow meter in presence Lijuan Wang (S’13) received the B.Eng. degree
of two-phase fluid,” in Proc. Int. Conf. Artif. Intell. Image Process., 2012, in computer science and technology from Qiqihar
pp. 192–196. University, Qiqihar, China, in 2010, and the Ph.D.
[15] V. A. Lari and F. Shabaninia, “Error correction of a coriolis mass degree in measurement and automation from North
flow meter in two-phase flow measurment using neuro-fuzzy,” in China Electric Power University, Beijing, China,
Proc. 16th CSI Int. Symp. Artif. Intell. Signal Process., May 2012, in 2015. She is currently pursuing the Ph.D. degree
pp. 611–616. in instrumentation and measurement with the Uni-
[16] Q. L. Hou, K. J. Xu, M. Fang, Y. Shi, B. B. Tao, and R. versity of Kent, Canterbury, U.K.
W. Jiang, “Gas–liquid two-phase flow correction method for digital Her current research interests include electrosta-
CMF,” IEEE Trans. Instrum. Meas., vol. 63, no. 10, pp. 2396–2404, tic sensing, flow measurement, and digital signal
Mar. 2014. processing.
868 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 66, NO. 5, MAY 2017

Jinyu Liu received the B.Eng. degrees in electrical Xue Wang received the B.Sc. degree in mathe-
engineering and automation from Tianjin Univer- matical statistics from Northwest University, Xi’an,
sity, Tianjin, China, and in electronic and computer China, in 1996, the M.Sc. degree in statistics
systems from the University of Kent, Canterbury, from Northwestern Polytechnic University, Xi’an,
U.K., in 2012, and the M.Sc. degree in energy in 1999, and the Ph.D. degree in statistics from
and sustainability with electrical power engineering The University of Nottingham, Nottingham, U.K.,
from the University of Southampton, Southampton, in 2006.
U.K., in 2013. He is currently pursuing the Ph.D. She was a Post-Doctoral Research Assistant with
degree in instrumentation and measurement with the the Statistics Group, University of Bristol, Bris-
University of Kent. tol, U.K., from 2005 to 2006. She is currently a
His current research interests include flow rig Lecturer with University of Kent, Canterbury, U.K.
construction, flow measurement, and digital signal processing. Her current research interests include bayesian nonparametrics, wavelet and
multiscale methods, and statistical signal processing.

Yong Yan (M’04–SM’04–F’11) received the B.Eng.

and M.Sc. degrees in instrumentation and con-
trol engineering from Tsinghua University, Beijing,
China, in 1985 and 1988, respectively, and the Ph.D.
degree in flow measurement and instrumentation Tao Wang received the B.Eng. and M.Sc. degrees
from the University of Teesside, Middlesbrough, in aerospace manufacturing engineering from
U.K., in 1992. Northwestern Polytechnic University, Xi’an, China,
He was an Assistant Lecturer with Tsinghua Uni- in 1996 and 1999, respectively, and the Ph.D. degree
versity in 1988. In 1989, he joined as a Research in engineering from the University of Cambridge,
Assistant with the University of Teesside. He was a Cambridge, U.K., in 2003.
Lecturer with the University of Teesside from 1993 He was a Research Associate with the University
to 1996, and a Senior Lecturer, a Reader, and a Professor with the University of Cambridge from 2002 to 2004 and a Lecturer
of Greenwich, Chatham Maritime, U.K., from 1996 to 2004. He is currently a with Brunel University, London, U.K., from 2007
Professor of Electronic Instrumentation, the Head of Instrumentation, Control to 2008. He is currently the Research and Metrol-
and Embedded Systems Research Group, and the Director of Research with ogy Manager with KROHNE Ltd., Wellingborough,
the School of Engineering and Digital Arts, University of Kent, Canterbury, U.K., and a Senior Visiting Research Fellow with the School of Engineering
U.K. His current research interests include multiphase flow measurement, and Digital Arts, University of Kent, Canterbury, U.K. His current research
combustion instrumentation, and monitoring and optimization of industrial interests include flow measurement and instrumentation, Coriolis flowmeter
processes. technology, numerical simulation, and manufacturing engineering.