Sunteți pe pagina 1din 13

Computers and Chemical Engineering 24 (2000) 921

Stabilization of nonlinear processes with input constraints


Navneet Kapoor
1
, Prodromos Daoutidis *
Department of Chemical Engineering and Materials Science, Uni6ersity of Minnesota, 421 Washington A6e. SE, Minneapolis, MN 55455, USA
Received 26 March 1998; received in revised form 4 January 2000; accepted 4 January 2000
Abstract
This work deals with nonlinear processes which, in the absence of input constraints, can be locally stabilized by a linearizing
static state feedback law. Such systems, under unconstrained linearizing control laws, possess a cascaded structure between a
(asymptotically or exponentially) stable nonlinear subsystem and an exponentially stable linear subsystem, which allows for a
straightforward stability analysis. In the presence of input constraints, however, this cascaded structure breaks down to an
interconnection between two nonlinear subsystems; analyzing the stability of such interconnections is a rather cumbersome task,
that typically results in conservative estimates of regions of stability. In this article, we present an analysis framework for the local
stabilization of such processes and the estimation of regions of closed-loop stability in the presence of input constraints. The
proposed approach entails: (i) specifying a region in statespace where the closed-loop system behaves effectively as a cascade and
asymptotic stability can be guaranteed in the presence of constraints, pro6ided that the states of the system remain in this region
for all times; and (ii) constructing invariant sets within this region that qualify as regions of closed-loop stability. A detailed case
study is carried out on a polymerization reactor example and the desirable features of the proposed methodology are aptly
illustrated. 2000 Elsevier Science Ltd. All rights reserved.
www.elsevier.com/locate/compchemeng
1. Introduction
All practical control systems are subject to input
constraints, whose presence may signicantly degrade
the stability/performance of the closed-loop system.
Common problems in the case of static feedback con-
trollers include restricted regions of closed-loop stabil-
ity and sluggishness of the closed-loop response. On the
other hand, static controllers are commonly augmented
with dynamics (e.g. due to integral action or an ob-
server for the process states). In the case of such
dynamic controllers, input constraints result in addi-
tional stability/performance degradation, that is com-
monly referred to as windup and manifests itself in the
form of undesired overshoots, oscillations and even
instability. In both cases, in order to ensure closed-loop
stability in the presence of constraints, one has to either
design controllers that are globally stabilizing or guar-
antee that the initial condition on the states lies within
the region of closed-loop stability; the latter in turn
involves estimating such regions of closed-loop
stability.
Over the last few decades various research directions
have been pursued to address the adverse effects of
input constraints. One direction has focused on pre-
venting excessive performance degradation due to
windup, for linear (A, stro m & Rundqwist, 1989; Hanus
& Kinnaert, 1989; Campo & Morari, 1990; Kothare,
Campo, Morari & Nett, 1994; Coulibaly, Maiti &
Brosilow, 1995; Kapoor, Teel & Daoutidis, 1998) and
nonlinear (Kendi & Doyle, 1997; Valluri & Soroush,
1998; Kapoor & Daoutidis, 1999) systems. Model pre-
dictive control formulations, which directly incorporate
the input constraints in the optimization problem that
determines the control action, have also been studied
extensively for linear (Lee & Cooley, 1997) and nonlin-
ear (Oliveira & Morari, 1994; Scokaert & Rawlings,
* Corresponding author. Tel: +1-612-6258818; fax: +1-612-
6267246.
E-mail address: daoutidi@cems.umn.edu (P. Daoutidis)
1
Current afliation: GE Corporate Research and Development,
P.O. Box 8, Schenectady, NY 12301, USA.
0098-1354/00/$ - see front matter 2000 Elsevier Science Ltd. All rights reserved.
PII: S0098- 1354( 00) 00296- 9
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 10
1995; Mayne, 1997; Valluri, Soroush & Nikravesh,
1998) systems.
Another direction of research, pursued mainly in the
systems literature, has focused on designing globally (or
semi-globally) stabilizing control laws. To this end, it
has been established that only stabilizable linear sys-
tems with eigenvalues in the closed left-half plane can
be globally stabilized using smooth constrained control
laws, and such static control laws have been con-
structed (Sussmann, Sontag & Yang, 1994). The global
stabilization of a special class of constrained nonlinear
systems that are Lyapunov stable has also recently been
addressed (Lin, 1994); tools to address global and
semi-global stabilizability for nonlinear systems that
conform to specic forms and/or satisfy certain Lips-
chitz-like conditions have also been developed (Lin,
Sontag & Wang, 1993; Teel & Praly, 1995; Teel, 1996).
However, for more general nonlinear systems, e.g. sys-
tems that are partially linearizable via unconstrained
static state feedback, global asymptotic stability can be
guaranteed only under very restrictive conditions, even
in the absence of constraints (Isidori, 1995).
The problem of estimating regions of closed-loop
stability in the presence of input constraints has also
attracted considerable attention. For linear systems,
research to this end has been carried out either within a
frequency-domain setting (Fuller, 1969; Glattfelder,
Eck & Schaufelberger, 1995) or within a Lyapunov
analysis setting (Gutman & Hagander, 1985; Gilbert &
Tan, 1991). Other recent results for linear systems
concern the estimation of regions of stability for un-
stable discrete-time systems (Zheng & Morari, 1995)
and the estimation of hypercuboidal regions of stability
under static state feedback laws (Kapoor & Daoutidis,
1997). For nonlinear systems under locally stabilizing
controllers, the requisite analysis is considerably more
complex. A Lyapunov analysis may be carried out in
principle, with the conservativeness of the results de-
pending strongly upon the choice of the Lyapunov
function candidates. One such effort has been made in
(Pappas, Lygeros & Godbole, 1995) leading to esti-
mates of regions of stability that comprised of level sets
wherein the manipulated input remained unconstrained.
In this work, we focus on nonlinear processes which,
in the absence of input constraints, can be locally
stabilized by a linearizing static state feedback law.
Such systems, under unconstrained linearizing control
laws, possess a cascaded structure between a (asymptot-
ically or exponentially) stable nonlinear subsystem and
an exponentially stable linear subsystem, which allows
for a straightforward stability analysis. In the presence
of input constraints, however, this cascaded structure
breaks down to an interconnection between two nonlin-
ear subsystems, which complicates the stability analysis
considerably. Motivated by this, we propose a stability
analysis framework that entails: (i) specifying a region
in state-space where the closed-loop system behaves
effectively as a cascade and asymptotic stability can be
guaranteed in the presence of constraints, pro6ided that
the states of the system remain in this region for all
times; and (ii) constructing invariant sets within this
region that qualify as regions of closed-loop stability. A
case study on a polymerization reactor example is used
to illustrate the application of the proposed framework
and the exibility that it provides in addressing trade-
offs between the size of the region of closed-loop stabil-
ity and closed-loop performance.
2. Preliminaries
We will consider nonlinear systems of the form:
x; =f(x) +g(x)sat(u) (1)
where x-XR
n
, X being an open and connected set,
denotes the vector of state variables, u-R denotes the
manipulated input, f(x) and g(x) are vector elds on
R
n
, and
sat(u) =

u
max
, uu
max
u, u
min
uu
max
u
min
, uu
min

(2)
is the standard mathematical representation of the satu-
ration operator.
Referring to the system of Eq. (1), it is assumed that
there exists a smooth scalar function h(x) with relative
order rn, well dened x-X. This guarantees the
existence of a local coordinate transformation (Isidori,
1995):

p
z
1

z
r

=(x) =

t(x)
h(x)

L
f
r1
h(x)

(3)
valid for x-MX, which transforms the system of Eq.
(1) into the following form:
p; =q(p, z)
z;
1
=z
2

z;
r1
=z
r
z;
r
=[L
f
r
h(x)]
x=
1
(p,z)
+[L
g
L
f
r1
h(x)]
x=
1
(p,z)
sat(u)
(4)
where p; =q(p, z) is the zero dynamics of the system of
Eq. (4). It is assumed that the origin is an equilibrium
point of the nonlinear system of Eq. (4), which can
always be ensured by working in suitable deviation
variables.
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 11
2.1. Local stabilization in the absence of constraints
In the absence of active input constraints, the local
stabilization of the origin of the system of Eq. (4) can
be achieved in a straightforward way subject to ap-
propriate stability conditions on the zero dynamics.
To this end, let us now make the following assump-
tion:
Assumption 1 The origin of the system p; =q(p, 0) is
locally exponentially stable.
The above assumption ensures that the states p
stay bounded for small initial conditions and small
z, i.e. that the zero dynamics possesses (at least lo-
cally) an input-to-state (ISS) stability property (see
(Khalil, 1996) for further details). Consider now a
linearizing static state feedback law of the form:
u=
[L
f
r
h(x)]
x=
1
(p,z)
+q
1
z
1
+q
2
z
2
+ ... +q
r
z
r
[L
g
L
f
r1
h(x)]
x=
1
(p,z)
(5)
where q
1
, , q
r
are adjustable controller parameters.
Under the above control law, the following closed-
loop system:
p; =q(p, z)
z;
1
=z
2

z;
r1
=z
r
z;
r
=q
1
z
1
+q
2
z
2
+ ... +q
r
z
r
(6)
which is a cascade of an r-dimensional linear system
and an (nr)-dimensional nonlinear system is ob-
tained. For a choice of the adjustable parameters q
i
such that the eigenvalues of the z subsystem, i.e. the
eigenvalues of the matrix
A=

0 1 0 0
0 0 1 0

q
1
q
2
q
3
q
r

(7)
have negative real parts, the local asymptotic stability
of the origin follows directly (Khalil, 1996). Let then
the region of stability (domain of attraction) of the
system of Eq. (6) be BMX.
2.2. Problem formulation
Consider now the closed-loop system under the
control law of Eq. (5), in the case that the input
constraints are active:
p; =q(p, z)
z;
1
=z
2

z;
r1
=z
r
z;
r
=[L
f
r
h(x)]
x=
1
(p,z)
+[L
g
L
f
r1
h(x)]
x=
1
(p,z)
sat
[L
f
r
h(x)]
x=
1
(p,z)
+q
1
z
1
+q
2
z
2
+ ... +q
r
z
r
[L
g
L
f
r1
h(x)]
x=
1
(p,z)

(8)
Note that in the presence of input constraints, the
control law of Eq. (5) can no longer cancel the non-
linearities in the z
r
subsystem and hence cannot im-
pose a cascaded, partially linear structure in the
closed-loop system. The system of Eq. (8) is, instead,
an interconnection of two nonlinear subsystems. Ana-
lyzing the stability of such systems clearly poses a
greater problem than that of cascaded systems.
Nonetheless, one can, in principle, construct appropri-
ate Lyapunov functions and carry out a stability
analysis of such systems with estimates of regions of
stability being appropriate level sets. One such analy-
sis has been carried out in (Pappas et al., 1995) where
level sets were constructed wherein the control input
stays necessarily unconstrained; such an approach is,
thus, inherently conservative. On the other hand, if
the interconnection could be transformed into a cas-
cade for an appropriate region in state space, the
analysis would be considerably simpler. Motivated by
this, in the rest of this manuscript we will focus on:
specifying a region in state-space where this system
behaves effectively as a cascade, with guaranteed
closed-loop asymptotic stability; and
characterizing the range of evolution of the states in
this region to obtain invariant sets that qualify as
regions of closed-loop stability.
The proposed approach will not enforce the control
input to be unconstrained and will, in addition, offer
the exibility of addressing trade-offs between the size
of the region of stability and the rate of convergence to
the desired steady state.
3. Local stabilization in the presence of constraints
For the rest of this manuscript, we will make the
following assumption on the eigenvalues of the matrix
A:
Assumption 2 The eigen6alues of the matrix A, are
negati6e real numbers, k
i
, i =1, , r.
Then (Kapoor & Daoutidis, 1997) the following
linear change of variables:
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 12

p
z
1
z
2

z
i

z
r

p
z
1
k
1
z
1
+z
2

(k
1
k
2
... k
i 1
)z
1
+

%
(i 1)
j =1
5
(i 1)
k=1,kj
k
k

z
2
+ ... +(k
1
+k
2
+ ... +k
i 1
)z
i 1
+z
i

(k
1
k
2
... k
r1
)z
1
+

%
(r1)
j =1
5
(r1)
k=1,kj
k
k

z
2
+ ... +(k
1
+k
2
+ ... +k
r1
)z
r1
+z
r

(9)
transforms the closed-loop system of Eq. (8) into the
following form:
p; =q(p, z )
z
;
1
=k
1
z
1
+z
2
z
;
2
=k
2
z
2
+z
3

z
;
r1
=k
r1
z
r1
+z
r
z
;
r
=F(p, z ) +L
g
L
f
r1
h(p, z )sat
F(p, z ) k
r
z
r
L
g
L
f
r1
h(p, z )

(10)
where F is, in general, a nonlinear function of the
states. Note that in the system of Eq. (10), the (r1)-
dimensional linear subsystem corresponding to the
states z
i
, 1i (r1) is exponentially stable (with
eigenvalues at k
1
, , k
r1
) and, hence, input-to-
state stable with respect to the forcing input z
r
. Further-
more, it is characterized by a feedback structure
(Kanellakopoulos, Kokotovic & Morse, 1991), with the
subsystem corresponding to each state z
i
, i =
1, , (r1) driven by the state z
i +1
. These facts allow
reducing the stabilization of the z subsystem to the
stabilization of the one-dimensional z
r
subsystem. Spe-
cically, as long as the state z
r
is driven to the origin
monotonically, the remaining z states will stay bounded
and decay to the origin.
Motivated by this, let us now focus on the stabiliza-
tion of the one-dimensional z
r
subsystem. To this end,
let Z denote the region in R
n
for which:
u
min
+
F(p, z )
L
g
L
f
r1
h(p, z )
u
max
(11)
where is a very small positive number. The term in
the denition of Z ensures that there is a nite input
force available to steer the z
r
subsystem to zero asymp-
totically (see proof of theorem 1 in the appendix). Since
the relative order is well-dened for all x-X, the sign of
the term L
g
L
r1
f
h(p, z ) is constant in the state-space of
operation, i.e. L
g
L
r1
f
h(p, z ) is invertible; hence, the
region Z is well-dened. We will assume, for ease of
exposition, that ZB. If this is not the case, the
region of interest is simply the intersection ZB.
Note that as long as the states of the system of Eq.
(10) evolve in Z,
the effect of the p subsystem on the z subsystem is
effectively cancelled and the system of Eq. (10) effec-
tively behaves as a cascade; this is because the effect
of the terms, F(p, z ) and L
g
L
r1
f
h(p, z ), through
which p drives the z subsystem, is effectively nullied
by the term F(p, z )/[L
g
L
f
r1
h(p, z )] in the z
r
sub-
system of Eq. (10);
the remaining component of the control law, i.e. the
term k
r
z
r
/[L
g
L
f
r1
h(p, z )], drives the state z
r
to the
origin.
The above arguments indicate that for (p, z ) -Z, the
origin of this system is asymptotically stable. This is
stated more precisely in the following theorem, the
proof of which is given in the appendix:
Theorem 1 Consider the system of Eq. (10) with
(p(0), z(0)) -Z. Then, if the solution of this system is
such that, t 0, (p(t), z(t)) -Z, then (p(t), z(t)) asymp-
totically decay to the origin.
It is important to note that the above region Z is not
the region in state space where the control law of Eq.
(5) is unconstrained. Instead, we allow the control law
to hit constraints throughout this analysis.
Remark 1 Note that the region Z depends on the rst
(r1) closed-loop poles, k
i
, 1i (r1), but not on
the last pole, k
r
. This implies that the rate of decay of the
state z
r
, in the absence of constraints, can be made as fast
as desired without affecting the size of the region of
stability. It furthermore implies that the location of the
closed-loop poles k
i
, 1i (r1) can be tuned in order
to enlarge the region Z.
We will now illustrate the above concepts through a
numerical example:
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 13
3.1. Example 1
Consider the system:
x;
1
=x
1
+x
2
x;
2
=x
3
x;
3
=x
1
2
+sat(u)
(12)
which is in the form of Eq. (4) with p=x
1
, z
1
=x
2
and
z
2
=x
3
. Consider now a feedback linearizing controller:
u=x
1
2
+q
1
x
2
+q
2
x
3
(13)
that globally asymptotically stabilizes the origin of the
closed-loop system with the closed-loop poles located at
( 1, k
1
, k
2
) in the absence of constraints. It can
be veried that the system of Eq. (12) under the control
law of Eq. (13) assumes the following form under the
change of variables x
2
=x
2
, x
3
=k
1
x
2
+x
3
(which is of
the form of Eq. (9)):
x;
1
=x
1
+x
2
x
;
2
=k
1
x
2
+x
3
x
;
3
=k
1
( k
1
x
2
+x
3
) +x
1
2
+sat( k
1
( k
1
x
2
+x
3
)
x
1
2
k
2
x
3
) (14)
Then, according to the previous analysis, as long as x
1
,
x
2
, x
3
evolve in Z, dened as the region where:
u
min
+k
1
( k
1
x
2
+x
3
) x
1
2
u
max
(15)
where is a small positive number, we are guaranteed
that the closed-loop system is asymptotically stable.
In the original coordinates, the region Z can be
described as:
u
min
+k
1
x
3
x
1
2
u
max
(16)
Inspecting the form of the region Z, it is obvious that
it is independent of the value of the pole k
2
, as
pointed out in remark 1. Moreover, as k
1
0, the
region Z assumes the form:
u
min
+x
1
2
u
max
(17)
which is exactly the region in state-space where the
destabilizing term x
2
1
can be cancelled by the control
law of Eq. (18). Thus, there is a trade-off between the
rate of convergence of the x
2
subsystem and the permis-
sible range of evolution of the states x
1
and x
3
. For
k
1
=k
2
=1, it can be seen that the control law of Eq.
(13), can be expressed as:
u=(x
3
+x
1
2
) x
2
x
3
(18)
which is a summation of the term (x
3
+x
2
1
) that
describes the region Z and cancels the destabilizing
term (x
3
+x
2
1
), and the stabilizing term (x
2
+x
3
). It
is clear that constraints are imposed only on the term
that is employed to cancel the destabilizing term (as is
expected); the full control law, on the other hand, is
allowed to hit constraints.
The previous analysis allows specifying a region in
state-space, where the closed-loop system behaves effec-
tively as a cascade, and where asymptotic stability can
be guaranteed in the presence of constraints, provided
that the states of the system remain in this region for all
times. This region, however, is not an invariant set by
construction; hence, it does not qualify as a region of
closed-loop stability. On the other hand, any invariant
set within Z would qualify as a region of closed-loop
stability for the system of Eq. (10). Motivated by this,
in the next section we will address the explicit construc-
tion of invariant sets within Z that qualify as such
regions of stability.
4. Closed-loop regions of stability
Referring to the closed-loop system of Eq. (10) (Eq.
(4) under the control law of Eq. (5)), the approach that
we will follow for estimating regions of closed-loop
stability within Z entails the following steps:
Initially, the feedback structure of the z subsystem
will be exploited, to characterize explicitly the range
of evolution of each individual z state, for a given
initial condition. This will provide a (relatively non-
conservative) bound on the evolution of the z states.
Given the bound on the evolution of the z states,
and the stability properties of the p subsystem (zero
dynamics), a bound on the evolution of the p states
will also be obtained, leading to an invariant set that
qualies as region of closed-loop stability as long as
it is contained in Z.
We begin with the following lemma (the proof of which
is presented in the appendix), which is a generalization
of a result for linear systems in (Kapoor & Daoutidis,
1997), and characterizes the largest range of values that
each state z
i
, 1i r, can assume with the constraints
of Eq. (11) being satised, for a given initial condition.
Lemma 1: Consider the system of Eq. (10). Let
{p
o
, z
1o
, ... , z
ro
}-Z denote an initial condition. Then if,
t 0, (p, z) -Z, then the following will hold t 0:
(i ) if z
ro
.0, then z
r
(t) -[0, z
ro
], and if z
ro
0, then
z
r
(t) -[z
ro
, 0]
(ii ) if i, z
io
are non-positi6e (non-negati6e), then z
i
will
also be non-positi6e (non-negati6e) for all i. Specically,
if z
i +1
(t) -[0, a], where a.0, then z
i
(t) -[0, max(z
io
, a/
k
i
)] and if z
i +1
(t) -[b, 0], where b0, then
z
i
(t) -[min(z
io
, b/k
i
), 0]
(iii ) if i, z
i
(t) -[a, b], where a0, b0, then z
(i
1)(t) -[a/k
i 1
, max(z
(i 1)o
,b/k
i 1
)], for z
(i 1)o
.0, and
z
(i 1)
(t) -[min(z
(i 1)o
,a/k
i 1
), b/k
i 1
], for z
(i 1)o
0
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 14
Lemma 1 allows us to specify, for a given initial
condition in Z, a set [a
i
, b
i
] where each state z
i
evolves
in the closed-loop system, provided that the state con-
straints of Eq. (11) (which depend on both the states p
and z ) are satised.
We will now focus on estimating a bound on the
range of values assumed by the p states, given an a
priori bound on the z states. To this end, dene sup
t 0
z (t) =z
sup
. Since the zero dynamics of the system of
Eq. (4) is locally exponentially stable, it possesses the
input-to-state stability property with respect to inputs z
that are small in magnitude. Exploiting this fact, we
will now concentrate on the computation of sup
t 0
p(t) =
p
sup
, with z
sup
known a priori.
Specically, given assumption 1, it follows from the
converse Lyapunov theorem that there exists a function
V(p): D
o
R that satises the following inequalities:
c
1
p
2
V(p) c
2
p
2
(V(p)
(p
q(p, 0) c
3
p
2
(19)
)(V(p)
(p
)
c
4
p,
where {p-D
o
:p r
o
} and c
1
, c
2
, c
3
, c
4
are positive
constants. It will be assumed that Z
(nr)
D
o
, where
Z
(nr)
is nr-dimensional subspace of Z correspond-
ing to the p-subsystem. Choosing V(p) as our Lya-
punov function candidate, the time-derivative of the
function V(p) can be expressed as:
V: c
3
p
2
+c
4
pq(p, z ) q(p, 0) (20)
where q(p, z ) has been written as:
q(p, z ) =q(p, 0) +[q(p, z ) q(p, 0)]
From the analyticity of the vector elds, it follows that
q(p, z ) is locally Lipschitz in (p, z ). Assuming that q(p,
z ) is locally Lipschitz (p, z ) -Z, we can express:
[q(p, z ) q(p, 0)] Lz , (21)
where L is a Lipschitz constant. This implies that
V: c
3
p
2
+c
4
Lpz (22)
or,
V: c
3
p(p c
4
/c
3
Lz ) (23)
For p .c
4
/c
3
Lz , we obtain
V: 0, p 0
provided p-Z
(nr)
. Thus, we can conclude that
p
sup
max

p(0),
'c
2
c
1
c
4
/c
3
Lz
sup

which follows from the rst inequality expression in Eq.


(19) (see Khalil, 1996 chapter 5 for more details). In the
above analysis, we exploited the fact that the term
q(p, z ) is locally Lipschitz in Z. Note, however, that if
(p, z ) -Z, M.0 such that
[q(p, z ) q(p, 0)] M (24)
then we have that
p
sup
max

p(0),
'c
2
c
1
c
4
/c
3
M

.
The above analysis provides an upper bound on the
norm of the states p for a given upper bound on the
norm of the states z, and allows constructing a hyper-
cuboidal set I of the form:
I:{z :a
i
z
i
b
i
, i =1, 2, ... , r, p:p
j
p
j
q
j
,
j =1, ... , (nr)} (25)
where (z
o
, p
o
) -I. This set qualies as a region of
closed-loop stability as long as it is contained in Z. The
above procedure can be directly employed to ascertain
whether an initial condition (z
o
, p
o
) -Z belongs to a
region of closed-loop stability, and, if yes, specify a
region of closed-loop stability. Note that if Z is a
simply-connected set, one could verify whether I-Z
by simply checking whether the points on the surfaces
of I satisfy Eq. (11).
Remark 2 Though the abo6e analysis was performed with
a general Lyapuno6 function candidate, V(p), assumption
1 guarantees that a standard quadratic function sufces
for this purpose. This allows us to preclude the task of
searching for an appropriate Lyapuno6 function candi -
date that satises the inequalities of Eq. (19).
Remark 3 It is assumed in the analysis that the term
q(p, z) is locally Lipschitz (p, z) -Z. In fact, it is suf-
cient that this term is Lipschitz only in I because the
states of the system of Eq. (10) will e6ol6e only in
IZ. Also, it is assumed in the abo6e analysis that
Z
(nr)
D
o
; instead, all we need to 6erify is that I
D
o
.
Remark 4 Throughout the abo6e analysis, it is assumed
that the zero dynamics is locally exponentially stable.
This was done purely for the sake of simplicity. In fact,
one can relax assumption 1 to the zero dynamics being
locally asymptotically stable. In this case, the zero dy-
namics will still be locally ISS and one will ha6e to
employ a different 6ersion of the con6erse Lyapuno6
theorem (Khalil, 1996) for the entire analysis to go
through.
We will now reconsider example 1, with the closed-
loop poles at s=1 each in the absence of constraints,
and employ the above analysis to obtain an estimate of
the region of closed-loop stability:
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 15
4.1. Example 1 (contd)
Consider the closed-loop system:
x;
1
=x
1
+x
2
x
;
2
=x
2
+x
3
x
;
3
=x
2
+x
3
+x
1
2
+sat( ( x
2
+x
3
+x
1
2
) x
3
)
(26)
with the initial condition (x
1
(0), x
2
(0), x
3
(0)) =
(x
10
, x
20
, x
30
) -Z, where x
10
.x
20
.x
20
+x
30
.0. Then,
it follows from lemma 1 that:
x
3
(t) x
30
=x
20
+x
30
, x
2
(t) x
20
=x
20
if x-Z, t 0. Note that the origin x
1
=0 of the zero
dynamics x;
1
=x
1
is globally exponentially stable. In
order to calculate x
1

sup
, note that the input driving the
forcing x
1
-dynamics is x
2
and that x
10
.x
20
. It can then
be inferred by mere observation that:
x
1
(t) x
10
provided x-Z, t 0. It then follows that the set:
I:0x
1
x
10
, 0x
2
x
20
, 0x
2
+x
3
x
20
+x
30
is a region of closed-loop stability as long as it belongs
to Z.
In the next section, we consider a polymerization
reactor system whose open-loop dynamics have been
studied in (Jaisinghani & Ray, 1984). This system was
also considered in (Kapoor & Daoutidis, 1999), where a
feedback linearizing controller with integral action was
designed such that the region of closed-loop stability in
the presence of constraints is limited only by the do-
main of attraction of the static feedback linearizing
controller. In what follows, we will address the estima-
tion of regions of closed-loop stability for this system
under a static feedback linearizing controller of the
form of Eq. (5).
5. Case study: polymerization reactor
Consider a CSTR where a free-radical methyl-
methacrylate polymerization reaction with AIBN
(azoisobutyronitrile) initiator is taking place. The feed
stream enters the reactor at ow rate q, temperature T
f
and monomer and initiator concentrations M
f
and I
f
,
respectively. The outlet stream leaves at temperature T
and the monomer and initiator exit concentrations are
M and I, respectively. There is a cooling jacket around
the reactor with the coolant temperature being T
c
.
Under standard modeling assumptions mentioned in
(Jaisinghani & Ray, 1984), the process dynamics is
described by the follouing set of equations:
dI
dt0
=
q
V
(I
f
I) k
d
I exp
E
d
RT

dT
dt0
=
q
V
(T
f
T) +
hA
c
(TT
c
)
lVC
p
+
( DH
p
)
lC
p
k
p
MP exp
E
p
RT

dM
dt0
=
q
V
(M
f
M) k
p
MP exp
E
p
RT

(27)
where P=

n=1
P
n
=( fk
d
I/k
t
)
1/2
, with P
n
being a grow-
ing polymer chain of length n. Here, k
d
, k
p
, k
t
denote
the reaction rate constants for the initiation step, the
propagation step and the termination steps, respec-
tively. E
d
, E
p
, E
t
denote the energies of activation for
the initiation step, the propagation step and the termi-
nation steps, respectively. The initiator efciency f is
assumed to be a constant. The reactor modeling equa-
tions can be put in dimensionless form by dening:
t =t0q/V, x
1
=M/M
f
, x
2
=
TT
f
T
f
E
p
RT
f

,
x
3
=I/M
f
, x
3f
=I/M
f
, w=P/M
f
k
p
=
E
p
RT
f
, k
t
=
E
t
E
p
, k
d
=
E
d
E
p
, B=
( DH
p
)M
f
lC
p
T
f
,
i=
hA
c
lC
p
q
, x
2c
=
T
c
T
f
T
f
E
p
RT
f

, p=
k
d
e
k
d
k
p
k
p
e
k
p
M
f
,
Da
p
=k
p
e
k
p
VM
f
/q,
Da
d
=k
d
e
k
p
k
d
V/q=pDa
p
, Da
t
=k
to
e
k
p
k
t
VM
f
/q
The resulting equations then take the form:
dx
1
dt
=1x
1
Da
p
wx
1
exp
x
2
1+x
2
/k
p

dx
2
dt
=x
2
+BDa
p
wx
1
exp
x
2
1+x
2
/k
p

ix
2
+Q (28)
dx
3
dt
=x
3f
x
3
Da
d
x
3
exp
k
d
x
2
1+x
2
/k
p

where w=[(2fDa
d
x
3
)/(Da
t
G(x))]
1/2
exp{[(k
d
k
t
)x
2
]/
[2(1+x
2
/k
p
)]} and Q=bx
2c
is the dimensionless heat
input to the reactor, which naturally has bounds on it.
The term G(x) is a nonlinear function of the monomer
conversion and the reactor temperature and it captures
the gel effect or the effect of viscosity on the kinetic rate
constants (see (Jaisinghani & Ray, 1984) for a detailed
expression). The state-space of interest is X:
(x
1
, x
2
, x
3
) -{[0, 1] [ 3, 3] [0, x
3f
]}. Note that
Da
p
.0, Da
d
.0, x
3f
.0 and w(x) .0, x-X. The
values of the process parameters (taken from Jaising-
hani & Ray, 1984) and the corresponding steady-state
values of the process variables are given in Table 1 and
they correspond to an unstable equilibrium point. The
open-loop process exhibits three steady states, two of
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 16
which are stable and one unstable for the range of the
manipulated input under consideration. It is desirable
to control the process at the unstable steady state due
to economic as well as practical reasons. The desired
steady state is x
sp
=(x
1
sp
, x
2
sp
, x
3
sp
) =(0.18, 1.42,
0.00422). We dene u=QQ
nom
, where Q
nom
=0.122
is the nominal value of the dimensionless heat input.
The input limits imposed are u
max
=0.058 and u
min
=
0.122. We will express the states and the input vari-
able in deviation from their nal steady-state values. To
this end, we dene x =xx
sp
, which ensures that the
nal equilibrium point corresponds to (x , u) =(0, 0).
The system of Eq. (28) now takes the following form:
dx
1
dt
=1x
1
sp
x
1
Da
p
w(x , x
sp
)(x
1
sp
+x
1
)exp
(x
2
+x
2
sp
)
1+(x
2
+x
2
sp
)/k
p

dx
2
dt
=x
2
sp
x
2
+BDa
p
w(x , x
sp
)(x
1
sp
+x
1
)exp
(x
2
+x
2
sp
)
1+(x
2
+x
2
sp
)/k
p

i(x
2
+x
2
sp
) +Q
nom
+u (29)
dx
3
dt
=x
3f
x
3
sp
x
3
Da
d
(x
3
sp
+x
3
)exp
k
d
(x
2
+x
2
sp
)
1+(x +x
2
sp
)/k
p

We will consider the reactor temperature T (x


2
in
deviation variables) as the output and will employ the
proposed analysis approach for the stabilization of the
system of Eq. (29) and the estimation of regions of
closed-loop stability. The relative order of this output
with respect to the manipulated input is 2. Thus, the
coordinate transformation:

p
z
1
z
2

=(x) =

x
3
h(x )
L
f
h(x )

(30)
which is well-dened for all x-X, transforms the system
of Eq. (29) into the following form:
p; =q(p, z
1
, z
2
)
z;
1
=z
2
z;
2
=F
1
(p, z) +L
g
L
f
h(p, z)sat(u)
(31)
where
z
1
=x
1
, z
2
=x
;
1
, q(p, z
1
, z
2
) =x
;
3
and F
1
(p, z) and L
g
L
f
(p, z) are smooth nonlinear
functions; it was veried that the sign of the term L
g
L
f
(p, z) is constant for x-X. The following partially
linearizing control law:
u=
F
1
(p, z) +q
1
z
1
+q
2
z
2
L
g
L
f
h(p, z)
(32)
was employed. Since Da
p
.0, Da
d
.0, x
3f
.0 and
w(x) .0, x-X, it can be inferred by analyzing the
system of Eq. (29) under the control law of Eq. (32)
that for (x
1
, x
2
) -Z, x
;
3
0 at x
3
=x
3f
x
3
sp
and x
;
3
.0
at x
3
=x
3
sp
. Hence, x
3
(t) -[ x
3
sp
, x
3f
x
3
sp
], t 0.
Also, the origin x
3
=0 of the unforced zero dynamics is
globally exponentially stable. It then follows that the
closed-loop system can be globally asymptotically stabi-
lized in the absence of constraints.
Two different choices of (q
1
, q
2
) were tried out to
study the effect of the location of the closed-loop poles
on the size of the region of stability. Initially, (q
1
, q
2
)
was chosen as (1, 2) which places the closed-loop poles
at s=1, 1 in the absence of constraints. Under a
change of variables of the form of Eq. (9), with k
1
=1
and k
2
=1, the closed-loop system assumes the follow-
ing form:
p; =q(p, z
1
, z
2
)
dz
1
dt
=z
1
+z
2
dz
2
dt
=F
2
(p, z ) +L
g
L
f
h(p, z )sat

F
2
(p, z ) +z
2
L
g
L
f
h(p, z )

(33)
where z
1
=z
1
, z
2
=z
1
+z
2
and F
2
(p, z ) and L
g
L
f
h(p, z )
are smooth nonlinear functions. The set Z is the state-
space region for which
u
min
+
F
2
(p, z )
L
g
L
f
h(p, z )
u
max
(34)
Table 1
Process parameters for the case study
k
p
=5.1E6 l mol
1
s
1
l mol
1
s
1
k
t
=7.1E8
k
d
=1.58E15 s
1
E
p
=6.3 kcal mol
1
E
t
=2.8 kcal mol
1
E
d
=30.8 kcal mol
1
B=12
i=6
x
10
=0.16
x
20
=1.46
x
30
=0.004217
M
f
=11.78 mol l
1
I
f
=0.05 mol l
1
T
c
=295 K
T
f
=295 K
Q
nom
=0.122
y
sp
=0.1805
Da
p
=6.60115E5
f =0.6
R=1.987E3 kcal mol
1
K
1
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 17
Fig. 1. Projections of the sets Z and I in the (z
1
, z
2
) space with (q
1
, q
2
) =(1, 2).
where was chosen as 0.000001. The initial condition:
(x
1o
, x
2o
, x
3o
) =(0.16, 1.446, 0.004217) which corre-
sponds to
(p
o
, z
1o
, z
2o
) =( 0.00003, 0.02, 0.024)
and belongs to the set Z was chosen. It then follows
from lemma 1 that as long as (p, z ) -Z, the states z
evolve in the set
A:{0.024z
1
0, 0.024z
2
0}
The zero dynamics or the x
3
-subsystem can be ex-
pressed as:
dx
3
dt
=x
3

1+Da
d
exp
k
d
x
2
sp
1+x
2
sp
/k
p

Da
d
(x
3sp
+x
3
)

exp
k
d
(x
2
+x
2sp
)
1+(x
2
+x
2sp
)/k
p

exp
k
d
x
2sp
1+x
2sp
/k
p

(35)
Given the exponential stability of the origin of the
unforced zero dynamics, there exists a quadratic Lya-
punov function candidate V=
1
2
x
2
3
which satises Eq.
(19), with D
o
=R
n
, c
1
=c
2
=1, c
3
=1+Da
d
exp(k
d
x
2sp
)/
(1+x
2sp
/k
p
) and c
4
=1. Note that:
q(p, z ) q(p, 0)
=
)
Da
d
(x
3sp
+x
3
)

exp
k
d
(x
2
+x
2sp
)
1+(x
2
+x
2sp
)/k
p

exp
k
d
x
2sp
1+x
2sp
/k
p
)
(36)
Da
d
x
3f
)
exp
k
d
(x
2
+x
2sp
)
1+(x
2
+x
2sp
)/k
p

exp
k
d
x
2sp
1+x
2sp
/k
p
)
M
Clearly then, p
sup
max(x
30
,
M
c
3
z
sup
). It was veried
that max(x
30
,
M
c
3
z
sup
) =x
30
. Thus, the set I has the
following form:
I:{0.0024z
1
0.0, 0.024z
2
0, 0.00003
p0.00003}
The term [F
2
(p, z )]/[L
g
L
f
h(p, z )] was computed along
the surfaces of I and was found to assume values
between u
min
and u
max
, thereby guaranteeing that I is
indeed a region of closed-loop stability. Projections of
the sets Z and I in the (z
1
, z
2
) space are shown in Fig.
1. A simulation run was performed starting from the
chosen initial condition, which belongs in I, under the
control law of Eq. (32). Fig. 2 shows the state and the
manipulated input proles for this run. It can be seen
that the states decay smoothly to the set-point whereas
the manipulated input stays saturated at u
min
for a brief
initial period of time before stabilizing to its nominal
value. These proles clearly reect the relative non-con-
servativeness of the proposed methodology, i.e. its abil-
ity to guarantee closed-loop stability even when the
control law is saturated.
The feedback linearizing controller was modied by
changing (q
1
, q
2
) to (100, 20) which places the closed-
loop poles at s=10, 10 in the absence of con-
straints, thereby increasing the rate of decay of the
states. Under a change of variables of the form of
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 18
Eq. (9), with k
1
=10 and k
2
=10, the closed-loop sys-
tem assumes the following form:
p; =q(p, z
1
, z
2
)
dz
1
dt
=10z
1
+z
2
dz
2
dt
=F
2
(p, z ) +L
g
L
f
h(p, z )sat

F
2
(p, z ) +10z
2
L
g
L
f
h(p, z )

(37)
where z
1
=z
1
, z
2
=10z
1
+z
2
. Fig. 3 shows projections of
the sets Z and I in the (z
1
, z
2
) space for the same
initial condition. Note that the initial condition as well
as the set I are no longer contained in Z which implies
that the latter may not be a region of closed-loop
stability. This was indeed veried by carrying out a
simulation run with the same initial condition as before
but under the faster controller. Fig. 4 shows the state
and the input proles and, as can be seen, the closed-
Fig. 2. State and manipulated input proles with (q
1
, q
2
) =(1, 2).
Fig. 3. Projections of the sets Z and I in the (z
1
, z
2
) space with (q
1
, q
2
) =(100, 20).
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 19
Fig. 4. State and manipulated input proles with (q
1
, q
2
) =(100, 20).
loop system is unstable and the states are driven to a
different set-point which is undesirable.
6. Conclusions
In this work, we presented a framework for stability
analysis of nonlinear systems subject to control laws that
are locally stabilizing in the absence of constraints.
Central to this analysis framework is a change of
variables which allows: (i) characterizing a state space
region where stabilization of a one-dimensional subsys-
tem implies closed-loop asymptotic stability; and (ii)
constructing hypercuboidal invariant sets in this region
that qualify as regions of closed-loop stability. The
desirable features of the proposed approach (relative
simplicity, nonconservativeness and exibility in address-
ing trade-offs between stability and performance) were
illustrated through a case study on a polymerization
reactor.
Acknowledgements
Financial support from the National Science Founda-
tion, CTS-9624725 is gratefully acknowledged.
Appendix A
A.1. Proof of theorem 1
Consider the system of Eq. (10). Also consider the
Lyapunov function candidate:
V=
1
2
z
r
2
(38)
whose time derivative along the trajectories of the
z
r
-subsystem yields:
V: =z
r

F(p, z ) +L
g
L
f
r1
h(p, z )sat

F(p, z ) +k
r
z
r
L
g
L
f
r1
h(p, z )

(39)
Given that (p, z ) -Z, t 0, it is clear that V: is of the
form:
V: (t) sz
r
2
(40)
where s= max
Z
)L
g
L
f
r1
h(p, z )
z
r
)
. Hence, V: 0, z
r
0.
Also, V: 0, z
r
=0. Hence, it follows from Lasalles
theorem that z
r
decays asymptotically to the origin. As
z
r
0, the z
r1
-subsystem takes the form:
z
;
r
=k
r1
z
r
(41)
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 20
which then decays asymptotically to the origin. Arguing
similarly for the subsystems z
i
, i =1, , (r2) we can
infer that the z-subsystem decays to the origin. With
z 0, the p-subsystem has the form:
p; =q(p, 0) (42)
the origin of which is asymptotically stable. Hence, the
proof.
A.2. Proof of lemma 1
(i) Since (p, z ) -Z, t 0, it can be inferred from the
proof of theorem 1 that z
r
decays monotonically to the
origin. Hence, the result follows.
(ii) Without loss of generality, let the initial condition
z
o
be non-negative. Considering the subsystem corre-
sponding to the state z
r1
, since the forcing input
z
r
(t) -[0, z
ro
], it follows that if z
ro
k
r1
z
(r1)o
, then
0z
r1
(t) z
(r1)o
and if z
ro
.k
r1
z
(r1)o
, then 0
z
r1
(t) z
ro
/k
r1
. Thus at all times, z
r1
(t)
-[0, max(z
(r1)o
, z
ro
/k
r1
)]. Arguing in a similar fashion
for each the states z
i
, i =(r2), , 1, we obtain the
stated result.
(iii) Let z
ro
.0 and z
(r1)o
0 In the event of z
r
decaying to zero faster than z
r1
, the latter lies in the
range [z
(r1)o
, 0]. In the case when z
r1
converges to
zero at a faster rate than z
r
, z
r1
may change sign and
assume a maximum value z
ro
/k
r1
, after which it will
start decreasing again. Thus, z
r1
(t) -[z
(r1)o
, z
ro
/k
r1
].
Now, considering the subsystem corresponding to z
r2
,
if z
(r2)o
.0 then z
r2
(t) -[z
(r1)o
/k
r2
, max(z
(r2)o
,
z
ro
/(k
r1
k
r2
))]. On the other hand if z
(r2)o
0, then
z
r2
-[min(z
(r1)o
/k
r2
, z
(r2)o
), z
ro
/(k
r1
k
r2
)]. Thus,
the stated result follows by arguing similarly for thepre-
ceding states.
Appendix B. Notation
Roman letters
vector eld f
g vector eld associated with u
scalar eld h
r relative order of h with respect to u
time t
u manipulated input
manipulated input in deviation variables 6
state vector x
inputoutput state variables z
inputoutput state variables in transformed z
coordinates
y output
A matrix
Lyapunov function candidate V
an invariant set I
Z region in state space that contains the esti-
mated region of stability
Z
0
set of initial conditions
Greek letters
zero-dynamics state variables p
Math symbols
saturation nonlinearity sat(.)
R
n
n-dimensional Euclidean space
transpose T
L
k
f
h
I
kth order Lie Derivative of h
i
along f
for all
Euclidean norm of vectors and the induced =
norm of matrices
supremum of the Euclidean norm of a sup
t0
vector
References
A, stro m, K. J. & Rundqwist, L. (1989). Integrator windup and how to
avoid it. Proceedings of the American Control Conference (p.
1693).
Campo, P. J., & Morari, M. (1990). Robust control of processes
subject to saturation nonlinearities. Computers & Chemical Engi -
neering, 14, 343.
Coulibaly, C., Maiti, S., & Brosilow, C. (1995). Internal model
predictive control. Automatica, 31, 1471.
Fuller, A. T. (1969). In the large stability of relay and saturated
control systems with linear controllers. International Journal of
Control, 10, 457.
Gilbert, E. G., & Tan, K. T. (1991). Linear systems with state and
control constraints: the theory and application of maximal ouput
admissible sets. IEEE Transactions on Automatic Control, 36,
1008.
Glattfelder, A. H., Eck, C., & Schaufelberger, W. (1995). Stability
analysis of two different generalized antiwindup systems. Proceed-
ings of the Third European Control Conference (p. 1492).
Gutman, P. O., & Hagander, P. (1985). A new design of constrained
controllers for linear systems. IEEE Transactions on Automated
Control, 30, 22.
Hanus, R. & Kinnaert, M. (1989). Control of constrained multivari-
ables systems using the conditioning technique. Proceedings of
The American Control Conference (p. 1711).
Isidori, A. (1995). Nonlinear control systems. Springer-Verlag.
Jaisinghani, R., & Ray, W. H. (1984). On the dynamic behavior of a
class of homogeneous continuous stirred tank polymerization
reactors. Chemical Engineering Science, 39, 1741.
Kanellakopoulos, I., Kokotovic, P., & Morse, A. S. (1991). System-
atic design of adaptive controllers for feedback linearizable sys-
tems. IEEE Transactions on Automated Control, 36, 1241.
Kapoor, N., & Daoutidis, P. (1997). Stabilization of systems with
input constraints. International Journal of Control, 66, 653.
Kapoor, N., & Daoutidis, P. (1999). An observer-based anti-windup
scheme for nonlinear systems with input constraints. International
Journal of Control, 72, 18.
Kapoor, N., Teel, A. R., & Daoutidis, P. (1998). An anti-windup
design for linear systems with input saturation. Automatica, 34,
559.
Kendi, T. A., & Doyle, F. J. (1997). An anti-windup scheme for
multivariable nonlinear systems. Journal of Process Control, 7,
329.
N. Kapoor, P. Daoutidis / Computers and Chemical Engineering 24 (2000) 921 21
Khalil, H. K. (1996). Nonlinear systems. Prentice-Hall.
Kothare, M. V., Campo, P. J., Morari, M., & Nett, C. N. (1994). A
unied framework for the study of anti-windup designs. Automat-
ica, 30, 1869.
Lee, J. H. & Cooley, B. (1997). Recent advances in model predictive
control and other related areas. In J. C. Kantor, C. E. Garcia &
B. Carnahan, AIChE Symposium Series No. 316, Chemical Process
Control -V, 93 (p. 201).
Lin, W. (1994). Input saturation and global stabilization by output
feedback for afne systems. Proceedings of the Thirty-Third Con-
ference on Dec. and Control (p. 1323).
Lin, Y., Sontag, E. D., & Wang Y. (1993). Recent results on
lyapunovtheoretic techniques for nonlinear stability. Technical
Report SYCON-93-09.
Maine, D. Q. (1997). Nonlinear model predictive control: an assess-
ment. In J. C. Kantor, C. E. Garcia & B. Carnahan, AIChE
Symposium Series No. 316, Chemical Process Control -V, 93 (p.
217).
Oliveira, S. L. & Morari M. (1994). Robust model predictive control
for nonlinear systems. Proceedings of the Thirty-Third Conference
on Dec. and Control (p. 3561).
Pappas, G. J., Lygeros J., & Godbole D. T. (1995). Stabilization and
tracking of feedback linearizbale systems under input constraints.
Proceedings of the Thirty-Fourth Conference on Dec. and Control
(p. 596).
Scokaert, P. O. M., & Rawlings J. B. (1995). Stability of model
predictive control under perturbations. Preperation of Nonlinear
Control Systems Design Symposium (p. 21).
Sussmann, H. J., Sontag, E. D., & Yang, Y. (1994). A general result
on the stabilization of linear systems using bounded control.
IEEE Transactions on Automated Control, 39, 2411.
Teel, A. R. (1996). A nonlinear small gain theorem for the analysis of
control systems with saturation. IEEE Transactions on Automated
Control, 41, 1256.
Teel, A. R., & Praly, L. (1995). Tools for semi-global stabilization by
partial state and output feedback. SIAM Journal of Control and
Optimization, 33, 1443.
Valluri, S., & Soroush, M. (1998). Analytical control of siso nonlinear
processes with input constraints. American Institute of Chemical
Engineers J, 44, 116.
Valluri, S., Soroush, M., & Nikravesh, M. (1998). Shortest-predic-
tion-horizon nonlinear model predictive control. Chemical Engi -
neering Science, 53, 273.
Zheng, A. & Morari, M. (1995). Control of linear unstable systems
with constraints. Proceedings of American Control Conference (p.
3704).
.

S-ar putea să vă placă și