Modeling and control of a food extrusion process

Modeling and control of a food extrusion process

I9gS iopotu pug Id oldm!s jo oouetuaojaod ot0 pue pop!Aoad uoql oau so!pros ioauoo ,Ouu!tu.qoad 'solqe!auA lO.nUOO £mmud oql se posn oa~ l~ql soanseot...

458KB Sizes 0 Downloads 52 Views

I9gS iopotu pug Id oldm!s jo oouetuaojaod ot0 pue pop!Aoad uoql oau so!pros ioauoo ,Ouu!tu.qoad 'solqe!auA lO.nUOO £mmud oql se posn oa~ l~ql soanseotu £1.qenb poldums £Iluonbo.~ju! oql aoju! ol po!Idd~ uooq s~q (L86I) lu lo lsnoptml!nD jo .~olmu!lso l¢!luoaoju! OA!ld~p~ ol~a-!llntu oql 'aOllOaUOO¢ jo uo!lmuomoldtu! OA.IlooJJO oql oIq~uo O.L "tu~.tp~d (dD) ~ulmm~a~oad o.nouo~ oql ~u!sn ~lep imuotu!aodxo tuo~j iopotu i~o!~OlOOqa jo luotudoloAOp oql ql!A~possoappt~ uooq st~q I~.UOItZtU ssoooad £qoaels Otll jo £~OlOOq~ xoldtuoo oq1 ~u!qposop jo tuolqoad oq,L "~l~p ImUOm!aodxo ol pong pu~ podoloAOp uooq st~q amp oouop.tsoa pue 0UInlOAog.toods '£1houo l~O!ueqootu og!oods 'uo!les!u!l~ioi~ lonpoad s~ qons solqe!.reA £1!I~nb i~u!leaodaoou! ssoooad oql jo lopom o!m~ugp OlUeS.IOAV "oanleaodtuol mnpoad s~ qons S0lq'el._reA Lmpuooos tt1~ql .loqllu solq~I.r~A lOalUOO 0ql oq II!A~soanseom £1!i~nb £amuud lUtll s! so.tpms ioauoo snotAoad tuo.Ij S.lOjjt.p loo.foad s!ql oaoqAk "~!l~asnv '£~OlOUqOo& pue oouo!oS pooA jo luotulaedo(I O8IS3 oqa ,e poleooI lueld loi!d e uo otuoqos loauoo ssoooad olqm!ns e luOmolduI! pue doloAOp ol st tute oletU[lln oq.L "ssoooad uo.tsnnxo ~UplOOOe o l sonb.mqool HSd jo uo!leo!idde oql uo suodoa uo!mq[auoo stq.l. '£ouo.to!jjo ssoooad poAoadm! snql pue oi~lseta jo uo!lonpoH •£ouols!suoo pu~ ,(l!I~nb lonpoad poAoadmI 'slonpoad pooj ,/(ougj,, pue lOAOU go luotudoIoAOp oq,L

guplooo ol px~oa uI 'smomOAo.tdtu! £xessooou aoj s!sgq oql stuaoj s.lql oou.ls 'pogoldmo sossoooad oql jo ~u[puglsaapun oql olu.t qoat3osoa pOl~A!lOtUsgq sa~o,~ luoooa m somedmoo pooj ls~uotug uo!l!lodtuoo osuoluI •oououodxo aolCaodo pue suo!l~lO~OO IeO!a!dtuo jo £io$a~I sls!suoo oSpolmOU31 uo!stmxo i~u!~oo3 luoaano 'llnsoa e sV "olq!ssod suo!l~an$~uoo o!p pue A~OaOS SSOII!tU!I lSOtUle oql pue 'suo!loeoa leo!moqo poolsaopun £1aood 'le!aoletu ssoooad oql jo ano!Aeqoq t~OlJ pue so!uodoad leOlfloiooq.t XOldmO3 oql ol onp ,(ll~O!ls!u~qoom lopotu ol llno~j!p ,(lOtUOaXOoae saopnaxo i~uploo3 'aeino!lxed u I "so!alsnpu! aoqlo osoql ueql poluo!ao ljeao pue qmeq oaoui qonm ~u!oq 'gasnpu! pooj oql jo punoa~ioeq oql ol onp ,(lOi~.mi s! s!q,L "so!.nsnpu! itto!tuoqooaod pue leO.ttuoqo oql £q se luolxo ~tues oql ol Lusnpu! pooj oql £q po!ldde uooq lou OAeq sonb!uqool (HSd) ~u!aoou!$uH smolsgs ssoooM uaopom 'oou~uodtu! o.n.uouooo pue osn peoadsop!A~a!oql ol!dso(I -spo~otu uo!lonpoad leuog!pea aoAo uop~s!I~UO!l~a ssoooad pu~ ssouon!loojjo lsoo poAoadm.t loj Ie!luolod oql u! sllnsoa £11.I.IlUS.IOApu~ £ouo!ogJO ,(i~aouo 'solea lndq~noatll qi~!q jo uol.l~u.lqtuoa oq,L 's~ls'~d pu~ SleOaoalsvJ)leoaq 'spooj £q~q 'spooj )IOeUS jo uo!lanpoad oql u! poqs!lqelso [IOA~,(p~oalv oae saopnalxo i~UplOOD "sjjmspoo2 2o oi~uea op!m ~ jo i~mssoooad snonu!luoa oql aoj su~otu luot.o[IJo tm sop!Aoad 1[ se sxeo£ luoooa u! Lrlsnpu! pooj oql mo,j uo!luolle olqvaop!suoo pola~alv s~q uo!snnxo i~U.DlOOD hIOI.I.DflflOH.I,NI

• • •

:splot.J 8u!A~OIIOJ0ql u! Iu0p.IA0 £i.relnoga~d s! SlU0UIOAo.Idttl.IloJ 0Al.lp 0ql 'uo!snaxo

•pox~dmoo s! soIqe.LmAg~!Ienb osoql jo IOa~UOOOA.tlo.tpoadlopotu pue Id oldtu!s jo oo~tuao~uod polelntu!s oq£ "ou!I-UO solq~!a~A £l!lenb poanseom £Imonbax3u! gugo!poad aoj (L86I "I~ lo ~snopuel!nO) aol~tu!lso l~!luoaoju! OA!ld~pe ue jo £1.qeo!lo~adoql OI~!lSOAU[Ol posn s! Iopom oq.l. " ~ p I~lUOtU[aodxool pond pug podoIOAOp uooq seq amp oouop!soa pu~ otunloA og!oods '£ihouo leO!ueqoom o~!oods 'uo!~s!u!lelOg lonpoad solqe!a~A £1!Ienb ot~ gu!m!poad ssoooad oq~ jo iopom o.n,ueu£p ol!leSaOAV "omoqos IOaUOO ssoooad ou!I-UO ue jo uopmuomoldm! pue luomdoloAOp oql spa~A~O1 gu!p~oI SSOOOaduo!srulxo pooj e jo s.tsLl~u~ ~tlUOOmi~uostuols,~s ssoooad oql uo suodo.~ uo!mq!xmoo s!q,L - laaalsqv lho%)ln'oe" Iou'uo.uolWdnq E6Eg E~E I6I 1717+:xed E17ELEEE I6I Ve+ :qd )In'oe'IOU@S!lI!~'~l~m 2,I_C1'I'I~IL I~a'q ou££ uodn OllS~O~oN jo £:LtSaOAl.Ufl ~u!aoou!~uH ssoooad pue l~O!tuoqDjo 'ldo(I dnoaD qoa~osoH uo!lus!tugdO a!Ioqtu£s+

17¢8E ICg I9+ :x~d g~¢E I ~ E I9+ :qd oP'0017p'~mqauq-m@ uosm~quodo.u nwnpo'p£sn'~uo'tuoqa@ {um.mq 'uoq '£OSlO} ~!l~asnv 900~; A~SN £oup£s jo £1!saoA!ufl ~u.uoou!I/uH I~O!tuoqDjo luotu~redo(I,

+SlqqlA~ NHVIAIpu~ ,NOZHVfl "AS.X~dd._qOBO ',XV)IOlAI NHfl ',NHSflVHNHdHI'8 DHOf ',XHSqH NI&Sflf

ssoaoad uo!snalxx poo I u jo lOalUOD puu U!lOpOIAI 00"0+00"LI$ L6/I'~I-8600 9-~/,000(/,6)1~$£I'8600S:IId u.~ult l~arrD u! poluud po~asoa slt#u IIV p~q aaua!~S aO!AOSI~[/.661 O L66[ '99£S- 19gS 'dd"|ddn S ' I E "IOA'3u3uE -ulat/a ,¢.lasndmoD

uomBBaad

$362

PSE '97-ESCAPE-7 Joint Conference

predictive controllers are compared. Finally, conclusions are drawn and further work outlined. EXTRUDER MODELLING PROCESS DESCRIPTION Extrusion is a shaping operation in which a material is pressurised by some means to force it through a die. The material is generally a solid at ambient temperature, and extrusion usually requires processing the material at a higher temperature, under which the material softens or melts to facilitate flow. A typical extruder consists of a barrel inside which one or more helical screws rotate to propel the feed material towards a die opening at the discharge end of the extruder, as illustrated in Figure 1.

FordHopper Be'el

~

Eleraentl

~1 i i l l l l l l l l l Y l l l l l ll,~.~,., ~.~Ji)l i i i ii i i i i i i i i iiiil".-i.

/////////////////II Figure 1: Schematic representation of a co-rotating twin-screw extruder.

The high pressure generated during this process forces the material to exit the extruder through the die. At the same time, heat is generated due to friction and the material is exposed to high rates of shear stress. This usually leads to some kind of material transformation and determines the shape, size and texture of the extrudate. MECHANISTIC MODELLING The steady state model presented by Kulshreshtha et al. (1991) and its dynamic version, Kulshreshtha etal. (1992), represent the most recent one-dimensional mechanistic modelling approach for twin-screw cooking extrusion. Both the steady state and the dynamic model are based on elemental heat and energy balances. Given the screw speed, feed rate, moisture content, feed temperature and barrel temperature profile this model calculates the temperature and pressure profiles along the barrel and the overall shaft-power required.

The model has been implemented in the MATLAB software environment with several modifications and extensions for the purposes of this project: 1. The model solution schemes and the numerical techniques suggested by Kulshreshtha et al. (1991,1992) are unnecessarily complex and inefficient, requiring repeated iteration to calculate the position of the moving boundary between each solids conveying zone (SCZ) and melt zone (MZ).

Instead, the partial differential equations describing mass and energy balances down the length of the extruder have been discretised and solved using a finite element technique. This is equivalent to dividing the extruder into a number of continuously stirred tank reactors (CSTRs) - with back-flow occurring in the MZ elements - connected in series. Using this method, the SCZ-MZ interface is simply determined, occurring where the CSTRs become completely filled. 2. The above discretisation procedure has positive implications for more accurate flow modelling. Kulshreshtha's basic model for simplification purposes assumes that a twin-screw extruder has plug-flow behaviour. The leakage back-flow between consecutive C-shaped chambers is simply treated as a reduction in forward flow, which is an unrealistic simplification (Jager et al., 1992). The discretisation procedure allows the separation of the forward flow (due to positive displacement by the screws) and the back-flow (due to leakage around the screw flights) leading to a more accurate flow representation. 3. The model has been extended to include variables which may be helpful for describing product quality. Product gelatinisation, specific volume, specific mechanical energy, and residence time are now predicted by the model. The methods employed to extend the model are described in more detail below THE GELATINISATION MODEL Starch gelatinisation during extrusion is a complex reaction for which the mechanisms are not fully understood. A review of the literature revealed that there have been only two different modelling approaches investigated, those suggested by Wang et al (1992) and Cai and Diosady (1993). Although it is not possible to conclude which of these approaches is the more appropriate at this stage, the model suggested by Cai and Diosady has been validated over a wider range of operating conditions, and is therefore the model chosen for our implementation.

Examples of the dynamic response of starch gelatinisation to changes in feed rate, screw speed and moisture content are shown in Figures 2(a) to 2(c), where inverse response characteristics as well as process dead-time can be observed.

PSE '97-ESCAPE-7 Joint Conference F = 20 to 30 [ kg/h ] 0.8.

0.75

=o

0.7

.+o .c

0.65

0.6

10

15 20 Time [ s ]

25

30

$363

The predicted residence time distribution of an inert tracer was compared with experimental RTD data obtained by injected sodium chloride at the feed port of the extruder. Figures 3(a) and 3(b) are a comparison between the predicted and measured concentrations of the tracer in the extrudate over time. It can be seen that increasing screw speed has the effect of reducing the "dead-time" of the process, but has little effect on the variance of the distribution. Increasing feed-rate, however, has little effect on the dead-time of the

(a) N = 400 to 500 [ rpm ]

RTD Curves, N=300 [rpm]

1

0.08

0.95

/ ....

0.9 .,=

0.85

~

0.7

0.6

0

5

10

15 Time[s]

20

25

~ i

o.o5t o.o4

°¢

0.03

0.75

0.65

.',

.... r •-

0.8 .~

I

,, ,,=1

i

i ;,',.

i

0.02t

i

'

"

30 u

(b)

0

M = 0.2 to 0.3 [ kg/kg ]

10

20

30

40

O''O--

50

60

Time [sl

O.E

(a)

0.7

RTD Curves, F=10 [kg/h]

~,~ 0.61

0.05

..,¢

N=200 [rpm] model N=200 [rpm] experimental N=300 [rpm] model N=300 [rpm] experimental

0.5i 0.04 :c-

0.4:

1~

0.3

0.2

F--10[kg/h] experimental F=20 [kg/h] model F=20 [kg/h} experimental

I + I- o

0.03 0

5

10

15 20 Time [ s ]

25

30

(c) Figure 2: starch gelatinisation trajectories as a function of changes in operating conditions.

0.02

i° 0.01

"-~..

01

10

20

30

40

50

60

Time [s]

Although the gelatinisation model parameters have yet to be determined experimentally for the cooking extruder being studied, parameters were estimated to match reported values of gelatinisation in similar extruders.

RESIDENCE TIME DISTRIBUTION (RTD) The effect of the residence time distribution on product quality is an important consideration, as a high variance for the RTD implies a highly inconsistent degree of cooking for the product. For this reason, it is advantageous to be able to predict the RTD for a given set of operating conditions, so that the extruder can be operated in such a way as to produce a homogeneously cooked product.

(b) Figure 3: Experimental and model predicted RTD curves. process, but leads to a narrower distribution. The excellent fit obtained between the experimental and model predicted RTD curves might be interpreted as further evidence of the general applicability of the model developed here.

EXTRUDATE EXPANSION Extruders are often used to produce puffed products such as breakfast cereals, flat breads and snacks. For such products the expansion ratio is clearly an important quality variable. The model of Fan et al. (1994) which describes the dynamics of bubble growth in starchy extrudates has been implemented in the

$364

PSE '97-ESCAPE-7 Joint Conference

general extruder model to provide predictions of the expansion ratio and bubble size.

A comparison of the model prediction experimental data is presented in Figure 4.

RHEOLOGICAL MODELS

with

x 1012

Dough mixes are non-Newtonian fluids, and their rheological behaviour is quite complex• The selection of a suitable theological model is crucial to the validity of the resulting extruder model, as it directly influences the flow behaviour, heat generation, and pressure development within the extruder. In a previous paper (Elsey et al., 1996) two different power law model structures (those of Kulshreshtha et al., 1991 and Vergnes et al., 1987) were compared for their accuracy at describing experimental extrusion data. However, the power law models that are typically applied often do not have a firm theoretical basis, and are therefore simply correlations of a form that appear to fit experimental data over a reasonable range of conditions.

8

'o ~'" ..••.""

7 ..'"

~ ~E ~. ~ ~ ~ :E

o

.•.•"

5

o

• - • •°~o

4 o

3 2 1 o

oo

°"

O °° •' '*~

o ...'•'" ...••• °O o i

i

i

i

i

i

1

2

3

4

5

6

Experimental eta/Bk [Pa.s.m-4]

8

x 1012

Figure 4: Comparison of viscosity model prediction From a process systems engineering point of view, the application of a non-linear data-based modelling technique is a much more appropriate approach to this problem. Genetic Programming (GP) (e.g. McKay et al., 1996) is a novel data-based modelling technique which allows the generation of a suitable model structure and determination of the model parameters simultaneously. Hence the accuracy of model predictions is no longer dependent on a pre-assumed model structure, Experiments were performed on an APV-Baker MPF 40 twin-screw cooking extruder to determine the pressure drop over a short section of the MZ. As this pressure gradient is directly related to the melt viscosity 1"1, leakage back-flow parameter Bk, and the conditions of the melt (temperature T, screw speed N, moisture content M), a model relating melt viscosity to the melt conditions could be developed. The GP algorithm was used to evolve 65 models for the melt viscosity. The resulting expressions were narrowed down to three which were outstanding on the basis of each having a correlation coefficient greater than 0.95 on the training data. After determining that the model residual errors were of zero mean and normally distributed, an F-test confirmed that neither of these three models was significantly better than the others, at the 95% confidence interval• The simplest model from this set was then chosen as being the most appropriate, yielding:

1"1= [-7.7718X10"]T+[1.195X101°]MT2 +.[1.64.1x10"1

B,

N

NM

+[.2.5184×IO"]M+[1.441×IO~]NM'+[2.7367x1012]

(1)

with experimental data.

It is interesting to note that this model form is efficient in that it contains only six parameters, comparable to the model of Vergnes et al. (1987) which employed seven• Perhaps one advantage of this modelling approach is that other process variables which are not typically considered in extrusion rheological models can be included• For example, if an appropriate sampling valve were used to collect samples of the melt, the degree of cook could be determined and included in the model for melt viscosity• Considering that typical power law rheology models require fitting of their parameters to specific feed materials in any case, there are clearly advantages in evolving an entire rheological model from this experimental data. Having developed a process model that compared well with experimental observations, process control schemes can now be investigated. ADAPTIVE INFERENTIAL EESTIMATION Typical control variables for a cooking extrusion process are product temperature and die temperature• These variables are typically chosen because they give some indication of the state of the process and are easily measured at rates suitable for on-line process control. Ideally however, real quality measures would be the control variables, such as degree of cook, mean residence time, or product expansion ratio. Unfortunately, the feasibility of on-line measurement of these variables is limited as either the instrumentation does not exist or the analysers require long cycle times. The resulting delays would prevent early detection of the effects of load disturbances, resulting in degraded process operation.

PSE '97-ESCAPE-7 Joint Conference Estimators that are capable of alleviating the problem of large measurement delays and irregular sampled feedback have been previously developed (Guilandoust et al., 1987; Tham et al., 1992). Since the primary controlled variable is usually related to other process outputs, the estimators make use of these secondary outputs to infer the state of the primary output each time the secondary outputs are measured. The estimators are implemented within an adaptive framework to ensure their applicability to time varying processes. The parameters of the algorithm are estimated whenever measurements of the primary output become available. As a result, estimates of the primary output are obtained at the faster rate at which the secondary outputs are measured. To assess the feasibility of this adaptive inferential estimator (AIE) for applications to extrusion control, the AIE algorithm of Guilandoust et al., (1987) was used to provide estimates of the product gelatinisation fraction of an extruded starchy material, as simulated by the extruder model described earlier in this work. Product gelatinisation fraction of extruded starchy products can be determined off-line using a rapid viscoanalyser. A typical analysis takes approximately 20 minutes, and thus this measurement would normally be unsuitable for process control purposes. Figure 4 shows the simulated response of product gelatinisation fraction when the extruder model inputs screw speed, moisture content and feed rate are stepped. The estimated output of the AIE algorithm is also shown, and clearly a reasonable estimate of the primary output has been obtained.

$365

PROCESS CONTROL A preliminary investigation of the suitability of PI and model predictive control (MPC) for single-input-singleoutput control has been performed. Simulations were performed using in turn moisture content, feed rate and screw speed as the manipulated variable to move the degree of gelatinisation between set-points. Results of these simulations are shown in Figure 5. It can be seen that the best control performance was obtained using screw speed as the control variable within a MPC scheme. The use of moisture as a control

Controlling Product Gela|inlaatton Fraction b y Manipulating Moisture Content

~ o.rs o,7

I 20

40

Controlling

~0

Product

80 Time (*)

G elatlnlaatlon

~00

Friction

120

140

180

by M | n l p u l t t l n g

Feed R a t e

I

os

jo. 07

ot

20

40

no

eo

100

*20

1,0

Tim • Is) Controlling

Product

Getatlnllatlon Fraction Screw Speed

by M a n i p u l a t i n g

0.9 i

0.8 0.7

,,,-

1

0.e~ oeo

0.6

~

8

0.5

"~

0.4

~ 0.3 ~ 0.2 ~ 0.1 0 0

07S 070

P

o.ss

zc

4o

eo

s

1oo

12o

1 o

Tim • (m)

,--.o . . . . . . ol

! s;0

Figure 5: A comparison of different strategies for controlling product gelatinisation fraction. 600

T i m e [s]

Figure 4: Estimate of product gelatinisation fraction is plotted with actual (simulated) values. With frequent estimates of infrequently measured quality variables now available, the control of an extrusion process using these quality variable estimates becomes feasible.

variable was unsuccessful as the range of variation required in the set-point changes could not be achieved via moisture content manipulation. In the case of using feed rate as the manipulated variable, PI control could not be used as the product gelatinisation fraction exhibits an inverse response to changes in feed rate In general, the performance of either moisture content or feed rate as the control variable was far from ideal due to the presence of dead-times between the control

$366

PSE '97-ESCAPE-7 Joint Conference

action and the process's response. Clearly though, the performance of MPC is far superior to that for PI control. Such a result is expected as MPC has the advantage of having an internal model of the process to use in optimising its control action. CONCLUSIONS An outline of the process systems engineering techniques relevant to the modelling and control of a food extrusion process has been presented. Emphasis has been directed towards the predicting and controlling realistic quality variables such as product gelatinisation fraction in preference to secondary process indicators such as temperature and pressure. A flexible dynamic model of the extrusion process has been developed, rheological modelling investigated, an adaptive inferential estimator for product ge!atinisation fraction developed, and simple control strategies investigated. It is expected that the control of real quality variables will result is a more reliable control scheme. This is the subject of current research. ACKNOWLEDGEMENTS The authors would like to thank Dr Ming Tham for his assistance with the adaptive inferential estimation algorithm implementation, and the CSIRO Department of Food Science and Technology Australia for providing the extrusion data used in theses studies. REFERENCES

Cai, W. and Diosady, L.L. (1993). A model for gelatinisation of wheat starch in a twin-screw extruder. J. Food Sci. 58:872-875, 887. Guilandoust, M.T., Morris, A.J. and Tham, M.T. (1987). Adaptive inferential control, lEE Proceedings, 134d-3, 171-179. Elsey, J., Riepenhausen, J., McKay, B., and Barton, G.W. (1996). Dynamic modelling of a cooking extruder. Chemeca 1996, Sydney, Australia. Fan, J., Mitchell, J.R. and Blanshard, J.M.V. (1994). A computer simulation of the dynamics of bubble growth and shrinkage during extrudate expansion. J. Food Eng. 23, 337-356. Jager, T., van Zuilichem, D.J., Stolp, W. (1992). Residence time distribution, mass flow, and mixing in a co-rotating, twin-screw extruder. In Food Extrusion Science and Technology, J.L. Kokini, C.T. Ho and M. V. Karwe, Eds., Marcel Decker, New York, pp. 165-176. Kulshreshtha, M. K., Zaror, C. A., Jukes, D. J. and Pyle, D. L. (1991). A generalised steady state model for twin screw extruders. Trans IChemE, Part C, Food and Bioproducts Proc., 69 (C4): 189199. Kulshreshtha, M.K. and Zaror, C.A. (1992). An unsteady state model for twin screw extruders. Trans IChemE, Part C, Food and Bioproducts Proc., 70 (C4): 21-28.

McKay, B., Willis, M.J., Barton, G.W., (1996). Steadystate modelling of chemical process systems using genetic programming. Accepted for publication in Comp. and Chem. Eng. Tham, M.T., Montague, G.A. and Morris, A.J., (1991). J. Proc. Cont., 1, 3-14. Wang, S. S., Chiang, W. C., Zheng, X., Zhao, B., Cho, M. H. and Yeh, A. (1992). Application of an energy equivalent concept to the study of the kinetics of starch conversion during extrusion. In Food Extrusion Science and Technology, J.L. Kokini, C. T. Ho and M. V. Karwe, Eds., Marcel Decker, New York, pp. 165-176.