43rd IEEE Conference on Decision and Control WeC09.

2
December 14-17, 2004
Atlantis, Paradise Island, Bahamas

MIMO Integral-Action Anti-Windup Controller Design and

Applications to Temperature Control in RTP Systems
A. N. Mete and A. N. Gündeş

Abstract— An integral-action controller synthesis is pre- proportional+integral+derivative (PID) form. Closed-loop

sented, where the controller achieves closed-loop stability and stability can be achieved using PID-controllers only for cer-
zero steady-state error due to step-input references. Closed- tain classes of plants, while many others cannot be stabilized
loop stability is maintained even when the integrators in all of
the controller channels are limited or completely switched off using PID-controllers. Standard observer-based integral-
to protect against integrator windup. The proposed method is action controller designs apply Linear Quadratic Regulator
applied to temperature control in Rapid Thermal Processing (LQR) or pole-placement methods to an augmented plant
systems, where critical requirements are imposed on tracking model, which includes the n states of the plant and the ny
of temperature profiles. states of the integrators [13], [11], [5]. An augmented full-
order observer-based design results in an (n + ny )-th order
I. I NTRODUCTION
controller. The integrators cannot be completely switched
We consider controller synthesis for linear, time-invariant off since the gain matrix corresponding to the integrator
(LTI) multi-input multi-output (MIMO) plants subject to states are part of the feedback loop that ensures stability.
input saturation. Our goal is to achieve closed-loop stability Instead of turning off the integral-action completely to
and asymptotic tracking of step-input references with zero protect against integrator windup, the integrators are limited
steady-state error. To achieve robust tracking, the controller so that the input does not reach the saturation limits.
is designed to have integral-action. In this paper, we propose an integral-action synthesis
The performance of integral-action controllers depends procedure based on adding integral-action onto the system,
on the system operating in a linear range. They suffer where an initially designed stabilizing controller is already
serious loss of performance due to a phenomenon called present in the feedback loop. This is achieved in two stages:
integral windup, which occurs when the actuators in the An initial stabilizing controller is designed for the original
control-loop saturate (see e.g. [6]). Limitations of actuators plant using any desired method (LQR, H∞ , etc.) and it does
put an upper bound on the amplitude of the control signal. not have integral-action. Then a PID-controller is designed
Actuators reach their saturation limit when a reference for a stable system associated with the plant. These two
signal requiring a control effort beyond that bound is applied blocks are configured so that the final controller achieves
to the system. The integrators in the controller continue to closed-loop stability and integral-action. Furthermore, all
integrate the error although the input is constrained, and integral-action controllers can be obtained from this con-
once the input comes out of the saturation limit, the initial troller by inclusion of a free controller parameter. If a
condition that the integrator has built up to causes a large full-order observer-based controller is chosen for the initial
transient response and serious performance degradation. stabilizing stage, then only the plant states are estimated and
Effects of integral windup can range from large overshoots the state feedback gains are only associated with the plant
in transient response to loss of stability. Therefore, integral- states without augmenting. The transfer-function of the final
action controllers should be modified by taking actuator controller once the PID block is added on is (n + 1). To
limitations into account. Most methods of dealing with protect against windup due to the integrators in the PID
the effect of integrator windup are based on the idea of block containing the integral-action, this second block can
turning off the integrators when the input reaches a limit be completely switched off without affecting closed-loop
and resetting the integrators’ states. Numerous anti-windup stability or it can be limited.
modifications have been proposed, many of which can be The proposed integral-action controller design is applied
considered as observer-based modifications (see e.g., [1]). A to temperature control in Rapid Thermal Processing (RTP)
synthesis procedure for designing observer-based controller systems (see e.g. [14], [3] for modelling and control of
gains in anticipation of windup can be found in [10]. RTP). The recent technology of RTP in integrated circuit
There are several ways to design LTI integral-action manufacturing is a fast and efficient multi-chamber single-
controllers (see e.g. [2], [12], [8], [9], [5] for decentralized wafer technology that uses a much smaller chamber than
and centralized integral-action controller synthesis). The a batch process. Single-wafer processing achieves more
simplest controller that achieves integral-action is in the uniform film thickness for larger wafers. During an RTP
process, it is crucial to maintain uniform temperature on the
This work was supported by the NSF Grant ECS-9905729 wafer surface at all times since small temperature variations
The authors are with Electrical and Computer Engineering, University
of California, Davis, CA 95616. anmete@ece.ucdavis.edu & can lead to large variations in reaction rates [3]. The benefits
gundes@ece.ucdavis.edu of RTP cannot be realized without meeting the stringent
0-7803-8682-5/04/$20.00 ©2004 IEEE 2590
temperature uniformity specifications. Fast tracking control G iff Ĉ is proper and Sys(G, Ĉ) is stable. iii) The stable
laws that achieve near uniform spatial temperature distribu- system Sys(G, Ĉ) is said to have integral-action iff Her
tions across a semiconductor wafer during both transient has blocking zeros at the origin.
and quasi steady-state phases of the process need to be Suppose that Sys(G, Ĉ) is stable and that step reference
developed for RTP systems, which are inherently nonlinear inputs r are applied to the system. The steady-state error
dynamic processes with actuator saturation. due to all step inputs goes to zero if and only if Her (0) = 0,
The paper is organized as follows: The problem descrip- i.e., the system has integral-action.
tion and definitions are in Section II. The main results Let G = Ng Dg−1 = D
−1 N
g , Ĉ = Ncr D−1 = D−1 Nc
g cr c
are in Section III: Section III-A explains the proposed be any RCF and LCF of the plant and the controller. Then
controller synthesis. Theorem 3.1 states that all integral- Ĉ stabilizes G if and only if ML in (1) equivalently, MR ,
action controllers can be realized in two stages, where the is unimodular [15], [7]:
integral-action is achieved using a separate PID block that
g Dcr + N

g Ncr =: ML , Dc Dg + Nc Ng =: MR (1)
can be switched off or limited without affecting closed- D
loop stability. One method of designing the PID stage is the Then Her can be written as in (2), and equivalently, (3):
systematic procedure in Proposition 3.1. Section III-B gives
a review of integral-action design using an augmented plant
g ,
Her = (Iny + GĈ)−1 = Dcr ML−1 D (2)
model. In Section IV, the results are applied to a linearized Her = Iny − GĈ(Iny + GĈ) −1
= Iny − Ng MR−1 Nc . (3)
MIMO RTP system model subject to lamp input saturation
nonlinearity as described in [4]. The proposed systematic By Definition 2.1, Sys(G, Ĉ) has integral-action iff
synthesis approach is compared with the standard observer- Her (0) = (Dcr ML−1 D
g )(0) = 0.
−1
based controller design with augmented state feedback. Definition 2.2: The controller Ĉ = Ncr Dcr is said to
Although we discuss continuous-time systems here, all be an integral-action controller iff Ĉ stabilizes G and the
results also apply to discrete-time systems with appropriate denominator matrix Dcr for any RCF of Ĉ has blocking
modifications. The following notation is used: Let C I , IR zeros at the origin, i.e., Dcr (0) = 0.
−1
denote complex and real numbers. The extended closed By Definition 2.2 and (2), if Ĉ = Ncr Dcr is an integral-
right-half complex plane is U = {s ∈ C I |Re(s) ≥ 0}∪{∞}; action controller, then Sys(G, Ĉ) has integral-action. Ob-
Rp denotes real proper rational functions of s; S ⊂ Rp viously, Dcr (0) = 0 is sufficient but not necessary for
is the stable subset with no poles in U; M(S) is the set Her (0) = (Dcr ML−1 D
g )(0) = 0. If G has poles at s = 0,
of matrices with entries in S ; In is the n × n identity rankD
g (0) < ny ; hence, the system may achieve integral-
matrix. The H∞ -norm of M (s) ∈ M(S) is M  := action even if Dcr (0) = 0. If G has no poles at s = 0, then
sups∈∂U σ̄(M (s)), where σ̄ is the maximum singular value −1
Sys(G, Ĉ) has integral-action if and only if Ĉ = Ncr Dcr
and ∂U is the boundary of U. We drop (s) in transfer ma- is an integral-action controller, i.e., Dcr (0) = 0.
trices such as G(s) whenever this causes no confusion. We Lemma 2.1 gives the necessary conditions imposed on
use coprime factorizations over S ; i.e., for G ∈ Rp ny ×nu , G due to the integral-action requirement:
G = Ng Dg−1 denotes a right-coprime-factorization (RCF), Lemma 2.1: Let G ∈ Rp ny ×nu . If the system
where Ng ∈ Sny ×nu , Dg ∈ Snu ×nu , det Dg (∞) = 0; Sys(G, Ĉ) has integral-action, then i) (normal) rankG =
G = D
−1 N

g denotes a left-coprime-factorization (LCF), ny ≤ nu ; ii) G has no transmission zeros at the origin.
g


where Ng ∈ Sny ×nu , D
g ∈ Sny ×ny , det D

g (∞) = 0. Proof: The stability of Sys(G, Ĉ) implies Her (0) =
Iny − GHwr (0) = 0, i.e., GHwr (0) = Iny . Therefore,
II. P ROBLEM D ESCRIPTION (normal) rank(GHwr ) = ny ≤ min{rankG, rankHwr }
Consider the LTI, MIMO unity-feedback system implies ny ≤ rankG ≤ min{ny , nu }. By (3), Her (0) = 0
Sys(G, Ĉ) in Fig. 1; G ∈ Rp ny ×nu and Ĉ ∈ Rp nu ×ny implies Ng (0)MR−1 (0)Nc (0) = Iny ; hence, rankNg (0) =
are the plant’s and the controller’s transfer-functions, re- rankNc (0) = ny .
spectively. Assume that the feedback system is well-posed,
III. M AIN R ESULTS
the plant and controller have no unstable hidden-modes, and
the plant G ∈ Rp ny ×nu is full normal rank. Let Her = The simplest integral-action controllers are in PID form.
(Iny + GĈ)−1 = Iny − GĈ(Iny + GĈ)−1 Iny − GHwr We consider a realizable form of proper PID-controllers,
denote the (input-error) transfer-function from r to e. where Kp , Ki , Kd ∈ IRnu ×ny are called the proportional,
Definition 2.1: i) The system Sys(G, Ĉ) is said to be the integral, and the derivative constant, respectively [5]:
stable iff the closed-loop transfer-function from (r, v) to Ki Kd s
(y, w) is stable. ii) The controller Ĉ is said to stabilize Cpid = Kp + + , (4)
s τd s + 1
v To implement the derivative term, a pole is typically added
r e w ? y
- f - Ĉ - f- G - to the derivative term (with τd > 0) so that Cpid in (4) is
−6 proper. The integral-action in Cpid is present when Ki = 0.
The controller in (4) is in proportional+integral (PI) form
Fig. 1. Unity-Feedback System Sys(G, Ĉ).
2591
Cpi = Kp + Ki /s when Kd = 0, integral+derivative (ID) A. Two stage integral-action design
form Cid = Ki /s + Kd s/(τd s + 1) when Kp = 0, pure Standard integral-action designs are generally based on
integral (I) form Ci = Ki /s when Kp = Kd = 0. augmenting the plant to include the integrator’s states in
Although PID-controllers are simple and low order, some state-feedback. This approach is briefly reviewed in Sec-
(unstable) plants G are not stabilizable with any Cpid . Since tion III-B. In this section, we propose an approach that
the plants considered here are not restricted to be stable, does not involve augmenting the plant’s states. Theorem 3.1
existence of stabilizing PID-controllers is not guaranteed. states that any integral-action controller for G = Ng Dg−1
Proposition 3.1 shows that stable systems can be stabilized can be expressed as the sum of two blocks: The first is
using PID-controllers, with Ki = 0, if and only if G has
designed using any
any stabilizing controller Cg = Y
−1 X,
no transmission zeros at the origin, and proposes a method method, and does not have integral-action. The second is
of selecting the constants Kp , Ki , Kd . Y
−1 Cpid , where Cpid is any PID-controller that stabilizes
Proposition 3.1: Let Ng ∈ Sny ×nu , (normal) rankNg = Ng ; it can be designed using any method including the
ny ≤ nu . i) There exist stabilizing PID-controllers with procedure in Proposition 3.1. This second block provides
nonzero integral constant Ki ∈ IRnu ×ny if and only if integral-action and even if it is switched-off, the closed-
rankNg (0) = ny . ii) Suppose rankNg (0) = ny . Let loop remains stable due to Cg still remaining in the loop.
Ng (0)I ∈ IRnu ×ny be any right-inverse of Ng (0). Choose Theorem 3.1: Let the plant be G ∈ Rp ny ×nu , (normal)
any K̂ p , K̂ d ∈ IRnu ×ny , τd > 0. With Ĉpd defined as in rankG = ny ≤ nu ; let G have no transmission zeros at
(5) below, let Kp = ρK̂ p , Kd = ρK̂ d , Ki = ρNg (0)I , s = 0. Let G = D
−1 N

g be any LCF, G = Ng Dg−1 be any
g
where ρ ∈ IR is any positive constant satisfying (6): RCF. Choose any controller Cgo ∈ Rp nu ×ny that stabilizes
K̂d s G. There exists an LCF Cgo = D
c of Cgo that satisfies

c−1 N
Ĉpd := K̂p + ; (5)
τd s + 1
c Dg + N

c Ng = In .
D u (13)
Ng (s)Ng (0) − I −1
I
0 < ρ <  Ng (s)Ĉpd +  . (6) Let Cpid be any PID-controller stabilizing Ng with Ki = 0.
s
Then Ĉ is an integral-action controller for G if and only if
Then Ng is stabilized by the PID-controller in (7):

c − QN
Ĉ = (D
g )−1 ( N

c + QD

g + Cpid ) , (14)
ρNg (0)I ρK̂ d s
Cpid = ρK̂ p + + . (7)
c − QN
g )(∞) = 0.
s τd s + 1 where Q ∈ Snu ×ny satisfies det(D
In (7), K̂ d = 0 gives a PI-controller; K̂ p = 0 gives an We prove that any Ĉ in (14) is an integral-action controller
ID-controller; K̂ d = K̂ p = 0 gives a pure I-controller. for G; a detailed proof that all integral-action controllers
Proof : For any positive a ∈ IR, define Z ∈ M(S) as are in the form given by (14) can be found in [8]:
s Kd s s Ki Proof : Define Z ∈ M(S) as in (8). Since Cpid stabilizes
Z := Cpid
s+a
= ( Kp + ) +
τd s + 1 s + a s + a
. (8) Ng , the matrix M in (9) is unimodular. Let Cgo = Nc Dc−1
be an RCF of Cgo that satisfies
Then Cpid = Z( s+a s
Iny )−1 is an RCF of Cpid . Since Cpid
stabilizes Ng , the matrix M defined by (9) is unimodular:
g Dc + N
D
g Nc = In . (15)
y

s   s
M := Iny + Ng Z = Iny + Ng Cpid In (9) Then all stabilizing controllers Cg for G can be expressed
s+a s+a y as Cg = Y
−1 X
= XY −1 , where
By (9), a = 0, rankM (0) = rank(a−1 Ng (0)Ki ) = ny ≤
min{rankNg (0), rankKi } ≤ min{ny , nu } = ny implies Y
:= (D

c − QN

g ), X

:= (N

c + QD

g ),
rankNg (0) = ny . Then there exist a right inverse Ng (0)I ∈ Y := (Dc − Ng Q), X = (Nc + Dg Q) , (16)
IRnu ×ny , i.e., Ng (0)Ng (0)I = Iny . Define M̂ as nu ×ny
c − QN
g )(∞) = 0
and Q ∈ S satisfies det(D
s+a s s
−1

[15], [7]. Using Y X = XY −1
and (9), Ĉ in (14)
M̂ = M= I + Ng Cpid . (10)
s+ρ s+ρ s+ρ can be written as: Ĉ = Y
−1 (X
+ Cpid ) = Y
−1 X
+
ρ
−1


−1

Y (Y Dg + XNg )Cpid = Y X(Iny + Ng Cpid ) +
In (10), add and subtract s+ρ Iny to obtain
Dg Cpid = XY −1 Iny + Ng Cpid + Dg Cpid = (X +
ρs Ng (s)Ng (0)I − Iny Dg Cpid
s
M −1 Y ) Y −1 M (
s
In )−1 . Therefore, an
M̂ = Iny + ( Ng Ĉpd + ) . (11) s+a s+a y
s+ρ s −1
RCF Ĉ = Ncr Dcr for the controller Ĉ is given by
Since  s+ρ
ρs
 = ρ, for any ρ > 0 satisfying (6), we have
−1 s
Ncr Dcr = (X + Dg ZM −1 Y )( M −1 Y )−1 , (17)
ρs Ng (s)Ng (0)I − Iny s+a
 ( Ng Ĉpd + )  < 1 . (12)
s+ρ s where Ncr , Dcr ∈ M(S) and Dcr is biproper. It follows
by (9), (15) and N
g Ng that ML = D

g Dg = D
g Dcr +
By (12), M̂ is unimodular; equivalently, M is unimodular s



Ng Ncr = Dg (

M Y ) + Ng (X + Dg ZM −1 Y ) =
−1
since a, ρ > 0; therefore Cpid stabilizes Ng . s+a
2592

g ( s M −1 Y + Ng ZM −1 Y ) + N
D
g X = D

g M M −1 Y + - Cpid v
s+a

g X = D
N
gY + N

g X = In is unimodular. The system r e w ?u y
y
-e -N

-?
e -D

−1 -e- G -
Sys(G, Ĉ) is stable since ML in (1) is unimodular. Since c c
s −6 6
Dcr (0) = ( M −1 Y )|s=0 = 0, Dcr has blocking zeros
s+a Q
−1
at s = 0. By Definition 2.2, any Ĉ = Ncr Dcr given by
6
(14) is therefore an integral-action controller for G. -D

−1 - e
g 
g N
The block diagram of Sys(G, Ĉ), with Ĉ as in (14), is in
Fig. 2. For Q = 0, the integral-action controller Ĉ becomes
Fig. 2. The system Sys(G, Ĉ) with integral-action controller Ĉ.

c−1 Cpid .
Cˆo = Cgo + D (18)
- s−1 I - Ki Cpid
The controller Ĉ in (14) is simplified for stable plants as
follows: Let G ∈ Sny ×nu , rankG(0) = ny ≤ nu . Let - sKd ?
-e
Cpid , with Ki = 0, be a PID-controller that stabilizes G. τd s + 1 6
Then Ĉ is an integral-action controller for G if and only - Kp
if Ĉ = (I − QG)−1 ( Q + Cpid ), where Q ∈ Snu ×ny v
satisfies det(I − QG)(∞) = 0. -Q
r e ?- e?w u y
The parametrization given in (14) can also be obtained -e - −K - e -?e-G -
using a state-space representation (A, B, C, D) of G ∈ −6 ?
Rp ny ×nu , where A ∈ IRn×n , (A, B) is stabilizable and e - −L
(C, A) is detectable. Let K ∈ IRnu ×n and L ∈ IRn×ny be 6
such that FL and FK defined in (19) are stable: e C  (sI − A)−1  ?
e
6 x̂ 6
FL := (sI −A+LC)−1 , FK := (sI −A+BK)−1 . (19) B
Then using G = Ng Dg−1 = D
g , where

g−1 N 6
D
Dg = I − KFK B , Ng = (C − DK)FK B + D,
D
g = I − CFL L , N

g = CFL (B − LD) + D , (20) Fig. 3. Sys(G, Ĉ) using full-order observer-based controller Cgo
.
a controller Cgo = D c

c = Nc D−1 is given by

−1 N
c
(C − DK)(A − BK)−1 (sI − A + BK)−1 sBNg (0)I
implies s−1 (Ng (s)Ng (0)I − I) = (C − DK)(A −

c = I + KFL (B − LD) , N
D
c = KFL L , BK)−1 FK BNg (0)I . The bound in (6) for ρ > 0 becomes
Dc = I + (C − DK)FK L , Nc = KFK L ,
ρ < Ng (s)Ĉpd +(C −DK)(A−BK)−1 FK BNg (0)I −1 .
Cgo = K ( sI − A + BK + L(C − DK) )−1 L . (21) (24)
With the nominal full-order observer-based controller Cgo Then Ng is stabilized by the PID controller Cpid in (7).
in (21), the expression for all integral-action controllers in B. Integral-action design based on plant augmentation
(14) of Theorem 3.1 becomes Ĉ = [I + KFL (B − LD) − We briefly review the well-known full-order observer-
Q(CFL (B−LD)+D)]−1 [ KFL L+Q(I −CFL L)+Cpid ], based integral-action controller synthesis, where the integra-
where Q ∈ Snu ×ny is such that det(I −Q(∞)D) = 0. With tor’s states are also included in state-feedback [5], [4]. The
Q = 0, the controller Cˆo in (18) is expressed as state-feedback matrix can be designed using pole-placement
Cˆo = K(sI − A + BK + L(C − DK))−1 L + D
c−1 Cpid . or LQR. Let (A, B, C, D) be a state-space representation of
(22) G ∈ Rp ny ×nu , where A ∈ IRn×n , (A, B) is stabilizable,
The block diagram of Sys(G, Ĉ), with Cgo as in (21), is (C, A) is detectable. Let L ∈ IRn×ny be such that FL ∈
in Fig. 3. The PID block for Ng = [(C − DK)FK B + Sn×n . Define the (n + ny )-th order augmented system as
D] can be designed using Proposition 3.1: Since G has no    
A 0 B
transmission zeros at s = 0, Ng (0) has a right-inverse Aa := , Ba := , Ca := [ C 0 ] . (25)
−C 0 −D
 I
Ng (0)I = D − (C − DK)(A − BK)−1 B . (23) The pair (Aa , Ba ) is stabilizable if and only if (A, B) is
stabilizable and G has no transmission zeros at the origin.
The expression (6) is simplified as follows: Since FK ∈
A state-feedback Ka = [ Kx Kξ ] is then determined for
M(S), FK (0) = (−A + BK)−1 exists. By (23), [Ng (s) −
the augmented system in (25), and the resulting (n + ny )-th
Ng (0)]Ng (0)I = [(C −DK)FK B +D−D+(C −DK)(A−
order observer-based controller is called Ĉ a , given by
BK)−1 B]Ng (0)I = (C − DK)(A − BK)−1 [(sI(A −
−1
BK)−1 − I)−1 + I]BNg (0)I = (C − DK)(A − Ĉ a = −Ka [sI − Aa + Ba Ka + La (Ca − DKa ) ] La ,
BK)−1 [(sI(A−BK)−1−I)−1 sI(A−BK)−1 ]BNg (0)I = (26)
2593
  
−L the PID block for Ng in (20) of the RCF G = Ng Dg−1 .
where La := . The block diagram of Sys(G, Ĉ a )
Iny Choosing K̂ p = 15I3 and K̂ d = 0, the inequality
with the (n + ny )-th order Ĉ a is in Fig. 4. The augmented (24) becomes 0 < ρ < Ng (s)Ĉpd + (C − DK)(A −
system is stabilized using Ka ; if Kξ = 0, the design BK)−1 FK BN I −1
does not guarantee stability with Kx acting alone. The ⎡ g (0)  = 0.3292. Choose ⎤ ρ = 0.3; with
4.44 −0.58 −0.05
integrators cannot be taken out of service completely. Ng (0)−1 = ⎣ 0.57 4.43 −0.35 ⎦, Cpi = ρ(K̂ p +
0.09 0.32 4.46
IV. A PPLICATIONS TO RTP SYSTEMS 1 0.3
−1
We apply the integral-action controller design procedure Ng (0) ) = 4.5I3 + Ng (0)−1 . Finally, with Q =
s s
of Section III-A to temperature control in RTP systems. 0, the integral-action controller Cˆo = Cgo + D
−1 Cpi =
c
We use the linearized model described in [4], with three (hij ) /dˆ, where dˆ = sd, is a fourth-order controller due
standard tungsten halogen lamps as actuators, and three to the third-order observer-based controller Cgo in the first
temperature sensors. Let x = [x1 x2 x3 ]T denote the stage and the first-order PI block in the second stage.
temperatures. The linearized MIMO system has a state- We also design a standard observer-based controller Ĉ a
space representation (A, B, C, D), where as in Section III-B using the augmented plant description
⎡ ⎤ (25), with n = 3, ny = 3. We use the same state-estimator
−0.0682 0.0149 0
gain L as in (27) and choose the augmented state weighting
A = ⎣ 0.0458 −0.1181 0.0218 ⎦ ,
Qa = diag[ 20I3 , I3 ] , the control weighting R = I3 .
0 0.04683 −0.1008
⎡ ⎤ The state feedback Ka = [ Kx Kξ ] is found using
0.3787 0.1105 0.0229 LQR. The integral-action controller Ĉ a is computed
B=⎣ 0 0.4490 0.0735 ⎦ , C = I3 , D = 0 . using (26). Fig. 5 shows the closed-loop step responses
0 0.0007 0.4177 of the three temperatures individually for Sys(G, Cˆo ) and
We design an integral-action controller following the two- Sys(G, Ĉ a ), with unit steps applied at each of the three
stage procedure of Section III-A; a full-order observer-based inputs. The step response characteristics are very similar
stabilizing controller is chosen for the first stage and a PI- for all three temperatures. The response of Sys(G, Cˆo )
controller for the second block. We choose to design the (solid line) displayed no overshoot and fast rise time
initial controller Cgo as in (21). The estimator is designed (less than 1 sec.) in the absence of actuator saturation.
here by pole placement, with the poles of FL located at Temperature uniformity on the wafer surface is also
{−3, −4, −5}. We use LQR to find a state-feedback gain maintained. The responses for Sys(G, Ĉ a ) are slower
K ∈ IR3×3 such that FK is stable. We choose the state and have 22% overshoot. Step responses of Sys(G, Cˆo )
weighting matrix Q̂ = 20I3 and the control weighting and Sys(G, Ĉ a ) when the integrators are turned off due
matrix R̂ = I3 . Then L, K ∈ IR3×3 are to actuator saturation are shown in Fig. 6. Saturation
⎡ ⎤ nonlinearities included in the control loop saturate at
2.9318 0.0149 0 ±0.3. The step responses of both systems are slowed
L = ⎣ 0.0458 3.8819 0.0218 ⎦ , (27) down because of the actuator saturations, with Sys(G, Cˆo )
0 0.0468 4.8992 displaying faster response. Although integral-action is
⎡ ⎤
4.2308 −0.4739 −0.0610 not available due to saturation, the steady-state errors are
K = ⎣ 0.6725 4.1515 −0.2611 ⎦ . (28) negligibly small for both systems. Another anti-windup
0.0966 0.4407 4.2242 1.5

−1
The controller = K(sI −A+BK +L(C −DK)) L =
Cgo 1
Y11

(nij ) /d, where d = (s + 4.609)(s + 5.921)(s + 6.809), 0.5

is a stabilizing controller for the plant G and it does not
0
have integral-action. We follow Proposition 3.1 to design 0 2 4 6 8 10 12
time (sec)
1.5

- s−1 I ξ-
−Kξ Ĉ a v 1
Y22

r e ? w u y
-e - −Kx -e -?
e- G - 0.5

−6 ?
e - −L
0
0 2 4 6 8 10 12
time (sec)
6 1.5

e C  (sI − A)−1  ?
e 1
6 x̂ 6
Y33

0.5
B
6
0
D  0 2 4 6
time (sec)
8 10 12

Fig. 5. Step responses of Sys(G, Cˆo ) and Sys(G, Ĉ a ).

Fig. 4. Sys(G, Ĉ a ) with augmented observer-based controller. 2594
technique limits the integral value when saturation occurs. 1

The step responses of the two systems are in Fig. 7; the

Y11
0.5
anti-windup gains (Kaw ) of integrator limiting feedback
loops for Sys(G, Cˆo ) and Sys(G, Ĉ a ) are selected as 0.18 0
0 5 10 15
and 1.2, respectively. Slightly faster responses are obtained time (sec)

compared to those in Fig. 6, with zero steady-state errors. 1

o
In the system Sys(G, Cˆo ), where Ĉ = Cgo + D
−1 Cpi ,
c

Y22
the entire PI block Cpi can be turned off when the system 0.5

has input saturation. Since Cgo designed to stabilize G is

0
still active when Cpi is set to zero, the closed-loop system 0 5 10 15
time (sec)
is still stable. The step response of Sys(G, Cgo ) under
1
actuator saturation is in Fig. 8. The response without the PI

Y33
block is reasonably fast and the steady-state error is small, 0.5
although it is not zero due to the absence of integral-action.
These simulation results indicate that Sys(G, Cˆo ) generally 0
0 5 10 15
has better step response than Sys(G, Ca ) in the absence time (sec)

of actuator saturation. The systems displayed very similar

Fig. 6. Step responses of Sys(G, Cˆo ), Sys(G, Ĉ a ) without integrators.
characteristics under actuator saturation: Responses for
both systems slowed down and temperature uniformity on
wafer surface was not maintained; the third channel shows 1

approximately 2 seconds more delay than the other two. Y11

0.5

0
0 2 4 6 8 10 12 14 16 18 20
time (sec)

R EFERENCES 1
Y22

[1] K. J. Åström, L. Rundqwist, “Integrator windup and how to avoid 0.5

it,” Proc. American Control Conference, pp. 1693-1698, 1989.

[2] P. J. Campo, M. Morari, “Achievable closed-loop properties of 0
0 2 4 6 8 10 12 14 16 18 20
systems under decentralized control: Conditions involving the steady- time (sec)
state gain,” IEEE Trans. Automat. Contr., Vol. 39, pp. 932-943, 1994.
1
[3] T. F. Edgar, S. Butler, W. J. Campbell, C. Pfeiffer, C. Bode, S. B.
Hwang, K. S. Balakrishnan, and J. Hahn, “Automatic control in mi-
Y33

croelectronics manufacturing: Practices, challenges and possibilities,” 0.5

Automatica, 36:11, pp. 1567-1603, 2000.

[4] G. F. Franklin, J. D. Powell, and A. Emami-Naeini, Feedback Control 0
0 2 4 6 8 10 12 14 16 18 20
of Dynamic Systems, 4th ed., Prentice Hall, 2002. time (sec)

[5] G. C. Goodwin, S. F. Graebe, and M. E. Salgado, Control System

Design, Prentice Hall, 2001. Fig. 7. Step responses of Sys(G, Cˆo ), Sys(G, Ĉ a ) with limited
[6] J. C. Doyle, R. S. Smith, D. F. Enns, “Control of plants with input integrators.
saturation nonlinearities,” Proc. American Control Conf., pp. 1034-
1039, 1987. 1

[7] A. N. Gündeş, C. A. Desoer, Algebraic Theory of Linear Feedback

Y11

Systems with Full and Decentralized Compensators, Lect. Notes in 0.5

Control and Inform. Sciences, 142, Springer, 1990.
[8] A. N. Gündeş, M. G. Kabuli, “Parametrization of stabilizing con-
0
trollers with integral action,” IEEE Trans. Automatic Control, 44:1, 0 5 10 15
time (sec)
pp. 116-119, 1999.
1
[9] A. N. Gündeş, M. G. Kabuli, “Reliable decentralized integral-action
controller design,” IEEE Trans. Automatic Control, Vol. 46, pp. 296–
Y22

301, 2001. 0.5

[10] N. Kapoor, A. R. Teel, P. Daoutidis, “An anti-windup desaign for

linear systems with input saturation,” Automatica, Vol. 34, No. 5, 0
0 5 10 15
pp. 559-574, 1998. time (sec)
[11] J. M. Maciejowski, Multivariable Feedback Design, Addison-Wesley, 1
1989.
[12] M. Morari, E. Zafiriou, Robust Process Control, Prentice Hall, 1989.
Y33

0.5
[13] K. Ogata, Modern Control Engineering, 3rd edition, Prentice Hall,
New Jersey, 1997.
[14] J. D. Stuber, I. Trachtenberg, T. F. Edgar, “Design and modeling of 0
0 5 10 15
rapid thermal processing systems,” IEEE Trans. Semicond. Manufac- time (sec)

turing, 11:3, pp. 442-457, 1998.

[15] M. Vidyasagar, Control System Synthesis: A Factorization Approach, Fig. 8. Step response of Sys(G, Cgo ) with saturated input.
Cambridge, MA: M.I.T. Press, 1985.

2595