D
A
11
,
D
A
12
,
D
A
21
, and
D
A
22
denote uncertainties;
j
A
R
n
1
þ n
2
denotes the non-repetitive disturbances. The boundary condition
of the Roesser’s system is
x
h
ð0, jÞ
x
v
ði, 0Þ
"#
,
where x
h
ð0, jÞ and x
v
ði, 0Þ represent the horizontal and vertical
boundary conditions, respectively. Detailed introductions on the
Roesser’s system can be found in literature (Kaczorek, 1985).
Definition 1 (Shi et al., 2005b). For any boundary conditions
x
h
ð0, jÞ and x
v
ði, 0Þ and all admissible uncertainties, if unforced
state response of 2D system (1) satisfies
lim
i, j-1
x
h
ði, jÞ
x
v
ði, jÞ
"#
¼
0
0
, ð2Þ
then, the 2D Roesser system (1) is robust asymptotically
stable (RAS).
To evaluate the influence of variations on the tracking error, a
definition is presented below.
Definition 2 (Shi et al., 2005b). The 2D system (1) is robust
asymptotically stable with an H
1
performance index
g
, if:
(a) when
j
0, it is asymptotically stable for all admissible
uncertainties
D
A
i, j
ði, j ¼ 1; 2Þ;
(b) when the boundary condition is zero, JZJ
2
r
g
J
j
J
2
.
The robust H
1
performance indicates the sensitivity of the 2D
system to external disturbances. A smaller value of
g
indicates
lower sensitivity to external disturbances. Therefore,
g
may be
taken as a design parameter to be minimized.
Lemma 1 (Shi et al., 2005b). If there exist a function VðÞ and a
scalar 0o
r
o 1 such that
(a) VðXÞ Z 0 for X A
R
n
1
þ n
2
, and VðXÞ¼0 iff X ¼ 0;
(b) VðXÞ-1 as JXJ-1;
(c) for all boundary conditions,
X
i þ j ¼ I
0
þ J
0
þ q þ 1
I
0
r i r I
0
þ q
J
0
r j r J
0
þ q
V
x
h
ði, jÞ
x
v
ði, jÞ
"# !
r
r
X
i þ j ¼ I
0
þ J
0
þ q
I
0
r i r I
0
þ q
J
0
r j r J
0
þ q
V
x
h
ði, jÞ
x
v
ði, jÞ
"# !
,
8I
0
Z 0, J
0
Z 0, q4 0: ð3Þ
then the system (1) is asymptotically stable.
Theorem 1. If there exist matrices P
1
4 0 and P
2
4 0 such that the
following linear matrix inequality (LMI ) holds
˚
A
T
P
˚
AP o 0, ð4Þ
where
˚
A ¼
4
A
11
A
12
A
21
A
22
"#
; P ¼
4
P
1
0
0 P
2
"#
, ð5Þ
then the 2D Roesser’s system (1) with
j
0 is asymptotically stable.
In addition, if x
h
ð0, jÞ0, there exists a scalar 0o
r
o 1 such that
X
I
0
i ¼ 0
V
P
2
ðx
v
ði, jþ 1ÞÞo
r
X
I
0
i ¼ 0
V
P
2
ðx
v
ði, jÞÞ,
8j4 0, 8I
0
4 0, 8x
v
ði, 0Þ: ð6Þ
Proof. The proof is presented in the Appendix. &
3. Problem formulation
3.1. System description
Consider the following multi-input multi-output (MIMO)
batch process:
xðt þ1, kÞ¼ðAþ
D
AÞxðt, kÞþðBþ
D
BÞuðt , kÞþwðt, kÞ
yðt, kÞ¼Cxðt, kÞþvðt, kÞ
(
xð0, kÞx
0
; t ¼ 0; 1, ..., T1; k ¼ 1; 2, ... ð7Þ
where t is the time index; k is the batch index; xðt, kÞA
R
n
, uðt, kÞA
R
p
,
and yðt, kÞA
R
m
represent, respectively, the states, outputs, and inputs
of the process at time t of the kth batch run; wðt, kÞA
R
n
denotes non-
repetitive load disturbances; vðt, kÞA
R
m
denotes non-repetiti ve mea-
surement noise; A, B,andC are the system matrices with appropriate
dimensions; x
0
is the identical initial condition for each batch;
D
A and
D
B respectively denote uncertain perturbations of matrices A and B,
which can be represented as
D
A ¼
4
F
A
D
A
C
A
,
D
B ¼
4
F
B
D
B
C
B
, ð8Þ
where f
F
A
,
C
A
g and f
F
B
,
C
B
g are known real constant matrices
characterizing the structures of uncertainties; while
D
A
and
D
B
are
uncertainties satisfying
D
T
A
D
A
o I,
D
T
B
D
B
o I: ð9Þ
The control objective is to determine a control law such that
the outputs track the given target, Y
R
ðtÞ, as closely as possible,
even if there exist uncertainties, disturbances, and noises. The
tracking error is defined as
eðt, kÞ¼
4
Y
R
ðtÞyðt, kÞ: ð10Þ
3.2. ILC-based PI control
In this section, a PI-type control is introduced as follows:
uðt, kÞ¼u
0
þK
1
eðt, kÞþK
2
I
e
ðt, kÞ, ð11Þ
where u
0
is the initial value of the control signal and other terms
are defined as below
eðt, kÞ¼
4
y
r
ðt, kÞyðt, kÞ,
I
e
ðt, kÞ¼
4
I
e
ðt1, kÞþeðt, kÞ,
I
e
ð0, kÞ¼
4
0, ð12Þ
Because of their different meanings, eðt, kÞ and
eðt, kÞ are termed as
the tracking error and offset, respectively. The set-point y
r
ðt, kÞ
may be varied in various batches, and it is updated by using ILC ,
as shown below
y
r
ðt, kÞ¼y
r
ðt, k1ÞþL
1
½I
e
ðt1, kÞI
e
ðt1, k1ÞþL
2
eðt þ1, k1Þ
ð13Þ
where L
1
and L
2
are the learning gain matrices. The key idea
behind (13) is optimizing the set-point value using the tracking
error in the previous batch and the variation of the offset integral
in the batch direction. For convenience, the following notation is
introduced:
dW
ðt, kÞ¼
4
W
ðt, kÞ
W
ðt, k1Þ,
W
¼ x, y, u, ... ð14Þ
hence
dW
is the variation of
W
in the batch direction. Then, (13) can
be rewritten as
y
r
ðt, kÞ¼y
r
ðt, k1ÞþL
1
d
I
e
ðt1, kÞþL
2
eðt þ1, k 1Þ: ð15Þ
Because (15) was included, the proposed algorithm consisting of
(11), (12), and (15) is termed as a modified PI controller, or ILC-
based PI controller.
Y. Wang et al. / Chemical Engineering Science 71 (2012) 153–165 155