Computers and Chemical Engineering 26 (2002) 187– 203

www.elsevier.com/locate/compchemeng

Integrated optimal actuator/sensor placement and robust control

of uncertain transport-reaction processes
Charalambos Antoniades, Panagiotis D. Christofides *
Department of Chemical Engineering, Uni6ersity of California, 405 Hilgard A6enue, Box 951592, Los Angeles, CA 90095 -1592, USA

Received 17 October 2000; received in revised form 6 September 2001; accepted 6 September 2001

Abstract

This paper focuses on transport-reaction processes with unknown time-varying parameters and disturbances described by
quasi-linear parabolic PDE systems, and addresses the problem of computing optimal actuator/sensor locations for robust
nonlinear controllers. Initially, Galerkin’s method is employed to derive finite-dimensional approximations of the PDE system
which are used for the synthesis of robust nonlinear state feedback controllers via geometric and Lyapunov techniques and the
computation of optimal actuator locations. The controllers enforce boundedness and uncertainty attenuation in the closed-loop
system. The optimal actuator location problem is subsequently formulated as the one of minimizing a meaningful cost functional
that includes penalty on the response of the closed-loop system and the control action. Owing to the boundedness of the state,
the cost is defined over a finite-time interval (the final time is defined as the time needed for the process state to become smaller
than the desired uncertainty attenuation limit), while the optimization is performed over a broad set of initial conditions and
time-varying disturbance profiles. Subsequently, under the assumption that the number of measurement sensors is equal to the
number of slow modes, we employ a standard procedure for obtaining estimates for the states of the approximate finite-dimen-
sional model from the measurements. The optimal location of the measurement sensors is computed by minimizing a cost function
of the estimation error in the closed-loop infinite-dimensional system. We show that the use of these estimates in the robust state
feedback controller leads to a robust output feedback controller, which guarantees boundedness of the state and uncertainty
attenuation in the infinite-dimensional closed-loop system, provided that the separation between the slow and the fast eigenvalues
is sufficiently large. We also establish that the solution to the optimal actuator/sensor problem, which is obtained on the basis of
the closed-loop finite-dimensional system, is near-optimal in the sense that it approaches the optimal solution for the
infinite-dimensional system as the separation between the slow and fast eigenvalues increases. The theoretical results are
successfully applied to a typical diffusion-reaction process with nonlinearites and uncertainty to design a robust nonlinear output
feedback controller and compute the optimal actuator/sensor locations for robust stabilization of an unstable steady state. © 2002
Elsevier Science Ltd. All rights reserved.

Keywords: Robust nonlinear control; Static output feedback control; Optimal actuator/sensor placement; Diffusion– convection-reaction processes

1. Introduction being nonlinear, parabolic PDE systems which model

diffusion –convection-reaction processes are uncertain
Transport-reaction processes with significant diffu- due to the presence of unknown or partially known
sive and dispersive mechanisms are typically character- process parameters and disturbances. Time-varying un-
ized by severe nonlinearities and spatial variations, and certain variables, if they are not appropriately ac-
are naturally described by quasi-linear parabolic partial counted for in the controller design, may cause
differential equations (PDEs). Examples include tubular deterioration of the nominal closed-loop performance,
and packed-bed reactors, as well as chemical vapor and even lead to closed-loop instability.
deposition and crystal growth processes. In addition to Parabolic PDE systems typically involve spatial dif-
ferential operators whose eigenspectrum can be parti-
* Corresponding author. Tel.: + 1-310-794-1015; fax: +1-310-206-
tioned into a finite-dimensional slow one and an
4107. infinite-dimensional stable fast complement (Friedman,
E-mail address: [email protected] (P.D. Christofides). 1976; Balas, 1979; Ray, 1981; Christofides, 2001). This

0098-1354/02/$ - see front matter © 2002 Elsevier Science Ltd. All rights reserved.
PII: S0098-1354(01)00740-2
188 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203

implies that the dynamic behavior of such systems can positions and times employing a state observer in the
be approximately described by finite-dimensional sys- presence of measurement noise. Early efforts for the
tems. Therefore, the standard approach to the control solution to this problem focused on linear PDE systems
of parabolic PDEs involves the application of (Yu & Seinfeld, 1973; Chen & Seinfeld, 1975; Kumar &
Galerkin’s method to the PDE system to derive ODE Seinfeld, 1978a; Omatu, Koide, & Soeda, 1978; Morari
systems that describe the dynamics of the dominant & O’Dowd, 1980) and the application of the results to
(slow) modes of the PDE system, which are subse- optimal state estimation in tubular reactors (Colan-
quently used as the basis for the synthesis of finite-di- tuoni & Padmanabhan, 1977; Kumar & Seinfeld,
mensional controllers (Balas, 1979; Ray, 1981). A 1978b; Harris, MacGregor, & Wright, 1980; Alvarez,
potential drawback of this approach is that the number Romagnoli, & Stefanopoulos, 1981; Waldraff,
of modes that should be retained to derive an ODE Dochain, Bourrel, & Magnus, 1998). The central idea
system that yields the desired degree of approximation to the solution involves the use of a spatial discretiza-
may be very large, leading to complex controller design tion scheme to obtain a lumped approximation of the
and high dimensionality of the resulting controllers. original distributed parameter system followed by the
Motivated by this, research efforts on control of formulation and solution of an optimal state estimation
parabolic PDE systems over the last decade have fo- problem which involves computing sensor locations so
cused on the problem of synthesizing low-order con- that an appropriate functional that includes penalty on
trollers on the basis of ODE models obtained through the estimation error and the measurement noise is
combination of Galerkin’s method with approximate minimized. Significant research efforts have also been
inertial manifolds (see the recent book of Christofides, made on the problem of optimal placement of con-
2001 for details and references). More recently, Lya- trollers and sensors for various class of linear dis-
punov-based control methods for quasi-linear parabolic tributed parameter systems (see Amouroux, Di Pillo, &
PDE systems with time-varying uncertain variables Grippo, 1976; Malandrakis, 1979; Omatu & Seinfeld,
have been developed which lead to the synthesis of state 1983 and the review paper Kubrusly & Malebranche,
(Christofides, 1998) and output (Christofides & Baker, 1985).
1999) feedback controllers that achieve closed-loop sta- Despite the progress on optimal sensor placement
bility and uncertainty attenuation. and the availability of results on the integration of
Even though the developed methods allow to system- linear feedback control with actuator placement for
atically design practically-implementable nonlinear linear parabolic PDEs, there are no results on the
feedback controllers for transport-reaction processes integration of nonlinear control with optimal placement
using parabolic PDE systems, there is no work on the of control actuators and measurement sensors for
integration of nonlinear controller design with optimal transport-reaction processes described by nonlinear
placement of control actuators and measurement sen- parabolic PDEs. Motivated by this, we have initiated a
sors for transport-reaction processes so that the desired line of work on the computation of optimal actuator/
control objectives are achieved with minimal energy sensor locations for nonlinear controllers in transport-
use. Regarding the problem of optimal placement of reaction processes. In previous work (Antoniades &
control actuators, the conventional approach is to se- Christofides, 2000, 2001), we proposed a general and
lect the actuator locations based on open-loop consid- practical methodology for the integration of nonlinear
erations to ensure that the necessary controllability output feedback control with optimal actuator/sensor
requirements for stabilization of the closed-loop system placement for transport-reaction processes described by
are satisfied. More recently, efforts have been made on a broad class of quasi-linear parabolic PDEs. Under
the problem of integrating feedback control and opti- the hypothesis that the process model is perfectly
mal actuator placement for certain classes of linear known, our approach leads to actuator/sensor locations
distributed parameter systems including investigation of which allow enforcing the desired performance objec-
controllability measures and actuator placement in os- tives in the closed-loop system with minimal control
cillatory systems (Arbel, 1981), optimal placement of energy use.
actuators for linear feedback control of parabolic PDEs In this paper, we consider transport-reaction pro-
(Xu, Warnitchai, & Igusa, 1994; Demetriou, 1999) and cesses with unknown time-varying parameters and dis-
of actively controlled structures (Rao, Pan, & turbances modeled by quasi-linear parabolic PDE
Venkayya, 1991; Choe & Baruh, 1992). systems, and address the problem of computing optimal
On the other hand, the problem of selecting optimal actuator/sensor locations for robust nonlinear con-
locations for measurement sensors in linear distributed trollers. Initially, Galerkin’s method is employed to
parameter systems has received very significant atten- derive finite-dimensional approximations of the PDE
tion over the last 30 years. The essence of this problem system which are used for the synthesis of robust
is to use a finite number of measurements to compute nonlinear state feedback controllers via geometric and
the best estimate of the entire distributed state for all Lyapunov techniques and the computation of optimal
 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203 189

(x̄
actuator locations. The controllers enforce boundedness C1x̄(h, t)+ D1 (h, t)= R1 (2)
(z
and uncertainty attenuation in the closed-loop system.
The optimal actuator location problem is subsequently (x̄
C2x̄(i, t)+ D2 (i, t)=R2
formulated as the one of minimizing a meaningful cost (z
functional that includes penalty on the response of the
closed-loop system and the control action. Owing to the and the initial condition:
boundedness of the state, the cost is defined over a x̄(z, 0)=x̄0(z) (3)
finite-time interval (the final time is defined as the time
needed for the process state to become smaller than the where x̄(z, t)= [x̄1(z, t) x̄2(z, t) … x̄n (z, t)]T denotes
desired uncertainty attenuation limit), while the opti- the vector of state variables, [h, i]¦ R is the domain
mization is performed over a broad set of initial condi- of definition of the process, z[h, i] is the spatial
tions and time-varying disturbance profiles. Sub- coordinate, t[0, ) is the time, u(t)=[u 1(t)
sequently, under the assumption that the number of u 2(t) … u l(t)]T Rl denotes the vector of manipulated
measurement sensors is equal to the number of slow inputs, q(t)= [q1(t) q2(t) … qq (t)]Rq denotes the vec-
modes, we employ a standard procedure for obtaining tor of uncertain variables, which may include uncertain
estimates for the states of the approximate finite-dimen- process parameters or exogenous disturbances, y ic R
sional model from the measurements. The optimal loca- denotes a controlled output, y sm R denotes a measured
tion of the measurement sensors is computed by output, (x̄/(z, ( 2x̄/(z 2 denote the first- and second-or-
minimizing a cost function of the estimation error in der spatial derivatives of x̄, f(x̄), W(x̄, r(z)q(t)) are vec-
the closed-loop infinite-dimensional system. We show tor functions, w, k, µ are vectors, A, B, C1, D1, C2, D2
that the use of these estimates in the robust state are matrices, R1, R2 are column vectors, and x̄0(z) is the
feedback controller leads to a robust output feedback initial condition.
controller, which guarantees boundedness of the state The properties and role of the functions
and uncertainty attenuation in the infinite-dimensional b(z), r(z), c i(z), s i(z) are as follows: b(z) is a known
closed-loop system, provided that the separation be- smooth vector function of z of the form b(z)=
tween the slow and the fast eigenvalues is sufficiently [b 1(z) b 2(z) · · · b l(z)], where b i(z) describes how the
large. We also establish that the solution to the optimal control action u i(t) is distributed in the interval [h, i]
actuator/sensor problem, which is obtained on the basis (e.g. point/distributed actuation), r(z)= [r1(z) · · · rq (z)],
of the closed-loop finite-dimensional system, is near-op- where rk (z) is a known smooth function of z which
timal in the sense that it approaches the optimal solu- specifies the position of action of the uncertain variable
tion for the infinite-dimensional system as the qk on [h, i], c i(z) is a known smooth function of z
separation between the slow and fast eigenvalues in- which is determined by the specification of the i-th
creases. The theoretical results are successfully applied controlled output in the interval [h, i], and s i(z) is a
to a typical diffusion-reaction process with nonlinear- known smooth function of z which is determined by the
ites and uncertainty to design a robust nonlinear output location and type of the measurement sensors (e.g.
feedback controller and compute the optimal actuator/ point/distributed sensing). Throughout the paper, O(m)
sensor locations for robust stabilization of an unstable denotes the order of magnitude notation (i.e. l(m)=
steady state. O(m) if there exist positive real numbers k1 and k2 such
that: l(m)5 k1m, Öm Bk2), · denotes the standard
Euclidean norm, and q denotes ess.sup.q(t), t]0
2. Preliminaries for any measurable function q:R ] 0 “ Rq. Moreover, we
will use the Lie derivative notation: Lfh denotes the Lie
2.1. Description of parabolic PDE systems with derivative of a scalar field h with respect to the vector
uncertainty field f, L kf h denotes the k-th order Lie derivative and
LgL kf − 1h denotes the mixed Lie derivative. Finally a
’
We consider quasi-linear parabolic PDE systems with function l ( ) is said to be of class.
uncertain variables of the form: Defining the state x in the infinite dimensional
(x̄ (x̄ ( 2x̄ Hilbert space H([h, i], Rn) (with H being the space of
= A +B 2 + wb(z)u +f(x̄)+W(x̄, r(z)q(t)) n-dimensional vector functions defined on [h, i] that
(t (z (z

&
(1) satisfy the boundary condition of Eq. (2) with inner
product and norm:
&
i
y ic = c i(z)kx̄ dz, i =1, …, l

&
i 1
h
i
(…1, …2)= (…1(z), …2(z))Rndz, …1 2 = (…1, …2)2
h
y sm = s s(z)µx̄ dz, s =1, …, p
h where …1, …2 are two elements of H and the notation
subject to the boundary conditions: (·,·)Rn denotes the standard inner product in Rn), as:
190 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203

x(t) = x̄(z, t), t\0, z  [h, i], (4) Remark 2. Referring to Assumption 1, the hypothesis
of finite number of unstable eigenvalues and discrete
and the operators as: eigenspectrum are always satisfied for parabolic PDE
systems defined in finite spatial domains, while the
(x̄ ( 2x̄
Ax = A + B 2, (5) assumption of existence of only a few dominant modes
(z (z that describe the dynamics of the parabolic PDE system
x  D(A) is usually satisfied by the majority of diffusion– convec-

!
= xH; C1x̄(h, t)+ D1
(x̄
(h, t)
tion-reaction processes (see the catalytic rod example of
Section 6).
(z

= R1, C2x̄(i, t)+D2

(x̄
(i, t) =R2
" 2.2. Galerkin’s method
(z
We derive an m-dimensional approximation of the
Bu =wbu, Cx =(c, kx), Sx = (s, wx)
system of Eq. (6) using Galerkin’s method. Letting Hs,
where c =[c 1 c 2 … c l], the system of Eqs. (1)–(3)takes Hf be two subspaces of H, defined as Hs =
the form: span{ƒ1, ƒ2, …, ƒm } and Hf = span{ƒm + 1, ƒm + 2, …},
and defining the orthogonal projection operators Ps
x; = Ax +Bu + f(x)+ W(x, q), x(0) = x0 (6) and Pf such that xs = Psx, xf = Pfx, the state x of the
system of Eq. (6) can be decomposed as:
yc = Cx, ym =Sx
x= xs + xf = Psx+ Pf x (8)
where f(x(t))=f(x̄(z, t)), W(x(t), q)= W(x̄, rq) and
x0 = x̄0(z). Applying Ps and Pf to the system of Eq. (6) and using
For A, the eigenvalue problem is defined as: the above decomposition for x, the system of Eq. (6)
can be equivalently written in the following form:
Aƒj =ujƒj, j=1, …,  (7) dxs
= Asxs + Bsu+ fs (xs, xf )+ Ws (xs, xf, q) (9)
where uj denotes an eigenvalue and ƒj denotes an dt
eigenfunction; the eigenspectrum of A, |(A), is defined dxf
as the set of all eigenvalues of A, i.e. |(A) = =Af xf + Bf u+ ff (xs, xf )+ Wf (xs, xf, q)
dt
{u1, u2, …, uj, …}. We will assume that |(A) satisfies
the following properties. yc = Cxs + Cxf, ym = Sxs + Sxf
Assumption 1. (Christofides, 2001) xs (0)= Psx(0)= Psx0, xf (0)= Pf x(0)= Pf x0
1. Re{u1}] Re{u2} ]··· ]Re{uj } ]···, where Re{uj }
denotes the real part of uj. where As = Ps A, Bs = Ps B, fs = Ps f, Ws =Ps W,
2. |(A) can be partitioned as |(A) =|1(A) + |2(A), Af = Pf A, Bf = Pf B, ff = Pf f, Wf = Pf W. In the
where |1(A) consists of the first m (with m finite) above system, As is a diagonal matrix of dimension
eigenvalues, i.e. |1(A) ={u1, …, um }, and Re{u1}/ m×m of the form As = diag{uj }, fs (xs, xf ), ff (xs, xf ),
Re{um } = O(1). are Lipschitz vector functions, Ws (xs, xf, q),
3. Re{um + 1}B 0 and Re{um }/Re{um + 1} = O(m), Wf (xs, xf, q) are Lipschitz matrix functions, and Af is a
where m =Re{u1}/Re{um + 1} B1 is a small positive stable unbounded differential operator. Using that m=
number. Re{u1}/Re{um + 1}, the system of Eq. (9) can be writ-
ten in the following form:
Remark 1. Referring to the system of Eq. (1), several dxs
remarks are in order: (a) the spatial differential opera- = Asxs + Bsu+ fs (xs, xf )+ Ws (xs, xf, q) (10)
dt
tor is linear; this assumption is valid for diffusion–con-
vection-reaction processes where the diffusion (xf
m =Afmxf + mBf u+ mff (xs, xf )+ mWf (xs, xf, q)
coefficient and the conductivity can be taken indepen- (t
dent of temperature and concentrations; (b) the manip-
yc = Cxs + Cxf, ym = Sxs + Sxf
ulated input enters the system in a linear and affine
fashion; this is typically the case in many practical where Afm is an unbounded differential operator
applications where, for example, the wall temperature is defined as Afm = mAf. Since m B1 and the operators
chosen as the manipulated input; and (c) the nonlinear- As, Afm generate semigroups with growth rates which
ities appear in an additive fashion (e.g. complex reac- are of the same order of magnitude, the system of Eq.
tion rates, Arrhenius dependence of reaction rates on (10) is in the standard singularly perturbed form (see
temperature). Kokotovic, Khalil, & O’Reilly, 1986 for a precise defin-
 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203 191

ition of standard form), with xs being the slow states kinetics include terms of the type 1/m the model reduc-
and xf being the fast states. Introducing the fast time- tion procedure can be extended to account for this type
scale ~=t/m and setting m =0, we obtain the following of behavior (see also Kumar, Christofides, & Daoutidis,
infinite-dimensional fast subsystem from the system of 1998 for results on this problem). Finally, regarding the
Eq. (10): accuracy of the finite-dimensional model, it can be
shown (Christofides, 2001) that when q(t) 0, the dis-
(xf
= Afmxf (11) crepancy between the solution of the finite-dimensional
(~ model of Eq. (12) and the solution of the parabolic
From the fact that Re{um + 1} B0 and the definition of PDE of Eq. (1) is of O(m). This result accounts for both
m, we have that the above system is globally exponen- the transport and reaction terms and was shown using
tially stable. Setting m =0 in the system of Eq. (10), we singular perturbation arguments and results for stabil-
have that xf = 0 and thus, the finite-dimensional slow ity of infinite-dimensional systems.
system takes the form:
dx̃s
=Asx̃s +Bsu+fs (x̃s, 0) +Ws (x̃s, 0, q) (12)
dt 3. Problem statement and solution framework
yc = Cx̃s, ym = Sx̃s
In this paper, we address the problem of computing
where the tilde notation in x̃s denotes that this is the optimal locations of point control actuators and point
state of a finite-dimensional system. measurement sensors associated with robust nonlinear
output feedback control laws of the following general
Remark 3. We note that the above model reduction form:
procedure which led to the approximate ODE system of
u=F(ym ) (13)
Eq. (12) can be also used, when empirical eigenfunc-
tions of the system of Eq. (6) computed through where F(ym ) is a nonlinear vector function and ym
Karhunen –Loéve expansion (see Christofides, 2001; denotes the vector of measured outputs, so that the
Atwell & King, 2001 for applications of this method in following properties are enforced in the closed-loop
the design of low-order controllers) are used as basis system: (a) boundedness of the state; (b) arbitrary
functions in Hs and Hf, instead of the eigenfunctions degree of attenuation of the uncertain variables on the
of A. state of the slow subsystem; and (c) the optimal actua-
tor/sensor locations, which are obtained on the basis of
Remark 4. Referring to the model reduction of the PDE the closed-loop finite-dimensional system, are near-opti-
system of Eq. (1) (diffusion– convection –reaction pro- mal in the sense that it approaches the optimal solution
cess) using Galerkin’s method, it is important to note for the infinite-dimensional system as the separation of
the following: while the solution of the parabolic PDE the slow and fast eigenmodes increases. To address this
system is expanded in terms of the eigenfunctions of the problem, we will initially synthesize, under the assump-
spatial differential operator, both the linear spatial tion of existence of bounding functions that capture the
differential operator (diffusion and convection terms) size of the uncertain terms, stabilizing robust nonlinear
and the nonlinear term f(x̄) and W(w̄,r(z)q(t)) (reac- state feedback controllers via Lyapunov techniques on
tion terms) are projected and used for the construction the basis of finite-dimensional approximations of the
of the finite-dimensional approximation. To clearly see PDE system. The optimal actuator location problem
this point, consider the terms fs (x̃s, 0) +Ws (x̃s, 0, q) in will be subsequently formulated as the one of minimiz-
the finite-dimensional system of Eq. (12); these terms ing a meaningful cost functional that includes penalty
are the projections of the nonlinear terms f(x̄)+ on the response of the closed-loop system and the
W(x̄, r(z)q(t)) that appear in the parabolic PDE of Eq. control action. Then, under the assumption that the
(1). Therefore, the finite-dimensional model not only number of measurement sensors is equal to the number
approximates the diffusion and convection phenomena of slow modes, we will employ a procedure proposed in
but also it approximates the reaction phenomena (mod- Christofides and Baker (1999) for obtaining estimates
elled by the nonlinear terms f(x̄) +W(x̄, r(z)q(t))). for the states of the approximate finite-dimensional
Moreover, the separation of the infinite set of ODEs model from the measurements. The estimates will be
into fast and slow according to the fast and slow combined with the robust state feedback controllers to
eigenvalues of the spatial differential operator is mean- derive robust output feedback controllers. The optimal
ingful because the kinetics terms ( f(x̄) +W(w̄,r(z)q(t))) location of the measurement sensors will be computed
do not contribute to the two-time-scale behavior of the by minimizing a cost function of the estimation error in
spatial modes of the parabolic PDE systems since they the closed-loop infinite-dimensional system. It will be
do not include terms of the type 1/m. Note that if the established by using singular perturbation techniques
192 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203

that: (a) the robust output feedback controllers enforce Ws (x̃s, 0, q) 5c(xs, t), qk (t)5 qbk, k= 1, 2, …, q
boundedness of the state and uncertainty attenuation in (15)
the infinite-dimensional closed-loop system provided
for all xs Hs, qRq, t] 0.
that the separation between the slow and the fast
Proposition 1 that follows provides the explicit for-
eigenvalues is sufficiently large; and (b) the solution to
mula for the state feedback controller that achieves the
the optimal actuator/sensor problem, which is obtained
aforementioned control objectives. The proof of the
on the basis of the closed-loop finite-dimensional sys-
proposition is given in the Appendix A.
tem, is near-optimal in the sense that it approaches the
Proposition 1. Consider the finite-dimensional system of
optimal solution for the infinite-dimensional system as
Eq. (12) for which Assumption 2 and Assumption 3
the separation between the slow and fast eigenvalues
hold. Then, the state feedback controller:
 
increases.
x̃s
u= Bs− 1 (Ls − As )x̃s − fs (x̃s, 0)− c(x̃s, t)
x̃s +ƒ
4. Integrating robust control and optimal actuator (16)
placement where Ls is a stable matrix, \ 1 and ƒ\ 0 are positi6e
parameters, guarantees boundedness of the state of the
4.1. Robust state feedback controller synthesis closed-loop finite-dimensional system and ensures that
limt “  x̃s (t)5 k(ƒ) where k(·) is a class K function.
In this section, we assume that measurements of the
states of the PDE system of Eq. (12) are available and Remark 5. The structure of the closed-loop finite-di-
address the problem of synthesizing robust nonlinear mensional system of Eq. (12) under the controller of
static state feedback control laws of the general form: Eq. (16) has the following form:
x̃s
u = F(za, x̃s ) (14) x̃; s = Lsx̃s − c(x̃s, t) + Ws (x̃s, 0, q) (17)
x̃s + ƒ
where F(za, x̃s ) is a nonlinear vector function and za From the above equation, it follows directly that the
denotes the vector of the actuator locations, that guar- response of this system depends on the stable matrix Ls,
antee boundedness of the state of the closed-loop finite- the initial condition, xs (0), the uncertain variable q(t),
dimensional system and make the effect of the and the matrix Bs which accounts for the actuator
uncertain variables on the state arbitrarily small by locations. This is completely different from the optimal
appropriate selection of the controller parameters. We actuator placement for PDE systems without uncertain
note that owing to the fact that we allow the uncertain variables where the structure of the closed-loop finite-
variables, q(t), to change the equilibrium of the finite- dimensional system was of the following form (Antoni-
dimensional system of Eq. (12), it is not possible to ades and Christofides, 2001):
synthesize a continuous controller that enforces asymp-
x̃; s = Lsx̃s (18)
totic stability in the closed-loop system. To proceed
with the synthesis of the controller, we will need the and thus, the response was independent of the actuator
following assumptions (see Remark 6 below for a dis- locations.
cussion on this assumption). Remark 6. The requirement l =m is sufficient and not
Assumption 2. l=m (i.e. the number of control actua- necessary, and it is made to simplify the solution of the
tors is equal to the number of slow modes), and the controller synthesis problem. Full linearization (when
inverse of the matrix Bs exists. q(t)0) of the closed-loop finite-dimensional system
Assumption 3 that follows requires the existence of a through coordinate change and nonlinear feedback can
nonlinear time-varying bounding function that captures be achieved for any number of manipulated inputs (i.e.
the size of the uncertain terms in the system of Eq. (12) for any l[1, m]), provided that an appropriate set of
and the existence of known bounds for the uncertain involutivity conditions is satisfied by the corresponding
variables qk (t). Such bounding functions are typically vector fields of the system of Eq. (12) (see Isidori, 1989
obtained from physical considerations, preliminary sim- for details).
ulations or experimental data. The requirement of exis- Remark 7. Referring to the role and selection of the
tence of bounding functions is standard in all parameters  and ƒ of the controller of Eq. (16), we
Lyapunov-based robust control methods (see Corless & note that  determines the gain of the controller term
Leitmann, 1981; Christofides, Teel, & Daoutidis, 1996). that compensates for the effect of the uncertainty and
Assumption 3. There exists a known function c(xs, t) should be chosen greater than unity to achieve global
and a known set of positive constants qbk such that the boundedness of the state of the closed-loop finite-di-
following conditions holds: mensional system (note that the choice = 1 would
 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203 193

provide local boundedness of the state). On the other and Athans, 1978; Antoniades and Christofides, 2000)
hand, ƒ determines the degree of asymptotic attenua- and consider an average cost over a set of m linearly
tion of the effect of the uncertainty on the state and independent initial conditions, x is(0), i= 1, …, m, of the
should be chosen to ensure uncertainty attenuation following form:
with smooth control action (note that as ƒ “0 the
controller of Eq. (16) becomes discontinuous and the J( s =
1 m
%
& Tf
i
((x̃ Ts (x is(0), t), Qsx̃s (x is(0), t))
control action non-smooth). Finally, it is important to mi = 1 0

point out that the robust controller of Eq. (16) compen-

+u T(x̃s (x is(0), t), za) Ru(x̃s (x is(0), t), za)) dt (20)
sates for the effect of uncertain variables on the state of
the closed-loop finite-dimensional system without using Note that different initial conditions correspond to
high-gain feedback; this is a consequence of our objec- different times Tfi. An additional issue that should be
tive to design a practically-implementable controller accounted for in the cost functional used to compute
which would not destabilize the fast modes of the the optimal actuator locations is the fact that the
closed-loop infinite-dimensional system (note that there evolution of the state, x̃s, of the closed-loop finite-di-
is guarantee that such a destabilization will be avoided mensional system depends on the profile of the vector
if high-gain feedback of the type 1/m is employed to of unknown variables q(t). To account for this depen-
compensate for the effect of the disturbances). dence, we modify the cost of Eq. (20) to average over a
large set of disturbance profiles as follows:

J. s
&
4.2. Computation of optimal location of control
actuators 1 1 K m Tf
is
= % % ((x̃ Ts (x is(0), q s, t), Qsx̃s (x is(0), q s, t))
m Ks = 1 i = 1 0
In this subsection, we compute the actuator locations
so that the robust state feedback controller of Eq. (16) + u T(x̃s (x is(0), q s, t), za) Ru(x̃s (x is(0), q s, t), za)) dt (21)
is near-optimal for the full PDE system of Eq. (9) with
respect to a meaningful cost functional which imposes Note again that different profiles of the vector of
penalty on the response of the closed-loop system and unknown variables, q(t), might also correspond to dif-
the control action. Owing to the boundedness of the ferent times Tfis. In contrast to the case of PDE systems
state, we consider a cost functional which is defined without uncertainty, the solution to optimal placement
over a finite-time interval (the final time is defined as problem, zam, for which the above cost is minimum can
the needed for the process state to become smaller than be only found numerically since it requires multiple
the desired attenuation limit), Tf. From the stability integrations of the closed-loop finite-dimensional sys-
bound x̃s (t)5 Kx̃s (0)e − h2t +k(ƒ), it follows directly tem of Eq. (17) for appropriate finite times, linearly
that given the desired attenuation limit, d, so that independent initial conditions and different disturbance
x̃s (t ] Tf )5d and the value of ƒ used in the con- profiles (Remark 8 below discusses a numerical ap-
troller, the values of Tf can be computed (note that Tf proach for computing the optimal actuator locations).
is the time in which the state x̃s (t) satisfies x̃s (t] Theorem 1 that follows establishes that the robust
Tf ) 5d for the first time). Specifically, we focus on the controller of Eq. (16) guarantees the desired bounded-
ODE system of Eq. (12) and consider the following cost ness and uncertainty attenuation properties in the
functional: closed-loop finite-dimensional system and the locations

Js =
& Tf
((x̃ Ts (xs (0), t), Qsx̃s (xs (0), t))
for the control actuators that minimize the cost of Eq.
(21) for the finite-dimensional closed-loop system are
0 also near optimal for the closed-loop infinite-dimen-
+u T(x̃s (xs (0), t), za ) Ru(x̃s (xs (0), t), za)) dt (19) sional system, provided that the degree of separation
between the fast and slow modes of the spatial differen-
where Qs and R are positive definite matrices. The cost tial operator of the PDE system is sufficiently large (the
of Eq. (19) is well-defined and meaningful since it proof of this theorem can be found in the Appendix A).
imposes penalty on the response of the closed-loop Theorem 1. Consider the infinite-dimensional system of
finite-dimensional system and the control action. How- Eq. (9) for which Assumption 1 holds, and the finite-di-
ever, a potential problem of this cost is its dependence mensional system of Eq. (12), for which Assumption 2
on the choice of a particular initial condition, xs (0), and and Assumption 3 hold. Then, there exist positi6e real
thus, the solution to the optimal placement problem numbers v1, v2 and m* such that if xs (0) 5v1,
based on this cost may lead to actuator locations that xf (0)2 5 v2, and m(0, m*], then the controller of Eq.
perform very poorly for a large set of initial conditions. (14):
To eliminate this dependence and obtain optimality 1. guarantees boundedness of the state of the closed-
over a broad set of initial conditions, we follow (Levine loop infinite-dimensional system, and
194 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203

2. the optimal locations of the point actuators obtained is terminated when the change in the optimal actuator
for the closed-loop finite-dimensional system are locations between two consecutive steps is smaller than
near-optimal for the closed-loop infinite-dimensional a desired accuracy. This approach was successfully used
system in the sense that: in the diffusion-reaction process example discussed in
Section 6.
J.

=
1 1 K m
% %
& Tf
is
((x Ts (x is(0), q s, t), Qsxs (x is(0), q s, t))
Remark 9. It is important to point out that in the case
of PDE systems without uncertainty, the closed-loop
finite-dimensional system takes the form x̃; s = Lsx̃s and
K ms = 1 i = 1 0

&
thus the penalty on the response of this system:
+(x Tf (x is(0), q s, t), Qf xf (x if(0), q s, t)) (22)

1 m
+ u (xs (x (0), q , t), za) Ru(xs (x (0), q , t), za)) dt“ J. s
T i
s
s i
s
s
J. xs = % (x̃ Ts (x is(0), t), Qsx̃s (x is(0), t)) dt (23)
mi = 1 0
as m“0
is finite and independent of the actuator locations.
where Qf is an unbounded positi6e definite operator and J. Therefore, the optimal actuator placement problem re-
is the a6erage cost function associated with the controller duces to the one of minimizing the following cost which

&
of Eq. (16) and the infinite-dimensional system of Eq. only includes penalty on the control action:
(9). 
1 m
Remark 8. Owing to the numerical complexity involved J. us = % u T(x̃s (x is(0), t), za) Ru(x̃s (x is(0), t), za ) dt
in computing the actuator locations that exactly mini- mi = 1 0

mize the cost of Eq. (21) (it involves search over an J. us is a function of multiple variables, za =
infinite number of locations), we initially assume a large [za1 za2 … zal ], and thus, it obtains its local minimum
number, say L, of equispaced locations along the length values when its gradient with respect to the actuator


of the spatial domain in which the control actuators are locations is equal to zero, i.e.:
possible to be placed. In determining this set of loca-
tions, we use standard controllability arguments for
(J. us
=
(J. us (J. us
···
n T
= [0 · · · 0]T (24)
parabolic PDEs (see Ray, 1981) to exclude actuator (za (za1 (zal
locations for which the resulting system is uncontrol- and 9z azaJ. us (zam )\ 0, where 9z az aJ. us is the Hessian ma-
lable. Since the disturbance vector q(t) is unknown, we trix of J. us and zam is a solution of the system of
then use a stochastic approach (which is inspired by the nonlinear algebraic equations of Eq. (24) (which in-
methodology proposed in Gazi, Ungar, Seider, and cludes l equations with l unknowns and can be obtained
Kuipers (1996), Gazi, Seider, and Ungar (1997) for by standard methods). The solution zam for which the
robust controller verification) to compute a large num- above conditions are satisfied and J. us obtains its
ber of randomly-generated disturbance profiles that smallest value (global minimum) corresponds to the
satisfy qk (t) 5qbk for all times (Assumption 2). We optimal actuator locations for the closed-loop finite-di-
finally compute the value of the cost of Eq. (21) for all mensional system.
possible combinations of the actuator locations to cal-
culate the optimal locations. While the computational
complexity of the above procedure is not very large for 5. Robust output feedback control with optimal sensor
a small number of control actuators, its computational placement
complexity grows very fast as the number of control
actuators increases, making the practical implementa- In this section, we proceed with the synthesis of
tion of this procedure very difficult. In this case, the robust output feedback controllers that enforce stability
following algorithm which employs a search procedure and uncertainty attenuation in the closed-loop infinite-
that uses the optimal actuator locations for the nominal dimensional system, and compute the optimal locations
system (see Remark 9 below for how this can be done) for the measurement sensors. Specifically, we consider
as initial guess can be followed. The procedure is as output feedback control laws of the general form:
follows: we initially fix l− 1 control actuators at the
u(t)= F(ym ) (25)
optimal locations computed on the basis of the nominal
system and perform a search for the optimal location of where F(ym ) is a nonlinear vector function and ym is
the remaining control actuator that minimizes the cost the vector of measured outputs. The synthesis of the
for the uncertain system. This process is then performed controller of Eq. (25) will be achieved by combining the
l −1 more times for each one of the other control state feedback controller of Eq. (16) with a procedure
actuators to compute a new set of actuator locations proposed in Christofides & Baker (1999) for obtaining
for the uncertain system. This set of actuator locations estimates for the states of the approximate ODE model
is subsequently used as a new initial guess to carry out of Eq. (12) from the measurements. To this end, we
the above procedure for a second time. The algorithm need to impose the following requirement on the num-
 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203 195

ber of measured outputs in order to obtain estimates of (b) the locations of the point actuators and measurement
the states xs of the finite-dimensional system of Eq. sensors are near-optimal in the sense that the cost
(12), from the measurements y sm, s = 1, 2, …, p. function associated with the controller of Eq. (27)
Assumption 4. p=m (i.e. the number of measurements and the system of Eq. (9) satisfies:
is equal to the number of slow modes), and the inverse
of the operator S exist, so that x̂s =S − 1ym. J.
We note that the requirement that the inverse of the
operator S exists can be achieved by appropriate
choice of the location of the measurement sensors (i.e.
=
1 1 K m
% %
& Tf
is
((x Ts (x is(0), q s, t), Qsxs (x is(0), q s, t))
K ms = 1 i = 1 0
functions s s(z)). The optimal locations for the measure-
ment sensors can be computed by minimizing an aver- + (x Tf (x is(0), q s, t), Qf xf (x if(0), q s, t)) (28)
age cost function of the estimation error of the
+u T(xs (x is(0), q s, t), za) Ru(xs (x is(0), q s, t), za)) dt“J. s
closed-loop infinite-dimensional system of the form:

J. (e) as m“ 0

=
1 1 K m
% %
& Tf
is
(xs (x is(0), q s, t) −x̂s (x is(0), q s, t)2) where J. and J. s are the a6erage cost functions of the
K ms = 1 i = 1 0 infinite-dimensional system of Eq. (9) and the finite-di-
mensional system of Eq. (12), respecti6ely, under the
dt (26)
robust output feedback controller of Eq. (27).
where xs is the slow state of the closed-loop infinite-di- Remark 10. Note that in contrast to the case of state
mensional system of Eq. (9), x̂s =S − 1ym, and e(t)= feedback control where the controller depends only the
xs − x̂s 2 is the estimation error. As in the case of the slow state xs, the output feedback controller of Eq. (27)
optimal location problem for the control actuators, the uses static feedback of the measured outputs y sm, s=
solution to the sensor placement problem requires the 1, 2, …, p, and thus, it feeds back both xs and xf. While
numerical integration of the closed-loop system (both this type of feedback of the state xf has been shown to
finite- and infinite-dimensional) in order to compute xs, lead to closed-loop instability for certain classes of
and x̂s (from the measurements y sm, s = 1, 2, …, p), and singularly perturbed systems (Kokotovic et al., 1986),
thus, it is computationally demanding (see Remark 11 in the present case, the large separation of the slow and
for how compute the optimal sensor locations). fast modes of the spatial differential operator (i.e. as-
Theorem 2 that follows establishes that the proposed sumption that m is sufficiently small), together with the
robust output feedback controller enforces stability in special structure of the fast subsystem of Eq. (10)
the closed-loop infinite-dimensional system and that the (where m multiplies the control input) and the fact that
solution to the optimal actuator/sensor problem, which the controller does not include terms of the form O(1/
is obtained on the basis of the closed-loop finite-dimen- m), do not allow such a destabilization to occur.
sional system, is near-optimal in the sense that it ap- Remark 11. Similar to the case of optimal actuator
proaches the optimal solution for the placement, we perform the search for the optimal sen-
infinite-dimensional system as the separation of the sor locations so that the cost of Eq. (26) is minimized
slow and fast eigenmodes increases. The proof of this over a finite set of equispaced locations along the length
theorem can be found in the Appendix A. of the spatial domain of the PDE.
Theorem 2. Consider the system of Eq. (9) for which
Assumption 1 hold, and the finite-dimensional system of
Eq. (12), for which Assumption 2, Assumption 3 and 6. Application to a diffusion-reaction process with
Assumption 4 hold, under the robust output feedback uncertainty
controller:

u= Bs− 1 (Ls −As )S − 1ym −fs (S − 1ym, 0)
Consider a long, thin rod in a reactor (Fig. 1). The
reactor is fed with pure species A and a zeroth order


exothermic catalytic reaction of the form A“ B takes
S − 1ym place on the rod. Since the reaction is exothermic, a
−c(S − 1ym, t) (27) cooling medium which is in contact with the rod is used
S − 1ym  +ƒ
for cooling. Under the assumptions of constant density
Then, there exist positi6e real numbers v1, v2 and m* such and heat capacity of the rod, constant conductivity of
that if xs (0) 5v1, xf (0)2 5v2, and m  (0, m*], then the the rod, and constant temperature at both ends of the
controller of Eq. (27): rod, the mathematical model which describes the spa-
(a) guarantees boundedness of the state of the closed- tiotemporal evolution of the dimensionless rod temper-
loop system, and ature consists of the following parabolic PDE:
196 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203

(x̄ ( 2x̄ ( 2x̄

= + iTe − k/(1 + x̄) +iU(b(z)u(t) −x̄) − iT,n e − k Ax = , xD(A) = {x H([0, y]; R);
(t (z 2 (z 2
(29)
x̄(0, t)= 0, x̄(y, t)= 0} (33)
subject to the Dirichlet boundary conditions:
can be solved analytically and its solution is of the
x̄(0, t)= 0, x̄(y, t) =0 (30) form:
and the initial condition:
uj = − j 2, ƒj (z)= ƒ( j (z)=
' 2
sin( j z),
x̄(z, 0) = x̄0(z) (31) y

where x̄ denotes the dimensionless temperature of the j =1, …,  (34)

rod, iT denotes a dimensionless heat of reaction (which where uj, ƒj, ƒ( j, denote the eigenvalues, eigenfunctions
is assumed to be unknown and time-varying; uncertain and adjoint eigenfunctions of Ai, respectively.
variable), iT,n denotes a nominal dimensionless heat of Even though the eigenvalues of A are all stable, the
reaction, k denotes a dimensionless activation energy, spatially uniform operating steady-state x̄(z, t)=0 of
iU denotes a dimensionless heat transfer coefficient, the system of Eq. (29) is an unstable one (i.e. the
and u denotes the manipulated input (temperature of linearization of the system of Eq. (29) around x̄(z, t)=
the cooling medium). The following typical values were 0 possesses one unstable eigenvalue due to the exother-
given to the process parameters: mic nature of the reaction). Fig. 2 shows the
iT,n =50.0, iU =2.0, k= 4.0 (32) spatiotemporal evolution of the open-loop dimensional
rod temperature when no uncertainty is present, start-
The eigenvalue problem for the spatial differential op- ing from initial conditions close to the steady state
erator of the process: x̄(z, t)= 0; the process moves to another stable steady
state characterized by a maximum in the temperature
profile, hot-spot, in the middle of the rod. We note that
in all the simulation runs, a 20-th order Galerkin
truncation of the system of Eq. (29) was used in our
simulations in order to accurately describe the process
(further increase on the order of the Galerkin trunca-
tion was found to give identical numerical results).
Therefore, the control objective is to stabilize the di-
mensionless rod temperature at the unstable steady-
state x̄(z, t)=0, in the presence of time-varying
uncertainty in the dimensionless heat of reaction iT.
We consider the first two Galerkin modes of the PDE
system as the slow modes and use Galerkin’s method to
construct a second-order ODE system which is used for
controller design and optimal actuator/sensor place-
ment. This approximate system has the following form:

 n 
Fig. 1. Catalytic rod.
x̃; s1 u1 − iU
;x̃s2 = 0
0 n nx̃s1
(35)
u2 − iU x̃s2

+ iU
ƒ( 1(za1) ƒ( 1(za2) n n
u1 
f (x̃ , 0)
+ (iT,n + q(t)) 1 s
n
ƒ( 2(za1) ƒ( 2(za2) u2 f2(x̃s, 0)

where za1 and za2 are the locations of the two point
actuators, q(t) is the uncertain term in the dimension-
less heat of reaction and the explicit form of the terms
f1(x̃s, 0) and f2(x̃s, 0) are omitted for brevity. We used
the second-order ODE system of Eq. (35) for the
synthesis of the nonlinear robust output feedback con-
troller through an application of the formula of Eq.
(16) (the explicit form is omitted for brevity). This
Fig. 2. Profile of evolution of rod temperature in the open-loop controller was employed in the simulations with =
system. 1.1, qb = 12.5 and ƒ=0.005.
 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203 197

+ u T(x̃s (x is(0), q s, t), za) Ru(xs (x̃ is(0), q s, t), za)) dt

where

Qs = R=
 n
1 0
,
0 1

T( f = max{Tfis }= 10 for i= 1, 2, s= 1, 2, …, 20,

x 1s (0)= ƒ1 and x 2s (0)= ƒ2, are the two linearly indepen-
dent initial conditions, and q s, s= 1, 2, …,20, are the
randomly generated disturbance profiles that satisfy
q(t)[− 0.25 iT,n, 0.25 iT,n ] and q(t+ 0.001)=q(t)+
0.04 r where r is a random number r[− 1, 1]. In Fig. 3,
we can see three of these randomly generated profiles of
the uncertainty. The optimal location of control actua-
tors were computed to be: za1 = 0.42y and za2 =0.59y.
Fig. 3. Profile of three randomly-generated disturbances. In Fig. 4 we see the contour plot of constant J. s
(0.87, 0.89, 0.93, 1.05, 1.2, 1.5, 2.5 and 6) for the closed-
loop system under the robust state feedback controller
for different locations of the two control actuators. The
plot is symmetric about the 45° line as the two actuator
locations are equivalent. We can see that the minimum
J. s = 0.87 is at the optimal locations that we have
computed.
Using the result of Theorem 2, the vector of mea-
sured outputs ȳm (t)R2 is defined as:

 n 
ȳm1
=
x(zs1, t) n (37)
ȳm2 x(zs2, t)

where zs1 and zs2 are the locations of the two point
sensors. The optimal sensor locations were computed
Fig. 4. Contour plot of constant J. s (0.87, 0.89, 0.93, 1.05, 1.2, 1.5, 2.5
by minimizing a cost functional of the estimation error
and 6) for the closed-loop system under the robust state feedback
controller for different locations of the two control actuators. which has the following form:
J. (e)
Table 1
Results for the actuator locations
=
1 20 2
% %
& T( f
(xs (x is(0), q s, t)− x̂s (x is(0), q s, t)2) dt
40s = 1 i = 1 0
Case Actuator locations J. s (38)
Optimal 0.42y, 0.59y 0.8678 where xs is obtained from the simulation of the high-or-
2 0.39y, 0.66y 0.9198 der discretization of the PDE, and x̂s is calculated from
3 0.32y, 0.68y 1.0380 the following equations:

 n  n 
4 0.25y, 0.75y 1.5060
x̂s1 ƒ( (z )
= 1 s1
ƒ( 1(zs2) −1
ym1(zs1, t) n
, (39)
Using the results of Theorem 1, the optimal locations x̂s2 ƒ( 2(zs1) ƒ( 2(zs2) ym2(zs2, t)
of the control actuators were computed by minimizing
The optimal location of the measurement sensors were
a cost that includes penalty on the control action and
computed to be: zs1 = 0.35y and zs2 = 0.66y.
averages over the set of two linearly independent initial
In the remainder, we will present some of our simula-
conditions and a large set of randomly generated dis-
tion results. We initially focus on the state feedback
turbance profiles q(t) which satisfy q(t) 5qb for all
control problem and compute, for the sake of compari-
times. The specific cost used in our calculations has the
son, the optimal locations of the control actuators, J. s
following form:
for the case where no uncertainty in the dimensionless

J. u =
1 20 2
% %
& T( f
((x̃ Ts (x is(0), q s, t), Qsx̃s (x is(0), q s, t))
heat of reaction is considered (iT = iT,nom). We also
compute J. s for two arbitrary actuator locations. The
computed results for these cases are shown in Table 1.
40s = 1 i = 1 0
(36) From the values in Table 1, one can see that a robust
198 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203

controller which uses the optimal actuator locations at

za1 =0.42y and za2 =0.59y, utilizes less control effort in
order to stabilize the system at the spatially-uniform
steady state x̃(z, t)=0, compared to the other three
actuator placements.
Fig. 5 shows the norm of the control effort, u, for
the four different actuator placements reported in Table
1, optimal case (solid line), case 2 (long-dashed line),
case 3 (short-dashed line), and case 4 (dotted line), for
xs (0) = ƒ1 and for two different, randomly generated,
disturbance profiles. We can see that the control effort
used for the optimal case is less than the other cases.
Fig. 6 presents the same results as Fig. 5 but for the
initial condition xs (0) = ƒ2. One observe that there is
no significant difference in the control effort needed for
stabilization and uncertainty attenuation among the
different actuator placements.
We now turn to the problem of optimal placement of
the measurement sensors. For the sake of comparison,
Table 2 shows J. (e), J. s for the optimal sensor place-

Fig. 6. Norm of the control effort, u for two different disturbance

profiles (top and bottom plots) for the optimal case (solid line), case
2 (long-dashed line), case 3 (short-dashed line), and case 4 (dotted
line) —xs (0) =ƒ2.

Table 2
Results for the sensor locations

Case Sensor locations J. (e) J. s

Optimal 0.35y, 0.66y 8.862e−4 0.8992

2 0.30y, 0.64y 2.131e−2 0.9523
3 0.38y, 0.68y 5.435e−1 4.3531

ment and two other placements. We can see from Table

2 that the computed optimal sensor locations at zs1 =
0.35y and zs2 = 0.66y, have much smaller estimation
error than the other two sensor locations. Fig. 7 shows
the norm of the closed-loop estimation error versus
time, for the optimal actuator/sensor locations, for
xs (0)= ƒ1 (top plot) and xs (0)= ƒ2 (bottom plot) for
five different, randomly generated, disturbances.
Clearly, the estimation error is very small for both
Fig. 5. Norm of the control effort, u for two different disturbance
profiles (top and bottom plots) for the optimal case (solid line), case
initial conditions.
2 (long-dashed line), case 3 (short-dashed line), and case 4 (dotted Fig. 8 displays the profiles of the evolution of the
line) —xs (0) =ƒ1. temperature of the rod, under robust output feedback
 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203 199

control, for the optimal actuator/sensor locations, for

xs (0) = ƒ1 (top plot), and xs (0) =ƒ2 (bottom plot). In
both cases, the controller stabilizes the closed-loop
PDE system at the spatially uniform operating steady-
state very quickly despite the presence of time-varying
uncertainty.
Finally, for the sake of comparison, we show in Figs.
9 and 10 the profiles of the evolution of the temperature
of the rod under an output feedback controller that
does not account for the presence of uncertainty for
xs (0) = ƒ1. In Fig. 9 we have used the optimal actuator/
sensor locations (za1 =0.42y, za2 =0.59y and zs1 =
0.35y, zs2 =0.66y) and in Fig. 10 we have used the
optimal actuator/sensor locations for the case where no
uncertainty is considered (za1 =0.39y, za2 =0.66y and
zs1 =0.31y, zs2 =0.72y). We have used the same ran-
domly generated disturbance profiles employed in the
simulations shown in Fig. 8 (top plot), Figs. 9 and 10 in
order for our comparison to be meaningful. We clearly
observe that the closed-loop performance is signifi-
cantly superior (in terms of uncertainty attenuation and

Fig. 8. Profile of evolution of the rod temperature under robust

output feedback control with the optimal actuator/sensor locations.
Top plot: xs (0) =ƒ1. Bottom plot: xs (0) =ƒ2.

Fig. 9. Profile of evolution of the rod temperature under output

feedback control without uncertainty attenuation using the optimal
actuator/sensor locations —xs (0) =ƒ1.

speed of convergence to the steady-state) when the

robust output feedback controller with the optimal
actuator/sensor locations is applied to the system.

7. Conclusions

This paper focused on transport-reaction processes

Fig. 7. Norm of the closed-loop estimation error e versus time, for with unknown time-varying parameters and distur-
the optimal actuator/sensor locations, for five different disturbance bances described by quasi-linear parabolic PDE sys-
profiles. Top plot: xs (0)= ƒ1. Bottom plot: xs (0)=ƒ2. tems, and addressed the problem of computing optimal
200 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203

actuator/sensor locations for robust nonlinear con- results were successfully applied to a typical diffusion-
trollers. Galerkin’s method was initially employed to reaction process with nonlinearites and uncertainty to
derive finite-dimensional approximations of the PDE design a robust nonlinear output feedback controller
system which were used for the synthesis of robust and compute the optimal actuator/sensor locations for
nonlinear state feedback controllers via geometric and robust stabilization of an unstable steady state.
Lyapunov techniques and the computation of optimal
actuator locations. The controllers enforce boundedness
and uncertainty attenuation in the closed-loop system. Acknowledgements
The optimal actuator location problem was subse-
quently formulated as the one of minimizing a mean- Financial support from the National Science Foun-
ingful cost functional that includes penalty on the dation, CTS-0002626, and the Air Force Office of
response of the closed-loop system and the control Scientific Research is gratefully acknowledged.
action. Owing to the boundedness of the state, the cost
was defined over a finite-time interval (the final time is
defined as the time needed for the process state to Appendix A
become smaller than the desired uncertainty attenua-
tion limit), while the optimization was performed over a
broad set of initial conditions and time-varying distur- Proof of Proposition 1. Substituting the control law of
bance profiles. Subsequently, under the assumption that Eq. (16) into the finite-dimensional system of Eq. (12),
the number of measurement sensors is equal to the we obtain:
number of slow modes, we employed a standard proce- x̃s
dure for obtaining estimates for the states of the ap- x̃; s = Lsx̃s − c(x̃s, t) + Ws (x̃s, 0, q) (40)
x̃s + ƒ
proximate finite-dimensional model from the
measurements. The optimal location of the measure- To prove boundedness of the state of the above system
and uncertainty attenuation, we use the following Lya-
ment sensors was computed by minimizing a cost func-
punov function candidate:
tion of the estimation error in the closed-loop
infinite-dimensional system. We showed that the use of 1
V= x̃ 2s (41)
these estimates in the robust state feedback controller 2
leads to a robust output feedback controller, which
Computing the time-derivative of this function along
guarantees boundedness of the state and uncertainty
the trajectories of the closed-loop system of Eq. (40),
attenuation in the infinite-dimensional closed-loop sys-
we obtain:
tem, provided that the separation between the slow and
the fast eigenvalues is sufficiently large. We also estab- x̃s
V: = x̃ Ts x̃; s = x̃ Ts [Lsx̃s − c(x̃s, t) + Ws (x̃s, 0, q)]
lished that the solution to the optimal actuator/sensor x̃s + ƒ
problem, which is obtained on the basis of the closed- (42)
loop finite-dimensional system, is near-optimal in the Using the stability property of Ls and the bound of
sense that it approaches the optimal solution for the Assumption 3, the following bound for V: can be
infinite-dimensional system as the separation between computed:
the slow and fast eigenvalues increases. The theoretical
x̃ 2s
V: 5 − h1x̃ 2s − c(x̃s, t) + x̃s c(x̃s, t) (43)
x̃s + ƒ
− x̃ 2s c(x̃s, t)+ x̃ 2s c(x̃s, t)+ ƒx̃s c(x̃s, t)
5 −h1x̃ 2s +
x̃s + ƒ
− (− 1)x̃ 2s c(x̃s, t)+ ƒx̃s c(x̃s, t)
5 − h1x̃ 2s +
x̃s + ƒ
where h1 is a strictly positive real number which is
smaller or equal to the smallest eigenvalue of the matrix
Ls. Since x̃s /x̃s + ƒB 1, we finally have:
V: 5 − h1x̃ 2s + (−(− 1)x̃s + ƒ)c(x̃s, t) (44)
From the above inequality, it follows that when x̃s ]
Fig. 10. Profile of evolution of the rod temperature, under output
ƒ/( − 1) (i.e. the state, x̃s, is outside of a ball which is
feedback control and optimal actuator/sensor placement which do centered at the origin), then − (− 1)x̃s + ƒ50, and
not account for the disturbances —xs (0)= ƒ1. thus, V: satisfies the following bound:
 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203 201

V: 5 −h1x̃ 2s (45) and thus the closed-loop of the finite-dimensional slow

system takes the form:
This implies (Khalil, 1992) that there exist positive real
numbers K, h2 and a class K function k(·) such that the dx̃s x̃s
= Lsx̃s − c(x̃s, t) + Ws (x̃s, 0, q) (50)
state of the system of Eq. (40) satisfies: dt x̃s + ƒ
x̃s (t)5 Kx̃s (0)e − h2t +k(ƒ) (46) In Proposition 1, we proved that the state of the above
slow subsystem is globally bounded. Since the state of
From the above inequality, it follows directly that the
the slow subsystem of Eq. (50) is bounded and the fast
state of the system of Eq. (40) remains bounded and
subsystem of Eq. (48) is globally exponentially stable,
asymptotically approaches a ball whose size is scaled by
there exist positive real numbers v1, v2, and m* such
the controller parameter ƒ, lim x̃s (t)5 k(ƒ).
t“ that if xs (t)5 v1, xf (t)2 5 v2, and m(0, m*], then the
system of Eq. (47) is bounded and the solution
Proof of Theorem 1. The proof of this theorem will be
xs (t), xf (t) of the system of Eq. (47) satisfies for all
obtained in two steps. In the first step, we will show
t [tb, Tf ):
boundedness and closeness of solutions for the closed-
loop system of Eq. (9), provided that the initial condi- xs (t)= x̃s (t)+ O(m) (51)
tions and m are sufficiently small. In the second step, we
xf (t)= x̃f (t)+ O(m)
will exploit the closeness of solutions result to show
that the cost associated with the closed-loop PDE sys- where tb = O(m) is the time required for xf (t) to ap-
tem approaches the optimal cost associated with the proach x̃f (t). x̃s (t) and x̃f (t) are the solutions of the
closed-loop finite-dimensional system under state feed- slow and fast subsystems of Eqs. (48) and (50), respec-
back control, when the initial conditions and m are tively (Christofides, 2001) and Tf = O(1) is a positive
sufficiently small, thereby establishing that the location constant.
of the control actuators obtained by using the finite-di- Near-optimality of the actuator locations: The cost for
mensional system is near-optimal. the closed-loop infinite-dimensional system can be writ-
Boundedness–closeness of solutions: Using that m=
&
ten as follows:
Re{u1}/Re{um + 1} and under the controller of Eq. 1 1 K m tb
(16), the closed-loop system of Eq. (9) takes the form: J. = % % ((x Ts (x is(0), q s, t), Qsxs (x is(0), q s, t))
K ms = 1 i = 1 0
dxs xs
=Lsxs −c(xs, t) +Ws (xs, xf, q)+fs (xs, xf ) +(x Tf (x is(0), q s, t), Qf xf (x if(0), q s, t)) (52)
dt xs +ƒ
+ u T(xs (x is(0), q s, t), za) Ru(xs (x is(0), q s, t), za)) dt

(x
−fs (xs, 0) (47)
+
1 1 K m
% %
& Tf
is
((x Ts (x is(0), q s, t), Qsxs (x is(0), q s, t))
m f =Afmxf + mf( f (xs, xf ) + mWf (xs, xf, q) K ms = 1 i = 1
(t tb

+ (x Tf (x is(0), q s, t), Qf xf (x if(0), q s, t))

where Afm is an unbounded differential operator
defined as Afm =mAf, and − f( f (xs, xf ) = Bf Bs− 1((As − +u T(xs (x is(0), q s, t), za) Ru(xs (x is(0), q s, t), za)) dt
Ls )xs +fs (xs, 0)−c(xs, t)(xs /xs  +ƒ)) + ff (xs, xf ).
From the closeness of solutions results of Eq. (51) and
Since m is a small positive number less than unity
picking Tfis 5 Tf (this is always possible by appropriate
(Assumption 1, part 3), the system of Eq. (47) is in the
choice of the controller parameters), it follows that for
standard singularly perturbed form, with xs being the
tb 5 t5Tfis (outside of the boundary layer):
slow states and xf being the fast states. Introducing the
fast time-scale ~ =t/m and setting m = 0, we obtain the xs (x is(0), q s, t)“ x̃s (x is(0), q s, t), xf (x if(0), q s, t) “0,
following infinite-dimensional fast subsystem from the
u(xs (x is(0), t), za)“ u(x̃s (x is(0), t), za) as m“ 0
system of Eq. (47):
(53)
(x̃f

&
= Afmx̃f (48) and hence:
(~
1 1 K m Tf
is
where the tilde symbol in x̃f, denotes that the state x̃f is % % (x Ts (x is(0), q s, t) Qsxs (x is(0), q s, t)
K ms = 1 i = 1 tb
associated with the approximation of the fast xf -subsys-
tem. From the fact that Re{um + 1} B0 and the defini- + (x Tf (x if(0), q s, t) Qf xf (x if(0), q s, t) (54)
tion of m, we have that the above system is globally
+ u T(xs (x is(0), q s, t), za) Ru(xs (x is(0), q s, t), za)) dt
exponentially stable. Setting m = 0 in the system of Eq.
(47) and using that the operator Afm is invertible, we
have that:
“
1 1 K m
% %
& Tf
is
(x̃ Ts (x is(0), q s, t) Qsx̃s (x is(0), q s, t)
K ms = 1 i = 1 tb

x̃f = 0 (49) + u T(x̃s (x is(0), q s, t), za) Ru(x̃s (x is(0), q s, t), za)) dt
202 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203

xs + xf
as m“0 − c(xs + xf, t) )+ mff (xs, xf )
xs + xf + ƒ
From the boundedness of the state of the closed-loop
+ mWf (xs, xf, q)
system, we have that there exists a positive real number
M that bounds the absolute values of the integrand of Using that m is a small positive number less than unity
Eq. (52) for t[0, Tf ]. Using the fact that tb =O(m), we (Assumption 1, part 3), and introducing the fast time-
then have: scale ~=t/m and setting m= 0, we obtain the following
1 1 K m
% %
& tb
(x Ts (x is(0), q s, t) Qsxs (x is(0), q s, t)
infinite-dimensional fast subsystem which describes the
fast dynamics of the system of Eq. (58):
K m s=1 i=1 0
(x̃f
+ (x Tf (x if(0), q s, t) Qf xf (x if(0), q s, t) (55) =Afmx̃f (59)
(~
+u T(xs (x is(0), q s, t), za) Ru(xs (x is(0), q s, t), za)) dt which is globally exponentially stable. Setting m=0 in

5
& tb
M dt5Mm =O(m)
the system of Eq. (58) and using that the operator Afm
is invertible, we have that:
0
x̃f = 0 (60)
Similarly, from the stability of the closed-loop finite-di-
mensional system of Eq. (12) and the fact that tb = and thus the closed-loop of the finite-dimensional slow
O(m), we have that there exists a positive real number system takes the form:
M% such that:
1 1 K m
% %
& tb
(x̃ Ts (x is(0), q s, t) Qsx̃s (x is(0), q s, t)
dx̃s
dt
= Lsxs − c(xs, t)
xs
xs + ƒ
+ Ws (xs, 0, q) (61)
K m s=1 i=1 0
which is globally bounded. Therefore, there exist posi-
(56)
tive real numbers v1, v1, and m* such that if xs (t)5v1,
+ u T(x̃s (x is(0), q s, t), za) Ru(x̃s (x is(0), q s, t), za)) dt xf (t)2 5 v2, and m(0, m*], then the system of Eq. (10)

5
& tb
M% dt 5 M%m=O(m)
is globally bounded and the solution xs (t), xf (t) of the
system of Eq. (58) satisfies the estimates of Eq. (58).
0 Given the stability and closeness of solutions results
Combining Eqs. (54)–(56), we obtain: for the closed-loop system, the near-optimality of the

1 1 K m
% %
& Tf
is
(x Ts (x is(0), q s, t) Qsxs (x is(0), q s, t)
control actuators and measurement sensors in the sense
described in Eq. (28) can be established by using similar
K m s=1 i=1 0 calculations to the ones in part 2 of the proof of
Theorem 1.
+ (x Tf (x if(0), q s, t) Qf xf (x if(0), q s, t) (57)
+u T(xs (x is(0), q s, t), za) Ru(xs (x is(0), q s, t), za)) dt

“
1 1 K m
% %
& Tf
is
(x̃ Ts (x is(0), q s, t) Qsx̃s (x is(0), q s, t) References
K m s=1 i=1 0
Alvarez, J., Romagnoli, J. A., & Stefanopoulos, G. (1981). Variable
+ u T(x̃s (x is(0), q s, t), za) Ru(x̃s (x is(0), q s, t), za)) dt
measurement structures for the control of a tubular reactor.
as m“ 0 Chemical Engineering Science, 36, 1695 – 1712.
Amouroux, M., Di Pillo, G., & Grippo, L. (1976). Optimal selection
This completes the proof of the theorem. of sensors and control location for a class of distributed parame-
ter systems. Ric. Automatica, 7, 92 – 108.
Antoniades, C., & Christofides, P. D. (2000). Computation of optimal
Proof of Theorem 2. Under the output feedback con- actuator locations for nonlinear controllers in transport-reaction
troller of Eq. (27), the closed-loop system takes the processes. Comp. & Chem. Eng., 24, 577 – 583.
form: Antoniades, C., & Christofides, P. D. (2001). Integrating nonlinear
output feedback control and optimal actuator/sensor placement
dxs xs +xf for transport-reaction processes. Chemical Engineering Science,
=Lsxs −c(xs + xf, t) +(As −Ls )xf
dt xs +xf  +ƒ 56, 4517 – 4535.
Arbel, A. (1981). Controllability measures and actuator placement in
+fs (xs, xf ) (58) oscillatory systems. International Journal of Control, 33, 565–574.
Atwell, J. A., & King, B. B. (2001). Proper orthogonal decomposition
− fs (xs + xf, 0)+Ws (xs, xf, q) for reduced basis feedback controllers for parabolic equations.
Math. & Comp. Model., 33, 1 – 19.
(xf Balas, M. J. (1979). Feedback control of linear diffusion processes.
m =Afmxf
(t International Journal of Control, 29, 523 – 533.
Chen, W. H., & Seinfeld, J. H. (1975). Optimal location of process
+mBf Bs− 1((As −Ls )(xs +xf ) + fs (xs +xf, 0) measurements. International Journal of Control, 21, 1003 –1014.
 C. Antoniades, P.D. Christofides / Computers and Chemical Engineering 26 (2002) 187–203 203

Choe, K., & Baruh, H. (1992). Actuator placement in structural Kubrusly, C. S., & Malebranche, H. (1985). Sensors and controllers
control. Journal of Guidance, 15, 40–48. location in distributed systems-a survey. Automatica, 21, 117–128.
Christofides, P. D. (1998). Robust control of parabolic PDE systems. Kumar, A., Christofides, P. D., & Daoutidis, P. (1998). Singular
Chemical Engineering Science, 53, 2949–2965. perturbation modeling and control of nonlinear two-time-scale
Christofides, P. D. (2001). Nonlinear and robust control of PDE processes with non-explicit time-scale multiplicity. Chemical Engi-
systems: methods and applications to transport-reaction processes. neering Science, 53, 1491 – 1504.
Boston: Birkhäuser. Kumar, S., & Seinfeld, J. H. (1978a). Optimal location of measure-
Christofides, P. D., Teel, A. R., & Daoutidis, P. (1996). Robust ments for distributed parameter estimation. IEEE Trans. Au-
semi-global output tracking for nonlinear singularly perturbed tomat. Contr., AC-23, 690 – 698.
systems. International Journal of Control, 65, 639–666. Kumar, S., & Seinfeld, J. H. (1978b). Optimal location of measure-
Christofides, P. D., & Baker, J. (1999). Robust output feedback ments in tubular reactors. Chemical Engineering Science, 33,
control of quasi-linear parabolic PDE systems. Syst. and Contr. 1507 – 1516.
Lett., 36, 307 – 316. Levine, W. S., & Athans, M. (1978). On the determination of the
Colantuoni, G., & Padmanabhan, L. (1977). Optimal sensor locations optimal constant output feedback gains for linear multivariable
for tubular-flow reactor systems. Chemical Engineering Science, system. IEEE Trans. Automat. Contr., 15, 44 – 48.
32, 1035 – 1049. Malandrakis, C. G. (1979). Optimal sensor and controller allocation
Corless, M., & Leitmann, G. (1981). Continuous state feedback
for a class of distributed parameter systems. International Journal
guaranteeing uniform ultimate boundedness for uncertain dy-
of Systematic Science, 10, 1283 – 1295.
namic systems. IEEE Trans. Automat. Contr., 26, 1139–1144.
Morari, M., & O’Dowd, J. (1980). Optimal sensor location in the
Demetriou, M. A. (1999). Numerical investigation on optimal actua-
presence of nonstationary noise. Automatica, 16, 463 – 480.
tor/sensor location of parabolic PDEs. Proceedings of the Ameti-
Omatu, S., Koide, S., & Soeda, T. (1978). Optimal sensor location
can Control Conference (pp. 1722 –1726). San Diego, CA.
problem for a linear distributed parameter system. IEEE Trans.
Friedman, A. (1976). Partial differential equations. New York: Holt,
Automat. Contr., AC-23, 665 – 670.
Rinehart & Winston.
Omatu, S., & Seinfeld, J. H. (1983). Optimization of sensor and
Gazi, E., Seider, W. D., & Ungar, L. H. (1997). A non-parametric
Monte-Carlo technique for controller verification. Automatica, 33, actuator location in a distributed parameter system. Journal of
901 – 906. Franklin Institute, 315, 407 – 421.
Gazi, E., Ungar, L. H., Seider, W. D., & Kuipers, B. J. (1996). Rao, S. S., Pan, T. S., & Venkayya, V. B. (1991). Optimal placement
Automatic analysis of Monte-Carlo simulations of dynamic chem- of actuators in actively controlled structures using genetic al-
ical plants. Comp. & Chem. Engng., 20(s), 987–992. gorithms. AIAA J., 29, 942 – 943.
Harris, T. J., MacGregor, J. F., & Wright, J. D. (1980). Optimal Ray, W. H. (1981). Ad6anced process control. New York: McGraw-
sensor location with an application to a packed bed tubular Hill.
reactor. American Institute of Chemical Engineering Journal, 26, Waldraff, W., Dochain, D., Bourrel, S., & Magnus, A. (1998). On the
910 – 916. use of observability measures for sensor location in tubular reac-
Isidori, A. (1989). Nonlinear control systems: an introduction (2nd tor. J. Proc. Contr., 8, 497 – 505.
ed.). Berlin-Heidelberg: Springer-Verlag. Xu, K., Warnitchai, P., & Igusa, T. (1994). Optimal placement and
Khalil, H. K. (1992). Nonlinear systems. New York: Macmillan gains of sensors and actuators for feedback control. J. Guid.
Publishing Company. Contr. Dyn., 17, 929 – 934.
Kokotovic, P. V., Khalil, H. K., & O’Reilly, J. (1986). Singular Yu, T. K., & Seinfeld, J. H. (1973). Observability and optimal
perturbations in control: analysis and design. London: Academic measurement location in linear distributed parameter systems.
Press. International Journal of Control, 18, 785 – 799.