P. 1
Landau RST

Landau RST

|Views: 53|Likes:
Published by pbaculima

More info:

Published by: pbaculima on Jun 12, 2011
Copyright:Attribution Non-commercial

Availability:

Read on Scribd mobile: iPhone, iPad and Android.
download as PDF, TXT or read online from Scribd
See more
See less

01/06/2013

pdf

text

original

*E-mail: landau@lag.ensieginpg.

fr
Control Engineering Practice 6 (1998) 155—165
The R-S-T digital controller design and applications
I.D. Landau*
Laboratoire d+Automatique de Grenoble (CNRS/INPG/UJF), ENSIEG — BP 46 — 38402 Saint-Martin d+He` res Cedex, France
Received September 1997; in revised form November 1997
Abstract
The two-degrees-of-freedom R-S-T digital controller is becoming a standard for computer control in industry. This paper presents
a methodology for the design of the R-S-T controller, which involves identification of the plant model from data, combined with
a robust control design. The performance of the controller can be further enhanced by plant model identification in a closed loop, and
re-tuning of the controller. For large parameter variations, adaptation has to be considered in order to maintain the performance.
Software packages are available for the design, implementation and commissioning of the R-S-T digital controllers. The methodology
is illustrated by its application to the control of deposited zinc in hot-dip galvanizing at SOLLAC(Florange, France). 1998 Elsevier
Science Ltd. All rights reserved.
Keywords: System identification; digital control; robust control; adaptation; software tools
1. Introduction
A ‘‘good’’ control system has in general an important
economic impact in industry. Fig. 1 illustrates the histo-
gram of a controlled variable for ‘‘poor’’ control and for
‘‘good’’ control, respectively.
If the variance of the controlled variable is high, a
significant number of the measurements will lie far from
the desired value. In a large number of applications,
a minimal acceptable value is imposed (e.g., the humidity
of the paper, the depth of the coating, etc.) and poor
quality of the control will require the choice of a higher
value for the reference. As a consequence, more energy or
material will be needed, and the direct consequence is an
increase in production costs.
If one has a ‘‘good’’ controller, which significantly
reduces the variance of the controlled variable around
the reference value, this will on the one hand improve the
quality and, on the other hand, will allow a reduction of
the reference value. This leads in general to energy and
material savings which correspond to a reduction in
production costs.
Therefore, the impact of ‘‘good’’ control is:
(1) Improvement in the quality of the products.
(2) Energy and material savings.
However, it is important that the investment return
gained by improvements in control can be clearly evalu-
ated, in order to justify the investment.
The question is: How does one improve the investment
return for high-performance control systems?
The answers to this question are:
(1) Reduction in the design cost.
(2) Reduction in the implementation cost (including
commissioning).
(3) Achievement of the desired performance.
Therefore, it is necessary to develop an efficient design
and implementation methodology. This development has
to be considered in the context of computer control,
which is now widespread in industry. All the advantages
and features of using computers for control have to be
taken into account. Among these aspects, those of system
identification and the introduction of a standard form for
a digital controller (the R-S-T controller) play a crucial
role.
Fig. 2 summarizes the basic principles of control
system design. In order to design and to tune a good
0967-0661/98/$19.00 1998 Elsevier Science Ltd. All rights reserved
PII S 0 9 6 7 - 0 6 6 1 ( 9 8 ) 0 0 0 1 6 - 1
Fig. 1. Histograms for good and poor control.
Fig. 2. Principle of controller design.
Fig. 3. Computer control system. D.A.C.: Digital-to-Analog Conver-
ter, Z.O.H.: Zero-Order Hold, A.D.C.: Analog-to-Digital Converter.
controller one needs:
(1) To specify the desired control-loop performances.
(2) To have a dynamic model of the plant to be con-
trolled (this can be obtained from real data by identi-
fication).
(3) To possess a suitable controller design methodology,
compatible with the desired performance and the
corresponding plant model.
(4) To have a procedure for controller validation and on-
site re-tuning.
(5) To have appropriate software packages with
real-time capabilities for data acquisition, system
identification, control design and on-site commis-
sioning.
This paper will present such a methodology for the
design and application of R-S-T digital controllers. The
methodology will be illustrated by its application to the
control of the deposited zinc in hot-dip galvanizing at
SOLLAC (Florange, France).
2. Identification of discrete-time models for
industrial processes
Fig. 3 illustrates an appropriate setting for a computer
control system. The set D.A.C.# plant # A.D.C. is
interpreted as a discretized system, whose control input is
the sequence ¦u(t)¦ generated by the computer, the out-
put being the sequence ¦y(t)¦ resulting from the A/D
conversion of the system output y(t). The discretized
plant is characterized by a discrete-time model, which
should be identified.
Note that the sampling frequency is selected in accord-
ance with the bandwidth of the continuous-time plant,
and more specifically in accordance with the desired
bandwidth of the closed loop. The basic rule is:
f
Q
"(6 to 25) f !J

where f
Q
is the sampling frequency and f !J

is the desired
bandwidth of the closed loop.
The discrete-time model of the plant to be controlled is
described in the time-domain by:
y(t)"!
L

G¯¹
a
G
y(t!i)#
L

G¯¹
b
G
u(t!d!i)
where d is the integer number of sampling periods con-
tained in the time-delay of the plant, t is the normalized
discrete time (0, 1, 2, 3,
2
) and corresponds to the
discrete time divided by the sampling periods ¹
Q
. The
discrete-time model of the plant to be controlled can
alternatively be represented by its pulse transfer operator
H(q¹):
y(t)"H(q¹) u(t).
H(q¹) is defined by:
G(q¹)"
qB B(q¹)
A(q¹)
"
qB¹ B*(q¹)
A(q¹)
where q¹ is the backward shift operator (q¹ y(t)"
y(t!1)) and
A(q¹)"1#a
¹
q¹#
2
a
L
qL
B(q¹)"b
¹
q¹#
2
b
L
qL"q¹B*(q¹).
156 I.D. Landau/ Control Engineering Practice 6 (1998) 155—165
Fig. 4. Parameter estimation of discrete-time models.
Fig. 5. The R-S-T canonical structure of a digital controller.
For models with constant parameters, replacing q by
z (the complex variable) in the expression of the pulse
transfer operator gives the pulse transfer function.
The principle of the identification of discrete-time
models is illustrated in Fig. 4.
A discrete-time model with adjustable parameters is
implemented on the computer. The error between the
system output at instant t, y(t), and the output predicted
by the model, yL(t), (known as the prediction error) is used
by a parameter-adaptation algorithm which, at each
sampling instant, will modify the model parameters in
order to minimize this error. The input is in general
a very low-level pseudo-random binary sequence, gene-
rated by the computer (sequence of rectangular pulses
with randomly variable duration). Once the model has
been obtained, an objective validation can be made by
carrying out statistical tests on the prediction error c(t)
and the predicted output yL(t). The validation test enables
the best algorithm to be obtained for the estimation of
the parameters.
This approach provides much more accurate models
than the methods based on step response or frequency
response. In addition, it requires an input signal of much
lower magnitude than those used for step or frequency
response.
The identification methodology includes four steps:
(1) Input-output data acquisition around an operating
point, using as input in general a centered pseudo-
random binary sequence (PRBS) of small magnitude,
(2) Estimation (choice) of the model complexity
(structure),
(3) Estimation of the model parameters,
(4) Validation of the identified model (structure and
values of the parameters).
One of the important facts to be emphasized is that the
plant measurements are generally noisy. Unfortunately,
no unique parameter-estimation method exists which
may be used successfully for all the types of disturbances,
such that the estimated parameters are always unbiased.
Therefore, a good identification of a plant model gene-
rally requires the use of an interactive system featuring
various parameter-estimation methods and the corre-
sponding validation techniques.
For a more detailed discussion see (Landau, 1990;
Ljung, 1987).
3. The R-S-T digital controller
The canonical structure of the R-S-T digital controller
is represented in Fig. 5.
This structure has two degrees of freedom, i.e., the
digital filters R and S are designed in order to achieve the
desired regulation performance, and the digital filter ¹ is
designed afterwards in order to achieve the desired track-
ing performance. This structure allows achievement of
different levels of performance in tracking and regulation.
The case of a controller operating on the regulation
error (which does not allow the independent specification
of tracking and regulation performance) corresponds to
¹"R. Digital PID can also be represented in this form,
leading to particular choices of R, S and T.
The equation of the R-S-Tcanonical controller is given
by:
S(q¹) u(t)#R(q¹) y(t)"¹(q¹) y*(t#d#1)
where u(t) and y(t) are the input and output of the plant
and y*(t#d#1) is the desired tracking trajectory,
which is either generated by a tracking reference model
(Bm/Am) or stored in the computer memory.
The polynomials R(q¹), S(q¹), ¹(q¹) have the
form:
R(q¹)"r
"
#r
¹

2
#r
L0
qL0
S(q¹)"s
"
#s
¹

2
#s
L1
qL1 (often s
"
"1)
¹(q¹)"t
"
#t
¹

2
t
L2
qL2.
The corresponding time-domain expression of the con-
trol law is given by (s
"
"1):
u(t)"!
L1

G¯¹
s
G
u(t!i)!
L0

G¯"
r
G
y(t!i)
#
L2

G¯"
t
G
y*(t#d#1!i).
I.D. Landau/ Control Engineering Practice 6 (1998) 155—165 157
The closed-loop control transfer operator (between r(t)
and y(t)) is given by:
H
''
(q¹)"
qB ¹(q¹) B(q¹)
A(q¹) S(q¹)#qB B(q¹) R(q¹)
"
qB ¹(q¹) B(q¹)
P(q¹)
,
and the behaviour with respect to an output disturbance
is given by the output sensivity function:
S
WN
(q¹)"
A(q¹) S(q¹)
A(q¹) S(q¹)#qB B(q¹) R(q¹)
"
A(q¹) S(q¹)
P(q¹)
where P(q¹) defines the desired closed-loop poles (regu-
lation behaviour).
The input sensitivity function which reflects the effects
of an output disturbance upon the plant input is given by:
S
SN
(q¹)"!
A(q¹) R(q¹)
P(q¹)
.
In general, the desired closed-loop poles are specified
in the form:
P(q¹)"P
"
(q¹) P
$
(q¹)
where P
"
(q¹) specifies the desired dominant poles of the
closed loop, and P
$
(q¹) specifies the auxiliary poles of
the closed loop.
Once the closed-loop poles have been defined, solving
the equation:
P(q¹)"A(q¹) S(q¹)#q° B(q¹) R(q¹)
allows the determination of S(q¹) and R(q¹), which
will ensure the desired closed-loop poles.
Let the degrees of polynomials A(q¹) and B(q¹) be
defined by:
n

"deg A(q¹); n

"deg B(q¹).
Then the above equation has a unique solution (assum-
ing that A(q¹) and B(q¹) do not have common factors)
for:
n
.
"deg P(q¹))n

#n

#d!1
n
1
"deg S(q¹)"n

#d!1
n
0
"deg R(q¹)"n

!1
in which:
S(q¹)"1#s
¹
q¹#
2
s
L1
qL1"1#q¹ S*(q¹)
R(q¹)"r
"
#r
¹
q¹#
2
r
L0
qL0.
However, in general the polynomials R(q¹) and S(q¹)
of the controller may, for various reasons, contain some
prespecified fixed parts. For this reason it is convenient
to factorize the polynomials R(q¹) and S(q¹) as
follows:
R(q¹)"H
0
(q¹) R' (q¹)
S(q¹)"H
1
(q¹) S' (q¹)
where H
0
(q¹) and H
1
(q¹) are prespecified polyno-
mials. These polynomials are defined by the performance
specifications (e.g., the integrator in the controller) and
by considerations of robustness. For example, the intro-
duction of an integrator in the controller requires one to
take H
1
(q¹)"1!q¹.
Many control strategies can be applied to the design of
the R-S-T controller by an appropriate reformulation.
See for example (Landau, 1990; Astro¨ m and Wittenmark,
1990; Landau et al., 1997), as well as the special issue of
European Journal of Control Vol. 1, N°2, 1995, dedicated
to a robust control benchmark.
However, the pole placement can be considered as the
basic design technique, and most of the various designs
can be related to it.
3.1. Pole placement
This control strategy can be used for plant models of
any order, with or without time delay, and featuring
stable or unstable zeros. The only assumption is that the
polynomials A(q¹) and B(q¹) characterizing the plant
model do not have common factors.
The controller polynomials S(q¹)"H
1
(q¹) S' (q¹)
and R(q¹)"H
0
(q¹) R' (q¹) are obtained by solving
the equation:
P(q¹)"A(q¹) H
1
(q¹) S' (q¹)
#qB¹ B*(q¹) H
0
(q¹) R' (q¹)
where P(q¹) defines the desired closed-loop poles.
The ¹(q¹) polynomial is chosen as:
¹(q¹)"P(q¹) /B(1)
(B(1)"B(q¹) for q"1), and the tracking behaviour is
described by the equation:
y(t)"qB¹
B*(q¹)
B(1)
y*(t#d#1).
In other words one follows the desired trajectory, filtered
through the plant zeros. For more details see (Landau,
1990, 1993).
Note that unstable discrete-time zeros occur when
a fractional delay larger than half of the sampling period
is present, or when a high sampling frequency is used for
158 I.D. Landau/ Control Engineering Practice 6 (1998) 155—165
Fig. 6. Modulus, gain and phase margins.
continuous-time models having a difference of degree
between denominator and numerator greater than or
equal to 2 (Astro¨ m and Wittenmark, 1990; Landau,
1990). Both phenomena lead to the conclusion that
the sampling frequency has to be chosen as low as pos-
sible, but in accordance with the desired closed-loop
performance.
3.2. Relationship with other control strategies
For plant models with stable zeros, one can use the
strategy called ‘‘tracking and regulation with indepen-
dent objectives’’, which can be viewed as a particular case
of the pole-placement strategy, where the desired closed-
loop poles contains the zeros of the plant, i.e.,
P(q¹)"P
"
(q¹) ) B*(q¹) ) P
$
(q¹).
‘‘Internal model control’’ (Morari, 1989) corresponds to
a pole-placement strategy where the desired closed-loop
poles contain the poles of the plant model, i.e.,
P(q¹)"A(q¹) ) P
$
(q¹).
Digital PID can be designed using pole placement. In
this case, the orders of the plant model are limited to
n

)2, n

)2, d"0.
Predictive control strategies and an LQ control strat-
egy using an appropriate formulation of the criterion to
be minimized lead to an R-S-T controller, and can be
viewed as an approximation of the pole placement in the
sense of a certain quadratic criterion (Landau et al.,
1997).
4. Robustness
Four indicators are generally used to express the
robustness of a design in terms of the minimal distance
with respect to the critical point [!1, j0] in the Nyquist
plane. These indicators are the gain margin (G), the
phase margin (), the modulus margin (M) and the
delay margin (t). The modulus margin and the delay
margin are the most interesting in applications. Fig. 6
illustrates the modulus, gain and phase margins.
The modulus margin is the minimal distance between
the critical point [!1, j0] and the Nyquist plot of the
open-loop transfer function:
H
''
(z¹)"z° B(z¹) R(z¹)/A(z¹) S(z¹).
The modulus margin M is defined as the radius of
a circle, centered on [!1, j0] and tangent to the Nyquist
plot of H
''
(eHU) (see Fig. 6).
The result is that:
M"¦1#H
''
(eHU) ¦
°'"
.
"¦S¹
WN
(eHU) ¦
°'"
.
"(¦S
WN
(eHU) ¦
°º`
)¹.
In other words, the modulus margin corresponds to the
inverse of the H
`
norm of the output sensitivity function.
Minimization of the H
`
norm of S
WN
(eHU) will maximize
the modulus margin.
To obtain the modulus margin, it is therefore sufficient
simply to plot the frequency characteristics of the modu-
lus (gain) of the output sensitivity function in dB. In this
case:
M dB"(¦S
WN
(eHU) ¦
°º`
)¹ dB
"!¦S
WN
(eHU) ¦
°º`
dB.
Note that a given value of the modulus margin will
guarantee certain phase and gain margins, while the
converse is not true (systems with good phase and gain
margins can pass very close to the critical point).
The delay margin is the maximal additional delay that
will be tolerated in the open-loop system without causing
an instability of the closed-loop system. If the Nyquist
plot of the open-loop system intersects the unit circle at
several cross-over frequencies cG
°'
, characterized by the
corresponding phase margins
G
, the delay margin of
the system is defined as:
t"min
G

G
cG
°'
.
Typical values of these robustness indicators are:
— modulus margin: M*0.5 (!6 dB)
— delay margin: t*¹
1
(sampling period)
— gain margin: G*2 (6 dB)
— phase margin: 30°))60°.
Note that M*0.5 implies G*2 and '29° (the
converse is not necessarily true).
Sensitivity functions are related to the robust stability
of the closed loop with respect to plant model uncertain-
ties, see (Doyle et al., 1992). Bounds on the magnitude of
the frequency-dependent model uncertainties convert
to upper constraints upon the moduli of the various
I.D. Landau/ Control Engineering Practice 6 (1998) 155—165 159
Fig. 7. Templates for the sensitivity functions: (a) output sensitivity
function; (b) input sensitivity function.
Fig. 8. Plant model identification in closed-loop operation.
sensitivity functions. These constraints, as well as those
on the modulus margin and delay margin, can be trans-
lated into desired templates for the sensitivity functions.
Conversely, constraints imposed on the sensitivity func-
tions can be translated into tolerated model uncertainties.
Typical templates for S
WN
and S
SN
are shown in Fig. 7.
The upper and lower bounds on the output sensitivity
function S
WN
in the high-frequency region come from the
translation of the delay margin constraints in the fre-
quency domain (Landau, 1995).
The input sensitivity function reflects both tolerance
with respect to additive uncertainties, and the activity of
the input in the presence of disturbances (high values in
certain frequency regions indicate low model uncertainty
tolerance and important stress on the actuator).
A design methodology that combines pole placement
with the shaping of the sensitivity function has been
developed in order to ensure both performance and
robustness. See (Landau, 1993; Landau et al., 1996).
5. Identification in a closed loop
In a number of practical situations, it may not be
possible to operate the plant in an open loop in order to
achieve system identification. Such situations are en-
countered, for example, when the plant contains an inte-
grator, or when important drifts of the operating point
may occur during input/output data acquisition. In
a number of other situations, a controller already exists,
but for various reasons cannot be disconnected. There-
fore, techniques for plant identification in closed-loop
operation should be used.
Significant progress in this area has been made in the
last few years. Newalgorithms dedicated to identification
in closed loops have been developed. For a detailed
presentation see (Van den Hof and Schrama, 1995;
Landau and Karimi, 1997; Landau et al., 1997).
A basic scheme for plant model identification in a
closed loop is shown in Fig. 8 (Landau and Karimi,
1997). The upper part represents the true closed-loop
system, and the lower part corresponds to an adjustable
predictor for the closed loop, re-parameterized in terms
of a known fixed controller and an adjustable plant
model. The error between the system output and the
closed-loop predictor output (called a closed-loop output
error) is used by a parameter-adaptation algorithm that
will drive the parameters of the estimated plant model in
order to minimize the error between the two closed-loop
systems. In other words, the model obtained in a closed
loop will allow for better prediction of the behaviour of
the closed-loop system.
6. Controller validation and on-line retuning
As indicated in Section 1, validation of the designed
controller is a key issue in assessing the effective perfor-
mance of the closed-loop system. Identification of the
closed loop will allow the performance achieved to be
compared with the designed performance by comparing
the achieved and the designed closed-loop poles. It will
also allow the robustness of the control schemes to be
assessed by comparing the designed and achieved sensi-
tivity functions in the frequency domain.
However, with the same data, acquired in a closed
loop, one can also identify a new plant model. The new
160 I.D. Landau/ Control Engineering Practice 6 (1998) 155—165
Fig. 9. ‘‘Closed loop’’ adaptive control.
Fig. 10. ‘‘Open-loop’’ adaptation.
model, identified in the closed loop, is then used for
re-tuning of the controller.
This procedure is used for two purposes:
— improvement of a previous design
— controller maintenance.
Examples can be found in (Zhang et al., 1995; Langer and
Landau, 1996; Landau et al., 1997).
7. Adaptation
When ‘‘system identification’’ plus ‘‘robust control
design’’ does not allow one to obtain a single linear
controller, giving acceptable performance for the whole
range of operating points because of the too-wide varia-
tions in the dynamic characteristics of the plant, one has
to consider the ‘‘adaptation’’ of the controller.
The term ‘‘adaptation’’ (adaptive control) refers to a set
of techniques for the automatic tuning of the controller in
real time, in order to maintain the desired performance
when the plant parameters vary.
One can distinguish between two basic adaptive con-
trol techniques:
(1) ‘‘closed-loop’’ adaptive control (Fig. 9)
(2) ‘‘open-loop’’ adaptive control (Fig. 10).
The ‘‘closed-loop’’ adaptive control system usually
combines a real-time identification algorithm with the
computation of the controller in real time, based on
a current estimation of the plant model and the desired
performance.
However, in a number of applications, the characte-
ristics of the dynamic model of the plant depend upon
a set of measured variables, which define an operating
point. In this case, one can use an ‘‘open-loop’’ adaptive
control (Fig. 10). The range of operating points is divided
into a number of operating intervals. For each interval,
a relevant operating point is selected and a correspond-
ing controller is designed, based on an identified model.
This controller assures the desired performance for all the
operating points located in the interval. The correspond-
ing controllers are stored in a table. When the plant is
operating at a certain point, the corresponding values of
the controller parameters will be used, according to the
table.
8. Hot-dip galvanizing at SOLLAC (Florange)
The objective of the galvanizing line is to obtain
galvanized steel with formability, surface quality and
weldability equivalent to uncoated cold rolled steel. The
variety of products is very large in terms of deposited zinc
thickness and steel strip thickness. The deposited zinc
may vary between 50 to 350 g/m` (each side), and the
strip speed may vary from 30 to 180 m/mn.
The most important part of the process is the hot-dip
galvanizing. The principle of hot-dip galvanizing is illus-
trated in Fig. 11. Preheated steel strip is passed through
a bath of liquid zinc, and then rises vertically out of the
bath through the stripping ‘‘air knives’’, which remove
the excess zinc. The remaining zinc on the strip surface
solidifies before it reaches the rollers, which guide the
finished product. The measurement of the deposited zinc
can be made only on the cooled, finished strip. The effect
of the air knives depends on the air pressure, the distance
between the air knives and the strip, and the speed of the
strip. Nonlinear static models have been developed for
computing the appropriate pressure, distance and speed
for a given desired value of the deposited zinc.
The objective of the control is to assure good unifor-
mity of the deposited zinc, whilst guaranteeing a mini-
mum value of the deposited zinc per unit area. Tight
control (i.e., a small variance in the controlled variable)
will allow a more uniform coating and a reduction of the
average quantity of deposited zinc per unit area. As
a consequence, in addition to an improvement in quality,
tight control of the deposited zinc per unit area has an
important commercial impact since the average con-
sumption for a modern galvanizing line is of the order of
40 tons per day (price+1 500 USD/ton).
I.D. Landau/ Control Engineering Practice 6 (1998) 155—165 161
Fig. 11. Process description.
9. Model of the process
For an analysis of the process, the model originally
proposed by Harvey and Carton (1974) and completed
by Jacobs (1991) can be used:
m"KD

»
P

K
where m is the deposited mass per unit area, K is a con-
stant of proportionality, D is the distance between the air
knives and the strip, P is the air pressure and » is the
strip speed. ¨
K
accounts for unpredictable effects and/or
modelling errors. At SOLLAC Ste Agathe, the control
variable is the air pressure.
A linearized model around an operating point (P
"
, »
"
,
D
"
) can be obtained using a standard Taylor series ex-
pansion for variations of pressure (P), speed (») and
distance (D). It has the form:
m"KD
"

»
"
P
"
#: D#[ »!j P#.
K
;
:, [, j'0.
It can be seen that, by using the pressure as the control
variable, one can compensate for the disturbances
created by variations in distance and speed as well as by
the term .
K
.
The pressure in the air knives is regulated through
a pressure loop, which can be approximated by a first-
order system. The delay of the process will depend
linearly on the speed. Therefore, a continous-time linear
dynamic model, relating variations in the pressure to
variations in the deposited mass, of the form:
H(s)"
GeQ
t
1#s¹
; t"
¸
»
can be considered, where ¸ is the distance between the air
knives and the transducers, and » is the strip speed.
When discretizing this model, the major difficulty comes
from the variable time-delay. In order to obtain a con-
troller with a fixed number of parameters, the delay of the
discrete-time model should remain constant. Therefore,
the sampling period is tied to the strip speed using the
formula:
¹
Q
"
(¸/»)#o
d
; (d"integer)
where d is a small additional time-delay due to the
implementation, and d is the discrete-time delay (an
integer).
The corresponding linearized discrete-time model will
be of the form:
H(q¹)"
qB (b
¹
q¹)
1#a
¹

.
The fractional delay (which corresponds to the presence
of an additional term b
`
q`) is negligible because of the
way in which the sampling period ¹
Q
is selected; this was
confirmed by the model identification procedure. How-
ever, the parameters of the model, given above, will
depend on the distance D and on the speed ».
10. Identification of the discrete-time plant model
The process comprises the air-pressure control loop
and the coating process. The control input to the process
is the reference of the air pressure control loop, and the
output of the process is the measured deposited mass per
unit area (see Fig. 12).
The PC used for data acquisition, identification and
control is connected to the process through an industrial
network. The identification has been done with the coat-
ing process operating in an open loop.
The sampling frequency has been chosen at each ope-
rating point, in order to have the discrete-time delay
d"7. The data acquisition is illustrated in Fig. 13. First,
an analog anti-aliasing filter is used, before a high-
frequency sampling is undertaken (a multiple of the
desired sampling frequency). A digital anti-aliasing filter
is inserted between the two samplers.
The input used was a P.R.B.S. (Pseudo Random
Linary Sequence) of a magnitude of $4% with respect
to the static pressure (P
"
). The P.R.B.S. was generated by
a shift register with N"5 cells and a clock frequency
equal to half of the sampling frequency (length of the
162 I.D. Landau/ Control Engineering Practice 6 (1998) 155—165
Fig. 12. Process block diagram.
Fig. 13. Data acquisition.
sequence: 64). 100 to 160 (average: 128) measurements
have been used for the various identifications made in the
different regions of operation. The choice made for the
P.R.B.S. allowed at least one full sequence to be sent for
each experiment, and yielded the largest pulse width
(10 ¹
'
) comparable with the rise time of the process. As
both sides of the steel strip have to be galvanised, and
because the positions and physical realisations of the two
actuators are not symmetrical, both ‘‘front’’ and ‘‘back’’
models have been identified.
As no unique identification method gives unbiased
results for all types of disturbances, the following recur-
sive identification methods have been used (Landau,
1990):
— Recursive least squares
— Extended least squares
— Recursive maximum likelihood
— Instrumental variable with auxiliary model
— Output error
— Generalized least squares
For validation of the identified models and compari-
son of the models obtained with the different methods,
a cross-correlation between the predicted output (using
an output error predictor) and the output error has been
used (Landau, 1990). (For the first four methods, the
whiteness test on the prediction error has also been used
for validation and comparison). ‘‘Output error’’ consis-
tently provided the best results for this application.
It was observed that a significant variability in the
parameters occurs with a change of the operating points.
This necessitated splitting the operation of the plant into
several regions. However, a variability of the parameters
is observed, even within a region of operation at a con-
stant distance and with relatively small speed variations.
One of the causes is the imperfect measurement of the
strip/air-knives distance. This variability will require
a robust control design.
11. Controller design and adaptation
The ‘‘tracking and regulation with independent objec-
tives’’ (which in this case is equivalent to the poles place-
ment, since the model does not have finite zeros) has been
used.
Robust control design using an identified model
and based on shaping of the sensitivity function
allowed a modulus margin greater than !6 dB, to
be obtained with a delay margin greater than 2¹
Q
.
These robustness margins assure satisfactory perfor-
mance in a region of operation, despite the variability of
the model.
In order to ensure satisfactory performance for all
regions of operation, an ‘‘open-loop adaptation’’ tech-
nique has been considered. The open-loop adaptation is
made with respect to:
— steel strip speed,
— distance between the air knives and the steel strip.
The strip speed directly affects the sampling period
according to the relationship:
¹
Q
"
¸#o
»
where o is the equivalent time-delay of the industrial
network and of the programmable controller used for
pressure regulation. The speed range and the distance
range have been split into 3 regions, giving a total of
9 operating regions. For each of these operating regions
an identification has been performed, and controllers
based on the identified model have been computed and
stored in a table.
Anti-wind-up procedures have been used for the imple-
mentation of the controller, and a smooth transfer from
open-loop to closed-loop operation has also been
assured. For a detailed presentation of this application
see (Fenot et al., 1993).
12. Results
Fig. 14 shows typical results, obtained when one of the
sides is under digital regulation and the other side is
I.D. Landau/ Control Engineering Practice 6 (1998) 155—165 163
Fig. 14. Typical performance of the digital regulation of the deposit zinc.
under computer-aided manual control (the operator has
on display a moving short-time history of the deposited
zinc and applied pressure).
A reduction in the dispersion of the coating is noticed
when closed-loop digital control is used. This provides
a better-quality finished product (extremely important in
the automotive industry, for example).
The average quantity of deposited zinc is reduced by
3% when closed-loop digital control is used, while still
guaranteeing the specifications for minimum zinc deposi-
tion. Taking into account the line production and the
price of the zinc, this corresponds to an annual saving
over 350 000 USD. The closed-loop operation also re-
duces the task of the operator, thereby creating better
working conditions.
13. Conclusions
A methodology for the design and tuning of R-S-T
digital controllers has been presented. These controllers
have already been used in a significant number of indus-
trial applications (Rolland and Landau, 1991). The eco-
nomic impact of the improved performance has largely
justified their use. Software packages are available for the
design, implementation and commissioning of the R-S-T
digital controllers.
References
Astro¨ m, K.J., Wittenmark, B., 1990. Computer Controlled Systems—
Theory and Design. Prentice Hall, N.J. (2nd Edition).
Doyle, J.C., Francis, B.A., Tannenbaum, A.R., 1992. Feedback Control
Theory. MacMillan, New York.
Fenot, C., Rolland, F., Vigneron, G., Landau, I.D., 1993. Open loop
adaptive digital control in hot-dip galvanizing. Control Engineering
Practice 1(5), 779—790.
Harvey, C.F., Carlton, A.J., 1974. Mathematical Modelling of Air Jet
Coating Mass, J. Lysagh Ltd. Res. Tech. Rep. Australia.
Jacobs, O.L.R., Chen, D.S., 1991. Regulation of Deposited Zinc in hot-
dip Galvanizing. Proceedings of the first European Control Confer-
ence, Grenoble, 1, 773—778.
Landau, I.D., 1990. System Identification and Control Design. Prentice
Hall, Englewood Cliffs, N.J.
Landau, I.D., 1993. Identification et Commande des Syste` mes (seconde
e´ dition), Herme` s, Paris.
Landau, I.D., Karimi, A., 1997. Recursive algorithms for identification
in closed loop — A unified approach and evaluation. Automatica
33(8).
Landau, I.D., Langer, J., Rey, D., Barnier, J., 1996. Robust control of
a 360° flexible arm using the combined pole/placement sensitivity
function shaping method. IEEE T-CST, 4(4), 369—383.
Landau, I.D., Lozano, R., M’Saad, M., 1997. Adaptive Control.
Springer Verlag, London.
Langer, J., Landau, I.D., 1996. Improvement of robust digital control
by identification in closed loop, application to a 360° flexible arm.
Control Engineering Practice 4(8), 1079—1088.
Ljung, L., 1987. System Identification. Theory for the user. Prentice
Hall, Englewood Cliffs, N.J.
164 I.D. Landau/ Control Engineering Practice 6 (1998) 155—165
Morari, M., Zafiriou, 1989. Robust process control. Prentice Hall, N.J.
Rolland, F., Landau, I.D., 1991. Pour mieux re´ guler, le PC va
vous aider. Mesures, n°640, pp. 71—73, De´ c.
Van den Hof, P., Schrama, R., 1995. Identification and control —
closed-loop issues. Automatica 31(12).
Zang, Z., Bitmead, R.R., Gevers, M., 1995. Iterative weighted least-
squares identification and weighted lqg control design. Automatica,
31(11).
I.D. Landau/ Control Engineering Practice 6 (1998) 155—165 165

D. The set D. Histograms for good and poor control.H.C. 3. Computer control system.156 I. The Q discrete-time model of the plant to be controlled can alternatively be represented by its pulse transfer operator H(q\): Fig. Note that the sampling frequency is selected in accordance with the bandwidth of the continuous-time plant.: Analog-to-Digital Converter. which should be identified. 3 illustrates an appropriate setting for a computer control system. This paper will present such a methodology for the design and application of R-S-T digital controllers.: Digital-to-Analog Converter. (2) To have a dynamic model of the plant to be controlled (this can be obtained from real data by identification).D. the output being the sequence +y(t). 3. Z. t is the normalized discrete time (0. is interpreted as a discretized system.  L Fig.2) and corresponds to the discrete time divided by the sampling periods ¹ .O. q\B B(q\) q\B\ B*(q\) G(q\)" " A(q\) A(q\) where q\ is the backward shift operator (q\ y(t)" y(t!1)) and A(q\)"1#a q\#2a q\L  L B(q\)"b q\#2b q\L "q\B*(q\). A.C. generated by the computer. . (4) To have a procedure for controller validation and onsite re-tuning.C.# plant # A. Identification of discrete-time models for industrial processes Fig.C. The discretized plant is characterized by a discrete-time model. 1. 2. (3) To possess a suitable controller design methodology.D.: Zero-Order Hold. (5) To have appropriate software packages with real-time capabilities for data acquisition. 1. Fig. Landau / Control Engineering Practice 6 (1998) 155—165 2. resulting from the A/D conversion of the system output y( ). whose control input is the sequence +u(t). H(q\) is defined by: controller one needs: (1) To specify the desired control-loop performances. and more specifically in accordance with the desired bandwidth of the closed loop. 2. y(t)"H(q\) u(t). France).D.A. system identification. compatible with the desired performance and the corresponding plant model. Principle of controller design. The basic rule is: f "(6 to 25) f !J Q where f is the sampling frequency and f !J is the desired Q bandwidth of the closed loop. control design and on-site commissioning. The methodology will be illustrated by its application to the control of the deposited zinc in hot-dip galvanizing at SOLLAC (Florange. The discrete-time model of the plant to be controlled is described in the time-domain by: L L y(t)"! a y(t!i)# b u(t!d!i) G G G G where d is the integer number of sampling periods contained in the time-delay of the plant.A.

at each sampling instant. The validation test enables L the best algorithm to be obtained for the estimation of the parameters. an objective validation can be made by carrying out statistical tests on the prediction error (t) and the predicted output y (t). which is either generated by a tracking reference model (Bm/Am) or stored in the computer memory. it requires an input signal of much lower magnitude than those used for step or frequency response. the digital filters R and S are designed in order to achieve the desired regulation performance. For a more detailed discussion see (Landau. i. generated by the computer (sequence of rectangular pulses with randomly variable duration). G G Fig. Landau / Control Engineering Practice 6 (1998) 155—165 157 Therefore.. (known as the prediction error) is used L by a parameter-adaptation algorithm which. 3. (4) Validation of the identified model (structure and values of the parameters). A discrete-time model with adjustable parameters is implemented on the computer. In addition. y (t). ¹(q\) have the form: R(q\)"r #r q\2#r q\L0   L0 S(q\)"s #s q\2#s q\L1 (often s "1)   L1  ¹(q\)"t #t q\2 t q\L2. Fig. The input is in general a very low-level pseudo-random binary sequence. Parameter estimation of discrete-time models. This approach provides much more accurate models than the methods based on step response or frequency response.D.   L2 The corresponding time-domain expression of the control law is given by (s "1):  L1 L0 u(t)"! s u(t!i)! r y(t!i) G G G G L2 # t y*(t#d#1!i). leading to particular choices of R. and the digital filter ¹ is designed afterwards in order to achieve the desired tracking performance. such that the estimated parameters are always unbiased. S(q\). 4. Digital PID can also be represented in this form. The error between the system output at instant t. The identification methodology includes four steps: (1) Input-output data acquisition around an operating point. (2) Estimation (choice) of the model complexity (structure). The case of a controller operating on the regulation error (which does not allow the independent specification of tracking and regulation performance) corresponds to ¹"R. will modify the model parameters in order to minimize this error. This structure has two degrees of freedom. The equation of the R-S-T canonical controller is given by: S(q\) u(t)#R(q\) y(t)"¹(q\) y*(t#d#1) where u(t) and y(t) are the input and output of the plant and y*(t#d#1) is the desired tracking trajectory.I. The polynomials R(q\). replacing q by z (the complex variable) in the expression of the pulse transfer operator gives the pulse transfer function. Once the model has been obtained. . For models with constant parameters. and the output predicted by the model. 5. (3) Estimation of the model parameters. The R-S-T digital controller The canonical structure of the R-S-T digital controller is represented in Fig. 1990. 1987). no unique parameter-estimation method exists which may be used successfully for all the types of disturbances. S and T. 4.e. y(t). 5. The R-S-T canonical structure of a digital controller. a good identification of a plant model generally requires the use of an interactive system featuring various parameter-estimation methods and the corresponding validation techniques. Unfortunately. This structure allows achievement of different levels of performance in tracking and regulation. Ljung. The principle of the identification of discrete-time models is illustrated in Fig. using as input in general a centered pseudorandom binary sequence (PRBS) of small magnitude. One of the important facts to be emphasized is that the plant measurements are generally noisy.

1 Many control strategies can be applied to the design of the R-S-T controller by an appropriate reformulation. The only assumption is that the polynomials A(q\) and B(q\) characterizing the plant model do not have common factors. N°2. These polynomials are defined by the performance specifications (e. the introduction of an integrator in the controller requires one to take H (q\)"1!q\. The ¹(q\) polynomial is chosen as: ¹(q\)"P(q\)/B(1) (B(1)"B(q\) for q"1). contain some prespecified fixed parts. Once the closed-loop poles have been defined. and most of the various designs can be related to it. Let the degrees of polynomials A(q\) and B(q\) be defined by: n "deg A(q\). For this reason it is convenient to factorize the polynomials R(q\) and S(q\) as follows: R(q\)"H (q\) R(q\) 0 S(q\)"H (q\) S(q\) 1 where H (q\) and H (q\) are prespecified polyno0 1 mials.1. for various reasons. the pole placement can be considered as the basic design technique. the desired closed-loop poles are specified in the form: P(q\)"P (q\) P (q\) " $ where P (q\) specifies the desired dominant poles of the " closed loop. and the tracking behaviour is described by the equation: y(t)"q\B\ B*(q\) y*(t#d#1). solving the equation: P(q\)"A(q\) S(q\)#q\ B(q\) R(q\) allows the determination of S(q\) and R(q\). However. as well as the special issue of European Journal of Control Vol.. with or without time delay. 1997). 1990. 1990. n "deg B(q\). B(1) In other words one follows the desired trajectory. See for example (Landau. P(q\) and the behaviour with respect to an output disturbance is given by the output sensivity function: A(q\) S(q\) S (q\)" WN A(q\) S(q\)#q\B B(q\) R(q\) A(q\) S(q\) " P(q\) where P(q\) defines the desired closed-loop poles (regulation behaviour). Astrom and Wittenmark.  n "deg S(q\)"n #d!1 1 n "deg R(q\)"n !1 0  in which: S(q\)"1#s q\#2 s q\L1"1#q\ S*(q\)  L1 R(q\)"r #r q\#2 r qL0. and featuring stable or unstable zeros.D. 1993).g. For example.   L0 However. Landau et al.  Then the above equation has a unique solution (assuming that A(q\) and B(q\) do not have common factors) for: n "deg P(q\))n #n #d!1 . ¨ 1990. 1995. Pole placement This control strategy can be used for plant models of any order. The controller polynomials S(q\)"H (q\) S(q\) 1 and R(q\)"H (q\) R(q\) are obtained by solving 0 the equation: P(q\)"A(q\) H (q\) S(q\) 1 #q\B\ B*(q\) H (q\) R(q\) 0 where P(q\) defines the desired closed-loop poles. and P (q\) specifies the auxiliary poles of $ the closed loop. Note that unstable discrete-time zeros occur when a fractional delay larger than half of the sampling period is present. SN P(q\) In general. which will ensure the desired closed-loop poles. 1. The input sensitivity function which reflects the effects of an output disturbance upon the plant input is given by: A(q\) R(q\) S (q\)"! . the integrator in the controller) and by considerations of robustness. For more details see (Landau.158 I. in general the polynomials R(q\) and S(q\) of the controller may. or when a high sampling frequency is used for . Landau / Control Engineering Practice 6 (1998) 155—165 The closed-loop control transfer operator (between r(t) and y(t)) is given by: q\B ¹(q\) B(q\) H (q\)" !* A(q\) S(q\)#q\B B(q\) R(q\) q\B ¹(q\) B(q\) " .. dedicated to a robust control benchmark. 3. filtered through the plant zeros.

while the converse is not true (systems with good phase and gain margins can pass very close to the critical point). P(q\)"P (q\) ) B*(q\) ) P (q\).. -*  ""S\(e\HU) " . the phase margin ( ). -* The modulus margin M is defined as the radius of a circle. it is therefore sufficient simply to plot the frequency characteristics of the modulus (gain) of the output sensitivity function in dB.5 implies G*2 and '29° (the converse is not necessarily true). Note that M*0..e. the delay margin of G the system is defined as: G. " $ ‘‘Internal model control’’ (Morari. "min G G  Typical values of these robustness indicators are: — modulus margin: M*0.D. and can be viewed as an approximation of the pole placement in the sense of a certain quadratic criterion (Landau et al. d"0. Fig. These indicators are the gain margin ( G). ¨ 1990). 4. The modulus margin and the delay margin are the most interesting in applications. gain and phase margins. centered on [!1. j0] and the Nyquist plot of the open-loop transfer function: H (z\)"z\ B(z\) R(z\)/A(z\) S(z\).I.e. The delay margin is the maximal additional delay that will be tolerated in the open-loop system without causing an instability of the closed-loop system. P(q\)"A(q\) ) P (q\). 1989) corresponds to a pole-placement strategy where the desired closed-loop poles contain the poles of the plant model. i. which can be viewed as a particular case of the pole-placement strategy. In other words. j0] and tangent to the Nyquist plot of H (e\HU) (see Fig. n )2. gain and phase margins. In this case. characterized by the  corresponding phase margins .5 (!6 dB) — delay margin: *¹ (sampling period) 1 — gain margin: G*2 (6 dB) — phase margin: 30°) )60°.2. one can use the strategy called ‘‘tracking and regulation with independent objectives’’. 6 illustrates the modulus. 1997). Landau. 6).  Minimization of the H norm of S (e\HU) will maximize  WN the modulus margin. If the Nyquist plot of the open-loop system intersects the unit circle at several cross-over frequencies G . Robustness Four indicators are generally used to express the robustness of a design in terms of the minimal distance with respect to the critical point [!1. 1992). "!"S (e\HU) " WN  Note that a given value of the modulus margin will guarantee certain phase and gain margins. Bounds on the magnitude of the frequency-dependent model uncertainties convert to upper constraints upon the moduli of the various ..  Predictive control strategies and an LQ control strategy using an appropriate formulation of the criterion to be minimized lead to an R-S-T controller.. The modulus margin is the minimal distance between the critical point [!1. j0] in the Nyquist plane. Relationship with other control strategies For plant models with stable zeros. Sensitivity functions are related to the robust stability of the closed loop with respect to plant model uncertainties. -* The result is that: M""1#H (e\HU) " . To obtain the modulus margin. Landau / Control Engineering Practice 6 (1998) 155—165 159 continuous-time models having a difference of degree between denominator and numerator greater than or equal to 2 (Astrom and Wittenmark. where the desired closedloop poles contains the zeros of the plant. the orders of the plant model are limited to n )2. the modulus margin corresponds to the inverse of the H norm of the output sensitivity function. In this case: M dB"("S (e\HU) " )\ dB WN  dB. 1990. WN  Fig. the modulus margin ( M) and the delay margin ( ). WN  "("S (e\HU) " )\. 3. Modulus. 6. i. but in accordance with the desired closed-loop performance. Both phenomena lead to the conclusion that the sampling frequency has to be chosen as low as possible. $ Digital PID can be designed using pole placement. see (Doyle et al.

when the plant contains an integrator. (b) input sensitivity function. Landau et al. 1997. it may not be possible to operate the plant in an open loop in order to achieve system identification.160 I. The upper part represents the true closed-loop system.D. Such situations are encountered. one can also identify a new plant model. 1997). Templates for the sensitivity functions: (a) output sensitivity function. and the lower part corresponds to an adjustable predictor for the closed loop. with the same data.. Fig. 8 (Landau and Karimi. A basic scheme for plant model identification in a closed loop is shown in Fig. for example. 1996). 8. Landau and Karimi. sensitivity functions. 1995). The new 5. . Significant progress in this area has been made in the last few years. 1995. However. Typical templates for S and S are shown in Fig. 7. WN SN The upper and lower bounds on the output sensitivity function S in the high-frequency region come from the WN translation of the delay margin constraints in the frequency domain (Landau. a controller already exists. 6. techniques for plant identification in closed-loop operation should be used. For a detailed presentation see (Van den Hof and Schrama. constraints imposed on the sensitivity functions can be translated into tolerated model uncertainties. Identification of the closed loop will allow the performance achieved to be compared with the designed performance by comparing the achieved and the designed closed-loop poles. In other words. A design methodology that combines pole placement with the shaping of the sensitivity function has been developed in order to ensure both performance and robustness. It will also allow the robustness of the control schemes to be assessed by comparing the designed and achieved sensitivity functions in the frequency domain. Controller validation and on-line retuning As indicated in Section 1. 1997). Conversely. 1993. the model obtained in a closed loop will allow for better prediction of the behaviour of the closed-loop system. or when important drifts of the operating point may occur during input/output data acquisition. can be translated into desired templates for the sensitivity functions. These constraints. See (Landau. New algorithms dedicated to identification in closed loops have been developed. Identification in a closed loop In a number of practical situations. The error between the system output and the closed-loop predictor output (called a closed-loop output error) is used by a parameter-adaptation algorithm that will drive the parameters of the estimated plant model in order to minimize the error between the two closed-loop systems. but for various reasons cannot be disconnected. acquired in a closed loop. 7. Landau / Control Engineering Practice 6 (1998) 155—165 Fig. validation of the designed controller is a key issue in assessing the effective performance of the closed-loop system. Therefore. In a number of other situations. Landau et al. re-parameterized in terms of a known fixed controller and an adjustable plant model. Plant model identification in closed-loop operation. as well as those on the modulus margin and delay margin.. The input sensitivity function reflects both tolerance with respect to additive uncertainties. and the activity of the input in the presence of disturbances (high values in certain frequency regions indicate low model uncertainty tolerance and important stress on the actuator).

The principle of hot-dip galvanizing is illustrated in Fig.I. The measurement of the deposited zinc can be made only on the cooled. the corresponding values of the controller parameters will be used. Nonlinear static models have been developed for computing the appropriate pressure. In this case. which remove the excess zinc. Tight control (i. Landau / Control Engineering Practice 6 (1998) 155—165 161 model. The most important part of the process is the hot-dip galvanizing. in a number of applications. 1995. Fig. a small variance in the controlled variable) will allow a more uniform coating and a reduction of the average quantity of deposited zinc per unit area. ‘‘Open-loop’’ adaptation. . finished strip. and then rises vertically out of the bath through the stripping ‘‘air knives’’. 10). the distance between the air knives and the strip. The objective of the control is to assure good uniformity of the deposited zinc. one can use an ‘‘open-loop’’ adaptive control (Fig. tight control of the deposited zinc per unit area has an important commercial impact since the average consumption for a modern galvanizing line is of the order of 40 tons per day (price+1 500 USD/ton).. The corresponding controllers are stored in a table. 1997). based on an identified model. 10). in addition to an improvement in quality. is then used for re-tuning of the controller. For each interval. one has to consider the ‘‘adaptation’’ of the controller. The effect of the air knives depends on the air pressure. The term ‘‘adaptation’’ (adaptive control) refers to a set of techniques for the automatic tuning of the controller in real time.. ‘‘Closed loop’’ adaptive control. the characteristics of the dynamic model of the plant depend upon a set of measured variables. 7. and the strip speed may vary from 30 to 180 m/mn. This controller assures the desired performance for all the Fig. 1996.D. operating points located in the interval. 11. which define an operating point.e. Preheated steel strip is passed through a bath of liquid zinc. However. The remaining zinc on the strip surface solidifies before it reaches the rollers. identified in the closed loop. Langer and Landau. a relevant operating point is selected and a corresponding controller is designed. Hot-dip galvanizing at SOLLAC (Florange) The objective of the galvanizing line is to obtain galvanized steel with formability. according to the table. One can distinguish between two basic adaptive control techniques: (1) ‘‘closed-loop’’ adaptive control (Fig. based on a current estimation of the plant model and the desired performance. Examples can be found in (Zhang et al. 8. 9) (2) ‘‘open-loop’’ adaptive control (Fig. Landau et al. As a consequence. Adaptation When ‘‘system identification’’ plus ‘‘robust control design’’ does not allow one to obtain a single linear controller. distance and speed for a given desired value of the deposited zinc. and the speed of the strip. This procedure is used for two purposes: — improvement of a previous design — controller maintenance. The variety of products is very large in terms of deposited zinc thickness and steel strip thickness. whilst guaranteeing a minimum value of the deposited zinc per unit area.. 10. 9. which guide the finished product. When the plant is operating at a certain point. giving acceptable performance for the whole range of operating points because of the too-wide variations in the dynamic characteristics of the plant. The ‘‘closed-loop’’ adaptive control system usually combines a real-time identification algorithm with the computation of the controller in real time. The range of operating points is divided into a number of operating intervals. in order to maintain the desired performance when the plant parameters vary. surface quality and weldability equivalent to uncoated cold rolled steel. The deposited zinc may vary between 50 to 350 g/m (each side).

When discretizing this model. '0. by using the pressure as the control variable. The identification has been done with the coating process operating in an open loop. 1#a q\  The fractional delay (which corresponds to the presence of an additional term b q\) is negligible because of the  way in which the sampling period ¹ is selected. relating variations in the pressure to variations in the deposited mass. one can compensate for the disturbances created by variations in distance and speed as well as by the term . The delay of the process will depend linearly on the speed. Identification of the discrete-time plant model The process comprises the air-pressure control loop and the coating process. The sampling frequency has been chosen at each operating point.R. The corresponding linearized discrete-time model will be of the form: q\B (b q\)  H(q\)" . Therefore. the major difficulty comes from the variable time-delay.B.R. However. 10. the model originally proposed by Harvey and Carton (1974) and completed by Jacobs (1991) can be used: m"KD  » # K P where m is the deposited mass per unit area. Therefore. and d is the discrete-time delay (an integer). identification and control is connected to the process through an industrial network. The control input to the process is the reference of the air pressure control loop. before a highfrequency sampling is undertaken (a multiple of the desired sampling frequency). the sampling period is tied to the strip speed using the formula: (¸/»)# ¹" . and the output of the process is the measured deposited mass per unit area (see Fig. » . In order to obtain a controller with a fixed number of parameters. in order to have the discrete-time delay d"7. which can be approximated by a firstorder system. First. speed ( ») and distance ( D). of the form: G e\Q H(s)" . D is the distance between the air knives and the strip. was generated by  a shift register with N"5 cells and a clock frequency equal to half of the sampling frequency (length of the  D# »! P# K . The data acquisition is illustrated in Fig. 11. A linearized model around an operating point (P . 12).   D ) can be obtained using a standard Taylor series ex pansion for variations of pressure ( P). K The pressure in the air knives is regulated through a pressure loop.D. Process description. the parameters of the model.B. accounts for unpredictable effects and/or K modelling errors. P is the air pressure and » is the strip speed. (Pseudo Random Linary Sequence) of a magnitude of $4% with respect to the static pressure (P ). an analog anti-aliasing filter is used. (d"integer) Q d where d is a small additional time-delay due to the implementation. given above. 1#s¹ ¸ " » can be considered. Model of the process For an analysis of the process. The PC used for data acquisition. A digital anti-aliasing filter is inserted between the two samplers. K is a constant of proportionality. the delay of the discrete-time model should remain constant. It can be seen that. the control variable is the air pressure. 9. Fig. and » is the strip speed. It has the form: m"KD  » # P  . Landau / Control Engineering Practice 6 (1998) 155—165 dynamic model. this was Q confirmed by the model identification procedure. will depend on the distance D and on the speed ». 13. .S. The input used was a P. a continous-time linear .S.162 I. The P. At SOLLAC Ste Agathe. where ¸ is the distance between the air knives and the transducers.

12. One of the causes is the imperfect measurement of the strip/air-knives distance. 1993). 14 shows typical results. This variability will require a robust control design. The speed range and the distance range have been split into 3 regions.D. As  both sides of the steel strip have to be galvanised. ‘‘Output error’’ consistently provided the best results for this application. 100 to 160 (average: 128) measurements have been used for the various identifications made in the different regions of operation. a variability of the parameters is observed. The strip speed directly affects the sampling period according to the relationship: ¸# ¹" Q » where is the equivalent time-delay of the industrial network and of the programmable controller used for pressure regulation. despite the variability of the model. For a detailed presentation of this application see (Fenot et al. both ‘‘front’’ and ‘‘back’’ models have been identified. and yielded the largest pulse width (10 ¹ ) comparable with the rise time of the process. an ‘‘open-loop adaptation’’ technique has been considered.I. However. a cross-correlation between the predicted output (using an output error predictor) and the output error has been used (Landau. Anti-wind-up procedures have been used for the implementation of the controller. Q These robustness margins assure satisfactory performance in a region of operation. Data acquisition.B. For each of these operating regions an identification has been performed. even within a region of operation at a constant distance and with relatively small speed variations.R. This necessitated splitting the operation of the plant into several regions. 12. 1990). the following recursive identification methods have been used (Landau. since the model does not have finite zeros) has been used. Controller design and adaptation The ‘‘tracking and regulation with independent objectives’’ (which in this case is equivalent to the poles placement. and because the positions and physical realisations of the two actuators are not symmetrical. giving a total of 9 operating regions. (For the first four methods.S. It was observed that a significant variability in the parameters occurs with a change of the operating points. Process block diagram. the whiteness test on the prediction error has also been used for validation and comparison). Results Fig. Fig. The choice made for the P. — distance between the air knives and the steel strip. 11. Landau / Control Engineering Practice 6 (1998) 155—165 163 Fig. 13. 1990): — Recursive least squares — Extended least squares — Recursive maximum likelihood — Instrumental variable with auxiliary model — Output error — Generalized least squares For validation of the identified models and comparison of the models obtained with the different methods. The open-loop adaptation is made with respect to: — steel strip speed. As no unique identification method gives unbiased results for all types of disturbances. to be obtained with a delay margin greater than 2¹ . In order to ensure satisfactory performance for all regions of operation.. and controllers based on the identified model have been computed and stored in a table. obtained when one of the sides is under digital regulation and the other side is . Robust control design using an identified model and based on shaping of the sensitivity function allowed a modulus margin greater than !6 dB. allowed at least one full sequence to be sent for each experiment. sequence: 64). and a smooth transfer from open-loop to closed-loop operation has also been assured.

I.A. Res. Wittenmark. . application to a 360° flexible arm. Identification et Commande des Systemes (seconde ` edition). 369—383. 1996. Robust control of a 360° flexible arm using the combined pole/placement sensitivity function shaping method. 14.. I. A.D. Rolland. MacMillan. Langer. Doyle. B. The economic impact of the improved performance has largely justified their use. Computer Controlled Systems — ¨ Theory and Design.. Hermes. Feedback Control Theory. Mathematical Modelling of Air Jet Coating Mass. A... J. Proceedings of the first European Control Conference.J. G. Open loop adaptive digital control in hot-dip galvanizing. Regulation of Deposited Zinc in hotdip Galvanizing. I. London. 1997. System Identification and Control Design..D.. This provides a better-quality finished product (extremely important in the automotive industry. Springer Verlag. 4(4)... Australia. Prentice Hall. Englewood Cliffs.. Landau / Control Engineering Practice 6 (1998) 155—165 Fig. B. J. Rep. Fenot. 1991). L.R.F. 1974. M. A. Prentice Hall. Jacobs. Typical performance of the digital regulation of the deposit zinc. C.. Landau. this corresponds to an annual saving over 350 000 USD.. These controllers have already been used in a significant number of industrial applications (Rolland and Landau.D.. J. Barnier. 1993. for example). Francis.D. K. Tannenbaum. New York. Landau. Vigneron. 1997. Landau. I.J. Karimi. O. Control Engineering Practice 4(8).D.J.. 1992.. Adaptive Control.D.J. N.C. 1.. 1990. F. Prentice Hall.. ´ ` Landau. Control Engineering Practice 1(5). 1993. 13. Lozano. 1079—1088. 1990. Automatica 33(8).. under computer-aided manual control (the operator has on display a moving short-time history of the deposited zinc and applied pressure).. Software packages are available for the design. while still guaranteeing the specifications for minimum zinc deposition. Carlton.. IEEE T-CST. The closed-loop operation also reduces the task of the operator.D. Improvement of robust digital control by identification in closed loop.L. C.D. J. Chen. Recursive algorithms for identification in closed loop — A unified approach and evaluation. Grenoble. D. M’Saad. I. thereby creating better working conditions. References Astrom. I. Landau.. J. 1996. Englewood Cliffs. N.. System Identification. Landau. implementation and commissioning of the R-S-T digital controllers. The average quantity of deposited zinc is reduced by 3% when closed-loop digital control is used. 1991.164 I.J... Theory for the user. (2nd Edition). Lysagh Ltd. Tech. 1987. D.R. Rey. 773—778. I.. Ljung.. A reduction in the dispersion of the coating is noticed when closed-loop digital control is used. Conclusions A methodology for the design and tuning of R-S-T digital controllers has been presented. N. Landau. Langer..S. Paris. 779—790.. Harvey. R. Taking into account the line production and the price of the zinc.

le PC va ´ vous aider.. 1995. Automatica 31(12). M. 1989. Prentice Hall. Iterative weighted leastsquares identification and weighted lqg control design. Landau. F.I. R. Robust process control. 1991..R. Bitmead. Zafiriou. n°640. Landau / Control Engineering Practice 6 (1998) 155—165 Morari. Rolland. Dec.. Schrama. N. Pour mieux reguler. R.J. Identification and control — closed-loop issues.. 31(11)..D. pp. 165 Zang.D.. M. Automatica. 71—73. ´ Van den Hof. Z. 1995. P. Gevers. .. Mesures.. I.

You're Reading a Free Preview

Download
scribd
/*********** DO NOT ALTER ANYTHING BELOW THIS LINE ! ************/ var s_code=s.t();if(s_code)document.write(s_code)//-->