You are on page 1of 10

Systems & Control Letters 44 (2001) 201210

www.elsevier.com/locate/sysconle
Iterative learning control design for Smith predictor
Qiuping Hu, Jian-Xin Xu

, Tong Heng Lee


Department of Electrical Engineering, The National University of Singapore, 10 Kent Ridge Crescent, Singapore 119260, Singapore
Received 24 April 2000; received in revised form 3 April 2001
Abstract
The Smith predictor has been used to improve the closed-loop performance for systems with time delays. This paper
proposes a frequency-domain method to design an iterative learning control to further improve the performance of Smith
predictor controller. For a time-invariant plant with multiplicative perturbations and a Smith predictor controller, we derive
a sucient and necessary condition (which has the same form as that of a general robust performance design problem)
for the iterative process to converge for all admissible plant uncertainties. In addition, the iterative learning controller
under plant uncertainty is designed. An illustrative example demonstrating the main result is presented. c 2001 Elsevier
Science B.V. All rights reserved.
Keywords: Smith predictor controller; Iterative learning control; Multiplicative uncertainty; Convergence analysis
1. Introduction
Most iterative learning control (ILC) schemes are designed to nd purely feedforward action depending
wholly on the previous control performance of an identical task. Although the purely feedforward control
scheme is theoretically acceptable, it is dicult to be applied to real systems without a feedback control due
to several reasons. One of the reasons is that it is not robust against disturbances which are not repeatable
with respect to iterations. Another reason is that the tracking error may possibly grow quite large in the
early stage of learning, though it eventually converges after a number of trials. In addition, an iterative
learning control is designed and analyzed with a mathematical model of the plant. Since modeling errors
are unavoidable, the real iterative learning control system may violate its convergence condition although
the iterative learning control satises the condition for a nominal plant model. Thus in practice, a feedback
control is commonly employed along with the iterative learning control for system robustness enhancement and
better performance [4,5]. In these control schemes, the feedback controller ensures closed-loop stability and
suppresses exogenous disturbances and the iterative learning controller provides improved tracking performance
over a specic trajectory utilizing past control results.
Some robust approaches to iterative learning control have been proposed in the literature. Roh et al. [7]
proposed a design method of learning controller for a class of nonminimum-phase systems and analyzed the
robustness of convergence condition under multiplicative unstructured uncertainty. De Roover [1] synthesized
an iterative learning control based on H

control under unstructured uncertainty. Doh et al. [2] presented a

Corresponding author. Tel.: +65-874-2566; fax: +65-779-1103.


E-mail address: elexujx@nus.edu.sg (J.-X. Xu).
0167-6911/01/$ - see front matter c 2001 Elsevier Science B.V. All rights reserved.
PII: S0167- 6911( 01) 00142- 6
202 Q. Hu et al. / Systems & Control Letters 44 (2001) 201210
Fig. 1. Smith predictor controller in standard form.
sucient condition for robust convergence and robust stability for uncertain linear systems and a synthesis
method for their conditions. Moon et al. [5] proposed a robust approach to iterative learning control design
for a conventional feedback control system. Park et al. [6] designed an iterative learning control algorithm
for a class of linear dynamic systems with uncertain time delays. Most of the existing literature has focused
mainly on systems without time delays. The problem of how to achieve an iterative learning controller for
uncertain plants with time delays has not been suciently tackled. Motivated by this problem, we propose
in this paper a systematic design method to obtain an iterative learning controller that assures convergence
and better performance for all admissible plant uncertainties. Learning control is conducted in a common
conguration of Smith predictor controller and the input update law is given in the frequency domain. A
sucient and necessary condition which has the same form as that of the general robust performance design
problem for the iterative learning control system to converge is derived. In addition, the procedure for the
design of the iterative learning controller under plant uncertainty is given.
This paper is organized as follows. In Section 2, we characterize the Smith controller to which the new
iterative learning control scheme is applicable. The design procedure and simple proof of stability are presented
in Section 3. In Section 4 we illustrate this procedure on a example. Some concluding remarks are given in
Section 5.
2. Problem statement
A standard conguration of Smith predictor controller is shown in Fig. 1. In this gure, y
d
(t) is the
reference trajectory, y(t) is the system output, u(t) is the feedback control input. C(s) is the transfer function
of the Smith predictor controller, and the transfer function of the plant P(s) is composed of a rational stable
transfer function P
0
(s) and a dead time, ,
P(s) =P
0
(s)e
s
; (1)
and a model of P(s) is described by

P(s)

P(s) =

P
0
(s)e

s
; (2)
where

is an estimate of . We assume that the plant P(s) is perturbed and described by the following
transfer functions of the form:
P =(1 +W
2
)

P; (3)
where W
2
is a xed stable transfer function as the weight; and is a variable stable transfer function satisfying

61. It is well known that C(s) permits higher gains to be used for the structure in Fig. 1. This
increase in the permitted gains comes from the elimination of the dead time from the characteristic equation of
the closed-loop system when a perfect model is available. The 2-norm and -norm for the transfer function
G are dened as
G
2
:=
_
1
2
_

|G(j!)|
2
d!
_
1=2
; (4)
Q. Hu et al. / Systems & Control Letters 44 (2001) 201210 203
Fig. 2. Iterative learning control of Smith predictor.
G

:=sup
!
|G(j!)|; (5)
respectively. Note that if G is stable, then by Parsevals theorem
G
2
=
_
1
2
_

|G(j!)|
2
d!
_
1=2
=
__

|G(t)|
2
dt
_
1=2
: (6)
If

61, then

P(j!)

P(j!)
1

6|W
2
(j!)|; !: (7)
This shows that |W
2
(j!)| provides the uncertainty prole. The objective of learning control is to progressively
achieve a perfect tracking which may be written mathematically as
lim
k
Y
d
(j!)e
j!

Y
k
(j!)
2
=0: (8)
Considering the iterative learning system shown in Fig. 2 which consists of a Smith predictor controller and
an iterative learning control input, we have
U
k
(s) =C[(Y
d
Y
k
) (

P
0


P)(U
k
+V
k
)]; (9)
and
Y
k
=P(U
k
+V
k
) +Y
0
k
; (10)
where V
k
(s) and Y
0
k
(s) denote the Laplace transforms of the iterative learning control input and the zero-input
response of the plant at the kth iteration, respectively. From (10), we get
U
k
=(Y
k
Y
0
k
)=P V
k
: (11)
Substituting the above equation into (9) and using (1) and (2) yield
Y
k
(s) =
CP
0
e
s
M
Y
d
(s) +
P
0
e
s
M
V
k
(s) +
1 +C(

P
0


P
0
e

s
)
M
Y
0
k
(s); (12)
where
M =1 +C(P
0
e
s
+

P
0


P
0
e

s
): (13)
204 Q. Hu et al. / Systems & Control Letters 44 (2001) 201210
Dene the sensitivity and complementary sensitivity functions associated with the nominal system
S(s) =
1
1 +C(s)

P
0
(s)
; T(s) =
C(s)

P
0
(s)
1 +C(S)

P
0
(s)
: (14)
Consider the input update law given in the frequency domain as
V
k+1
(s) =W
1
(s)[V
k
(s) +U
k
(s)]; (15)
where U
k
(s) denotes the Laplace transform of the output of C(s) at the kth iteration. W
1
(s) is the transfer
function of the learning controller. The following assumptions are made:
A1. The track signal Y
d
(t) is bounded and piecewise smooth for all t 0.
A2. The initial state of the plant P(s) is invariant with respect to k and equal to zero, i.e., Y
0
k
(s) =Y
0
(s) =0
for k =1; 2; : : : :
A3. The Smith predictor controller provides robust stability, this implies [3],
W
2
(j!)T(j!)

61: (16)
We wish to nd a W
1
(s) such that the iterative learning control system shown in Fig. 2 with the input update
law (15) converges for every plant in the family of (7).
3. Iterative learning control design
Dene the equations
e

(t) =L
1
[E

(s)] =L
1
__
e

CP
0
e
s
M W
1
_
Y
d
_
: (17)
With the help of (3), (13) and (14), we have
e

(t) =L
1
__
e

(1 +W
2
)C

P
0
e

s
1 +C

P
0
+W
2
C

P
0
e

s
W
1
_
Y
d
_
=L
1
__
e

(1 +W
2
)Te

s
1 +W
2
Te

s
W
1
S
_
Y
d
_
; (18)
where L
1
denotes the inverse Laplace operator. The remaining error at kth iteration is
e
k
(t) =L
1
[E
k
(s)] =L
1
[Y
d
e

s
Y
k
]: (19)
Theorem 1. The ILC system is convergent in the sense that
lim
k
e
k
e

2
=0 (20)
with the input update law (15) if and only if
_
_
_
_
W
1
S
1 +W
2
Te
j!

_
_
_
_

1: (21)
Proof. () From (21), it follows that
W
1
S

1 + W
2
Te
j!

: (22)
From this and (18), we can see that E

is well dened. From (19), (12) and Assumption A2, we have


E
k
=
_
e

CP
0
e
s
M
_
Y
d

P
0
e
s
M
V
k
; (23)
Q. Hu et al. / Systems & Control Letters 44 (2001) 201210 205
From the above equation and (17), we get
E
k+1
E

=
_
e

CP
0
e
s
M
_
Y
d

P
0
e
s
M
V
k+1

_
e

CP
0
e
s
M W
1
_
Y
d
=
_
CP
0
e
s
M W
1

CP
0
e
s
M
_
Y
d

P
0
e
s
M
V
k+1
: (24)
Since Y
0
k
=0 (Assumption A2), (11) and (12) can be written as
U
k
=
Y
k
P
0
e
s
V
k
; (25)
Y
k
(s) =
CP
0
e
s
M
Y
d
(s) +
P
0
e
s
M
V
k
(s); (26)
respectively. Substituting (26) into (25) yields
U
k
=
C
M
Y
d

M 1
M
V
k
: (27)
Substituting this into (15) gives
V
k+1
=
CW
1
M
Y
d
+
W
1
M
V
k
: (28)
Substituting (28) into (24) gives
E
k+1
E

=
_
CP
0
e
s
M W
1

CP
0
e
s
M

CP
0
W
1
e
s
M
2
_
Y
d

P
0
W
1
e
s
M
2
V
k
=
W
1
M
__
CP
0
e
s
M W
1
M
W
1

CP
0
e
s
W
1
+
CP
0
e
s
M
_
Y
d
+
_
e

CP
0
e
s
M W
1
_
Y
d
E

P
0
e
s
M
V
k
_
=
W
1
M
__
e

CP
0
e
s
M
_
Y
d

P
0
e
s
M
V
k
E

_
=
W
1
M
(E
k
E

) =
W
1
S
1 +W
2
Te

s
(E
k
E

): (29)
By Parsevals theorem, we have
e
k+1
(t) e

(t)
2
=E
k+1
E

2
6
_
_
_
_
W
1
S
1 +W
2
Te
j!

_
_
_
_

E
k
E

2
: (30)
Consequently, e
k+1
(t) e

(t)
2
0 as k .
() Let !
0
be such that
_
_
_
_
W
1
S
1 +W
2
Te
j!

_
_
_
_

W
1
S
1 +W
2
Te
j!

!=!
0
1: (31)
Select V
0
and Y
d
such that
|E
0
(j!) E

(j!)| =
_

_
c
_

2
if |! !
0
| or |! +!
0
| ;
0; otherwise
(32)
where 1 0 and c 0 are constants. It is known that
e
0
(t) e

(t)
2
=E
0
(j!) E

(j!)
2
=
_
1
2
_

|E
0
(j!) E

(j!)|
2
d!
_
1=2
=c: (33)
206 Q. Hu et al. / Systems & Control Letters 44 (2001) 201210
Then, when becomes very small
e
1
(t) e

(t)
2
2
=E
1
(j!) E

(j!)
2
2
=
1
2
_

W
1
S
1 + W
2
Te
j!

2
|E
0
(j!) E

(j!)|
2
d!

c
2
2
_

W
1
S
1 + W
2
Te
j!

2
!=!
0
+

W
1
S
1 + W
2
Te
j!

2
!=!
0

_
=
_
_
_
_
W
1
S
1 + W
2
Te
j!

_
_
_
_
2

c
2
c
2
=e
0
(t) e

(t)
2
2
: (34)
Similarly, we can show E
2
E

2
E
1
E

2
, etc. This clearly shows that (21) is the necessary condition
for (20).
Remark 1. Theorem 1 has given a necessary and sucient condition such that lim
k
e
k
=e

(t). If e

(t) =0,
we have lim
k
=0. The purpose of this paper is to choose W
1
such that under the condition (21) e

(t) =0
for the interested frequency region of the tracking signal.
Remark 2. From (18), we can see that if W
1
=1 and W
2
=0, i.e., no modeling error exists, then e

(t) =0.
This means that lim
k
=e
k
(t) =0, and therefore no remaining error for the ILC.
Remark 3. If the model is not perfect, we can still choose W
1
such that e

(t) 0 and the convergence


condition (21) of the ILC is satised for the frequency region of interest. In the following, we will discuss
how to select W
1
.
Clearly, the convergence condition should therefore be
W
2
T

1 and
_
_
_
_
W
1
S
1 + W
2
Te
j!

_
_
_
_

1; : (35)
This is the same form as the robust performance condition for basic feedback control system where W
1
is the
performance weighting function [3]. Theorem 1 presents a sucient and necessary condition for the iterative
learning control to converge. However, this result is not directly applicable to iterative learning control design.
It is not clear how to obtain an iterative learning control so that the condition is assured for all admissible
uncertainties. The next theorem gives an equivalent condition for the iterative learning control to converge in
terms of the function
s |W
1
S| +|W
2
T|:
Theorem 2. A sucient and necessary condition for (35) is
|W
1
S| +|W
2
T|

1: (36)
Proof. () Assume (36), or equivalently,
W
2
T

1 and
_
_
_
_
W
1
S
1 |W
2
T|
_
_
_
_

1: (37)
Since
1 =|1 + W
2
Te
j!

W
2
Te
j!

| 6|1 + W
2
Te
j!

| +|W
2
T|; (38)
we have
1 |W
2
T| 6|1 + W
2
Te
j!

|: (39)
Q. Hu et al. / Systems & Control Letters 44 (2001) 201210 207
This implies that
_
_
_
_
W
1
S
1 |W
2
T|
_
_
_
_

_
_
_
_
W
1
S
1 + W
2
Te
j!

_
_
_
_

: (40)
This and (37) yield
_
_
_
_
W
1
S
1 + W
2
Te
j!

_
_
_
_

1: (41)
() Assume that
W
2
T

1 and
_
_
_
_
W
1
S
1 + W
2
Te
j!

_
_
_
_

1; : (42)
Pick a frequency ! where
|W
1
S|
1 |W
2
T|
is maximum. Now pick so that
1 |W
2
T| =|1 + W
2
Te
j!

|: (43)
Now we have
_
_
_
_
W
1
S
1 |W
2
T|
_
_
_
_

=
|W
1
S|
1 |W
2
T|
=
|W
1
S|
|1 + W
2
Te
j!

|
6
_
_
_
_
W
1
S
1 + W
2
Te
j!

_
_
_
_

: (44)
From this and (42) there follows (37).
Remark 4. It is worthwhile to point out that, the tracking convergence of most robust control and ILC methods
are ensured with sucient condition only. This in general leads to a conservative design. In comparison, in
above theorems both the necessary and sucient conditions are provided for ILC convergence.
Returning to (18) again, we consider e

(t). Clearly, if we set


W
1
=1 +
W
2
T(e

s
1)
S
(45)
and the convergence condition is satised, then perfect tracking, i.e., e

(t) =0, is achievable. Though W


2
in (45) is not available in practice, later in this section we will show that W
1
can be realized by means of
using previous cycle system information. In this case, the convergence condition (36) becomes
|S + W
2
T(e
j!
1)| +|W
2
T|

1: (46)
If W
2
=0, i.e., no modeling error exists, the above equation becomes
|S| +|W
2
T|

1: (47)
Theorem 3. A necessary condition for (47) is
W
2
(j!) 1: (48)
Proof. Assume that 1 6W
2

. Then
1 =|S +T| 6|S| +|T| 6|S| +|W
2
T|: (49)
This contradicts (48).
208 Q. Hu et al. / Systems & Control Letters 44 (2001) 201210
Eq. (48) is too restrict a condition to satisfy for most Smith predictor controllers. Therefore, the selection
of W
1
in (45) is not acceptable even in the presence of perfect model. If we choose
W
1
=F
_
1 +
W
2
T(e

s
1)
S
_
(50)
where F(s) is low-pass lter, so that the convergence condition is guaranteed, W
1
will be close to the right
of (45) in the low-frequency. Note that in this case the tracking error will not converge to zero perfectly as
seen in (18). However, the performance degradation caused by the introduction of F(s) may be negligible.
In practice, any reference trajectory to be tracked remains in the low-frequency region, accurate tracking is
normally only needed up to the bandwidth of interest.
Theorem 4. The ILC system show in Fig. 2 with W
1
of (50) is uniformly convergent in the L
2
-norm sense
if
|F|
1 |W
2
T|
|S| + 2|W
2
T|
; !: (51)
Proof. (51) implies
|F|(|S| + 2|W
2
T|) +|W
2
T| 1; !: (52)
Note that
|F|(|S| + 2|W
2
T|) |F|(|S| +|W
2
T||e

s
1|)
|F|[|S + W
2
T(e

s
1)|] =|W
1
S|; !: (53)
Combining (52) and (53) yields (36).
W
1
cannot be constructed directly because the plant uncertainty is involved in it. However, V
k+1
can still
be obtained with the previous plant operation data. Substituting (50) into (15) gives
V
k+1
=F
_
1 +
W
2
T(e

s
1)
S
_
(V
k
+U
k
) (54)
Substituting W
2
in (3) and S; T in (14) into the above equation gives
V
k+1
=F
_
1 +C

P
0
_
P
0
e
s

P
0
e

s
1
_
(e

s
1)
_
(V
k
+U
k
) (55)
From Fig. 2, P
0
e
s
(V
k
+U
k
) =Y
k
, so the above equation changes to
V
k+1
=F
_
1 +C

P
0
(1 e

s
)
_
(V
k
+U
k
) +CF(1 e

s
)Y
k
: (56)
Note that (1 e

s
) is noncausal but (1 e

s
)Y
k
is realizable as it uses previous cycle operation data only.
4. Illustrative example
Consider the plant with the model

P(s) =

k
s + 1
e

s
; (57)
where all three parameters in the model are uncertain
0:8 6k 61:2; (58)
0:7 6 61:3; (59)
0:8 6 61:2: (60)
Q. Hu et al. / Systems & Control Letters 44 (2001) 201210 209
Fig. 3. Bode plot of |F| and (1 |W
2
T|)=(|S| + 2|W
2
T|) (dash).
The multiplicative uncertainty bound W
2
(s) can be obtained as
W
2
(s) =
2s + 1
0:7s + 5
; (61)
which is an increasing function of frequency. The existing Smith predictor controller is a PI controller,
C(s) =2
_
1 +
1
s
_
: (62)
Following above design procedures, we take
F(s) =
1
2s + 1
; (63)
which is a rst-order low-pass lter. As shown in Fig. 3, (51) is satised, so the convergence condition
is assured. Suppose that the desired trajectory is composed of a sinusoidal signal with 0:1 Hz fundamental
frequency and its harmonics, i.e.,
y
d
(t) =3 sin(0:2t) + 0:75 sin(0:4t) + 0:5 sin(0:8t) (64)
and shown in Fig. 4.
Simulations with high representative sample parameters (i.e., k =1:2; =1:3 and =1:2) from the set
(58)(60) are shown in Fig. 5. The learning controller was initialized as V
1
=0, thus the tracking error at
the rst trial was only the result of the Smith predictor controller. By adding the learning control action,
the tracking error is signicantly reduced as iteration number increases. Within expectation, although the ILC
cannot track the higher frequency signals due to the presence of F, from this result, it can be veried that
the performance of the Smith predictor controller can be greatly improved by the added iterative learning
controller.
5. Concluding remarks
An ILC design method for uncertain Smith predictor control systems was presented in this paper. The
iterative learning controller was added to an existing stable Smith predictor control system to improve the
tracking accuracy. We derive sucient and necessary conditions for the ILC system to converge for all
admissible plant uncertainties. The simulation result demonstrated the signicant improvement of tracking
accuracy of Smith predictor control system with the addition of ILC.
210 Q. Hu et al. / Systems & Control Letters 44 (2001) 201210
Fig. 4. Desired trajectory.
Fig. 5. Tracking errors at 1st (dot), 2nd (dash) and 4th (solid) iteration.
References
[1] D. De Roover, Synthesis of a robust iterative learning controller using approach, Proceeding of the 35th Conference on Decision and
Control, 1996, pp. 30443049.
[2] T.Y. Doh, J.H. Moon, K.B. Jin, M.J. Chung, Robust iterative learning control with current feedback for uncertain linear systems,
Internat. J. Systems Sci. 30 (1999) 3947.
[3] J.C. Doyle, B.A. Francis, A.R. Rannenbaum, Feedback Control Theory, Maxwell Macmillan, London, 1992.
[4] T. Kavli, Frequency domain synthesis of trajectory learning controller for robot manipulators, J. Robot. Systems 9 (1992) 663680.
[5] J.H. Moon, T.Y. Doh, M.J. Chung, A robust approach to iterative learning control design for uncertain systems, Automatica 34 (1998)
10011004.
[6] K.H. Park, Z. Bien, D.H. Hwang, Design of an ILC for linear systems with time-delay and initial state error, in: Z. Bien,
J.X. Xu (Eds.), Iterative Learning Control: Analysis, Design, Integration and Application, Kluwer Academic, Netherlands, 1998,
pp. 147164.
[7] C.L. Roh, M.N. Lee, M.J. Chung, ILC for non-minimum phase system, Internat. J. Systems Sci. 27 (1996) 419424.

You might also like