Vous êtes sur la page 1sur 20

Applied Mathematical Modelling 30 (2006) 884903

www.elsevier.com/locate/apm

CFD-based predictive control of melt temperature in plastic injection molding


A.G. Gerber *, R. Dubay, A. Healy
Department of Mechanical Engineering, University of New Brunswick, P.O. Box 4400, Fredericton, Canada E3B 9N4 Received 1 August 2004; received in revised form 1 May 2005; accepted 27 June 2005 Available online 26 August 2005

Abstract A unique method of coupling computational uid dynamics (CFD) to model predictive control (MPC) for controlling melt temperature in plastic injection molding is presented. The methodology is based on using CFD to generate, via open-loop testing, a temperature and input dependent system model for multi-variable control of a three-heater barrel on an injection molding machine. Results clearly show the benet of temperature and input dependent system models for MPC control, and that CFD can be used to dramatically reduce the time associated with open-loop testing through physical experiments. 2005 Elsevier Inc. All rights reserved.
Keywords: Computational uid dynamics; Model predictive control; Plastic injection molding; Melt temperature prole

1. Introduction In the manufacture of plastic components using injection molding, online process optimization is imperative for minimizing energy consumption, part rejects, machine setup duration, and obtaining and maintaining high product quality. Its importance is vital when producing high

Corresponding author. Tel.: +1 506 453 4513; fax: +1 506 453 5025. E-mail address: agerber@unb.ca (A.G. Gerber).

0307-904X/$ - see front matter 2005 Elsevier Inc. All rights reserved. doi:10.1016/j.apm.2005.06.001

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

885

Nomenclature A element face area (m2) G open loop response J sum of controlled variable errors to be minimized N prediction horizon P plant model dynamic matrix Q heater load (W/m3) S volumetric source term (in energy equation W/m3) T temperature (K) TSP set-point temperature prole (K) V control-volume (m3) a CFD coecients of linearized equations 1 A(z ), B(z1) ARX regression coecients e controlled variable error (K) specic-heat at constant pressure (J/kgK) cp h enthalpy (J/Kg) k thermal conductivity (W/mk) or dimensionless scaling factor (Eq. (12b)) _ m melt mass ow (kg/s) control horizon nu q control moves (percent heater load) (%) t simulation time (s) u, v, w uid (melt) velocity (m/s) u manipulated variable y controlled variable response Greek / q C k symbols dependent variable in CFD solution density (kg/m3) diusion coecient control move relaxation

Subscripts/superscripts P central node in computational molecule SS steady-state value nb neighboring nodes in computational molecule o indicates old time level in CFD solution j open loop process k time instant m manipulated variable (heater inputs) r temperature regime

886

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

precision components for applications such as optical lenses, medicinal and pharmaceutical devices, prosthetics, micro-machines, solar-energy systems, sensors, and in many electrical and electronic devices. Consequently, research on developing adaptive controllers is essential for process optimization and is a natural candidate for achieving better control performance. An example of a system that would benet from adaptive control is the heating of a barrel on an injection-molding machine (IMM), which is typically divided into individually controlled heater zones. The heating zone at the rear of the chamber near the hopper, serves to pre-heat the plastic, while the center heating zone ensures the polymer is fully plasticized, and the front zone is controlled to keep the material at the desired melt temperature required for injection [1]. Several researchers [24] have investigated the use of conventional three-term PID controllers, including Dahlin and Smith predictor control schemes for controlling melt temperature. Deterministic and stochastic models were found by manipulating the heater power applied to the heater zones independently and recording their response. The barrel heating dynamic characteristics such as time constants, process gains and dead time were obtained from the numerous open-loop tests performed. In general, PID control resulted in large overshoot and settling times, as well as substantial oscillations in the melt temperature. The Smith predictor, which included a dead time compensator reduced the oscillation in melt temperature as compared to the PID control, however overshoot and process damping remained a problem despite exhaustive open-loop testing. During plasticization the polymer within the barrel is a continuous medium, which interacts with the heater zones by direct contact with the wall of the barrel. Kamel et al. [2] demonstrated interaction between zone melt temperatures. This indicated conductive heat ow between the barrel and nozzle heating zones as well as heat transfer through the barrel casing. The models derived for melt temperature control simulations neglected these interactions and were therefore single input single output (SISO). The resulting temperature oscillations during closed loop control with SISO [25] can result in thermal deterioration in the polymer and longer operator setup times on the IMM. The formulation of advanced controllers such as ones derived from the predictive control methodology [510], as well as conventional approaches [24], all depend on open-loop testing for model identication. The development of multi-inputmulti-output (MIMO) predictive control for melt temperature was conducted by Dubay et al. [11,12], incorporating the interactions between the barrel zone temperatures. This multi-variable control including process interactions resulted in better control performance when compared to PID based control methodologies. The heater zone interactions were determined experimentally by applying a step input to each individual zone, and capturing its response as well as any eects sensed by the neighbouring zones. This lengthy experimental procedure captured the dynamics of polymer melting at only one operating regime, resulting in a process model that was not temperature or input dependent. The major drawback was the inability to redene the controller parameters for changes in the processing conditions and polymer. Negligible oscillations were achieved for set-point changes by implementing suppression factors on the manipulated variable, which was the electrical power to the heaters. Computational uid dynamics is a powerful form of computer-based simulation used in many industrial applications to capture important process characteristics and responses [13]. It is based on solving (for this application) conservation equations for mass, momentum and energy to

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

887

determine the spatial and temporal distribution of ow velocity and temperature while retaining much of the geometric complexity. A few researchers have begun to utilize CFD with process control [1416] to improve the controller performance. CFD has the ability to model the spatial and temporal complexity of uid ows while process control ensures that processing conditions are tightly followed even when dynamic changes occur within a system. Designing a controller however, requires knowledge of the process or a system model for evaluating the controller parameters and conducting simulations, which is where CFD becomes valuable as a modelling tool. As an example Peng and van Paassen [14] uses a CFD simulation to predict the temperature distribution for a room divided into ve zones. The zonal temperature responses are then converted to a state space model used to control the temperature of the rooms air supply. Similar investigations by Bezzo et al. [15] demonstrated how CFD and process simulation complement one another in control of a semi-batch reactor. In their study, CFD is used rst to compute the ow eld for a semi-batch reactor, and this information is then used for process simulation. Two commercial software packages are interfaced, and continuously exchange critical process parameters during the simulation. The result is the ability to harness updated or current process characterization to conduct simulation in order to better control the batch reactor. In another study Yang et al. [16] conceptually proposed a scheme to use CFD simulation results to generate a database of predictions of various processing situations, which can then act as a soft sensor for process optimization and control. The CFD model was used as an o-line tool to capture the physical/chemical process and dynamic behaviour in a rotary kiln for chemical waste incineration. Their research was motivated by the lack of information and variability of the input waste composition and the inability to sample process conditions directly inside the kiln. To the authors knowledge, a CAE tool such as CFD, has not been used to dene a temperature and input dependent process model for predictive control of the melt temperature in the barrel and nozzle regions of an IMM. This paper presents a unique strategy for integrating CFD to a predictive controller on a three-heater zone IMM. A CFD model of a three-heater zone IMM will act as a slave to the MPC algorithm by providing the necessary system identication data required to design the controller and conduct control simulations. This paper presents some important theoretical aspects of CFD and MPC, and evaluates the CFD modelling ability to generate a process model of the three-heater zones. The paper also presents a new architecture for the CFD-based controller and investigates the improvements of the CFD-based controller over a wide range of operating regimes.

2. Computational uid dynamics A CFD model was created to represent the heating of the polymer and the physical components on an actual Engel reciprocating screw IMM with a 34 g shot size and 222 KN clamping force. The model includes the screw, barrel, three electrical heater zones and the polymer melt regions. For this study, screw rotation or translation is not applied and the uid region is held stagnant or specied with a xed mass ow obtained via an initial CFD solution. The simplication of using a xed mass ow (which implies a steady melt velocity prole) of the melt was for the purpose of reducing computational time in the transient simulations. The CFD model used is fully functional

888

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

in that by changing input parameters, uid ow can be simulated dynamically in time. In addition the melt viscosity was held constant, no viscous heating and no phase change model was included in the melt region. While the CFD methodology is general enough to include these eects, they were not deemed necessary for the purpose of demonstrating the CFD-based controller methodology. The methodology in this case emphasizes the heater zone interactions. 2.1. Geometry and boundary conditions CFX-TASCow is used for the CFD simulations. It uses a nite-element/nite-volume approach for discretizing the governing mass, momentum and energy equations for the uid region and the energy equation for the CHT (conjugate heat transfer) solid domains [17]. A three-dimensional representation of the geometry is shown in Fig. 1a. In this wedge shaped geometry it is assumed that the temperature distribution is axisymmetric. Two electrical heater bands encase the barrel zones while a third heater is on the nozzle section. The heater bands all act independently

Fig. 1. (a) CFD geometry with heater locations and monitoring points and (b) melt region.

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

889

and are driven by the predictive controller (in closed loop mode) to provide a variable volumetric heat source. All uid and solid regions shown in Fig. 1a are implicitly coupled during the solution and divided into a grid comprised of linear nite-element volumes (ux elements). Three temperaturemonitoring points are located at the center of each heater zone, situated one grid spacing below the barrel (or nozzle) surface into the melt region. The melt region of the CFD geometry is indicated in Fig. 1b. The boundary conditions involve a specied mass ow (0 and 2.54 104 kg/s for the two cases considered) and temperature (300 K) at the melt inlet. The outer boundaries of the barrel were treated as adiabatic stationary walls in consideration of the insulation applied to the actual physical machine. Symmetry conditions were applied at the faces of the wedge geometry based on the assumption of axisymmetric conditions. All interior walls (along the barrel, nozzle and screw) are treated as stationary no slip walls. Finally the barrel and screw regions of the model are given the material properties of plain carbon steel. The electrical heater bands on the actual machine that the CFD model represents are made of a refractory material (which are wrapped in insulation). General purpose polystyrene is used to represent the melt, or uid region. Contact resistances between the heaters and barrel are not considered, but could be added in a subsequent CFD model of the IMM. 2.2. Governing equations The transport of a uid property in unsteady ow can be represented by a convectiondiusion equation of the form: oq/ divq/u divCgrad/ S / ; 1 ot where the uid property, /, represents quantities such as u, v, or w velocity components or energy which can be arranged in terms of temperature (T) in the solid region or enthalpy (h) in the uid region. The diusion coecient, C, represents molecular transport properties such as dynamic viscosity and thermal conductivity. The rst term of Eq. (1) is the transient term, representing the control volumes rate of change of / with time. The second term captures the transport by convection of the property, /, out of the uid element. This term is not present within the solid regions of the domain (heater bands, barrel and screw) where the velocity is zero. The third and fourth terms represent the spatial change in / due to diusion and any volumetric sources respectively. In the present case, enthalpy h is used to describe the thermal energy transport (i.e. / = h) in the uid and the temperature at each node is obtained by inverting the relationship: Z T cp dT : 2 h
T ref

As only temperature can be diused (1st term from RHS of Eq. (1)), the diusion coecient for the liquid region takes an approximate form of: C k ; cp 3

890

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

where k and cp are the thermal conductivity and specic heat, respectively. In the solid regions / = T and C = k, and there is no convection term as already mentioned. All material properties for the energy equations (k, cp and density, q) can be determined as a function of temperature, however, for the purpose of this study, have been taken as constant and chosen appropriate for the temperature regimes considered. To obtain the discrete equations CFX-TASC ow performs a conservative control volume integration of Eq. (1) at the vertices of each nite element in the domain. A discrete conservation of mass, momentum and/or energy statement results for every control volume, which when coupled to its neighbours leads to a system of equations that can be solved for the momentum or enthalpy distribution in the uid region (or temperature in the solid). Applying Gauss divergence theorem to the convective and diusive terms of Eq. (1), volume integration results in: Z   Z tDt Z o n qu/ dA dt q/ dt dV ot t t CV A  Z t D t Z Z tDt Z n Cgrad/ dA dt S / dV dt; Z
tDt t A t CV

where An adA represents the surface integral of the conserved quantity a (i.e. qu/) normal to the control volume surface. The time domain integration is evaluated using a rst order backward Euler scheme. Completing the nite volume integration process for Eq. (4) results in a discrete equation representing all important transport processes in and between control volumes [18]. The general discretized form of the unsteady conductiondiusion equation in nodal coecient form is: X o anb /nb ao 5 aP /P P /P S / ; where aP X anb ao P 6 7

ao P q

DV : Dt

In Eqs. (5) and (6) the nb subscript denotes the neighboring east, west, north, south, top and bottom coecients relative to a hexahedral volume element. Each set of a nodal coecients represent a balance of the convective mass ux per unit area and the diusion conductance at the neighboring control volume face; details can be found in [13,18]. The superscript o refers to the uid property at the previous time step. The source S/, within the heater regions, refers to the applied volumetric heat source re-evaluated at each time step by the control strategy in closed-loop mode or changed in a prescribed manner in open-loop mode. Details on how S/ is calculated will be discussed later. The implicit scheme used by CFX-TASC ow is rst-order accurate in time, nominally second-order in space, and unconditionally stable for any time step, Dt. At each time step of the transient solution, the system of nodal equations (5) is solved referencing the previous nodal solution.

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

891

3. Model based predictive control Since its development and introduction in forms such as model predictive heuristic control [19,20], and as dynamic matrix control [7,21], MPC has gained widespread use and acceptance both in academic research and industrial applications [10,2225]. MPC is a strategy that uses a dynamic process model to predict the process response trajectory over a prediction horizon N, to past and future control moves in order to minimize an error vector. The error vector e is determined over horizon N as the dierence between the process prediction and its set-point trajectory. The minimization of these sum of errors J, subject to process and manipulated variable (electrical heater power) constraints, provides a set of control actions (current and future) which are to be inputted to the electrical heaters on the IMMs barrel. The general structure of MPC is shown in Fig. 2 [26]. A dynamic matrix P constitutes the plant model used for evaluating the future process response and the control moves Dq. 3.1. Multi-step testing and process prediction MPC uses two horizons, a control horizon nu and a prediction horizon N, in order to evaluate a prediction of the controlled variable output at discrete time instants. The value of N is chosen based on open-loop settling times determined by exciting the process to be controlled by a step or an impulse in the manipulated variable. In this paper, 95% of the time to reach steadystate is used to evaluate N which represents the number of discrete sampling instants to reach this time. In the case of MIMO control, the multi-variable dynamic matrix P shown in Eq. (8) containing the dynamic responses of each process 1 to j that are aected by any of the m manipulated variables or controller outputs over the horizons nu and N. The subscripts on the open-loop response coecients in P represent the process to the controlled (1, 2, . . . , j), and the location on N. The rst of the two superscripts indicate the region of operation, and the second indicate which heater input was active to obtain the process output responses, during open-loop testing.

Fig. 2. MPC structure.

892

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

2 b 1 ;1 p 6 11 6 b 1 ;1 6 p12 6 6 6 6 6 6 b 1 ;1 6 p 1N 6 6 6 6 b 6 P 6 6 6 6 6 bj ;1 6 pj1 6 6 6 bj ;1 6 pj2 6 6 6 6 4 pjN


bj ;1

0
b 1 ;1 p11

0 0

b 1 ;m p11 b 1 ;m p12

0
b 1 ;m p11

. . . p 1 ;N 1
b 1 ;1

. . . p1;N nu 1
b 1 ;1

p 1N . . . . . .

b 1 ;m

p1;N 1

b 1 ;m

0 pj1 . . . pj;N 1
b j ;1 b j ;1

0 0

b j ;m

pj1 pj2

0
b j ;m

b j ;m

pj1 . . .

p1;N nu 1

b j ;1

pjN

b j ;m

pj;N 1

b j ;m

7 7 0 7 7 7 7 7 7 7 b1 ;m p1;N nu 1 7 7 7 7 7 7 7 7 7 7 7 0 7 7 7 7 0 7 7 7 7 7 5 pj;N nu 1
bj ;m j:Nxm:nu

In Eq. (8), the dynamic responses can exhibit variable dead time corresponding to each of the m inputs, which would result in rows of discrete zero values in the dynamic matrix. For convenience, dead time is not shown in the matrix. This dynamic matrix with its (j m) partitioned blocks can be compactly expressed without the extensive subscripting as: 3 2 G11 G12 G1m 7 6G b 6 21 G22 G2m 7 7: 6 9 P 6 .. . 7 . 5 4 . . Gj1 Gj2 Gjm The commas on the subscripts and the superscripts are removed for convenience, with the rst subscript indicating the process, and the second the heater input m during open-loop testing. In this study, for the number of manipulated variables and the process to be controlled, m and j are equal to three. MPC minimizes J for the changes in heater input at each of the barrel zones. Details can be found in [12] while the changes in heater input Dq at time instant k can be expressed as: 1 Dqk P T P kI P T ek : 10 The weighting parameter k relaxes the control moves if desired and is set to zero in this investigation. The sequence of the m control actions resulting from the minimization of J is given as: T 11 Dq Dq11 Dq12 . . . Dq1nu . . . Dqm1 Dqm2 . . . Dqmnu ; where only the rst element of each of the m control sets is sent to each of the j processes or heating zones. The remaining control moves in each of the m sets are used for predicting the future behaviour of the heating process of each barrel zone.

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

893

4. CFD/MPC coupling and system identication This research combines two disciplines; MPC and CFD, to improve polymer melt temperature control within the barrel and nozzle region. The strategy is tested considering a three-heater zone injection-molding machine. A CFD model of a three-heater zone IMM will act as a slave to the MPC algorithm by providing the necessary system identication data required for process modelling and control simulation. Presently, it is dicult to use a CFD model concurrently with process control to drive an online process. The computation time to simulate one open-loop test is much too long for direct control purposes but much faster than experimental open-loop testing, which can take up to 8 h in time. Therefore, one challenge is how to eectively integrate the CFD model with the process controller. The CFD predictions need to be linked to the actual predictive controller without jeopardizing the time step of the desired control actions to the controlled variables. In practice the open-loop simulations would be conducted separately, and the responses (P-matrix) stored in a database from which the MPC algorithm can extract the dynamic matrix most appropriate for the immediate temperature/input regime. It should be noted that the simulations conducted in this work are all open-loop simulations followed by closed-loop responses to a new set-point regime, avoiding any access to a database. This sequence was repeated for any number of set-point change scenarios as part of one complete simulation. The combining of the open and closed loop simulations into one large simulation was for convenience in evaluating the concept. 4.1. CFD simulation strategy As mentioned previously, MPC requires a process model generated from open-loop testing to optimize the inputs applied to the heating zones. In the present example the system model is obtained by capturing the transient response of the CFD model subjected to open-loop testing. For this study, three-zone melt temperatures are the controlled variables with the three electrical heater inputs as the manipulated variables. The melt temperature at each zone is monitored at nodes located in the uid domain, one grid control volume beyond the barrel-uid interface as previously depicted in Fig. 1. At present, open-loop tests to each heating zone in succession are simulated, followed by closed-loop predictive control to converge to a new set-point regime with the CFD model acting as a plant. While in open-loop control, each heater zone is individually excited with a step-input. Its response as well as the interactive responses felt by the neighboring monitoring points are normalized and stored. The open-loop testing procedure is outlined in Fig. 3, where, for example, G12 is the discrete response matrix for TMP(1) due to a step input applied to heater zone 2. Beginning from a steady-state temperature prole, each zone is individually excited by applying a percent increase to the steady-state heat load. This in turn is a function of the total or maximum heater wattage available. After completing the open-loop tests the simulation switches to closed-loop predictive control, and using the recently stored P-matrices responds to a set-point prole change in melt temperature. At each time step during closed-loop control, the predictive controller provides the CFD model, acting now as a plant, with three distinct volumetric heat sources to apply to three-heater regions in Fig. 1a. This value is a positive or negative percent change, Dq, compared to the total wattage.

894

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

Fig. 3. Open-loop testing procedure to determine P-matrix coecients.

The closed-loop control algorithm developed to maintain or achieve a desired set-point prole is outlined as follows: Retrieve temperature and input specic normalized open-loop step response data for each processpassed from CFD simulation. Map and store the open-loop data into the dynamic P-matrix. Store the pseudo-inverse system matrix (PTP)1. Initialize the predicted proles to the current steady-state value of each process output over the prediction horizon 0, . . . ,N. Set the desired set-point trajectories over the prediction horizon 0, . . . ,N. Sample the current process outputs and adjust the prediction prole over the prediction horizon 0, . . . ,N. Calculate the future errors, e. Calculate the desired control actions in the heater power Dq using the control law of Eq. (10). Constrain the Dq vector and send the rst control action to the three-heater zones of the CFD model. Calculate the process prediction prole PDq. Advance the process prediction prole one time instant. Update any changes in the set-point or reference trajectory and repeat sampling.

4.2. Applying the volumetric heat load The heat load at each heater zone of the CFD model is applied through the source term of the solid region, ST, of the convectiondiusion equation (Eq. (1) where / = T and u = 0). On the actual Engel IMM, the electrical heater bands that surround the barrel zones are each rated at 1100 W while the nozzle heater is rated at 300 W. For the CFD model the total wattage applied to each zone is divided evenly over all of the control volumes within a specic heater region. This

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

895

provides a total available heat load of 15.21 MW/m3 for zones 1 and 2 and 2.37 MW/m3 for zone 3. During operation these maximum heat loads are never reached, and a percentage of these total loads are continuously varied depending on the control algorithm. During open-loop mode, the source term, ST, can be expanded as a function of the simulation time t for each heater zone where S T tj Btj ; Btj 1 k j QjSS ; 0 6 t < N Dt Btj QjSS ; N Dt 6 t < 2N Dt  1  DT r oQj r j kj r ; QjSS oT j
1 1 r Tr DT r j jSS T jSS r1 oQj Qr jSS QjSS ; r1 oT j T r jSS T jSS

12a

12b

12c

and QjSS is the steady-state volumetric power and TjSS the steady-state temperature. Each heater zone (j = 1, 2, 3) is individually excited, by a factor 1 + kj, for N time steps, from a steady-state heat load, and then relaxed for the same duration. The subscript SS refers to the steady-state power or temperature. While one zone is being excited, the source term for the remaining neighboring zones is kept at its previous steady-state value. Open-loop mode is maintained for 6 N time steps to cover excitation and relaxation to all three-zones of duration 2 N time steps, with a sampling time of Dt. The superscript r refers to the temperature regime at which the open-loop tests are simulated. The factor kj is a scaling factor applied to the open-loop heater excitation. It is 1 based on the future set-point change DT r j , and past heater changes in response to previous set-point changes (Eq. (12c)). Fig. 4 has been supplied to highlight the relationship between temperatures and heater loads across set-point changes (or temperature regimes). During closed-loop control the MPC algorithm computes the change in volumetric power to be applied to each heater zone (j = 1, 2, 3) at each time step. In this case, the source term ST for the solid region is expanded as: 13 S T tj Qj t Dt QjMAX Dqj ; where Qj is the volumetric power applied to each zone at the previous time step and Dqj is the applied control action evaluated for the next time step using the MPC controller. 4.3. System identication In order to conduct control simulations, system identication using recursive time-series ARX (autoregressive exogenous) models are formulated from the open-loop data obtained by exciting the CFD model. The ARX models are banded, meaning that their formulation is dependent on the band or region of the process response, and therefore on the size and direction of the heater input excitation. These models simulate the plant dynamic response characteristics and are used during closed-loop MIMO MPC. The ARX model (using the backward shift operator z1) can therefore be expressed as:

896

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

T, Q for zone j
r +1 +1 T jSS , Qr jSS

open loop region


r T jSS , Qr jSS

r+1

r
t=6Nx t

r 1 1 T jSS , Qr jSS

r-1

Closed loop mode, N time steps

time

Fig. 4. Zonal temperature regime change and transitions to open- and closed-loop control.

Az1 y z Bz1 uz:


1 1

14

The ARX regression coecients A(z ) and B(z ) are tted (using Sigma Plot). from the openloop response data generated by the CFD model. For a three inputthree output MIMO system, the melt temperature response at any instant for each zone is now a function of the combined eect of all process heater power manipulations. This is depicted in Fig. 5 and is used in control
Step Input Zone 1 + + G12 G13 +

G11

y1

G21 Step Input Zone 2 + G22 + G23 + y2

G31 G32 Step Input Zone 3 + G33 + + y3

Fig. 5. MIMO system schematic for ARX model.

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

897

simulations to evaluate the magnitude of each zone melt temperature. The ARX models are also used in controller comparisons to the CFD driven predictive controller.

5. Simulation results The CFD model was applied in open-loop mode followed by closed-loop MPC for the two melt ow conditions; zero mass ow and a set mass ow resulting in a conservative steady-state velocity prole. The two cases were chosen to highlight the methodology as well as to show the inuence of ow. During the overall injection cycle, there are periods when the melt in the barrel is stationary and when it is in motion. Before beginning open-loop mode a constant volumetric heat load is initially applied to each heater zone until a steady-state temperature prole is achieved. To determine the P-matrix coecients (Eq. (9)) the heat load at each zone undergoes a step change according to Eq. (12). The perturbation of the heat load is applied to each zone separately until a steady-state condition is satised, and then cooled to its initial condition. During this process the melt temperatures were sampled from the three monitoring points within the melt region; TMP(1), TMP(2) and TMP(3) as shown in Fig. 1a. Fig. 6 shows the initial open-loop melt temperature responses, for all three monitoring points, generated by the CFD model with no uid ow. Fig. 7 compares the open-loop responses at individual monitoring points when a velocity eld is present. In Figs. 6 and 7, zones 1 and 2 are shown to experience strong interactions with each other, while zone 3, being the stiest zone, is not inuenced as strongly by any temperature changes in the other two barrel heater zones if mass ow is

440
Step zone 3

430

Step zone 1 G33 G11

420

G21

Temperature (K)

Step zone 2

410

G32 G32 G22 TMP(3)

400
G12 G23 TMP(2) TMP(1)

390
G13

380 0
1000 2000

Time (min)

Fig. 6. CFD simulated open-loop tests (zero mass ow).

898

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

G11

40 Normalized temperature response (K)

30 Normalized temperature response (K)


Zone 1 open-loop response no fluid flow with steady state velocity profile

G21 Zone 2 open-loop response no fluid flow with steady state velocity profile

30

20
G22

20

G12

10

10

G13

G23

500

1000 1500 Time (min)

2000

500

1000 1500 Time (min)

2000

30
Zone 3 open-loop response no fluid flow with steady state velocity profile

Normalized temperature response (K)

G33

20
G31

10

G32

500

1000 1500 Time (min)

2000

Fig. 7. Open-loop response with zero/xed mass ow for (a) zone 1, (b) zone 2 and (c) zone 3.

not present. The presence of uid ow is shown to cool the melt region so that for zones 1 and 2 the open-loop response is similar but with lower temperatures. For zone 3 (see Fig. 7c), upstream of the other zones and normally quite insensitive to their temperature changes, the response to zones 1 and 2 is greatly increased. This is a result of the transport of energy by uid ow to zone 3, which occurs much faster than the time scales associated with diusion only, as in the case when the melt region is stagnant. The response shown in Fig. 7c clearly indicates that mass ow of the polymer will signicantly modify the interactions between heater zones, and can be readily included in the MPC controller as an interacting disturbance model for larger screw sizes. For the results in Figs. 6 and 7 each zone open-loop test was excited for 25,000 s, with the melt

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903 Table 1 Set-point schedule Zone TMP(1) TMP(2) TMP(3) TSP(1) (K) 393 413 423 TSP(2) (K) 413 433 443 TSP(3) (K) 423 443 433 TSP(4) (K) 443 453 463

899

temperature sampled at each zone every 20 s. This resulted in a prediction horizon, N, of 1250 time steps. After completing the open-loop tests, their responses are mapped into P-matrix with closed-loop predictive control simulations commencing, driving the CFD model to the assigned zone melt set-point temperatures. The CFD-based model predictive controller was then tested over four set-point temperature regimes as outlined in Table 1, in both open- and closed-loop control for two mass ow conditions; zero ow and xed mass ow. Each heater zone was initially excited by a 10% step-input (kj = 1.1 in Eq. (12a)) to generate the rst P-matrix used to converge to the rst set-point temperature prole, TSP(1). Thereafter, the step-inputs were scaled to represent the future set-point change to be experienced by that zone according to Eq. (12b). The discrete open-loop temperature responses are normalized and mapped into the current dynamic P-matrix subsequent to the next closed loop sequence. At each new set-point regime, temperature and input dependent CFD open-loop tests are performed to represent the upcoming set-point change. A complete simulation is shown in Fig. 8 for the CFD-based open-loop tests followed by closed-loop MPC over the specied setpoint regimes as outlined in Table 1. The CFD-based and ARX-model based closed-loop predictive responses are compared in Fig. 9, both acting as the plant. The purpose of this comparison is mainly to emphasise the need for temperature and input dependent process models when dealing with multi-variable, non-linear
460 440 420 Temperature (K)
2 3 1 3 2 2 3 3 2 1 1 3 2 2 3 1 3 1 3 2 2

400 380 360 340 320 300


1 2 3

1 1

Closed loop control (MPC MIMO)

Zone 1 Zone 2 Zone 3

5000 Time (min)

10000

Fig. 8. Complete CFD simulation (zero mass ow).

900
430 425 420 Temperature Response (K) 415 410 405 400 395 390 385
TSP(1)

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903


450 445
Zone 3 TSP(2)

Zone 3

440 Temperature Response (K) 435 430 425 420 415 410 405
Zone 1 Zone 1 Zone 2

Zone 2

400
CFD-based MPC ARX-model MPC

395 390 0 100 200 Time (min)

CFD-based MPC ARX-model MPC

100

(a)
450 445 440 435 430 425 420 415 410
TSP(3)

200 Time (min)

300

400

300

400

(b)
470 465
Zone 3

TSP(4)

Zone 3

460
Temperature Response (K)

Temperature Response (K)

455 450 445 440 435 430

Zone 2

Zone 2

Zone 1

Zone 1

CFD-based MPC ARX-model MPC

CFD-based MPC ARX-model MPC

425 420 0 100 200 Time (min) 300 400

100

200 Time (min)

300

400

(c)

(d)

Fig. 9. CFD-based closed-loop MPC (zero mass ow): (a) TSP(1), (b) TSP(2), (c) TSP(3) and (d) TSP(4).

processes such as temperature control in injection molding. Both the CFD-based and ARX-model responses are tuned similarly. In each case, the MPC controlled CFD simulation response is denoted by a solid line, while the dash-dotted line refers to the discrete ARX-model based on the initial set of open-loop tests performed by the CFD simulation. The CFD-based responses converging to TSP(2), TSP(3) and TSP(4) are using the temperature and input dependent open-loop response data when formulating the updated P-matrix at each new temperature regime, whereas the ARX-model uses a constant P-matrix based on the initial open-loop test. The CFD-based MPC response resulted in slight overshoot while converging to TSP(1), which used the generic response data to formulate the MPC. However, when using the up-dated process model, as was the case for closed-loop control to the second, third and fourth temperature regimes, this overshoot was reduced or eliminated. On the other hand, the ARX-model response becomes more sluggish as the set-point temperatures deviate from the operating regime at which the initial open-loop tests where simulated, emphasising the inherent non-linearities.

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

901

As part of the set-point schedule in Table 1, a negative open-loop response (for zone 3) is mapped into P to control a negative set-point change while neighboring zones undergo a positive change. In closed loop control zone 3 experienced a slight undershoot when subjected to the negative set-point change but responded much more rapidly than the ARX model. Zone 3 is a very sti system, with a very small process gain, as seen in the G33 open-loop curve shown previously. It is also not strongly inuenced by the manipulated variables for zones 1 and 2 (particularly when mass ow is not present). The electrical heater on zone 3 is a smaller capacity heater, however the volume of melt it controls is the same as the in the barrel, controlled by heater zones 1 and 2. Also, the thermal conductivity of the melt is small relative to the heater domain making a negative setpoint change (cooling the melt) dicult when simultaneously heating the melt in the upstream region of zone 2. The set-point schedule TSP(3) is a bit unrealistic for the actual process, however it demonstrates the CFDs ability to capture the process dynamics due to a decrease in the heat load. The main focus has been on the CFD simulation with zero mass ow however, Fig. 10 shows a comparison between two mass ows while converging to TSP(2). As demonstrated previously, when mass ow is present it can change considerably the open-loop response characteristics between the heater zones (see Fig. 7c) due to the convective transport of energy. Fig. 10 shows the dierences in closed-loop control when mass ow in the system is dierent from the quiescent case. In both cases the P-matrix has been computed by CFD previously with the relevant mass ow. As expected the response in each case is good showing only a slight overshoot. However for zone 3, when mass ow is present, the response is more sluggish since zones 1 and 2 have far more inuence through zone interactions now enhanced due to the moving melt. The MIMO controller tends to over-compensate its actions to zone 3 resulting in a slower response. The above simulations utilized the CFD control structure as shown in Fig. 11. As indicated, the strategy has the functionality of both on and o-line CFD model generation, which is to be used
445 440 435 Temperature response (K) 430 425 420 415 410 405 400 395 390 0 100 200 Time (min)
TSP 2 zero velocity steady state velocity TMP(1) TMP(2) TMP(3)

300

400

Fig. 10. Closed-loop MPC to TSP(2) comparing the two mass ows.

902

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

Fig. 11. Strategy for CFD based process control.

by the controller. The process information or input scenarios are the processing conditions and its corresponding set-point changes. Validation of the CFD models can be conducted by short experimental open-loop tests. This generic architecture can be applied to any size IMM, mold and polymer.

6. Conclusions This paper demonstrates the use of transient CFD simulations to develop a dynamic process model required in model predictive control based on expected process set-point changes in melt temperature. The general framework as shown in Fig. 11 is demonstrated using a CFD model of a three-heater zone IMM to generate open-loop data, which is then used to dynamically retune the predictive controllers dynamic matrix resulting in better closed-loop responses. The CFD control strategy has the ability to generate and store temperature and input dependent open-loop response data in a form suitable for MPC re-formulation, without the rigors of exhaustive experimental testing. The MPC algorithm can then query the database and use the most appropriate data depending on the process operating conditions and the magnitude of the set-point change. This provides the MPC with the process model that can clearly describe any parameter interaction and non-linear material behavior. The CFD simulations are run o-line but can be constantly queried by the controller on-line as outlined in Fig. 11.

Acknowledgement The work has been supported under NSERC Discovery Grants held by the rst two authors.

A.G. Gerber et al. / Applied Mathematical Modelling 30 (2006) 884903

903

References
[1] J. Brown, Injection Moulding of Plastic Components, A guide to Eciencies, Fault Diagnosis and Cure, McGrawHill Book Company (UK) Ltd., Berkshire, England, 1979. [2] M.R. Kamal, W.I. Patterson, V.G. Gomes, An injection molding study, Part I: melt and barrel temperature dynamics, Poly. Eng. Sci. 26/12 (1986) 854. [3] V.G. Gomes, W.I. Patterson, M.R. Kamal, An injection molding study, Part II: evaluation of alternative control strategies for melt temperature, Poly. Eng. Sci. 26/12 (1986) 867. [4] W.I. Patterson, M.R. Kamal, V.G. Gomes, SPE ANTEC 25 (1990) 754. [5] A. Zheng, Robust stability analysis of constrained model predictive control, J. Process Control 9 (1999) 271278. [6] A. Zheng, M. Morari, Stability of model-based predictive control with mixed constraints, IEEE Trans. Aut. Control 40 (1995) 18181823. [7] C.R. Cutler, B.L., Ramaker, Dynamic matrix controla computer control algorithm, in: Proc. Joint Amer. Con. Conf., 1980, Paper WP5-B. [8] D.M. Prett, R.D. Gillette, in: Proc. Joint Amer. Con. Conf., 1980, Paper WP5-C. [9] C.E. Garcia, A.M. Morshedi, Quadratic programming solution of dynamic matrix control (QDMC), Chem. Eng. Commun. 46 (1986) 7387. [10] C.R. Cutler, R.B. Hawkins, Constrained multi-variable control of a hydrocarbon reactor, Proc. Amer. Con. Conf. (1987) 10141020. [11] R. Dubay, Y.P. Gupta, A.C. Bell, Predictive control of plastic melt in heated zones with insulation, J. Injection Molding 2 (1) (1998) 3745. [12] R. Dubay, A.C. Bell, Y.P. Gupta, Control of plastic melt temperature: a multiple inputmultiple output model predictive approach, Polym. Eng. Sci. J. 37 (9) (1997) 15501563. [13] H.K. Versteeg, W. Malalasekera, An Introduction to Computational Fluid Dynamics, the Finite Volume Method, Prentice Hall, Malaysia, TCP, 1995. [14] X. Peng, A.H.C. van Paassen, A state space model for predicting and controlling the temperature responses of indoor air zones, Energy Buildings 28 (1998) 197203. [15] F. Bezzo, S. Macchietto, C.C. Pantelides, A general framework for the integration of computational uid dynamics and process simulation, Comput. Chem. Eng. 24 (2000) 653658. [16] Y. Yang, M.A. Reuter, D.T.M. Hartman, CFD modelling for control of hazardous waste incinerator, Control Eng. Practice 11 (2003) 93101. [17] Theory Documentation, CFX-TASCow V2.12, AEA Technology ESL, June, 2001. [18] G.E. Schneider, M.J. Raw, Control volume nite-element method for heat transfer and uid ow using allocated variables1. Computational Procedure, Numer. Heat Transfer 11 (1987) 363390. [19] J. Richalet, A. Rault, J. Tsetud, J. Papon, Model predictive heuristic control: applications to industrial processes, Automatica 14 (1978) 413428. [20] J. Richalet, A. Rault, J. Testud, J. Papon, in: 4th IFAC Symposium on Identication and System Parameter Estimation, Tbilisi, USSR, 1976. [21] C.R. Cutler, Dynamic matrix control, an optimal multi-variable control with constraints, PhD. Dissertation, University of Houston, Houston, TX, 1983. [22] C.E. Garcia, M. Morari, Internal model control. 1a unifying review and some new results, Ind. Eng. Chem. Process Des. Dev. 21 (1982) 308323. [23] C.E. Garcia, D.M. Prett, M. Morari, Model predictive control: theory and practicea survey, Automatica 25 (1989) 335348. [24] K.R. Muske, T.B. Rawlings, Model predictive control with linear models, AIChE J. 39 (1993) 262287. [25] J.B. Froisy, Model predictive control: past, present and future, ISA Trans. 33 (1994) 235. [26] E.F. Camacho, C. Bordons, Model Predictive Control in the Process Industry, Advances in Industrial Control, Springer Verlag, London, 1995.

Vous aimerez peut-être aussi