Journal of Process Control 17 (2007) 509–521

www.elsevier.com/locate/jprocont

q
Looper and tension control in hot rolling mills: A survey
a,*
I.S. Choi , J.A. Rossiter b, P.J. Fleming b

a
POSCO, Technical Research Laboratories, Process Control Research Group, 1, Goedong-dong, Nam-gu, Pohang, Gyungbuk 790-785, South Korea
b
University of Sheffield, Department of Automatic Control & Systems Engineering, Sheffield S1 3JD, UK

Received 6 October 2005; received in revised form 27 June 2006; accepted 17 December 2006

Abstract

Looper and tension control is important in hot strip mills because they affect the strip quality as well as strip threading. Moreover, the
most difficult challenge in controller design and control performance arises from the interaction between looper angle and strip tension.
Disturbances from several sources cause a deterioration in control performance, and thus a major role of the control algorithm is effec-
tive disturbance rejection. Many authors have proposed and applied a variety of control schemes to this control problem, but neverthe-
less, the increasingly strict market demand for strip quality requires further improvements in this control area. Hence, this paper critically
surveys the strengths and weaknesses of several existing academic and industrial approaches and explores the potential for development
in this area.
 2007 Elsevier Ltd. All rights reserved.

Keywords: Hot rolling mills; Looper and tension control; Coordinated control; Model predictive control; Multivariable control; Internal model control

1. Introduction order to produce thinner steel sheet. However, the quality

of these thinner sheets is seriously affected by any quality
This paper considers some of the control issues in hot defects arising during hot rolling. Therefore, hereafter this
rolling mills. Hence to set the context, first we give a brief paper focuses solely on hot rolling.
description of a typical process. The role of a hot strip mill The major specifications which should be satisfied in hot
process is to roll the slabs produced in a continuous casting rolling mills are the mechanical properties, surface and
process and produce strips of thickness 0.8–20 mm. Fig. 1 dimensional quality of a strip, and stable operation of the
shows a typical layout of a hot strip mill. Slabs are heated process. We consider only the specifications associated with
up to around 1200 C in reheating furnaces and then rolled control problems in this paper, hence focussing on the
on two reversing roughing mills which reduce the thickness dimensional quality and proper mass flow of a strip. The
of bars to around 30 mm. After rough rolling, the strip dimensional quality is controlled mainly in the finishing
passes through the crop shear before entering the finishing mills by dedicated control systems such as AGC (Auto-
mills. At this point the strip is further rolled by finishing matic Gauge Control), AWC (Automatic Width Control),
mills composed of typically six or seven rolling stands. ASC (Automatic Shape Control) and APC (Automatic
The rolled strip is cooled down by spraying water in a Profile Control).1 On the other hand, stable mill operation
run-out table and coiled by a down coiler. The final prod- through smooth threading of the strip is achieved by mass
uct may be processed further through cold rolling mills in flow control which is used to balance the input and output
flow of a strip in a stand.
Looper and tension control, shown in Fig. 2, is important
q
Briefer form was presented at IFAC World Congress (Prague, July in hot strip mills because they affect both the dimensional
2005).
*
Corresponding author. Tel.: +82 54 220 6352; fax: +82 54 220 6914.
1
E-mail address: i.s.choi@posco.com (I.S. Choi). Also define ASR (Automatic Speed Regulator).

0959-1524/$ - see front matter  2007 Elsevier Ltd. All rights reserved.
doi:10.1016/j.jprocont.2006.12.005
510 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521

Reheating Roughing Crop Finishing Run-out Down

Furnace Mills Shear Mills Tabl e Coiler

Fig. 1. Hot strip mill layout.

ith stand i+1th stand degrades the control performance and stability and the
SISO loop does not properly tackle this. Various research
tension
papers seeking to improve the performance and stability
looper
looper roll
strip
of this scheme have been published, for instance, [1,8,16],
but nevertheless the main weakness of this control scheme
remains, that is, the ‘neglect’ of the interaction.
M looper To overcome this weakness and enabled by recent
motor main
main
motor M motor M advances in the technology of sensor applications, there
has been an increased installation of tension meters. For
ASR ASR/ACR ASR
instance, [6] developed a tension meter using load cells
and applied it to width control problems and one alterna-
looper -tension
control system tive suggested the use of a soft sensor [3,35] based on esti-
mation theory and applied to tension control problems.
Fig. 2. Tension and looper control in finishing mills: ASR—Automatic The availability of tension measurements has enabled ten-
Speed Regulator, ACR—Automatic Current Regulator.
sion feedback and the reduction of interaction by changing
control pairings such that the rotating speed of a mill
quality and the mass flow of a strip. The high strip tension motor controls strip tension while angular velocity of a
between stands induces width shrinkage, thickness reduc- looper motor controls looper angle. Using this scheme
tion and moreover can produce an edge wave on a strip. makes the process model a two-input and two-output mul-
However, it does make the mass flow more stable and hence tivariable system, and many advanced control algorithms
the use of strip tension produces a tradeoff between dimen- [2,5,7,18] based on this control structure have been applied
sional quality and mass flow. Therefore, the strip tension and shown to give improved control performance. Never-
should be kept to a desired value during operation to ensure theless, increasingly strict demands by the market for strip
proper product quality and strip threading. quality in hot strip mills requires yet further advances in
A looper installed at inter-stand positions reduces ten- the control approaches in this area.
sion variations by changing its angle, so it can contribute In the meantime, there have been some previous
to the quality of products. It can also enable stable opera- approaches [37,34,38] to compare different control synthe-
tion of the process by absorbing an excessive loop of the sis for looper–tension control. Petit et al. [37] proposed a
strip arising from a mass flow unbalance. For example, in comparative synthesis of different multivariable control
the case of low tension, the looper angle increases to get structures such as multivariable PI control, decoupling
proper tension, resulting in stable threading of a strip, control, optimal control and H1 control. They described
while, in the case of high tension, the looper angle decreases brief design procedures, and also compared nominal per-
to reduce strip tension. Ideally, the looper angle needs to formance and robustness for these methods via simulation.
keep a desired value during operation to reduce the tension Anbe et al. [34] implemented several multivariable control
variation and to have the flexibility to absorb large changes schemes and compared their performance. Randall et al.
in loop length during an abnormal rolling condition. [38] describes several design synthesis including SISO (Sin-
Thus, the specifications of dimensional quality and mass gle Input and Single Output) and some multivariable
flow in hot rolling mills can be satisfied by simultaneous approaches such as RFN (Reverse Frame Normalisation),
control of the strip tension and looper angle. Traditionally, LQG (Linear Quadratic Gaussian) and H1 control. How-
looper angle control has been performed by changing the ever, none of these previous approaches provided a thor-
rotating speed of an upstream main motor but there has ough review over substantially large number of schemes
been no tension feedback control because of the difficulty and a suggestion for a fruitful research direction.
of installing and maintaining tension meters at inter-stand This paper provides a critical review with strengths and
positions. The conventional PI control is based on this con- weaknesses for various looper–tension control algorithms
trol scheme and has been used most widely in industry; this proposed in literature and industry. Furthermore it gives
is described in Section 3. The main advantage of this specific comparison based on four major assessment crite-
scheme is that it does not use tension meters, and hence ria and also explores some future technology in this area.
it is easy to design a controller as this reduces to a SISO The control problem is described in Section 2, a survey
system. On the other hand however, there is significant of existing looper–tension control strategies is discussed
interaction between the tension and looper angle, which in Section 3, a summary and comparison is given in
 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521 511

Section 4, future trends for looper–tension control are dis- action. AGC systems are deployed to ensure higher gauge
cussed in Section 5 and conclusions are given in Section 6. quality; they reject thickness disturbances due to set-up
mismatch, skid marks, roll eccentricity and so on, by
2. Control problems changing the roll gap using hydraulic screw down system.
However, while this hydraulic gap control system enables
This section gives an overview of the tension and looper a faster response for thickness control, it often creates a
model, the control issues caused by disturbances and the disturbance to the tension control system because of the
controller specifications. This forms the background and mass flow change caused by roll gap movement.
setting for the remainder of the paper. Another disturbance comes from a set-up mismatch at
the finishing mills. Before the strip arrives at the finishing
2.1. Tension and looper model mill, a supervisory computer calculates initial set-up values
such as the roll gap, rolling force and rolling speed for the
Fig. 3 shows an outline of a looper. Clearly from the stands to achieve the required specifications. However, any
figure, the loop length between stands is L0 ðhðtÞÞ ¼ mismatch between the real processed values and the set-up
pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi values creates a constant disturbance in the tension and
ðx2 þ y 2 Þ þ ðL0  xÞ2 þ y 2 .
Rt thickness.
Define L ¼ L0  0 ðvin;iþ1  vout;i Þdt, that is the accumu- Disturbances also occur for downstream loopers at coil-
lated loop length which changes due to the speed difference ing [7]. When the lead end of a strip is coiled, a large ten-
of the strip between stands. sion between the last stand and a down coiler may arise.
From L 0 and L, inter-stand strip tension is defined as This can cause inter-stand tension fluctuations at the finish-
follows [21]: ing mill.
 0 
L ðhðtÞÞ  LðtÞ All these disturbances discussed here influence the strip
rðtÞ ¼ E ð1Þ tension and looper angle control performance, thus affect-
LðtÞ
ing strip thickness, width, flatness and mass flow.
where, E represents Young’s modulus. The looper model is
derived by applying Newton’s second law with an inertia 2.3. Specifications for controller design
JL, motor torque M and load torque T.
J L €h ¼ M  T ð2Þ The objective of the looper and tension controller is to
maintain the tension and looper angle at their desired
Load torque T includes the torque by strip tension, the values regardless of disturbances, modelling errors and
torque from strip and looper weight, the torque to bend process interactions. Many control strategies have been
the strip, the frictional damping torque and so on [11]. proposed to satisfy the following design specifications:
The strip tension in Eq. (1) depends on looper angle,
while the looper angle is affected by the strip tension in • The stability of the closed-loop system is guaranteed
Eq. (2). Therefore there is an interaction between the under interactions, parameter uncertainties and distur-
looper angle and strip tension. This interaction makes it bances.
difficult to design a controller and the obvious consequence • Proper closed-loop performance is obtained by quicker
is degraded control performance and stability. control action for reference changes and disturbances;
proper strip quality and stable mill operation is ensured.
2.2. Disturbances • The interaction effects are minimized by having a proper
control structure and systematic handling of interactions.
There are tension or mass flow disturbances from • Tension control performance is more important than
several sources which affect both control performance looper angle performance; tension has a direct affect
and stability. The main disturbance comes from the AGC on the width, shape and thickness of a strip, and so
looper angle performance can be sacrificed for quicker
P(x,y) tension recovery.

i i+1 The following section gives a survey of several looper–

vout,i tension control schemes which have been designed to
vin,i+1
l
α β satisfy the specifications listed above. These comprise
(0,0) schemes contributed by industry as well as the academic
θ (L0,0)
community.
(x0,y0)
3. Survey of looper–tension control technology
L0

Fig. 3. Outline of a looper: vout,i is the exit strip speed at the (i)th stand, Several control algorithms have been applied success-
vin,i+1 is the entry strip speed at the (i + 1)th stand. fully, in industry, to looper–tension control. This section
512 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521

describes and gives a brief critique of some more recent The mass flow in this control scheme is controlled by the
strategies as well as a more conventional PI scheme. looper height control (LHC) loop which changes the inter-
stand strip length by the rotating speed of a mill motor.
3.1. The conventional PI control scheme The LHC performs PI (Proportional and Integral) control
action to get rid of the angle difference between the target
A conventional control scheme has been adopted most and measured looper angles.
widely in industry because of its simplicity and, perhaps The control mechanism is explained as follows. In the
more significantly, it does not require tension meters. steady state, the load torque, including the strip tension,
Fig. 4 shows the block diagram of a conventional control- balances the torque supplied by a looper motor, so the
ler [21]. In this scheme the torque of the looper motor is looper angle is stationary. However, if the strip tension is
adjusted according to looper angle to maintain strip ten- increased by a disturbance, then the load torque becomes
sion to a desired value. bigger than looper motor torque, so the looper angle
The CRCC (Current Reference Calculation Controller) decreases. When a decreased angle is detected, the LHC
computes the reference current from (2) to balance the tor- increases the rotating speed of an upstream main motor
que on a looper motor against the load torque at the given to recover the looper position. Therefore, the loop length
angle; the given angle depends on the strip tension, strip between stands increases, and the looper angle (height)
weight, looper weight and so on. Fig. 5 indicates the inputs subsequently increases.
to a CRCC block. The main problem with this scheme is an interaction
between looper angle and strip tension, which is amplified
θref Tension
by using DC motors with large inertia. This results in slow
Main
LHC
ASR
responses and large deviations in both the tension and
- looper angle. If the controller gains are increased with
Looper
the aim of improving control performance, the plant can
System Looper become unstable due to the interaction; an illustration of
σref angle this is given in Fig. 6 which presents time responses of ten-
CRCC
Looper ςφl
ACR sion and looper angle for a constant mass flow disturbance.
In summary, high gain causes instability whereas low gain
gives poor performance.

Fig. 4. Conventional PI control scheme. 3.2. Robust–adaptive looper control

Strip tension can be stabilized through precise looper

σ ref
re
Looper Torque angle control, because angle variation induces tension
href
re 1 Icref variation, but uncertainty due to changes in the material
breref Calculation ξφl characteristics and the slip between roll and strip may make
Block a conventional looper control system, such as the PI con-
θ
troller, unstable. A logical way forward is to try robust con-
trol [32] and hence this was proposed based on internal
Fig. 5. A block diagram of CRCC: rref—tension reference [N/mm2], href—
thickness reference [mm], bref—width reference [mm], h—actual looper model control. However, the problem of slow response
angle [rad], n/l—current torque coefficient of a looper motor [N mm/A], times remained because the controller was constrained by
Icref—reference current of a looper motor [A]. the need to ensure robust stability over all frequencies.

Fig. 6. (a) Tension responses for a mass flow disturbance. (b) Looper angle responses for a mass flow disturbance.
 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521 513

Estimator approaches have some advantages: (i) they do not require

Tf = f(k2,m2,∈)
(ζ) a formal model and (ii) they can make effective use of the
Looper considerable system knowledge of mill operators.
angle ref. For example, a tension control scheme [12] using fuzzy
R(s) + PM(s) -Ls
Y(s)
+
F(s) P(s)e logic was proposed for strip thickness control. The design
- P'(s) Looper
+ Filter Looper angle was based on the fact that tension is one of the most impor-
System
Main tant parameters affecting the thickness variation, and is
Controller
often adjusted by using the operator’s empirical knowledge
-L's
during operation. For the design of a fuzzy controller they
P'(s)e
chose as input variables: (i) the entry strip thickness (hin)
Looper Model
and (ii) carbon equivalent (C%) of the material, and as
Robust Controller
for output variables: (i) forward tension (rf) and (ii) back-
Fig. 7. Structure of adaptive robust looper control: PM(s)—transfer
ward tension (rb). They constructed 16 fuzzy rules based
function of the reference looper model, Tf—filter time constant, (>0), m2, on process experts’ knowledge with the following form:
k2—parameters derived from satisfaction of sensitivity and complemen-
tary sensitivity conditions. Ri : If hin is X i1 and C% is X i2 then
Drf is Y i1 and Drb is Y i2 ð3Þ
In order to improve control performance, some authors
[1] incorporated an adaptive control method into the con- where, Ri is the ith fuzzy rule, X i1 and X i2 are the linguistic
ventional robust control scheme. They changed robust con- variables for hin and C%, respectively, and Y i1 and Y i2 are
trol gains according to looper parameter changes. Fig. 7 the linguistic variables for Drf and Drb respectively.
shows the structure of a robust–adaptive looper control The simulation results using the proposed controller
system. In the figure the damping coefficient f, which is show a good agreement with the real data and an intuitive
the most changeable looper parameter over the fre- mechanism. However this control also has some restric-
quency in terms of rolling condition, is estimated under tions: (i) it can be applied to the tension set-up of finishing
the assumption that the plant model is a second order sys- mills but not to on-line tension control; (ii) its application
tem. The first order filter F(s) is selected to make the com- is restricted to thinner strips because tension does not have
plementary sensitivity function T(s) small only over the a a dominant affect on thickness for thick strips and (iii) it
frequencies with low damping coefficient whereas, with can be used for thickness control, but not for pure tension
conventional robust control, F(s) is usually designed to control or mass flow control.
be small for all frequencies; thus this method is differ- An intelligent design technique for looper control [8]
ent from conventional robust design and motivated the was also suggested to overcome the difficulties of modelling
incorporation of adaptive control. Moreover, in order to as well as the conventional weakness of robustness. The
reduce the influence of the disturbances, the filter was cho- proposed system applies to looper height (angle) control
sen such that the sensitivity function S(s) is small at low with height measurements used as an indication of tension.
frequencies. The resulting robust stability condition of For the fuzzy controller design, they selected as input vari-
the filter depends largely on the damping ratio, and so con- ables: (i) looper height error and (ii) its error change, and as
trol performance can be improved if this ratio can be esti- for output variable: exit strip speed in the upper stand.
mated reliably. Fuzzy rules were developed heuristically to facilitate the
In summary, the main contribution was to utilize a shaping of looper’s response and to simplify controller
robust controller which stabilized the looper control system tuning. They also incorporated self-tuning methods for
under changes of operation condition as well as rejected membership functions, rule-base, and aggregation (T)
disturbances. Moreover, by incorporating an adaptive ele- operators into a neuro-fuzzy controller to tackle the short-
ment, it also solved the problem of the slow response of a comings of typical fuzzy controllers, such as: (i) redundant
conventional robust design. However, the weakness of this or insufficient rules might be specified; (ii) a controller with-
approach is that angle control accuracy depends on the out a self-tuning mechanism gives unsatisfactory perfor-
estimation error for the ‘approximate’ damping ratio and mance under system parameter changes and (iii) multiple
there are still inherent performance limitations due to the tunings of the membership function, rules and operators
use of a SISO system design, that is, there still exists an makes optimal tuning more difficult. The block diagram
interaction problem between tension and angle. for this scheme is shown in Fig. 8. A comparison with con-
ventional PI and PID controllers demonstrated the
3.3. Neuro-fuzzy control improved performance of the proposed controller for dis-
turbance rejection. However, this scheme has some limita-
Adaption has also been deployed as a part of very differ- tions: (i) its design is based only on a looper height control
ent strategies. For instance, some researchers [8,12] have loop, and so interaction between tension and looper angle
used artificial intelligence and designed controllers with causes a deterioration in system performance and (ii)
learning, intelligence and an adaptation capability. These regardless of its on-line tuning capability, it still requires
514 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521

Δε
-1 State ∧
Tuning
Jk Z
Feedback ΔV Observer1
Algorithm ∧
Gain1 dv
-1
Looper -1 Z dv
Z +
Δe(k)
angle ref. - Vref + Tension ΔT
yr e(k) + Process y ASR
Neuro-Fuzzy
speed Control ΔV Dynamics
+ Controller u Looper
- e(k)
angle
Interaction
δv1 δv2 dτ
Δ ref +
Δθ
ATR
+ Looper
Δτ Dynamics
∧
Δθ
Fig. 8. A block diagram of the neuro-fuzzy control system: Jk(=(oy/ou)k), State ∧
dv1, dv2—Jacobian, the exit speed variation in stand i and the entry speed Feedback Δθ Observer2
∧
variation in stand i + 1, respectively. Gain2 dτ

Fig. 9. Structure of observer-based tension control: DVref, Dsref, DT, Dh—

an appropriate training set to provide sufficient accuracy work roll speed reference, looper motor torque reference, tension variation
until appropriate tuning is acquired. ^_ ^
and looper angle variation respectively. D^e, D Vb , d^v, Dh, Dh, d^s—estimates
Apart from the use of neuro-fuzzy scheme as a loop con- of strip strain variation by speed difference between stands, roll speed
troller for looper height control, neural network was uti- variation of upper stand, speed disturbance, looper angle velocity
variation, looper angle variation and torque disturbance respectively.
lised for the gain modulation of the controller. Konishi
ATR—Automatic Torque Regulator.
et al. [39] proposed the use of the RNN (Recurrent Neural
Network) for PID gain modulation of looper height con-
troller in order to cope with control parameter changes turbance. Moreover, it is easy to tune the controller thanks
during operation. They adopted the RNN model to make to the analytical derivation of the control parameter. How-
use of past experiences of the modulation as human does, ever, with this scheme there remains an interaction due to
and this was implemented by using past outputs in the the limitation of model accuracy and a lack of systematic
input layer. As a means to avoid the complexity of the net- interaction handling. This interaction can give violent tor-
work structure they proposed the use of four estimation que fluctuations in the case of a large initial tension distur-
functions of dynamic looper behaviours such as overshoot bance. Such an event may be followed by instability of the
value, mean value of overshoot, dead beat performance process at strip threading time. This problem was solved in
and overshoot time. Numerical experiments for the gain an ad hoc manner by setting the looper angle response
modification of the looper height controller showed that slower than the tension response.
the identified RNN model can modify PID gains of the
controller in an appropriate way.
3.5. Internal model control + impedance control
3.4. Observer-based control
Some authors [2] developed a tension and looper control
An observer-based multivariable tension control [13,14] scheme based on decentralization and coordination. They
was developed to improve tension control accuracy at strip chose the manipulated variables to be the rotation speed
threading time and applied to aluminium hot rolling mills. of a main motor for tension control and the angular velocity
The approach used control pairings such that the changing of a looper motor for angle control; these control pairings
rolling speed of the upper stand controls strip tension while gave reduced interaction effects which can be verified by
torque of a hydraulic looper controls looper angle; this is an interaction measurement [24] using structured singular
different from the conventional control pairings. The struc- values, hence demonstrating the validity of designing two
ture of this control scheme is shown in Fig. 9. The control- decentralised controllers. Each controller was designed
ler was designed such that: based on the two-degree-of-freedom IMC (Internal Model
Control) structure with reference tracking and low fre-
• A state feedback was designed to improve the dynamic quency disturbance rejection. However, this control scheme
response characteristics of speed and looper angle. weakened the coordinated action of the looper; the looper
• Additional control inputs were added to the state feed- control loop tried to keep the looper angle constant regard-
back via feed-forward to counterbalance the distur- less of tension variation. Therefore, in order to improve
bances in steady state. transitional response of tension control they incorporated
• The feed-forward control was computed by using the an impedance controller [23], which includes the looper
disturbance estimates from the observer. model, to calculate the desired looper position in accordance
with tension variations. The resultant control structure is
This control scheme achieved improved dynamic shown in Fig. 10. The main advantages of this control
response as well as offset-free tracking under constant dis- scheme are:
 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521 515

Target
IMC tension
Controller F11 σref +
- K i11 + V ref 1 Vr Tension σ
^ s - 1+T v s Dynamics
P11 + - + tension
- Main Motor
- Vref σ K i11
σref +
C 11
+
P11
ASR
+
+
s State
C1 Feedback Interaction
P12 Gains
Impedance Controller Target K i12
P21 looper s Looper Motor
angle ASR looper
ω ref + angle
-1 θ ref + θref + K i22 + - Looper
G G rθ C22 P22 θ PI Dynamics θ
+
-
+
- -
s + ωref -
^ + - ω
IMC
P22
Controller F22
Fig. 12. Block diagram of H1 looper–tension control.

Fig. 10. Structure of looper–tension control based on decentralization and

b
K posed of two PI controllers and cross control gains. How-
coordination: G ¼ 2 gr , G1 rh —an approximated inverse transfer func-
bJ s þb
D sþb
K ever, tension control performance was unsatisfactory in
tion of the transfer function from href to h, C11 and C22—PI type
controllers, F11 and F22—free parameters.
this scheme rather than angle performances; this because
of a lack of angle co-ordination with tension control objec-
tives. Therefore, a H1 looper–tension controller [7] was
• It is easy to design the controllers because this scheme proposed to improve tension control performance. For
consists of two SISO (Single Input and Single Output) controller design the weighting functions were chosen to
subsystems and an impedance controller. let the sensitivity function be small in low frequencies to
• It enables stepwise commissioning of control systems by ensure a good disturbance rejection ability whilst letting
adding control modules such as the IMC and impedance the complementary sensitivity function be small at high
control to PI controllers. frequencies to reduce the effects of noise and plant uncer-
• On-line tuning can be done intuitively because the con- tainty. They adopted a state-feedback-type controller
troller parameters allow physical interpretations such as shown in Fig. 12 instead of output-feedback-type because
a disturbance observer and mechanical impedance. it has lower order and does not imply pole-zero can-
cellation.
However, for effective interaction reduction this control The proposed controller satisfied the requirements for
scheme depends only on the process model and control both stability and disturbance rejection. However, the ten-
pairings. Therefore, in case of a looper system with large sion control performance was not satisfactory compared
inertia and slow response motors, the ignored interaction, with angle control. Therefore to improve tension control
that is the off-diagonal elements of the process model, give performance they introduced a cross parameter C1 which
restrictions on the efficacy of controller design. enables the designer to change the angle reference to share
the tension control.
3.6. H1 control The major advantages of this control scheme are its abil-
ity to reject low frequency disturbance such as skid marks,
The NIC (Non-interactive control) scheme [15] shown in to ensure robust stability for noise and model uncertainty,
Fig. 11 was widely used in industry due to the simplicity of and to co-ordinate the tension control by moving the angle
design and effective reduction of an interaction. It is com- reference. However, the main disadvantage is the complex-
ity of the controller, accompanied by the difficulty of tun-
Target ing. Moreover, the use of the cross parameter C1 seems ad
tension
σref +
- 1+T T s + Vref 1 Vr Tensio n σ
hoc and needs to be incorporated in a more systematic way.
s/K T + 1+T v s Dynamics tension
Tension Control Main Motor
H 12 ASR 3.7. Non-interactive control + H1 control
Cross Controller Interaction
One of the problems in a conventional NIC scheme is
Target
looper
H 21 Looper Motor
that the control gains of the PI controllers cannot be
Angle Control ASR looper
angle
θref + 1+Tps + ω ref Looper angle increased enough to reduce variations in tension and
PI θ
s/Kp + - Dynamics looper angle sufficiently. To tackle this problem, the mod-
-
ified NIC scheme [17] with disturbance compensation
shown in Fig. 13 was proposed. The disturbance compen-
Fig. 11. Block diagram of NIC-based looper–tension control: Tv—time
sator K(s) was constructed by using a H1 control design:
constant of main motor ASR, Vref—work roll peripheral speed reference,
Vr—work roll peripheral speed, wref—looper motor angular velocity appropriate weighting functions were determined to reduce
reference, w—looper motor angular velocity, PI—proportional and the disturbance effects in the low frequency and, to sup-
integral controller. press the noise and modelling errors in the high frequency.
516 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521

Main Motor Gain

Target ASR
tension selector
- σ
σref + + Vref 1 Vr Tension
Cσ(s)
+ 1+T v s Dynamics tension Target
tension
σ ref + - Vref 1 Vr Tension
Target H(s) Interaction
looper Looper Motor +
- 1+T v s Dynamics σ tension
ASR looper
angle Main Main Motor
ωref angle
θref + + Looper cont-
ASR
Cθ(s) PI θ State
+
- Dynamics roller
- Feedback Interaction
ω F
Target KI
- looper s Looper Motor
angle ASR
K(s) + - - ω
θ ref + ref Looper
+ +
PI Dynamics θ looper
-
- ω angle
Q(s)

Fig. 13. Structure of looper–tension control based on NIC with distur-

bance compensation: Ch(s), Cr(s)—PI type controllers, H(s)—cross Fig. 14. Block diagram of optimal multivariable looper–tension control.
controller, Q(s)—process model, K(s)—disturbance compensator.

because the increased weighting on tension control implies

The resultant scheme has independent controls for refer- that under abnormal states, the looper angle fluctuations
ence tracking by the non-interactive control and robust sta- may be bigger and might subsequently induce process
bility by the disturbance compensator. instability. Moreover, the use of models with low orders
The advantage of the scheme is the ease of controller in the controller design, to allow easy implementation,
tuning due to the independent controllers structure. The can cause a deterioration in control performance due to
controllers can be adjusted as PI controllers C(s) for refer- model mismatch.
ence tracking, the cross controllers H(s) for interaction Another optimal design scheme was proposed based on
reduction and the disturbance compensator K(s) for dis- interaction decoupling [9]. They set up a model to include
turbance rejection. However, when disturbance effects interactions among gauge, looper angle and strip tension
are large or reference tracking performance needs to be for an entire seven stands. The decoupling of the model is
improved, the gain of C(s) and/or K(s) needs to be realized by means of a similarity transformation, and
increased and this may cause to decrease robust stability; enables the treatment of the model as a set of units which
complementary sensitivity T(s) depends on both C(s) and do not affect each other. Using the decoupled model, they
K(s) in this scheme. Hence, robust stability is limited in this constructed an independent optimal controller for each
scheme which means that in order to increase robust stabil- stand which is composed of an integral controller, state feed-
ity margins, it is necessary to sacrifice the capacity of the back gains and similarity transformation matrix. The con-
disturbance suppression. troller termed ‘Local Autonomous Control’ has some
benefits: (i) weighting matrices can be selected so as to opti-
3.8. Optimal control mize each unit without taking account of the system’s total
optimality; (ii) the control system of each unit can start inde-
A looper–tension control [10,36] utilizing integral-type pendently without considering the other units. However, in
optimal regulator design technique was suggested. Seki this scheme the physical meaning of the states may not be
et al. implemented two distinct controllers using two set preserved because of the state transformation for decoupling
weighting matrices to improve tension control accuracy interacted-states, and so this removes the direct link to (and
under large disturbances such as caused by temperature therefore ease of) conventional controller tuning rules.
fluctuations; one is controlled to have a minimum looper
angle fluctuation within permissible tension range under 3.9. Predictive control
normal operation condition, and the other is designed for
abnormal circumstances with a large tension fluctuation Recently, there have been some approaches to this area
which exceeds the permissible range. The scheme controls investigating the potential benefits of MPC (Model Predic-
the looper angle actively in order to return the tension fluc- tive Control) [26–29].
tuation to within its permissible range quickly. This control Some authors [16] suggested an MPC controller for
shown in Fig. 14 is comprised of an integral controller, a mass flow control design by taking account of constraints.
state feedback controller and a control gain selector. They assumed constraint violation of output variables for
The major advantage of this scheme is the co-operation difficult rolling materials such as HSLA (High Strength
between the looper and tension control by giving tension Low Alloy) and took this into account in the controller
more weighting, which improves the tension performance design. The simulation results with their proposed control-
effectively under abnormal conditions. Also, steady state ler shows better performance than a conventional PI con-
error is eliminated by using integral controller. However, trol. However, the design was applied to a SISO loop,
there may exist some restriction in controller gain selection, looper angle control, and consequently some of the major
 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521 517

benefits of the MPC for multivariable system design such schemes such as those described in Sections 3.1–3.3 adopt
as interaction handling were not utilised. conventional pairings; the speed of an upstream main
Others [18] investigated the efficacy of an MPC scheme for motor controls the looper angle, whilst the torque of a
the looper–tension control problem by using a MIMO model. looper motor controls the strip tension.2 The problem with
They defined constraints on looper angle and strip tension to these pairings is that significant interaction remains.
ensure quality specification and stable mill operation. The With tension meters, normally two sorts of control pair-
existence of large mass flow disturbances which cause ings have been used. They are opposite to the conventional
abnormal operating conditions and hence constraint viola- pairings; the speed of an upstream main motor controls the
tions were assumed in the controller design. They proposed strip tension, whilst the torque (or angular velocity) of a
a linear quadratic optimal MPC [20,33] to ensure guaranteed looper motor controls the looper angle. These pairings
stability and constraint satisfaction. The incorporation of can decouple two loops such as a tension control loop
integral action enabled offset-free tracking with disturbances. and a looper angle control loop. Section 3.4 adopts torque
A comparison of the simulation results with PI control control of a looper motor3, while Sections 3.5–3.9 adopt
demonstrated that the MPC control scheme can be a useful angular velocity control of a looper motor in looper angle
design strategy for looper–tension control problem in that control4.
it handles constraints as well as interaction systematically. A comparison of the CC and SC schemes using the l–
interaction measure [2] shows that both schemes are satis-
4. Summary and comparison factory for a thinner strip with high tension reference.
However, for a thicker strip with low tension reference
This section provides a summary and comparison of the frequency responses of the CC scheme are restricted
the various looper–tension control schemes described in in some frequency ranges. A physical explanation is that
Section 3. For ease of comparison, Table 1 represents a ‘‘the looper ASR (Automatic Speed Regulator) in the SC
summary of different control approaches proposed for scheme suppresses torque disturbances caused by tension
looper–tension control. The following subsections then deviations by keeping the angular velocity of the looper at
focus on some specific comparison. zero; this means that the process itself has a function for alle-
viating the interactions. However, in the CC scheme, the pro-
4.1. Industry application cess itself does not have such a function for suppressing the
disturbances, so the performance of the looper angle depends
The conventional PI control has been used most widely only on the looper angle controller’’ [2]. The advantage of
in industry because of its simplicity, and because it does not the SC control scheme compared to the CC scheme is also
require tension meters. In Fig. 4, CRCC can be easily shown in the [10]; ‘‘the process characteristics of the SC
designed given model (2) and usually has a tabular form. scheme are relatively stable regardless of the rolling condi-
The only design parameters are PI gains in the LHC. How- tions because the SC scheme acts on the looper speed control
ever, it tends to give slow responses and large control errors loop so that it can absorb changes in the looper load torque
of tension and looper angle due to the interaction. caused by tension and looper angle fluctuation’’. However,
In cases where there are tension meters available at the SC scheme may not be applied to looper control sys-
inter-stand positions, one successful scheme in industry tems with slow response motors because contact between
has been the NIC. The main advantage is the ability, via the looper and the strip may be lost in the case of a large
the cross controller, to reduce interaction and moreover disturbance being added to the system.
it is easy to design. From Fig. 11, the main design param-
eters are two PI controllers and two cross gains. The PI 4.3. Design methodologies
controllers are designed and tuned independently, and the
cross gains are then tuned to satisfy stability and control All of the schemes in the previous section are motivated
performance. The problems are unsatisfactory tension con- to satisfy specific control objectives and under different
trol performance and a limitation to the magnitude of the operation conditions, even though they share a common
control gains under large disturbances. overarching goal to ensure stability and proper control per-
More advanced control schemes described in Sections formance. In simple terms they can be grouped into two
3.4–3.6 and 3.8 have also been applied to industry success- major design methodologies, frequency and time domain
fully. They were developed to solve some specific problems design.
with conventional control, but most of them introduced a Some approaches, such as those described in Sections
complex control structure and subsequently can be difficult 3.4, 3.8 and 3.9, use time domain design methodologies,
to design and tune. or in fact more precisely, they are multivariable designs

4.2. Control pairings 2

Note that there is no tension feedback control in this pairings. Tension
is controlled by changing looper motor torque via feed-forward.
Control pairings are closely related to the use of ten- 3
This is called a CC (Current Control) scheme.
4
sion meters. Without tension meters, most of the control This is called a SC (Speed control) scheme.
518 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521

Table 1
Summary of different control approaches proposed for looper–tension control
Method Industry application Control pairings Methodology (domain) Coordinated control
Conventional CC SC Time Frequency
p p p
Conventional PI control
p p p
Non-interactive control
p p
Robust–adaptive control
p p
Neuro-fuzzy control
p p p
Observer-based control
p p p p
IMC + Impedance control
p p p p
H1 control
p p
NIC + H1 control
p p p p
Optimal control
p p p
Predictive control
CC—the current control scheme, SC—the speed control scheme.

arising from optimal control laws selected to minimize explicit inclusion of guaranteed stability for disturbance
some performance index. The major control objective is and parameter uncertainty. However, being worst case
to improve transient responses in the case of large distur- design procedures, these designs can give conservative per-
bances and implementation is via state feedback, integral formance and tuning may not be intuitive. Therefore, in
controllers and disturbance estimators. The major advan- order to improve control performance they may require
tages of optimal control strategies are systematic handling an incorporation of an adaptive element such as described
of interactions and moreover simple techniques to ensure in Section 3.2.
offset-free tracking under constant disturbances. The disad-
vantages are that there is no systematic way for selecting 4.4. Coordinated control
the weighting functions in the performance index and it is
probably impractical to implement full state feedback. In spite of its simplicity and ability to reduce interaction,
The approaches of Sections 3.2 and 3.5–3.7 utilize fre- the NIC scheme has a performance limitation. It does not
quency domain design methodologies. The major objec- provide satisfactory tension control performance compared
tives are to reject low frequency disturbances, and to to looper angle performance, because there is a lack of
ensure stability against high frequency noise and model some coordinated action of angle to tension control. As
uncertainty and therefore, H1 and IMC techniques are stated in Section 2.3, tension control is more important
adopted. The main advantage of these techniques is the than looper angle control in some operation conditions.

Fig. 15. An example of coordinated control in an MPC scheme with constrained optimization.
 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521 519

Therefore, looper angle performance can be sacrificed for trade off between mass flow control and gauge control by
quicker tension recovery under large disturbances by changing the output weightings. They constructed weight
changing its position within a acceptable range. Coordi- changing rules so that the dynamic response (controller
nated action is the topic of several sections. Section 3.5 sug- bandwidth, etc.) is independent of the strip parameters
gests the use of an impedance controller which includes the such as entry/exit thickness, force, temperature etc. There-
looper model to calculate the desired looper position in fore, optimum control gains are product dependent and
accordance with tension variation. Section 3.6 incorporates can be changed on-line but in a coil-by-coil basis.
a cross parameter C1 to change angle reference to share the One research direction with significant potential is coor-
tension control, which improves tension control perfor- dinated control using an MPC algorithm which integrates
mance significantly. Section 3.8 and 3.9 propose optimal gauge, looper and tension models across all the stands. Sys-
control with coordinated actions of angle for a large ten- tematic on-line coordinations between mass flow and strip
sion fluctuation. They compute angle demands in an intrin- quality can be implemented by the constraint handling
sic way during optimization in order to return the tension facility of MPC. Mass flow limits can be defined as hard
fluctuation to within its permissible range quickly. Fig. 15 constraints which should not be violated in any case
shows an example of the coordinated action of an MPC whereas tension and thickness specifications can be defined
scheme (Section 3.9) with constrained optimization. An as soft constraints. In the case where prediction gives large
unknown constant disturbance would induce constraint AGC actions and abnormal conditions which induce severe
violation of the tension so MPC introduces non–linear con- mass flow unbalance and constraint violations, the MPC
trol perturbations such as c which are added to the linear could compute controls to avoid this condition by allowing
control laws; u = Kx + c. Their values are computed in some slack in the control of thickness. Therefore, the pro-
a constrained optimization in order to avoid a predicted cess should operate with best quality within permissible
violation of the tension. c1 and c2 act in a coordinated mass flow performance. The incorporation of a gauge
fashion to keep strip tension within constraints; c1 acts to model to the looper–tension control model also enables
increase the mill motor speed while c2 works to decrease one to achieve better observation of the disturbances
the looper motor angle. As a consequence, tension perfor- and therefore better control performance than with just
mance can be improved systematically by sacrificing looper looper–tension control. However, in order to apply this
angle performance. scheme to industry, a significant and major obstacle to
In summary, the evidence suggests that the incorpora- overcome is the implied computational burden associated
tion of coordinated actions in the looper–tension control with the prediction and constraint handling. Moreover
design is essential to improve tension performance for large there is a need for good prediction models and the formu-
disturbances. lation of a robust MPC algorithm which satisfies state and
input constraints under parameter uncertainty and
5. Future trends for looper–tension control disturbances.
Another important future research direction, which is
This section gives some outlines of what we believe to be strongly related to looper–tension control, is looperless
possible areas of fruitful development. control. Some techniques [22] have already been applied
Up to now, many advanced control algorithms have in industry, but these were restricted to short time periods
been proposed and applied to looper–tension control prob- before looper control starts. The control law is based on
lems. However, it is notable from Sections 3 and 4 that tension estimation by using mill motor torque and rolling
none of these schemes is entirely satisfactory. Nevertheless, force etc. and manipulates the speed of an upstream main
it has become clear that one of the most important design motor. The major advantages are a reduction in the main-
issues is the rejection of unknown mass flow disturbances; tenance requirements of loopers as well as improved
these mostly arise from AGC action and can sometimes tension control accuracy at the top part of the strip.
induce severe quality defects of a strip as well as process Unsurprisingly, removal of the looper control loop also
instability [18]. That means AGC and looper–tension con- simplifies a MIMO control problem to a SISO problem,
trol have been implemented independently; AGC is used so there is no longer an interaction problem induced
only for thickness control and its effect on mass flow is between tension and looper angle. However, a main obsta-
not considered. Therefore, improvement would be possible cle to the efficacy of this technique is the need for accurate
by better coordination among thickness, tension and tension estimation without availability of the looper
massflow. parameters [25].
Some multivariable design approaches [4,9,30,31] have
considered interactions among gauge, tension and looper, 6. Conclusions
and have included a gauge model in the looper–tension
control design. However, none of them had a systematic In order to improve control performance and stability for
coordination between mass flow and thickness. Recently, looper–tension control, various control algorithms have
[19] emphasized the importance of the interaction between been developed. Conventional PI control schemes have been
loop length and exit thickness, and tried a performance used most widely in industry regardless of the performance
520 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521

limitations, because they do not need tension meters. Where [15] Y. Kotera, F. Watanabe, Multivariable control of hot strip mill
a tension measurement is available, many multivariable con- looper, in: Proceedings of 1981 IFAC Eighth World Congress, 1981,
XVIII, pp. 1–6.
trol algorithms have also been applied, but, due to increased [16] J. Schuurmans, T. Jones, Control of mass flow in a hot strip mill using
complexity, these are far more difficult to tune. This paper model predictive control, in: Proceedings of the 2002 IEEE Interna-
gives a survey of all these various control algorithms, draw- tional Conference on Control Applications, 2002, pp. 379–384.
ing together the strengths and weaknesses and hence demon- [17] M. Shioya, N. Yoshitani, T. Ueyama, Noninteracting control with
strating some areas of potential future development. In disturbance compensation and its application to tension looper
control for hot strip mill, in: Proceedings of the 1995 IECON, 1995,
particular it is noted that none of the schemes as yet pro- pp. 229–234.
posed in the literature combine simplicity and effectiveness [18] I.S. Choi, J.A. Rossiter, P.J. Fleming, An application of the model
to a degree that is desirable. based predictive control in a hot strip mill, in: Proceedings of the 11th
We feel that one avenue that has been substantially IFAC Symposium on Automation in Mining, Mineral and Metal
under explored is model predictive control. This has the Processing, 2004, Wed-C1.
[19] G. Hearns, P. Reeve, T.S. Bilkhu, P. Smith, Multivariable gauge and
facility to handle the large interactions (both inter- and mass flow control for hot strip mills, in: Proceedings of the 11th IFAC
intra-stand) within finishing mills and at the same time to Symposium on Automation in Mining, Mineral and Metal Process-
take proper account of the hard constraints within the sys- ing, 2004, Wed-C1.
tem. Moreover, it is possible in principle to take systematic [20] J.A. Rossiter, B. Kouvaritakis, M.J. Rice, A numerically robust state-
account of both the known and unknown disturbances, the space approach to stable predictive control strategies, Automatica 34
(1998) 65–73.
rejection of which constitutes a main control challenge. The [21] POSCO Ltd., HITACHI Ltd., Development of dynamic simulator on
main immediate goals in future work are the design of a hot finishing mills, Technical Research Report, POSCO, 1997.
suitable model and then the formulation of a robust [22] POSCO Ltd., Development of tension control techniques on hot
MPC algorithm which can be implemented at fast enough rolling mills, Technical Research Report, POSCO, 2001.
sample rates to be useful in the real industrial environment. [23] T. Umeno, T. Kaneko, Y. Hori, Robust servosystem design with two
degrees of freedom and its application to novel motion control of
robot manipulators, IEEE Transactions on Industry Electronics 40
References (1993) 473–485.
[24] K. Asano, M. Morari, Interaction measure of tension–thickness
[1] H. Asada, A. Kitamura, S. Nishino, M. Konishi, Adaptive and robust control in tandem cold rolling, Control Engineering Practice 6 (1998)
control method with estimation of rolling characteristics for looper 1021–1027.
angle control at hot strip mill, ISIJ International 43 (2003) 358–365. [25] H. Katori, R. Hirayama, T. Ueyama, K. Furura, On the possibility of
[2] K. Asano, K. Yamamoto, T. Kawase, N. Nomura, Hot strip mill looperless rolling on hot rolling process, in: Proceedings of the 1999
tension–looper control based on decentralization and coordination, IEEE International Conference on Control Applications 1999, MA4-
Control Engineering Practice 8 (2000) 337–344. 4, pp. 18–22.
[3] G. Hearns, M.J. Grimble, Inferential control for rolling mills, IEE [26] B. Bulut, A.W. Ordys, M.J. Grible, Application of efficient nonlinear
Proceedings of Control Theory and its Applications 147 (2000) 673–679. predictive control to a hot strip finishing mill, in: Proceedings of the
[4] G. Hearns, M.J. Grimble, Robust multivariable control for hot strip 2002 IEEE International conference on Control Applications 2002,
mills, ISIJ International 40 (2000) 995–1002. pp. 373–378.
[5] T. Hesketh, D.J. Clements, D.H. Buttler, R. Lann, Controller design [27] B. Bulut, D. Greenwood, M.J. Grible, Load distribution ration as a
for hot strip finishing mills, IEEE Transactions on Control Systems benchmark for flatness using predictive control, in: Proceedings of the
Technology 6 (1998) 208–219. 41st IEEE Conference on Decision and Control, 2002, TuM05—5,
[6] W.K. Hong, P.H. Kim, Y.H. Moon, J.J. Yi, Hot strip width control pp. 555–560.
method by using looper tension measuring system in finishing mill, [28] D. Uduehi, A.W. Ordys, M.J. Grible, A generalized predictive control
Journal of materials processing technology 111 (2001) 74–78. benchmark index for MIMO systems, in: Proceedings of the 2002
[7] H. Imanari, Y. Morimatsu, K. Sekiguchi, H. Ezure, R. Matuoka, A. IEEE International Conferences in Control Applications, 2002, pp.
Tokuda, H. Otobe, Looper H-Infinity control for hot strip mill, IEEE 1213–1218.
Transactions on Industry Applications 33 (1997) 790–796. [29] M.J. Grimble, P. Martin, Restricted structure adaptive predictive
[8] F. Janabi-Sharifi, A neuro-fuzzy system for looper tension control in control of nonlinear systems, in: Proceedings of the 2002 Interna-
rolling mills, Control Engineering Practice 13 (2005) 1–13. tional Symposium on Computer Aided Control Systems Design
[9] M. Okada, K. Murayama, A. Urano, Y. Iwasaki, A. Kawano, H. (2002) pp 663–668.
Shiomi, Optimal control system for hot strip finishing mill, Control [30] Y. Kadoya, T. Ooi, Y. Washikita, Y. Seki, Strip gage and tension
Engineering Practice 6 (1998) 1029–1034. control at cold tandem mill based on I.L.Q. design theory, in:
[10] Y. Seki, K. Sekiguchi, Y. Anbe, K. Fukushima, Y. Tsuji, S. Ueno, Proceedings of the 1999 IEEE International Conference on Control
Optimal multivariable looper control for hot strip finishing mill, Applications, 1999, MA4-5, pp. 23–28.
IEEE Transactions on Industry Applications 27 (1991) 124–130. [31] R. Takahashi, State of the art in hot rolling process control, Control
[11] J.C. Price, The hot strip mill looper system, IEEE Transactions on Engineering Practice 9 (2001) 987–993.
Industry Applications, IA 9 (5) (1973) 551–562. [32] Y. Fujisaki, A. Kitamura, H. Asada, Robust control of looper angle
[12] J.Y. Jung, Y.T. Im, Fuzzy algorithm for the prediction of tension for hot strip mills, Transactions of Society of Instrument and Control
variations in hot rolling, Journal of Materials Processing Technology Engineers 27 (1991) 723–725.
96 (1999) 163–172. [33] J.A. Rossiter, Model-based predictive control: A practical approach,
[13] I. Hoshino, Y. Okamura, H. Kimura, Observer-based multivariable CRC Press, 2003.
tension control of aluminum hot rolling mills, in: Proceedings of the 1996 [34] Y. Anbe, K. Sekiguchi, H. Imanari, Tension control of a hot strip mill
of the 35th Conference on Decision and Control, 1996, pp 1217–1222. finisher, in: Proceedings of 13th IFAC World Congress, 1996, pp.
[14] I. Hoshino, H. Kimura, Observer-based multivariable control of 439–444.
rolling mills, in: Proceedings of 8th IFAC Symposium on Automation [35] J. Fetyko, E. Hutnı́k, J. Čverčko, V. Fedák, Observer based
in Mining, Mineral and Metal Processing, 1998, pp. 251–256. tension control in hot-strip finishing mill, in: Proceedings of 11th
 I.S. Choi et al. / Journal of Process Control 17 (2007) 509–521 521

International Conference on Power Electronics and Motion Control, [38] A. Randall, L. Balmer, K.J. Burnham, M.J. Chapman, Hot strip mill
2004, pp. 94–99. tension control-multivariable techniques, Systems Science 21 (4)
[36] Z.B. Jamaa, B. Petit, P. Borne, Multivariable controls of a hot strip (1995) 89–100.
finishing mill interstand, Studies in Informatics and Control 5 (1) [39] M. Konishi, S. Imajo, J. Imai, T. Nishi, Modelling of gain tuning
(1996) 41–48. operation for hot strip looper controller by recurrent neural network,
[37] B. Petit, P. Borne, M. Verge, Comparative analysis of multivariable in: Proceedings of the 2004 IEEE International Conference on
advanced controls applied to the hot rolling mill interstand process, Control Applications, 2004, pp. 890–895.
in: Proceedings of 7th International Conference on Steel Rolling,
1998, pp. 73–80.