MODELLING AND CONTROL OF AN ALUMINIUM STRIP

UNWINDER-REWINDER
Sylvain Leirens1 , Joseph Pierquin
Pechiney - Centre de Recherches de Voreppe
725, rue Aristide Bergès BP 27 - 38341 Voreppe Cedex - France
joseph.pierquin@pechiney.com

Keywords: Aluminium strip processing, modelling, tension drives the rewinder: it provides torque to the roll with high dy-
control, decoupling control. namic through a rotor field oriented control. A pneumatic brake
is used for the unwinder. Two load cells measure the tension in
the aluminium strip. An incremental encoder implemented on a
Abstract
non-motorized roll measures the linear speed. Both acquisition
In an aluminium strip processing line, the main concern is to and control systems are implemented in real time on a Matlab-
control independently velocity and tension. This paper deals Simulink/dSpaceTM 1103 card software-hardware package.
with two control systems applied to a basic unwinder-rewinder
system: the first one is based on classic PI controllers includ-
ing specific compensations. The second is a LQG control.
Simulations demonstrate the effectiveness of both solutions.

1 Introduction
Modern high-speed process lines demand good tension control
for maximum efficiency [1][2]. In an aluminium strip process-
ing line, tension control is required in many operations (rolling,
coating, washing...). For such systems, it is very important to
prevent the occurrence of strip break by decoupling the strip
tension and the strip velocity [1]. The strong coupling between
both variables is induced by the strip elasticity: a physical mod-
elling of this process is given in section 2. Generally, on in-
dustrial production lines, the control is based on simple PID Figure 1: Laboratory prototype
techniques and the know-how of the operators [3]: it is often
non-optimal with low dynamic performances. The first con-
trol strategy presented in this paper is an optimised PI control A simplified diagram is given in figure 2. The gear ratio be-
including specific disturbance compensations: it is based on tween motor and roll is one to one. It is assumed that there is
the inversion principle of the physical modelling deduced from no slip between roll and strip.
the general laws of physics (c.f. section 3). The second is a V1 V2
T
LQG control based on an optimal statistical Kalman filtering Ω1 Ω2
(c.f. section 4). Both strategies are compared and show an Web (l, e)

ability to decouple velocity and tension. r1 L r2

2 System Modelling C1 C2
Unwinder (J1, f1) Rewinder (J2, f2)
A continuous aluminium strip processing line typically can be
described as a system of rolls and a strip stretched between
them [4]. This is the simplest unit of a continuous strip pro-
cess. At each side, a motor provides torque to the roll. Our Figure 2: Unwinding-Rewinding system
laboratory prototype exhibits the inherent problems of metal The aim of this study is to build a model of this system. First,
transport system (see figure 1). An induction motor (3 kW) a mathematical ed
model of both rolls and strip tension dynamics
1 That work has been performed to meet requirements of Sylvain Leirens’s is established. Second, this model is developed under a state-
end of studies internship with Pechiney. Sylvain Leirens is now with the ASH model form.
Team of Supelec (sylvain.leirens@supelec.fr), Avenue de la Boulaie, BP 81127
35511 CESSON-SEVIGNE Cedex - France.
Nomenclature considering that the tension force in the strip of the unwinder
roll is neglected in comparison with T .
L : Span length (m)
When the span is not stretched, Au = A1u = A2u and (6) can
l : Span width (m)
be simplified:
e : Span thickness (m)
A : Cross-sectional area of web (mm2 ) "Z #
L
ρ : Density of aluminium strip (kg/m3 ) d 1 V2 (t)
dx = V1 (t) − . (7)
E : Modulus of elasticity (Pa) dt 0 1 + ²(x, t) 1 + ²2 (t)
Vi : Strip velocity (m/s)
T : Strip tension force (N) Assuming that ² ¿ 1, high order terms can be neglected:
ri : Roll radius (m) 1/(1 + ²) ≈ 1 − ². Equation (7) becomes:
r0 : Empty roll radius (m) "Z #
L¡ ¢ ¡ ¢
Ji : Polar moment of inertia of roller (kg.m2 ) d
1 − ²(x, t) dx = V1 (t) − V2 (t) 1 − ²2 (t) . (8)
Ji0 : Polar moment of inertia of empty roller (kg.m2 ) dt 0
fi : Friction coefficient (kg.m2 /s.rad)
Ωi : Roll rotational speed (tr/mn) Meanwhile supposing that the strain ² does not vary with x, i.e.
Ci : Torque (N.m) ²(x, t) ≈ ²2 (t), (8) can be written as:

d²2 ¡ ¢
2.1 Strip tension force −L = V1 (t) − V2 (t) 1 − ²2 (t) , (9)
dt
The system under consideration is the piece of strip between so:
the unwinder and the rewinder whose length L is assumed to d²2
be constant. It is a typical mass flow system. L = V2 − V1 − ²2 V2 (10)
dt
Moreover, the strip is assumed not to be strained beyond its
limit of elasticity. omitting time t.
The mass balance equation is: Using Hooke’s law, i.e. ²2 = T /AE and notation d
dt ( ) = (˙),
à " #! equation (10) gives:
d mass stored
= LṪ = AE(V2 − V1 ) − T V2 . (11)
dt in the system
à ! à !
input output
− . (1) Considering small variations around a steady-state operating
mass flow mass flow point:
The strip density is assumed to be constant. Equation (1) gives:
"Z # V1 = V10 + ∆V1 ,
L
d V2 = V20 + ∆V2 ,
A(x, t) dx = A1 (t)V1 (t) − A2 (t)V2 (t) (2)
dt 0 T = T0 + ∆T.
where subscripts 1 and 2 denote unwinder and rewinder respec- Then we have:
tively.
Considering an infinitesimal element of span (length dx), ele- d∆T
L + V20 ∆T = AE(∆V2 − ∆V1 ). (12)
ment stretching can be written as: dt
¡ ¢
dx = 1 + ²(x, t) dxu . (3) With s the Laplace variable, (12) becomes:
Both unstretched (subscript u ) and stretched element volume (Ls + V20 )∆T = AE(∆V2 − ∆V1 ), (13)
are equal:
dv = A(x, t) dx = Au (x, t) dxu . (4) which gives:

A combination of (3) and (4) takes the following form: AE

∆T = (∆V2 − ∆V1 ). (14)
Ls + V20
Au (x, t)
A(x, t) = . (5)
1 + ²(x, t) Omitting the notation ”∆” to improve readability, Equation
Equation (2) gives: (14) can be written as a classic first order transfer:
"Z #
d L
Au (x, t) Kb
dx = T = (V2 − V1 ) (15)
dt 0 1 + ²(x, t) Tb s + 1

A2u (t) where Kb = AE/V20 et Tb = L/V20 .

A1u (t)V1 (t) − V2 (t) (6)
1 + ²2 (t)
2.2 Roll dynamics 2.3 Models

Details are given for the unwinder (which is similar to the Nonlinear time-varying model
winder). Radii vary with time and their expressions take the
following form by writing roll surfaces Si [5]: Using results above, the model below was built:

S1 (t) = π r12 (t), (16)  J˙1 Ω1 + J1 Ω̇1 = T r1 + C1 − f1 Ω1




 J˙2 Ω2 + J2 Ω̇2 = C2 − T r2 − f2 Ω2
then differentiating with respect to time: 
 J1 = a1 + b1 r14



 J2 = a2 + b2 r24
Ṡ1 = 2π r1 ṙ1 (17) 



 J˙1 = 2πρ l ṙ1 r13
omitting time t. J˙2 = 2πρ l ṙ2 r23 (33)
And 


 ṙ1 = − Ω2π
1e

Ṡ1 = −V1 e = −r1 Ω1 e. (18) 




 ṙ2 = Ω2π
2e

Hence: 
 V1 = r 1 Ω 1


Ω1 e = −2π ṙ1 , (19) 
 V2 = r 2 Ω 2


so: LṪ = AE(V2 − V1 ) − T V2
Ω1 e
ṙ1 = − . (20) where : ½
2π
ai = Ji0 − 12 πρ l r04
As radius r1 , inertia J1 is time-varying too:
bi = 12 πρ l
(J1˙Ω1 ) = J˙1 Ω1 + J1 Ω̇1 (21) with i = 1, 2.
with: This model requires knowledge of J10 , J20 , r0 , r10 et r20 .
J1 = J1motor + J1roll (22)
Linear time-invariant model
so:
1 1
J1 = J10 + πρ l r14 − πρ l r04 . (23) Assuming that inertiae change slowly compared to the strip dy-
2
| {z 2 } namics, therefore radii and inertiae are considered as constant
wound strip now. In addition, we detail the model by considering torque
J1 can be written as: loops of our asynchronous motors. Calling Ci∗ the torque ref-
erence, a first order system is used to represent this loop :
J1 = J1constant + J1variable (24)
Ci 1
which is the sum of a constant inertia J1constant and a variable ∗ = (34)
inertia J1variable . Hence: Ci Tc s + 1

1 where i = 1, 2 and Tc is the time constant of the torque loop.

J1constant = J10 − πρ l r04 , (25)
2
1
J1variable = πρ l r14 . (26) As before, the results above are collected and the following
2 model is (see figure 3):
Differentiating (26) with respect to time:  1

 Ω1 = J1 s+f (C1 + T r1 )
J˙1 = 2πρ l ṙ1 r13 . (27) 

 Ω = 1
1


 2 J2 s+f2 (C2 − T r2 )

 1 ∗
The dynamic principle applied to a rotation movement gives:  1C = Tc s+1 C1
1
C2 = Tc s+1 C2∗ (35)
(J1˙Ω1 ) = C1d − C1c (28) 

 V1 = r 1 Ω 1



 V2 = r 2 Ω 2
where C1d et C1c denote driving and load torques respectively: 

 T = Kb (V − V )
1+Tb s 2 1
C1d = C1 + T r1 , (29)
C1c = f1 Ω1 . (30) where : 
 Ji = ai + bi ri4
It gives: ai = Ji0 − 12 πρ l r04 .

(J1˙Ω1 ) = C1 + T r1 − f1 Ω1 . (31) bi = 12 πρ l
Similarly, we obtain for the rewinder: with i = 1, 2.
This model requires knowledge of V10 , J10 , J20 , r0 , r1 et r2 .
ṙ2 = Ω2π
2e
,
J˙2 = 2πρ l ṙ2 r23 , (32)
(J2˙Ω2 ) = C2 − T r2 − f2 Ω2 .
 Enrouleur-dérouleur

1 C1 1 Ω1 V1 3.1 Control design

C1* r1
Tc s + 1 + + J 1 s + f1
Tr1 −
Torque loop Unwinder roll
Kb To control the strip linear velocity V2 , we choose to work on
r1
dynamics
T
Tb s + 1 the motor of the rewinder (see figure 4). This choice is classic
+
Web dynamics
[1]. The strong problem comes from the coupling between V2
r2 and T through T r2 . The influence of T r2 must be compensated
Tr2 in the tension control loop (see figure 4). Then a PI-controller
1 C2 − 1 Ω2 V2
C2* r2 (called Speed PI-2) is tuned (c.f. section 3.2). Its tuning is
Tc s + 1 + J 2 s + f2
Torque loop Rewinder roll
based on mechanical and torque loop dynamics (modelled as
dynamics
first order transfers). Using this type of controller, the static
Enrouleur : boucle de vitesse error (step response) is cancelled.
Dérouleur : boucle de tension
Figure 3: Linear time-invariant model
In the same way, to control the strip tension T , we choose to
K p2 +
1 1 ∆V * V1* 1
V *
C2* T* Kp + work V1+ throughK C + 1 . Therefore weC1* use a similar structure
2
+ − Ti 2 s + + + − Ti s − on+ T s
p1
+ −
− i1
Figure 3 shows thePIstrong
Vitesse coupling between the speed V2PI and
Tension to the rewinder speed loop and tune two PI-controllers (called
PI Vitesse

V2
the strip tension T . Whenever rspeed changes, it affects
T tension
2
SpeedV2(*)PI-1 and
V1 Tension-PI). The r first one is calculated by us-
1

trough the difference ∆V = V1 − V2 . Variations of tension ing mechanical and torque loop dynamics (similar to V2 speed
T
disturb the speed through the load torque T r2 as well. loop). The second one is basedTon strip and V1 speed loop dy-
namics. The controller output ∆V ∗ = V2∗ − V1∗ (∗ denotes
Finally, the state-form of the linear time-invariant model can be
reference)
ComInvTens
allows to obtain V1∗ using the reference V2∗ .
written. ComInvVit
With vectors: This control is based on the compensation of load torques T r1
  and T r2 . These torques disturb loops significantly and the
C1 quality of compensation is very important. We use the strip
 C2  µ ∗ ¶ µ ¶
  C1 T tension model to improve the dynamic of these compensations
X=  Ω 1
, U =
 and Y = , (see figure 4). Furthermore, it is necessary to use reference V2∗
 Ω2  C2∗ V2
to obtain efficient feedforward action on the tension controller
T output.
and matrix: T V1
 
− T1c 0 0 0 0 − 1 ∆V
*
V1* − 1 C1*
T* Kp + K p1 +
  + −
 0 − T1c 0 0 0  Ti s
+ + Ti1 s +−
 1
0 − Jf11 0 r1  Tension PI Speed PI-1
A= J1 J1 , * r1
  V
− Jf22
2
 0 1
J2 0 − Jr22  V2* − V1* Kb
0 0 − KTbbr1 Kb r2
Tb − T1b Tension control loop Tb s + 1

 1 
Tc 0 r2
 0 1  µ ¶ + C2*
 Tc  0 0 0 0 1 V2* K p2 +
1
B=
 0 0 
 and C= , + − Ti 2 s +
 0 0 0 0 r2 0
0  Speed PI-2

0 0 Speed control loop V2

the state-model is obtained:

½ Figure 4: PI-based control scheme
Ẋ = AX + BU
. (36)
Y = CX
3.2 PI controllers tuning
3 PI-based control The model of the system is rewritten as follows:
An optimised PI-based control is proposed here. PID control K
techniques are often considered as non-adapted to control ve- Hbo (s) = (37)
(T1 s + 1)(T2 s + 1)
locity and tension independently [3]. In fact industrial control
schemes are non-optimised. The proposed strategy is based on where T1 et T2 denote the small (parasitic) and large time con-
the model inversion principle [6]. Modelling of physical phe- stants respectively.
nomena has allowed a good understanding of the way the en- PI controllers have the following form:
ergy is converted in the system. Therefore, the system can be
efficiently controlled by appropriate corrections and compen- Kp (Ti s + 1)
sations [6]. CP I (s) = (38)
Ti s
where Kp et Ti denote proportional gain and integral constant 3.3 Simulation results
(s) respectively.
Results for step responses, 100 N for the tension and 1 m/s for
the velocity are given figure 6.
The tuning of controllers must be effective to obtain good de-
coupling between tension force and velocity. The compensated −3
x 10
open loop transfer function is given by: 120
6
100
4
Kp (Ti s + 1) K 80

Tension (N)

Tension (N)
HCbo (s) =
2
(39) 60
Ti s (T1 s + 1)(T2 s + 1) 40
0
−2
20
−4
and corresponding Bode diagram is represented on figure 5. 0
−6
−20
0 0.5 1 3.1 3.2 3.3 3.4 3.5
Gain Time (s) Time (s)

−5
-1 x 10
10

1
-2
0.8

Velocity (m/s)

Velocity (m/s)
5
-1 0.6
ω 0.4
0
0
0.2
-2
0
HC −5 −0.2
bo 0 0.5 1 3.1 3.2 3.3 3.4 3.5
Time (s) Time (s)
Phase
1/T2 1/Ti ωc 1/T1
0 ω
Figure 6: PI-based control
-90
Figure 6 shows that the decoupling between tension and ve-
∆φ locity is efficient. Interactions can be neglected compared to
-180 nominal values of velocity and tension.
Typical strategies are not able to decouple these variables with
Figure 5: Compensated open loop Bode diagram such a performance. It is important to note that this result
could be achieved with standard PLC but requires accurate con-
To reach a compromise between fast response and disturbance trollers tuning as well as the computation of compensations.
reject, the crossover pulsation is placed in the middle of the This level of performance is easily obtained with advanced con-
segment [1/Ti , 1/T1 ] which has a -1 slope [7]. The time con- trol strategies (see next section) but their implementation on
stant Ti is calculated as follows: standard material is more difficult.
Ti − T1
sin(∆φ) = . (40)
Ti + T1 4 State feedback control
For a phase margin ∆φ equal to 37o which means an overshoot In this part, a state approach is studied to build the controller.
approximately equal to 30%, we obtain Ti = 4T1 . Then the For this purpose, we use the state model (36) established in
crossover pulsation is equal to: section 2.
1 1
ωc = √ = . (41) 4.1 Control design
Ti T1 2T1
At this pulsation, we make the following approximation: For this purpose, a LQG control has been developed with an
optimal statistical Kalman filtering to estimate the state of the
KKp system. The problem is formulated in standard form and the
=1 (42)
T2 s regulator is optimised using models of the process and its en-
which means: vironment: the state model is enlarged taking into account pre-
T2 dictive models of references and disturbances (see figure 7).
Kp = . (43)
2KT1
+ U X
= AX + BU Y +

Theoretically, the relation Ti /T1 = 4 can be considered as an u  yo
− Y = CX
 + y
optimum (c.f. Kessler’s symmetrical optimum [7]). But in State model
yc
practice, this tuning is too fast and oscillations may appear. u0
d> = 0 y0
Therefore we can choose a factor greater than 4 ; the sys- −
r> = 0 +
tem will react more slowly. We can add a reference filter to r e
compensate the zero induced by the PI controller too and de-
crease the overshooting. Such a filter has been implemented
Figure 7: Deterministic standard model
(not showed on the figure for better readability).
These predictive models introduce an uncontrollable state part. Figure 8 shows the decoupling between tension and velocity.
The rejection of these modes allows to cancel the static error The step responses are faster than those for the PI-based con-
(presence of integrators). trol.
To reach the compromise between performance and robustness,
two high level synthesis parameters are used. The choice of 5 Conclusion
weighting matrix is based on calculating partial grammians ac-
cording to [8]: the tuning needs to specify two horizons, To and For optimal control of an aluminium strip unwinder-rewinder,
Tc for observation and control respectively. two control structures have been outlined. The first one is based
Writing the gap e = Ce X between outputs and references, the on PI controllers : it is easily implementable on standard PLCs
controller has to minimize the criterium: and gives good results. The second one uses a state approach
Z ∞ and requires a powerful hardware/software system. The state
J= (eT Sc e + uT Rc u)dt (44) feedback controller gives better results especially for the strip
0 tension force. In both cases, the proposed controllers allow a
where Sc and Rc have been chosen as follows: good decoupling between strip tension and velocity.
Z Tc
Sc = I, Rc = Tc (Ce eAt B)T (Ce eAt B)dt. (45) References
0
[1] S. H. Jeon, J.-M. Kim, K.-C. Jung, S.-K. Sul, and J. Y.
To calculate the Kalman filter, we use noise variance matrix So Choi. Decoupling control of bridle rolls for steel mill drive
and Ro as follows: system. IEEE Transactions on Industry Applications, vol.
· Z To ¸−1 35, no1:pp. 119–125, January/Febraury 1999.
T
So = To eA t C T CeAt dt , Ro = I. (46) [2] D. Jouve and D. Bui. Digital servo drives for material
0
tension control system. In PCIM Conference, Nuremberg
The horizon To tunes mainly the response time for the distur- (Germany), pp. 71-78, May 1996.
bance reject. The horizon Tc allows to manage the compromise
between noise sensitivity of the control inputs and gain/phase [3] H. Koç, O. Knittel, M de Mathelin, and G. Abba. Mod-
margins. elling and robust control of winding system for elastic
webs. IEEE Transaction on Control Systems Technology,
vol. 10, no2:pp. 197–208, March 2002.
4.2 Simulation results
[4] P. R. Pagilla. Robust observer-based control of an alu-
Analysing the results for step responses, 100 N for the tension
minium strip processing line. IEEE Transactions on In-
and 1 m/s for the velocity:
dustry Applications, vol. 36, no3:pp. 865–870, May/June
120 3
−3
x 10 2000.
100

80
2
[5] M. Humpich. Contribution à l’étude d’une commande op-
Tension (N)

Tension (N)

60 1
timale multivariable en laminage monocage à froid des
40 0 tôles minces. PhD thesis, Ecole des mines de Paris, Uni-
20

0
−1 versité de Nice (France), 1988.
−20 −2
0.1 0.2 0.3 0.4 3.1 3.2 3.3 3.4
Time (s) Time (s) [6] S. Charlemagne, A. Bouscayrol, I. Slama-Belkhodja, and
3
−3
x 10 J.P. Hautier. Flatness based control of non-linear textile
2.5 1 multimachine process. In Proc. European Power Electron-
2 0.8 ics Conference EPE 2003, CD-ROM, Toulouse (France),
Velocity (m/s)

Velocity (m/s)

0.6
1.5
September 2003.
0.4
1
0.2
0.5
0
[7] C. Kessler. Das symmetriche optimum. Regelungstechnik,
0

0.1 0.2 0.3 0.4

−0.2
3.1 3.2 3.3 3.4
vol. 6:pp. 432–436, 1958.
Time (s) Time (s)

[8] P. de Larminat. Contrôle d’état standard. Collection

Figure 8: State feedback control pédagogique d’automatique. HERMES, 2000.