IDENTIFICATION OF GAS-LIQUID FLOW REGIMES IN A HORIZONTAL FLOW USING NEURAL NETWORK

66Journal of Hydrodynamics Ser. B 2005 ,17( 1 ) 166 -73China Ocean Press , Beijing - Printed in ChinaIDENTIFICATION OF GAS-LIQUID FLOW REGIMES IN A HORIZONTALFLOW USING NEURAL NETWORK。JIA Zhi-hai , NIU Gang , W ANG JingSchool of Mechanical and Power Engineering , Shanghai Jiaotong University , Shanghai 200030 , China , e-mail :wangj@ sjtu. edu. cn( Received Sept. 18 ,2003 )ABSTRACT : The knowledge of flow regimes is very important ent behaviour of the flow which gives rise to difficultiesin the study of a two-phase flow system. A new flow regime identi-in instrumentations and measurement. The characteristicfication method based on a Probability Density Function( PDF )parameters closely relating to flow regimes are the pres-and a neural network is proposed in this paper. The instantaneousdifferential pressure signals of a horizontal flow were acquired withsure , pressure drop and void fraction. Because pressurea differential pressure sensor. The characters of differential pres-or pressure drop signals are acquired easily in most ca-sure signals for different flow regimes are analyzed with the PDF. ses. They are widely used as the main signals in manyThen , four characteristic parameters of the PDF curves are dmeasurement processes. For instance ，Tutul1] ，Mat-fined , the peak number ( K ), the maximum peak valuesuif23] , Wu et al.[4] acquired the differential pressure( K2 ),the peak position( Ks ) and the PDF variance( K; ).signals of the flow pipeline and applied the fractal meth-The characteristic vectors which consist of the four characteristicparameters as the input vectors train the neural network to classifyod to analyze the signals to identify flow regimes. Voidthe flow regimes. Experimental resuls show that this novel meth- fraction is also used to identify the flow regimes by Jonesod for identifying air-water two-phase flow regimes has the advan- and Zuber51 , Vince and Laheyf61 , Costigan and Whal-tages with a high accuracy and a fast response. The results clearly ley7] and Lefteri et al.[8]demonstrate that this new method could provide an accurate iden-Subjective judgments and objective indication aretification of flow regimes.two methods to identify flow regimes. Subjective judg-KEY WORDS : flow regime identifcation , Probability Densityment is usually made by the flow visualization , such asFunction( PDF ) , neural network , two-phase flow , flow regimework by Mishima and Ishif9] , Lowe and Rezkallah 10].However , flow visualization may not be reliable if theflow could be seen through or the flow is fast and the1. INTRODUCTIONvoid fraction is high , or the pipe is opaque. The pic-Two-phase flow is becoming increasingly importanttures are often confusing and difficult to be analyzed e-in many engineering processes ，such as petroleumplants , chemical processes ，power engineering andven a high speed camera is used. Therefore , the methodmany heat exchange equipments with a changing inusually acts as an auxiliary way to observe flow regimesphase. Flow regime investigation is essential to designin the laboratory. .Objective identification usually identifies the flowthese equipments , analyze the heat transfer and calcu-late the pressure drop of the pipeline.regim中国煤化工c:essing methods to ana-Many measurement techniques are used to acquire lyzenals. Therefore , signalCNMHG'character signals of two-phase flows because the transi- processing metnod 1s one 01 the crucial important aspectsProject supported by the National High Technology and Research Development Program Special Fund of China( Grant No :2002AA616050 ).Biography行方熬据i( 1975-) , Male , Ph. D.67of a two-phase flow. The PDF and the Power Spectral 125 ). There are two pressure gauges at the bottom ofDensity( PSD ) function are usually used to extract the the pipe. Through pipe , a differential pressure sensor ischaracters of the differential signals or void fraction sig- linked with these two pressure gauges and used to meas-nals. Then , the peak values and the shape characteris- ure the pressure drop of the air-water two-phase flow be-tics of PDF or PSD are used to classfy the flow regimes，tween the two gauges. The distance between the two .such as Jones and Zuber-S] , Vince and Lahey61 and gauges is 320mm( 10D ).Elkow and Rekallah1].As an advanced theory , a neural network plays animportant role in classifying the regimes. The neuralnetwork simulates human mind and shows high intelli-tE的gence. It can be trained to learn the correct output andclassify for each of the train samples. However , it isnecessary to obtain known inputs that are used to train .neural network. After training , the neural network canclassify the unknown similar regimes and have high ac-curacy. For instance , Lefteri et al. [81 studied the neuro-母宫轫fuzzy methodology to identify the multiphase flow re-gimes. Wu et al.4J used fractal dimensions to train BP1- -Water tank ,2- -Water pump ,3- -Orificeneural network to judge flow regimes. Mi et al.[12 J131]3 lap-flow meter for water loop , 4-Air compres-plied flow regime maps and two-phase flow models tosor ,5- -Air filter ,6- -Air pressure buffer ,train supervised network to identify flow regimes.7- -Orifice flow meter for air loop , 8-AirThe two-phase flow is very complicated and thewater mixer , 9- -Test section , 10- Air-wa-knowledge of flow regimes is crucial in modeling and op-ter-separatoreration of two-phase systems. In this paper , the signalsFig. 1 Air-water two-phase flow experiment loopare acquired with a differential pressure sensor in a hori-zontal flow. The method combining PDF and neuralIn this experimental loop , the superficial velocitieswork is proposed to identify flow regimes.of air and water were varied in order to span four differ-ent flow regimes : stratified , churn ,slug flow and annu-2. EXPERIMENTAL SETUPlar flow. The data acquisition card is AT-MIO-64E-32.1 Test loopmade by NI. The output signals of the differential pres-The experimental loop of a horizontal air-water two- sure between two gauges were sampled at 200Hz for aphase flow is shownin Fig. 1. Water is supplied froma periodof 15s to 30s. The flow regimes were observed .water tank ( 1 ) and flows through a calibrated orifice through transparent part of the test section.flow meter( 3 ) into an air-water mixer( 8 ). Air is sup-回Phplied from a compressor( 4 ) , and flows through an airpressure buffer( 6 ) and a calibrated orifice flow meter(7 ) into an air-water mixer( 8 ). The air and water2800t 509232|mixture then flows through a test section( 9 ) into an air-water separator( 10 ) where the air is separated to theFig.2 Schematic of the test section( mm )atmosphere and the water flows into the water tank( 1 ),for cycling utilization. .中国煤化工2. 2The test section ,as shown in Fig.2 ,is made up ofHCNMHGHhie guar uis siuuy was to develop a flow regimea plexiglass tube with an inner diameter of 32mm( D ).identification method. It is insensitive to flow unsteadi-The inlet length of test section is 2. 8m( L/D = 87.ness and transients. Hence , no special care is taken5 ). The length of test section is 1.6m( L/D=50 ).The outlef test section is 2.5m( L/D = 78.duringtheexperiments to ensure either a particular flow68regime or a developed steady-state flow. The situation is That means the variance of the PDF of the stratified flowcommon in the multiphase flow system. In the experi- is rather bigger in values.mental loop , the water flow rate varies in the range of8.3.2 Churn flow and its PDF characters3 x10*to5.6x 103 m'/s. The air flow rate varies in theAs the air velocity increases , the interface beginsrange of0 to 0. 02m'/s. The environmental temperature violently fluctuant and becomes instable. The flow re-is 18C and the environmental pressure is 1.01 x 10’ gimes from stratified flow changes into a churn flow. ThePa.differential pressure time series of the churm flow and thePDF character have been shown in Fig. 4( a ) and Fig.TYPICAL FLOW REGIMES AND PDF 4( b).CHARACTERS2.6 7Each flow regime has an intrinsic effect on manyflow parameters , such as void fraction , pressure drop,heat and mass transfer. Therefore ,it is important in themodeling and the operation of two-phase flow systems to1.0+812售1know the instantaneous flow regimes. In this paper , fourtypical flow regimes are identified. The PDF is used to .Fig.4( a) Time series of a typicalanalyze the characters of these four different flow re-chum flowgimes.0.01243.1 Stratified flow and its PDF characters .0.008-When the water velocity and the air velocity are0.004small , the stratified flows can be observed in the testsection. There exists a clear interface between the air00.T322sAP/Kpaand the water. The interface is smooth. A typical timeseries of a stratified flow is shown in Fig. 3( a). In thisFig.4( b) The PDF of a typicalfigure , the differential pressure time series centers on achurm flowlow average differential pressure with small fluctuations.1.5-The differential pressure signals of a churn flowhave a bigger fluctuation than these in a stratified flow asshown in Fig. 4( a ). Then the bigger fluctuation at theinterface between the air and water can be observed.4812 i6The PDF curve for churn flow has a peak at the biggertvoltage than those for stratified flow. The maximum val-Fig.3( a) Time series of a typicalue of the peak is almost the same as a stratified flow.stratified flowThe variance is smaller than those in stratified flow.0.0161Comparing with a stratified flow , the maximum peak val-0.012 H台0.00ue is smaller and the peak position is bigger.0.004-3.3 Slug flow and its PDF charactersSlug flow is a very important flow regime in desig-0505~T1522sSPIKpaning and analyzing a two-phase system. Violent slugflow can cause a huge wallop and affect the efficiency ofFig.3( b) The PDF of a typical strati-a pum中国煤化工gineering fields ,a slugfied flowflow sMHC N M H Gated. A typical unit ofaThe PDF character of a typical stratified flow isslug flow consits of a Taylor bubble and a liquid slug ,shown in Fig. 3(b). In Fig. 3( b), the PDF curve hasin which small bubbles are surrounded by continual liq-a peak atp h数pltage. The peak is abrupt in shape.uid. A typical time series of a slug flow is shown in Fig.5( a).69The time series of a slug flow is totally different than those stratified flow.from a stratified flow and a churn flow. Each unit is likea ladder shape for a slug flow. The PDF character of a0.55slug flow is shown in Fig. 5( b ). The PDF curve of slugflow has two peaks. It is shown that a slug flow is char-0.45acterized by a bimodal PDF with a low-void peak corre-0.35 -sponding to a Taylor bubble and a high-void peak due to4812/sthe liquid slugf 13].Fig. 6( a) Time series of annular flow0.016 {2.2-{0.012瓷008}0.004To 1s2 2s2 160P/KpaFig.6(b) The PDF of a typical annu-Fig.5( a) Time series of a typicallar flowslug flow4. FLOW REGIME IDENTIFICATION USING0.034RBF NEURAL NETWORKy0.02In this section , four characteristic parameters of0.01PDF curves are defined. Then , the characteristic vec-h16120 22 24tors which consist of four defined parameters are used assP/Kpathe input samples to train a neural network to identifyFig.5( b) The PDF of a typicalflow regimes.4.1 Characteristic parameters of a PDFTo train a neural network , four character parame-ters of PDF curves are defined. They are the peak num-ber( K ) , the maximum peak value3.4 Annular flow and its PDF characters( K2 ),the peak position( K; ) and the PDF varianceAs the air velocity continually increases , flow re- ( K4 ) respectively. They consist of the input vectors ofgime changes into an annular flow. There exists a gas a neural network.core and a liquid film around the inner wall of pipe. A(1 ) K ,the peak number of a PDFtypical time series of an annular flow is shown in Fig. 6K is defined as the peak number of a PDF. From .( a). The time series of an annular flow is similar to a above PDF curves , a stratified flow and a churn flowstratified flow in shape. However , the differential pres- have one peak and a slug flow has two peaks. There-sure time series with smaller fluctuation of an annular fore , the peak number can be used as one of the charac-flow have a lower average values than those in stratified teristic parameters.flow.中国煤化工_ak value ofa PDFAn annular flow PDF character is shown in Fig. 61H.CNMHGak values corresponding(b ). There is a peak in the low differential pressure to diffterem 11w Tegnes. 50 , une maximum peak valuevalue. The maximum peak value of an annular flow is of a PDF ,K, ,can be used as the second characteristicthe almost same as in stratified flow. However , the peak parameter.shape is b座upt and the PDF variance is bigger(3) K; ,the peak position of a PDF70Table 1 Part of the sample input of the RBF neural networkIdentifcationCaseWate(以)Air(ogn)FlowKK2K,Result ofno.( m/s)regimeRBF0. 03450. 1380. 0200. 8833. 90x105StratfedStratified0.5180.0190. 9023.78 x 1051. 036.10. 01700. 9953.69x 10-Stratified .0. 0172.033.65 x 10-50. 01802.023.61 x1051.0360.01502.043.54x10-Stratifed0. 06912.8430.0171.322.65 x 10-5ChurnChur5. 188 .0. 01601.452. 62x10*Chum6. 9080.015 .1.692.59x105Churm0. 1042. 84301.412. 46x 105!10.104 .5. 18801.592.45x10-20.1041.742.25x1050. 0141. 91 x10*40. 138.0.01401.601. 80x 10551.781.69 x 10560.3450. 3450.015 0.021 1.82 2. 111.73 x 10-Slug70. 6910.015 0.010 1. 752.08.2.29x10-1. 3820.025 0.012 1.62 2. 183.70x10590. 8642 0.015 0.012 1.71 2.071.95x10520.01 0. 03.1.882.203.20x 10-0.03 0.01 1.68 2. 184. 34 x 1050.016 0.027 1.72 2. 082.33x1051.3820.018 0.030 1 702.083.49x1052420.0200.0321.762.124.75x10-0. 010420. 7290. 0150.484.02 x 10-5Annular22. 4540. 015.00.494.11 x1052724. 1780.0164.06x105820.7290.514.08x105291 0.01600.52中国煤化工CNMHGTable 2 Identification results of RBF neural networkStratifiedFlow regimeSlugAnnularflowcategory Chur flowNumber of flow regimes45418Number of right identification39505116Identification accuracy rating ( % )92. 992.694.4 .88. 8The peak position is different for four typical flow re-To enhance the network accuracy and decrease thegimes. Therefore , the peak position can be used as the hidden nodes , improved Gauss function is used as thethird characteristic parameters.radial basis function by follow expression(4) K、, the PDF varianceThe PDF curves of above four flow regimes are dif-ferent in shape. The PDF peak shape of a stratfied flowR(x)=exp[-(x-c,)K(x-c,)] .(2)and an annular flow are abrupt and the peak shape of achurm flow is not. The PDF variance ,K。, is used as In the function( 2 ) ,K can be calculated bythe fourth characteristic parameter to describe the char-acterK = E[(x-c)(x-c; )I-'(3 )The characters of differential pressure signals areanalyzed by a PDF. Parts of the characteristic vectors where E is a unit matrix.consisted of four defined parameters are shown in TableIn the function( 1 ) ,C; is the center of the No. i ra-dial basis function. It can be calculated by c cluster4.2 Flow identification with RBF neural networkmethod. Givena sampley; ,(y;∈T;rj=1 ,2,3..，Flow regime classification is difcult to realizeby a M T; is the No. j cluster )traditional classification method. In this paper , a RadialBasis Function( RBF ) neural network is used as the y; = {y1 2 V3... vv},(i = 1 ,2 ,3.... N) (4)flow regime classifier. The RBF neural network hasthree layers. The input nodes pass the input values to its mean value is given bythe connecting arcs , and the first layer connections arenot weighted. Thus , each hidden node receives each in-N. Ey(5)put value unaltered. The hidden nodes are the radial ba-INiy;er;sis function units. The transfer function for the hiddennodes is non-monotonic Gauss function in contrast to the J. is the error square sum and is defined bymonotonie sigmoid function of Back Propagation( BP )networks. The second layer of the connections is weigh-J。=22. ly-,I2(6)ted , and the output nodes are simple summations. Thisparticular architecture of the RBF has been proved to中国煤化工improve directly the training performance.J. isu: dfferent clusterI; ,J。The Gauss function is given by,is d:MYHc N M H Ginimum ,c, get its opi-mal result. By these equations , the centers and width ofR(x)=exp[--(1)the radial basis functions R( x )can be adjusted and a2σ?useful RBF neural network model can be acquired. .The reason for using the RBF is due to finding a .72number of drawbacks of BP networks in the particularapplication. Comparing to the BP neural network , RBF 5. CONCLUSIONSneural network has a better quickly learning ability , a( 1 )A new method for the flow regime identifica-pattern recognition ability and a classification abilty tion based on PDF and RBF neural network is proposedthan a BP neural network.in this paper. The features of the differential pressureNeural network learning algorithm can be divided signal are extracted by PDF. The RBF neural network isinto the trained with or without supervised. In this pa- used to classify different flow regimes.per , flow regimes are identified by supervised RBF neu-(2 ) The PDF plays an important role in analyzingral network systems. The flow regime classification the signals. It is valid and quick to use the PDF to ex-structure of RBF neural network is shown in Fig. 7. Let tract signal characters. The PDF can easily characterizethe output vector of a stratified flow , a churn flow , a different flow regimes.slug flow and an annular flow be( 0 ,0 ,0 ,1 ),(0 ,0，( 3 ) RBF neural network is applied to identify dif1 ,0),(0,1 ,0 ,0)and( 1 ,0 ,0 ,0 ) respectively. ferent flow regimes. The neural network has a better andThe characteristic vectors consisted of the four defined faster learning ability ,a pattern recognition ability and aPDF parameters are used as the input samples to train classification ability than a BP neural network. The ap-RBF neural network. The input samples of 42 stratified proach has been tested on a practical system of differen-flows ,54 churn flows , 54 slug flows and 18 annular tial pressure data. The results show that the applicationflows are used to train the neural network. Parts of the of the method with traditional statistics analysis and aninput samples are shown in Table 1. The self-organiza- advanced neural network can quickly and automaticallytion training method is used to train RBF neural net- identify the flow regimes with a high accuracy.work. This method doesn' t need know the number of( 4 ) To improve the identification accuracy and ex-the hidden layer units at the beginning. The training tend the application fields , the sample database of flowprocess was performed 3000 episodes for a whole set of regimes for different flow regimes need be established.sample vectors. The system might be over-trained if thetotal number of episodes is large and will have a poorextending ability.REFERENCES[T.0,0.0,1)-Stratified[ 1 ] TUTU N. K. Pressure fluctuations and flow regime recogni-charactertion in vertical two-phase gas-liquid flows[ J ] Int. J.k0,0,1,0)-ChumMutiphase Flow ,1982 ,8( 4 ) :443-447.K1,K2,K3,20.10.0)-Slug[2] MATSUI G. Identifcation of flow regimes in vertical gas-k1.0,0,0)-Annularliquid two-phase flow using differential pressure fluctuations[J] Int. J. Multiphase Flow ,1984 , 10( 6 ):711-720.SamplesRBF neural networkResults[ 3 ] MATSUI G. Automatic identification of flow regime in ver-tical two-phase flow using differential pressure fluctuationsFig. 7 Schematic of RBF neural network identifica-[ J] Nuclear Engineering and Design , 1986 ,95 :221-tion system231.After training the RBF neural network , the weights[4] WU Hao-jiang ,ZHOU Fang-de , WU Yu-yuan. Inelligentidentification system of flow regime of oil-gas-water multi-and bias in each neural are adjusted so that the neuralphase flow[J] Int. J. Multiphase Flow ,2001 ,27(3 ):network is able to match the statistics inputs with their459-475.corresponding outputs associated with the flow regimes.[5]中国煤化工e iterelation between voidThe same number of test samples for stratified flowchurn flows , slug flows and annular flows are identified.0HCNMHGFpatterns in two-phase flow[」ml. J muiupiase rIuw ,1975 ,2( 3 ):273-306.Identification results of the neural network after training [6] VINCE M. A. ,LAHEY R. T. On the development of anare shown in Table 1 and Table 2. The results show thatobjective flow regime indicator[ J ] Int. J. Multiphasethe identification ability of the RBF neural network isFlow , 1982 ,8( 12) :93-124.pretty goof?73[7] COSTICAN C. , WHALLEY P. B. Slug flow regime iden-457.tification from dynamic void fraction measurements in verti- [ 11 ] ELKOW K. J. , REZKALLAH K. S. Statistical analysis ofcal air-water flow[ J] Int. J. Multiphase Flow , 1997 ,void fluctuations in gas-liquid flows under 1-G and μ-G23(2):263-282.conditons using a capacitance senso[J] Int. J. Multi-[8] LEFTERI H. , TSOUKALAS Mamoru Ishii and MI Y. Aphase Flow , 1997 ,23( 5 ) :831-844.neurofuzzy methodology for impedance-based muliphase [ 12 ]MI Y. ,ISHII M. TSOUKALAS L. H. Vertical two-phaseflow identifcatio[ J] Eng. Appl. Artif. Intel. , 1997 ,flow identification using advanced instrumentation and neu-10( 6):545-555.ral networks[ J ] Nuclear Engineering and Design,[9] MISHIMA K. ,ISHII M. Flow regime transition criteria for1998 ,184( 3 ) :409 420.upward two-phase flow in vertical tube[ J] Int. J. Heat [ 13] MI Y. ,ISHII M. ,TSOUKALAS L. H. Flow regime identi-Mass Transfer 1983 ,27 :723-737.fication methodology with neural network and two-phase. [ 10 ]LOWE D. C. , REZKALLAH K. S. Flow regime identifca-flow models[ J ] Nuclear Engineering and Designtion in microgravity two-phase flows using void fraction sig-2001 ,204( 1 ):87-100.nal[ J] Int. J. Multiphase Flow ,1999 ,25( 3 ):433-中国煤化工MHCNMHG