Sie sind auf Seite 1von 10

Automatica 40 (2004) 19871996

www.elsevier.com/locate/automatica
Brief paper
Linear LMI-based external anti-windup augmentation for stable
linear systems

Gene Grimm
a
, Andrew R. Teel
b
, Luca Zaccarian
c,
a
Raytheon Company, Space and Airborne Systems, 2000 E. El Segundo Blvd, El Segundo, CA 90245, USA
b
Department of Electrical and Computer Engineering, University of California, Santa Barbara, CA 93106, USA
c
Dipartimento di Informatica, Sistemi e Produzione, University of Rome, Tor Vergata, 00133 Rome, Italy
Received 21 March 2003; received in revised form 16 June 2004; accepted 27 June 2004
Abstract
We study linear anti-windup augmentation for linear control systems with saturated linear plants in the special case when the anti-windup
compensator can only modify the input and the output of the windup-prone linear controller. We also measure the arising performance in
terms of the nite L
2
gain from exogenous inputs to selected performance outputs. Our main results are a system theoretic feasibility
characterization for xed order anti-windup design and a linear matrix inequality (LMI) formulation for optimal static and plant-order anti-
windup design. Interpretations of lower bounds on the achievable performance are also given. The effectiveness of the design procedure
is demonstrated on a simulation example.
2004 Elsevier Ltd. All rights reserved.
Keywords: Windup; Saturation control; Linear matrix inequalities
1. Introduction
Anti-windup designs were characterized already in the
1950s as augmentations to a prespecied linear controller
(which induces a highly desirable closed-loop behavior
when interconnected to a linear plant without saturation)
with the goal of guaranteeing two properties from the aris-
ing augmented nonlinear closed-loop system: (1) as long as
the actuators do not saturate, the response coincides with the
linear, unconstrained response; (2) if the actuators saturate,
stability is preserved and performance is recovered as much
as possible (within the limits imposed by the saturation

This paper was not presented at any IFAC meeting. This paper was
recommended for publication in revised form by Associate Editor Alessan-
dro Astol under the direction of Editor Hassan Khalil. This work was
supported in part by AFOSR under Grants F49620-00-1-0106 and F49620-
03-1-0203, NSF under Grants ECS-9988813 and ECS-0324679, ASI and
MIUR through PRIN project MATRICS and FIRB project TIGER.

Corresponding author. Tel.: +39-06-7259-7429; fax: +39-06-7259-


7460.
E-mail addresses: grimm@raytheon.com (G. Grimm),
teel@ece.ucsb.edu (A.R. Teel), zack@disp.uniroma2.it (L. Zaccarian).
0005-1098/$ - see front matter 2004 Elsevier Ltd. All rights reserved.
doi:10.1016/j.automatica.2004.07.001
constraint). During the 1970s and the 1980s, many solutions
to particular instances of the windup problem were proposed
with an industrial application oriented viewpoint. Some ap-
proaches, however, were already leaning towards a more for-
malized understanding of the windup problem and toward
an increased generality of the proposed solutions, with their
consequent applicability to a large class of windup-prone
control systems (see, e.g., Hanus (1988), Kothare, Campo,
Morari, & Nett (1994) for surveys of these approaches).
Only in the last decade, the problem has been addressed and
solved in a rigorous manner, thus leading to constructive so-
lutions with formal stability guarantees for the arising non-
linear control system and sometimes characterization and
optimization of performance, via different measures.
These recently introduced modern techniques range from
linear dynamic generalizations of the linear static approach
widely characterized in the excellent survey Kothare et al.
(1994), up to fully nonlinear compensation schemes, which
typically lead to improved closed-loop performance at the
cost of more complicated constructions. In particular, lin-
ear dynamic anti-windup techniques denote cases where
1988 G. Grimm et al. / Automatica 40 (2004) 19871996
z
y
x p
x c
x aw
q
-
w
u
[()
v2
v1
yc
uc
P
C
AW
Fig. 1. The external anti-windup augmentation scheme.
a linear lter is driven by the mismatch between controller
output and plant input (see Fig. 1), whereas nonlinear anti-
windup denotes nonlinear generalizations of this scheme.
When restricting the attention to linear anti-windup schemes,
it is useful to characterize the degrees of freedom available
to the anti-windup compensator for injecting modications
in the predesigned controller dynamic equations. In partic-
ular, in the so-called full-authority case, the anti-windup
compensator can inject signals at each state equation and
at the output equation of the predesigned controller. On the
other hand, the so-called external case corresponds to the
stricter situation where only the input and output signals
of the predesigned controller are accessible for anti-windup
modications and the full authority over all the controller
states is not anymore allowed to the anti-windup signals.
1
Full-authority linear anti-windup characterizes implicitly
the architecture of the very rst anti-windup schemes (such
as, e.g., the back-calculation of Fertik & Ross (1967)) where
in PI (or PID) controllers, the integrator input (namely, the
controller state equation) was modied by a signal propor-
tional to the plant input mismatch. Most of such original
schemes were summarized in Kothare et al. (1994) where a
framework based on factorizations of the predesigned con-
troller was used for unifying purposes. Static full-authority
anti-windup was successfully employed in Mulder, Kothare,
and Morari (2001), where convex synthesis tools based
on linear matrix inequalities (LMIs) were used to pro-
vide globally stabilizing static anti-windup gains inducing
global nite-gain input/output L
2
performance properties.
Static full-authority designs were also used in Cao, Lin,
and Ward (2002a,b) to guarantee large operating regions
and (local) nite gain L
2
performance in the exponentially
unstable case. The full-authority approach for global nite-
gain L
2
performance was fully characterized in Grimm,
Hateld, Postlethwaite, Teel, Turner, and Zaccarian (2003),
where dynamic compensators are synthesized via LMIs. In
Grimm et al. (2003) it is shown that asymptotic stability of
the plant is a necessary condition for such L
2
performance
properties to be achievable. Moreover, the limits of the pre-
vious static designs of Mulder et al. (2001) are claried by
providing system theoretic characterizations of their feasi-
1
This can happen, e.g., in cases where part of the controller is
implemented in an analog way.
bility. Dynamic full-authority anti-windup for exponentially
unstable systems was also recently addressed in Wu and Lu
(2004) based on narrowed sector techniques.
In this paper, we address the problem of linear external
anti-windup augmentation for nite L
2
input/output gain.
As compared to previous linear external techniques, such
as those proposed
2
in Zheng et al. (1994), Park and Choi
(1995), Teel and Kapoor (1997), Zaccarian and Teel (2002),
this approach is more appealing because: (1) it provides tools
for static or reduced order anti-windup designs, whereas
those approaches are only plant-order ones; (2) it allows to
determine optimality-based solutions by convex optimiza-
tion, thereby resulting in very desirable closed-loop behav-
ior, whereas most of the above-mentioned ones dont ad-
dress the performance of the compensated closed-loop. Note
that solutions to the convex optimization problems charac-
terized here can be readily computed by using commercially
available software, such as Gahinet, Nemirovski, Laub, and
Chilali (1995), thus making the algorithms herein proposed
very simple to apply even for the non expert. As compared
to previous nonlinear external anti-windup techniques, the
advantage of this approach stands in its simplicity of design
and implementation. For example, the framework of Teel
and Kapoor (1997) has been employed for optimality-based
(therefore, high performance) nonlinear external compen-
sation schemes in Bemporad, Teel, and Zaccarian (2004),
where the overall scheme is sampled-data and in Teel (1999)
and Zaccarian and Teel (2004) where the compensation law
is nonlinear. Although these last nonlinear approaches are
typically associated to very desirable closed-loop behavior,
their structure may sometimes be too complicated for imple-
mentation. Indeed, neither of the above mentioned solutions
addresses static external anti-windup design,
3
whereas that
problem is among those addressed and solved here. It should
be pointed out, however, that simple anti-windup schemes
for MIMO systems have been already proposed without sta-
bility and performance guarantees (see, e.g., Peng, Vran ci c,
Hanus, & Weller (1998)) and are applicable when formal
closed-loop guarantees are not a concern.
As compared to the full-authority approaches of Kothare
et al. (1994); Mulder et al. (2001); Grimm et al. (2003), this
paper can be seen as the external counterpart of the results
of Grimm et al. (2003) where LMI-based constructions were
given for static and dynamic full-authority anti-windup. Also
for the external case we are able to provide LMI-based tools
for static and dynamic design, thus making it possible to
use convex optimization to solve the linear external anti-
windup problem too. In this sense, the full-authority anti-
windup results of Grimm et al. (2003) can now be seen as
2
Note that to show that the approaches in Zheng, Kothare, and Morari
(1994) and Park and Choi (1995) are external, one needs to suitably
represent the compensation schemes by loop transformations.
3
Static external anti-windup design is the simplest anti-windup
scheme that one can implement because it corresponds to selecting the
anti-windup lter as a static gain.
G. Grimm et al. / Automatica 40 (2004) 19871996 1989
a special case of the external compensation scheme where
the predesigned controller input matrix is the identity.
The paper is organized as follows: in Section 2 we dene
the problem. In Sections 3 and 4, respectively, we character-
ize external anti-windup feasibility and provide procedures
for optimal external anti-windup synthesis. In Section 5,
we illustrate the proposed design on an example where the
external construction leads to the same performance level
achievable by the full-authority scheme. Finally, in Section
6 we give interpretations of performance bounds and in Sec-
tion 7 we report the proofs of the formal statements of the
previous sections.
Notation. Given a matrix B, B

denotes a matrix of full


column rank that spans the null space of B (namely B
T
B

=
0). Given a square matrix X, He(X) := X +X
T
.
Denition 1. Given any symmetric positive denite matrix
W R
n
u
n
u
and two matrices W
1
, W
2
R
n
u
r
, dene
the W-product of W
1
and W
2
as W
1
, W
2

W
:= W
T
1
WW
2
.
A function f : R
n
u
R
n
u
is said to belong to the sector
[0, I]
W
if f (w), wf (w)
W
0 for all w R
n
u
. A func-
tion f : R
n
u
R
n
u
is said to belong to the incremental
sector [0, I]
W
if Jf (y), I Jf (y)
W
0 for almost all y
R
n
u
, where Jf (y) denotes the Jacobian of f evaluated at y.
Denition 2. A function [ : R
n
u
R
n
u
is said to belong
to 1
W
if the function [() is locally Lipschitz, belongs to
the incremental sector [0, I]
W
and [(0) =0.
Remark 1. If [() belongs to 1
W
then [() belongs to the
sector [0, I]
W
. When W =I, the W-product W
1
, W
2

W
coincides with the standard product W
T
1
W
2
. The sector
[0, I]
I
property coincides with the sector [0, I] property
dened in (Khalil, 1996, p. 403) (see also (Grimm et al.,
2003, Remark 1)). Moreover, the nonlinearity [() could be
a decentralized saturation function which belongs to 1
W
for
all diagonal positive denite W (see also (Grimm et al.,
2003, Remark 2)).
2. Problem denition
Consider a linear stable plant given by
P
_
x
p
=A
p
x
p
+B
p,u
u +B
p,w
w,
y =C
p,y
x
p
+D
p,yu
u +D
p,yw
w,
z =C
p,z
x
p
+D
p,zu
u +D
p,zw
w,
(1)
where x
p
R
n
p
is the plant state, u R
n
u
is the control
input, w R
n
w
is the exogenous input (possibly containing
disturbance, reference and measurement noise), y R
n
y
is
the plant output available for measurement, z R
n
z
is the
performance output (possibly corresponding to a weighted
tracking error) and A
p
, B
p,u
, B
p,w
, C
p,y
, D
p,yu
, D
p,yw
, C
p,z
,
D
p,zu
and D
p,zw
are matrices of suitable dimensions.
Assume also that an unconstrained controller of the form
C
_
x
c
=A
c
x
c
+B
c,w
w +B
c,y
u
c
,
y
c
=C
c
x
c
+D
c,w
w +D
c,y
u
c
,
(2)
is given (where x
c
R
n
c
is the controller state, y
c
R
n
u
is the controller output and A
c
, B
c,w
, B
c,y
, C
c
, D
c,w
and
D
c,y
are matrices of suitable dimensions). Also assume that
the controller interconnection to the linear plant through the
equations
u =y
c
, u
c
=y, (3)
is well-posed and guarantees internal stability of the arising
unconstrained closed-loop system (1)(3).
Suppose the control input of the plant is subject to the
nonlinearity [(), where [() belongs to 1
W
. Also suppose
that we can modify the input of the controller and the input
of the plant with the signals v
1
and v
2
, respectively, through
the equations
u =[(y
c
+v
2
) and u
c
=y +v
1
. (4)
Given an integer n
aw
0, in this paper we address the prob-
lem of designing an order n
aw
linear external anti-windup
compensator
AW
_
x
aw
=A
aw
x
aw
+B
aw
(y
c
+v
2
u),
v =
_
v
1
v
2
_
=C
aw
x
aw
+D
aw
(y
c
+v
2
u),
(5)
(where x
aw
R
n
aw
is the anti-windup state, v R
n
v
(with
n
v
:= n
y
+ n
u
) is the anti-windup output, and the matri-
ces A
aw
, B
aw
, C
aw
and D
aw
are of suitable dimensions) that
guarantees a desirable L
2
relationship between the exoge-
nous input w and the performance output z. Interconnection
(1), (2), (4), (5) will henceforth be called the (external) anti-
windup closed-loop system and is shown in Fig. 1.
Note that, since the block AW is linear, the structure rep-
resented in Fig. 1 automatically achieves one important re-
quirement of anti-windup augmentation. In particular, pro-
vided the interconnection of the block AW (correspond-
ing to Eq. (4)) is well-posed (namely, solutions exist) and
x
aw
(0) = 0, as long as [(y
c
()) y
c
(), the response of
the augmented closed-loop reproduces the unconstrained re-
sponse corresponding to the unconstrained interconnection
(3) between plant (1) and the unconstrained controller (2).
In addition to this, according to the design goals of Grimm
et al. (2003), we want to characterize here desirable proper-
ties that are associated with the behavior of the augmented
closed-loop in Fig. 1 for all those responses corresponding
to activation of the saturation nonlinearity. A desirable sta-
bility and performance property can be stated in terms of
Lyapunov analysis tools. In particular, the following de-
nition entails (sufcient) conditions for internal stability of
the anti-windup closed-loop system and for nite L
2
gain
from w to z for all [() that belong to 1
W
(see Grimm
et al. (2003, Remark 3) for details).
1990 G. Grimm et al. / Automatica 40 (2004) 19871996
Denition 3. Given the linear plant P in (1) and the uncon-
strained controller C in (2), a linear external anti-windup
compensator (5) of order n
aw
guarantees well-posedness
and quadratic performance of level if the external anti-
windup closed-loop system (1), (2), (4), (5) is such that, for
all [() that belong to 1
W
,
1. interconnection (4) is well-posed,
2. there exists a scalar c >0 and a quadratic Lyapunov
function V(x) = x
T
Px (with x := [x
T
p
x
T
c
x
T
aw
]
T
and
P = P
T
>0) such that its time derivative

V along the
dynamics of (1), (2), (4), (5) satises

V <cx
T
x
1

z
T
z +w
T
w, (x, w) = 0. (6)
3. External anti-windup feasibility
Given a generic plant-controller pair of type (1), (2), in
this section we will formulate an important feasibility re-
sult stating necessary and sufcient conditions for the ex-
istence of a static and, respectively, of a plant-order exter-
nal anti-windup compensator that guarantees well-posedness
and quadratic performance for some performance level (in
the sense of Denition 3). To this aim, we introduce two
systems that are special congurations of the plant and con-
troller. The rst system can be interpreted as a realization
of the unconstrained closed-loop system or as a realization
of the external anti-windup closed-loop system where [()
is replaced by the identity function ([ I) and the ex-
ternal anti-windup compensator output is identically zero.
This system has state x
cl
:= [x
T
p
x
T
c
]
T
R
n
cl
, where n
cl
:=
n
p
+n
c
and has the state-space representation
x
cl
=A
cl
x
cl
+B
cl,w
w,
z =C
cl,z
x
cl
+D
cl,zw
w, (7)
where A
cl
, B
cl,w
, C
cl,z
and D
cl,zw
are uniquely determined
by the matrices in (1) and (2).
The second system can be interpreted as an interconnec-
tion of (1), (2), (4) (just as in the anti-windup closed-loop
system but) where v is considered as an exogenous input and
where [() is replaced by the zero function. In particular,
choose a somewhat specialized exogenous input v
1
=v
1
+y
and note that [() 0 implies that v
2
does not alter the dy-
namics of the system. Dene the state x
ol
:= [x
T
p
x
T
c
]
T

R
n
cl
. Then the system with inputs w and v
1
and output z,
(with u 0) can be rewritten in state-space form as
x
ol
=A
ol
x
ol
+B
ol,w
w +B
ol
v
1
,
z =C
ol,z
x
ol
+D
ol,zw
w,
(8)
where the matrices A
ol
, B
ol
, B
ol,w
, C
ol,z
and D
ol,zw
are
uniquely determined by the matrices in (1) and (2). For com-
pleteness, explicit expressions for these matrices are given
in the following equations:
(9)
Denition 4. Let X = X
T
>0 and W : R
n
cl
R be such
that for all R
n
cl
, W() =
T
X. The function W() is
said to be a quadratic Lyapunov function for system (7) if
w=0 and x
cl
= 0 imply W(x
cl
), x
cl

I
<0. The function
W() is said to be a quadratic control Lyapunov function
for system (8) if w = 0, x
ol
= 0 and B
T
ol
Xx
ol
= 0 imply
W(x
ol
), x
ol

I
<0.
We are now ready to state the feasibility result of this
paper (whose proof is reported in Section 7).
Theorem 1. Given the plant P and the unconstrained con-
troller C in (1) and (2),
1. there exists a static external anti-windup compensator
that guarantees well-posedness and quadratic perfor-
mance of the closed-loop system if and only if any of
the following (equivalent) conditions hold:
(a) there exists a quadratic function that is both a
Lyapunov function for system (7) and a control
Lyapunov function for system (8);
(b) Given any B
ol
of full column rank that spans
the null space of B
T
ol
, there exists R = R
T
>0
such that
B
ol
T

(RA
T
ol
+A
ol
R)B
ol
<0,
RA
T
cl
+A
cl
R<0,
2. there exists a plant-order external anti-windup compen-
sator that guarantees well-posedness and quadratic per-
formance of the closed-loop system if and only if the
plant is asymptotically stable (A
p
is Hurwitz).
4. External anti-windup synthesis
In this section, we will formulate procedures for anti-
windup synthesis, which solve the external anti-windup de-
sign problem whenever it is feasible (according to Theorem
1) by means of solutions to always feasible LMIs. More-
over, these LMIs can be solved while globally minimizing
a performance index which corresponds to the gain in (6)
through the convex optimization tools provided by the LMI
setting.
According to the scheme in Fig. 1, dene q =y
c
+v
2
u.
Then, the interconnection of the plant and controller via the
equations
u =y
c
+v
2
q and u
c
=y +v
1
, (10)
where q is considered an additional exogenous input,
can be written as the following linear system with state
G. Grimm et al. / Automatica 40 (2004) 19871996 1991
x
cl
=[x
T
p
x
T
c
]
T
and state-space representation
x
cl
=A
cl
x
cl
+B
cl,w
w +B
cl,q
q +B
cl,v
v,
z =C
cl,z
x
cl
+D
cl,zw
w +D
cl,zq
q +D
cl,zv
v,
y
c
+v
2
=C
cl,y
x
cl
+D
cl,yw
w +D
cl,yq
q +D
cl,yv
v,
(11)
where the matrices of this system are uniquely determined
by the matrices in (1) and (2).
A rst procedure allows us to construct a static external
anti-windup compensator which optimizes the performance
level in Denition 3. As stated in Theorem 2, at the end
of this section, all the steps of the procedure can be accom-
plished if and only if the feasibility conditions in item 1 of
Theorem 1 are satised.
Procedure 1 (Static external anti-windup).
Step 1: Solve the feasibility LMIs. Find a solution
(R, ) to the following set of LMIs (possibly minimizing ):
L
T
_
_
RA
T
ol
+A
ol
R B
ol,w
RC
T
ol,z
B
T
ol,w
I D
T
ol,zw
C
ol,z
R D
ol,zw
I
_
_
L<0, (12a)
_
_
RA
T
cl
+A
cl
R B
cl,w
RC
T
cl,z
B
T
cl,w
I D
T
cl,zw
C
cl,z
R D
cl,zw
I
_
_
<0, (12b)
R =R
T
=
_
R
11
R
12
R
T
12
R
22
_
>0, (12c)
where L := diag(B
ol
, I, I).
Step 2: Construct and solve the anti-windup com-
pensator LMI. Choose any o R, o >0 and dene
U = oW
1
. Using R from Step 2, construct the matrices
H R
n
v
(n
cl
+n
u
+n
w
+n
z
)
, G R
n
u
(n
cl
+n
u
+n
w
+n
z
)
and
T R
(n
cl
+n
u
+n
w
+n
z
)(n
cl
+n
u
+n
w
+n
z
)
, as follows:
H=[B
T
cl,v
D
T
cl,yv
0
n
v
n
w
D
T
cl,zv
],
G=[0
n
u
n
cl
U 0
n
u
n
w
0
n
u
n
z
],
T=He
_
_
_
_
_
_
_
_
_
_
A
cl
R B
cl,q
U +RC
T
cl,y
B
cl,w
RC
T
cl,z
0 D
cl,yq
U U D
cl,yw
UD
T
cl,zq
0 0

2
I D
T
cl,zw
0 0 0

2
I
_

_
_
_
_
_
_
.
Finally, solve the following LMI for
R
n
v
n
u
:
T+G
T

T
H +H
T
G<0, (13)
and compute the static anti-windup gain as D
aw
=.
A similar procedure can be formulated for the construc-
tion of a plant order external anti-windup compensator which
optimizes the performance level in Denition 3. Theo-
rem 2 below states that the steps of the procedure can be
accomplished if and only if item 2 of Theorem 1 is fullled,
namely if and only if the plant is asymptotically stable.
Procedure 2 (Plant order external anti-windup).
Step 1: Solve the feasibility LMIs. Find a solution
(R
11
, S, ) to the following set of LMIs (possibly minimiz-
ing ):
L
T
_
_
RA
T
ol
+A
ol
R B
ol,w
RC
T
ol,z
B
T
ol,w
I D
T
ol,zw
C
ol,z
R D
ol,zw
I
_
_
L<0, (14a)
_
_
SA
T
cl
+A
cl
S B
cl,w
SC
T
cl,z
B
T
cl,w
I D
T
cl,zw
C
cl,z
S D
cl,zw
I
_
_
<0, (14b)
R =R
T
=
_
R
11
S
12
S
T
12
S
22
_
>0, (14c)
S =S
T
=
_
S
11
S
12
S
T
12
S
22
_
>0, (14d)
R
11
S
11
>0, (14e)
where L : =diag(B
ol
, I, I).
Step 2: Construct the matrix Q. Set n
aw
=n
p
. Using the
solution (R
11
, S, ) from Step 1, construct the matrix N :=
[
N
1
0
] R
n
cl
n
aw
, where N
1
=N
T
1
is determined by N
1
N
1
=
(R
11
S
11
)(S
11
S
12
S
1
22
S
T
12
)
1
(R
11
S
11
) +(R
11
S
11
),
whose right-hand side can be proven to be positive de-
nite. Construct also the symmetric matrix M R
n
aw
n
aw
as
M := N
1
(R
11
S
11
)
1
N
1
. Finally, dene the matrix Q
R
(n
cl
+n
aw
)(n
cl
+n
aw
)
as
Q :=
_
R N
N
T
M
_
. (15)
Step 3: Build required matrices. Set n = n
p
+ n
c
+ n
aw
.
Construct the matrices A
0
R
nn
, B
q0
R
nn
u
, C
y0

R
n
u
n
, D
yq0
R
n
u
n
u
, C
z0
R
n
z
n
, D
zq0
R
n
z
n
u
,
B
w
R
nn
w
, D
zw
R
n
z
n
w
and D
yw
R
n
u
n
w
as
Step 4: Anti-windup compensator LMI. Choose o R,
o >0 and dene U = oW
1
. Based on Steps 2 and 3,
construct the matrices H R
(n
aw
+n
v
)(n+n
u
+n
w
+n
z
)
, T
R
(n+n
u
+n
w
+n
z
)(n+n
u
+n
w
+n
z
)
and G R
(n
aw
+n
u
)(n+n
u
+n
w
+n
z
)
as follows:
T=He
_
_
_
_
_
_
_
_
A
0
Q B
q0
U +QC
T
y0
B
w
QC
T
z0
0 D
yq0
U U D
yw
UD
T
zq0
0 0

2
I D
T
zw
0 0 0

2
I
_

_
_
_
_
_
,
1992 G. Grimm et al. / Automatica 40 (2004) 19871996
Finally, solve LMI (13) in the unknown
R
(n
aw
+n
v
)(n
aw
+n
u
)
, and compute the matrices of the external
anti-windup compensator (5) by partitioning the solution as
follows:
_
A
aw
B
aw
C
aw
D
aw
_
=.
Theorem 2. All the steps of Procedure 1 (respectively, Pro-
cedure 2) can be accomplished if and only if the conditions
at item 1 (respectively, item 2) of Theorem 1 are satised.
Moreover, the arising solution satises Denition 3 with
performance level .
Proof. See Section 7.
Remark 2. The static and plant order external anti-windup
constructions can be generalized to the case of arbitrary (but
xed) order external anti-windup compensation. However,
paralleling the full-authority results of Grimm et al. (2003),
in the general case the LMIs at Step 1 of Procedures 1 and 2
transform into nonlinear matrix inequalities due to the pres-
ence of an extra rank condition thereby losing the nice con-
vexity property of the previous constructions. Nevertheless,
the rest of the procedure remains almost unchanged, so that
once Step 1 is accomplished (possibly using approximate
methods), the rest of the construction is straightforward. In
particular, for this general case one may follow Procedure 2
by replacing (14) in Step 1 with
(14a)(14c), (16a)
R S0, (16b)
rank(R S)n
aw
(16c)
and by replacing step 2 with step 2 of (Grimm et al., 2003,
Procedure 1).
5. Simulation example
Consider the electrical network, introduced in Grimm,
Postlethwaite, Teel, Turner, and Zaccarian (2001). The non-
linearity [() is a decentralized saturation function, as dis-
cussed in Remark 1, with unitary saturation limit M
1
= 1.
An optimal static external anti-windup compensator con-
structed according to Procedure 1 minimizing , is found
to be D
aw
=
_
5.1635
376.8937
_
, which induces performance level
=86.07. Furthermore, an external anti-windup compensator
of order equal to that of the plant can be constructed accord-
ing to Procedure 2 while minimizing which leads to a com-
pensator that induces performance level = 58.81 and has
0 5 10 15 20 25 30
2
1
0
1
2
P
l
a
n
t

o
u
t
p
u
t
Time [s]
0 5 10 15 20 25 30
1.5
1
0.5
0
0.5
1
1.5
P
l
a
n
t

i
n
p
u
t
Time [s]
Fig. 2. Responses of the unconstrained (solid), saturated (dotted), full
authority static anti-windup (dashed), full authority dynamic (dash-dotted)
and static and dynamic external anti-windup closed-loop systems (stars
and crosses +, respectively) to the reference prole introduced in
Grimm et al. (2001).
state-space representation
It is interesting to compare these external anti-windup com-
pensators with the full-authority anti-windup compensators
proposed in Grimm et al. (2003) which use the same per-
formance index. First, as reported in Grimm et al. (2001),
the full-authority anti-windup compensators induce perfor-
mance levels 85.20 and 58.66 for static and dynamic anti-
windup compensation, respectively. It is notable that the full-
authority anti-windup performance level is nearly identical
to that which can be induced by external anti-windup. Fur-
thermore, using the same reference prole as in Grimm et
al. (2001), the response of the external anti-windup closed-
loop system is virtually indistinguishable from the response
of the full-authority anti-windup closed-loop system in both
the static and plant order cases. This is shown in Fig. 2
where the unconstrained response is represented by the solid
line, the saturated (without anti-windup) response is dot-
ted, the full-authority optimal static anti-windup system re-
sponse is dashed and the full-authority optimal plant order
anti-windup system response is dash-dotted. However, these
last two curves are hardly visible because they are perfectly
matched up (therefore, hidden) by the samples of the re-
sponses using the optimal external static and plant order anti-
windup compensators (which are represented, respectively,
by stars and crosses +).
G. Grimm et al. / Automatica 40 (2004) 19871996 1993
0 5 10 15 20 25 30
3
2
1
0
1
2
3
P
l
a
n
t

o
u
t
p
u
t
Time [s]
Fig. 3. Responses of the unconstrained (solid), saturated (dotted), full
authority dynamic anti-windup (dash-dotted) and conditioning technique
(dashed).
Generally speaking, the class of anti-windup problems
solved by the procedures given in Section 4 is larger than
the class of problems solved by most of the classical anti-
windup schemes used in the industrial environment (see,
e.g., the anti-windup methods for PID controllers surveyed
in Bohn & Atherton (1995)). Indeed, our procedures are ap-
plicable to generic MIMO linear control systems, whereas
most of those simple approaches are for SISO schemes with
PID controllers.
4
Nevertheless, for this example, where a
PID controller is used and the plant is a SISO one, it is pos-
sible to compare the responses of Fig. 2 to the response re-
sulting from a simple anti-windup strategy. Here, we select
the tracking anti-windup technique summarized in Bohn
and Atherton (1995) and corresponding to the condition-
ing technique (see, e.g., Hanus (1988)). For a SISO PID
controller, tracking anti-windup can be interpreted as an ex-
ternal compensation scheme where the modication signal
required at the integrator input is actually moved at the input
of the controller (corresponding to the signal v
1
) and its con-
sequent effect at the controller output through the derivative
and proportional branches can be canceled out by a suitable
dynamic selection of the signal v
2
.
Since the controller (see Grimm et al. (2001)) approxi-
mates a PID controller with K
p
=80, K
d
=1, K
i
=20, we can
compare our responses to that obtained applying the con-
ditioning technique and selecting the anti-windup parame-
ter T
t
using the rule of thumb T
t
= T
i
suggested in Bohn
and Atherton (1995) (with reference to the tracking anti-
windup notation). (Note that one advantage of our approach
versus the conditioning technique is that the former has a
formal proof of closed-loop stability, whereas the latter does
not rely on any formal stability statement.) This comparison
is reported in Fig. 3, where the conditioning technique re-
sponse (dashed) is able to induce graceful convergence to the
desired set-point but the optimality-based selection of our
4
In fairness, a notable exception to this is the conditioning technique
used in Peng et al. (1998) where multivariable control systems are ad-
dressed, although no stability or performance guarantee is provided using
that approach.
anti-windup gains (dash-dotted) guarantees a faster conver-
gence (thus corresponding to a more desirable performance).
6. Interpretations of performance bounds
Interesting interpretations of performance bounds can be
deduced from conditions (12), (14) and Theorem 2 by way
of the following statements.
Denition 5. Let X =X
T
>0 and W : R
n
cl
R be such
that for all R
n
cl
, W() =
T
X. The function W()
is said to be a quadratic disturbance attenuation Lyapunov
function for system (7) which guarantees L
2
attenuation
if (w, x
cl
) = 0 implies W(x
cl
), x
cl

I
<
1

z
T
z +w
T
w.
The function W() is said to be a quadratic disturbance
attenuation control Lyapunov function for system (8) which
guarantees L
2
attenuation if
_
x
ol
w
_
= 0 and B
T
ol
Xx
ol
= 0
imply W(x
ol
), x
ol

I
<
1

z
T
z +w
T
w.
The proof of the following statement is based on simple
LMI transformations and is omitted due to space constraints.
Lemma 1. Let B
ol
be dened as any matrix of full column
rank that spans the null space of B
T
ol
. Given the function
W(x
ol
) = x
T
ol
Xx
ol
, where X = X
T
>0, the following state-
ments are equivalent:
1. The function W() is a quadratic disturbance attenu-
ation control Lyapunov function for system (8) which
guarantees L
2
attenuation .
2. There exists a matrix K such that the following matrix
inequality holds:
He
_
_
_
_
(A
ol
+B
ol
K)X
1
B
ol,w
X
1
C
T
ol,z
0

2
I
n
w
D
T
ol,zw
0 0

2
I
n
z
_
_
_
_
<0.
3. The following matrix inequality holds (with L
:= diag(B
ol
, I, I)):
L
T
_
_
X
1
A
T
ol
+A
ol
X
1
B
ol,w
X
1
C
T
ol,z
B
T
ol,w
I
n
w
D
T
ol,zw
C
ol,z
X
1
D
ol,zw
I
n
z
_
_
L<0.
By applying Lemma 1 to Theorem 2, there exists a static
external anti-windup compensator that guarantees well-
posedness and quadratic performance of level if and only
if there exists a quadratic function that is both a disturbance
attenuation control Lyapunov function for system (8) which
guarantees L
2
attenuation and a disturbance attenuation
Lyapunov function for system (7) which guarantees L
2
at-
tenuation . A similar statement can be made for n
aw
>0 by
considering Remark 2; the difference being that the single
quadratic function is replaced by two quadratic functions
that must satisfy the coupling conditions (16b) and (16c).
By observing the structure of the LMI constraints (14a)
and (14b), we also deduce similar performance bounds as
1994 G. Grimm et al. / Automatica 40 (2004) 19871996
those observed in Grimm et al. (2003).
5
As a matter of
fact, based on the Bounded Real Lemma, (14b) constrains
the performance level to be greater than the L
2
gain of the
closed-loop system, while by the structure of the matrix B
ol
(see Eq. (9)) it is easily seen that a sufcient (and necessary)
condition for (14a) to hold is that the performance level is
greater than the L
2
gain of the plant.
7. Proof of the main results
Similar to the construction of system (11), by dening the
state x := [x
T
p
x
T
c
x
T
aw
]
T
, the interconnection of (1), (2), (5),
(10) can be written with exogenous input q as
x =Ax +B
q
q +B
w
w,
y
c
+v
2
=C
y
x +D
yq
q +D
yw
w,
z =C
z
x +D
zq
q +D
zw
w,
(17)
where the matrices of this system are uniquely determined
by the matrices in (1), (2) and (5). If
q =y
c
+v
2
[(y
c
+v
2
). (18)
then the interconnection of (17) and (18) corresponds to the
external anti-windup closed-loop system.
We report next three preliminary statements needed for the
proof of Theorems 1 and 2. The rst two are straightforward
generalizations of (Grimm et al., 2003, Theorems 1 & 3),
respectively, and are reported without proof.
Proposition 1. Given a symmetric positive denite matrix
W, the external anti-windup closed-loop system (17), (18)
and a desired performance level R, >0, the external
anti-windup closed-loop system is well-posed and guaran-
tees quadratic performance of level if and only if there
exists a matrix Q=Q
T
>0 and o R, o >0 such that the
following matrix inequality
_
_
_
_
QA
T
+AQ B
q
U +QC
T
y
B
w
QC
T
z
UB
T
q
+C
y
Q D
yq
U +UD
T
yq
2U D
yw
UD
T
zq
B
T
w
D
T
yw
I D
T
zw
C
z
Q D
zq
U D
zw
I
_

_
<0,
with U := oW, is satised.
Proposition 2. Given the plant P in (1), the unconstrained
controller C in (2), an integer n
aw
0 and a scalar , there
exists a linear external anti-windup compensator of order
n
aw
that guarantees well-posedness and quadratic perfor-
mance of level if and only if there exist matrices R and S
that satisfy (16).
5
Indeed, by constraining the authority of the anti-windup compensator
to only act externally on the controller state equation, we expect not
to achieve a better performance than that of the construction in Grimm
et al. (2003). On the other hand, our scheme does not require authority
on the internal states of the controller.
Remark 3. Proposition 2 generalizes Theorem 3 of Grimm
et al. (2003). Indeed, if B
c,y
has full-row rank then one
should expect the external anti-windup problem can be re-
duced to the full-authority anti-windup problem discussed
in Grimm et al. (2003). In fact, if B
c,y
has full row rank
then a valid choice for B
ol
is B
ol
T

=[I 0] and Proposition


2 exactly corresponds to Theorem 3 of Grimm et al. (2003).
Furthermore, if B
c,y
does not have full row rank then Propo-
sition 2 differs from Theorem 3 of Grimm et al. (2003) be-
cause the dimension of the null space of B
ol
is larger than
n
p
and the matrix condition (14a) is more restrictive. Due to
this similarity, the proof of Proposition 2 follows the same
steps as the proof of Theorem 3 of Grimm et al. (2003): it
appeals to Proposition 1 and essentially follows the proof
technique of the main result in Gahinet and Apkarian (1994).
To prove Theorems 1 and 2 we also need the following
two technical results (the rst result is a special case of
Lemma 1 and its proof is omitted).
Lemma 2. Let B
ol
be dened as any matrix of full column
rank that spans the null space of B
T
ol
. Given the function
W(x
ol
) = x
T
ol
Xx
ol
, where X = X
T
>0, the following state-
ments are equivalent.
1. The function W() is a quadratic control Lyapunov func-
tion for system (8).
2. There exists a matrix K such that the following ma-
trix inequality holds: X
1
(A
T
ol
+ K
T
B
T
ol
) + (A
ol
+
B
ol
K)X
1
<0.
3. The following matrix inequality holds: B
ol
T

(X
1
A
T
ol
+
A
ol
X
1
)B
ol
<0.
Lemma 3. The LMI conditions (14) have a feasible solution
(R, S, ) if and only if the plant is asymptotically stable (A
p
is Hurwitz).
Proof. Based on the internal stability assumption of the
unconstrained closed-loop system, A
cl
is Hurwitz. There-
fore, there exists a matrix S =S
T
>0 such that (14b) is sat-
ised for some =
1
. Consequently, there exists a matrix
S =S
T
>0 such that
SA
T
cl
+A
cl
S =He
__
m
11
m
12
0 m
22
__
<0, where (19)
m
11
=(A
p
+B
p,u
A
u
D
c,y
C
p,y
)S
11
+S
12
(B
p,u
A
u
C
c
)
T
m
12
=S
11
(B
c,y
A
y
C
p,y
)
T
+S
12
(A
c
+B
c,y
A
y
D
p,yu
C
c
)
T
+(A
p
+B
p,u
A
u
D
c,y
C
p,y
)S
12
+(B
p,u
A
u
C
c
)S
22
m
22
=B
c,y
A
y
C
p,y
S
12
+(A
c
+B
c,y
A
y
D
p,yu
C
c
)S
22
By Lemma 2, (and using (14c)) the inequality B
ol
T

(RA
T
ol
+
A
ol
R)B
ol
<0 holds if and only if there exists a matrix
G. Grimm et al. / Automatica 40 (2004) 19871996 1995
K =[K
1
K
2
] such that
(20)
With the particular choice
K =[K
1
K
2
] =[0 A
y
(D
p,yu
C
c
+C
p,y
S
12
S
1
22
)],
inequality (20) gains two nice traits: the lower right block
of (20) is the same as the lower right block of (19) (and
hence negative denite) and the off diagonal blocks of (20)
do not depend on R
11
. Thus, R
11
can be chosen to make
R
11
A
T
p
+ A
p
R
11
negative denite and sufciently large
to make the whole matrix in (20) negative denite. Thus
B
ol
T

_
RA
T
ol
+A
ol
R
_
B
ol
<0. With this choice of R and S,
= can be chosen sufciently large to satisfy the rest of
(14a) and (14b).
Proof of Theorem 2. When n
aw
= 0, Proposition 2 guar-
antees that there exists a static linear external anti-windup
compensator that induces quadratic performance of level
if and only if there exist matrices R and S that satisfy (16).
However, (16c) implies R = S, and the remaining condi-
tions in (16) reduce to (12).Thus there exists a static linear
external anti-windup compensator that guarantees quadratic
performance of level if and only if there exists a matrix R
that satises (12). Moreover, it is easy to check that there ex-
ists that satises (12) if and only if there exists a matrix R
such that B
ol
T

(RA
T
ol
+A
ol
R)B
ol
<0 and RA
T
cl
+A
cl
R<0.
Therefore, to complete the proof of the static case we only
need to show that, once step 1 is accomplished, there exists
a solution to the LMI of Step 2 of Procedure 1.
As for the plant order case, by Lemma 3, A
p
is Hurwitz if
and only if the LMI conditions (14) have a feasible solution
(R, S, ). Thus all that needs to be proven for this second
case is that if Step 1 of Procedure 2 can be accomplished
then the remaining steps can be accomplished too.
To show that Procedure 1 (respectively, Procedure 2) can
be completed once Step 1 is accomplished, the same steps
used in the proof of (Grimm et al., 2003, Theorem 4) can be
followed. These are based on the fact that (13) is feasible if
Q
1
=
_
R N
N
T
M
_
1
=
_
S
1
P
12
P
T
12
P
22
_
(where P
12
and P
22
are
matrices of appropriate dimensions) and that, by the results
in Packard (1994) and the formulae for the inversion of
block matrices (Zhou, Doyle, & Glover, 1995, p. 23), this
last inequality always holds if we select Q=R for Procedure
1 (respectively, Q as in (15) for Procedure 2).
Proof of Theorem 1. 1. The proof of Theorem 2 establishes
the existence of a static linear external anti-windup compen-
sator that guarantees well-posedness and quadratic perfor-
mance of the closed-loop system if (b) holds. The proof is
completed by noting that Lemma 2 shows that (a) and (b)
are equivalent.
2. The (constructive) proof follows directly from
Theorem 2.
8. Conclusions
In this paper, we characterized linear external anti-windup
synthesis for linear windup-prone control systems with in-
put saturation. We give a system theoretic characterization
of the feasibility of this anti-windup augmentation problem,
in addition to constructive LMI-based procedures for opti-
mal anti-windup synthesis in the static and plant-order cases.
Simulation results on a physically motivated example con-
rm that the proposed construction is an effective approach
to the solution of the windup problem.
References
Bemporad, A., Teel, A., & Zaccarian, L. (2004). Anti-windup synthesis
via sampled-data piecewise afne optimal control. Automatica, 40(4),
549562.
Bohn, C., & Atherton, D. (1995). An analysis package comparing PID
anti-windup strategies. IEEE Control Systems Magazine, 15(2), 3440.
Cao, Y., Lin, Z., & Ward, D. (2002a). An antiwindup approach to enlarging
domain of attraction for linear systems subject to actuator saturation.
IEEE Transactions on Automatic Control, 47(1), 140145.
Cao, Y., Lin, Z., & Ward, D. (2002b). Antiwindup design for linear
systems subject to input saturation. Journal of Guidance Navigation
and Control, 25(3), 455463.
Fertik, H., & Ross, C. W. (1967). Direct digital control algorithm with
anti-windup feature. ISA Transactions, 6(4), 317328.
Gahinet, P., & Apkarian, P. (1994). A linear matrix inequality approach to
H

control. International Journal of Robust and Nonlinear Control,


4, 421448.
Gahinet, P., Nemirovski, A., Laub, A., & Chilali, M. (1995). LMI control
toolbox. The MathWorks Inc.
Grimm, G., Hateld, J., Postlethwaite, I., Teel, A., Turner, M., & Zaccarian,
L. (2003). Antiwindup for stable linear systems with input saturation:
An LMI-based synthesis. IEEE Transactions on Automatic Control,
48(9), 15091525.
Grimm, G., Postlethwaite, I., Teel, A., Turner, M., & Zaccarian, L. (2001).
Case studies using linear matrix inequalities for optimal anti-windup
synthesis. In Proceedings of the European control conference. Porto,
Portugal.
Hanus, R. (1988). Antiwindup and bumpless transfer: A survey. In
Proceedings of the 12th IMACS world congress (pp. 5965), Vol. 2.
Paris, France.
Khalil, H. (1996). Nonlinear systems. (2nd ed), Englewood Cliffs, USA:
Prentice-Hall.
Kothare, M., Campo, P., Morari, M., & Nett, N. (1994). A unied
framework for the study of anti-windup designs. Automatica, 30(12),
18691883.
Mulder, E., Kothare, M., & Morari, M. (2001). Multivariable anti-windup
controller synthesis using linear matrix inequalities. Automatica, 37(9),
14071416.
Packard, A. (1994). Gain scheduling via linear fractional transformation.
Systems and Control Letters, 22(2), 7992.
Park, J., & Choi, C. (1995). Dynamic compensation method for
multivariable control systems with saturating actuators. IEEE
Transactions on Automatic Control, 40(9), 16351640.
1996 G. Grimm et al. / Automatica 40 (2004) 19871996
Peng, Y., Vran ci c, D., Hanus, R., & Weller, S. (1998). Anti-windup designs
for multivariable controllers. Automatica, 34(12), 15591565.
Teel, A. (1999). Anti-windup for exponentially unstable linear systems.
International Journal of Robust and Nonlinear Control, 9, 701716.
Teel, A., & Kapoor, N. (1997). The L
2
anti-windup problem: Its denition
and solution. In Proceedings of the 4th ECC. Brussels, Belgium.
Wu, F., & Lu, B. (2004). Anti-windup control design for exponentially
unstable LTI systems with actuator saturation. Systems and Control
Letters, 52(34), 304322.
Zaccarian, L., & Teel, A. (2002). A common framework for anti-
windup bumpless, transfer and reliable designs. Automatica, 38(10),
17351744.
Zaccarian, L., & Teel, A. (2004). Nonlinear scheduled anti-windup design
for linear systems. IEEE Transactions on Automatic Control.
Zheng, A., Kothare, M. V., & Morari, M. (1994). Anti-windup design
for internal model control. International Journal of Control, 60(5),
10151024.
Zhou, K., Doyle, J., & Glover, K. (1995). Robust and optimal control.
Englewood Cliffs, NJ: Prentice-Hall.
Gene Grimm received his Ph.D. degree in
electrical and computer engineering from
the University of California, Santa Barbara,
in 2003. Since then he has been with the
Space and Airborne Systems business unit
of Raytheon Company in El Segundo, where
he has worked on line-of-sight control de-
sign and analysis.
Andrew R. Teel received his Ph.D. degree
in Electrical Engineering from the Univer-
sity of California, Berkeley, 1992. He is
currently a professor in the Electrical and
Computer Engineering Department at the
University of California, Santa Barbara
where he is also director of the Center for
Control Engineering and Computation. He
is a Fellow of the IEEE.
Luca Zaccarian received his Ph.D. degree
in Computer Science and Control Engineer-
ing in 2000 from the University of Rome,
Tor Vergata, where he is currently an assis-
tant professor (ricercatore). In 19982000 he
spent approximately two years as a visiting
researcher at the CCEC of the UC, Santa
Barbara (USA). During the summers 2003
and 2004 he was a visiting professor at
the EEE department of the University of
Melbourne (Australia).

Das könnte Ihnen auch gefallen