AE Assisted Process Monitoring

8/3/2019 AE Assisted Process Monitoring

http://slidepdf.com/reader/full/ae-assisted-process-monitoring 1/10

Acoustic emission data assisted process monitoring

Gary G. Yen, Haiming Lu Intelligent Systems and Control Laboratory, School of Electrical and Computer Engineering, Oklahoma State University,

Stillwater, OK 74078, USA

Received 6 March 2001; accepted 25 June 2001

Abstract

Gas-liquid two-phase flows are widely used in the chemical industry. Accurate measurements of flow parameters,such as flow regimes, are the key of operating efficiency. Due to the interface complexity of a two-phase flow, it is verydifficult to monitor and distinguish flow regimes on-line and real time. In this paper we propose a cost-effective andcomputation-efficient acoustic emission AE detection system combined with artificial neural network technology torecognize four major patterns in an air-water vertical two-phase flow column. Several crucial AE parameters areexplored and validated, and we found that the density of acoustic emission events and ring-down counts are twoexcellent indicators for the flow pattern recognition problems. Instead of the traditional Fair map, a hit-count map isdeveloped and a multilayer Perceptron neural network is designed as a decision maker to describe an approximatetransmission stage of a given two-phase flow system. © 2002 ISA—The Instrumentation, Systems, and AutomationSociety.

Keywords: Acoustic emission; Process monitoring; Nondestructive testing; Artificial neural network

1. Introduction

Modern engineering technology is leading to in-creasingly complex chemical processes with evermore demanding performance criteria. Imminentneeds in optimizing the production yield and costfor global economic competition call for an evenhigher standard in performance reliability. A criti-cal need with a supplementary sensory systembased upon nonintrusive sensory signatures surely

exists to assist the monitoring decision by opera-tors. The research dedicated to process industry,such as the one proposed herein, will promote anultimate enabling tool appropriate for on-linehealth monitoring and decision making. To sub-stantiate the feasibility, a generic gas-liquid two-phase vertical column is considered to validate thetechnology proposed.

Gas-liquid two-phase flows are defined as theflow of a mixture of two homogeneous phases, gasand liquid, through a system. Since they would aidthe description of heat and mass transfer mecha-nisms in a system, they play a very important roleand are widely used in petrochemical and chemi-cal process industries 1. An example of thiswould be the pipe flow in the gas/liquid two-phaseconveying process. It has been proven that the op-erating efficiency of such a process is closely re-

lated to accurate measurement of flow parameters,such as flow regimes and multiple flow velocities2. Generally speaking, flow patterns are classi-fied as bubbly 3, slug 4, churn 5, and annular6. These flow regimes typically have distinctflow characteristics and heat and mass transfermechanisms, which are very critical for detailedstudy in this field. Some detection techniques were

ISA

TRANSACTIONS ®

ISA Transactions 41 2002 273–282

0019-0578/2002/$ - see front matter © 2002 ISA—The Instrumentation, Systems, and Automation Society.



applied to monitor and detect flow pattern imagesor measure flow velocity in previous research. Xuand Xu 7 established a mathematical model,two-value 0/1 logical back-projection filtering

algorithm combined with a transmission-mode ul-trasound computerized tomography system. Itspurpose is to reconstruct the image of a distribu-tion of bubbles over a two-dimensional 2D crosssection of a pipe for both parallel beam scanningand fan-shape beam scanning geometry. Albusaidiand Lucas 8 proposed a technique that consistsof mounting an array of 64 axially separated con-ductivity sensors in a vertical pipe through whichan air/water mixture is flowing, to obtain the meanCap bubble or Taylor bubble velocity and hencean estimation of the mean gas velocity by cross

correlation of the output signals. These detectionapproaches based upon ingenious sensor innova-tions are often local in nature, passive, and laborintensive. The developed prototype instrumentsare heavy, expensive, and fault prone. Thus it isdifficult to implement these methods in a trans-portable, on-board, automatic, real-time, and glo-bally health assessment tool. In this paper, we pro-pose a cost-effective, nonintrusive monitoringapproach based upon acoustic emission AE sen-sors combined with an artificial neural network tostudy a vertical air-water, two-phase flow systemand classify the four major flow patterns of this

phenomenon.Acoustic emission is a term describing a class of

phenomena whereby transient elastic waves aregenerated by the rapid release of energy from lo-calized sources within a material. AE has devel-oped rapidly over the last two decades as a non-destructive evaluation technique and as a tool formaterial research. It is a highly sensitive approachfor detecting active microscopic events in a mate-rial and has been successfully used in the field of monitoring the welding or crack in solid materials,such as metal, glass, and ceramic under stress 9.

Some acoustic emission sensors were designed formonitoring the kinetics of chemical reactions 10.Additionally, an acoustic emission monitoring sys-tem was built to estimate the size of suspendedsolids in an agitated vessel, and estimate the yieldfrom a crystallizer 11. In this paper, a systemwith AE methods was applied to detect and clas-sify four major regimes, namely, bubbly, slug,churn, and annular of a vertical air-water two-phase flow.

The remainder of this paper is organized as fol-lows. Section 2 investigates four significant flowregimes. The well-regarded Fair regime map is in-troduced to describe traditional regime classifica-

tion technique. Section 3 discusses the principle of acoustic emission techniques and defines someimportant parameters for the experiment. Section4 proposes our AE based air-water two-phase flowregime classification system. System hardwareconfiguration is illustrated, and a multiplayer Per-ceptron MLP neural network is designed to lo-cate the detected signal at the correct position onthe AE hit-count map. Section 5 presents the ex-perimental result by our designed system. Section6 provides some concluding remarks along withpertinent observations.

2. Flow regimes of a gas-liquid two-phase

column

The description of a two-phase flow in pipes ishighly intricate due to the various existence of theinterface between the two phases. For gas-liquidtwo-phase flows, the variety of interface forms de-pends on the flow rates, phase properties of thefluid and on the inclination, and the geometry of the tube. Generally, for vertical gas-liquid two-phase flows, the flow regimes are mainly deter-

mined by the phase flow rates. In this case, bubbly,slug, churn, and annular are four significant re-gimes that can be recognized as standard patternsin the chemical industry. The characteristics of these four patterns are shown in Fig. 1. Each of these four patterns has a distinguished air/waterdensity and flow speed ratio. The calculation of the flow rates is required to ensure that all the flowpatterns could be observed. In order to obtain allthe required flow rates with the equipment, theflow regime map developed by Fair 12 was used.The map, shown in Fig. 2, is a plot of a Martinelli

parameter 13 X tt , given by

X tt 1 x

x

0.9

g l

0.5

l

g

0.1

1

versus the total mass velocity GT , defined as

GT

mgm l

S, 2

274 Gary G. Yen, Haiming Lu / ISA Transactions 41 (2002) 273 –282



where quantity x is defined as the mass fraction of the gas phase in the two-phase mixture and isgiven by

xm g

mgm l

. 3

Other parameters involved are defined as

m g : gas flow rate, in lb/s;m l : liquid flow rate, in lb/s; g : gas density, in lb/ft3; l : liquid density, in lb/ft3;

g : gas viscosity, in lb/ft s; l : liquid viscosity, in lb/ft s; andS: area of cross section of the pipe, in ft2.

In our experiment, the phase flow rates can bemeasured by given meters. The pipe radius isknown, and the density and viscosity values of airand water are given in standard lookup tables.Therefore we can map the obtained data to theexact position on the Fair map, which provides areference to the flow regimes of an AE classifica-tion system.

3. Acoustic emission technique

Acoustic emission testing is a powerful methodfor examining the behavior of materials in which atransient elastic wave is generated by a rapid re-

Fig. 1. Water/air flow ratio of four major two-phase vertical flow patterns.

Fig. 2. Fair map—four major regimes.

275Gary G. Yen, Haiming Lu / ISA Transactions 41 (2002) 273 –282



lease of energy. During the AE test, the sensors onthe test piece produce any number of transient sig-nals. A signal from a single, discrete event isknown as a burst-type signal. This type of signalhas a fast rise time and a slower decay, as illus-trated in Fig. 3. Burst-type signals vary widely inshape, size, and rate of occurrence, depending onthe structure and test conditions. Several param-

eters of an AE signal need to be defined as fol-lows:

a AE threshold: a predefined value to indicatethe occurrence of an AE hit and a number of AEcounts.

b AE hit event: occurs when the amplitudevalue of the sensor output signal is higher than thepredefined threshold.

c AE signal duration: the period between hitstarting and ending points.

d AE ring-down count: the number of thethreshold-crossing pulses.

Figure 4 shows the frequency spectrum of an

AE signal.Since AE signals in our experiment are of rela-

tively short durations less than 1 msec, reachmaximum amplitude early in the signal alwaysassume 0, and decay exponentially, as shown in

Fig. 3. Standard AE event hit signal in time domain.

Fig. 4. Standard AE event hit signal in frequency domain.




Fig. 3, we can calculate the sensor output as

V t V 0ert sin wt , 4

where

V (t ): output voltage of sensor;V 0 : initial signal amplitude;

r : decay constant (0);t : time; andw: signal frequency.

Since threshold voltage V * has been set up, wecan count the number of times the sensor voltageexceeds it. This technique is known as ring downcounting. For the signal represented by Eq. 4, thenumber of counts ( N ) to the nearest integer isgiven by

N t *

2 / w

w

2 r ln

V 0

V *, 5

where

V *V 0ert *; and t *1

r ln

V 0

V *. 6

For the given air-water two-phase flow classifica-tion problem, we obtain the average values of eachsecond for those related parameters introducedabove. Table 1 shows the comparison results forthe four major patterns.

From the data analysis summarized in Table 1,we can see that the number of AE ring-downcounts, occurring in one second, is the most reli-able indicator for the given pattern classificationproblem. To ensure the reliability of the final clas-sification result, we also combine the AE hitevent number, occurring in one second, to be

another indicator. By mapping one-second data toa point on the hit-count map, we can classify thefour major flow regimes, which will be shown inSec. 5. For a more complicated gas-liquid verticalcolumn to be monitored, all features discussedabove can be integrated into the decision-makingprocess.

4. AE classification system

The hardware configuration of the proposedreal-time AE air-water two-phase flow classifica-

tion system is shown in Fig. 5. The system in-cludes data acquisition, signal processing, dataanalysis, and decision making. In Fig. 5, sensor Ais an AE sensor which is attached on the pipe fordetecting AE signals that occurs in flows, and sen-sor B is the same type of piezoelectric AE sensor,which is located near the sensor A and designatedfor detecting background noise. Related AE hard-ware parameters are listed in Table 2. The outputdata from AE sensors are amplified and filtered

Table 1AE parameter values of four major flow regimes. The values in are the value range of the given AE parameters in onesecond.

Bubbly Slug Churn Annular

Average number of AEhits occurs in one

second

50–58

8325–150

185134–243

394–92

Average number of AE

counts occurs in one

second

87

0–32

749

17–4923

8192

2487–14236

30521

13510–44673

Average value of

amplitude for the AE

hits occurs in one

second dB

61.56

60.17–61.92

60.59

59.81–60.7

61.43

59.93–61.65

61.13

60.11–61.96

Average Rise time for

the AE hits occurs in

one second s

17.7

1–59

20.49

1–429

26.47

1–711

35.37

1–717

Average Duration timefor the AE hits occurs

in one second s

751–264 127.91–8468 314.61–43012 430.61–132259




before the A/D conversion. In data analysis anddecision making parts, the input of the MLP neuralnetwork is the discrete data from the AEDSP cardmanufactured by Physical Acoustics Corporation.

The network output determines the current posi-tion on the hit-count map, which can be the indi-cator of the decision-making Graphic User Inter-face GUI software.

5. Experimental results

In the experiment, we designed a continuousregime-changing process which includes 12 steadystates and 11 transient states during 20-min dataacquisition time. Each steady state keeps a pair of typical air and water flow rates Fig. 6 for about

30 seconds and then transfers to another steady

state. The state between two steady states is the

transient state, which has unstable phase flow

rates. The phase flow rates of all the steady and

transient states were recorded to generate our

reference Fair map Fig. 7. Meanwhile, the AE hit

and count number for each second period are also

measured and stored during the entire

process.

Figure 8 shows the corresponding AE hit-count

map, and each marked point represents the sum-

mation AE hit number and count number for one

second. From this hit-count map, we can see that

regime ‘‘annular’’ is clearly separated from the

other three regimes. Regimes ‘‘churn’’ and ‘‘slug’’

are also well separated, although there exists a

small overlapping between these two regimes. The

regime ‘‘bubbly’’ and ‘‘slug’’ cannot be linearly

separated since the overlapping area between them

is not trivial. However, we can train a nonlinear

classifier, such as MLP neural network to classify

them, because the majority of these two regimes

are separable.

In our experiment, we use a two-hidden-layer

neural network with ten neurons in each hidden

layer. We randomly select 600 data points for

Fig. 5. Configuration of experimental AE detection system.

Table 2AE hardware parameter values.

AE sensor resonant frequency 150 kHz

Sampling rate of DSP board 1 MHz

Gain of amplifier 40 dB

Time window of each AE hit 1024 points 256 s

Threshold voltage 0.0586 V 35 dB




training and the other 600 data points for testing.The Levenberg-Marquardt with Bayesian regular-ization algorithm is applied, and the learning rateis set to be 0.1. The training epoch is 1000, and thestopping mean-square error is 1e-5. Figs. 9a andb show the training result hit-count map and thetraining target hit-count map. Figs. 10a and bshow the testing result hit-count map and testingtarget hit-count map.

Comparing the training and testing result hit-count maps with the target maps in Figs. 9 and 10,

we can see that the overlapping areas between ad-

jacent regimes have disappeared due to the clus-tering character by the neural network. In thiscase, the designed neural network can improve theclassification performance by reducing the mis-classification rate for the given air-water two-phase flow regime classification problem. The out-put of the neural network is transferred to thedecision making and Graphic User Interface soft-ware. This software is designed to have the fol-lowing functions: 1 indicate the current flow re-gime; 2 control the starting and stopping of the

data acquisition process by user; and 3 perform

Fig. 6. Air and water flow rates for steady states and transient states.

Fig. 7. Reference Fair map for the four major regimes.




database management for history record inquiry.Fig. 11 shows the GUI software. The developedinterface allows the operator to visually monitorthe process to facilitate the expert decision-making. By integrating with concepts, such as dis-tributed virtual instrumentation, the operator notonly can remotely monitor the process, but acti-vate the control law for reconfiguration via Ether-net 14.

6. Conclusion

The application of one of the nondestructivetesting techniques i.e., AE and neural networksNN on vertical air-water two-phase flow patternrecognition problems was proposed and discussed.In this study, several AE parameters were ex-tracted from four major two-phase flow patternsignals, and the results were discussed. AE

Fig. 8. AE hit-count map corresponding to the Fair map in Fig. 7.

Fig. 9. a Neural network training result hit-count map and b target training hit-count map.




hits events and ring-down counts density can becombined as a stable and excellent indicator todescribe flow patterns accurately. They form theinput stream of multilayer perceptron neural net-work. After training the network, the system out-put can tell the continuous flow stage includingfour major patterns and transient states on-lineand real time. This AE combined NN detectionsystem may be easily transferred to other gas/

liquid two-phase flow regime classification prob-lems, such as saturated steam flow, which iswidely related to different industrial processes forheat energy transfer, power source, sanitary flush-ing, etc. Some common flow regimes for saturatedsteam flows are uniform density regime, annularregime, slug regime, and asymmetric density re-gime. While not demonstrated, we fully believethat the proposed acoustic emission monitoring

Fig. 10. a Neural network testing result hit-count map and b target testing hit-count map.

Fig. 11. GUI software of the AE flow regime detection system.




system will work equally well for saturated steamand other similar two-phase flow systems.

References1 Hewitt, G. F., Measurement of Two Phase Flow Pa-

rameters. Academic, London, 1978.2 Galegar, W. C., Stovall, W. B., and Huntington, R. L.,

More data on two-phase vertical flow. Pet. Refineries33, 208–219 1954.

3 Rhodes, E., Gas-liquid flow on video tape, HeatTransfer & Fluid Flow Services, Atomic Energy Re-search Establishment. Harwell, London, 1982.

4 Fernandes, R. C., Semiat, R., and Dukler, A. E., Hy-drodynamic model for gas-liquid slug flow in verticaltubes. AIChE J. 29, 981–995 1983.

5 Taitel, Y., Bornea, D., and Dukler, A. E., Modellingflow pattern transmissions for steady upward gas-liquid flow in vertical tubes. AIChE J. 26, 345–359

1980.6 Hewitt, G. F., Disturbance waves in annular two-phase

flow. Proc. Ind. Mech. Eng. 18, 142–149 1969.7 Xu, L. J. and Xu, L. A., Ultrasound tomography sys-

tem used for monitoring bubbly gas/liquid two-phaseflow. IEEE Trans. Ultrason. Ferroelectr. Freq. Control44, 67–74 1997.

8 Albusaidi, K. H. and Lucas, G., Measurement of mul-tiple velocities in multiphase flow. IEE Colloq. Adv.Sens. Fluid Flow Meas. 12, 1–4 1995.

9 Bassim, M. N., Dudar, M. P., Rifat, R., and Roller, R.,Application of acoustic emission for nondestructiveevaluation of utility inductive reactors. IEEE Trans.Power Deliv. 8, 281–284 1993.

10 Bang, S. W., Lec, R. M., Genco, J. M., and Ransdell,

J. C., Acoustic emission chemical sensor. Proc.-IEEEUltrason. Symp. 1, 439–443 1993.11 Bouchard, J. G., Payne, P. A., and Szyszko, S., Non-

invasive measurement of process states using acousticemission techniques coupled with advanced signalprocessing. IChemE Trans. 72, 20–25 1994.

12 Fair, J. R., What you need to design thermosiphonreboilers. Pet. Refineries 39, pp. 105–116 1960.

13 Martinelli, R. C. and Nelson, D. B., Prediction of pres-sure drop during forced-circulation boiling of water.Trans. ASME 70, 695–704 1948.

14 Yen, G. G., Sensory-based expert monitoring and con-trol, Proc. Intl. Conf. Inf. Fusion. 953–959 1999.

Gary G. Yen received thePh.D. degree in electrical andcomputer engineering from theUniversity of Notre Dame,Notre Dame, IN in 1992. He iscurrently an Associate Profes-sor in the School of Electricaland Computer Engineering,Oklahoma State University,Stillwater, OK. Before he

joined OSU in 1997, he waswith the Structure Control Di-vision, U.S. Air Force Re-search Laboratory in Albuquer-

que, NM. His research is supported by the DoD, DoE, EPA, NASA,

NSF, and Process Industry. His research interest includes intelligentcontrol, computational intelligence, conditional health monitoring, sig-nal processing, and their industrial/defense applications. Dr. Yen wasan associate editor of the IEEE Transactions on Neural Networks andthe IEEE Control Systems Magazine during 1994–1999. He is cur-rently an associate editor for the IEEE Transactions on Control Sys-

tems Technology and the IEEE Transactions on Systems, Man and

Cybernetics, Part C: Applications and Reviews. On behalf of ISA, Dr.Yen has served as a Society Review Chair for American Control Con-ferences since 1998.

Haiming Lu received the B.S.degree in electrical engineer-ing from Tsinghua University,China in 1995. He is currentlyworking toward the Ph.D. de-

gree in electrical engineeringat Oklahoma State University.His interest of research in-cludes evolutionary algo-rithms, neural networks, multi-objective optimization, signalprocessing, and their applica-tions in process industry.


Documents

AE Assisted Process Monitoring