134 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO.

2, MARCH 1998

Extrusion Process Control: Modeling,

Identification, and Optimization
Brian R. Tibbetts, Member, IEEE, and John Ting-Yung Wen, Senior Member, IEEE

Abstract—Our work has been focused on developing a method-

ology for process control of bulk deformation—specifically, extru-
sion. A process model is required for control. We state the difficul-
ties which exist with currently available models. An appropriate
decomposition and associated assumptions are introduced which
permit a real-time process model to be constructed via spectral
approximation to solve the axi-symmetric two-dimensional (2-
D) transient heat conduction equation with heat generation and
loss. This solution, simulation results, and model execution times
are provided. We use this model development plus output re-
lationships from previous model development work to develop
parameter identification and open-loop control methodologies.
These methodologies are motivated by the structure and proper-
ties of the developed model. We demonstrate that the parameters
and control variables enter into the model equations so that the
identification and open-loop optimization problems are tractable.
An example with plant trial data is provided.
Index Terms— Approximation methods, identification, metal
industry, modeling, optimization methods, partial differential
equations, process control, spectral approximation.
I. THE PROCESS
T HE direct extrusion process is illustrated in Fig. 1. The
ram (1) applies force to the dummy block (2), transmitting
the force to the billet (5). Initially, this force causes upset—the
whole billet is deformed to fit the container (3) shape and
there is initial flow from the die (4). After upset, the extrusion
process takes place, producing the desired extrudate (6). The
metal for the extrudate comes from the deformation zone
(5a) within the billet. The material within the dead metal
zone (5b) remains within the container throughout the process.
During the entire process, the material flow and shape of the
deformation and dead metal zones are evolving. After most of
the billet has been extruded, the flow contorts to such a degree
that extrusion is no longer feasible. The remainder of the billet
is removed and the entire process is repeated.
The range of extruded products is typically large; therefore,
the control objectives vary greatly from product to product.
Manuscript received January 1997; revised October 1997. This work was
supported by the Aluminum Processing Program at Rensselaer Polytechnic
Institute, a consortium of ten Aluminum Extruders, a U.S. Department
of Energy Fellowship for Integrated Manufacturing, and the New York
State Center for Advanced Technology (CAT) in Automation, Robotics, and
Manufacturing. The CAT is partially funded by a block grant from the New
York State Science and Technology Foundation.
B. R. Tibbetts is with the New York State Center for Advanced Technology
in Automation, Robotics, and Manufacturing, Rensselaer Polytechnic Institute,
Troy, NY 12180-3590 USA.
J. T.-Y. Wen is with the Electrical, Computer, and Systems Engineering
Department, Rensselaer Polytechnic Institute Troy, NY 12180-3590 USA.
Publisher Item Identifier S 1063-6536(98)01819-3.
1063–6536/98$10.00  1998 IEEE
Fig. 1. Schematic of direct extrusion.
These objectives may include, for example, surface quality,
microstructure uniformity, and production efficiency. Some
specific examples are:
• stock shapes for secondary processing—production effi-
ciency is very important since these items have a com-
modity pricing structure;
• furniture components—surface quality is key;
• architectural components—surface quality and mi-
crostructural properties are both critical since strength
and aesthetics are required.
The goal of our work is to develop cost-effective strategies
for control of this process. These strategies need to address a
wide range of product geometries, processing conditions, qual-
ities requirements, and productivity goals. A process model is
required to achieve this goal.
The process model enters explicitly into an identifier or
observer, operating in real-time plus it may be directly in-
corporated into a feedback controller. Therefore, this model
must be designed to capture the response of very different
systems (for example, round rod with area reduction of ten
times or rectangular tubing with area reduction of 500 times)
and require modest computational power.
II. BACKGROUND
Process control of bulk deformation has received relatively
little attention from the control systems community. Much
of the work found has been limited by the existing models.
For example, Meyer and Wadley [3] developed a model-
based feedback control for material properties and inequality
constraints for part yielding. This approach could only be
applied to relatively slow processes due to model complexity;
in this particular case, the hot isostatic process required 2 h
to perform.
Previously reported work has focused on problems where
simplifications are appropriate which are not broadly applica-
TIBBETTS AND WEN: EXTRUSION PROCESS CONTROL
ble. For example, there has been an ongoing effort in control of
isothermal thermal forging. The researchers involved include
J. Adams, S. S. Banda, J. Berg, R. Chang, A. Chaudhary, R.
V. Grandhi, D. Irwin, A. Kumar, J. C. Malas, III, M. Mears,
C. A. Schwartz, and V. Seetharaman. Most of this body of
work [4]–[10] assumes that there is no temperature change
in the workpiece during the process. The condition does not
apply to metal extrusion and is not expected to be applicable
to many other deformation processes such as drawing. Also,
it is notable that even with the above the simplifications, the
resulting models were not suitable for real-time control.
There has been some work which is based on removing
the need for a model. The previously cited group proposed
using artificial neural networks to emulation the complex
finite element method (FEM)-based model and open-loop
optimization [7], [9], [10]. This effort did not remove the
constant temperature restriction or faced the prospect of com-
putationally expensive network retraining. Another example
is cyclic learning control for extrusion developed by Pandit
and Buchheit [11], [12]. This work employed a black-box
model and required strong assumptions on process linearity,
actuation behavior, and constancy of disturbances. Many typ-
ical deformation processes will violate one or more of these
assumptions.
Based on our above assessment of the available work and
its possible application to extrusion, an existing model was
sought. These models could be classified into two broad areas:
FEM-based models (detailed) or averaged-variable models
(coarse, approximation-based). The FEM-based models pro-
vide the needed information, but cannot be implemented in
real time [13]–[15]. The averaged-variable models are com-
putationally tractable; however, they do not provide localized
information needed for control [16], [17]. Therefore, we chose
to develop a model of the extrusion process specifically for
real-time control.
III. THE MODEL FOR CONTROL DESIGN
The thermo-mechanical response of the billet during ex-
trusion is governed by a set of nonlinear, coupled partial
differential equations given by 1) heat transfer and 2) defor-
mation approximation
(1)
s.t. (2)
These equations are coupled via the flow stress of the material
which is part of the heat generation expression and
the deformation power . and the development of one possible family of flow velocities
Our purpose requires that these equations be solved; how-
ever, direct solution cannot be done in real time. The following
assumptions are introduced.
1) The deformation field determined with constant flow
stress will closely approximate the optimum field in the
set of possible fields.
135
Fig. 2. Thermo-mechanical model components.
2) The set of possible fields can be restricted to include
only fields which are linear in ram velocity.
3) Model errors introduced by using nonlocal flow stress
are acceptable.
Assumption 1) accomplishes the decoupling of (1) and (2) by
not considering the temperature dependence of the material
flow stress. Assumption 2) allows the model description to be
independent of the control, the ram velocity. Assumptions 2)
and 3) permit the most intense calculations to be done off-line.
The decomposition of the thermo-mechanical model is
shown in Fig. 2. Development of the heat transfer and de-
formation components are described below. The heat transfer
model component establishes the extrusion model framework
or basic model structure. The deformation model compo-
nent, in conjunction with the structure of the heat transfer
model, provides the unique attributes for a given extrusion.
The material response can be included directly by using the
temperature predicted by the heat transfer model and the
strain rate predicted by the deformation model to determine
an appropriate flow stress value. This flow stress value, in
turn, influence the subsequent temperature distributions.
Our model is based on an axi-symmetric extrusion, round
billet to round extrudate. This restriction has been chosen since
even with the simplifications introduced thus far, there is little
hope of a real-time solution for an arbitrary shape. We expect
that the analytical model will capture a large portion of the
response for an arbitrary extrusion shape and the black-box
model will capture the remainder.
This problem is naturally described in cylindrical coordi-
nates. The -axis is the centerline through the center of the
die port and the billet. The -axis extends radially from the
-axis. The origin is at the die entrance. The current billet
length (positive direction) is denoted by . The coordinate
system, main geometric parameters, and the flow boundary
function are illustrated in Fig. 3.
The remainder of this section will be focused on solving (1).
This solution will rely on the linear dependence of the flow
velocity field from the ram velocity; however, no specific form
will be assumed. The next section will review the basis for (2)
which satisfy this equation.
A. State Equation Development
There are two physical quantities from which the states and
their evolution equations are developed: the billet size and the
billet temperature distribution. The system input is the ram
136 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 2, MARCH 1998

in ) allows the desired boundary conditions to be imposed.

The temperature is approximated by

(6)

based on the following eigenvalue problems.

Eigenvalue Problem #1:

Fig. 3. Problem coordinates.

Solution:

speed, . The billet size state can be represented by the

billet length, , with the evolution equation
where
(3) zeroth order Bessel function, first kind;
first-order Bessel function, first kind;
Development of the states and evolution equation for the bil-
zeros of .
let temperature distribution begins with the governing equation
for heat transfer. The governing equation is Inner Product:

(4)

with the following boundary conditions: Eigenvalue Problem #2:

where is the thermal diffusivity, is the density of Solution:

aluminum, is the specific heat capacity, is temperature
is the heat generation term, and is represents the
billet boundary.
The changing boundary and the loss of mass via the where
extrudate adds an interesting component to this problem. Inner Product:
We have transformed this equation so that one can apply
finite-difference or spectral approximation to obtain a solution
according to the following rules:

The original equation, (5), is transformed from a partial

The governing equation becomes
(5)
where the flow velocity is represented as
We have chosen spectral approximation for the solution;
the partial differential equation is transformed into an infinite-
dimension system of ordinary differential equations. An appro-
priate set of basis functions (Bessel functions in and cosines
differential equation to a finite system of ordinary differential
equations. Equation (5) is projected onto the basis set of the
temperature approximation using the inner product operations
defined previously. Integration by parts is used to exchange
higher order derivatives of the temperature field for higher
order derivatives of the basis vectors. The stabilizing influence
of the conduction term permits the infinite dimensional system
of ODE’s to be truncated.
Integration by parts of the conduction term also requires
that the temperature gradient boundary conditions be incorpo-
rated into the solution. Gradient boundary conditions which
consist of temperature-dependent and constant terms can be
incorporated.
After the above manipulation and rearrangement into a
matrix form, the final state equations become
(7)
TIBBETTS AND WEN: EXTRUSION PROCESS CONTROL 137

where dummy block

The heat transfer boundary term is derived from three

functions, , , and , which characterize the
boundary. The following decomposition is introduced so that
specific parameters are available for identification and the
parameters enter into the equation linearly

Using the above decomposition, the boundary condition

term in the temperature state equation can be ex-
pressed as

..
. (8)
.. ..
. .
.. where
.

is the “projected temperature state”, is coefficient

matrix for the conduction term, is the coefficient matrix
for the advective term (rate of change in temperature due to
motion of a bulk with a temperature gradient), is the
coefficient matrix for the heat generation term, is the
boundary condition contribution, is the ram velocity, and
is the ram position. The variable represents the configuration adjustable com-
This system of equations is in a very desirable form for the ponent of the boundary conditions. For example, the simu-
control purpose and has the following properties. lation results shown subsequently assume perfect insulating
• Temperature evolution is nonlinear with respect to the conditions at the extrudate and constant leaky conditions at
ram position (measurable). the die. The coefficients represent the magnitude of these
• Temperature evolution is bilinear with respect to the functions.
temperature state, , and the control, .
• There are free coefficients in and which C. Heat Generation Term Definition
may be determined by through identification. A similar decomposition can be performed on the heat
• The elements of and are determined solely generation term . This term is constructed from deforma-
by problem geometry and the inner product operation for tion, friction, and shearing sources; the following equation
the spectral approximation. illustrates this calculation:
• The elements of and are determined by the
deformation component , and and the
spectral approximation.
• The stabilizing, conduction term dominates as higher (9)
order basis functions are included in the approximation,
permitting the infinite-dimension system of equations to
be truncated to a very low order. The heat generation functions have the following dependen-
cies:

B. Boundary Condition Definition

The model given in Section III-A is for the generalized
boundary condition, . In this section, the
model is made more specific by chosing the gradients to
be independent of temperature. Three gradient functions are where , and represent the scaling of the
specified at 1) the container wall; 2) the die face; and 3) the assumed flow stress. Therefore, the heat generation term can
138 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 2, MARCH 1998

be represented as represented as follows:

(10)
The specific heat generation functions, , , and
, will be developed in Section IV.
D. Output Equations
There are many possible output relations which one can
derive from this model. For our purposes, we will present three
particular equations, the extrudate temperature, the extrusion
load, and the maximum strain rate within the billet.
1) Extrudate Temperature: The output equation for the ex-
trudate temperature depends greatly on the method of measure-
ment or on the use of the variable. Generically, this variable
can be calculated as follows:
(11)
The specific load functions, , , and , will be
developed in Section IV. A bias term has also been included
to account for various losses such as friction in the press and
errors in load measurement.
3) Maximum Strain Rate: Our investigation into the phe-
nomenological effects of the extrusion process on the quality
of the extrudate show that the strain rate and, in particular,
the local strain is one indicator of quality (see Shepard [18]).
This investigation did not, however, indicate precisely how this
local strain rate should be defined so that one has a parameter
which is a good indicator of product quality. We, therefore,
suggest that the maximum strain rate within the billet be used
as an approximation to this indicator.
By construction of the flow field description (Section IV),
this parameter is linear in

If one wishes to consider the process performance or a research (12)

setup which uses an instrumented die with a thermal couple,
then the coefficients, , are time-invariant and can where represents the functional description of the strain
be written as . The typical measurement rate.
capability in the production environment requires a more
complicated description.
E. Model Summary
The production environment may use either a pyrometer or
an infrared camera for determining the extrudate temperature, The model is summarized in (13)
if this temperature is measured at all. Either of these devices
will measure temperature on the extrudate where the extrudate
exits the die bolster (a substantial support structure which (13a)
holds the die in place). This point is approximately 3 to 4 ft (13b)
from the die exit. A very thorough description would require
that one consider the continued conduction throughout the (13c)
extrudate, losses to the environment, and the movement of (13d)
the extrudate from the die exit to the point of measurement. (13e)
A reasonable approximation is to consider the following.
1) There is no heat loss to the environment. where
2) There is no heat conduction axially. “projected temperature state;”
3) The heat completely equalizes radially due to the high ram position;
conductivity of aluminum. ram velocity;
Thus, the coefficients must contain a time delay, conduction;
dependent on , and the projection of the billet boundary heat loss;
modes at the die face onto the zeroth radial mode of the advection (rate of change in temperature due to
extrudate motion of a bulk with a temperature gradient);
heat generation;
extrudate temperature;
output matrix for ;
extrusion load;
2) Extrusion Load: The same phenomena which generates maximum strain rate;
heat also requires the press exert a force on the billet to cause map, linear in ram velocity and nonlinear in ram
the billet to be extruded. Generically, the relationship can be position and billet location, to strain rate.
TIBBETTS AND WEN: EXTRUSION PROCESS CONTROL 139

IV. FLOW FIELD DEVELOPMENT functions without affecting the bulk material flow. This result
We have fully developed the deformation model component is given below
in earlier work [1], [2]. A brief summary of these results is
provided for completeness.
We chose upper bound analysis (UBA) as the basis of the
deformation model component. UBA is the search for the
kinematically admissible flow field which requires the
minimum power
s.t.

According to the UBA theorem [19], this flow field is the

actual flow field. Conventional applications of UBA were
specifically for determining required process power. Typically,
simple flow fields have been used which predict the power;
however, they bare little resemblance to the physical flow (14)
fields observed experimentally. This is not suitable for our
purpose. subject to the following constraints:
Han et al. [20] developed a flow field approximation for
use in UBA for a contoured die, very light reduction extrusion
problem. This particular flow field did appear to be a closer
approximation to the physical situation than the typical UBA
flow field. The flow field includes a bulk component (fixed)
and a variational component. The parameters of the variational where
component of the flow field are chosen to minimize the process
power via a numerical optimization algorithm.
We have made the following improvements in the original
work for application to typical industrial extrusion control
problems.
• Determined required conditions on the flow boundary for The strain rate tensor components are defined as
application to flat-faced dies, , shown in Fig. 3.
• Generalized the structure of the variational component of
the flow field.
• Determined required conditions on the flow field de-
scription so that it is linear in ram velocity (the control
variable).
• Implemented a hybrid optimization algorithm for deter-
mining the free parameters in the flow field descrip-
tion. The algorithm includes line searches, modified line
searches, and steepest descent methods. Increases in speed and the strain rate magnitude, , is defined as
and robustness against local minima are observed.
With the conditions specified, the optimization can be per-
formed off-line, independent of the control.
The conventional UBA applied to flat-face dies required
that the material ultimately going into the die (region 5a
in Fig. 1) be artificially divided into an undeformed region
The power optimization is described by (15) where is
(far from the die) and a deformation zone. This division is
the minimum power resulting from a flow field described in the
difficult to specify and generates local discontinuities. These
previous section with the most favorable set of free parameters
difficulties are avoided with our approach. We demonstrated
and boundary description,
good qualitative prediction of experimental measurements of
a flow field with our deformation model. (15)
The kinematically admissible velocity field description is
constructed by forming the axial component, , first and There are three contributing factors to the power used during
then directly calculating the radial velocity component . extrusion: power directly required to deform the material,
The first term (bulk component) in the axial velocity function , power require to overcome friction at the container wall,
guarantees that the same amount material flows through all , and power required to shear the material in the dead
axial cross sections. The remaining terms with constraints are zone-deformation zone interface, . All of these terms
constructed to locally modify the flow via polynomial basic depend on the velocity of the flow field and, therefore, on the
140 IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 2, MARCH 1998

free parameters introduced above. The power components are which are not specifically defined by the modeling effort. The
conduction and advective terms in the temperature state equa-
tion (13a) are completely defined by the modeling process;
the conduction term is solely a function of the billet geometry
and the advective term is defined by the flow field. The
heat boundary condition describes the gradient at the
boundary and is not specified. The heat generation term
depends on 1) the flow field (fixed by construction
of the deformation model component) and 2) the material flow
stress and coefficients of friction and shear. The coefficients of
(16) friction and shear are not specified and, therefore, suitable for
identification. The flow stress is also suitable for identification
where is the height of the dead zone, , , and since it appears as a constant. This constant represents an
are the same scaling factors as in (10), and is the material aggregate of the constitutive relationship operating on the flow
flow stress. field and the temperature distribution.
By the development of the velocity field, the ram velocity The ram position state transition equation (13b) does not
, enters into the velocity field and strain rate parameters have parameters for identification. Similarly, the temperature
linearly output equation (13c) does not have parameters requiring
identification since this equation depends only on the projected
temperature state and the workpiece geometry.
The extrusion load equation does have the same parameters
From (17), the load terms are defined as for identification that the heat generation term has. The load
equation can be defined in terms of the heat generation term

(17)

The parameters for identification are expressed for conve-

The velocity field, via the definitions for the extrusion load, nience as
is used to construct the heat generation terms which result
from deformation, internal shear, and friction. The following
terms are used in (9):

A. Estimated Extrudate Temperature

Identification will use measured temperature, , and its
estimate, . This measurement is discrete with samples.
Therefore, and

The heat generation kernel due to deformation is the integrand

from the extrusion load analog. The heat generation kernels
due to friction and shear are slightly more complicated since The estimate for measured temperature is calculated by using
they are not defined throughout the volume. The introduction the forth-order Runge–Kutta integration algorithm to transform
of the Dirac delta function and application of appropriate the continuous state equation (13a) into a discrete representa-
scaling permit these function to be used in the spectral tion
approximation.
(18)
V. IDENTIFICATION
The model description in (13) is physically motivated; where is the output matrix,
therefore, parameters for identification should be quantities and are the number of
TIBBETTS AND WEN: EXTRUSION PROCESS CONTROL
basis functions in and , respectively, and
..
.
.. ..
. .
where
Also, note that represents the conduction matrix at the
th instantance . Similarly, represents the
advection matrix at the th instantance . manner.
It is important to note that (18) is linear in the parameters to
be identified, . This linearity will be maintained as long as:
• flow stress is not treated as an explicit function of
temperature (an assumption in the model development);
• boundary temperature gradients are not temperature de-
pendent.
The linearity of this relationship greatly improves the solvabil-
ity of the problem over a typical nonlinear description.
Equation (18) is linear in the initial state, , also. This
property will prove very useful in solving the open-loop
control problem discussed in Section VI-C.
B. Estimated Extrusion Load
The estimated extrusion load is given by
(19)
VI. RESULTS
In this section, we first present simulation results for the
model, including predicted internal billet temperature distri-
bution. Next, our initial parameter identification technique for
this model is shown. Finally, using the equations developed for
the identification, we solve an open-loop control problem. Ex-
perimental data from an experimental extrusion press and from
a production press is used for comparison and identification.
A. Model
The temperature predictions shown in Fig. 4 and in Fig. 5
are based on experimental extrusion of AA6060 by Lefstad
[21]. The billet is 16.5-cm length and 10.0-cm diameter. The
extrudate is 1.04-cm diameter and the ram velocity throughout
the push is 2.0 cm/s. Experimentally measured extrudate
temperature is shown in Fig. 4.
141
Fig. 4. Predicted extrudate temperature.
The model-generated results were produced in the following
1) Constructed a flow field for the extrusion geometry and
tabulated by current billet length (off-line).
2) Calculate and which depend on the flow
field (off-line).
3) Calculate and , (off-line).
4) Solve the temperature evolution equation (on-line).
These results reasonably approximate the experimental mea-
surements for most of the extrusion. Also, the model predicts
that the core will be cooler than the extrudate surface, consis-
tent with available information.
The isothermal plots for different times during the push
show that the majority of the temperature increase is generated
by the intense shearing between the deformation zone and dead
zone, regions 5a and 5b in Fig. 1, respectively. Note, the heavy
boundary represents the current billet shape and the extrudate
size (toward the bottom).
The off-line calculations for the simulation shown required
2 h with a 586, 100 MHz running Windows NT 3.51. The
results shown required 1.5 min to run the 7.0-s extrusion.
These results are for a 64-state model (8 8 basis functions).
Many of the features predicted by this model precision are
present in a 16-state model (4 4 basis functions). The
reduced-state model for this example executes in 3.0 s on the
same platform. Although some care was exercised in code
development to optimize for speed, much more can be done
for a real-time implementation. Based on number of operations
and raw processor performance, improvements up to a factor
of ten might be realizable on this platform. In addition, low-
cost hardware alternatives exist which might realize similar
improvements. Therefore, the high-resolution model can be
implemented in real-time.
B. Parameter Identification
The identification problem will use both measured tem-
perature and extrusion load. The problem will be cast as
142
Fig. 5. Predicted temperature history of billet during extrusion. Note: Thinner lines are used to represent lower temperatures.
TABLE I
PARAMETER IDENTIFICATION: PARAMETER CONSTRAINTS
a constrained quadratic optimization problem. Earlier efforts
used a least squares approach which did generate parameters;
however, these parameters were not physically realizable. The
identification problem requires that the following minimization
be solved [22]:
where is the set of permissible parameter values and the
cost function is defined as
where represents the weighting on the measured temper-
ature error and is the weighting on the extrusion load
error.
The weighting parameters have two purposes. First, they can
be used to reflective relative confidence in the measurement ac-
curacy. Second, the force and temperature measurement have
dissimilar units and different magnitudes. These differences
imply a weighting factor, perhaps an inappropriate weighting.
Explicit inclusion of a weighting factor and an expectation of
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 2, MARCH 1998
TABLE II
PARAMETER IDENTIFICATION: PARAMETER VALUES
the magnitudes of the different data items permits adjustment
of the weighting parameters for a balanced cost function.
The nonlinear programming algorithm, constr() in MAT-
LAB was used with the constraints shown in Table I to
minimize the cost function. The weightings in the cost function
were set with temperature measured in degrees Celius and
force in kiloNewtons: and .
The resulting parameter values are provided in Table II.
(20) The plant data used for the identification and the result model
predictions are shown in Fig. 6. The comparison of actual data
and model prediction with errors generally below 10%. A trend
is noted in the temperature prediction being positively biased
and the load being negatively biased.
(21) The identified parameters were used in the model with a new
plant trial data set to gage the success of the identification
procedure. The extrusion which generated this new data set
used a different initial billet temperature profile and a different
ram speed trajectory. Fig. 7 shows the experimental data and
the model predictions. The prediction of trends and general
agreement are good. Although still within 10%, the error in
temperature prediction has increased compared with the error
present in the identification run. The error in load prediction
has decreased.
TIBBETTS AND WEN: EXTRUSION PROCESS CONTROL
Fig. 6. Comparison of data and model prediction after identification.
Fig. 7. Model predictions for the reference extrusion.
Figs. 8 and 9 illustrate the differences in ram speed trajec-
tory and initial billet temperature profile between the extrusion
used for identification and the extrusion used as reference to
check the identification.
Both the identification run and the verification run show
temperature predictions above the measured temperature. We
believe that there are two sources of this discrepancy. First,
improvements in the deformation model for short billets are
require. In general, the force predictions for short billets are
too high which will result in elevated temperature prediction.
Second, constant flow stress is not accurate and is expected
to have a substantial impact on the temperature prediction.
The validation run started with a billet which was 40 to 70 C
hotter than the billet used for identification. Also, the extrudate
temperature was 20 C hotter. Therefore, the flow stress of this
billet will substantially lower than the identification billet, thus
producing less heat during extrusion. To see the effects, the
verification run was rerun with a 15% reduction in the flow
stress parameter. The results are shown in Fig. 10. We see that
143
Fig. 8. Comparison between identification and reference extrusions: Ram
speed trajectory.
Fig. 9. Comparison between identification and reference extrusions: Initial
billet temperature profile.
Fig. 10. Model predictions for the reference extrusion with reduced flow
stress.
this validation run has significantly less error than the original
validation run.
One source of difficulty for using this data set for identifica-
tion is the lack of variation in the input signal (the ram speed,
shown in Fig. 8). It is highly likely that this signal does not
satisfy the persistency of excitation condition, resulting in only
partial identification, rather than unique determination, of the
parameters.
C. Open-Loop Control
The range of extruded products and their applications are
large. As such, the applicable measures of quality and their
required levels vary from product to product. These require-
ments can be mapped to specific values of the physical process
variables [9], [18], [5]. We have chosen constant maximum
144
Fig. 11. Comparison of the optimized and reference ram speed and maxi-
mum strain rate trajectories.
Fig. 12. Predicted extrudate temperature using optimized ram speed trajec-
tory and initial billet temperature. The reference extrudate temperature is also
shown.
strain rate and constant extrudate temperature to represent
good quality for this example.
The maximum strain rate, given by (13e), is linear in the ram
speed, ; however, it is nonlinear in the billet length. Also, it
is not dependent on the projected temperature state variables.
The temperature evolution equation (13a) depends on both
the initial projected temperature state and the ram speed
trajectory . Therefore, the suggested design procedure is
to determine the desired ram speed trajectory which results
in the appropriate maximum strain rate. With the ram speed
determined, one can then find an initial billet temperature
profile which achieve a constant extrudate temperature for the
ram speed trajectory.
The required ram speed trajectory can be calculated easily
since the strain rate output equation (13e) is a static map.
Therefore, the desired ram speed trajectory is given by
The choice of desired maximum strain rate must still be made.
For this particular example, we adjusted the value so that our
optimized extrusion finished approximately 10% faster than
the extrusion shown in the previous section for identification
validation. The ram speed profiles and the associated maxi-
mum strain rate predictions for the referenced plant trial and
the open-loop control design are shown in Fig. 11.
IEEE TRANSACTIONS ON CONTROL SYSTEMS TECHNOLOGY, VOL. 6, NO. 2, MARCH 1998
Fig. 13. Comparison of the optimized and reference initial billet temperature
profiles.
TABLE III
OPTIMIZATION: INITIAL BILLET TEMPERATURE MODE CONSTRAINTS AND VALUES
To determine the desired initial billet temperature profile,
we start by recalling the equation for the estimated extrudate
temperature (18)
(23)
Note, we have expressed the implicit dependence of the
and on the ram speed trajectory. Since ram speed
is known, all parts of (23) are known except for . An
approach analogous to the parameter identification problem for
determining is motivated since enters into (23) linearly.
Finding the best initial billet temperature profile requires
that the following minimization be solved:
(24)
where is the set of permissible initial projected temperature
states and the cost function is defined as
(25)
Exactly as was done for the identification problem, the non-
linear programming algorithm, constr() in MATLAB was used
with the constraints shown in Table III to minimize the cost
function, . The constraints on the initial billet temperature
were driven by the actuation capabilities. We assumed that
the best one could hope is that a “reasonably smooth” profile
(22) could be imposed axially and that there is no radial variability.
The first six states represent the constant radial variable axial
basis functions. Since the greater index represents a mode with
higher spatial frequency, these modes were subject to tighter
restrictions to impose the “reasonably smooth” criteria.
The control problem was solved with desired extrudate
temperature being arbitrarily specified at 520 C. The required
initial projected temperature modes are given in Table III.
Fig. 12 shows the predicted extrudate temperature from this
TIBBETTS AND WEN: EXTRUSION PROCESS CONTROL
initial billet temperature profile and the previously developed
ram speed trajectory. Finally, the required initial axial billet
temperature is shown in Fig. 13.
VII. CONCLUSION
We have presented a model which is derived directly
from the mathematical description of the physical phenomena
present. This model has demonstrated real-time performance.
The construction of the model required only a minimum of
assumptions.
The resulting temperature state equation is bilinear in the
temperature state and the control, ram velocity plus nonlinear
in the billet length state (measurable).
This paper successfully demonstrates the use of a physically
motivated model for identification and open-loop control of the
extrusion process. For this example, we picked one particular
objective function for the open-loop design. The model form
can support many different objective functions or constraint
conditions; therefore, the model can be useful for a wide
variety of products.
Further refinements are required; particularly noteworthy is
the change in the flow stress model. Other future developments
may include investigation of excitation persistency for identi-
fication plus observer and feedback controller development.
ACKNOWLEDGMENT
The authors would like to thank the Werner Co. for per-
forming the identification plant trials and Williamson Corp.
providing a temperature sensor for these trials. They would
also like to thank M. Lefstad, SINTEF, University of Trond-
heim, for allowing the use of his extrusion data.
REFERENCES
[1] B. Tibbetts and J. Wen, “Application of modern control and modeling
techniques to extrusion processes,” in Proc. 4th IEEE Conf. on Contr.
Applicat., 1995, pp. 85–90.
[2] , “Control framework and deformation modeling of extrusion
processes: An upper bound approach,” in Proc. 6th Int. Alum. Extrusion
Tech. Sem., 1996, vol. 1, pp. 375–386.
[3] D. Meyer and H. Wadley, “Model-based control feedback control of
deformation processing with microstructure goals,” Metallurgical Trans.
B, vol. 24B, pp. 289–300, Apr. 1993.
[4] A. Kumar, R. Grandhi, A. Chaudhary, and D. Irwin, “Processing
modeling and control in metal forming,” in 1992 Amer. Contr. Conf.,
Chicago, IL, June 1992, pp. 817–821.
[5] J. Malas III and V. Seetharaman, “Using material behavior models to
develop process control strategies,” JOM, vol. 44, pp. 8–13, June 1992.
[6] R. V. Grandhi, A. Kumar, A. Chaudhary, and J. Malas III, “State-space
representation and optimal control of nonlinear material deformation
using the finite element method,” Int. J. Numerical Methods, vol. 36,
pp. 1967–1986, June 1993.
[7] J. Berg, J. Adams, J. Malas III, and S. Banda, “Design of ram velocity
profiles for isothermal forging via nonlinear optimization,” in 1994
Amer. Contr. Conf., Baltimore, MD, June 1994, pp. 323–327.
[8] J. Berg, J. Malas III, and A. Chaudhary, “Open-loop control of a
hotforming process,” in Proc. 5th Int. Conf. on Numerical Methods in
Industrial Forming Processes—NUMIFORM’95, Ithaca, NY, June 1995,
pp. 539–544.
[9] J. Berg, R. Adams, J. Malas III, and S. Banda, “Nonlinear optimization-
based design of ram velocity profiles for isothermal forging,” IEEE
Trans. Contr. Syst. Technol., vol. 3, pp. 269–278, Sept. 1995.
[10] C. A. Schwartz, J. M. Berg, R. Chang, and M. Mears, “Neural-
network control of an extrusion,” in 1995 Amer. Contr. Conf., 1995,
pp. 1782–1786.
145
[11] M. Pandit and K. Buchhei, “Control schemes for a cyclically operating
aluminum extruder plant,” in Proc. 3th IEEE Conf. on Contr. Applicat.,
1994.
[12] M. Pandit and K. Buchheit, “A new measurement and control system
for isothermal extrusion,” in Proc. 6th Int. Alum. Extrusion Tech. Sem.,
1996, vol. 1, pp. 79–86.
[13] K. Mori, “Extrusion and drawing,” in Numerical Modeling of Material
Deformation Processes, P. Hartley, I. Pillinger, and C. Sturgess, Eds.
London: Springer-Verlag, 1992, pp. 303–317.
[14] E. Atzema and J. Huetink, “Finite element analysis of forward–backward
extrusion using ALE techniques,” in Proc. 5th Int. Conf. on Numerical
Methods in Industrial Forming Processes—NUMIFORM’95, Ithaca, NY,
June 1995, pp. 383–388.
[15] D. Vucina and I. Duplancic, “Flow formulation FE metal forming
analysis with boundary friction via penalty function,” in Proc. 6th Int.
Alum. Extrusion Tech. Sem., 1996, vol. 1, pp. 69–72.
[16] K. Laue and H. Stenger, Extrusion. Metals Park, OH: Amer. Soc.
Metals, 1976.
[17] G. Dieter, Mechanical Metallurgy, 3rd ed. New York: McGraw-Hill,
1986.
[18] T. Sheppard, “The application of limit diagrams to extrusion process
control,” in Proc. 2nd Int. Alum. Extrusion Tech. Sem., 1977, vol. 1, pp.
331–337.
[19] W. Prager and P. G. Hodge, Theory of Perfectly Plastic Solids. New
York: Wiley, 1951.
[20] C. H. Han, D. Y. Yang, and M. Kiuchi, “A new formulation for
three-dimensional extrusion and its application to extrusion of clover
sections,” Int. J. Mech. Sci., vol. 28, no. 4, pp. 201–218, 1986.
[21] M. Lefstad, “Metallurgical speed limitations during the extrusion of
AlMgSi-alloys,” Ph.D. dissertation, Univ. Trondheim, Norway, Nov.
1993.
[22] L. Ljung, System Identification: Theory for the User. Englewood Cliffs,
NJ: Prentice-Hall, 1987.
Brian R. Tibbetts (M’97) received the B.S. de-
gree in mechanical engineering from Rensselaer
Polytechnic Institute, Troy, NY, in 1986 and the
M.S. degree in electric engineering at the University
of Texas at Arlington in 1993. He is currently
pursuing the Ph.D. degree in computer and systems
engineering at Renselaer on a U.S. Dept. of Energy
Fellowship for Integrated Manufacturing for work
on aluminum extrusion process control.
He worked for General Dynamics, Fort Worth
Division/Lockheed Fort Worth from May 1986 to
January 1994. During this period, he worked on many aspects of flight
control hardware and system architecture on a variety of concept and prototype
aircraft, notably the YF-22. He returned to Rensselaer in January 1994 and
joined the N.Y.S. Center for Advanced Technology in Automation, Robotics,
and Manufacturing. His current research interests include modeling, control,
and software architecture for systems and processes.
John Ting-Yung Wen (S’79–M’81–SM’93)
received the B.Eng. degree from McGill University,
Montreal, Canada, in 1979, the M.S. degree from
the University of Illinois, Urbana-Champaign,
in 1981, and the Ph.D. degree from Rensselaer
Polytechnic Institute, Troy, NY, in 1985, all in
electrical engineering.
From 1981 to 1983, he was a System Engineer
with Fisher Control Company, Marshalltown, IA,
where he worked on the coordination control of a
pulp and paper plant. From 1985 to 1988, he was
a member of technical staff at the Jet Propulsion Laboratory, Pasadena, CA,
where he worked on the modeling and control of large flexible structures and
multiple-robot coordination. Since 1988, he has been with the Department
of Electrical, Computer, and Systems Engineering at Rensselaer Polytechnic
Institute, where he is currently a Professor. He is also a member of the
New York Center for Advanced Technology in Automation, Robotics and
Manufacturing. His current research interests are in the area of modeling
and control of multibody mechanical systems, including flexible structures,
robots, and vehicles, and material processing systems such as extrusion and
resistive welding.