You are on page 1of 9

2396 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO.

10, OCTOBER 2014

Gas–Liquid Two-Phase Flow Correction


Method for Digital CMF
Qi-Li Hou, Ke-Jun Xu, Min Fang, Yan Shi, Bo-Bo Tao, and Rong-Wei Jiang

Abstract— There are large measurement errors in Coriolis compressed air may be used to purge the line between batches.
mass flowmeter (CMF) when it is used to measure gas–liquid This may exacerbate gas–liquid condition at the beginning and
two-phase flow without compensation. In particular, the flowtube end of the batch and air entrainment is the most commonly
of CMF may stall when using analog transmitters and CMF
cannot operate normally because of the high damping caused by occurring case of gas–liquid flow. When gas–liquid two-phase
gas–liquid two-phase. To solve above problems, a gas–liquid two- flow condition occurs, the density difference of the gas and
phase flow rig is developed to perform the two-phase flow experi- liquid causes the relative movement between the gas particle
ments. A digital drive method using multiplying digital-to-analog and liquid. The friction between the gas particle and the
converter and direct digital synthesizer is applied to CMF to fluid consumes energy so the power of the tube driver rises
maintain the flowtube oscillation under two-phase flow condition.
A digital zero-crossing detection method based on Lagrange to compensate the energy loss. In addition, because of the
interpolation is adopted to calculate the frequency and phase existence of relative movement, not all the particles in the
difference of the sensor output signals in order to synthesize the tube are laterally moved with the tube amplitude and the mass
digital drive signal and realize the flow measurement. The artifi- flow reading is less [7]. Experimental data shows that even
cial neural network using back propagation algorithm is applied small amount of entrained air can cause significant errors [1].
to correct measurement errors. A digital Coriolis mass flow
transmitter is developed with a digital signal processor to control What’s worse, relatively large amount of entrained air can
the digital drive, and realize the two-phase flow measurement induce very high and rapidly fluctuating damping. If the driver
and correction. By connecting the developed transmitter with cannot provide enough power in time, then the tube stalls and
a CMF025 type primary instrument, the gas–liquid two-phase no measurement can be provided. Therefore, the measurement
flow experiments are conducted to validate the performance of of gas–liquid two-phase flow is still difficult for CMF.
the transmitter. The testing results show that the transmitter can
maintain the flowtube oscillation and reduce the metering errors. To maintain the flowtube oscillation under two-phase flow
condition, many digital drive methods have been studied
Index Terms— Artificial neural network, Coriolis mass
flowmeter (CMF), digital drive method, digital zero-crossing and applied to CMF. Maginnis [8] introduced an initial-
detection method, gas–liquid two-phase flow. ization algorithm for drive control of CMF. Shimada [9]
combined the analog drive method with the digital wave
I. I NTRODUCTION
synthesis drive method based on phase-locked loop. Henry

C ORIOLIS mass flowmeter (CMF) has been widely used


in industry applications because it can measure mass flow
directly and obtain the density simultaneously [1]. Researchers
and Zamora [10] studied many drive methods of CMF, like
random sequence mode, positive feedback mode, and digital
synthesis mode. The digital synthesis mode was realized by
have studied many digital signal processing methods and field programmable gate array (FPGA) in [11]. Hou et al.
developed digital transmitters, which have been applied to the [12] studied the drive method using multiplying digital-to-
measurement of single-phase flow successfully in [2]–[6]. analog converter (MDAC) and direct digital synthesizer (DDS)
However, in industrial applications, gas–liquid two-phase chip and developed a digital transmitter to realize the drive
flow condition occurs inevitably. For example, during oil control and flow measurement. These digital drive meth-
transportation, the gas–liquid condition happens as the fluid ods provide the premise for the measurement of two-phase
contains natural gas. During food processing, high-pressure flow.
Manuscript received October 8, 2013; revised January 17, 2014; accepted To reduce the metering errors of two-phase flow, additional
January 24, 2014. Date of publication March 28, 2014; date of current correcting methods are needed. Hemp and Yeung [13] pro-
version September 11, 2014. This work was supported in part by the posed a bubble model to correct the measurement errors. This
National High Technology Research and Development Program under Con-
tract 2009AA04Z128 and in part by the China Aviation Industry Cooperative model predicts monotonic, negative errors which are a function
Innovation Program under Contract CXY2011HFGD23. The Associate Editor only of the gas void fraction (GVF). But Liu et al. [14] found
coordinating the review process was Dr. Niclas Bjorsell. (Corresponding that the actual response observed from a commercial CMF was
author: K.-J. Xu.)
Q.-L. Hou, M. Fang, Y. Shi, B.-B. Tao, and R.-W. Jiang are with the School very different from this model. The experimental results of [7]
of Electrical and Automation Engineering, Hefei University of Technology, and [14]–[16] show that the measurement errors of two-phase
Hefei 230009, China. flow are affected not only by GVF, but also by fluid velocity,
K.-J. Xu is with the School of Electrical and Automation Engineering, Hefei
University of Technology, Hefei 230009, China, and also with the Engineering flowtube orientation, and geometry. Their experimental results
Technology Research Center of Industrial Automation, Hefei 230009, China of metering errors turn out to be complicated nonlinearity
(e-mail: dsplab@hfut.edu.cn). and nonmonotonic. Therefore, Henry et al. [14]–[16] used
Color versions of one or more of the figures in this paper are available
online at http://ieeexplore.ieee.org. multilayer perceptron and the radial basis function (RBF)
Digital Object Identifier 10.1109/TIM.2014.2310031 network to correct the metering errors.
0018-9456 © 2014 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
HOU et al.: GAS–LIQUID TWO-PHASE FLOW CORRECTION METHOD FOR DIGITAL CMF 2397

Fig. 1. Schematic of two-phase flow experimental rig. Fig. 2. Picture of gas–liquid two-phase flow experimental rig.

Gysling [17] proposed an aeroelastic model to address the valves. Two CMF025 type primary instruments made by Micro
effects of compressibility and inhomogeneity introduced by Motion are installed on the pipeline. The upstream primary
aerated fluids. However, this model is under the assumption instrument is connected with Model 2700 transmitter made by
that bubbles are well mixed within the liquid continuous Micro Motion to form a reference flowmeter. The downstream
phase. As a result, this model cannot address slugging and/or primary instrument is connected with our transmitter and is the
stratified flows within Coriolis phenomena that can often occur flowmeter under test. A picture of this rig is shown in Fig. 2.
in start-up and batching processes, but rather it is intended to When performing the experiments, the water is pumped
address a class of application in which the amounts of gas is from the tank, flows through the pipelines and the two primary
relatively small (GVF is smaller than 10%). In addition, there instruments, then returns to the tank. The air flow is generated
is no correcting result, so the correcting accuracy is unknown. by the air compressor and added to pipeline through the valve.
Lari and Shabaninia [18] proposed using Neuro-Fuzzy for The direction of the arrows in Fig.1 shows the direction of
the error correction of two-phase flow measurement. However, water and gas flow. Note that the air flow is added to the
the two-phase flow experiments were not performed and the pipeline that between the upstream primary instrument and
actual correcting accuracy is unknown. the downstream primary instrument, so that the upstream CMF
In this paper, the neural network using back propagation (the reference CMF) measures the single-phase (water) flow,
(BP) algorithm is utilized to correct the measurement errors while the downstream CMF (the CMF under test) measures
caused by gas–liquid two-phase flow. A gas–liquid two-phase the gas–liquid two-phase flow. By comparing the mass flow
flow experimental rig is designed and developed to perform readings of the two CMFs, the measurement errors of the
two-phase flow experiments. A digital Coriolis flow transmitter CMF under test can be obtained. The air flow added to
with digital drive and digital signal processing technology the downstream primary instrument is measured by the gas
is also developed to realize the drive control as well as reference rotameter while the true water flowrate is given by
flow measurement and correction. Then by connecting our the water reference transmitter.
transmitter with CMF025 type primary instrument, gas–liquid The drive control and signal processing method used in
two-phase flow experiments are conducted on the developed the transmitter under test will be discussed in the following
flow rig to validate the digital drive, flow measurement, and section.
correction techniques.
III. D IGITAL CMF T RANSMITTER
II. G AS –L IQUID T WO -P HASE F LOW E XPERIMENTAL R IG
The digital drive and signal processing methods are studied
To correct the measurement errors of two-phase flow, the for the measurement of gas–liquid two-phase flow. The digital
errors should firstly be modeled. The complicated nonlinearity drive method based on MDAC and DDS is used to maintain
and nonmonotonic behavior of the measurement errors makes the flowtube oscillation, and the digital zero-crossing detection
it difficult to be modeled through mechanism modeling meth- method based on Lagrange interpolation is adopted to calculate
ods but only can be modeled through experimental modeling the frequency and phase difference of the sensor output signals
way, that is to find out the relationship between system to synthesize the digital drive signal and obtain the mass
input and output based on experimental data. To obtain the flowrate [12].
two-phase flow experimental data, a gas–liquid two-phase flow
experimental rig is designed to perform the experiments. The
schematic of the two-phase experimental rig is shown in Fig. 1. A. Digital Drive Method Based on MDAC and DDS
The experimental rig is mainly consisted with a tank, The block diagram of the digital drive method is shown
a pump, an air compressor, pipelines, Coriolis primary instru- in Fig. 3. The frequency and phase of the drive signal is
ments, CMF transmitters, piezometers, gas rotameters, and controlled by DSP through the DDS chip, and the amplitude
2398 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO. 10, OCTOBER 2014

Fig. 3. Block diagram of digital drive method based on MDAC and DDS.

is controlled through the MDAC chip. Comparing with the


FPGA-based drive methods described in [10] and [11], this
MDAC and DDS-based drive method can synthesize the sine
wave with required frequency and phase information automati-
cally. Therefore, the large amount of computation to synthesize
the sine waves can be reduced.
The DSP collects the sensor signal and calculates its ampli-
tude, frequency, and phase continuously. Then DSP determines Fig. 4. Schematic diagram of zero-crossing detection.
the frequency, phase, and amplitude of the synthesized drive
waveform, which is required to match with the sensor sig-
nal. The initial phase value of DDS output signal must be because the core of our processor is DSP rather than FPGA
determined to match the phase relationship between the drive combined with microprocessor. In addition, when we applied
signal and sensor signal. Before updating the drive signal, this signal processing method combined with digital drive
the zero-crossing points of sensor signal are calculated and method to CMF using DSP chip, it was found that the algo-
the time delays of hardware (analog circuitry and analog-to- rithm cannot be realized in real time. Instead, a digital zero-
digital converter (ADC)) and software (digital filter) must be crossing detection method based on Lagrange interpolation
compensated so as to obtain the actual-phase value of the is used to calculate the frequency and phase difference. The
sensor output signal. Then according to the phase relationship water flow calibration results we have performed show that
between the drive signal and sensor signal and the delays this algorithm possesses the accuracy of 0.1% on mass flow
caused by digital-to-analog converter (DAC) and drive cir- measurement and compared with the Fourier-based processing
cuitry are also considered, the initial phase of the drive signal method, the advantages of this approach are smaller amount
is calculated. of calculation involved and good tracking speed because the
A nonlinear amplitude control method is used to calculate signal parameters can be obtained in one signal cycle and
the drive gain [19]. Suppose the calculated amplitude of the measurements can be generated twice per cycle.
sensor signal is Amp(t). A desired amplitude value Ampd is A schematic diagram of the zero-crossing detection is shown
predetermined and the error AmpErr(t) is given as in Fig. 4. As shown in Fig. 4, given the zero-crossing times
AmpErr(t) = ln(Ampd ) − ln(Amp(t)). (1) t1L , t2L , t3L , t1R , t2R , t3R , the signal frequency is
 
The error AmpErr(t) is set into a proportional–integral (PI) f = 1/ t3R − t1R . (3)
control block to calculate the gain of drive signal

The time difference and phase difference of two channels are,
K (t) = K p AmpErr(t) + K i AmpErr(t)dt (2) respectively
where K p is the proportional coefficient and K i is the integral  
t = t2R − t2L (4)
coefficient.
The DSP writes the new defined frequency and phase into φ = 360 f t. (5)
the DDS, and amplitude gain into the MDAC through the
digital interface. Accordingly, the DDS outputs the sinusoid In practice, DSP processes the discrete signals that are
signal with specific frequency and phase to the MDAC. This sampled and converted by ADC. Due to signal discretization,
signal is then multiplied by the amplitude gain from the digital the possibility of sampling at zero-crossing points directly is
port. Thus, a signal containing the required frequency, phase, very small. Consequently, it is necessary to perform curve
and amplitude information is generated by the MDAC. At last fitting for the sampled data so as to obtain the zero-crossing
stage, this signal is amplified to drive the flowtubes. points of the signal. Here, Lagrange quadratic interpolation
method is employed to fit the curve.
Given the sampling sequence of the sensor signal x(n), if
B. Digital Zero-Crossing Detection Method x(n − 1)x(n) < 0, then there is a zero-crossing time point
In [3], DFT was used to process the sensor output signals. between (n −1) and n. The Lagrange interpolating polynomial
We also studied the Fourier-based technique, and it is really is supposed as
very precise. However, the problem is that the amount of
the calculation of Fourier-based method may be some large x = at 2 + bt + c (6)
HOU et al.: GAS–LIQUID TWO-PHASE FLOW CORRECTION METHOD FOR DIGITAL CMF 2399

Fig. 6. Block diagram of system software.

Fig. 5. Block diagram of system hardware.

where t is time, and


a = 0.5x(n − 2) − x(n − 1) + 0.5x(n) (7)
b = − 0.5x(n − 2)(2n − 1) + x(n − 1)(2n − 2)
− 0.5x(n)(2n − 3) (8)
c = 0.5x(n − 2)(n − 1)n − x(n − 1)(n − 2)n
+ 0.5x(n)(n − 2)(n − 1). (9)
The zero-crossing points can be obtained by solving above
equations. Note that the roots out of the range of [n − 1, n]
must be discarded.

C. Digital Transmitter Based on DSP


It is necessary to select a kind of microprocessor with
high-speed processing ability and abundant peripherals
for implementing various functions of CMF transmitter,
such as calculation, control, and interfaces. Therefore, the
TMS320F28335, a DSP chip made by TI Company, which has
the processing speed of 150 MIPS and abundant peripherals, Fig. 7. Flowchart of system software.
like multichannel buffered serial port (McBSP), serial periph-
eral interface (SPI), serial communication interface (SCI),
external interface (XINTF), pulse-width modulation (PWM), the sensor signals. Next, DSP begins to call the algorithm to
and so on, is chosen to develop a CMF transmitter. calculate the mass flowrate and control the flowtube oscillation
A block diagram of the transmitter hardware is shown continuously. During this procedure, DSP also controls the
in Fig. 5. This is a system configured differently from the LCD display, keyboard input, and SCI communication.
FPGA-based transmitter architecture in [11], and it has only The repeatability experiments mentioned in [3] are per-
one DSP chip used to realize the signal sampling, sig- formed and the calculate results of frequency, phase difference,
nal processing, drive control, and human–machine interface. and amplitude are saved per second and the experiments last
Furthermore, this transmitter consists of a digital drive system, 75 s (about 10 000 signal cycles) to obtain the standard devi-
analog signal amplifying and filtering circuits, three 24-bit ations. The repeatability results, with frequency varies from
ADCs, an external static random access memory, a ferro- 135 to 156 Hz, phase difference varies from 0.0052° to 2.6°,
electric nonvolatile random access memory (FRAM), a liquid amplitude 4.3 V, are shown in Table I.
crystal display (LCD), a keyboard, a SCI interface, a low- Fig. 8 shows the sensor signal under gas–liquid two-phase
voltage supervisor, a DAC, a voltage-to-current circuit, and a flow condition, which is sampled by the oscilloscope. In Fig. 8,
pulse outputting circuit. the water flowrate is 10.0 L/min, and the air flowrate is
A block diagram of the transmitter software is shown 6.0 L/min.
in Fig. 6. It is designed with modularized scheme, and Fig. 8(a) shows that the amplitude of the sensor signal
mainly consists of a main monitoring program, an initialization suddenly becomes small when the gas is just mixed. After
module, an interrupt module, a FRAM module, a watchdog about 0.6 s, the amplitude of the sensor signal is gradu-
module, an algorithm module, an outputting module, and a ally stabilized under the system feedback control. Fig. 8(b)
human–machine interface module. shows a partial enlarged curve of Fig. 8(a). Although the
The flowchart of the transmitter software is shown in Fig. 7. amplitude of the sensor signal is very small, its waveform
After powered up, the main monitoring program calls the ini- is still the sinusoidal wave, and can be used for subsequent
tialization module, and starts the conversion of ADC to collect computation.
2400 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO. 10, OCTOBER 2014

TABLE I drop in density is given as


R EPEATABILITY R ESULTS
D1
Drop_ in_density = 1 − . (10)
D0
Above all, the effect of the apparent mass flowrate and drop in
density to the measurement errors is considered in our study
and the steps of the two-phase flow experiments are as follows.
1) Change the open degree of the liquid valve 1 to get a
proper liquid flowrate. The flowrate value can be read
from the reference CMF.
2) Vary the gas valve 2 to get a proper gas flowrate. Both
the parameter of drop in density calculated by the CMF
under test and the gas rotameter reading can indicate the
amount of entrained gas.
3) Record the true mass flowrate, true liquid density from
reference CMF, and record the gas flowrate from gas
rotameter. Then, the true GVF and true mixture density
can be obtained. Before calculating the true GVF, the
gas flowrate mixes into the flowtube under test should
be estimated from the gas rotameter reading according to
the pressure and temperature at the inlet of the flowtube
under test and the gas rotameter. But one defect in
our experimental rig is that there is no space left for
installing another piezometer for the pressure measure-
Fig. 8. Sensor signal under two-phase condition. (a) Global view of sensor ment at the inlet of the flowtube under test. Therefore,
signal. (b) Partial view of sensor signal. the assumption that the gas flowrate mixes into the
flowtube under test equals the gas rotameter reading is
IV. G AS –L IQUID T WO -P HASE F LOW E XPERIMENTS made. This assumption may cause measurement errors
The two-phase flow measurement errors are affected by for the true GVF and true mixture density, but it does
many factors, such as liquid mass flowrate, GVF, flowtube not influence the validation of two-phase flow correction
orientation, and geometry, which may result in different fluid methods, for the experiments of sampling training data
situations, like bubble flow, slugging flow, stratified flow, and and online correction are performed under the same
so on. In our experimental rig, the type of primary instru- condition.
ment is CMF025, and it is installed in horizontal orientation. 4) When the experiment begins, a data acquisition board
Therefore, only the factors of liquid mass flowrate and GVF made by National Instruments is used to count the pulse
are considered. number, which stands for the mass flowrate from the
reference CMF and the CMF under test simultaneously.
A. Apparent Liquid Mass Flowrate When the experiment ends, the data acquisition board
stops counting. By comparing the pulse numbers of the
The true mass flowrate is not available within the transmitter
two CMFs, the mass flow error of the CMF under test
under test itself when it is subject to two-phase flow. Therefore,
can be obtained. During this process, the CMF under
the observed (faulty) flowrate, rather than the true mass
test also uploads the data, like apparent mass flowrate,
flowrate, is considered to be an influential factor.
apparent mixture density, apparent density drop, signal
frequency and temperature, and so on, to PC for analysis
B. Apparent Drop in Density through SCI interface continuously. The uploaded data
In gas–liquid two-phase flow condition, GVF defined by sequence of apparent mixture density is averaged and
the proportion of gas by volume is often used to indicate the then compared with the true mixture density and the
amount of entrained gas. In practice, outside of laboratory, density error of the CMF under test is obtained. Mean-
on-line assessment of GVF is rarely possible without using while, the data sequence of apparent mass flowrate and
extra instrument. However, the observed fluid density can apparent density drop is also averaged and the average
provide an indirect measure of GVF by assuming that the values combined with the mass flow error or density
true liquid density is known. Knowledge of the true single- error are used to train the network. After the network
phase liquid density can be learned online or configured by has been trained, the raw data sequence can be treated
user [14]. Transmitter calculates the fluid density and then the as new samples to validate the network.
drop in density, which is a monotonic indicator of GVF, can 5) Repeat 2)–4), change the gas flowrate to other values
be obtained. and record the mass flow errors and density errors of
Assuming the fluid density calculated through the sensor the CMF under test with the same liquid flowrate but
frequency is D1 , and the true liquid density is D0 , then the different gas flowrate.
HOU et al.: GAS–LIQUID TWO-PHASE FLOW CORRECTION METHOD FOR DIGITAL CMF 2401

Fig. 11. Structure of the feed-forward network.

The input layer contains two nodes, p1 and p2 , which


Fig. 9. Raw mass flow error under two-phase flow.
represent the liquid flowrate and drop in density, respectively.
The output is the measurement error, which is the flowrate
error or density error.
The output of hidden neuron is given as
 r 
a1i = f 1 w1i j p j + b1i i = 1, 2, 3 . . . , s1 (11)
j =1

where r is the number of input nodes and r = 2; s1 is the


number of the hidden layer nodes; w1i j is the input to hidden
layer weights; b1i is the hidden layer biases; and
1
f 1(n) = . (12)
1 + e−n
For output layer, the output of the node is given as
 s2

a2k = f 2 w2ki a1i + b2k k = 1, 2, 3 . . . , s2 (13)
i=1
where w2ki is the hidden to input layer weights; b2k is the
Fig. 10. Raw mixture density error under two-phase flow. output layer biases; s2 is the number of output layer nodes;
and
6) Repeat 1)–5), change the liquid flowrate to other values
and record the mass flow errors and density errors of f 2(n) = n. (14)
the CMF under test with different liquid flowrate and
The mean square performance function is given as
different gas flowrate.
1
s2
The experimental data is shown in Figs. 9 and 10. The
E(W, B) = (tk − a2k )2 (15)
results show that, for CMF025 type primary instrument, when 2
k=1
the liquid flowrate varies from 1.5 to 15.5 kg/min and the GVF
varies from 0% to 30%, the mass flow error and the density where tk is the training target.
error can reach to −20%. The Levenberg–Marquardt (LM) algorithm is applied for
the training of network to short the training time and avoid
local optimum.
V. F EED -F ORWARD N EURAL N ETWORK The experimental data obtained in Section IV, shown in
U SING BP A LGORITHM Figs. 9 and 10, is used as the training data. The input of the
After the experimental data has been collected, a feed- network is
q
forward neural network is used for the prediction and correc- P = [Flow1 Flow2 . . . . . . Flowq ; Dd1 Dd2 . . . . . . Dd ]. (16)
tion of two-phase flow measurement errors. It has been proved
that anyone of continuous functions in a closed interval can The training target is
be approximated by the feed-forward neural network with one
T = [Err1 Err2 . . . . . . Errq ] (17)
hidden layer. Thus, the feed-forward network consists of one
input layer, one hidden layer, and one output layer is used. where Flow and Dd are the liquid flowrate and drop in density,
The training accuracy is improved by adding hidden neurons. Err is the corresponding measurement error, and q is the
The structure of the feed-forward network is shown in Fig. 11. number of sampling data.
2402 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO. 10, OCTOBER 2014

Fig. 13. Testing results of RBF network.


Fig. 12. Testing results of BP algorithm-based feed-forward network.

Increasing the number of hidden layer nodes can improve


the training accuracy, but too many hidden nodes can result
in long training time and poor performance of unseen input
data. The experienced formula for determining the number of
hidden nodes is given as

s1 = n i + n o + m (18)

where n i is the number of input nodes and n o is the number of


output nodes, m is a constant whose value is between 1 and 10.
Based on this formula, the final number of hidden nodes is
chosen as 15 by simulation. The training goal of mean square
error is set as 0.5.
Here, the RBF network is also simulated to make compar- Fig. 14. Steps of two-phase flow measurement and correction.
ison. The simulation results show that to achieve the same
training accuracy, the RBF network needs much more nodes,
for example, when the number of hidden layer nodes is chosen
as 25, the training goal only can achieve 2.4. The increasing When applying the correction for two-phase flow measure-
of nodes may increase the amount of computation in the ment, first, calculate the sensor frequency and phase differ-
following online correction. What’s more, when testing the ence, and then the raw mass flowrate can be obtained. Second,
network using the data sequence uploaded from the CMF calculate the density of the mixture fluid, and the drop in
under test discussed in Section IV, it is found that the RBF density can be calculated according to (10). Third, these two
network has a larger prediction error than that of BP algorithm- parameters, raw mass flowrate and drop in density, are set into
based feed-forward network, as shown in Figs. 12 and 13. the neural network as the input, after calculation, the output of
From Figs. 12 and 13, the testing errors of BP algorithm- network, that is the predicted mass flow error or the predicted
based feed-forward network are mostly within ±2%, while density error can be obtained. Finally, correct the raw mass
the testing errors of RBF network are mostly within ±4% flow and mixture density according to the predicted error. The
when the number of hidden layer nodes is 25 and simulation steps are shown in Fig. 14.
results show that increasing the number of layer nodes does Repeat the experiment in Section IV and the result of
not improve the testing results too much. Therefore, the online correction for two-phase flow measurement is shown
BP algorithm-based feed-forward network is selected for the as Figs. 15 and 16. Note that to validate the prediction ability
following online correction. of network for new inputs, the selected liquid flowrate is
in the range of training data but different from the training
data.
VI. O NLINE C ORRECTION FOR T WO -P HASE
Figs. 15 and 16 show that, for CMF025 type primary instru-
F LOW M EASUREMENT
ment, when the liquid flowrate varies from 3 to 15 kg/min
After training the neural network, the parameter of the net- and the GVF varies from 0% to 30%, the corrected mass flow
work is set into the transmitter program for online prediction errors can be kept within ±3.5%, while the corrected density
and correction of measurement errors. errors can be kept within ±1.5%.
HOU et al.: GAS–LIQUID TWO-PHASE FLOW CORRECTION METHOD FOR DIGITAL CMF 2403

optimum. Comparing with RBF network, to achieve


the same training accuracy, BP algorithm-based feed-
forward network needs less neuron than that of RBF, and
this largely reduces the computation power consumption
during online correction.
4) The gas–liquid two-phase flow experiment is performed
on the experimental rig by connecting the transmitter
under test with the CMF025 type primary instrument.
The online correction result shows that when water
flowrate varies from 3 to 15 kg/min, with GVF up to
25%, the flowrate errors can be kept within ±3.5%,
while density errors can be kept within ±1.5%.
5) However, limitation of the presented work still exists.
There is no piezometer to measure the pressure in the
flowtube and the two-phase flow experiment is only
Fig. 15. Corrected mass flow error under two-phase flow. applied for CMF025 type primary instrument under
gas–water condition. The experiments with other type
primary instruments and other application situations, like
gas–oil and oil–water, have not been performed. In our
future work, we will improve the experimental rig and
take the two-phase experiments with different primary
instruments under different application situations to fur-
ther validate our transmitter.

R EFERENCES
[1] J. R. Reizner, “Coriolis-the almost perfect flow meter,” IEE Comput.
Control Eng., vol. 14, no. 4, pp. 28–33, Aug./Sep. 2003.
[2] B. S. Freeman and N. C. Ashevillc, “Digital phase locked loop sig-
nal processing for Coriolis mass flow meter,” U.S. Patent 5 804 741,
Sep. 8, 1998.
[3] M. P. Henry et al., “A self-validating digital Coriolis mass-flow meter:
An overview,” Control Eng. Pract., vol. 8, no. 5, pp. 487–506, May 2000.
[4] Y.-Q. Tu and H.-T. Zhang, “Method for CMF signal processing based
on the recursive DTFT algorithm with negative frequency contribution,”
IEEE Trans. Instrum. Meas., vol. 57, no. 11, pp. 2647–2654, Nov. 2008.
Fig. 16. Corrected mixture density error under two-phase flow. [5] K. Chen, D.-Z. Zheng, S.-C. Fan, and Z. Jianhui, “Novel Coriolis
mass flowmeter signal processing algorithms based on DFT and digital
correlation,” in Proc. IEEE Symp. ISIEA, Oct. 2010, pp. 56–60.
[6] Q.-L. Hou, K.-J. Xu, M. Fang, W.-J. Xiong, and C. Liu, “A DSP-based
VII. C ONCLUSION signal processing method and system for CMF,” Measurements, vol. 46,
no. 7, pp. 2184–2192, Aug. 2013.
1) A gas–liquid two-phase flow experimental rig is devel- [7] M. Seeger, “Coriolis flow measurement in two phase flow,” IEE Comput.
Control Eng., vol. 16, no. 3, pp. 11–16, Jun./Jul. 2005.
oped in this paper. On this rig, two CMFs are installed on [8] R. L. Maginnis, “Initialization algorithm for drive control in a Coriolis
the pipeline. One CMF is used as a liquid flow reference flowmeter,” U.S. Patent 013 330 7A1, Sep. 19, 2002.
meter, while the other is the CMF under test for two- [9] H. Shimada, “Coriolis flowmeter,” U.S. Patent 026 848 4A1,
Oct. 21, 2010.
phase flow measurement. By comparing with the ref- [10] M. P. Henry and M. E. Zamora, “Startup and operational techniques for
erence CMF reading, the two-phase flow measurement a digital flowmeter,” U.S. Patent 695 076 0B2, Sep. 27, 2005.
errors of the CMF under test can be obtained. [11] M. Zamora and M. P. Henry, “An FPGA implementation of a digital
Coriolis mass flow metering drive system,” IEEE Trans. Ind. Electron.,
2) A CMF transmitter is developed with digital drive vol. 55, no. 7, pp. 2820–2831, Jul. 2008.
and digital signal processing technology based on [12] Q.-L. Hou, K.-J. Xu, M. Fang, C. Liu, and W.-J. Xiong, “Development
TMS320F28335 chip. The digital wave synthesis drive of Coriolis mass flowmeter with digital drive and signal processing
method is implemented using MDAC and DDS to technology,” ISA Trans., vol. 52, no. 5, pp. 692–700, Sep. 2013.
[13] J. Hemp and H. Yeung, “Coriolis meters in two phase conditions,” IEE
maintain the flowtube oscillation under two-phase flow Comput. Control Eng., vol. 14, no. 4, pp. 1–36, Aug./Sep. 2003.
condition. The digital zero-crossing detection method [14] R. P. Liu, M. J. Fuent, M. P. Henry, and M. D. Duta, “A neural
based on Lagrange interpolation is adopted to synthesize network to correct mass flow errors caused by two-phase flow in a
digital Coriolis mass flowmeter,” Flow Meas. Instrum., vol. 12, no. 1,
the drive signal and measure the mass flowrate. pp. 53–63, Mar. 2001.
3) The BP algorithm-based feed-forward neural network [15] M. P. Henry, M. Tombs, M. D. Duta, and M. J. Fuent, “Two-phase flow
is used for the measurement error prediction and cor- metering of heavy oil using a Coriolis mass flow meter: A case study,”
Flow Meas. Instrum., vol. 17, no. 6, pp. 399–413, Dec. 2006.
rection. The LM algorithm is applied to train the net- [16] M. P. Henry and M. J. Fuent, “Correcting for two-phase flow in a digital
work to reduce the training time and avoid the local flowmeter,” U.S. Patent 6 505 519, Jan. 14, 2003.
2404 IEEE TRANSACTIONS ON INSTRUMENTATION AND MEASUREMENT, VOL. 63, NO. 10, OCTOBER 2014

[17] D. L. Gysling, “An aeroelastic model of Coriolis mass and density meters Yan Shi received the bachelor’s degree from the
operating on aerated mixtures,” Flow Meas. Instrum., vol. 18, no. 2, Hefei University of Technology, Hefei, China, in
pp. 69–77, Apr. 2007. 2012, where she is currently pursuing the master’s
[18] V. A. Lari and F. Shabaninia, “Error correction of a Coriolis mass degree with the School of Electrical and Automation
flow meter in two-phase measurement using neuro-fuzzy,” in Proc. 16th Engineering.
CSI Int. Symp. Artif. Intell. Signal Process., Shiraz, Iran, May 2012, Her current research interests include measurement
pp. 611–616. technology and digital signal processing applica-
[19] D. W. Clark, “Non-linear control of the oscillation amplitude of a tions.
Coriolis mass-flow meter,” Eur. J. Control, vol. 4, no. 3, pp. 196–207,
1998.

Qi-Li Hou received the bachelor’s and master’s Bo-Bo Tao received the bachelor’s degree from the
degrees from the Hefei University of Technology, Hefei University of Technology, Hefei, China, in
Hefei, China, in 2008 and 2011, respectively, where 2012, where he is currently pursuing the master’s
he is currently pursuing the Ph.D. degree with the degree with the School of Electrical and Automation
School of Electrical and Automation Engineering. Engineering.
His current research interests include electrical His current research interests include measurement
measurement technology. technology and digital signal processing applica-
tions.

Ke-Jun Xu received the Ph.D. degree in electrical Rong-Wei Jiang received the bachelor’s degree
engineering and automation from the China Univer- from the Hefei University of Technology, Hefei,
sity of Mining and Technology, Xuzhou, China, in China, in 2012, where he is currently pursuing the
1988. master’s degree with the School of Electrical and
He has been with the School of Electrical and Automation Engineering.
Automation Engineering, Hefei University of Tech- His current research interests include measurement
nology, Hefei, China, since 1989. He was a Professor technology and digital signal processing applica-
in 1994. He has authored 14 books in Chinese, more tions.
than 260 papers in domestic and international peri-
odicals in Chinese or English, and obtained 29 Chi-
nese patents. His current research interests include
sensor technology, process instrumentation, and digital signal processing.

Min Fang received the bachelor’s and master’s


degrees from the Hefei University of Technology,
Hefei, China, in 2008 and 2011, respectively, where
he is currently pursuing the Ph.D. degree with the
School of Electrical and Automation Engineering.
His current research interests include electrical
measurement technology.

You might also like